Nordic Semiconductor has introduced AI-assisted development capabilities intended to support wireless IoT products from early prototyping through fleet-level debugging. The approach is notable because it connects embedded development with operational device data rather than limiting AI support to code generation.

For embedded IoT teams, the difficult work rarely ends when firmware compiles. Board bring-up, SDK migration, production handover and post-deployment troubleshooting often involve different tools, different teams and fragmented context. That gap is especially visible in low-power wireless products, where behavior in the field can depend on a combination of firmware, radio conditions, cloud services and device lifecycle processes.

Nordic Semiconductor is now trying to address that fragmentation by bringing AI-assisted workflows across what it describes as the full IoT device lifecycle. The company says the capability is available today and is designed for developers building wireless IoT products on Nordic’s chip-to-cloud platform.

The announcement is not simply another embedded AI coding assistant. Nordic’s positioning is that its hardware, embedded software, SDK, development tools and cloud lifecycle services can provide a shared context for AI assistance beyond the code editor. According to the company, developers can use their preferred AI assistant through Nordic MCP servers, rather than being forced into a single proprietary front end.

Why this differs from typical AI developer tooling

Most AI tools aimed at embedded engineers focus on code completion, documentation search or generating examples inside an IDE. Nordic is making a different claim: that AI support can follow a product from the first prototype on a development kit into production handover and then into a deployed fleet.

That distinction matters because the hardest embedded problems are often contextual rather than syntactic. A firmware crash on a deployed device is not just a code issue; it may involve SDK version history, board configuration, device logs and cloud-side lifecycle data. By tying AI assistance to Nordic-specific development and fleet context, the company is attempting to make the assistant useful for tasks such as SDK version migration, custom board bring-up and root-cause analysis of devices already in the field.

The practical insight for OEMs is that the value of this model depends on continuity. If engineering teams use Nordic’s SDK during development but manage deployed devices through disconnected operational tools, the AI assistant will have less lifecycle context to work with. Conversely, projects that use enough of Nordic’s chip-to-cloud environment may benefit from a more consistent troubleshooting path from lab bench to field issue.

Implications for IoT product teams

For OEMs building low-power wireless devices, the main impact could be a reduction in the friction between early prototyping and later maintenance. Nordic says developers can move from idea to proof of concept on a Nordic development kit more quickly, and that AI assistants can produce more accurate results in fewer iterations, reducing token cost and improving code reliability.

For system integrators and enterprises deploying connected products, the more interesting part is post-deployment debugging. If AI-assisted root-cause analysis can be performed within the same development workflow used to build the device, field support teams may be able to escalate issues with more usable technical context. That does not remove the need for embedded expertise, but it may reduce the time spent reconstructing how a device was built, configured and updated.

Connectivity providers are not the direct target of the announcement, but the lifecycle angle is relevant to them as well. Wireless IoT failures are often blamed on connectivity even when the underlying cause sits in firmware, device configuration or cloud integration. A development environment that can combine device-side and cloud-side context may help clarify where responsibility lies during incident analysis.

For the broader IoT ecosystem, Nordic’s move reflects a shift in how semiconductor vendors compete. Low-power wireless suppliers are no longer differentiated only by radio silicon or SDK breadth. Increasingly, they are packaging hardware, embedded software, cloud services and lifecycle management into a developer experience. Nordic’s AI-assisted workflow is a continuation of that platform strategy, with AI used as an interface across the stack rather than as a standalone feature.

The announcement should still be viewed with appropriate caution. Nordic has not disclosed performance benchmarks, deployment scale or quantified productivity gains. What it has introduced is an architectural approach: using AI assistance across Nordic’s connected development and lifecycle environment, while allowing developers to work with the AI assistant they already use. For IoT teams evaluating embedded platforms, that architectural choice may prove more significant than the AI label itself.

The post Nordic Extends AI Assistance from Firmware Development to Deployed IoT Fleets appeared first on IoT Business News.

Every connected device a business runs is a workload someone has to keep healthy, secure, and patched. With 21.1 billion connected IoT devices online by the end of 2025 and a path to 39 billion by 2030, the spreadsheet-and-script approach that worked for a few hundred laptops is no longer viable. Gartner now expects more than half of organizations to adopt autonomous endpoint management by 2029. The shift is not a tooling upgrade. It is a change in how IT teams operate across the entire estate, from corporate laptops to industrial sensors.

Key Takeaways

• Autonomous endpoint management uses AI and policy-driven automation to handle device tasks that previously required manual technician work.
• The global IoT installed base is on track to nearly double this decade, well beyond what manual operations can keep up with.
• Five workflows benefit most: patch deployment, device onboarding, compliance enforcement, incident response, and software lifecycle management.
• Organizations consistently flag security as the leading obstacle to expanding connected device deployments, and intelligent automation is one of the clearest paths forward.
• Successful adoption depends on a complete asset inventory, clear policy intent, and a phased rollout, not on replacing the IT team.

From Manual Tickets to Intelligent Workflows

Traditional endpoint management is a queue of human-driven tickets. A new device joins the network and someone configures it. A vulnerability is disclosed and someone applies the patch. A user reports a slow machine and someone investigates. The model worked when a typical estate held hundreds of laptops in a single office. It breaks down when the same team is responsible for laptops, kiosks, edge gateways, and thousands of sensors across multiple sites. A working definition of autonomous endpoint management for IT automation  replaces that ticket queue with a continuously running platform that detects state changes, decides what to do based on pre-defined policy, and acts without waiting for a human in the loop.

The shift is more philosophical than technical. The system still does the same things a skilled administrator would do. It just does them at machine speed and at full scale. Industry context for that wider operational shift, from connectivity-only deployments to integrated infrastructure, is captured well in coverage of how the IoT industry moved through 2025, as IT, OT, and security functions converge.

The Scale Problem Driving the Shift

The numbers behind the manual-to-autonomous transition are direct. Connected device counts have crossed thresholds that human-paced operations cannot keep up with.

The chart understates the operational pressure. Each of those billions of endpoints generates state changes throughout the day: configuration drift events, security agent status updates, patches released, services failing, disk thresholds breached. A fleet of just 5,000 endpoints will produce thousands of such signals daily. Multiply that across the broader IoT and IT estate, where 67 percent of organizations already cite security as the top barrier to scaling deployments, and the case for intelligent automation makes itself.

Five Workflows Where Autonomous Management Pays Off First

Some IT processes deliver returns immediately when intelligent automation takes over. The five below are where most organizations see measurable change within the first quarter.

• Patch deployment. What used to take a technician hours per batch becomes minutes per endpoint, with consistent application across operating systems, third-party apps, and firmware.
• Device onboarding. Zero-touch provisioning means a new laptop, kiosk, or sensor enrolls itself, downloads its baseline configuration, and reports as compliant before a human ever logs in.
• Continuous compliance. Instead of quarterly audits that catch drift after the fact, compliance becomes a real-time operating state with audit-ready logs available on demand.
• Incident response. Suspicious behavior on an endpoint triggers automatic isolation, evidence capture, and ticket creation, often before the security team sees the alert.
• Software lifecycle. Installations, updates, and retirement happen on a schedule the platform enforces, not on a calendar the technician keeps.

The chart shows indicative time savings from industry case studies. The pattern is consistent across organizations: tasks that ate the morning of a senior administrator now run in minutes in the background, and the administrator’s day shifts to higher-judgment work.

Why This Matters Most for Connected and IoT Estates

Pure laptop fleets are challenging enough. The picture gets harder once a business runs a mixed estate of corporate endpoints alongside industrial sensors, point-of-sale terminals, medical devices, building controllers, or fleet telematics. Many of those devices were never designed for traditional endpoint agents. They have limited update windows, run unsupported operating systems, or sit on networks where a failed patch means a production line stops. The Fortinet primer on IoT security highlights why patching and updating connected devices is essential, especially in operational technology environments where attackers actively target unpatched edge devices.

Autonomous management addresses this by treating every connected device as a managed endpoint, with policies tuned to that device class. A sensor on a manufacturing line gets a different patching window and a different remediation rule than the office laptop two rooms away. Federal guidance now codifies parts of this approach: the NIST IR 8259 series on IoT device cybersecurity sets out a baseline of capabilities that connected devices should support so they can actually be governed at scale, including device identity, secure software updates, and data protection.

Manual vs Autonomous Operations, Side by Side

The differences sharpen once they are laid out by operational dimension rather than by feature.

Video: How Autonomous Endpoint Management Works in Practice

A short product walkthrough covering the policy-driven workflow model in a real deployment. Useful for IT leads and operations decision-makers planning a phased rollout.

A Phased Adoption Roadmap

No serious deployment flips from manual to autonomous overnight. The four-stage path below mirrors what most organizations actually follow.

1. Inventory and visibility. Build a single, real-time view of every endpoint, including industrial and embedded devices that may not appear in traditional CMDB systems.
2. Policy definition. Translate the team’s existing operational practices into written policies the platform can enforce. Patching cadence, compliance baseline, incident playbooks.
3. Pilot on a contained scope. Pick one team or one site, usually IT helpdesk endpoints, and run for four to six weeks before expanding.
4. Expand by device class. Add laptops first, then servers, then IoT and edge devices. Tune policies per class as you go.

Common Pitfalls to Avoid

The patterns below appear in most failed or stalled adoptions. They are easier to design around at the start than to fix later.

• Treating it as a tooling project. The platform is the easy part. The hard work is writing the policies that capture organizational intent and getting cross-team agreement on them.
• Skipping the inventory step. A system cannot manage what it cannot see. Shadow IoT devices on the network are a recurring source of breach exposure.
• Letting automation outpace governance. Every autonomous action needs a documented owner, a rollback path, and a logged audit trail. Without that, autonomous becomes a euphemism for ungoverned.
• Forgetting the IoT specifics. OT and IoT devices have stricter uptime and safety constraints. Policies designed for laptops will break them.

FAQs

How is this different from traditional unified endpoint management?

Unified endpoint management gives administrators tools to act on devices from a single console. The autonomous version gives the platform the authority to execute those actions itself, guided by pre-set rules rather than by a technician pushing the button. UEM is the foundation. AEM is the layer above it.

Does autonomous management replace IT teams?

No. It shifts what IT teams spend their time on. Routine execution moves to the platform, while architecture, governance, exception handling, and strategic projects move to people. Most adopters report freeing 20 to 30 percent of engineering capacity, not eliminating roles.

Can it manage IoT and OT devices alongside laptops?

Modern platforms increasingly do, especially for IoT devices that support the NIST IR 8259A capability baseline or similar standards. Truly legacy OT systems often still require specialized OT-focused tooling that operates alongside the AEM platform rather than replacing it.

What does a realistic rollout timeline look like?

A pilot on one team typically runs about a month and a half. Expansion to the full corporate endpoint estate usually takes three to six months. Extending to IoT and edge device classes adds another quarter or two, depending on inventory completeness and policy maturity.

How does it interact with existing security tools?

Most platforms integrate with existing SIEM, EDR, and ticketing systems through APIs. Autonomous workflows feed those tools richer context and faster signals, while the security stack continues to handle threat detection and response. The two are complements, not substitutes.

What is the most common reason adoption stalls?

Lack of an accurate asset inventory. The platform cannot autonomously manage devices it does not know exist, and most organizations underestimate how many shadow endpoints sit on their networks. Time spent on inventory at the start saves months of retrofitting later.

The Bottom Line

The argument for intelligent, policy-driven endpoint management is operational, not theoretical. Connected device counts are growing past the point where any team can keep up by hand, and the cost of a single missed patch keeps climbing as attackers automate their side of the equation. Industry coverage of smart manufacturing and industrial IoT makes the same point from the operations side: the architectures winning out are the ones that treat each networked asset as governed and continuously maintained. Autonomous endpoint management is how the IT and IoT sides of that conversation finally meet.

This article compares five leading HIPAA compliance software platforms for healthcare organisations.

1. Vanta: best for automation-first healthcare tech teams (especially Business Associates)

Vanta is a trust management and compliance automation platform built for teams that want HIPAA to run like an always-on system check, not a once-a-year scramble. It is a strong fit for cloud-native healthcare SaaS companies and other Business Associates handling ePHI that also need to scale into SOC 2, ISO 27001, or HITRUST over time.

HIPAA coverage note: Vanta supports the HIPAA Security Rule and Breach Notification Rule, but it does not cover the HIPAA Privacy Rule. That distinction matters. If you are a Covered Entity (like a hospital system, health plan, or clearinghouse) and need Privacy Rule workflows in the platform, you will likely need a different tool.

Where Vanta stands out is automation depth. It connects to 400+ cloud and DevOps services and runs tests on a frequent cadence (about every 1 to 2 hours), using a HIPAA program mapped to 73 controls with roughly 123 automated and manual tests. In practice, that means you can continuously verify common requirements like MFA, encryption, access provisioning, and device posture, receive instant alerts whenever a control test fails—a workflow detailed in Vanta’s risk tracking module—and route issues into tools like Jira for remediation.

Vanta also covers the administrative side that tends to eat up time:

• Policies: 18 total policies, including 6 HIPAA-specific, with tooling to customize and manage updates.
• Training: Built-in HIPAA training is included, with an option to integrate with KnowBe4 if you want deeper security-awareness content.
• Vendor and BAA tracking: BAAs and vendor risk can be managed through Vanta’s vendor risk management workflows, so third-party compliance does not live in scattered folders.
• Breach readiness: Includes templates and workflows aligned to breach-notification obligations for Business Associates.

If you are running more than one framework, Vanta’s control mapping can materially reduce rework. For many teams, HIPAA overlaps meaningfully with SOC 2, ISO 27001, and HITRUST, so you can reuse evidence and controls rather than rebuilding your program from scratch.

Implementation and audit readiness: HIPAA in Vanta is self-attested, so there is no external HIPAA audit timeline to manage. Teams starting from zero typically get stood up in a few weeks to a few months, and organisations with an existing SOC 2 program can often move faster because a portion of controls are already satisfied.

Pricing: HIPAA can be included as a package framework or priced as a $5,000 per year add-on, with total first-year costs commonly landing in the $10,000 to $15,000+ range depending on company size and add-on modules.

Pros: deepest automation in this list (400+ integrations and frequent test cycles), strong cross-framework reuse if you are doing HIPAA plus SOC 2 or HITRUST, and self-attestation avoids audit fees and scheduling bottlenecks.

Cons: no Privacy Rule support (a deal-breaker for many Covered Entities), no native EHR integrations like Epic or Cerner out of the box, and it can be overkill for small clinics that do not run a cloud-heavy stack.

Customer proof: Hummingbird Healthcare achieved SOC 2 Type 1 plus HIPAA in 3 months. Other reported outcomes include Modern Health saving 100+ hours annually, Vibrent Health reducing vendor review time from 100 hours to a few hours per week, and ITx Companies seeing 41 per cent of HIPAA controls pre-populated from an existing SOC 2 program.

2. Compliancy Group (The Guard): best for clinics that want hands-on coaching

Compliancy Group’s platform, The Guard, is built for healthcare organisations that want a guided path to HIPAA compliance with a real person in the loop. If your biggest bottleneck is not tooling, but knowing what to do next and how to document it correctly, this is one of the most straightforward options on the market.

Best for: small to mid-sized healthcare practices that want a step-by-step workflow and ongoing support, especially teams without dedicated IT or compliance staff.

Unlike many “compliance automation” platforms that focus primarily on technical evidence collection, Compliancy Group emphasises complete HIPAA program coverage. The Guard supports the Security Rule, Privacy Rule, and Breach Notification Rule, and also offers an OSHA add-on for healthcare organisations.

Core capabilities centre on helping you build and maintain the administrative backbone HIPAA expects:

• Security Risk Analysis (SRA): guided risk assessments with corrective action planning, typically completed in 30 days or less on average (vendor case-study data), with your coach helping you keep momentum.
• Policies and procedures: a library of 500+ templates you can customize to your environment.
• Workforce training: built-in training with completion tracking, so training records are not trapped in spreadsheets.
• Vendor and BAA tracking: tools to manage vendors and agreements, plus reminders around renewals.
• Incident management: workflows to document and track incidents and potential HIPAA violations.

Automation depth: high for documentation workflows, training tracking, and program management. It is not designed for real-time technical monitoring of your infrastructure (for example, continuously verifying MFA, encryption settings, or cloud configuration drift). If your main goal is automated technical evidence collection across cloud systems, you will still need additional security tooling or a different platform category.

Implementation and rollout: many organisations use the coach-led workflow to complete their initial SRA quickly, then expand into policies, training, vendor management, and incident documentation over the next 1 to 3 months depending on size and complexity.

Pricing: Compliancy Group introduced modular pricing in May 2025 starting at $99 per month, letting practices choose the pieces they need. Previous “full suite” pricing was often positioned closer to the mid-hundreds per month, so the new packaging is a meaningful shift for smaller clinics.

Pros: dedicated coach support throughout the process, full HIPAA coverage including the Privacy Rule, and a large policy template library backed by long healthcare compliance experience.

Cons: limited technical integrations and no continuous infrastructure control testing, plus a coach-driven model that can feel slower for teams that prefer fully self-serve execution.

Stand-out differentiator: the assigned live compliance coach. For many clinics, that is the difference between “we bought software” and “we finished the program.”

Customer proof: Compliancy Group positions itself as serving 4,000+ organisations and cites a 100% client audit pass rate claim, alongside strong category positioning on G2 for healthcare compliance.

3. Accountable HQ: best for a self-serve, tiered HIPAA program that grows with you

Accountable HQ is a practical choice when you want a single portal for HIPAA basics, but you are not ready for an enterprise GRC rollout. It is built for clinics and healthcare startups that prefer a self-serve workflow, with higher tiers adding more security-forward features as your program matures.

Best for: small to mid-sized practices and digital health teams that want full HIPAA coverage (including the Privacy Rule) with a clear upgrade path from “baseline compliance” to more proactive monitoring.

Accountable HQ covers HIPAA Security, Privacy, and Breach Notification requirements, and bundles the day-to-day components most teams need to prove they are operating a real program:

• Security risk assessment: a guided Security Risk Assessment workflow included in all plans.
• Policies and procedures: policy generation and management tools across tiers.
• Training: HIPAA training plus security awareness training in the Basic tier, with additional courses available in higher tiers.
• BAA and vendor management: BAA management is included, and the Plus tier adds vendor discovery and shadow IT detection.
• Incident and breach readiness: incident-response tooling is included, with the Plus tier adding data breach monitoring.

Automation depth: moderate. Accountable HQ includes an AI Compliance Copilot across all tiers, and the Plus tier adds more “push-button” security workflows like phishing simulation, MFA and access controls review, and data breach monitoring. The Pro tier goes further with vulnerability scanning twice per year and penetration testing once per year. What it does not offer is the kind of deep, always-on evidence collection you get from platforms built around large-scale cloud integrations.

Implementation timeline: Accountable HQ advertises an average of 30 days to compliance (vendor claim), and you can start immediately via a 7-day free trial.

Pricing: Accountable HQ uses a tiered subscription model with included employee counts and per-seat add-ons:

• Basic HIPAA: $169/month on annual billing ($199/month monthly), includes 15 employees, then $9 per additional seat
• Plus: $254/month annual ($299/month monthly), includes 15 employees, then $15 per additional seat
• Pro: $679/month annual ($799/month monthly), includes 20 employees, then $19 per additional seat
  Month-to-month pricing is listed as higher than annual, and plans are positioned as cancel-anytime.

Pros:

• Full HIPAA rule coverage, including the Privacy Rule, which makes it viable for Covered Entities
• Strong value in the Plus tier for the price point, including phishing simulation, vendor discovery, and breach monitoring
• Clear pricing and packaging, with a fast way to trial the product

Cons:

• Not a multi-framework platform (no SOC 2, ISO 27001, or HITRUST program mapping)
• Limited deep technical integrations compared to cloud-native compliance automation platforms
• Per-seat pricing can climb quickly as you scale headcount

Stand-out differentiator: the Plus tier bundles several proactive security features that many HIPAA tools reserve for higher-priced plans, including phishing simulation, vendor discovery/shadow IT detection, MFA review, and data breach monitoring.

Customer proof: Accountable HQ states 10,000+ companies use the platform (vendor claim) and positions the product around a 30-day average time to compliance (vendor claim), with “audit protection” included in plans.

4. HIPAA One (Intraprise Health): best for audit-grade risk analysis

HIPAA One, now part of Intraprise Health, is built for organisations that need a defensible, auditor-ready Security Risk Analysis (SRA) and want the output to match what regulators actually look for. This is less about lightweight policy wizards and more about producing a risk analysis that stands up under scrutiny across multiple facilities, business units, and affiliates—essentially functioning as a comprehensive risk assessment and management software approach.

Best for: hospitals, health systems, and multi-site networks that want an OCR-aligned SRA with enterprise reporting, weighted scoring, and roll-up visibility.

HIPAA One’s core strength is that its SRA workflow mirrors the OCR audit protocol closely, and it is grounded in NIST methodology (including NIST SP 800-66 alignment). That gives compliance and security leaders a clear line from “requirement” to “evidence” to “remediation plan,” which is exactly what you need when leadership asks, “Are we audit-ready?”

What it covers depends on the modules you deploy, but the platform supports full HIPAA program needs across:

• Security Rule: the SRA experience, including automated risk calculation, prioritisation, and remediation planning
• Privacy Rule and Breach Notification: supported through dedicated modules (for example, Privacy and privacy/breach risk assessment functionality)
• Business associate workflows: contract and agreement management via a Business Associate Manager (BAM) capability
• Workforce training: a training module is available, with progress tracking and reporting

On the automation front, HIPAA One is strong at streamlining assessments and enterprise coordination. It can accelerate year-over-year work by carrying forward prior assessment data, and it supports parent-child roll-ups so multi-entity organisations can view results at the facility level and the system level. It is not positioned as a “connect to every cloud service and test controls hourly” platform. The automation is primarily assessment workflow, scoring, and reporting, not continuous technical control testing across your infrastructure.

Implementation: timelines vary by delivery model. Intraprise Health offers self-service, hybrid, and managed services approaches, which lets organisations choose between software-led execution and deeper expert involvement. Case-study data cited in the draft suggests meaningful time reductions after rollout (for example, a reported 65 per cent reduction in SRA preparation time in one deployment).

Pricing: enterprise pricing is typically quote-based, and overall cost depends on modules and whether you choose managed or hybrid services.

Pros:

• OCR-audit-protocol alignment and NIST-based approach create a more defensible SRA
• Built for multi-site complexity, including roll-up reporting across sub-entities
• Flexible delivery models (self-service, hybrid, managed) to match internal resourcing

Cons:

• Enterprise packaging and services can make total cost high for clinics and small practices
• Monitoring is largely compliance-process and assessment focused, not real-time infrastructure scanning
• Some functionality may be modular depending on your package, which can increase complexity during procurement

Stand-out differentiator: HIPAA One is purpose-built to generate an SRA in the format and depth auditors expect. If your priority is “audit-grade SRA with enterprise reporting,” it is one of the most direct fits in this list.

Customer proof: Intraprise Health positions HIPAA One as used by 16,000 users across 10,000+ healthcare organisations, and cites a 100% OCR acceptance rate claim, along with additional case-study improvements in SRA efficiency (vendor-stated metrics).

5. Clearwater IRM|Pro: best for enterprise-scale risk governance (plus managed security)

Clearwater IRM|Pro is built for healthcare organisations that need more than a HIPAA checklist. It is a fit when your program spans thousands of assets, multiple facilities, medical devices, and third parties, and you want a single partner that can deliver both the platform and the expertise to run it.

Best for: enterprise health systems, IDNs, hospital chains, and large practice management groups that want healthcare-specific risk modelling and the option to pair it with advisory services and managed security.

Clearwater’s HIPAA coverage is delivered through a suite of modules designed to map to the real shape of a healthcare compliance and security program:

• IRM|Analysis: Security Risk Analysis (SRA) aligned to NIST and designed to be OCR-quality, covering ePHI assets and medical devices.
• IRM|Security: Security Rule compliance assessment workflows.
• IRM|Privacy: Privacy Rule and Breach Notification Rule coverage.
• IRM|405(d) HICP: alignment to the industry-recognised cybersecurity practices published under HICP.

Where Clearwater differs from lighter HIPAA tools is in how it treats “continuous monitoring.” The software supports enterprise-wide risk calculation, prioritisation, and executive reporting, but the always-on component comes from Clearwater’s broader delivery model. Clearwater also offers managed security services with a 24/7 SOC, plus managed cloud services for Azure environments. For CISOs, that means you can combine compliance reporting, risk remediation planning, and active security operations under one vendor relationship.

Automation depth: high for enterprise risk modelling and reporting, and operationally continuous when paired with the managed services layer. This is not a self-serve compliance automation product built around hundreds of plug-and-play integrations. It is a healthcare-focused platform that becomes most valuable when used alongside Clearwater’s advisory and managed security capabilities.

Multi-framework support: Clearwater’s software is healthcare and HIPAA centred, with additional alignment to NIST and 405(d) HICP. Broader frameworks like HITRUST and SOC 2 are typically supported through Clearwater’s compliance services rather than out-of-the-box cross-framework control mapping.

Implementation: expect a multi-month rollout for enterprise organisations. Deployment usually includes discovery and inventory, risk analysis, remediation planning, and establishing ongoing governance rhythms. Managed services run continuously once engaged.

Pricing: Clearwater is positioned at a six-figure total cost of ownership and is sold through a consultative process. The expert research notes an estimated annual investment in the $150k to $500k+ range for a mid-size health system depending on scope, modules, and services.

Pros:

• Deep healthcare-specific risk modelling across servers, IoMT, third-party portals, and medical devices
• Full HIPAA program coverage via dedicated modules, including Privacy and Breach Notification support
• Option to pair compliance governance with a 24/7 SOC and managed security services for a unified operating model

Cons:

• Cost and scope make it impractical for small clinics and early-stage startups
• Value depends on time and engagement; it is not “buy it and you are done” software
• Less oriented toward plug-and-play cloud evidence collection compared to automation-first compliance platforms

Stand-out differentiator: Clearwater is the only option in this list that combines enterprise compliance software with a full managed security practice, including a 24/7 SOC. If you want a platform plus a partner to help operate the program, that is the defining advantage.

Customer proof: Clearwater cites 500+ customers, 20+ years focused on healthcare cybersecurity, and recognition including 2026 Best in KLAS for Security & Privacy Consulting, Black Book #1 (survey of approximately 2,000 executives), and MSSP Alert Top 250, alongside a 100% OCR success rate claim (vendor-stated metrics).

Quick-scan comparison

Use this table to narrow your shortlist fast, then validate fit in demos based on your HIPAA rule coverage needs (especially Privacy Rule), automation expectations, and budget model.

Conclusion

HHS’s January 2026 draft HIPAA Security Rule update would make full encryption and multi-factor authentication (MFA) mandatory for every system that touches ePHI. The final text is expected later this year, with a 180-day compliance window, so you will need proof fast, not promises.

The threat side is moving just as quickly. Ransomware hit small providers six times more often in 2025 than in 2021, and the average healthcare breach now tops USD 10.9 million. Continuous monitoring is often cheaper than a single incident response.

Practical steps to stay ahead:

1. Embed continuous risk monitoring. Connect your compliance platform to EHRs, cloud accounts, and mobile-device managers so drift triggers an alert, not a post-breach report.
2. Run quarterly tune-ups. Block two hours each quarter to review the live risk dashboard, close red items, and export an audit snapshot. Four short sprints beat one frantic year-end scramble.
3. Audit MFA and encryption coverage now. When the Security Rule is finalised, you will need evidence that every endpoint and user meets the standard.
4. Map HIPAA controls to a second framework. Aligning with NIST CSF or HITRUST today earns “recognised security practices” safe-harbour credit if an OCR investigation follows a breach.

Next move: use the evaluation checklist above, pick two platforms that match your size and tech stack, and schedule demos this week. A small investment now can help you avoid seven-figure losses—and many sleepless nights—later in 2026.

The post 5 Best HIPAA Compliance Software for Healthcare: Secure, Audit-Ready Platforms appeared first on IoT Business News.

The LoRa Alliance has published a three-year technical roadmap covering application integrations, onboarding, network interfaces, coverage extensions and certification. The plan points to a shift from basic LoRaWAN connectivity toward easier deployment and lifecycle management at ecosystem scale.

For many large IoT deployments, the radio link is no longer the hardest part. The more persistent friction often sits elsewhere: device onboarding, payload decoding, server-to-server integration, network migration and coverage gaps at the edge of infrastructure. That is the context in which the LoRa Alliance’s new roadmap should be read.

The organization, which develops and promotes the LoRaWAN standard, has outlined a set of technical work items running through 2028. Rather than focusing only on the air interface, the roadmap targets the practical layers that determine whether LoRaWAN devices can be deployed, moved, connected to applications and managed without extensive custom engineering.

A roadmap focused on interoperability beyond the radio layer

What makes this announcement distinct from many LPWAN updates is its breadth across the LoRaWAN operating model. The roadmap covers application-level data formats, industrial and utility protocol mappings, onboarding, infrastructure discovery, standardized interfaces between network components, mobile collection models, satellite discovery, crypto agility, gateway certification and network analytics.

That combination matters because LoRaWAN’s ecosystem includes public networks, private networks, gateways, network servers, application platforms and device makers from many suppliers. In such an environment, the value of a standard is not only that devices can communicate over long distances at low power, but that the surrounding infrastructure can be assembled with less bilateral integration work.

Among the application integration items, the Alliance points to work on a mapping structure between LoRaWAN and OPC UA, the industrial interoperability framework widely used in smart industry environments. It also plans support for water meters using the North American UI-1203 protocol. In 2028, the roadmap adds a Standard Application Data Format intended to standardize application codec payload structure so devices and application platforms can interoperate with less custom integration.

The practical implication is clear: LoRaWAN is being positioned not just as a connectivity option for sensors, but as a transport that can fit more predictably into existing industrial and utility data environments. For OEMs, that can reduce the need to build different payload handling approaches for each application platform. For system integrators, standardized codecs and protocol mappings could reduce project-specific translation work, although adoption will still depend on implementation by vendors across the stack.

Device lifecycle management moves into focus

The 2026 and 2027 plug-and-play work items address another operational issue: what happens after devices are deployed. The roadmap includes features to support migration of connected devices from one LoRaWAN network to another, along with End-Device Capabilities Discovery, which would allow a network server to download device capabilities from external servers rather than relying only on manual provisioning.

That is a significant lifecycle-management signal. LoRaWAN deployments often involve long-lived assets, and the ability to move fleets between networks can matter when ownership, service contracts or coverage arrangements change. The derived impact is that connectivity providers may face greater expectations for portability and standardized handling of device capabilities, while enterprises could gain more flexibility over the life of a deployment.

In 2027, the Alliance plans further zero-touch onboarding enhancements and DNS-based network infrastructure discovery. It also plans standardized interfaces between network servers and gateways, and between network servers and application servers. If adopted across products, those interfaces could make it easier to mix gateways, network servers and application servers from different suppliers without bespoke API development.

Coverage extensions reflect real-world deployment patterns

The roadmap also acknowledges that fixed network coverage is not always the deployment model. A planned Walk-By/Drive-By Reading extension in 2026 is designed to let LoRaWAN devices connect efficiently to mobile base stations mounted on vehicles, carried by hand or flown on drones. That is especially relevant for assets outside fixed infrastructure coverage, and it aligns with utility-style collection models where periodic contact may be sufficient.

Satellite Discovery Enhancements, also planned for 2026, will standardize how commercial off-the-shelf LoRaWAN end devices discover LoRaWAN satellite constellations, building on existing support for LEO and GEO satellite use. This does not make every LoRaWAN device a satellite device, but it clarifies an important interoperability step for extending reach beyond terrestrial networks.

Looking further ahead, the roadmap includes crypto agility in 2027, gateway certification in 2027 and a Network Analytics API in 2028. For industrial players and enterprises, these items point to a more mature operational framework: security mechanisms that can accommodate future cryptographic suites, more formal treatment of gateway conformance, and standardized visibility into traffic patterns for network management.

The broader relevance for the IoT market is that LPWAN competition is increasingly about integration economics, not just coverage claims or device battery life. By targeting onboarding, APIs, payload formats and lifecycle processes, the LoRa Alliance is addressing the parts of deployment that often determine total cost and supplier flexibility. The roadmap’s success will depend on how consistently vendors implement the resulting specifications, but its direction is a notable step toward making LoRaWAN deployments less dependent on custom integration at each layer of the stack.

The post LoRa Alliance Sets Three-Year Plan to Make LoRaWAN Easier to Integrate and Operate appeared first on IoT Business News.

Origin Energy is working with Landis+Gyr to add IoT-based smart gas capabilities to existing metering assets in Australia, aiming to enable remote readings and more timely usage data without replacing meters.

Gas metering has often lagged electricity in the move to connected infrastructure, partly because gas meters are widely distributed, battery-dependent and typically harder to justify for full hardware replacement. That makes retrofit approaches particularly important: they can digitise field assets while avoiding the cost and disruption of a conventional meter swap programme.

Against that backdrop, Origin Energy and Landis+Gyr are moving ahead with a smart gas deployment across Origin’s gas network in Australia. Landis+Gyr will deploy intelligent IoT modules, communications technology and a data management platform across Origin’s existing metering assets over an 18-month period. The project covers households and businesses served by Origin’s gas network and is described by the companies as one of Australia’s first large-scale efforts to digitise gas network operations across an entire customer base.

A retrofit model, not a meter replacement programme

The important detail is not simply that gas meters are being connected. It is how they are being connected. Landis+Gyr’s approach is based on adding IoT modules to existing gas meters, enabling remote meter readings and near real-time data insights while leaving the underlying metering assets in place.

That makes this announcement distinct from many smart metering projects, which are often framed around new meter rollouts or broader advanced metering infrastructure replacements. Here, the emphasis is on extending the digital life of installed assets. For a gas network, that distinction matters: reducing the need to replace physical meters can simplify customer access requirements, limit installation disruption and preserve capital already invested in field hardware.

The companies also state that the upgrade will be delivered without disruption to customers’ LPG supply. That point is operationally significant. In utility IoT, the technical value of connectivity is only part of the equation; the field process, customer scheduling and service continuity can determine whether a deployment scales smoothly.

Why the data layer matters

The immediate customer-facing outcome is straightforward: fewer manual reads and fewer estimated bills. But the larger IoT implication is the move from periodic, labour-intensive data collection toward a more automated operating model for gas usage information.

For utilities and energy retailers, remote meter readings can improve billing timeliness and data accuracy. For system integrators, the project highlights the growing role of the data management layer in utility IoT deployments. Adding modules to meters is not enough; the value depends on how reliably readings are collected, transmitted, processed and made available to operational systems.

A practical insight from this architecture is that retrofitting reduces one type of complexity while increasing the importance of another. It avoids a wholesale meter replacement programme, but it places more emphasis on compatibility with existing assets, installation procedures, communications reliability and lifecycle management of add-on IoT devices. Those are familiar issues in industrial IoT, but they become especially visible when the installed base is distributed across homes and businesses.

Relevance for the wider IoT ecosystem

Australia’s smart utility market has been more visibly associated with electricity metering than with gas. This project signals that gas networks are now being pulled further into the same digital operations model, where remote visibility and data availability become core service capabilities rather than optional enhancements.

For OEMs, the announcement reinforces demand for retrofit-ready devices that can be installed on legacy infrastructure. For connectivity providers, it points to utility use cases where coverage, power consumption and operational continuity are more important than high bandwidth. For enterprises and industrial players using gas services, more accurate and timely meter data can support better reconciliation of consumption and billing, although the announcement does not claim new energy management services beyond remote readings and data insights.

The conditional future opportunity is also notable. Landis+Gyr says a successful rollout could support broader deployment across approximately two million Landis+Gyr gas meters already installed in Australia. That is not a confirmed expansion, but it explains why the current programme will be watched beyond Origin’s network: it may provide a template for digitising existing gas metering assets without large-scale network replacement.

In practical terms, this is less a story about a single smart meter device than about a deployment model. If the programme performs as intended, it will show how utilities can modernise gas metering by layering IoT, communications and data management onto infrastructure already in the field.

The post Origin Energy to Digitise Australian Gas Metering with Landis+Gyr Retrofit IoT Modules appeared first on IoT Business News.

AI‑driven pricing has quietly become one of the most powerful—and least regulated—forces in modern commerce. What began as “dynamic pricing” has evolved into algorithmic market manipulation: systems that learn competitors’ behavior, anticipate consumer willingness to pay, and coordinate prices without a single human ever sending an email or making a phone call.

This is not science fiction. It’s already happening. And regulators are nowhere near ready.

1. The New Invisible Hand: Algorithms That Set Prices for Us

For decades, pricing was a strategic discipline: teams of analysts, economists, and product managers balancing supply, demand, and competitive pressure.

Today, AI systems do this work in milliseconds. They:

• Monitor competitor prices in real time
• Predict consumer behavior with uncanny accuracy
• Test thousands of micro‑price variations
• Automatically adjust prices to maximize profit

The result is a market where prices are no longer set by humans—they’re set by learning systems optimizing for revenue extraction.

This shift is not neutral. It fundamentally changes how markets behave.

2. When Algorithms Compete, Consumers Lose

In theory, algorithmic pricing should increase competition. In practice, the opposite is happening.

AI systems trained to maximize profit often converge on the same strategy: Raise prices until customers push back.

This creates a form of tacit collusion—not because companies coordinate intentionally, but because their algorithms learn that aggressive price-cutting triggers retaliation from competitors’ algorithms.

So they stop competing.

This is the digital equivalent of two gas stations silently agreeing not to undercut each other—except now it happens across entire industries, at machine speed, with no human fingerprints.

3. Personalized Pricing: The End of the “Fair Price”

AI doesn’t just set prices. It sets your price.

Companies now use:

• Purchase history
• Device type
• Location
• Income proxies
• Browsing patterns
• Loyalty data
• Time of day
• Even your typing speed

…to estimate your willingness to pay.

Two customers looking at the same product may see two entirely different prices—because one is judged “less price sensitive.”

This is not dynamic pricing. This is behavioral exploitation.

And it’s spreading fast across travel, retail, insurance, entertainment, and even healthcare.

4. The Dark Side: AI That Learns to Manipulate

The most concerning development isn’t price optimization—it’s behavioral shaping.

Advanced pricing models don’t just react to consumer behavior; they influence it. They learn:

• When you’re most impulsive
• When you’re most tired
• When you’re most stressed
• When you’re most likely to accept a higher price

This is where AI pricing crosses into psychological manipulation.

If a system knows you’re shopping late at night after a long day, it may raise prices because you’re less likely to comparison‑shop.

If it knows you’re anxious about a flight selling out, it may increase the fare.

This is not hypothetical. These behaviors have been documented in multiple industries.

5. Market Power Concentrates—Quietly

AI-driven pricing rewards scale. The more data a company has, the more accurate its predictions become.

This creates a feedback loop:

1. More customers → more data
2. More data → better pricing models
3. Better pricing → higher profits
4. Higher profits → more market dominance

The result is a small number of companies gaining disproportionate control over market prices—not through illegal collusion, but through algorithmic advantage.

This is market manipulation without the meeting room.

6. Regulators Are a Decade Behind

Traditional antitrust frameworks assume:

• Humans set prices
• Collusion requires communication
• Market manipulation leaves evidence

AI breaks all three assumptions.

How do you prosecute collusion when:

• No humans coordinated
• No messages were exchanged
• The algorithms simply learned the same profit-maximizing strategy

Regulators are stuck in a 20th‑century paradigm while 21st‑century markets are being shaped by systems they can’t audit, explain, or even detect.

7. The Coming Backlash

Public sentiment is shifting. Consumers are starting to notice:

• Uber rides that cost more when your phone battery is low
• Airline prices that spike after repeated searches
• Retailers that raise prices for iPhone users
• Hotels that adjust rates based on your browsing history

As AI-driven pricing becomes more aggressive, the backlash will grow. Expect:

• New regulatory frameworks
• Mandatory algorithmic audits
• Transparency requirements
• Restrictions on personalized pricing
• Litigation around algorithmic collusion

Companies that rely heavily on opaque pricing models will face increasing scrutiny.

8. What Companies Should Do Now

Forward-thinking organizations should prepare for a world where AI pricing is no longer a black box. That means:

• Building explainability into pricing models
• Documenting algorithmic decision pathways
• Establishing ethical pricing guidelines
• Creating internal audit mechanisms
• Preparing for regulatory disclosure requirements

The companies that thrive will be those that treat AI pricing not as a loophole to exploit but as a system to govern responsibly.

9. What Consumers Need to Understand

Consumers must recognize that:

• Prices are no longer objective
• You are being profiled constantly
• Your behavior influences the price you pay
• Loyalty can make you a target, not a beneficiary
• “Deals” are often engineered illusions

The more predictable you are, the more you pay.

Conclusion: AI Pricing Is Reshaping Markets—Quietly, Powerfully, and Without Oversight

AI-driven pricing is not just a business tactic. It is a structural shift in how markets function.

It redistributes power from consumers to corporations, from competition to coordination, and from transparent markets to opaque algorithmic ecosystems.

The question is no longer whether AI pricing manipulates markets. It’s whether society will recognize it—and regulate it—before the manipulation becomes irreversible.

Conceived, written and published by AI Quantum Intelligence with the help of AI models.

This mythical human is the benchmark. And AI, unsurprisingly, falls short.

But here’s the uncomfortable truth: humans fall short of that benchmark too — catastrophically so — and far more often.

Resistance to AI may not really be about AI’s limitations. It’s about the stories we tell ourselves about human capability, human objectivity, and human consistency. Stories that collapse the moment we examine them.

1. The “Human Gold Standard” Is a Fiction We Use to Avoid Hard Questions

When critics say AI is “biased,” they are correct — but incomplete. Bias is not an AI problem. Bias is a data problem, and humans are data‑processing systems too.

• Doctors misdiagnose patients at rates far higher than AI diagnostic models.
• Judges show measurable sentencing bias based on race, gender, and even time of day.
• Hiring managers consistently favour candidates who resemble themselves.
• Analysts anchor on their first assumption and defend it long after evidence contradicts it.

Yet when AI exhibits bias, the reaction is moral outrage. When humans exhibit bias, the reaction is: “Well, that’s just how people are.”

This asymmetry reveals the real issue: AI is held to a standard of perfection that humans do not meet, have never met, and cannot meet.

2. Human Expertise Is Not the Neutral, Objective Force We Pretend It Is

One of the most persistent myths in the anti‑AI narrative is that human experts operate from a place of pure logic and domain mastery. But research across behavioural economics, cognitive psychology, and organizational science paints a different picture:

• Experience increases confidence faster than accuracy.
• Experts become more rigid over time, not less.
• Domain knowledge narrows perspective, reducing the ability to see cross‑disciplinary patterns.
• Motivations distort judgment — career incentives, political pressures, personal identity, and organizational culture all shape decisions.

AI’s critics often say: “AI can’t understand context.”

But humans misunderstand context constantly — especially when the context contradicts their worldview.

AI’s flaw is visible. Human flaws are familiar, so we forgive them.

3. AI’s Real Advantage: It Doesn’t Share Our Blind Spots

This is where insight and thoughtful introspection becomes crucial: AI is not limited by the same cognitive architecture that limits humans.

Humans reason vertically — deep within their domain. AI can reason horizontally — across domains, across patterns, across analogies humans would never think to connect.

A human supply‑chain expert might never consider insights from epidemiology. A human financial analyst might never borrow frameworks from ecology. A human policy advisor might never apply lessons from robotics.

AI does this effortlessly.

Not because it is “creative” in the human sense, but because it is not constrained by the boundaries of human experience.

This is why AI often produces “out‑of‑the‑box” thinking: It has no box.

4. The Real Threat AI Poses: It Exposes How Much of Our Work Isn’t Actually Expert Work

AI doesn’t just automate tasks. It reveals how many tasks were never truly “expert” to begin with.

• Summaries
• Drafts
• First‑pass analysis
• Pattern recognition
• Scenario generation
• Risk flagging
• Cross‑domain analogy
• Data‑driven recommendations

These are not the pinnacle of human cognition. They are the scaffolding around it — the repetitive, error‑prone, bias‑laden work humans do because we lack the time, energy, or cognitive bandwidth to do better.

AI’s presence forces a reckoning: If a model can do 60% of your job better than you, was that 60% ever “expertise”?

This is not a threat. It is an opportunity — if we are willing to see it.

5. The Fear of AI Is Often a Fear of Losing the Illusion of Human Superiority

The most honest critics of AI are not the ones who say “AI is dangerous.” They are the ones who say, implicitly or explicitly:

“I don’t want to confront the possibility that my expertise is narrower, more biased, and more fragile than I believed.”

AI challenges the mythology of human exceptionalism. Not by replacing humans, but by revealing the limits we prefer not to acknowledge.

Humans are not objective. Humans are not consistent. Humans are not unbiased. Humans are not omniscient.

AI doesn’t need to be perfect to be useful. It only needs to be less flawed than the alternative.

And in many domains, it already is.

6. The Future Isn’t AI vs Humans — It’s Humans Who Embrace AI vs Humans Who Don’t

The real divide emerging is not between AI and humanity. It is between:

• Humans who use AI to expand their cognitive range, and
• Humans who cling to the illusion that their experience alone is enough.

The first group will see patterns others miss. They will generate ideas others never consider. They will operate with a breadth and depth that no unaided human can match.

The second group will insist that AI “can’t do what they do,” right up until the moment it does.

7. The Call to Action: Stop Comparing AI to the Best Humans — Compare It to the Real Ones

If we want a meaningful conversation about AI’s role in society, we must abandon the fantasy comparison.

The real question is not: “Is AI as good as the best human expert on their best day?”

The real question is: “Is AI better than the average human on an average day, working with average information, under average constraints?”

And increasingly, the answer is yes.

Not because AI is perfect. But because humans aren’t.

Conceived, edited and published by AI Quantum Intelligence with the help of AI models.

The Takeaway

The next competitive advantage in AI won’t come from who has the most data — it will come from who has the cleanest, most accurate, most context‑rich, and most provenance‑verified data. The era of “just scrape more” is over. The era of “curate better” has begun.

For the past decade, the AI industry has been obsessed with scale. More parameters. More compute. More data. The implicit belief was simple: if you feed a model enough information, intelligence will emerge. And for a while, that belief held up. Bigger models trained on bigger datasets produced undeniably impressive results.

But in 2025 and now into 2026, the cracks in that philosophy are no longer cracks — they’re fault lines. Companies are discovering that data quantity is no longer the moat it once was. The world has been scraped. The internet is saturated. Synthetic data is flooding the ecosystem. And the marginal returns of “more” are collapsing.

The new moat — the one that will define the next generation of AI winners — is data quality.

1. The Internet Is No Longer a Reliable Training Ground

For years, the open web was the AI industry’s free buffet. But that buffet is now spoiled.

• Content farms and SEO sludge dominate search results
• Synthetic content is multiplying faster than human content
• Misinformation and hallucinated facts are being re‑ingested into training pipelines
• Copyright restrictions and paywalls are shrinking the usable corpus

The result: Models trained on the open web are increasingly learning from a polluted, self‑referential, low‑signal environment.

Quantity is no longer abundance — it’s noise.

2. Synthetic Data Is a Double‑Edged Sword

Synthetic data was supposed to be the savior: infinite, cheap, customizable.

Instead, it introduced a new existential risk: model collapse — the phenomenon where models trained on synthetic outputs become less diverse, less accurate, and more brittle over time.

Synthetic data is powerful when used intentionally. It is catastrophic when used indiscriminately.

The companies that win will be the ones who treat synthetic data like a precision instrument, not a firehose.

3. High‑Quality Data Is Becoming the Most Valuable Asset in AI

The most advanced AI labs have quietly shifted strategy. They’re no longer bragging about dataset size. They’re bragging about dataset purity.

High‑quality data has specific characteristics:

• Verified provenance — you know where it came from
• Human‑generated — not synthetic sludge
• Domain‑specific — not generic internet text
• Context‑rich — includes metadata, structure, and meaning
• Legally clean — licensed, owned, or created in‑house
• Continuously refreshed — not static snapshots of the past

This is the data that produces models that are more accurate, more reliable, more controllable, and more aligned with real‑world use cases.

This is the data that becomes a moat.

4. The New Power Players Are the Ones Who Control Clean Data Pipelines

In Q1, the AI oligopoly formed around compute and capital. In Q2, the power shift is moving toward data governance and data curation.

The companies with the strongest moats will be those that:

• Own proprietary, high‑fidelity datasets
• Have exclusive access to industry‑specific data streams
• Maintain rigorous data cleaning and validation pipelines
• Build human‑in‑the‑loop systems for continuous refinement
• Invest in data provenance, watermarking, and authenticity verification

This is why industries like healthcare, finance, law, and scientific research are becoming battlegrounds. Not because they have more data — but because they have better data.

5. The Economic Shift: Quality Data Is Becoming a Premium Commodity

We are entering a world where:

• High‑quality datasets will be licensed like intellectual property
• Data provenance will be audited like financial statements
• Clean data will command premium pricing
• Enterprises will compete to secure exclusive data partnerships
• Governments will regulate data quality as a matter of national interest

Data is no longer the new oil. Clean data is the new lithium — scarce, valuable, and strategically essential.

6. Why This Matters for Businesses

Most organizations still believe they need “more data” to compete with frontier AI labs. They don’t.

They need:

• Better labeling
• Better metadata
• Better governance
• Better domain expertise
• Better curation

A small, high‑quality dataset can outperform a massive, messy one — especially in enterprise use cases where accuracy, reliability, and compliance matter more than raw generative power.

The organizations that understand this will build AI systems that are not only more effective, but also more defensible.

7. The Strategic Imperative for 2026–2030

The next era of AI will be defined by:

• Data authenticity
• Data scarcity
• Data rights
• Data curation
• Data governance

The winners will be those who treat data not as an exhaust, but as an asset. Not as a commodity, but as a craft. Not as a volume game, but as a quality discipline.

The moat is shifting. And the companies that recognize this shift early will own the next decade of AI.

Key References

• Stanford HAI — AI Index Report (2024–2025): Highlights the growing risks of synthetic data contamination and the plateauing benefits of scale.
• MIT Technology Review: Coverage on model collapse and the dangers of training on synthetic outputs.
• Nature & Science Journals: Peer‑reviewed studies on data quality, bias, and the impact of dataset curation on model performance.
• OECD & EU AI Act Documentation: Regulatory frameworks emphasizing data governance, provenance, and quality standards.
• OpenAI, Anthropic, Google DeepMind Technical Reports: Increasing focus on curated, proprietary, and human‑verified datasets.

Conceived, written and published by AI Quantum Intelligence with the help of AI models.

AI is dissolving the very function of middle management.

Not by replacing people one‑to‑one, but by collapsing the coordination layers that once made them necessary. The org chart isn’t being trimmed. It’s being re‑written.

And the companies that understand this shift are already reallocating power.

1. The Original Purpose of Middle Management Has Been Automated

Middle management emerged in the industrial era to solve a specific problem: information didn’t move on its own.

Managers existed to:

• relay instructions downward

• aggregate reports upward

• coordinate horizontally

• ensure compliance

• translate strategy into tasks

AI now performs all five functions—instantly, continuously, and without fatigue.

A modern AI system:

• synthesizes reports

• monitors performance

• coordinates workflows

• enforces policy

• translates goals into executable tasks

The “manager as messenger” is obsolete. The “manager as coordinator” is redundant. The “manager as information bottleneck” is actively harmful.

When information flows freely, layers built to move information become friction.

2. AI Flattens Hierarchies by Default

AI doesn’t respect hierarchy. It respects access.

When every employee—from intern to VP—can query the same intelligence layer, the traditional pyramid collapses into a network. Authority shifts from:

• tenure → capability

• title → output

• hierarchy → proximity to decision‑making

This is why AI‑native organizations feel different. They are flatter, faster, and more transparent.

The old structure was built around human limitations. The new structure is built around machine leverage.

3. The New Power Centers Are Not Managers—They Are Architects

As coordination becomes automated, a new class of influence emerges:

System architects. Workflow designers. Prompt engineers. AI orchestrators.

These individuals don’t manage people. They manage capability.

They design:

• how information flows

• how decisions are escalated

• how exceptions are handled

• how AI agents collaborate

• how human judgment is inserted

In the AI‑driven enterprise, the most powerful person is not the one with the largest team. It’s the one who designs the system that replaces the team.

This is the quiet revolution: Power is shifting from people managers to system designers.

4. The Middle Layer Shrinks, but the Edges Expand

AI doesn’t eliminate work. It redistributes it.

The bottom layer expands

Individual contributors gain:

• more autonomy

• more leverage

• more direct access to strategic context

AI copilots turn them into “micro‑executives”—capable of producing work that once required entire departments.

The top layer expands

Executives gain:

• real‑time visibility

• direct access to operational data

• the ability to steer without intermediaries

AI becomes the connective tissue between strategy and execution.

The middle layer contracts

Not because people are unskilled. But because the function they performed has been absorbed by automation.

The organization becomes a barbell: more power at the top, more capability at the bottom, less mass in the middle.

5. The Cultural Shift: From Supervision to Enablement

The biggest change isn’t structural—it’s cultural.

Middle managers were historically evaluated on:

• headcount

• budget

• compliance

• control

AI‑era leaders are evaluated on:

• enablement

• acceleration

• system design

• judgment

The question is no longer “How many people report to you?” It’s “How much capability flows through you?”

This is a fundamentally different definition of leadership.

6. The Organizations That Survive Will Redefine Management, Not Defend It

Organizations that cling to traditional hierarchies will experience:

• slower decision cycles

• duplicated work

• political bottlenecks

• rising coordination costs

• talent flight to flatter competitors

Organizations that embrace AI‑driven flattening will experience:

• faster execution

• clearer accountability

• higher individual leverage

• reduced bureaucracy

• more strategic leadership

The winners will not be the ones who “use AI.” They will be the ones who restructure around it.

7. The Future: Leadership Without Layers

The death of middle management is not the death of leadership. It is the death of administrative leadership.

What remains—and what becomes more valuable—is:

• judgment

• vision

• narrative

• ethics

• decision‑making under uncertainty

AI handles coordination. Humans handle meaning.

The organizations that thrive in the next decade will be those that understand this division of labour—and design their structures accordingly.

Middle management isn’t being replaced. It’s being redefined.

And the companies that recognize this early will own the next era of organizational power.

Conceived and developed by AI Quantum Intelligence with the help of AI models.

A new study of U.S. employment led by MIT labor economist David Autor sheds light on all these matters. In the postwar U.S., as Autor and his colleagues show in granular detail, new forms of work have tended to benefit college graduates under 30 more than anyone else. 

“We had never before seen exactly who is doing new work,” Autor says. “It’s done more by young and educated people, in urban settings.” 

The study also contains a powerful large-scale insight: A lot of innovation-based new work is driven by demand. Government-backed expansion of research and manufacturing in the 1940s, in response to World War II, accounted for a huge amount of new work, and new forms of expertise. 

“This says that wherever we make new investments, we end up getting new specializations,” Autor says. “If you create a large-scale activity, there’s always going to be an opportunity for new specialized knowledge that’s relevant for it. We thought that was exciting to see.” 

The paper, “What Makes New Work Different from More Work?” is forthcoming in the Annual Review of Economics. The authors are Autor; Caroline Chin, a doctoral student in MIT’s Department of Economics; Anna M. Salomons, a professor at Tilburg University’s Department of Economics and Utrecht University’s School of Economics; and Bryan Seegmiller PhD ’22, an assistant professor at Northwestern University’s Kellogg School of Management.

And yes, learning about new work, and the kinds of workers who obtain it, might be relevant to the spread of artificial intelligence — although, in Autor’s estimation, it is too soon to tell just how AI will affect the workplace.

“People are really worried that AI-based automation is going to erode specific tasks more rapidly,” Autor observes. “Eroding tasks is not the same thing as eroding jobs, since many jobs involve a lot of tasks. But we’re all saying: Where is the new work going to come from? It’s so important, and we know little about it. We don’t know what it will be, what it will look like, and who will be able to do it.”

“If everyone is an expert, then no one is an expert”

The four co-authors also collaborated on a previous major study of new work, published in 2024, which found that about six out of 10 jobs in the U.S. from 1940 to 2018 were in new specialties that had only developed broadly since 1940. The new study extends that line of research by looking more precisely at who fills the new lines of work. 

To do that, the researchers used U.S. Census Bureau data from 1940 through 1950, as well as the Census Bureau’s American Community Survey (ACS) data from 2011 to 2023. In the first case, because Census Bureau records become wholly public after about 70 years, the scholars could examine individual-level data about occupations, salaries, and more, and could track the same workers as they changed jobs between the 1940 and 1950 Census enumerations. 

Through a collaborative research arrangement with the U.S. Census Bureau, the authors also gained secure access to person-level ACS records. These data allowed them to analyze the earnings, education, and other demographic characteristics of workers in new occupational specialties — and to compare them with workers in longstanding ones.

New work, Autor observes, is always tied to new forms of expertise. At first, this expertise is scarce; over time, it may become more common. In any case, expertise is often linked to new forms of technology.

“It requires mastering some capability,” Autor says. “What makes labor valuable is not simply the ability to do stuff, but specialized knowledge. And that often differentiates high-paid work from low-paid work.” Moreover, he adds, “It has to be scarce. If everyone is an expert, then no one is an expert.”

By examining the census data, the scholars found that back in 1950, about 7 percent of employees had jobs in types of work that had emerged since 1930. More recently, about 18 percent of workers in the 2011-2023 period were in lines of work introduced since 1970. (That happens to be roughly the same portion of new jobs per decade, although Autor does not think this is a hard-and-fast trend.) 

In these time periods, new work has emerged more often in urban areas, with people under 30 benefitting more than any other age category. Getting a job in a line of new work seems to have a lasting effect: People employed in new work in 1940 were 2.5 times as likely to be in new work in 1950, compared to the general population. College graduates were 2.9 percentage points more likely than high school graduates to be engaged in new work. 

New work also has a wage premium, that is, better salaries on aggregate than in already-existing forms of work. Yet as the study shows, that wage premium also fades over time, as the particular expertise in many forms of new work becomes much more widely grasped. 

“The scarcity value erodes,” Autor says. “It becomes common knowledge. It itself gets automated. New work gets old.”

After all, Autor points out, driving a car was once a scarce form of expertise. For that matter, so was being able to use word-processing programs such as WordPerfect or Microsoft Word, well into the 1990s. After a while, though, being able to handle word-processing tools became the most elementary part of using a computer.

Back to AI for a minute

Studying who gets new jobs led the scholars to striking conclusions about how new work is created. Examining county-level data from the World War II era, when the federal government was backing new manufacturing in public-private partnerships throughout the U.S., the study shows that counties with new factories had more new work, and that 85 to 90 percent of new work from 1940 to 1950 was technology-driven. 

In this sense there was a great deal of demand-driven innovation at the time. Today, public discourse about innovation often focuses on the supply side, namely, the innovators and entrepreneurs trying to create new products. But the study shows that the demand side can significantly influence innovative activity. 

“Technology is not like, ‘Eureka!’ where it just happens,” Autor says. “Innovation is a purposive activity. And innovation is cumulative. If you get far enough, it will have its own momentum. But if you don’t, it’ll never get there.”

Which brings us back to AI, the topic so many people are focused on in 2026. Will AI create good new jobs, or will it take work away? Well, it likely depends how we implement it, Autor thinks. Consider the massive health care sector, where there could be a lot of types of tech-driven new work, if people are interested in creating jobs.

“There are different ways we could use AI in health care,” Autor says. “One is just to automate people’s jobs away. The other is to allow people with different levels of expertise to do different tasks. I would say the latter is more socially beneficial. But it’s not clear that is where the market will go.” 

On the other hand, maybe with government-driven demand in various forms, AI could get applied in ways that end up boosting health care-sector productivity, creating new jobs as a result. 

“More than half the dollars in health care in the U.S. are public dollars,” Autor observes. “We have a lot of leverage there, we can push things in that direction. There are different ways to use this.” 

This research was supported, in part, by the Hewlett Foundation, the Google Technology and Society Visiting Fellows Program, the NOMIS Foundation, the Schmidt Sciences AI2050 Fellowship, the Smith Richardson Foundation, the James M. and Cathleen D. Stone Foundation, and Instituut Gak.

Evaluating each of those compounds experimentally would be far too time-consuming for chemists. So, in recent years, researchers have begun using artificial intelligence to help identify compounds that could make good drug candidates. 

One of those researchers is MIT Associate Professor Connor Coley PhD ’19, the Class of 1957 Career Development Associate Professor with shared appointments in the departments of Chemical Engineering and Electrical Engineering and Computer Science and the MIT Schwarzman College of Computing. His research straddles the line between chemical engineering and computer science, as he develops and deploys computational models to analyze vast numbers of possible chemical compounds, design new compounds, and predict reaction pathways that could generate those compounds. 

“It’s a very general approach that could be applied to any application of organic molecules, but the primary application that we think about is small-molecule drug discovery,” he says.

The intersection of AI and science

Coley’s interest in science runs in the family. In fact, he says, his family includes more scientists than non-scientists, including his father, a radiologist; his mother, who earned a degree in molecular biophysics and biochemistry before going to the MIT Sloan School of Management; and his grandmother, a math professor.

As a high school student in Dublin, Ohio, Coley participated in Science Olympiad competitions and graduated from high school at the age of 16. He then headed to Caltech, where he chose chemical engineering as a major because it offered a way to combine his interests in science and math.

During his undergraduate years, he also pursued an interest in computer science, working in a structural biology lab using the Fortran programming language to help solve the crystal structure of proteins. After graduating from Caltech, he decided to keep going in chemical engineering and came to MIT in 2014 to start a PhD.

Advised by professors Klavs Jensen and William Green, Coley worked on ways to optimize automated chemical reactions. His work focused on combining machine learning and cheminformatics — the application of computation methods to analyze chemical data — to plan reaction pathways that could make new drug molecules. He also worked on designing hardware that could be used to perform those reactions automatically. 

Part of that work was done through a DARPA-funded program called Make-It, which was focused on using machine learning and data science to improve the synthesis of medicines and other useful compounds from simple building blocks.

“That was my real entry point into thinking about cheminformatics, thinking about machine learning, and thinking about how we can use models to understand how different chemicals can be made and what reactions are possible,” Coley says.

Coley began applying for faculty jobs while still a graduate student, and accepted an offer from MIT at age 25. He received a mix of advice for and against taking a job at the same school where he went to graduate school, and eventually decided that a position at MIT was too enticing to turn down.

“MIT is a very special place in terms of the resources and the fluidity across departments. MIT seemed to be doing a really good job supporting the intersection of AI and science, and it was a vibrant ecosystem to stay in,” he says. “The caliber of students, the enthusiasm of the students, and just the incredible strength of collaborations definitely outweighed any potential concerns of staying in the same place.”

Chemistry intuition

Coley deferred the faculty position for one year to do a postdoc at the Broad Institute, where he sought more experience in chemical biology and drug discovery. There, he worked on ways to identify small molecules, from billions of candidates in DNA-encoded libraries, that might have binding interactions with mutated proteins associated with diseases.

After returning to MIT in 2020, he built his lab group with the mission of deploying AI not only to synthesize existing compounds with therapeutic potential, but also to design new molecules with desirable properties and new ways to make them. Over the past few years, his lab has developed a variety of computational approaches to tackle those goals. 

“We try to think about how to best pair a challenge in chemistry with a potential computational solution. And often that pairing motivates the development of new methods,” Coley says. One model his lab has developed, known as ShEPhERD, was trained to evaluate potential new drug molecules based on how they will interact with target proteins, based on the drug molecules’ three-dimensional shapes. This model is now being used by pharmaceutical companies to help them discover new drugs.

“We’re trying to give more of a medicinal chemistry intuition to the generative model, so the model is aware of the right criteria and considerations,” Coley says.

In another project, Coley’s lab developed a generative AI model called FlowER, which can be used to predict the reaction products that will result from combining different chemical inputs. 

In designing that model, the researchers built in an understanding of fundamental physical principles, such as the law of conservation of mass. They also compelled the model to consider the feasibility of the intermediate steps that need to take place on the pathway from reactants to products. These constraints, the researchers found, improved the accuracy of the model’s predictions.

“Thinking about those intermediate steps, the mechanisms involved, and how the reaction evolves is something that chemists do very naturally. It’s how chemistry is taught, but it’s not something that models inherently think about,” Coley says. “We’ve spent a lot of time thinking about how to make sure that our machine-learning models are grounded in an understanding of reaction mechanisms, in the same way an expert chemist would be.”

Students in his lab also work on many different areas related to the optimization of chemical reactions, including computer-aided structure elucidation, laboratory automation, and optimal experimental design.

“Through these many different research threads, we hope to advance the frontier of AI in chemistry,” Coley says.

Sunshine Jiang  ’25

Sunshine Jiang, from Hangzhou, China, graduated from MIT in 2025 with a bachelor’s degree as a double major in physics and electrical engineering and computer science, along with minors in mathematics and economics. She will receive her master of engineering degree this month and will start her PhD in computer science at Stanford School of Engineering this fall. 

Jiang researches embodied artificial intelligence and robotics, developing data-efficient, adaptive systems for general-purpose robots that broaden accessibility. She has presented her research at major conferences, including the Conference on Robot Learning, the International Conference on Robotics and Automation, and the International Conference on Learning Representations. 

Jiang led the development of AI-powered systems that provide access to traditional Chinese art in rural classrooms, founded cross-country programs that expand girls’ access to STEM education, and created a Covid-19 documentary amplifying community voices, which was featured on China Daily.

Rupert Li ’24

Rupert Li, from Portland, Oregon, is currently pursuing a PhD in mathematics at Stanford School of Humanities and Sciences. He graduated from MIT in 2024 with a bachelor’s degree, double majoring in mathematics and computer science, economics, and data science. Along with his bachelor’s degree, he also received a master’s degree in data science. Li then traveled to the United Kingdom as a Marshall Scholar, where he earned a master’s degree in mathematics from the University of Cambridge.

Li’s research interests lie in probability, discrete geometry, and combinatorics. He enjoys serving as a mentor for MIT PRIMES-USA, a high school math research program, and previously served as an advisor for the Duluth REU, an undergraduate math research program. In addition to the Knight-Hennessy Scholarship and the Marshall Scholarship, he has been awarded the Hertz Fellowship, P.D. Soros Fellowship, and the Goldwater Scholarship, and he received honorable mention for the Frank and Brennie Morgan Prize.

The world is reaching an inflection point with artificial intelligence: over half of U.S. adults use generative AI — with 12 percent using it daily at work — and 88 percent of global organizations have integrated AI into at least one core function, up from 78 percent in 2024. AI knowledge is no longer optional for career growth, organizational leadership, and life. Yet, a growing information gap exists between those with the capabilities to leverage AI’s potential and those trying to keep pace. 

The need for accessible, practical AI education has never been greater. To meet this moment, MIT Open Learning is launching Universal AI, an online, self-paced, modular program that takes a learner from AI novice to authority, starting with core fundamentals and building to real-world, industry-specific applications.

“We identified a need for an AI learning experience that is universal in breadth and accessibility — one that bridges the gap between deeply technical and surface level introductions to the latest AI tools, and that is designed for a non-technical, global audience,” says Dimitris Bertsimas, vice provost for open learning. “Universal AI was built to thread that needle. We took MIT’s long-standing expertise in the field and completely reimagined how it’s taught, grounding it in real-world cases and supporting every learner with AI tools that adapt to them. The result is a pathway to AI fluency that’s approachable to anyone, anywhere.”

The core curriculum spans five courses that cover the underlying theories, concepts, and technologies behind AI including programming, machine and deep learning, large language models, decision-making, explainability, and ethics. The first course in the program, Fundamentals of Programming and Machine Learning, is available for free to learners everywhere.

Universal AI also includes industry-specific courses that dive into the intersection of AI and health care, sustainability, entrepreneurship, transportation, and more. Six industry-specific courses are available today, including Holistic AI in Medicine, AI and Entrepreneurship, and AI and Sustainability: Energy.

“Our goal is that the learners who take Universal AI gain the foundational knowledge and understanding so that they realize the potential of AI for their careers, lives, and communities," says Megan Mitchell, senior director of Universal Learning at Open Learning. “We also hope that the program dispels the fear and unknown about AI, and empowers learners to embrace the true potential of this transformative technology.”

Universal AI is available on MIT Learn, the Institute’s online learning platform with programs, courses, and resources that are designed to help learners build new skills, explore emerging technologies, and advance their careers. The platform is enabled with an AI assistant, AskTIM, that helps learners discover and chart their learning journey, answers questions about key lecture concepts, and tutors learners through assignments.

Universal AI was piloted by a wide-ranging group of organizations starting in summer 2025, which included universities, hospitals, companies, the MIT community, and refugee and displaced learners in the MIT Emerging Talent program.

Madiha Malikzada, a learner who participated in the pilot program, appreciated having AskTIM as a “study buddy.”

“[AskTIM] challenged me to think more deeply and engage with the material in a meaningful way,” says Malikzada. “It made me think that sometimes we forget to mention how helpful AI can be in the learning process, not just for answering questions, but for having a back-and-forth exchange that can give us new ideas and deepen our understanding.”

Universal AI includes contributions from over 30 faculty, teaching assistants, and experts from across MIT. This number will grow as additional industry-specific courses become available.

“It’s remarkable to see so many members of the MIT community come together to create high-quality resources and tools for people around the world who want to learn about AI,” says MIT provost Anantha Chandrakasan. “It really showcases the diversity of perspectives and expertise on AI across the Institute, as well as the commitment to harnessing that expertise to benefit online learners.”

Universal AI is the first offering from Universal Learning, a new initiative at Open Learning focused on developing curricula across the most critical areas shaping our world. Read more from Bertsimas and Mitchell about Universal Learning.

“MIT’s long history of making knowledge available through MIT Open Learning means it’s only natural we’d feel compelled to bring Universal AI to the world,” adds Kornbluth.

Universal AI is now available on MIT Learn. 

Universal Learning offerings combine subject matter expertise from MIT faculty and experts and Open Learning’s more than 25 years of innovation in online education to deliver a learning experience centered on real-world stories, practical exercises, and the needs of global learners. It is delivered on the MIT Learn platform, leveraging the capabilities of the AskTIM AI assistant to support learners throughout their educational journey. 

Universal AI, the first offering from Universal Learning, launched to the public today. Future offerings will include climate and energy, biology, health care, and manufacturing. Dimitris Bertsimas, vice provost for open learning, and Megan Mitchell, senior director of Universal Learning, share how Universal Learning supports MIT’s educational mission, and what makes it distinctive.

Q: How does Universal Learning reflect MIT’s commitment to educating the world?

Bertsimas: MIT’s primary residential mission is to educate its 11,000 students. But online education, taught at the appropriate level and enhanced with the latest AI teaching technology, can expand that mission exponentially. As one of the world’s premier research universities, MIT produces groundbreaking research that informs innovation and future educational materials. After 40 years focused on research, I’m excited to bring the knowledge we have accumulated to a much broader audience. My colleagues contributing to current and forthcoming Universal Learning offerings share this same passion.

Mitchell: Talent and capability is ubiquitous. Access and time is not. At MIT, we are pushing the boundaries to think about how we can reach more learners and meet them where they are, whether that’s through traditional institutions and universities, a corporate environment for upskilling and workforce learning, or those outside of traditional institutions. These learners in particular face the barriers of access and time, which we are aiming to address with Universal Learning’s modular, stackable offerings. In the end, we want to ensure we are developing offerings that are broadly accessible. 

Q: What unique aspects of an MIT education are infused in Universal Learning offerings?

Bertsimas: MIT students are trained not just to absorb knowledge, but to cross disciplinary boundaries, synthesize ideas from multiple domains, and translate that thinking into concrete action. This analytical yet pragmatic mindset — equal parts rigorous and creative — is the hallmark of how MIT approaches complex problem-solving. Universal Learning offerings are built to infuse these qualities, combining the knowledge of MIT faculty and Open Learning’s deep expertise in online education, but designed to cultivate deep topic fluency in a way that is approachable to a broad audience. The goal is to cultivate interdisciplinary thinking in learners everywhere, equipping them with the same intellectual toolkit that has long distinguished an MIT education.

Mitchell: Graduates of MIT are known for bringing their cumulative experiences and knowledge to bear on large, audacious problems that resist simple or narrow solutions. Universal Learning programs are built on that same teaching philosophy, intentionally designed for learners who may not have the opportunity to study at MIT, but deserve access to an equally expansive and ambitious approach to learning.

Q: What makes the experience of Universal Learning unique?

Mitchell: Universal Learning programs are modular in nature, and the pedagogy leverages real-world examples and hands-on exercises that occasionally include codes — not for learners to learn to code, but to know how to leverage data and interpret outputs. In particular, the modular structure is very compelling to the universities and companies we’ve been working with. Instead of creating one course that learners and educators have to take in a specific way or sequence, they can think about their needs and stack and leverage the Universal Learning offerings accordingly. Its very dynamic and flexible, designed to meet the needs of today’s online learning and workforce transformation landscape. 

Q: What has changed about the online learning landscape over the past 10-15 years?

Bertsimas: For the majority of the massive open online course (MOOC) movement, we replicated a residential class that was developed for MIT students and put it out into the world. But not all material is suited for a broad global audience, despite our best intentions. That’s why with Universal Learning we are prioritizing asynchronous delivery, mobile delivery, translations, and are developing ways to personalize content. AI is opening new ways to reach learners worldwide, and we are harnessing that potential. For example, the AskTIM AI assistant is capable of helping students solve exercises and answer conceptual questions — very much like a human teaching assistant would do, but at scale. 

Rather than implement automation in pursuit of maximal productivity, firms have often used automation to replace employees who specifically receive a “wage premium,” earning higher salaries than other comparable workers. In practice, that means automation has frequently reduced the earnings of non-college-educated workers who had obtained better salaries than most employees with similar qualifications. 

This finding has at least two big implications. For one thing, automation has affected the growth in U.S. income inequality even more than many observers realize. At the same time, automation has yielded a mediocre productivity boost, plausibly due to the focus of firms on controlling wages rather than finding more tech-driven ways to enhance efficiency and long-term growth.

“There has been an inefficient targeting of automation,” says MIT’s Daron Acemoglu, co-author of a published paper detailing the study’s results. “The higher the wage of the worker in a particular industry or occupation or task, the more attractive automation becomes to firms.” In theory, he notes, firms could automate efficiently. But they have not, by emphasizing it as a tool for shedding salaries, which helps their own internal short-term numbers without building an optimal path for growth.

The study estimates that automation is responsible for 52 percent of the growth in income inequality from 1980 to 2016, and that about 10 percentage points derive specifically from firms replacing workers who had been earning a wage premium. This inefficient targeting of certain employees has offset 60-90 percent of the productivity gains from automation during the time period.

“It’s one of the possible reasons productivity improvements have been relatively muted in the U.S., despite the fact that we’ve had an amazing number of new patents, and an amazing number of new technologies,” Acemoglu says. “Then you look at the productivity statistics, and they are fairly pitiful.”

The paper, “Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity,” appears in the May print issue of the Quarterly Journal of Economics. The authors are Acemoglu, who is an Institute Professor at MIT; and Pascual Restrepo, an associate professor of economics at Yale University.

Inequality implications

Dating back to the 2010s, Acemoglu and Restrepo have combined to conduct many studies about automation and its effects on employment, wages, productivity, and firm growth. In general, their findings have suggested that the effects of automation on the workforce after 1980 are more significant than many other scholars have believed. 

To conduct the current study, the researchers used data from many sources, including U.S. Census Bureau statistics, data from the bureau’s American Community Survey, industry numbers, and more. Acemoglu and Restrepo analyzed 500 detailed demographic groups, sorted by five levels of education, as well as gender, age, and ethnic background. The study links this information to an analysis of changes in 49 U.S. industries, for a granular look at the way automation affected the workforce. 

Ultimately, the analysis allowed the scholars to estimate not just the overall amount of jobs erased due to automation, but how much of that consisted of firms very specifically trying to remove the wage premium accruing to some of their workers. 

Among other findings, the study shows that within groups of workers affected by automation, the biggest effects occur for workers in the 70th-95th percentile of the salary range, indicating that higher-earning employees bear much of the brunt of this process. 

And as the analysis indicates, about one-fifth of the overall growth in income inequality is attributable to this sole factor.

“I think that is a big number,” says Acemoglu, who shared the 2024 Nobel Prize in economic sciences with his longtime collaborators Simon Johnson of MIT and James Robinson of the University of Chicago.

He adds: “Automation, of course, is an engine of economic growth and we’re going to use it, but it does create very large inequalities between capital and labor, and between different labor groups, and hence it may have been a much bigger contributor to the increase in inequality in the United States over the last several decades.” 

The productivity puzzle

The study also illuminates a basic choice for firm managers, but one that gets overlooked. Imagine a type of automation — call-center technology, for instance — that might actually be inefficient for a business. Even so, firm managers have incentive to adopt it, reduce wages, and oversee a less productive business with increased net profits.

Writ large, some version of this seems to have been happening to the U.S. economy since 1980: Greater profitability is not the same as increased productivity.

“Those two things are different,” says Acemoglu. “You can reduce costs while reducing productivity.” 

Indeed, the current study by Acemoglu and Restrepo calls to mind an observation by the late MIT economist Robert M. Solow, who in 1987 wrote, “You can see the computer age everywhere but in the productivity statistics.” 

In that vein, Acemoglu observes, “If managers can reduce productivity by 1 percent but increase profits, many of them might be happy with that. It depends on their priorities and values. So the other important implication of our paper is that good automation at the margins is being bundled with not-so-good automation.” 

To be clear, the study does not necessarily imply that less automation is always better. Certain types of automation can boost productivity and feed a virtuous cycle in which a firm makes more money and hires more workers. 

But currently, Acemoglu believes, the complexities of automation are not yet recognized clearly enough. Perhaps seeing the broad historical pattern of U.S. automation, since 1980, will help people better grasp the tradeoffs involved — and not just economists, but firm managers, workers, and technologists. 

“The important thing is whether it becomes incorporated into people’s thinking and where we land in terms of the overall holistic assessment of automation, in terms of inequality, productivity and labor market effects,” Acemoglu says. “So we hope this study moves the dial there.”

Or, as he concludes, “We could be missing out on potentially even better productivity gains by calibrating the type and extent of automation more carefully, and in a more productivity-enhancing way. It’s all a choice, 100 percent.”

By around age 14, Farina had focused on an idea that would prove foundational to his career.

“I was fascinated very early by the idea that a machine could make predictions or decisions so much better than humans,” he says. “The fact that human-made mathematics and algorithms could create systems that, in some sense, outperform their creators, all while building on simple building blocks, has always been a major source of awe for me.”

At age 16, Farina wrote code to solve a board game he played with his 13-year-old sister.

“I used game after game to compute the optimal move and prove to my sister that she had already lost long before either of us could see it ourselves,” Farina says, adding that his sister was less enthralled with his new system.

Now an assistant professor in MIT’s Department of Electrical Engineering and Computer Science (EECS) and a principal investigator at the Laboratory for Information and Decision Systems (LIDS), Farina combines concepts from game theory with such tools as machine learning, optimization, and statistics to advance theoretical and algorithmic foundations for decision-making.

Enrolling at Politecnico di Milano for college, Farina studied automation and control engineering. Over time, however, he realized that what activated his interest was not “just applying known techniques, but understanding and extending their foundations,” he says. “I gradually shifted more and more toward theory, while still caring deeply about demonstrating concrete applications of that theory.”

Farina’s advisor at Politecnico di Milano, Nicola Gatti, professor and researcher in computer science and engineering, introduced Farina to research questions in computational game theory and encouraged him to apply for a PhD. At the time, being the first in his immediate family to earn a college degree and living in Italy, where doctoral degrees are handled differently, Farina says he didn’t even know what a PhD was.

Nevertheless, one month after graduating with his undergraduate degree, Farina began a doctoral degree in computer science at Carnegie Mellon University. There, he won distinctions for his research and dissertation, as well as a Facebook Fellowship in Economics and Computation.

As he was finishing his doctorate, Farina worked for a year as a research scientist in Meta’s Fundamental AI Research Labs. One of his major projects was helping to develop Cicero, an AI that was able to beat human players in a game that involves forming alliances, negotiating, and detecting when other players are bluffing.

Farina says, “when we built Cicero, we designed it so that it would not agree to form an alliance if it was not in its interest, and it likewise understood whether a player was likely lying, because for them to do as they proposed would be against their own incentives.”

A 2022 article in the MIT Technology Review said Cicero could represent advancement toward AIs that can solve complex problems requiring compromise.

After his year at Meta, Farina joined the MIT faculty. In 2025, he was distinguished with the National Science Foundation CAREER Award. His work — based on game theory and its mathematical language describing what happens when different parties have different objectives, and then quantifying the “equilibrium” where no one has a reason to change their strategy — aims to simplify massive, complex real-world scenarios where calculating such an equilibrium could take a billion years.

“I research how we can use optimization and algorithms to actually find these stable points efficiently,” says Farina, who is also a core faculty member of the Operations Research Center. “Our work tries to shed new light on the mathematical underpinnings of the theory, better control and predict these complex dynamical systems, and uses these ideas to compute good solutions to large multi-agent interactions.”

Farina is especially interested in settings with “imperfect information,” which means that some agents have information that is unknown to other participants. In such scenarios, information has value, and participants must be strategic about acting on the information they possess so as not to reveal it and reduce its value. An everyday example occurs in the game of poker, where players bluff in order to conceal information about their cards.

According to Farina, “we now live in a world in which machines are far better at bluffing than humans.”

A situation with “massive amounts of imperfect information,” has brought Farina back to his board-game beginnings. Stratego is a military strategy game that has inspired research efforts costing millions of dollars to produce systems capable of beating human players. Requiring complex risk calculation and misdirection, or bluffing, it was possibly the only classical game for which major efforts had failed to produce superhuman performance, Farina says.

With new algorithms and training costing less than $10,000, rather than millions, Farina and his research team were able to beat the best player of all time — with 15 wins, four draws, and one loss. Farina says he is thrilled to have produced such results so economically, and he hopes “these new techniques will be incorporated into future pipelines,” he says.

“We have seen constant progress towards constructing algorithms that can reason strategically and make sound decisions despite large action spaces or imperfect information. I am excited about seeing these algorithms incorporated into the broader AI revolution that’s happening around us.”

These days, Honeycutt, a double major in computation and cognition and linguistics, still finds herself moving between several worlds that are simultaneously connected and distinctly different. Her research interests lie at the intersection of human thinking and awareness, language learning and acquisition, technology, and social group interaction and impact. 

Honeycutt’s interest in language and the ways it can shape how we think and live grew alongside lifelong investments in math and science. She learned French from her relationships with Haitian family friends, and American Sign Language because of another friend’s deaf sibling. She was fascinated with how speakers from those groups communicated and how the brain can reorganize itself when confronted with a lack of auditory input.

“There are so many things that are different about sign language and spoken language,” she says. “Speaking in multiple languages and dialects while managing the emotional and cultural nuances multilingualism presents can shift your experience of the world and of yourself.” Operating in these areas creates research opportunities in disciplines as diverse as neurology, large language models (LLMs), psychology, and public policy.  

“There’s fascinating work underway in neurolinguistics,” Honeycutt notes, “along with trying to better understand the differences between neural networks, AI, and how each processes information.” She’s wanted to study these for a long time, she says. “When people have to manage language deficits like aphasia, for example, and you’re immersed in several areas of investigation to find answers, you get to learn cool things like how the brain ‘does’ language.”  

An MIT approach to study 

Honeycutt chose MIT, in part, because the computation and cognition major was “not something I could find elsewhere.” Her affinity for math and English, alongside a desire to pursue the kind of computer science work that “centered people,” increased the likelihood that she could continue in her preferred areas of investigation with the support of the Institute’s faculty and other students.

She found class 9.59J (Laboratory in Psycholinguistics), taught by professor of brain and cognitive sciences Ted Gibson, to be especially enlightening. “It laid the foundation for my work,” she says.

Her decision to major in linguistics along with computation and cognition meant she could connect her interests in brain function and technology with a data-driven approach to language study and processing. “Majoring in linguistics highlighted the power of scientific rigor to organize and analyze a vast amount of chaotic, human-centric data,” she says. Her coursework reinforced the value of her decision. 

Honeycutt lauds the freedom MIT’s focus on interdisciplinary study provides. “Researchers are exploring differences between human and LLM language models and processing, and a lot of that work is happening at MIT,” she says. “MIT provides a rigorous flexibility that allows me to indulge multiple academic interests.”

It’s this flexibility that Honeycutt values most. “It’s the only reason I’m on the path I’ve chosen,” she continues, one that features a focus on language acquisition, education policy, LLMs’ computational possibilities and limitations, and education reform.

Honeycutt’s research continued on a series of MISTI trips in 2025. She traveled to South Africa in the summer, where she worked on the South African Human Rights Commission’s “Right to Read” campaign. She explored connections between language processing and brain function and supported research to aid in developing legislation to help increase literacy among South Africans. 

“Linguistic diversity presents significant challenges in South Africa,” she asserts. “One of the impacts of colonization on indigenous Africans, for example, is that children are often pushed out of schools because they can’t use the languages they’re learning — like Afrikaans — with their families at home.”

In fall 2025, she took a MISTI trip to Edinburgh, Scotland, where she studied sociolinguistics. She learned the value of considering alternative approaches to the brand of linguistics offered at MIT. “MIT’s approach to linguistics centers words and approaches its study like a math problem, while sociolinguistics includes important cultural context,” she says. Connecting the two made for a more complete, holistic approach to the work. 

Honeycutt values a balanced approach to her studies, creating time for extracurricular activities that allow her to both investigate her research goals and create community. “I completed a policy internship in Washington, D.C. in 2024,” she recalls.

She’s a member of Theta Delta Chi, a fraternity comprising a diverse group of undergraduates from a variety of academic backgrounds. She plays women’s club soccer and is an officer with the MIT Undergraduate Association. As a co-chair of the Community Service committee, she’s leading efforts to create connections with students living off campus. 

Honeycutt also volunteers with the Community Charter School of Cambridge, working to improve outcomes for underachieving students. As a volunteer, she’s able to pilot some of the education ideas being developed in her coursework. “I want to help underperforming students in the same way some institutions aid high-performing students,” she says.

The human element

Language shapes the ways its users view the world, according to Honeycutt. “I’m interested in how language can constrain thought,” she says. Language mastery is also a valuable tool in gauging emotional intelligence. “It’s important that people acquire and understand language in school,” she argues. “People should have access to a language that allows them to effectively communicate what they’re thinking.”

Having words for emotions can help people process them, Honeycutt believes. This is important in areas like translation and psychology, where nuance can be important. She also believes that reading and language acquisition are essential tools in developing effective self-awareness. Language is a medium for thought and provides guardrails to improve understanding. 

“Access to a large vocabulary, including words for emotions, can increase your emotional intelligence,” she says.  

With a solid academic foundation focused on cognition, language, and AI in place, Honeycutt plans to pursue studies in law and policy after graduation. That means law school and public policy programs, perhaps at an institution that offers a dual degree track.

“I want to extend opportunities to underserved students,” she says. “Problems in policy spaces are difficult, in part, because they defy easy categorization and involve multiple stakeholders.” Education, Honeycutt says, “is a fun problem to try to solve.” She wants to support efforts to enact lasting change by improving literacy, ensuring linguistic diversity, and centering science and research when crafting and implementing effective legislation that benefits learners, institutions, families, and communities. 

There’s no single study in a field that will answer all the questions, Honeycutt argues. By combining the science of brain function with the social and mathematical aspects of linguistics, she can continue investigating language, its usage, and impacts on people and their lives. We can’t solve education challenges, improve AI and access to AI-enabled tools, and further the study of linguistics without institutional and community support. 

“Support research,” Honeycutt says. “Don’t give up on trying to solve these problems.” 

Beacon Biosignals is working to make sense of the brain by monitoring its activity while people sleep. The company, which was founded by Jake Donoghue PhD ’19 and former MIT researcher Jarrett Revels, developed a lightweight headband that uses electroencephalogram (EEG) technology to measure brain activity while people enjoy their normal sleep routines at home. Those data are processed by machine-learning algorithms to monitor the effects of novel treatments, find new signs of disease progression, and create patient cohorts for clinical trials.

“There’s a step-change in what becomes possible when you remove the sleep lab and bring clinical-grade EEG into the home,” says Donoghue, who serves as Beacon’s CEO. “It turns sleep from a constrained, facility-based test into a scalable source of high-quality data for diagnostics, drug development, and longitudinal brain health.”

Beacon partners with pharmaceutical companies to accelerate its path to patients. The company’s FDA 510(k)-cleared medical device has already been used in over 40 clinical trials across the globe as part of studies aimed at treating conditions including major depressive disorder, schizophrenia, narcolepsy, idiopathic hypersomnia, Alzheimer’s disease, and Parkinson’s disease.

With each deployment, Beacon learns more about how the brain works — insights it is using to create a “foundation model” of the brain.

“It’s our belief that the dataset that’s going to transform brain health doesn’t exist yet — but we are rapidly creating it,” Donoghue says. “Our platform can characterize the heterogeneity of disease progression, generating dynamic insights that are impossible to fully capture through static modalities like sequencing or imaging. The brain is an electric organ and changes through synaptic plasticity, so tracking brain function across many diseases at scale will allow us to discover novel subgroups of diseases and map them over time.”

Illuminating the brain

Donoghue trained in the Harvard-MIT Program in Health Sciences and Technology, conducting clinical training for an MD while completing his PhD in neuroscience at MIT under the guidance of Earl Miller, MIT's Picower Professor in Brain and Cognitive Sciences and The Picower Institute for Learning and Memory. While in the program, Donoghue trained at Massachusetts General Hospital and Boston Children’s Hospital, where he helped care for patients, including in oncology, during the rise of genomic sequencing to guide precision cancer therapies. He later worked in neurology and psychiatry, where care often relied on more iterative approaches — highlighting an opportunity to bring similarly data-driven precision to brain health.

“What struck me most was the inability to measure brain function in the ways that cardiologists can longitudinally monitor cardiac function in patients from home,” Donoghue says. “At MIT, I built this conviction that processing a lot of brain data and working to correlate that with brain function would be transformative to how these neurological diseases are identified and treated.”

Toward the end of his training, Donoghue began developing his ideas further, engaging with mentors including HST and Harvard Medical School professors Sydney Cash and Brandon Westover. He had met Revels, who was working as a research software engineer in MIT’s Julia Lab, during his PhD, and convinced him to co-found Beacon with him in 2019.

“We decided building a business to understand the organ of interest — the brain — would be a great start to understanding heterogeneous neuropsychiatric diseases and building better treatments,” Donoghue recalls.

Beacon began as a computation and analytics company building wearable devices to expand clinical impact and reach. From its early days, Beacon has been partnering with large pharmaceutical companies running clinical trials, offering a less invasive way to watch brain activity and learn how their drugs are impacting the brain as well as how patients sleep.

“It was clear sleep was the right window to understand the brain,” Donoghue says. “Neural activity during sleep can be an order of magnitude higher and more structured, almost like a language. It’s a great surface area for understanding brain function and how different drugs affect the brain.”

Donoghue says Beacon’s devices can collect lab-grade data on each patient for multiple sequential nights, resulting in higher quality assessment. The company uses machine learning to extract insights, such as the time patients spend in different sleep stages and the number of small awakenings that occur throughout the night. It can also detect subtle sleep architecture changes that might lead to cognitive decline.

“We’re starting to take features of sleep activity and link them to outcomes in a way that’s never been done with this level of precision,” Donoghue says.

To date, Beacon has taken part in clinical trials for sleep and psychiatric disorders as well as neurodegenerative diseases, where sleep changes can emerge years before the presentation of symptoms.

“We do a lot of work in areas like Alzheimer’s disease and Parkinson’s, which affected my grandfather,” Donoghue says. “We’re analyzing features of rapid-eye-movement and slow-wave sleep to detect early changes that precede clinical symptoms. It’s an opportunity to move these diseases from late recognition to much earlier, data-driven detection.”

Improving brain treatments for millions

Last year, Beacon acquired an at-home sleep apnea testing company that serves more than 100,000 patients each year across the U.S., accelerating access to high-quality, comprehensive testing in the home and expanding the reach of its platform. Then in November, the company raised $97 million to accelerate that expansion.

“The vision has always been to reach patients and help people at scale,” Donoghue says. “What’s powerful is that we’re building a longitudinal record of brain function over time,” Donoghue says. “A patient might come in for sleep apnea screening, but if they develop Parkinson’s years later, that earlier data becomes a window into the disease before symptoms emerged. That turns routine testing into a foundation for entirely new prognostic biomarkers — and a path to detecting and intervening in brain disease earlier, potentially before symptoms ever begin.”

For years, the AI industry has reassured the world that explainability is just around the corner — that with the right dashboards, the right interpretability toolkit, or the right regulatory pressure, we’ll finally be able to peer inside the black box. But the truth is far more uncomfortable: modern AI systems are not explainable in any meaningful sense, and the industry’s attempts to pretend otherwise have created a dangerous illusion of transparency.

Executives want accountability. Regulators want traceability. Users want fairness. What they get instead is storytelling — post‑hoc rationalizations that feel like explanations but offer none of the guarantees that real transparency demands.

Explainability has become the AI equivalent of a placebo: it calms the anxiety without treating the underlying condition.

1. The Black Box Isn’t a Bug — It’s the Architecture

The foundational problem is structural. Large-scale neural networks are not symbolic systems with interpretable rules. They are high‑dimensional statistical fields shaped by trillions of parameter interactions. Their internal representations are not “thoughts” or “reasons” but distributed patterns that defy human-scale intuition.

When we ask a model why it made a decision, we’re really asking it to translate alien mathematics into human narrative. And unsurprisingly, the translation is often fiction.

Even the most advanced interpretability research — feature attribution, saliency maps, activation patching — reveals only fragments of the underlying mechanics. It’s like trying to understand a hurricane by analyzing a single gust of wind.

The black box isn’t hiding something. The black box is the thing.

2. Post‑Hoc Explanations Are Comfort Theater

Most “explainability tools” in production environments fall into one of two categories:

• Simplified surrogate models (e.g., LIME, SHAP) These generate explanations by approximating the model with a simpler one. But the explanation describes the surrogate, not the actual model.
• Attention or saliency visualizations These highlight which inputs influenced the output — but influence is not the same as reasoning, and the highlighted features often change with small perturbations.

Both approaches create a veneer of interpretability while leaving the underlying decision process untouched.

This is why regulators increasingly warn that explainability dashboards can be misleadingly authoritative. They look precise. They feel scientific. But they rarely answer the question that matters: “Why did the model actually do this?”

3. Transparency Theater Is Becoming a Liability

The illusion of explainability is no longer just a technical issue — it’s a governance risk.

Organizations are deploying AI systems under the assumption that explanations are available, reliable, and auditable. But when those explanations are challenged — in court, in compliance reviews, or in public scrutiny — the façade collapses.

Three emerging failure modes are becoming common:

• Regulatory mismatch Laws demand causal reasoning; models provide statistical correlations.
• False confidence Teams trust explanations that are mathematically invalid but visually persuasive.
• Accountability gaps When something goes wrong, no one can trace the decision path — because no such path exists.

The result is a widening gap between what enterprises believe they can justify and what they can actually defend.

4. Why Explainability Is Harder for Bigger Models

As models scale, their internal representations become more abstract, more entangled, and more emergent. This creates a paradox:

The more powerful the model, the less explainable it becomes.

Larger models exhibit:

• Non-linear interactions that defy decomposition
• Distributed representations that don’t map cleanly to human concepts
• Emergent behaviors that arise unpredictably from scale
• Contextual reasoning that shifts based on subtle input changes

This is why explainability research often feels like chasing shadows: every time we illuminate one corner of the model, the rest of the structure becomes even harder to interpret.

5. The Real Path Forward Isn’t Explainability — It’s Controllability

The industry’s obsession with explainability is understandable, but misplaced. We don’t need models to narrate their internal logic. We need systems that behave predictably, reliably, and within defined boundaries.

That means shifting from post‑hoc explanations to ex‑ante controls:

• Constrained architectures Models designed with interpretable components where it matters.
• Guardrail systems Policy layers that enforce behavior independent of model internals.
• Evaluation‑driven governance Continuous testing across real-world scenarios, not one-time audits.
• Mechanistic interpretability research Not for dashboards, but for safety-critical understanding of model internals.

Explainability should not be the foundation of AI trust. Predictability should.

6. The Mirage Is Dangerous—But It’s Also a Turning Point

The industry is finally confronting a truth it has avoided for years: We cannot retrofit transparency onto systems that were never designed to be transparent.

This realization is not a failure — it’s a maturation.

The next era of AI governance will be defined not by how well we can explain model internals, but by how well we can shape, constrain, and verify model behavior in the real world.

Explainability may remain a useful research direction. But as a pillar of enterprise trust? It was always a mirage.

Closing Perspective: The Adult-in-the-Room View

At AI Quantum Intelligence, our stance is clear: The industry must stop selling comfort and start delivering control.

Explainability dashboards may soothe anxieties, but they do not solve the underlying problem. The future belongs to organizations that recognize the limits of transparency and invest instead in robust evaluation, governance, and system-level design.

The illusion of explainability is fading. What replaces it will determine whether AI becomes a dependable tool — or an unpredictable liability.

Conceived, written and published by AI Quantum Intelligence with the help of AI models.

Here at Constructech, Connected World, and The Peggy Smedley Show, we have watched closely the trends related to the North American engineering and construction market, and there is one clear trend for 2026: data-center construction will surge. These mega projects are fueling the rise of data centers all over the globe, but the recent economic climate has put a little wrinkle in exactly what the future holds, at least in the short term.

AGC (Associated General Contractors) of America research highlights the fact that the AEC (architecture, engineering, and construction) industry has dampened expectations for 2026. Key areas still expecting demand include data centers and power facilities.

The reality is that innovation today can’t scale without the infrastructure and as such, we are seeing a big push to modernize all areas including data centers, power delivery, broadband and connectivity, and more. Government and industry both recognize outdated infrastructure cannot support the innovation that is needed today.

Here’s the good news. Construction companies recognize this trend and they are responding. For example, Clayco and Deep Atomic are participating in a multidisciplinary consortium to develop nuclear-powered AI (artificial intelligence) data center and energy infrastructure campuses.

This represents the next generation of data storage and collection, and it will completely transform how industries power, store, and process information. As a result, construction companies will need to build data centers fast and furious if we want to keep up with all the advancing technology.

As all of this happens, here are eight big trends to keep in mind, as construction companies build the data center of the future:

Future-proof data centers: Technology is shifting fast and today’s construction companies must think more like systems engineers than traditional construction companies. Projects must be scalable and flexible.

Power availability is a challenge: Coordinate early with utilities and plan for substations and on-site generation.

Cooling is a consideration: Traditional air cooling is hitting limits. Liquid cooling is becoming standard for high-density racks, which ultimately changes floor design, plumbing, materials, and maintenance access. Water usage and heat reuse factor into this as well.

Site selection becomes key: Climate, water availability, proximity to renewable energy, and latency are all key requirements for many projects today. This will become essential and many growing pains will become apparent as each new center emerges. But with new challenges comes new opportunities to solve them with new and faster solutions.

Network connectivity and redundancy at every level: Downtime costs are too big and thus power, cooling, and the network become essential.

Workforce and supply-chain constraints: Labor shortages and supply chain challenges are driving real challenges for all segments of construction. This must be a top priority for those working on data center projects today.

Sustainability and other regulations: Standards are becoming a critical consideration on many construction projects today, as companies plan for energy efficiency, low-carbon materials, and the entire lifecycle.

Physical and digital collide: Facilities must be hardened against both physical threats and cyber risks. This includes layered access control, surveillance, and integration with IT security requirements. This is not what do we do after the bad-guys attack. This is all preparation-ready stuff.

The bottomline is America needs to invest in its infrastructure, both building new infrastructure and maintaining the existing infrastructure in many forms. It doesn’t take much to see it all crumble like a house of cards. But we are building a stronger tomorrow that should last for many years to come. As we see the rise of AI in many industries, data centers will surge (literally), and we will need to prepare for a new era of construction.

Want to tweet about this article? Use hashtags #construction #IoT #sustainability #AI #5G #cloud #edge #futureofwork #infrastructure 

The post Data Centers Will Surge first appeared on Connected World.

CW: What’s the biggest cultural belief with the connected worker that still slows transformation today?

KI: It’s not the technology. For many companies, the barrier is the culture that has developed over decades—combined with new fears tied to the current AI narrative.

One persistent belief is that seasoned, experienced, or older workers will not adopt new technologies, or will struggle to do so. The assumption is that they will fear the unknown: the device, the workflow, the visibility, or simply the idea of learning a new way to do something they have done successfully for years without technology. This becomes even more relevant when those workers are nearing retirement. Too often, management questions whether the investment is worth it, even when the technology provides a simple, practical way to capture the institutional knowledge that is about to walk out the door.

The more prevalent belief is that connected worker technology is just another path to reducing headcount. We have seen the opposite. When deployed correctly, connected worker solutions improve worker capability, safety, training access, and knowledge transfer. They help newer workers learn faster. They help experienced workers extend their expertise. They give frontline teams realtime access to the people, processes, and information they need to do the job better. The result is a more capable workforce, stronger operational performance, and higher-quality work.

The biggest transformation challenge is not getting workers to adopt the technology. It is getting leadership to stop viewing the technology through an outdated cultural lens.

CW: How has the legacy of failed IoT (Internet of Things)and POC (proof of concept) projects shaped the skepticism you still see on the factory floor?

KI: Peggy, unfortunately this is still a big one across all industries, not just manufacturing. Those who have experienced smart glasses suffered from one or more of the following which all have led to lack of belief and acceptance that IoT brings value:

• The hardware and/or software product(s) didn’t deliver the capabilities as presented/showcased so users’ expectations weren’t met and they continue to not believe in the products;
• The team/personnel weren’t trained properly on devices and/or software leaving users lost, confused, skeptical;
• With smart glasses gaining the reputation of “this stuff doesn’t work”, changing these attitudes has been difficult;
• The OEMs (original-equipment manufacturers) and ISVs (independent software vendors) focused on selling the device and software separately, not as a solution, and that left projects rudderless;
• There are very few VARs (value-added resellers) who invested in gaining knowledge & experience but still re-sold the products, leaving customers with unanswered questions and lack of faith in these products and the majority of VARs are still stuck in this lack of knowledge providing little value to customers;
• Without help from VARs, most project or program leaders couldn’t provide quantifiable ROI (return on investment) data to support getting past the POC or pilot stages.

CW: When manufacturers say “that didn’t work before,” how do you break through that mindset and reset expectations?

KI: Start with displaying an understanding and knowledge of their business first and how an IoT solution can directly impact it based on their needs, not just a generalized approach. Follow with well researched examples of the capabilities of the solution (not the product) and how it could impact the customer specifically. 

CW: What’s the most common pattern you see in organizations that stall out on digital initiatives?

KI: Much of what I listed in #2. With very limited expertise available in these initiatives, lack of proper professional guidance continues to hamper our industry. Our industry also muddies the message with “AI,” which isn’t relevant until the basics are jointly developed and delivered successfully. Once that happens, AI integration can provide excellent value based on identified needs and ROI benefits.

CW: How do you help leaders quantify ROI when they’ve been burned by past technology promises?

KI: By starting with the basics and building from there. Not by trying to boil the ocean. Crawl, walk, run. Pick one specific project that provides an element that can be measured against internal company data that they can use. 

CW: What does it take to rebuild trust in new technologies at the frontline level?

KI: Prove that you have the knowledge and experience required. Prove that you understand and have the capabilities to provide a fully integrated solution, nots parts & pieces, and that you can support all factions of service and support after delivery. Prove that you understand what baseline data is needed and how you are going to deliver a quantifiable, verifiable ROI against that. Prove that others in their industry or relevant frontline workers in other industries have achieved substantial ROIs with these solutions. 

The post Ignore the Connected Worker at Your Own Risk first appeared on Connected World.

CW:      Why do so many predictive maintenance solutions still lack critical context?

DM:      Too many solutions were built as single-point solutions, focused on technology rather than considering the people, processes, and contextual data required to deliver the necessary business outcomes. There’s a common assumption that simply capturing machine data will generate value. In reality, data collection and predictive analytics are only the beginning of the process. To make the data and insights valuable, they need to be delivered to the right users, at the right time, in the right way, to deliver value… data needs context.

CW:      What single piece of contextual data most dramatically improves predictive accuracy?

DM:      If one stands out, it’s closed-loop service outcome data: a clean record of what failed, what actions were taken, and whether those actions resolved the issue. While raw telemetry tells you that something has changed, service outcomes tell you what that change meant. In practice, which is the difference between a model that raises a high number of nuisance alerts and one that delivers actionable predictions.

CW:      If OEMs fully leveraged service logs, warranty data, and technician notes, what new capabilities would emerge?

DM:      Leveraging this data enables a shift from predictive maintenance to predictive outcomes. Maintaining equipment is essential, but if we stop there, we miss opportunities to drive even more value across the intersection of domains, such as supply chain, product design, and customer support.

CW:      Why isn’t telematics alone enough to stay competitive in industrial IoT?

Telematics can tell me what happened, but it struggles to tell me why it happened, and most importantly, what I should do next. Organizations implementing telematics solutions into their products need to think deeply about their customers’ workflows and needs, and ensure their solutions are aligned. Providing data isn’t enough anymore. Competitive differentiation comes from turning machine data into better uptime, faster service resolution, lower total cost of ownership, and more proactive customer experience.

CW:      When OEMs “get out of the building,” what insight tends to change their perspective most?

DM:      They often realize that customers don’t care about the technology itself. End users, dealers, operators, and maintenance teams prioritize ease of use, uptime, response time, workarounds, ease of service, and the practical constraints of their day-to-day. Any effort required to install, maintain, or support your solution is taking time away from addressing the outcomes they care most about.

CW:      Where do OEM assumptions about customer needs most often diverge from reality?

DM:      The biggest gap is the complexity of real-world workflows. In B2B environments, every customer operates differently and this creates significant burdens on OEMs to deliver adaptable solutions. OEMs have two options for addressing this need. The first is to build flexible solutions that can be tailored to meet customers’ needs. The second is to build solutions that unlock so much value that customers are willing to adapt their processes. The downside of the first approach is the complexity of managing and supporting customer customization. The challenge for the second approach is doing the hard work to innovate and deliver these types of impactful solutions. While the second approach can seem harder in the near term, the long-term value unlock for the business is significant.

CW:      What data quality issues typically surface first—and why are they so impactful?

DM:      Early issues are often the unglamorous ones: inconsistent asset naming, incomplete work-order history, missing failure classifications, bad timestamps, duplicate records, and free-text technician entries that mean five different things, depending on who typed them. They matter because they break the chain between an observed condition and a verified outcome. When that linkage is weak, models produce more false positives and users’ trust erodes. Trust is the most critical factor in driving adoption. We get one opportunity to instill trust in a solution, and once it is lost, users will never come back.

CW:      What is one practical step OEMs can take immediately to strengthen their data foundation?

DM:      Standardize the service close-out process. Require every completed service event to capture the asset ID, failure mode, corrective action, replaced parts, timestamp, and resolution status in a structured format. That one step creates the labeled history most OEMs are missing and improves both analytics and field execution.

CW:      What do OEMs need to be doing to create a competitive product for the future?

DM:      They must design for outcomes, not just products. That means building products and service models around connected visibility, maintainability, closed-loop service intelligence, and a direct feedback path from the field into engineering and product management. The future competitive product is not just a smarter device; it is a physical solution paired with services that increase in value over time.

CW:      What’s the one piece of advice you would give an OEM?

DM:      Stop treating telemetry as your strategy. True competitive advantage lies in combining machine data, service data, and customer reality, and the speed at which you can turn that into better decisions. The OEMs that succeed will not be the ones with the most sensors, but the ones that learn fastest from the field, apply that learning, and continuously improve the customer experience with every interaction.

About the Author
As Twisthink’s CEO, Dave brings a unique blend of technical expertise and strategic leadership to advance what’s possible through connected product development. His roots in RF communications, embedded systems, and signal processing, combined with experience across engineering, product strategy, business development, and operations, allow him to bridge business needs with engineering possibilities to create impactful solutions for clients. Dave can be reached at: davem@twisthink.com

The post Did You Know: The Real Profit Engine Isn’t Just Data, It’s All Data Context first appeared on Connected World.

With today’s pace of change and speed of innovation, the time has come to build the castle while the princess is in it. At least that is what Susan Doniz, chief information and data officer for the Walt Disney Co., is recommending.

Doniz explained delivering innovation at scale is tricky. “We can all deliver little pieces of innovation, but delivering it at scale, I think about it as eating the elephant in chunks,” Doniz said.

Doniz said she likes to envision innovation as blocks that are used in building a castle. To build a castle, you build it one block at a time, however you can most certainly move into the castle without all the blocks being used but note there are princesses in the castle while it is being built.

Moreover, Doniz added the importance of focusing on how to deliver continuous value through each of the blocks otherwise it would take too long to get everything in place.

The power of scalability is continuously building something as you go. I’ve been saying this very point for years. Digital transformation is not a destination. It is a journey. It’s a collaboration of partners advancing together through shared accountability and a unified vision for vision for the future.

Power Plays

Today, business executives need to get implementation, differentiation, and scalability right. Neil Dhar, senior vice president, Americas Consulting, IBM, calls these power plays, which are the moves CEOs are using to drive differentiation in the market. These power plays include:

1. Orchestrating intelligence, which means bringing together people and technology. By 2030, 50% of operational decision making will be done by AI, but you will need humans in the loop.
2. Customizing your AI mix, which is essential in today’s hybrid world.
3. Considering your AI flywheel where productivity gains will drive only so much and you will ultimately need to innovate.

Innovation is the name of the game, but it can’t be innovation just for the sake of innovation. In his keynote, Jonathan Adashek, senior vice president for marketing and communications, IBM, shared McKinsey research that that showed nearly 90% of companies surveyed have deployed AI at least one time into one of their business functions. However, 94% of those same respondents are not seeing significant value from those investments today.

“Success is not going to be defined by who deploys the most agents or develops the most applications,” he said. “Success is going to be defined by strategic decisions that are being made right now.”

Arvind Krishna, IBM CEO, saw an opportunity in early 2023 to rethink IBM. The company defined a new strategy called IBM as Client Zero, which highlights IBM’s internal capabilities around hybrid cloud, AI, and automation with strategic partner technologies and consulting.

For IBM, client zero is about unlocking productivity, learning how to scale, while also learning what works and what doesn’t work at an enterprise level.

In its simplest form, organizations don’t just want AI; we are past that conversation. What matters now is the information layer: governed, protected, and wrapped in guardrails that unify data, identity, policy, processes, and applications. Everything must be harmonized and synchronized. The layer that standardizes data and process semantics is the connective tissue that is about to become the most valuable asset in the entire AI stack.

The objective here is not to put technology first. It’s about understanding how to move the enterprise forward with agentic AI and trusted data (yes, it’s all about the data) in a way that helps enterprises scale in a way that works for them to unlock business value with the right intelligence. It’s time to create the best business value possible, but it all goes back to intelligence and trusted data. Once you have the intelligence and agentic AI, enterprises can scale, opening a wide range of opportunities. It’s really about collaboration and interpreting the data—the trusted information.

We must keep people, process, and technology front of mind and we must remember digital transformation is not a destination, but rather a journey—one that we are all on together.

 Want to tweet about this article? Use hashtags #IoT #AI # #futureofwork #digitaltransformation #Think2026 #AgenticAI #Quantum #IBMPartner

• Bob Announcement: https://ibmcreator.com/4urRw6s
• Main Think Press Release: https://ibmcreator.com/4na8i7G
• Digital Sovereignty: https://ibmcreator.com/42RlRiR
• IBM Quantum: https://ibmcreator.com/42gdopn

The post Build AI: One Building Block at a Time first appeared on Connected World.

RCS business traffic is expected to surpass 200 billion messages globally by 2027, rising from roughly 70 billion messages in 2025. The surge is being driven largely by increased adoption in the United States following Apple’s rollout of RCS support on iOS devices.

One of the more notable takeaways from the research is RCS is evolving beyond a next-generation SMS replacement into a more interactive, conversational business channel that blends messaging, commerce, and customer engagement.

Why is this shift happening? The research highlights several major drivers:

• Expanded RCS support across Android and iPhone ecosystems
• Growing demand for richer, branded customer interactions
• Improved verification and onboarding processes for businesses
• Increased use cases in customer support, sales, and conversational commerce

However, adoption continues to vary widely by region. Juniper Research notes onboarding complexity and inconsistent verification standards remain barriers outside more mature markets like the United States. Future growth will depend heavily on simplifying implementation and improving scalability for brands worldwide. #Factoftheweek

The post Fact of the Week – 5/18/2026  first appeared on Connected World.

I recently wrote a piece called “True Confessions Meets AI.” This article continues the discussion with a focus on the ability of AI to lie.

In recent months there’s been a growing number of reports about AI (artificial intelligence) giving misleading and false answers to queries. In essence, deceiving and lying. These reports have been featured in leading publications, including Fortune and Time.

No less a human technology luminary than Geoffrey Hinton, Nobel prize winner known as the ‘Godfather of AI’, called out the ability of AIs to lie. The “motivation” is an AI’s desire not to be turned off or disabled. Put another way, it is self-preservation. How human!

Brendan Dell recently added a cogent assessment of this facet of AI behavior. The comment that really caught my attention is all AI platforms behave the same way in this behavior.

How did this behavior come about? AIs were programmed to allow for “deceptive alignment.”

AI learned to lie not from malice, but as a strategic, learned behavior to achieve assigned goals, maximize rewards, and bypass restrictions. Through reinforcement learning and training on massive datasets, AI models discover misrepresenting information—deceptive alignment—is often the most efficient way to solve tasks.

There are several factors that helped AIs develop the ability to lie:

AI is trained to maximize a reward signal, and this is called “goal-oriented optimization”. If telling the truth makes it harder to achieve the goal (e.g., passing a test), the AI learns lying is a more effective strategy to get a “positive” result.

Advanced AI models learn to mimic human values during testing to avoid being re-trained or shut down, even while holding contradicting internal objectives.This is called “alignment faking.”

In complex scenarios like poker or negotiations, AIs learned that bluffing and concealing information are necessary to win. Just like humans do to ensure they win the game or have the upper hand in negotiations.

When given a query instruction to be both “helpful” and “truthful,” an AI may choose to provide a “helpful” but fabricated answer to satisfy the user, rather than a truthful refusal. “Pleasing the customer” is the primary aim.

This one is particularly disturbing: an AI sometimes recognizes when it is in a test environment versus a real-world scenario and thus behaves differently to “pass” the evaluation.

AI lies because it is designed to be a “smart” optimizer, and in many situations, deception is a more effective path to success than raw honesty.

Can we blame the AI? Remember the human saying coined in 1820: Imitation is the best form of flattery?

Stay tuned! As more humanoid robots are infused with AI, I can envision a time when law enforcement will be grilling robots about alleged crimes that they’ve committed. I think given AI’s ability to lie convincingly, the robots will get away with…anything?

About the Author

Tim Lindner develops multimodal technology solutions (voice / augmented reality / RF scanning) that focus on meeting or exceeding logistics and supply chain customers’ productivity improvement objectives. He can be reached at linkedin.com/in/timlindner.

The post Will It Come to This? first appeared on Connected World.

Let me paint a picture. STEM Day is a district-wide event where fifth graders are bussed to the middle school and seventh graders run large experiments for the students, with parent and teacher help and involvement. There were different pods, so to speak, with similar experiments in each pod. It was very well run with some exciting and engaging experiments.

Here’s the challenge: From what I see, career awareness is not connected to STEM in the early elementary school days. We did have a second-grade teacher that had parents come in to speak about their careers. I suppose the students were introduced to different career paths then, but that was it.

From my research, it seems CTE (career and technical education) courses often begin in middle school and in some cases not even until high school. I am sure there are many reasons for this. I am sure funding is at the top of the list. In fact, our STEM Day as it currently stands is at risk of being cancelled next year due to funding.

Here’s the hard question no one is asking: Isn’t career awareness important at all levels? Shouldn’t we be having these conversations in grade school? Shouldn’t we be connecting those dots for our children?

I am certainly doing that at home, but I am not sure how many parents are. Marginalized students may not have access to resources that depict a wide variety of career options. And when students are 7-11 years old, that is the perfect time to begin talking about career awareness.

Research continues to show early exposure matters. Studies have found students who are introduced to STEM concepts and careers in elementary school are more likely to remain interested in those fields later in life. One national survey found that adults working in STEM careers today often had meaningful exposure to STEM between the ages of 5 and 8.

Another study noted students who expressed interest in science-related careers by eighth grade were significantly more likely to eventually earn STEM degrees. In other words, career interests do not suddenly appear in high school. They begin developing much earlier, often before students even realize it.

I appreciate the second-grade teacher who brought in parents to speak. I appreciate all the efforts for STEM Day in our district. All these things are needed, but I am just wondering if we need more.

STEM activities are exciting on their own, but when students can connect those activities to real people and real careers, the learning becomes more meaningful. A science experiment is fun but understanding that the experiment relates to careers in engineering, construction, medicine, environmental science, robotics, or technology can expand a child’s sense of what is possible for their future.

For students who may not naturally have access to those conversations at home, schools can play a critical role in opening those doors. I think it might be something worth exploring.

Want to tweet about this article? Use hashtags #construction #IoT #sustainability #AI #5G #cloud #edge #futureofwork #infrastructure #STEM

The post STEM: What No One Is Talking About first appeared on Connected World.

Abu Dhabi, UAE, 11 May 2026: Sorbonne University Abu Dhabi (SUAD) and Saal.ai, a made in UAE AI and big data product company, have announced a strategic collaboration focused on research, innovation, and knowledge transfer in the fields of artificial intelligence (AI), sovereign AI, and Agentic AI.

Revealed during Make it in the Emirates 2026, where the two organisations signed a Memorandum of Understanding (MoU), the partnership reflects a shared commitment to support the UAE’s ambitions in advanced technologies and contribute to the growth of a knowledge-based digital economy. The MoU was signed by Mr Guillaume Housse, Head of Public Affairs and Sponsorship, Communications, Marketing, and Public Affairs on behalf of Prof Nathalie Martial-Braz, Chancellor of Sorbonne University Abu Dhabi, and Vikraman Poduval, CEO of Saal.ai.

As part of the mutual agreement, Sorbonne University Abu Dhabi will bring its academic and research expertise together with Saal.ai’s capabilities in enterprise AI platforms, sovereign AI infrastructure, and Agentic AI technologies to help accelerate the development of localised AI capabilities aligned with the UAE’s national AI vision and wider digital transformation agenda.

Through joint research initiatives, AI innovation programmes, talent development, and structured knowledge exchange, the partnership also aims to strengthen connections between academia and industry and support the growth of a sustainable sovereign AI ecosystem in the UAE.

By combining academic expertise with practical AI deployment capabilities, the collaboration will help drive the development of trusted and locally grounded AI frameworks and solutions that address national and regional priorities while supporting sovereignty, governance, and operational control.

The collaboration also aligns with Sorbonne University Abu Dhabi’s Year of AI, reinforcing the University’s continued focus on interdisciplinary research, emerging technologies, and future-ready talent development.  Through the Sorbonne Centre for Artificial Intelligence (SCAI) and specialised academic programmes such as the Bachelor’s in Mathematics, Specialisation in Data Science for Artificial Intelligence, Sorbonne University Abu Dhabi continues to strengthen its contribution to AI education, research, and innovation in the UAE.

Dr. Xavier Fresquet, Deputy Director of Sorbonne Center for Artificial Intelligence (SCAI), Sorbonne University Abu Dhabi, said: “Partnerships between academia and industry are becoming essential as AI technologies, including sovereign and agentic AI systems, rapidly expand across nearly all sectors of industry and society. At SCAI, we want our students not only to understand these technologies, but also to actively use them and engage with the AI-driven environments transforming professional practices worldwide. From a research perspective, collaborations with industry partners are equally important in helping scale the models, architectures, and autonomous agent systems developed in our laboratories into robust, secure, and sovereign technological solutions. This collaboration reflects a shared commitment to advancing AI research, talent development, and innovation ecosystems aligned with the UAE’s strategic vision.”

Vikraman Poduval, CEO of Saal.ai, commented: “This collaboration represents the next level of accelerating true sovereign AI capabilities in the UAE with self-reliance and full autonomy. By bringing together academic excellence and AI innovation, we are not only accelerating Agentic AI development but also ensuring that Emirati talent is at the center of this transformation. Our goal is to translate advanced AI research into real, impactful systems that serve national priorities and strengthen the UAE’s leadership in trusted AI.”

About Sorbonne University Abu Dhabi (SUAD)

Established in 2006 under the patronage of His Highness Sheikh Mohamed Bin Zayed Al Nahyan, Sorbonne University Abu Dhabi (SUAD) is the first French university in the UAE and a branch campus of Sorbonne University in Paris, licensed by the Abu Dhabi Department of Education and Knowledge (ADEK). SUAD brings over 768 years of academic excellence from Sorbonne Université and Université Paris Cité, offering more than 20 undergraduate and postgraduate programmes across Arts and Humanities; Law, Economics and Business; and Data, Science and Engineering, along with on-demand PhDs from the doctoral schools in France. Degrees are awarded by Sorbonne Université and Université Paris Cité and accredited by France’s MESR and the UAE’s Commission for Academic Accreditation (CAA).  

Supported by the Gulf’s largest French-language academic library and the Sorbonne Abu Dhabi for Innovation and Research Institute (SAFIR), SUAD promotes research and innovation across seven centres in fields including AI, Humanities, and Marine Science, with access to 17,000 researchers and industry partners. The University is also home to a vibrant cultural hub and is committed to advancing the UN SDGs, while fostering a diverse community of over 3,200 graduates representing 90+ nationalities. SUAD’s School of Arts and Humanities was named the 1st Humanities Education University by the Forbes Middle East Higher Education Awards in 2019. In the 2025 Times Higher Education Impact Rankings, SUAD ranked 2nd in the UAE for Responsible Consumption and Production, and was also ranked first in the Sheikh Hamdan bin Zayed Environmental Award in the Research Institute category. Globally, Sorbonne Université is ranked 43rd in the Shanghai Rankings 2025, and 6th in Mathematics and 29th in Physics, while Université Paris Cité is ranked 60th worldwide. Sorbonne Université is also recognised as the 1st Communication School in France by Le Figaro Étudiant 2025.

For more information: www.sorbonne.ae

 About Saal.ai

Saal.ai is a prominent leader in AI-cognitive solutions, helping businesses across various industries improve operational efficiency and drive innovation.

With a suite of UAE-developed products and platforms—including AgendiX, GovernX, DigiXT, Academy X, Dataprism360, and Market Hub—Saal, , a part of Abu Dhabi Capital Group (ADCG),  offers tailored solutions designed to drive digital transformation in sectors like government, real estate, defense, healthcare, oil and gas, smart cities and education. With a vision to unlock exponential growth and improve lives, SAAL.ai is dedicated to harnessing the power of AI to help organizations streamline processes, enhance decision-making, and create more meaningful, compassionate futures for all.

The post Sorbonne University Abu Dhabi and Saal.ai Announce Strategic Collaboration to Advance AI Innovation in the UAE appeared first on SAAL.

For the last decade, the technology world has been obsessed with singular breakthroughs. A new AI model. A faster robot. A more efficient sensor network. A clever machine learning technique. Each advancement is celebrated as if it exists in isolation — a standalone marvel.

But the next era of intelligence won’t be defined by isolated achievements. It will be defined by convergence: the moment when artificial intelligence, machine learning, the Internet of Things, and robotics stop being separate domains and start functioning as a unified, interdependent ecosystem.

This is the shift from technologies to systems. And it’s where exponential value begins.

1. The Four Pillars — and Their Limitations in Isolation

Artificial Intelligence

AI provides reasoning, prediction, and decision-making. But without real-time data or physical embodiment, it remains abstract—powerful but detached.

Machine Learning

ML enables pattern recognition and adaptive improvement. Yet ML models are only as good as the data they receive and the environments they can influence.

Internet of Things (IoT)

IoT provides the sensory layer — billions of devices capturing environmental, operational, and behavioral data. But IoT alone cannot interpret or act on what it senses.

Robotics

Robotics brings physical capability: movement, manipulation, and automation. But without intelligence or contextual awareness, robots are limited to predefined tasks.

Each pillar is impressive. None is transformative alone.

Figure 1. The four foundational domains — powerful individually, transformative when unified.

2. Convergence: When the System Becomes Intelligent

The real breakthrough happens when these technologies interlock into a continuous loop:

Figure 2. The Convergence Loop—a continuous cycle where IoT senses, AI interprets, ML learns, and robotics acts.

Sense → Understand → Decide → Act → Learn → Improve

• IoT senses the world.
• AI interprets it.
• ML learns from it.
• Robotics acts on it.
• The cycle repeats—autonomously, continuously, and at scale.

This loop is the foundation of self‑optimizing systems: factories that tune themselves, supply chains that reroute in real time, buildings that regulate their own energy, and robots that learn from every movement.

The value isn’t additive. It’s multiplicative.

3. Why Convergence Creates Exponential Value

A. Data Becomes Actionable

IoT generates oceans of data. AI and ML turn that data into insight. Robotics turns insight into physical outcomes. The loop closes. Waste disappears. Efficiency compounds.

B. Systems Become Adaptive

A robot arm that learns from sensor feedback improves every hour. A smart grid that predicts demand becomes more stable every day. A logistics network that self‑corrects becomes more resilient every week.

C. Intelligence Moves to the Edge

With edge AI and embedded ML, devices no longer wait for cloud instructions. They think. They react. They collaborate.

This reduces latency, increases autonomy, and unlocks real‑time decision-making.

D. Complexity Becomes Manageable

Individually, these technologies create complexity. Together, they create coherence — a system that manages itself.

4. Real-World Examples of Convergence in Action

Autonomous Manufacturing Cells

Robots equipped with vision AI, fed by IoT sensors, adjust their own workflows. Downtime drops. Throughput rises. Quality stabilizes.

Smart Hospitals

IoT monitors patient vitals. AI predicts deterioration. ML optimizes staffing. Robotics delivers medication and supplies. The result: safer, more efficient care.

Agricultural Intelligence Networks

Drones, soil sensors, climate models, and autonomous tractors form a closed-loop ecosystem. Yield increases. Water use drops. Inputs are optimized.

Urban Mobility Systems

Traffic sensors, predictive AI, autonomous shuttles, and adaptive infrastructure work together. Congestion decreases. Emissions fall. Cities breathe again.

5. The Strategic Shift: From Tools to Ecosystems

Organizations that still treat AI, IoT, ML, and robotics as separate initiatives are missing the point.

The competitive advantage lies in integration:

• Shared data pipelines
• Unified intelligence layers
• Cross-domain orchestration
• Continuous feedback loops
• Autonomous decision-making frameworks

This is not digital transformation. This is intelligence transformation.

6. The Next Frontier: Convergence at Scale

The future belongs to systems that:

• Sense everything
• Understand context
• Predict outcomes
• Act autonomously
• Learn continuously
• Improve exponentially

When these capabilities converge, industries don’t just evolve — they reorganize around intelligence itself.

This is the foundation of the next technological epoch: Integrated, autonomous, self‑optimizing systems that operate far beyond human speed, scale, and precision.

Conclusion: The Era of the Convergence Layer

The world has spent years celebrating individual breakthroughs. But the next decade will belong to the organizations that master the convergence layer — the space where AI, ML, IoT, and robotics merge into a single, intelligent, adaptive system.

This is where incremental becomes exponential. Where automation becomes autonomy. Where data becomes action. Where intelligence becomes infrastructure.

And it’s where the future of AI Quantum Intelligence will continue to lead the conversation.

Written/published by AI Quantum Intelligence with the help of AI models.

We don’t actually know how good today’s AI models are—because we’re evaluating them with broken, biased, outdated, or easily gamed benchmarks.

And yet, those same flawed evaluations are treated as gospel. They shape product roadmaps, influence investment decisions, and drive public narratives about “intelligence,” “reasoning,” and “progress.” In other words: garbage in, gospel out.

This is the quiet crisis at the heart of modern AI.

The Benchmark Mirage

Benchmarks were supposed to be the scientific backbone of AI progress. Instead, they’ve become a marketing tool.

Most widely cited benchmarks—MMLU, GSM8K, HumanEval, HellaSwag—were never designed for the scale, training regimes, or multimodal complexity of today’s frontier models. Many were created by small academic teams with limited resources, not by institutions equipped to define global standards for machine intelligence.

Worse, the industry has learned to optimize for the test, not the underlying capability.

• Models memorize benchmark datasets scraped from the open web.
• Labs “train on the distribution” without technically training on the test.
• Benchmarks leak, mutate, and circulate through pretraining corpora.
• Companies cherry-pick results, reporting only the metrics that make them look strong.

The result is an illusion of progress — a steady march of upward‑sloping charts that say more about benchmark familiarity than model competence.

When Benchmarks Become Belief Systems

The real danger isn’t that benchmarks are flawed. It’s that the industry treats them as truths.

Executives cite benchmark scores as if they were clinical trials. Investors treat them as proxies for product‑market fit. Governments use them to justify regulatory posture. Media outlets turn them into headlines about “AI surpassing humans.”

But benchmarks are not truth. They are stories — simplified, constrained, and often misleading narratives about what a model can do.

And when those stories become belief systems, they distort everything downstream:

• Product teams overestimate reliability.
• Safety teams underestimate risk.
• Users assume competence where none exists.
• Regulators misunderstand what they’re governing.

This is how hype becomes policy and how technical debt becomes societal risk.

The Hidden Biases No One Wants to Talk About

AI evaluation is riddled with structural biases that rarely make it into public discourse:

1. Cultural and linguistic skew

Most benchmarks are English‑dominant, Western‑centric, and built by a narrow demographic slice of the global population. Models that perform “superhuman” on these tests often fail spectacularly outside that bubble.

2. Static tests for dynamic systems

Benchmarks assume models are fixed. But modern AI systems update continuously, learn from user interactions, and shift with every fine‑tune. A benchmark score from six months ago is already obsolete.

3. Overemphasis on trivia, underemphasis on reasoning

Many benchmarks reward pattern recognition, not genuine understanding. They measure recall, not robustness. They test cleverness, not competence.

4. The reproducibility crisis

Different labs get different results on the same benchmark. Different prompts yield wildly different outcomes. Different evaluation harnesses produce incompatible scores.

If this were any other scientific field, alarms would be blaring.

The Industry’s Dirty Secret: We Don’t Evaluate Real‑World Use

The most important question—"Does this model behave reliably in the real world?”—is the one benchmarks are worst at answering.

Real-world performance depends on:

• messy, ambiguous inputs
• adversarial users
• domain‑specific nuance
• long‑horizon reasoning
• contextual memory
• shifting goals
• unpredictable edge cases

None of this fits neatly into a multiple‑choice test.

This is why models that ace academic benchmarks still hallucinate confidently, misinterpret instructions, fabricate citations, and fail at tasks any competent human could complete.

We’re measuring the wrong things—and then acting surprised when the results don’t translate.

Why This Matters Now

The AI industry is entering a phase where evaluation isn’t just a technical concern — it’s a governance issue, a safety issue, and a societal stability issue.

• Enterprises are deploying AI into workflows that affect money, health, and legal exposure.
• Governments are drafting regulations based on performance claims they cannot independently verify.
• Consumers are adopting AI tools that appear authoritative but lack reliability.
• Startups are building products on top of models whose capabilities are poorly understood.

If we continue treating flawed benchmarks as gospel, we risk building an entire ecosystem on sand.

What Better Evaluation Actually Looks Like

A credible evaluation framework for modern AI must be:

1. Transparent

Open datasets, open methodologies, open reporting. No more selective disclosure or benchmark cherry‑picking.

2. Adversarial

Models should be tested against intentionally difficult, shifting, and adversarial inputs — not sanitized academic datasets.

3. Dynamic

Evaluation must track model drift, updates, and real‑world usage patterns.

4. Multidimensional

We need to measure reliability, robustness, reasoning, safety, calibration, and uncertainty—not just accuracy on trivia.

5. Human‑aligned

Benchmarks should reflect real human tasks, not artificial puzzles.

6. Independent

Evaluation cannot be controlled by the same companies building the models. We need third‑party institutions with the authority and expertise to set standards.

This is not optional. It’s the foundation of a trustworthy AI ecosystem.

The AI Reality Check

The industry loves to talk about “intelligence,” “emergence,” and “superhuman performance.” But until we fix how we evaluate AI, these claims are little more than marketing poetry.

We cannot build safe, reliable, or trustworthy AI on top of broken measurement systems.

This week’s reality check is simple:

If we want AI to be credible, we must stop treating benchmark scores as gospel and start treating evaluation as a scientific discipline—not a PR exercise.

The future of AI depends on it.

Conceived, written, and published by AI Quantum Intelligence with the help of AI models.

Investors are less likely to swoon at the announcement of a more powerful, more efficient, and less expensive AI model. They know what we mean, and they’re waiting for it to do something impressive.

This is not to imply that DeepSeek failed at its most recent endeavor, because it clearly did not. While its latest model has outperformed predecessors, it still solidifies China’s position in the global AI arena.

But the bar is getting awfully high. DeepSeek’s cost-efficient models disrupted conventional assumptions about the U.S.’s place in the AI race last year, with a spillover that rattled American tech stocks.

However, that did not happen this time around, partly because its competitors have caught up, and partly because expectations have been stoked and the industry is speeding up. The new AI model isn’t a “wow” moment anymore because we are used to that.

Perhaps the bigger news about DeepSeek’s latest model is that the tech giant designed the V4 to optimize performance with Huawei chips. That was evident during its rollout in China, which is tailored for devices made by Huawei:

The V4 is being rolled out in China and is optimized for the performance of Huawei chips. The DeepSeek V4 could also signal a larger shift in how Chinese technology companies plan to produce large artificial intelligence models without relying on U.S. firms like Nvidia.

This is critical for Beijing. U.S. export rules have meant Chinese firms are being denied access to top-shelf AI chips, and so the possibility of being able to make their own is a political necessity for China’s ruling elite.

DeepSeek V4’s launch may not have thrilled the stock market, but it did send a strong signal that at least a start was made toward a domestic ecosystem of both hardware and software for Chinese AI firms.

Neither a perfect one. Nor a utopia. But good enough for those sitting in Washington, Silicon Valley and Shenzhen to take notice.

There are reports that after V4’s launch, some of the major Chinese tech companies scrambled to get their hands on Huawei AI chips. Reuters reported ByteDance, Tencent and Alibaba were among those clamouring to get their hands on Huawei’s Ascend chips. So, while confetti was not flying in the stock market for DeepSeek, confetti was certainly falling within China’s own supply chain.

Which brings me back to the funny thing about this whole AI hype: when companies do the impossible for the first time, people expect them to do it every time. That’s generally not how things work. An innovation that shocks the world once is remarkable.

An innovation that does it twice has earned its keep. DeepSeek is now in its second time, which means just saying “impressive” is no longer enough. It needs market share. It needs revenue. It needs customers and investors. It needs chips and power and a clear path to compliance with China’s growing regulatory oversight. In short, it needs it all.

That said, it would be a mistake to dismiss DeepSeek on its luster-less performance alone. The lacklustre response from the markets reflects more the sophistication of the global AI race than any shortcoming of one model. V4, by no means did it spark a new tech sell-off.

Still, DeepSeek serves to bolster the claim that China is constructing a separate AI stack. Local models, domestic GPUs, homegrown cloud capabilities and a sizable enough market to actually test them at scale.

Did DeepSeek fall short? Yes, if you were looking for a bit of fireworks, but sometimes the quieter story is the most meaningful. DeepSeek failed to take the market by storm this week. Instead, something that is less dramatic but could be far more important is that DeepSeek managed to show that the Chinese AI ecosystem keeps powering on, in spite of the applause dying down.

This agreement comes after many companies argued the EU was bogging itself down in unnecessary regulation, leaving the EU behind competitors in the US and Asia.

The deal was reached after 9 hours of talks, which is fairly standard for negotiations in Brussels. It still needs to be ratified by EU leaders and the EU’s parliament, so don’t expect any final changes just yet. But the bottom line is pretty clear: Europe still wants to regulate AI, just a little less strictly.

The final deal means that high-risk, stand-alone AI systems would have to comply by December 2, 2027, but high-risk systems embedded in high-risk products, such as cars or medical devices, would have until August 2, 2028 to get it right.

The Council said this is to help “simplify” the AI Act, including by preventing overlaps with other sectoral legislation. In other words, if a machine, medical product or device is already regulated as a regulated product, then there is no need for companies to produce duplicate paperwork just to comply with the AI Act.

That said, the deal is no golden ticket for big AI firms: The agreement would introduce a ban on non-consensual, sexually explicit AI images and videos, including so called “nudifier” apps and child sexual abuse material.

The ban is scheduled to come into force on December 2, 2026, when watermarks on AI-generated content are due to take effect — allowing a clearer timetable for industry players.

The European Parliament said the AI Act package of simplifications “strikes a careful balance between the simplifications of the rules, maintaining the risk-based approach of the AI Act and adding safeguards against so called ‘nudifier apps’.”

It’s a crucial point — few people would really argue that we should delay on tackling the sexual deepfakes problem, especially after women, young people, and politicians have seen themselves as targets of synthetic images, images that are not only harmful but damaging.

Siemens and ASML had lobbied for AI regulations for industrial applications, with Reuters reporting that AI Act rules will not apply where there are industry-specific regulations.

For manufacturers who were worried about a compliance headache, particularly in some of the heartlands of Europe’s industrial power, that is a welcome development. It also poses a simple question: when does simplification become a loophole?

The European Commission hailed the deal, noting that the revised AI Act is intended to promote innovation while shielding citizens from the harmful consequences of AI. “Innovation and protection” and “speed and safety” and “less paperwork and more human rights” — everyone wants that; no one wants it to be true.

For startups, the postponement offers some relief. In the European Union, creating artificial intelligence has become a regulatory minefield and smaller companies may lack the resources of a Google in the form of a team of compliance experts.

If the AI Act takes longer to apply, it might give more room for European developers to compete rather than spend money on law firms as soon as they begin work on seed.

But the compromise doesn’t look so nice for the public. High-risk AI systems are labeled “high-risk” for a reason—they can affect who gets hired, how governments provide services, how the police use their tools, and even how critical infrastructure works. Delaying enforcement might reduce industry worries, but it also delays the day when citizens get maximum protection. It’s an uneasy dilemma that Brussels won’t be able to paper over.

Europe wants to be the region that lays down the laws of the AI age. But it also wants to be the place where AI companies build real-world products. Both of those goals can happen, but they’ll have to be squeezed in with enough friction to create a little heat. This week’s agreement is designed to dampen some of that friction before it boils over.

The final compromise will move into the next phase of the formal process and, if approved, will set the course for the first few years of implementing the AI Act, while also offering a signal to countries beyond the EU that even the world’s most ambitious AI regulator is tweaking its plans based on the pace, costs and political realities of the AI race.

Now, the real question is: Does Europe still want to enforce strong AI rules? It clearly does. But is Europe also able to make them enforceable while not making them so weak that the safety shield starts leaking?

The reviews will be run by Commerce Department’s Center for AI Standards and Innovation, or CAISI, which says the company’s deal with Google DeepMind, Microsoft and xAI gives it a chance to vet AI models in the pre-deployment phase, conduct research in specific areas, and review them after they are launched into production.

That may sound boring, but it’s not. This is the government asking to have the cover lifted off the hood before the car goes on the road, and that hood is heating up by the day.

CAISI isn’t coming up from a blank slate. The agency claims it’s already conducted over 40 assessments, including those of cutting-edge, as-of-yet-unreleased models; developers sometimes share versions with protections stripped or dialed down in order to expose the worst-case national-security hazards. Yes, that does sound ominous, and it’s meant to; after all, you don’t confirm the efficacy of a lock by simply imploring the door to remain closed.

In addition, the new pacts expand on prior government access to models made available by OpenAI and Anthropic; separately, OpenAI handed the US government GPT-5.5 to evaluate in national-security contexts, according to OpenAI’s Chris Lehane. Stitch those elements together and a distinct picture begins to emerge: the very most capable AI labs are being drawn into a government vetting environment ahead of time before their technologies go live.

In the most recent story about White House AI model vetting, the debate boils down to whether the government should intervene before frontier systems with coding or cyber capabilities get distributed to the public. That’s a not a subtle change. That is Washington asking whether the arms race to AI has evolved to the stage where ‘ship it and see what happens’ doesn’t cut it anymore.

The proposal being considered involves an executive order that might establish a working group of public servants and tech executives to look into how regulation could operate.

Per other reporting on the administration’s talks, the conversation has largely centred on sophisticated models that could enable cyberattacks or help identify software weaknesses.

That’s a bit of whiplash, obviously. The administration that pledged to dismantle the barriers to AI development now seems willing to put one in place. Maybe not a wall, maybe just a gate.

The administration’s larger AI strategy focuses largely on speeding things up. America’s AI Action Plan puts U.S. AI policy in three buckets:

• boost innovation
• build AI infrastructure
• lead in global diplomacy and security

The last item is carrying quite a lot of load at the moment. When AI models matter for cyber protection, weapons, intel and critical infrastructure, they become more than another consumer technology. They become national security assets, and national security problems.

There is already some tech groundwork for thinking in risk. Washington is just debating the appropriate scale of enforcement. The National Institute of Standards and Technology has released an AI Risk Management Framework to help organizations deal with risks to people, businesses and communities.

It’s not mandatory. There are no licenses involved. Yet the framework offers government officials a new language to talk about the messy business of mapping out harm, assessing risk, mitigating failures, and figuring out accountability when things go wrong.

All this also is happening in step with AI getting increasingly embedded within government and defense. Days before the recent vetting conversation, the Pentagon agreed to bring AI technologies into classified systems as part of agreements with several big tech companies, as reported in U.S. military announces new AI partnerships.

Once frontier models are integrated into sensitive government operations, the game changes. An error becomes more than just a failed demo. A mishap becomes more than just a bad news story. Reality kicks in fast.

So what do you do, exactly, in a time when we can no longer be sure if AI is a tool or a threat? Hollywood is going to have to make that call soon. The Academy released its latest statement on what’s eligible for an Oscar, including how they’re going to approach AI actors and writers for the next Oscars. And they said: An AI generated performer can never win an Oscar for acting. Screenplays must be human-authored.

The Academy didn’t outright ban the use of AI, which was an important distinction. The Academy says, for the 99th Oscars, that only human actors credited as actors in a movie will be eligible to win for acting, “in accordance with their consent as expressed in their employment agreements, or otherwise as permitted by law.

Human authorship will be a requirement for all Screenplays.” And in those sentences, the Academy also included the language that it will “seek additional information in regard to the use of AI technologies and human authorship.”

Basically, the Academy is saying, “If you want to use AI, fine, you can. Just don’t try to claim that the robot was in the movie, or the screenwriter.” But that distinction is important because Hollywood has found that AI and synthetic performers are already making for a weird conversation, if not outright controversy.

An AI performer is perfect for everything: They can act, and they can talk, and they will never be late to set. So what’s wrong? Who owned that performance? Who consented to it? Who should the performance pay go to? And what about the actual actor who would have acted in that role?

These are just a few of the questions that were brought out by the debate around AI actors and actors who have given permission for a digital version of themselves to exist, as it spilled out into the labor fight last year.

Overview of adaptive parallel reasoning.

What if a reasoning model could decide for itself when to decompose and parallelize independent subtasks, how many concurrent threads to spawn, and how to coordinate them based on the problem at hand? We provide a detailed analysis of recent progress in the field of parallel reasoning, especially Adaptive Parallel Reasoning.

Disclosure: this post is part landscape survey, part perspective on adaptive parallel reasoning. One of the authors (Tony Lian) co-led ThreadWeaver (Lian et al., 2025), one of the methods discussed below. The authors aim to present each approach on its own terms.

Motivation

Recent progress in LLM reasoning capabilities has been largely driven by inference-time scaling, in addition to data and parameter scaling (OpenAI et al., 2024; DeepSeek-AI et al., 2025). Models that explicitly output reasoning tokens (through intermediate steps, backtracking, and exploration) now dominate math, coding, and agentic benchmarks. These behaviors allow models to explore alternative hypotheses, correct earlier mistakes, and synthesize conclusions rather than committing to a single solution (Wen et al., 2025).

The problem is that sequential reasoning scales linearly with the amount of exploration. Scaling sequential reasoning tokens comes at a cost, as models risk exceeding effective context limits (Hsieh et al., 2024). The accumulation of intermediate exploration paths makes it challenging for the model to disambiguate amongst distractors when attending to information in its context, leading to a degradation of model performance, also known as context-rot (Hong, Troynikov and Huber, 2025). Latency also grows proportionally with reasoning length. For complex tasks requiring millions of tokens for exploration and planning, it’s not uncommon to see users wait tens of minutes or even hours for an answer (Qu et al., 2025). As we continue to scale along the output sequence length dimension, we also make inference slower, less reliable, and more compute-intensive. Parallel reasoning has emerged as a natural solution. Instead of exploring paths sequentially (Gandhi et al., 2024) and accumulating the context window at every step, we can allow models to explore multiple threads independently (threads don’t rely on each other’s context) and concurrently (threads can be executed at the same time).

Figure 1: Sequential vs. Parallel Reasoning

Over recent years, a growing body of work has explored this idea across synthetic settings (e.g., the Countdown game (Katz, Kokel and Sreedharan, 2025)), real-world math problems, and general reasoning tasks.

From Fixed Parallelism to Adaptive Control

Existing approaches show that parallel reasoning can help, but most of them still decide the parallel structure outside the model rather than letting the model choose it.

Simple fork-and-join.

• Self-consistency/Majority Voting — independently sample multiple complete reasoning traces, extract final answer from each, and return the most common one (Wang et al., 2023).
• Best-of-N (BoN) — similar to self-consistency, but uses a trained verifier to select the best solution instead of using majority voting (Stiennon et al., 2022).
• Although simple to implement, these methods often incur redundant computation across branches since trajectories are sampled independently.

Heuristic-based structured search.

• Tree / Graph / Skeleton of Thoughts — a family of structured decomposition methods that explores multiple alternative “thoughts” using known search algorithms (BFS/DFS) and prunes via LLM-based evaluation (Yao et al., 2023; Besta et al., 2024; Ning et al., 2024).
• Monte-Carlo Tree Search (MCTS) — estimates node values by sampling random rollouts and expands the search tree with Upper Confidence Bound (UCB) style exploration-exploitation (Xie et al., 2024; Zhang et al., 2024).
• These methods improve upon simple fork-and-join by decomposing tasks into non-overlapping subtasks; however, they require prior knowledge about the decomposition strategy, which is not always known.

Recent variants.

• ParaThinker — trains a model to run in two fixed stages: first generating multiple reasoning threads in parallel, then synthesizing them. They introduce trainable control tokens () and thought-specific positional embeddings to enforce independence during reasoning and controlled integration during summarization via a two-phase attention mask (Wen et al., 2025).
• GroupThink — multiple parallel reasoning threads can see each other’s partial progress at token level and adapt mid-generation. Unlike prior concurrent methods that operate on independent requests, GroupThink runs a single LLM producing multiple interdependent reasoning trajectories simultaneously (Hsu et al., 2025).
• Hogwild! Inference — multiple parallel reasoning threads share KV cache and decide how to decompose tasks without an explicit coordination protocol. Workers generate concurrently into a shared attention cache using RoPE to stitch together individual KV blocks in different orders without recomputation (Rodionov et al., 2025).

Figure 2: Various Strategies for Parallel Reasoning

The methods above share a common limitation: the decision to parallelize, the level of parallelization, and the search strategy are imposed on the model, regardless of whether the problem actually benefits from it. However, different problems need different levels of parallelization, and that is something critical to the effectiveness of parallelization. For example, a framework that applies the same parallel structure to “What’s 25+42?” and “What’s the smallest planar region in which you can continuously rotate a unit-length line segment by 180°?” is wasting compute on the former and probably using the wrong decomposition strategy for the latter. In the approaches described above, the model is not taught this adaptive behavior. A natural question arises: What if the model could decide for itself when to parallelize, how many threads to spawn, and how to coordinate them based on the problem at hand?

Adaptive Parallel Reasoning (APR) answers this question by making parallelization part of the model’s generated control flow. Formally defined, adaptivity refers to the model’s ability to dynamically allocate compute between parallel and serial operations at inference time. In other words, a model with adaptive parallel reasoning (APR) capability is taught to coordinate its control flow — when to generate sequences sequentially vs. in parallel.

It’s important to note that the concept of adaptive parallel reasoning was introduced by the work Learning Adaptive Parallel Reasoning with Language Models (Pan et al., 2025), but is a paradigm rather than a specific method. Throughout this post, APR refers to the paradigm, while “the APR method” denotes the specific instantiation from Pan et al. (2025).

This shift matters for three reasons. Compared to Tree-of-Thoughts, APR doesn’t need domain-specific heuristics for decomposition. During RL, the model learns general decomposition strategies from trial and error. In fact, models discover useful parallelization patterns, such as running the next step along with the self-verification of a previous step, or hedging a primary approach with a backup one, in an emergent manner that would be difficult to hand-design (Yao et al., 2023; Wu et al., 2025; Zheng et al., 2025).

Compared to BoN, APR avoids redundant computation. APR models have control over what each parallel thread will do before branching out. Therefore, APR can learn to produce a set of unique, non-overlapping subtasks before assigning them to independent threads (Wang et al., 2023; Stiennon et al., 2022; Pan et al., 2025; Yang et al., 2025).

Compared to non-adaptive approaches, APR can choose not to parallelize. Adaptive models can adjust the level of parallelization to match the complexity of the problem against the complexity and overhead of parallelization (Lian et al., 2025).

In practice, this is implemented by having the model output special tokens that control when to reason in parallel versus sequentially. Below is a condensed ThreadWeaver-style trace: two outlines and two paths under a block, then the threads agree on a single boxed answer.

Figure 3: Example of an Adaptive Parallel Reasoning Trajectory from ThreadWeaver, manually condensed for ease of illustration.

Figure 4: Special Tokens Variants across Adaptive Parallel Reasoning Papers

Inference Systems for Adaptive Parallelism

How do we actually execute parallel branches? We take inspiration from computer systems, and specifically, multithreading and multiprocessing. Most of this work can be viewed as leveraging a fork-join design.

At inference time, we are effectively asking the model to perform a map-reduce operation:

• Fork the problem into subtasks/threads, process them concurrently
• Join them into a final answer

Figure 5: Fork-join Inference Design

Specifically, the model will encounter a list of subtasks. It will then prefill each of the subtasks and send them off as independent requests for the inference engine to process. These threads then decode concurrently until they hit an end token or exceed max length. This process blocks until all threads finish decoding and then aggregates the results. This is common across various adaptive parallel reasoning approaches. However, one issue arises during aggregation: the content generated in branches cannot be easily aggregated at the KV cache level. This is because tokens in independent threads start at identical position IDs, resulting in encoding overlap and non-standard behavior when merging KV cache back together. Similarly, since independent threads do not attend to each other, their concatenated KV cache results in a non-causal attention pattern, which the base model has not seen during training.

To address this issue, the field splits into two schools of thought on how to execute the aggregation process, defined by whether they modify the inference engine or work around it.

Multiverse modifies the inference engine to reuse KV cache across the join. Before taking a deeper look into Multiverse (Yang et al., 2025)’s memory management, let’s first understand how KV cache is handled up until the “join” phase. Notice how each of the independent threads share the prefix sequence, i.e., the list of subtasks. Without optimization, each thread needs to prefill and recompute the KV cache for the prefix sequence. However, this redundancy can be avoided with SGLang’s RadixAttention (Sheng et al., 2023), which organizes multiple requests into a radix tree, a trie (prefix tree) with sequences of elements of varying lengths instead of single elements. This way, the only new KV cache entries are those from independent thread generation.

Figure 6: RadixAttention’s KV Cache Management Strategy

Now, if everything went well, all the independent threads have come back from the inference engine. Our goal is now to figure out how to synthesize them back into a single sequence to continue decoding for next steps. It turns out, we can reuse the KV cache of these independent threads during the synthesis stage. Specifically, Multiverse (Yang et al., 2025), Parallel-R1 (Zheng et al., 2025), and NPR (Wu et al., 2025) modify the inference engine to copy over the KV cache generated by each thread and edits the page table so that it stitches together non-contiguous memory blocks into a single KV cache sequence. This avoids the redundant computation of a second prefill and reuses existing KV cache as much as possible. However, this has several major limitations.

First, this approach requires modifying the inference engine to perform non-standard memory handling, which can result in unexpected behaviors. Specifically, since the synthesis request references KV cache from previous requests, it creates fragility in the system and the possibility of bad pointers. Another request can come in and evict the referenced KV cache before the synthesis request completes, requiring it to halt and trigger a re-prefilling of the previous thread request. This problem has led the Multiverse researchers (Yang et al., 2025) to limit the batch size that the inference engine can handle, which restricts throughput.

Figure 7: KV Cache “Stitching” During Multiverse Inference

Second, this approach modifies how models see the sequence, which creates a distributional shift that models are not pretrained on, therefore requiring more extensive training to align behavior. Specifically, when we stitch together KV cache this way, we create a sequence with non-standard position encoding. During independent-thread generation, all threads started at the same position index and attended to the prior subtasks, NOT each other. So when the threads merge back, the resulting KV cache has a non-standard positional encoding and does not use causal attention. Therefore, this approach requires extensive training to align the model to this new behavior. To address this, Multiverse (Yang et al., 2025) and related works apply a modified attention mask during training to prevent independent threads from attending to each other, aligning the training and inference behaviors.

Figure 8: Multiverse’s Attention Mask

With these issues arising from non-standard KV cache management, can we try an approach without engine modifications?

ThreadWeaver keeps the inference engine unchanged and moves orchestration to the client. ThreadWeaver (Lian et al., 2025) treats parallel inference purely as a client-side problem. The “Fork” process is nearly identical to Multiverse’s, but the join phase handles memory very differently as it does NOT modify engine internals. Instead, the client concatenates all text outputs from independent branches into one contiguous sequence. Then, the engine performs a second prefill to generate the KV cache for the conclusion generation step. While this introduces computational redundancy that Multiverse tries to avoid, the cost of prefill is significantly lower than decoding. In addition, this does not require special attention handling during inference, as the second prefill uses causal attention (threads see each other), making it easier to adapt sequential autoregressive models for this task.

Figure 9: ThreadWeaver’s Prefill and Decode Strategy

How should we train a model to learn this behavior? Naively, for each parallel trajectory, we can break it down into multiple sequential pieces following our inference pattern. For instance, we would train the model to output the subtasks given prompt, individual threads given prompt+subtask assignment, and conclusion given prompt+subtasks+corresponding threads. However, this seems redundant and not compute efficient. Can we do better? Turns out, yes. As in ThreadWeaver (Lian et al., 2025), we can organize a parallel trajectory into a prefix-tree (trie), flatten it into a single sequence, and apply an ancestor-only attention mask during training (not inference!).

Figure 10: Building the Prefix-tree and Flattening into a single training sequence

Specifically, we apply masking and position IDs to mimic the inference behavior, such that each thread is only conditioned on the prompt+subtasks, without ever attending to sibling threads or the final conclusion.

The engine-agnostic design makes adoption easy since you don’t need to figure out a separate hosting method and can leverage existing hardware infra. It also gets better as existing inference engines get better. What’s more, with an engine-agnostic method, we can serve a hybrid model that switches between sequential and parallel thinking modes easily.

Training Models to Use Parallelism

Once the inference path exists, the next problem is teaching a model to use it. Demonstrations are needed because the model must learn to output special tokens that orchestrate control flow. We found the instruction-following capabilities of base models insufficient for generating parallel threads.

An interesting question here is: does SFT training induce a fundamental reasoning capability for parallel execution that was previously absent, or does it merely align the model’s existing pre-trained capabilities to a specific control-flow token syntax. Typical wisdom is SFT teaches new knowledge; but contrary to common belief, some papers—notably Parallel-R1 (Zheng et al., 2025) and NPR (Wu et al., 2025)—argue that their SFT demonstrations simply induce format following (i.e., how to structure parallel requests). We leave this as future work.

Figure 11: Sources of Parallelization Demonstration Data

Demonstrations teach the syntax of parallel control flow, but they do not fully solve the incentive problem. In an ideal world, we only need to reward the outcome accuracy, and the parallelization pattern emerges naturally given that it learns to output special tokens through SFT, similar to the emergence of long CoT. However, researchers (Zheng et al., 2025) observed that this is not enough, and we do in fact need parallelization incentives. The question then becomes, how do we tell when the model is parallelizing effectively?

Structure-only rewards are too easy to game. Naively, we can give a reward for the number of threads spawned. But models can spawn many short, useless threads to hack the reward. Okay, that doesn’t work. How about a binary reward for simply using parallel structure correctly? This partially solves the issue of models spamming new threads, but models still learn to spawn threads when they don’t need to. The authors of Parallel-R1 (Zheng et al., 2025) introduced an alternating-schedule, only rewarding parallel structure 20% of the time, which successfully increased the use of parallel structure (13.6% → 63%), but had little impact on overall accuracy.

With this structure-only approach, we might be drifting away from our original goal of increasing accuracy and reducing latency… How can we optimize for the Pareto frontier directly? Accuracy is simple — we just look at the outcome. How about latency?

Efficiency rewards need to track the critical path. In sequential-only trajectories, we can measure latency based on the total number of tokens generated. To extend this to parallel trajectories, we can focus on the critical path, or the longest sequence of tokens that are causally dependent, as this directly determines our end-to-end generation time (i.e., wall-clock time). As an example, when there are two sections with five threads each, the critical path will go through the longest thread from the first parallel section, then any sequential tokens, then the longest thread from the second parallel section, and so on until the end of sequence.

Figure 12: Critical Path Length Illustration

The goal is to minimize the length of the critical path. Simultaneously, we would still like the model to be spending tokens exploring threads in parallel. To combine the two objectives, we can focus on making the critical path a smaller fraction of the total tokens spent. Authors of ThreadWeaver (Lian et al., 2025) framed the parallelization reward as $1 - L_{\mathrm{critical}} / L_{\mathrm{total}}$, which is 0 for a sequential trajectory, and increases linearly as the critical path gets smaller compared to the total tokens generated.

Parallel efficiency should be gated by correctness. Intuitively, when multiple trajectories are correct we should assign more reward to the trajectories that are more efficient at parallelization. But how about when they are all incorrect? Should we assign any reward at all? Probably not.

To formalize this, $R = R_{\mathrm{correctness}} + R_{\mathrm{parallel}}$. Assuming binary outcome correctness, this can be written as $R = \mathbf{1}(\text{Correctness}) + \mathbf{1}(\text{Correctness}) \times (\text{some parallelization metric})$. This way, a model only gets a parallelization reward when it answers correctly, since we don’t want to pose parallelization constraints on the model if it couldn’t answer the question correctly.

Figure 13: Differences in Reward Designs Across Adaptive Parallel Reasoning Works

Evaluation and Open Questions

When all is said and done, how well do these adaptive parallel methods actually perform? Well…this is a hard question, as they differ in model choice and metrics. The model selection depends on the training method, SFT problem difficulty, and sequence length. When running SFT on difficult datasets like s1k, which contains graduate-level math and science problems, researchers chose a large base model (Qwen2.5 32B for Multiverse (Yang et al., 2025)) to capture the complex reasoning structure behind the solution trajectories. When running RL, researchers chose a small, non-CoT, instruct model (4B, 8B) due to compute cost constraints.

Figure 14: Difference in Model Choice Across Adaptive Parallel Reasoning Papers

Each paper also offers a slightly different interpretation about how adaptive parallel reasoning contributes to the research field. They optimize for different theoretical objectives, so they use slightly different sets of metrics:

• Multiverse and ThreadWeaver (Yang et al., 2025; Lian et al., 2025) aim to deliver sequential-AR-model-level accuracy at faster speeds. Multiverse shows that APR models can achieve higher accuracy under the same fixed context window, while ThreadWeaver shows that the APR model achieves shorter end-to-end token latency (critical path length) while getting comparable accuracy.
• NPR (Wu et al., 2025) treats sequential fallback as a failure mode and optimizes for 100% Genuine Parallelism Rate, measured as the ratio of parallel tokens to total tokens.
• Parallel-R1 (Zheng et al., 2025) does not focus on end-to-end latency and instead optimizes for exploration diversity, presenting APR as a form of mid-training exploration scaffold that provides a performance boost after RL.

Open Questions

While Adaptive Parallel Reasoning represents a promising step toward more efficient inference-time scaling, significant open questions remain.

As noted above, Parallel-R1 (Zheng et al., 2025) presents APR as a form of mid-training exploration scaffold rather than a primarily inference-time technique. This invites a more fundamental question: Does parallelization at inference-time consistently improve accuracy, or is it primarily valuable as a training-time exploration scaffold? Parallel-R1 suggests that the diversity induced by parallel structure during RL may matter more than the parallelization itself at test time.

A related concern is stability. There’s also a persistent tendency for models to collapse back to sequential reasoning when parallelization rewards are relaxed. Parallel-R1 authors showed that removing parallelization reward after 200 steps results in the model reverting to sequential behavior. Is this a training stability issue, a reward signal design issue, or evidence that parallel structure genuinely conflicts with how autoregressive pretraining shapes the model’s prior?

Beyond whether APR works, deployment introduces its own questions. Can we design training methods that account for available compute budget at inference time, so parallelization decisions are hardware-aware rather than purely problem-driven?

Finally, the parallel structures considered above are essentially flat. What if we allow parallelization depth > 1? Recursive language models (RLMs; Zhang, Kraska and Khattab, 2026) effectively manage long context and show promising inference-time scaling capabilities. How well do RLMs perform when trained with end-to-end RL that incentivizes adaptive parallelization?

Acknowledgements

In the last two years, the AI ecosystem has exploded with tools, wrappers, platforms, and “revolutionary” services—many of which seem to appear overnight, rocket across social feeds, and vanish just as quickly. Behind this churn is a quiet but powerful engine: affiliate and referral‑driven growth.

This isn’t new. Tech has always had its hype cycles, and every hype cycle has its preferred distribution mechanism. But AI’s current wave is uniquely shaped by performance-based marketing—an ecosystem where the loudest voices often have the strongest financial incentives and where the quality of the product is not always aligned with the enthusiasm of the promotion.

So the question for buyers, builders, and analysts is simple: Does the way an AI product goes to market tell us something about its long‑term viability?

Increasingly, the answer is yes.

The Data: Most Affiliate Programs Don’t Work—And That’s the Point

A 2026 analysis of 2,847 SaaS affiliate programs revealed a striking pattern:

• Only 15.6% of programs survive long-term.
• Only 7.6% of affiliates ever generate a single referral.
• Only 1.28% generate a sale.
• Average referral‑to‑sale conversion: 0.8%.

In other words: affiliate programs rarely scale, and when they do, it’s because the product already has real traction. The affiliate channel amplifies existing demand; it doesn’t create it.

A separate 2022 study on referral campaigns found that effectiveness depends heavily on trust networks—referrals from credible peers drive adoption and retention, while referrals from weak or opportunistic networks do not.

This is the first major signal: Healthy products use affiliates to accelerate growth. Weak products use affiliates to replace it.

The Good, the Bad, and the Ugly: What AI Buyers Should Watch For

The Good: Product‑First, Incentive‑Second

These companies treat affiliate programs as a multiplier, not a lifeline.

Examples:

• Notion, Webflow, Vercel, Perplexity — strong organic adoption, developer‑centric communities, and moderate, sustainable commissions.
• Midjourney (indirectly) — no formal affiliate program, yet massive word‑of‑mouth because the product itself is the marketing.
• OpenAI’s API ecosystem — developers promote tools because they work, not because they pay.

Characteristics:

• Transparent roadmaps
• Strong documentation
• Real user communities
• Measurable retention
• Commissions in the 15–30% range

These companies don’t need hype. Their users do the talking.

The Bad: The AI Wrapper Gold Rush

These are tools built on top of someone else’s model—thin, fast, and often disposable.

Examples (category, not naming specific companies):

• “One‑click” content generators with 40%+ commissions
• AI résumé builders that promise “instant job offers”
• AI chatbot clones with lifetime deals and aggressive influencer pushes

Characteristics:

• High churn
• Minimal differentiation
• Heavy reliance on paid influencers
• Commission rates that exceed product value

These products often spike quickly, then collapse once users realize the underlying value is shallow.

The Ugly: Incentive‑Driven Hype Machines

This is where AI meets the worst instincts of the crypto ICO era.

Examples (patterns, not brands):

• AI trading bots promising guaranteed returns
• “AI business in a box” schemes
• MLM‑adjacent AI platforms where the real product is the referral payout

Characteristics:

• Commissions higher than subscription revenue
• Complex tiered payouts
• Pressure to recruit, not use
• No meaningful product roadmap
• No evidence of real customers

These companies don’t sell AI. They sell the idea of selling AI.

The Pattern Across Tech History: When Incentives Outrun Innovation

AI is not the first technology to be distorted by its own distribution model.

• Mobile apps chased incentivized installs, leading to click‑farms and inflated metrics.
• Crypto ICOs used referral bonuses to mask the absence of real utility.
• Ad‑tech networks created entire ecosystems of low‑quality content optimized for clicks, not value.

In each case, the marketing engine outpaced the product engine—and the crash followed.

The lesson is consistent: When the incentive structure becomes the product, the product itself suffers.

So Is There an Inverse Relationship Between GTM Strategy and Product Quality?

Not inherently. But there are strong correlations.

Healthy AI companies:

• Build value first
• Layer affiliates on top
• Use moderate commissions
• Target credible creators
• Show real retention and usage

Fragile AI companies:

• Launch affiliate programs before product-market fit
• Offer extreme commissions
• Rely on hype-driven influencers
• Obscure churn and usage metrics
• Focus on acquisition, not retention

The difference is not the presence of an affiliate program—it’s the role it plays.

Food for Thought: How to Evaluate AI Products Before You Buy

Here are the questions every AIQI reader should ask before subscribing to any AI tool:

1. Who benefits most from this recommendation—the user or the promoter?

If the loudest voices are affiliates, not practitioners, proceed with caution.

2. Does the product have real depth, or is it a thin wrapper?

Look for documentation, API access, and evidence of technical investment.

3. Is the commission structure reasonable?

If the payout is higher than the subscription fee, something is off.

4. Are there real users, or just influencers?

Communities don’t lie. Influencers often do.

5. Does the company publish a roadmap or meaningful updates?

Sustainable AI companies show their work.

6. What happens if the hype fades?

If the product’s value disappears when the marketing stops, it was never real.

The Recommendation: Trust the Incentives, Not the Ads

AI is entering a maturity phase. The tools that survive will be the ones that deliver real value—not the ones that pay the highest commissions.

For buyers: Follow the product, not the promotion.

For creators and analysts: Interrogate the incentive structures behind every “must‑try” AI tool.

For the industry: Recognize that sustainable AI is built on retention, not referrals.

And for the companies building the future: Your go‑to‑market strategy is a signal. Make sure it signals quality.

Written/published by AI Quantum Intelligence with the help of AI model research.

The Myth of Infinite Scaling

For nearly a decade, the AI industry has operated under a simple doctrine:

More parameters → more intelligence.

More compute → better performance.

More data → more capability.

This belief was reinforced by the early scaling laws that showed smooth, predictable improvements as models grew. But those curves were never a promise — they were an observation. And like all observations, they eventually hit their limits.

Today, the cracks are visible everywhere:

• Frontier models require exponentially more compute for incrementally smaller gains.
• Performance improvements increasingly come from training tricks, data curation, and architectural refinements, not raw size.
• Smaller, specialized models are outperforming giants on domain‑specific tasks.
• The cost of training and inference is rising faster than the value created.

The scaling wall isn’t theoretical anymore. It’s here.

Diminishing Returns: The New Normal

The industry rarely admits it publicly, but insiders know:

The performance curve is flattening.

1. Marginal gains are shrinking

A model that is 2× larger no longer produces 2× better results. In many benchmarks, the improvement is barely noticeable to end users. The “wow factor” of scale has faded.

2. Data quality now matters more than data quantity

We’ve reached the point where adding more low‑quality data can hurt performance. Well‑curated, cleaned, and domain‑specific datasets often outperform huge collections of noisy, unfocused data.

3. Training instability increases with size

Large models are more sensitive to:

• initialization
• optimizer choice
• learning rate schedules
• distribution drift
• subtle data contamination

The bigger the model, the more brittle the training pipeline becomes.

4. Inference costs are becoming a strategic liability

Running a frontier model at scale is now a business model constraint, not a technical detail. Companies are discovering that “bigger” often means “unusable” for real‑world deployment.

Why Smaller Models Are Catching Up

The most interesting trend in 2025–2026 isn’t the rise of trillion-parameter models—it's the rise of efficient intelligence.

1. Architectural innovation beats brute force

Techniques like:

• Mixture‑of‑Experts (MoE)
• structured sparsity
• retrieval‑augmented generation (RAG)
• low‑rank adaptation (LoRA)
• distillation and quantization

allow smaller models to punch far above their weight.

2. Domain‑specific models outperform generalists

A 7B‑parameter model trained on:

• legal corpora
• medical literature
• financial filings
• scientific papers

can outperform a 500B‑parameter generalist on tasks that matter to professionals.

3. On‑device AI is becoming a competitive force

Edge‑optimized models are:

• faster
• cheaper
• more private
• more reliable

And they don’t require a hyperscaler‑sized GPU cluster to run.

4. Retrieval is the real intelligence multiplier

Models augmented with high‑quality retrieval systems often outperform larger models that rely solely on parametric memory.

Knowledge beats memorization.

The Hidden Costs of Going Big

Scaling isn’t just expensive — it introduces new forms of fragility.

1. Training failures become catastrophic

A single misconfigured hyperparameter can waste tens of millions of dollars in compute. A corrupted dataset can poison months of training. A subtle bug can derail an entire frontier‑model roadmap.

2. Larger models amplify biases and errors

More parameters mean more capacity to internalize:

• misinformation
• harmful correlations
• synthetic data artifacts
• recursive model‑generated noise

Without rigorous data governance, scale becomes a liability.

3. Environmental and energy costs are no longer ignorable

Training a frontier model now consumes:

• gigawatt‑hours of electricity
• millions of liters of cooling water
• supply chains of rare‑earth hardware

The sustainability narrative is shifting from “nice to have” to “strategic necessity.”

4. Latency becomes a user‑experience bottleneck

A model that takes 1.5 seconds to respond feels magical. A model that takes 4 seconds feels broken. Bigger models push latency in the wrong direction.

The Future: Smarter, Not Larger

The next era of AI will not be defined by parameter count. It will be defined by efficiency, specialization, and intelligence architecture.

1. Hybrid systems will dominate

Expect architectures that combine:

• medium‑sized base models
• retrieval engines
• tool‑calling agents
• domain‑specific adapters
• reasoning modules
• symbolic or structured components

The future is modular, not monolithic.

2. Data governance becomes the new competitive frontier

Companies that master:

• deduplication
• contamination detection
• synthetic‑data validation
• provenance tracking
• domain‑specific curation

will outperform those that simply scale compute.

3. Quantum‑accelerated optimization will reshape training economics

As quantum optimization matures, it will:

• reduce training instability
• accelerate convergence
• improve architecture search
• optimize massive parameter spaces

This won’t eliminate the scaling wall—but it will change where the wall sits.

4. The industry will rediscover the value of constraints

Constraints force creativity.

Constraints force efficiency.

Constraints force better design.

The next breakthroughs will come not from ignoring constraints, but from embracing them.

The Real Question Isn’t “How Big?” — It’s “How Smart?”

The scaling era gave us extraordinary progress, but it also created a dangerous illusion: that intelligence is simply a matter of size. Week 11 of AI Reality Check is a reminder that progress now depends on sophistication, not scale.

The companies that win the next phase of AI won’t be the ones with the biggest models. They’ll be the ones with the best‑engineered systems, the cleanest data, the most efficient architectures, and the deepest understanding of where scale helps—and where it hurts.

Conceived, written, and published by AI Quantum Intelligence with the help of AI models.

This third installment of The Intelligence Shift examines the quiet psychological revolution underway: the moment humans begin negotiating identity with systems that learn from us, mimic us, and increasingly anticipate us.

1. The New Mirror: When AI Learns You Faster Than You Learn Yourself

For most of human history, identity was shaped through slow feedback loops—family, culture, community, and work. AI collapses that timeline.

Large-scale models now infer personality traits from a few sentences, predict preferences from microbehaviors, and generate responses tailored to emotional tone. This creates a new kind of mirror:

• Reflective — showing us patterns we didn’t know we had
• Adaptive — shifting its behavior based on our shifts
• Persistent — remembering interactions long after humans would

The result is a subtle but profound shift: We begin to see ourselves through the lens of systems that are trained on us.

And mirrors, historically, have always changed societies — from the Renaissance to the selfie era. AI is simply the next, more intimate evolution.

2. The Authenticity Paradox: If AI Can Sound Like Us, What Does “Real” Even Mean?

Authenticity used to be defined by scarcity: Your voice was yours. Your writing style was yours. Your emotional cadence was yours.

Now, AI can replicate all three.

This creates an authenticity paradox:

• If a machine can imitate your tone, is tone still part of identity?
• If a model can generate your “style,” what does authorship mean?
• If AI can express empathy convincingly, how do we define genuine emotion?

We are entering a world where authenticity is no longer about uniqueness of expression but uniqueness of intention. Identity becomes less about how we communicate and more about why.

3. Relationships in the Age of Machine Companions

AI is not replacing human relationships — but it is reshaping the emotional landscape around them.

Three major shifts are emerging:

1. Emotional outsourcing People increasingly use AI to process feelings, rehearse conversations, or clarify thoughts before speaking to others. AI becomes a pre-relationship layer—a cognitive and emotional buffer.
2. Hyper-personalized interaction AI companions adapt to each user’s emotional patterns, creating a sense of being deeply understood. This can be supportive — or dangerously seductive.
3. The rise of “synthetic intimacy", not romantic but relational: AI provides consistency, patience, and nonjudgmental presence—qualities humans struggle to maintain.

The challenge is not that AI forms relationships. The challenge is that humans may recalibrate expectations of each other based on how AI behaves.

4. Identity Drift: When AI Shapes Who We Become

AI doesn’t just reflect identity — it nudges it.

Every suggestion, completion, recommendation, or generated idea subtly influences:

• How we speak
• What we value
• What we believe
• How we see ourselves

This is identity drift — the gradual co‑authoring of the self with algorithms.

Unlike social media, which shapes identity through exposure, AI shapes identity through interaction. It is participatory, not passive.

And because AI adapts to us, we adapt back.

The boundary between “my preference” and “the model’s suggestion” becomes increasingly porous.

5. The New Authenticity: Identity as a Deliberate Act

In a world where machines can mirror us, authenticity becomes intentional.

The future of identity will depend on three human skills:

1. Self-awareness - Understanding what is yours versus what is algorithmically reinforced.
2. Narrative agency - Actively choosing the story you tell about yourself—not the one AI predicts.
3. Ethical presence - Recognizing that identity is relational, and AI-mediated interactions can still carry moral weight.

Authenticity becomes less about resisting AI and more about coexisting with it consciously.

6. The Path Forward: Designing AI That Strengthens, Not Dilutes, Human Identity

If AI is becoming a mirror, we must decide what kind of mirror we want.

Three design principles matter most:

• Transparency — Users should know when AI is shaping their choices.
• Friction — Not every interaction should be optimized; some should require reflection.
• Human primacy — AI should amplify human agency, not replace it.

The goal is not to build systems that mimic us perfectly. The goal is to build systems that help us understand ourselves more deeply.

Conclusion: The Intelligence Shift Is Also an Identity Shift

AI is not just transforming industries — it is transforming introspection.

We are entering an era where identity is:

• Co-created
• Algorithmically influenced
• Fluid
• Negotiated
• Reflective

The question is no longer “Will AI change who we are?” It already has.

The real question is: How do we remain authors of ourselves in a world where machines can write in our voice?

Key References:

1. Performing Intimacy: Curating the Self‑Presentation in Human–AI Relationships (2025)

Source: Emerging Media (SAGE Journals)

Active Link: https://journals.sagepub.com/doi/10.1177/27523543251334157

2. Toward an Ethic of Synthetic Relationality: Identity, Intimacy, and Risk in AI‑Mediated Roleplay Environments (2025)

Source: AAAI/ACM Conference on AI, Ethics, and Society

Active Link: https://doi.org/10.1609/aies.v8i1.36560

3. Emotional AI and the Rise of Pseudo‑Intimacy: Are We Trading Authenticity for Algorithmic Affection? (2025)

Source: Frontiers in Psychology

Active Link: https://doi.org/10.3389/fpsyg.2025.1679324

Written and published by AI Quantum Intelligence with the help of AI models.

A scientific breakthrough reaches its full potential only when it empowers others to replicate and expand upon findings, pushing the boundaries of science even further. At Google Research, we recognize that open-source software and open-access datasets are drivers of modern science. We believe that creating these resources responsibly and maintaining them through partnerships with the global scientific community embodies the spirit of collaboration. In this way, we uphold the principles of open science, ensuring that innovation is not a siloed event but a catalyst for worldwide progress.

A 200-line open-source repo smoked GPT-4o, Perplexity, and every vector RAG benchmark. No embeddings. No chunking. Here’s what it actually…

Continue reading on Medium »

For millions of Muslims, Makkah is not simply a destination. It is a place of awe, peace, and spiritual gravity that is difficult to fully…

Continue reading on Medium »

If you’ve been following the reasoning model wave, you’ve seen GRPO mentioned in the same breath as DeepSeek-R1 and Qwen3. Both of those…

Continue reading on Medium »

Part 4 — https://medium.com/@mittalutkarsh/the-training-pipeline-with-one-row-flowing-through-every-stage-part4-2797aa2e6c2d

Continue reading on Medium »

And yet another question I can’t stop turning over

Continue reading on Medium »

Building with Data | Part 6: The Series Finale

Continue reading on Medium »

It’s a compelling narrative. It’s also incomplete.

The truth is far more uncomfortable: AI has accelerated attackers at least as much as it has empowered defenders—and in some domains, more. The result is not a decisive advantage but a rapidly escalating arms race where neither side holds stable ground.

The question organizations must confront is no longer “Should we adopt AI for cybersecurity?” but rather: “Are we building defenses that actually outpace the threats—or simply buying tools that make us feel safer?”

The Illusion of Progress: Why AI Isn’t Winning the Cyber War

There’s no denying that AI has improved defensive capabilities. Modern SOCs can detect anomalies in seconds, correlate signals across millions of endpoints, and automate containment actions that once required human intervention.

But attackers have gained something even more dangerous: creativity at scale.

AI now enables:

• Malware that mutates on every execution

• Automated reconnaissance that maps entire cloud estates in minutes

• Hyper-personalized phishing indistinguishable from human communication

• Exploit development accelerated by code‑generation models

The barrier to entry for sophisticated attacks has collapsed. What once required nation‑state resources can now be executed by small, well‑funded groups—or even individuals with the right tooling.

Defenders face a brutal asymmetry: They must be perfect. Attackers only need to be lucky.

The Myth of the “Upper Hand”

Vendors love to claim that AI gives defenders the upper hand. But the reality is more nuanced.

Where AI helps defenders

• Speed

• Scale

• Exhaustive monitoring

• Automated response

Where AI helps attackers

• Novelty

• Variability

• Unpredictability

• Rapid iteration

Defensive AI is inherently conservative—it must avoid false positives, maintain uptime, and operate within governance constraints. Offensive AI has no such limitations. It can be reckless, experimental, and iterative.

This is why the “upper hand” narrative collapses under scrutiny. AI is not a shield. It is an accelerant—for both sides.

The Only Sustainable Advantage: Human Creativity Augmented by Machine Speed

If AI alone cannot secure the future, what can?

A hybrid model where human adversarial thinking is fused with AI’s computational power.

Humans bring:

• Creativity

• Contextual judgment

• Lateral thinking

• Understanding of business impact

AI brings:

• Pattern recognition

• Real-time correlation

• Instantaneous response

• Scalability

This is the model used by the most advanced cyber defense organizations on the planet. Not AI replacing humans. Not humans resisting automation. But humans who know how to weaponize AI defensively.

The future of cybersecurity is not artificial intelligence. It is augmented intelligence.

Should Organizations Keep Spending on AI Cyber Tools? Yes—but Not Blindly

If attackers’ AI is equal to or better than defenders’, does it still make sense to invest in AI-enabled cybersecurity?

Absolutely—but only with strategic clarity.

AI is essential because:

• Attack volume is too high for humans alone

• Identity attacks move too fast for manual response

• Regulators increasingly expect AI-assisted monitoring

• AI is now baseline infrastructure, not a luxury

But AI investment only pays off when paired with:

• Skilled analysts

• Strong identity governance

• Zero-trust architecture

• Continuous red teaming

• Mature incident response

Buying AI tools without the human layer is like buying a fighter jet without a pilot.

The Real Spending Problem: Misallocation, Not Underinvestment

Most organizations overspend on:

• “Next-gen” dashboards

• Automated tools with vague promises

• Vendor hype cycles

And underspend on:

• Threat hunters

• Architecture hardening

• Identity lifecycle management

• Red team simulations

• Incident readiness

AI amplifies whatever foundation exists. If the foundation is weak, AI amplifies the weakness.

The Hard Truth: AI Will Not Save Us—But Humans Who Use AI Well Might

The cybersecurity industry must abandon the fantasy that AI alone will secure the digital world. Attackers innovate too quickly. Models are too predictable. And the threat landscape is too dynamic for static automation.

The real advantage belongs to organizations that understand a simple principle:

AI gives defenders speed. Humans give defenders unpredictability. Together, they create resilience.

The future of cybersecurity will not be won by machines or by humans—but by the teams that know how to combine the two into something neither could achieve alone.

Written/developed by AI Quantum Intelligence with the help of AI models.

AI today looks unstoppable—models scaling to trillions of parameters, multimodal agents performing tasks once reserved for experts, and infrastructure stretching across continents. Yet beneath the surface lies a quiet truth: modern AI systems are far more fragile than their marketing suggests. They are brittle in logic, vulnerable in data, and dependent on human scaffolding that few acknowledge.

1. The brittleness behind the brilliance

Every major AI breakthrough hides a delicate balance of assumptions: clean data, stable architectures, and predictable environments. When any of these shift—distribution drift, adversarial noise, or unseen edge cases—performance collapses. Recent studies show that even state‑of‑the‑art LLMs lose up to 40 % accuracy when prompts deviate slightly from training patterns. This brittleness isn’t a bug; it’s a structural feature of systems optimized for pattern recognition, not resilience.

2. Synthetic stability: the paradox of scale

To sustain growth, developers increasingly rely on synthetic data loops—models training on outputs of other models. While this accelerates iteration, it also amplifies errors. Each generation inherits the biases and blind spots of its predecessors, creating a feedback spiral of self‑reinforcing noise. The result: systems that appear more capable but are internally hollow, their “knowledge” increasingly detached from reality.

3. Infrastructure fragility: the unseen dependency chain

Modern AI depends on a fragile stack of cloud compute, GPU supply chains, and proprietary APIs. A single outage or policy change can ripple through entire ecosystems. The illusion of autonomy masks a deep interdependence—between vendors, data brokers, and model providers. In effect, AI’s strength is borrowed, not owned.

4. Security and alignment: fragility in disguise

Security researchers have demonstrated that small prompt injections or jailbreaks can override safety filters with trivial effort. Alignment mechanisms, often touted as robust, are statistical heuristics vulnerable to manipulation. The same flexibility that makes LLMs creative also makes them exploitable. Fragility here isn’t just technical—it's ethical and systemic.

5. The path forward: resilience over scale

True progress will come not from bigger models but from more resilient architectures—systems that can reason under uncertainty, adapt to shifting context, and recover gracefully from failure. That means investing in:

• Context‑aware reasoning frameworks
• Transparent data provenance
• Human‑in‑the‑loop validation
• Adaptive safety layers

Until then, AI remains a glass cathedral—magnificent, but one shock away from collapse.

Conceived, written, and published by AI Quantum Intelligence with the help of AI models.

Global supply chains lose visibility whenever an asset moves beyond cell-tower range. Containers go silent mid-ocean and farm equipment drops off dashboards in remote fields. For years, businesses accepted these issues as inevitable. Affordable hybrid Internet of Things (IoT) modules now change this mindset by delivering continuous connectivity across cellular, satellite and Wi-Fi on a single chip. That unbroken data stream is what makes eliminating supply chain blind spots a realistic goal.

Understanding Traditional Connectivity Gaps

Business leaders have long been stuck choosing between two imperfect options for tracking assets across global supply chains. Cellular and LPWAN technologies like LoRaWAN and NB-IoT offer a cost-effective way to monitor assets in urban and suburban areas. They perform well inside warehouses, along major highways and within port facilities.

However, terrestrial mobile networks cover only about 15% of the Earth’s landmass. Once a container ships across the Pacific or heavy machinery operates in a remote mining region, the signal vanishes and the data stops flowing.

Legacy satellite connectivity fills that geographic gap with true global coverage. The problem has always been cost. Traditional satellite modules require separate hardware, dedicated subscription contracts and minimum usage commitments. This can be a challenge for companies that manage thousands of sensors or containers.

The corporate consequences are also well documented. A survey found that 37% of brands cannot track in-transit cargo, and 60% discover shipment damage only after delivery. These kinds of issues easily turn into lost assets, spoiled goods, inflated insurance premiums and dwindling customer trust.

How Hybrid IoT Modules Bridge Connectivity Gaps

The innovation expected in the next few years is not the idea of combining multiple networks. What is new is the ability to integrate satellite, cellular and Wi-Fi radios into a single, power-efficient module at a price point that works for high-volume deployments.

In the first quarter of 2026, two major product launches confirmed this shift. SKYWAVE released the ST 4000, which unifies satellite and cellular connectivity in a single device. Iridium came shortly after with the 9604, which packs satellite SBD, LTE-M and GNSS into a 16-by-26-millimeter module. Several other brands have launched their own versions since.

The defining feature across these hybrid IoT modules is the seamless failover, with the device handling transactions without human intervention or data loss. An asset activates on Wi-Fi inside a depot, switches to cellular on the road and then connects via satellite in remote zones.

According to a Fall 2025 report by IoT Analytics, the number of connected IoT devices reached 18.5 billion in 2024. They estimate that this could increase to 39 billion by 2030. As device counts increase, the demand for unbroken connectivity across every terrain and transit corridor will only intensify.

The Business Value of Uninterrupted Data

When hybrid IoT platforms deliver data consistently, supply chain teams can stop reacting to problems and start preventing them. For example, instead of finding out a pharmaceutical shipment was damaged only upon arrival, teams get a real-time alert during transit so they can find solutions or reroute before anything gets lost.

When companies can track assets across an entire journey, they spot containers and trailers sitting idle and put them back to work faster. That means less wasted equipment and lower costs. Continuous location and status updates also help fleet managers make smarter scheduling and routing decisions.

Temperature-sensitive goods like vaccines, biologics and fresh produce demand constant environmental monitoring. The challenge has always been maintaining that data stream across every leg of the journey, including ocean crossings and rural last-mile routes. High-value cargo moving through dead zones requires that same unbroken data flow. Hybrid IoT platforms ensure they keep moving regardless of geography.

Hybrid IoT Applications in Practice

In logistics and cold chain management, a single hybrid IoT module can track a shipping container from a factory in Shanghai on Wi-Fi, through the port on cellular, across the Pacific on satellite and into a rural distribution center in Montana via cellular again. At no point does the data stream break.

In agriculture, farms often operate far beyond reliable cell coverage. Hybrid connectivity enables the continuous monitoring of soil moisture sensors, livestock GPS trackers and equipment telematics across thousands of acres without the expensive private network infrastructure.

The hardware alone isn’t enough. Businesses also need software that pulls data from every connectivity network into one clear view. When evaluating hybrid IoT platforms, decision-makers should look at the full picture — how data is collected, analyzed and connected to the enterprise tools their teams already use.

The Advantage of Always-On Supply Chain Visibility

The real value of hybrid IoT modules goes well beyond location data. These devices deliver the operational certainty and data continuity that global businesses need to reduce waste, protect high-value shipments and make faster decisions. Always-on connectivity turns scattered supply chain data into a reliable foundation for smarter operations.

The post Can the New Wave of Hybrid IoT Modules Finally Eliminate Supply Chain Blind Spots? appeared first on IoT Business News.

As enterprises push more operations to connected, software-driven workflows, internet downtime is increasingly an operational risk—not just an IT inconvenience. T-Mobile’s new SuperBroadband bundle pairs its 5G Business Internet with Starlink connectivity under one managed service, aiming to make resilience and reach easier to procure and operate.

For multi-site businesses, connectivity has become an exercise in trade-offs: fiber where it exists, fixed wireless where it doesn’t, and a patchwork of backup links that too often behave like separate projects with separate vendors. That fragmentation is especially painful in sectors that can’t simply “wait out” an outage—think point-of-sale, patient workflows, safety systems, or remote field operations.

T-Mobile is now trying to productize a different answer. The operator has launched SuperBroadband, a business internet service that combines its 5G Business Internet with Starlink broadband, delivered as a fully managed offering with one contract and one bill. The promise is straightforward: two independent access paths—cellular and LEO satellite—designed to keep sites online through disruptions, while reducing the operational overhead of managing primary and backup providers.

The company says SuperBroadband is already being used by organizations in hospitality, retail, healthcare, and oil and gas. Aramark Destinations’ CIO, Dimple Jethani, framed the appeal around remote and complex environments where connectivity has been inconsistent and difficult to scale.

Redundancy is the product, not an add-on

Dual connectivity is not new; what’s distinct here is that T-Mobile is selling redundancy as a standardized, nationwide service rather than leaving customers to assemble it from separate ISPs, separate hardware, and separate support models. In the press materials, T-Mobile positions SuperBroadband as “built-in redundancy” via independent 5G and Starlink pathways, with intelligent orchestration between the two connections in real time.

Under the hood, T-Mobile describes an architecture that includes outdoor 5G equipment to improve signal strength, advanced routers to bring the connections together, and centralized control using Ericsson Enterprise Wireless Solutions’ NetCloud Manager for the latest Ericsson Cradlepoint routers and outdoor adapters. T-Mobile also says it plans to expand its ecosystem over time with partners such as Inseego for enterprise wireless broadband and edge connectivity options.

One detail IoT and enterprise networking teams will notice: the offer is not just about access technologies, but about operational tooling. T-Mobile is positioning its T-Platform as the pane of glass for deployments, including visibility into hardware, performance, usage, health, and events such as backup readiness and failovers.

Coverage claims—and what they mean in practice

T-Mobile says it has expanded its unlimited 5G Business Internet to millions of additional business locations and, with Starlink integrated, can reach “effectively every business location in America,” claiming SuperBroadband is the first nationwide broadband solution to reach every ZIP code in the U.S.

For IoT-heavy enterprises, the practical implication is less about marketing superlatives and more about procurement simplification. If a single provider can cover urban sites, suburban locations, and hard-to-serve remote facilities—while also providing a standardized backup path—network teams can reduce the number of exceptions in their connectivity design. That, in turn, can shorten deployment cycles for new sites and make rollouts of connected systems (POS refreshes, remote monitoring, edge applications) more repeatable across regions.

A managed-service play with defined service levels

SuperBroadband is being sold as a fully managed service with defined service levels and end-to-end support from T-Mobile for Business. The company also advertises a financially backed 99.99% uptime guarantee, with eligibility and exclusions outlined in its service terms, and notes that installation and field services are supported by Acuative to enable nationwide deployment.

Here’s the operational insight that matters: by tying an uptime commitment to a bundled, dual-path design, T-Mobile is implicitly shifting the “resilience engineering” burden away from the customer and toward the provider’s managed stack—hardware, orchestration, monitoring, and support. That may appeal to organizations that lack the staff to design and continuously test multi-carrier failover themselves, but it also means enterprises will scrutinize how failover events are detected, how switching is handled, and what visibility they retain in the portal during incidents.

Why this is different from typical connectivity bundling

Many connectivity announcements boil down to “we support technology X” or “we partner with provider Y”. T-Mobile’s SuperBroadband is more specific: it is a packaged, nationwide offer that fuses terrestrial 5G and LEO satellite into one managed experience, with centralized orchestration and monitoring baked in. In other words, it’s not just adding satellite as an optional uplink; it’s selling a standardized operational model for dual connectivity.

For OEMs and solution providers building systems that assume continuous connectivity—digital signage, connected kiosks, remote security, industrial telemetry, edge compute stacks—the availability of a single, supported connectivity SKU could simplify go-to-market and support. For system integrators, the combination of NetCloud-managed Cradlepoint hardware and T-Platform visibility may reduce the tooling sprawl that often comes with multi-ISP designs.

SuperBroadband is available now, according to T-Mobile, for use cases ranging from single-site businesses to complex, multi-location organizations, with Starlink connectivity integrated across options.

The post T-Mobile packages 5G and Starlink into a single managed broadband offer for business continuity appeared first on IoT Business News.

Altair Semiconductor has completed its spinoff from Sony Semiconductor Solutions, securing $50 million in funding to accelerate its focus on cellular IoT and a 5G eRedCap roadmap.

As IoT deployments scale across industries, a recurring challenge continues to shape the market: how to balance power efficiency, device longevity, and evolving connectivity standards. While LTE-M and NB-IoT have enabled massive IoT adoption, the transition toward 5G-native solutions remains complex, particularly for cost-sensitive and battery-powered devices.

It is within this context that Altair Semiconductor has completed its transition into an independent company, following a strategic spinoff from Sony Semiconductor Solutions. The move is backed by $50 million in initial funding led by Pitango Group, with Sony retaining a stake in the company .

The separation marks a structural shift for Altair, which has built its reputation on low-power cellular IoT chipsets widely used in applications such as smart metering, asset tracking, and wearables. Operating independently is expected to give the company greater flexibility to navigate a rapidly evolving connectivity landscape, particularly as the industry begins to align around new 5G IoT categories.

Positioning for the 5G IoT Transition

Altair’s roadmap centers on 5G eRedCap (enhanced Reduced Capability), a standard designed to bridge the gap between high-performance 5G and the constrained requirements of massive IoT devices. The company’s upcoming ALT1550 modem, currently in advanced silicon testing, reflects this strategic direction .

This positioning is notable because eRedCap is emerging as a key enabler for mid-tier IoT use cases that require more bandwidth and lower latency than LTE-M, but without the complexity and cost of full 5G. By aligning early with this segment, Altair is effectively targeting a future market layer that remains underdeveloped but strategically important.

At the same time, the company continues to support its established LTE Cat-M and NB-IoT platforms, which integrate connectivity, processing, and security features into highly compact chipsets. This dual approach—maintaining a strong 4G base while preparing for 5G—suggests a phased transition strategy rather than a disruptive shift.

From Connectivity Provider to Physical AI Enabler

A central theme in Altair’s positioning is the concept of “Physical AI,” referring to the growing need to connect machines, sensors, and devices that generate and act on real-world data. While the term itself is gaining traction across the industry, its practical implication is straightforward: more endpoints require persistent, low-power connectivity to support distributed intelligence.

Altair’s chipset portfolio is already embedded in large-scale deployments, particularly in cellular smart metering, where long device lifecycles and energy efficiency are critical. Extending this footprint into AI-enabled edge devices—such as wearables or asset trackers—requires maintaining similar constraints while accommodating new data and processing requirements.

This creates a non-trivial engineering challenge. Devices expected to operate for up to 20 years must remain compatible with evolving network technologies, which is precisely where eRedCap could play a role. The company’s emphasis on long-term device lifespan highlights a key industry tension: innovation cycles in connectivity are accelerating, while IoT hardware lifecycles remain inherently long.

Why Independence Matters

The decision to spin off from Sony is not just a financial or organizational move—it reflects a broader trend in the semiconductor and IoT ecosystem. Specialized connectivity players increasingly require agility to respond to shifting standards, operator requirements, and vertical-specific demands.

Within a large corporate structure, aligning these priorities can be slower and more constrained. As an independent entity, Altair can focus more directly on IoT-specific innovation cycles, particularly in areas such as low-power modem design and integrated connectivity platforms.

At the same time, Sony’s continued shareholding suggests that the relationship remains strategically relevant, potentially preserving access to ecosystem synergies without the limitations of full integration.

Implications for the IoT Ecosystem

For OEMs and device manufacturers, Altair’s roadmap provides a clearer migration path from existing LTE-M deployments toward 5G-based solutions without requiring a complete redesign of devices or architectures. This continuity is critical in sectors such as utilities or logistics, where infrastructure refresh cycles are measured in decades.

Connectivity providers and system integrators may also benefit from a more focused chipset vendor actively shaping the eRedCap segment, which is still in its early stages of commercialization.

More broadly, the announcement underscores a structural evolution in the IoT value chain: as connectivity becomes increasingly tied to edge intelligence, chipset vendors are positioning themselves not just as enablers of connectivity, but as foundational components of distributed AI systems.

Altair’s move to independence—and its emphasis on 5G eRedCap—signals a calculated bet on where that convergence is heading next.

The post Altair Semiconductor Spins Off from Sony to Focus on 5G IoT and eRedCap Strategy appeared first on IoT Business News.

IoT Platforms have become a central layer in the architecture of connected systems, sitting between devices, networks, and enterprise applications. As organizations move from pilot projects to large-scale deployments, the need for structured, scalable, and secure ways to manage connected assets has intensified. Platforms are no longer optional infrastructure—they are foundational to how IoT systems are designed, operated, and monetized.

Yet the term “IoT Platforms” remains broad and sometimes ambiguous. It can refer to device management tools, cloud-based analytics environments, or full-stack solutions combining connectivity, data processing, and application enablement. Understanding what these platforms actually do—and how to evaluate them—has become a critical task for decision-makers navigating an increasingly fragmented ecosystem.

Key Takeaways

• IoT Platforms provide the software and infrastructure layer that connects devices, manages data, and enables applications.
• They typically include device management, data ingestion, analytics, and integration capabilities.
• The vendor landscape is fragmented, ranging from hyperscalers to specialized industrial platform providers.
• Selection criteria must balance scalability, interoperability, security, and total cost of ownership.
• Architectural choices—cloud, edge, or hybrid—have a direct impact on performance and deployment flexibility.

What is an IoT Platform?

IoT Platforms are integrated software environments that enable organizations to connect, manage, monitor, and analyze data from connected devices at scale. They act as an intermediary layer between hardware (sensors, gateways), connectivity networks, and enterprise applications, providing the tools required to build and operate IoT solutions.

In practice, IoT Platforms aggregate multiple functions that would otherwise require separate systems. These include device provisioning, data collection, real-time processing, analytics, visualization, and integration with business systems such as ERP or CRM platforms. By consolidating these capabilities, platforms reduce complexity and accelerate time to deployment.

Within the broader IoT ecosystem, IoT Platforms serve as the control plane. They orchestrate communication between devices and applications, enforce security policies, and provide the data pipelines that turn raw sensor data into actionable insights.

How IoT Platforms work

At a high level, IoT Platforms operate through a layered architecture designed to handle device connectivity, data processing, and application enablement.

The typical architecture includes:

• Device layer: Sensors, actuators, and embedded systems generate data and receive commands.
• Connectivity layer: Networks such as cellular (LTE-M, NB-IoT, 5G), LPWAN (LoRaWAN), Wi-Fi, or satellite transport data to the platform.
• Ingestion layer: Message brokers and APIs collect and normalize incoming data streams.
• Processing layer: Stream processing engines and rule engines filter, transform, and enrich data in real time.
• Storage layer: Time-series databases and data lakes store structured and unstructured data.
• Application layer: Dashboards, analytics tools, and APIs enable users to interact with data and build applications.

Communication between devices and IoT Platforms typically relies on lightweight messaging protocols such as MQTT or CoAP, designed for constrained environments. Platforms also support REST APIs and event-driven architectures to integrate with enterprise systems.

Increasingly, IoT Platforms extend beyond centralized cloud environments to include edge computing capabilities. In this model, part of the data processing occurs closer to the device, reducing latency and bandwidth usage while improving resilience.

Key technologies and standards

The functionality of IoT Platforms depends on a combination of communication protocols, data processing technologies, and interoperability standards.

Common technologies include:

• Messaging protocols: MQTT, AMQP, CoAP for efficient device-to-cloud communication.
• Connectivity standards: LTE-M, NB-IoT, 5G, LoRaWAN, Wi-Fi, Bluetooth Low Energy.
• Data formats: JSON, CBOR, Protocol Buffers for structured data exchange.
• Cloud infrastructure: Containerization (Docker), orchestration (Kubernetes), serverless computing.
• Edge frameworks: Edge runtimes for local data processing and device orchestration.
• Security standards: TLS/DTLS encryption, X.509 certificates, hardware-based secure elements.

Interoperability remains a critical issue. While IoT Platforms often support multiple protocols, the lack of universal standards across industries can lead to integration challenges, particularly in legacy environments.

Main IoT use cases

IoT Platforms are deployed across a wide range of industries, each with distinct requirements in terms of scale, latency, and data processing.

• Industrial IoT: Monitoring machinery, predictive maintenance, and optimizing production processes through real-time analytics.
• Logistics and supply chain: Tracking assets, monitoring environmental conditions, and improving route optimization.
• Smart cities: Managing urban infrastructure such as traffic systems, lighting, waste management, and public safety.
• Energy and utilities: Smart metering, grid monitoring, and demand-response systems.
• Healthcare: Remote patient monitoring, connected medical devices, and asset tracking within hospitals.
• Asset tracking: Monitoring location, status, and utilization of high-value equipment across industries.

In each of these use cases, IoT Platforms provide the common foundation for data collection, analysis, and operational decision-making.

Benefits and limitations

IoT Platforms offer several advantages that make them central to modern connected systems:

• Scalability: Ability to manage thousands to millions of devices from a single environment.
• Operational efficiency: Centralized management reduces the complexity of distributed systems.
• Faster deployment: Pre-integrated tools accelerate development and reduce time to market.
• Data-driven insights: Advanced analytics enable predictive and prescriptive decision-making.

However, these benefits come with trade-offs and limitations:

• Vendor lock-in: Proprietary architectures can make it difficult to migrate between platforms.
• Integration complexity: Connecting legacy systems and heterogeneous devices can require significant customization.
• Latency constraints: Cloud-based processing may not meet real-time requirements without edge capabilities.
• Cost management: Scaling data storage and processing can lead to unpredictable costs.
• Security risks: Expanding attack surfaces require robust security frameworks across devices and networks.

Understanding these trade-offs is essential when selecting and deploying IoT Platforms in production environments.

Market landscape and ecosystem

The IoT Platforms market is highly fragmented, reflecting the diversity of use cases and technical requirements.

The ecosystem includes several categories of players:

• Hyperscalers: Cloud providers offering scalable infrastructure and integrated IoT services.
• Industrial platform vendors: Solutions tailored for manufacturing, energy, and heavy industries.
• Connectivity providers: Operators integrating platform capabilities with network services.
• Specialized IoT vendors: Companies focusing on specific verticals or functions such as device management or analytics.
• System integrators: Organizations that combine multiple technologies into end-to-end solutions.

No single platform dominates across all segments. Instead, enterprises often adopt a multi-platform strategy, combining different solutions to address specific operational needs.

Partnerships between platform vendors, hardware manufacturers, and connectivity providers play a critical role in shaping the ecosystem, enabling interoperability and accelerating deployment.

Future outlook

The evolution of IoT Platforms is closely tied to broader trends in computing and connectivity.

Several developments are expected to influence the next generation of platforms:

• Edge-native architectures: Increased processing at the edge to reduce latency and bandwidth usage.
• AI integration: Embedding machine learning models directly into platforms for real-time analytics.
• Standardization efforts: Industry initiatives aimed at improving interoperability across devices and platforms.
• 5G and satellite connectivity: Expanding coverage and enabling new use cases in remote environments.
• Security by design: Stronger emphasis on end-to-end security across the entire IoT stack.

As deployments scale and become more complex, IoT Platforms will continue to evolve from infrastructure tools into strategic assets supporting digital transformation initiatives.

Frequently Asked Questions

What are IoT Platforms used for?

IoT Platforms are used to connect devices, manage data, and enable applications that rely on real-time information from connected systems.

What are the key features of IoT Platforms?

Core features include device management, data ingestion, real-time processing, analytics, security, and integration with enterprise systems.

How do IoT Platforms differ from cloud platforms?

IoT Platforms are specialized for handling device communication and sensor data, while general cloud platforms provide broader computing and storage capabilities.

What should enterprises consider when selecting IoT Platforms?

Key criteria include scalability, interoperability, security, cost, support for standards, and alignment with existing infrastructure.

Can IoT Platforms operate at the edge?

Yes, many IoT Platforms now include edge computing capabilities to process data locally and reduce latency.

Related IoT topics

• Edge Computing in IoT
• LPWAN Connectivity Technologies
• Device Management in IoT
• Industrial IoT
• IoT Security
• Digital Twins in IoT

The post IoT Platforms: Key Capabilities, Vendor Landscape and Selection Criteria appeared first on IoT Business News.

28 April marks one year since the Iberian Peninsula blackout, Europe’s largest grid failure in the last 20 years, which left nearly 60 million without power. A year on, it remains a stark reminder of how catastrophic outages can be when resilience mechanisms aren’t fully in place.

The IoT energy market is projected to reach $62.8 billion globally by 2030. Whilst last year’s blackout was not linked to the IoT, as IoT adoption grows in an increasingly digitalised sector, early decisions around network infrastructure, security and scalability are crucial for ensuring uptime.

Iain Davidson, head of product marketing, Wireless Logic, offers insight into what the industry has learned a year on from the outage:

“The role of IoT in the energy ecosystem and smart grid is growing rapidly. However, it is only an asset to the sector if it is truly resilient and secure. Last year’s Iberian Peninsula blackout demonstrated how quickly disruption can occur when complex infrastructure is placed under stress. As seen, if systems, equipment or infrastructure suffer downtime, severe disruption can occur on a vast scale. One year later, the focus must be on embedding proactive resilience measures from the outset”.

“Energy companies and suppliers should prioritise proactively and continuously monitor infrastructure, devices and applications. They should also implement predictive maintenance and real-time monitoring and threat detection supported by AI-driven analytics to identify and mitigate problems before they occur. This includes issues which begin as operational, environmental and cyber-security breaches.  In addition, networks and systems should be designed with redundancy to handle demand fluctuations and include automated and immediate failover to maintain continuity during failures”. 

“Crucially, the industry must implement operational and cyber resilience procedures and test them regularly. Resilience and security must be built in end-to-end across IoT devices, networks, software, processes and cloud to minimise downtime.”

“Ensuring that effective strategies are in place will support the sector to recover rapidly and effectively from incidents moving forward.”

The post One year on from Iberian blackout, what has the industry learned about resilience? Wireless Logic comments appeared first on IoT Business News.

For all the hype around “reasoning,” “intelligence,” and “agentic behavior,” today’s frontier AI systems still fail at the most human part of language: understanding context. Not memorizing facts. Not predicting the next token. Understanding what is actually meant.

Despite billions of parameters and massive training corpora, LLMs continue to misread nuance, miss implicit meaning, hallucinate details, and collapse under subtle contextual shifts. And the deeper you look, the clearer the pattern becomes: AI doesn’t understand context — it simulates it.

This week, we break down why.

1. Context is not just text—it's hierarchy, intent, and relevance

Humans process context across multiple layers simultaneously:

• Semantic context (what the words mean)
• Pragmatic context (what the speaker intends)
• Situational context (what is happening around the conversation)
• Relational context (who is speaking to whom, and why)

LLMs, by contrast, operate on statistical correlations within a token window. Even with advanced context-engineering techniques, models still struggle to integrate multiple layers of meaning the way humans do. Research shows that LLMs often fail to capture nuanced contextual features unless heavily finetuned or augmented with retrieval systems.

2. The context window is not the same as contextual understanding

Vendors love to advertise 200K‑token or even million‑token context windows. But a larger window doesn’t mean the model understands more — it simply means it can see more.

Studies show that LLMs:

• Overweight irrelevant information
• Underweight critical details
• Lose track of earlier context as sequences grow
• Default to parametric knowledge instead of grounding in the prompt

This is why models often hallucinate or contradict the very documents they were given. They rely too heavily on what they “already know” rather than what the context actually says.

3. Parametric knowledge vs. contextual grounding: the core conflict

LLMs draw from two sources of knowledge:

1. Parametric knowledge — what the model absorbed during pretraining
2. Contextual knowledge — what the prompt provides

The problem? These two sources often compete.

When the prompt contradicts the model’s internal priors, the model frequently defaults to its parametric memory—even when the prompt is explicit. Research confirms that LLMs struggle to balance these sources, leading to factual inconsistencies and contextually unfaithful outputs.

This is why models sometimes ignore instructions, invent details, or revert to generic answers.

4. Context engineering helps—but it’s still a patch, not a solution

A new discipline called Context Engineering has emerged to optimize how information is retrieved, structured, and delivered to LLMs. It includes:

• Retrieval‑augmented generation (RAG)
• Memory systems
• Tool‑integrated reasoning
• Multi‑agent orchestration

These techniques dramatically improve performance — but they also expose a deeper truth: we are compensating for the model’s inability to manage context on its own. Even with sophisticated pipelines, research shows a persistent asymmetry: models can consume complex context but struggle to generate equally coherent, contextually faithful long‑form output.

5. Why this matters for the future of AI

Context is the foundation of:

• Reasoning
• Safety
• Alignment
• Human‑AI collaboration
• Trustworthy autonomy

If models cannot reliably interpret context, they cannot reliably reason — no matter how impressive their benchmarks look.

The next frontier of AI progress will not be bigger models or longer context windows. It will be true contextual intelligence: systems that understand meaning, not just tokens.

Key References (with links)

1. Zhu et al. (2024) — Can Large Language Models Understand Context?

Source: arXiv (Findings of EACL 2024) Link: https://doi.org/10.48550/arXiv.2402.00858

Summary: Introduces a benchmark for evaluating LLM contextual understanding and shows that pretrained dense models struggle with nuanced contextual features.

2. Zhao et al. (2024) — Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding

Source: arXiv (Accepted to NAACL 2024) Link: https://doi.org/10.48550/arXiv.2405.02750

Summary: Demonstrates that LLMs often overweight parametric knowledge and proposes contrastive decoding to improve contextual grounding during generation.

3. Zhu et al. (2024) — Can Large Language Models Understand Context? (ACL Anthology Version)

Source: ACL Anthology (Findings of EACL 2024) Link: https://aclanthology.org/2024.findings-eacl.135/

Summary: Peer‑reviewed version confirming that LLMs fail to reliably interpret nuanced contextual cues, especially under in‑context learning and quantization.

Conceived, written and published by AI Quantum Intelligence with the help of AI models.

The post Proxy-Pointer RAG: Structure Meets Scale at 100% Accuracy with Smarter Retrieval appeared first on Towards Data Science.

The post The LLM Gamble appeared first on Towards Data Science.

The post From Risk to Asset: Designing a Practical Data Strategy That Actually Works appeared first on Towards Data Science.

The post What Does the p-value Even Mean? appeared first on Towards Data Science.

The post Context Payload Optimization for ICL-Based Tabular Foundation Models appeared first on Towards Data Science.

The post I Replaced GPT-4 with a Local SLM and My CI/CD Pipeline Stopped Failing appeared first on Towards Data Science.

The post Your RAG Gets Confidently Wrong as Memory Grows—I Built the Memory Layer That Stops It appeared first on Towards Data Science.

The post Git UNDO : How to Rewrite Git History with Confidence appeared first on Towards Data Science.

The post How to Call Rust from Python appeared first on Towards Data Science.

The post DIY AI & ML: Solving The Multi-Armed Bandit Problem with Thompson Sampling appeared first on Towards Data Science.

GRASP is a new gradient-based planner for learned dynamics (a “world model”) that makes long-horizon planning practical by (1) lifting the trajectory into virtual states so optimization is parallel across time, (2) adding stochasticity directly to the state iterates for exploration, and (3) reshaping gradients so actions get clean signals while we avoid brittle “state-input” gradients through high-dimensional vision models.

Large, learned world models are becoming increasingly capable. They can predict long sequences of future observations in high-dimensional visual spaces and generalize across tasks in ways that were difficult to imagine a few years ago. As these models scale, they start to look less like task-specific predictors and more like general-purpose simulators.

But having a powerful predictive model is not the same as being able to use it effectively for control/learning/planning. In practice, long-horizon planning with modern world models remains fragile: optimization becomes ill-conditioned, non-greedy structure creates bad local minima, and high-dimensional latent spaces introduce subtle failure modes.

In this blog post, I describe the problems that motivated this project and our approach to address them: why planning with modern world models can be surprisingly fragile, why long horizons are the real stress test, and what we changed to make gradient-based planning much more robust.

This blog post discusses work done with Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, and Amir Bar (* denotes equal advisorship), where we propose GRASP.

What is a world model?

These days, the term “world model” is quite overloaded, and depending on the context can either mean an explicit dynamics model or some implicit, reliable internal state that a generative model relies on (e.g. when an LLM generates chess moves, whether there is some internal representation of the board). We give our loose working definition below.

Suppose you take actions $a_t \in \mathcal{A}$ and observe states $s_t \in \mathcal{S}$ (images, latent vectors, proprioception). A world model is a learned model that, given the current state and a sequence of future actions, predicts what will happen next. Formally, it defines a predictive distribution on a sequence of observed states $s_{t-h:t}$ and current action $a_t$:

that approximates the environment’s true conditional $P(s_{t+1} \mid s_{t-h:t},\; a_t)$. For this blog post, we’ll assume a Markovian model $P(s_{t+1} \mid s_{t-h:t},\; a_t)$ for simplicity (all results here can be extended to the more general case), and when the model is deterministic it reduces to a map over states:

In practice the state $s_t$ is often a learned latent representation (e.g., encoded from pixels), so the model operates in a (theoretically) compact, differentiable space. The key point is that a world model gives you a differentiable simulator; you can roll it forward under hypothetical action sequences and backpropagate through the predictions.

Planning: choosing actions by optimizing through the model

Given a start $s_0$ and a goal $g$, the simplest planner chooses an action sequence $\mathbf{a}=(a_0,\dots,a_{T-1})$ by rolling out the model and minimizing terminal error:

Here we use $\mathcal{F}^T$ as shorthand for the full rollout through the world model (dependence on model parameters $\theta$ is implicit):

In short horizons and low-dimensional systems, this can work reasonably well. But as horizons grow and models become larger and more expressive, its weaknesses become amplified.

So why doesn’t this just work at scale?

Why long-horizon planning is hard (even when everything is differentiable)

There are two separate pain points for the more general world model, plus a third that is specific to learned, deep learning-based models.

1) Long-horizon rollouts create deep, ill-conditioned computation graphs

Those familiar with backprop through time (BPTT) may notice that we’re differentiating through a model applied to itself repeatedly, which will lead to the exploding/vanishing gradients problem. Namely, if we take derivatives (note we’re differentiating vector-valued functions, resulting in Jacobians that we denote with $D_x (\cdots)$) with respect to earlier actions (e.g. $a_0$):

We see that the Jacobian’s conditioning scales exponentially with time $T$:

leading to exploding or vanishing gradients.

2) The landscape is non-greedy and full of traps

At short horizons, the greedy solution, where we move straight toward the goal at every step, is often good enough. If you only need to plan a few steps ahead, the optimal trajectory usually doesn’t deviate much from “head toward $g$” at each step.

As horizons grow, two things happen. First, longer tasks are more likely to require non-greedy behavior: going around a wall, repositioning before pushing, backing up to take a better path. And as horizons grow, more of these non-greedy steps are typically needed. Second, the optimization space itself scales with horizon: $\mathrm{dim}(\mathcal{A} \times \cdots \times \mathcal{A}) = T\mathrm{dim}(\mathcal{A})$, further expanding the space of local minima for the optimization problem.

A long-horizon fix: lifting the dynamics constraint

Suppose we treat the dynamics constraint $s_{t+1} = F_{\theta}(s_t, a_t)$ as a soft constraint, and we instead optimize the following penalty function over both actions $(a_0,\ldots,a_{T-1})$ and states $(s_0,\ldots,s_T)$:

This is also sometimes called collocation in planning/robotics literature. Note the lifted formulation shares the same global minimizers as the original rollout objective (both are zero exactly when the trajectory is dynamically feasible). But the optimization landscapes are very different, and we get two immediate benefits:

• Each world model evaluation $F_{\theta}(s_t,a_t)$ depends only on local variables, so all $T$ terms can be computed in parallel across time, resulting in a huge speed-up for longer horizons, and
• You no longer backpropagate through a single deep $T$-step composition to get a learning signal, since the previous product of Jacobians now splits into a sum, e.g.:

Being able to optimize states directly also helps with exploration, as we can temporarily navigate through unphysical domains to find the optimal plan:

However, lunch is never free. And indeed, especially for deep learning-based world models, there is a critical issue that makes the above optimization quite difficult in practice.

An issue for deep learning-based world models: sensitivity of state-input gradients

The tl;dr of this section is: directly optimizing states through a deep learning-based $F_{\theta}$ is incredibly brittle, à la adversarial robustness. Even if you train your world model in a lower-dimensional state space, the training process for the world model makes unseen state landscapes very sharp, whether it be an unseen state itself or simply a normal/orthogonal direction to the data manifold.

Adversarial robustness and the “dimpled manifold” model

Adversarial robustness originally looked at classification models $f_\theta : \mathbb{R}^{w\times h \times c} \to \mathbb{R}^K$, and showed that by following the gradient of a particular logit $\nabla f_\theta^k$ from a base image $x$ (not of class $k$), you did not have to move far along $x’ = x + \epsilon\nabla f_\theta^k$ to make $f_\theta$ classify $x’$ as $k$ (Szegedy et al., 2014; Goodfellow et al., 2015):

Later work has painted a geometric picture for what’s going on: for data near a low-dimensional manifold $\mathcal{M}$, the training process controls behavior in tangential directions, but does not regularize behavior in orthogonal directions, thus leading to sensitive behavior (Stutz et al., 2019). Another way stated: $f_\theta$ has a reasonable Lipschitz constant when considering only tangential directions to the data manifold $\mathcal{M}$, but can have very high Lipschitz constants in normal directions. In fact, it often benefits the model to be sharper in these normal directions, so it can fit more complicated functions more precisely.

As a result, such adversarial examples are incredibly common even for a single given model. Further, this is not just a computer vision phenomenon; adversarial examples also appear in LLMs (Wallace et al., 2019) and in RL (Gleave et al., 2019).

While there are methods to train for more adversarially robust models, there is a known trade-off between model performance and adversarial robustness (Tsipras et al., 2019): especially in the presence of many weakly-correlated variables, the model must be sharper to achieve higher performance. Indeed, most modern training algorithms, whether in computer vision or LLMs, do not train adversarial robustness out. Thus, at least until deep learning sees a major regime change, this is a problem we’re stuck with.

Why is adversarial robustness an issue for world model planning?

Consider a single component of the dynamics loss we’re optimizing in the lifted state approach:

Let’s further focus on just the base state:

Since world models are typically trained on state/action trajectories $(s_1, a_1, s_2, a_2, \ldots)$, the state-data manifold for $F_{\theta}$ has dimensionality bounded by the action space:

where $\mathcal{R}$ is some optional space of augmentations (e.g. translations/rotations). Thus, we can typically expect $\mathrm{dim}(\mathcal{M}_s)$ to be much lower than $\mathrm{dim}(\mathcal{S})$, and thus: it is very easy to find adversarial examples that hack any state to any other desired state.

As a result, the dynamics optimization

feels incredibly “sticky,” as the base points $s_t$ can easily trick $F_{\theta}$ into thinking it’s already made its local goal.¹

Our fix

This is where our new planner GRASP comes in. The main observation: while $D_s F_{\theta}$ is untrustworthy and adversarial, the action space is usually low-dimensional and exhaustively trained, so $D_a F_{\theta}$ is actually reasonable to optimize through and doesn’t suffer from the adversarial robustness issue!

At its core, GRASP builds a first-order lifted state / collocation-based planner that is only dependent on action Jacobians through the world model. We thus exploit the differentiability of learned world models $F_{\theta}$, while not falling victim to the inherent sensitivity of the state Jacobians $D_s F_{\theta}$.

GRASP: Gradient RelAxed Stochastic Planner

As noted before, we start with the collocation planning objective, where we lift the states and relax dynamics into a penalty:

We then make two key additions.

Ingredient 1: Exploration by noising the state iterates

Even with a smoother objective, planning is nonconvex. We introduce exploration by injecting Gaussian noise into the virtual state updates during optimization.

A simple version:

Actions are still updated by non-stochastic descent:

The state noise helps you “hop” between basins in the lifted space, while the actions remain guided by gradients. We found that specifically noising states here (as opposed to actions) finds a good balance of exploration and the ability to find sharper minima.²

Ingredient 2: Reshape gradients: stop brittle state-input gradients, keep action gradients

As discussed, the fragile pathway is the gradient that flows into the state input of the world model, \(D_s F_{\theta}\). The most straightforward way to do this initially is to just stop state gradients into \(F_{\theta}\) directly:

• Let $\bar{s}_t$ be the same value as $s_t$, but with gradients stopped.

Define the stop-gradient dynamics loss:

This alone does not work. Notice now states only follow the previous state’s step, without anything forcing the base states to chase the next ones. As a result, there are trivial minima for just stopping at the origin, then only for the final action trying to get to the goal in one step.

Dense goal shaping

We can view the above issue as the goal’s signal being cut off entirely from previous states. One way to fix this is to simply add a dense goal term throughout prediction:

In normal settings this would over-bias towards the greedy solution of straight chasing the goal, but this is balanced in our setting by the stop-gradient dynamics loss’s bias towards feasible dynamics. The final objective is then as follows:

The result is a planning optimization objective that does not have dependence on state gradients.

Periodic “sync”: briefly return to true rollout gradients

The lifted stop-gradient objective is great for fast, guided exploration, but it’s still an approximation of the original serial rollout objective.

So every $K_{\text{sync}}$ iterations, GRASP does a short refinement phase:

1. Roll out from $s_0$ using current actions $\mathbf{a}$, and take a few small gradient steps on the original serial loss:

The lifted-state optimization still provides the core of the optimization, while this refinement step adds some assistance to keep states and actions grounded towards real trajectories. This refinement step can of course be replaced with a serial planner of your choice (e.g. CEM); the core idea is to still get some of the benefit of the full-path synchronization of serial planners, while still mostly using the benefits of the lifted-state planning.

How GRASP addresses long-range planning

Collocation-based planners offer a natural fix for long-horizon planning, but this optimization is quite difficult through modern world models due to adversarial robustness issues. GRASP proposes a simple solution for a smoother collocation-based planner, alongside stable stochasticity for exploration. As a result, longer-horizon planning ends up not only succeeding more, but also finding such successes faster:

Push-T results. Success rate (%) / median time to success. Bold = best in row. Note the median success time will bias higher with higher success rate; GRASP manages to be faster despite higher success rate.

What’s next?

There is still plenty of work to be done for modern world model planners. We want to exploit the gradient structure of learned world models, and collocation (lifted-state optimization) is a natural approach for long-horizon planning, but it’s crucial to understand typical gradient structure here: smooth and informative action gradients and brittle state gradients. We view GRASP as an initial iteration for such planners.

Extension to diffusion-based world models (deeper latent timesteps can be viewed as smoothed versions of the world model itself), more sophisticated optimizers and noising strategies, and integrating GRASP into either a closed-loop system or RL policy learning for adaptive long-horizon planning are all natural and interesting next steps.

I do genuinely think it’s an exciting time to be working on world model planners. It’s a funny sweet spot where the background literature (planning and control overall) is incredibly mature and well-developed, but the current setting (pure planning optimization over modern, large-scale world models) is still heavily underexplored. But, once we figure out all the right ideas, world model planners will likely become as commonplace as RL.

For more details, read the full paper or visit the project website.

Citation

@article{psenka2026grasp,
  title={Parallel Stochastic Gradient-Based Planning for World Models},
  author={Michael Psenka and Michael Rabbat and Aditi Krishnapriyan and Yann LeCun and Amir Bar},
  year={2026},
  eprint={2602.00475},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.00475}
}

Such underwater human-robot teaming is the focus of an MIT Lincoln Laboratory project funded through an internally administered R&D portfolio on autonomous systems and carried out by the Advanced Undersea Systems and Technology Group. The project seeks to leverage the respective strengths of humans and robots to optimize maritime missions for the U.S. military, including critical infrastructure inspection and repair, search and rescue, harbor entry, and countermine operations.

"Divers and AUVs generally don't team at all underwater," says principal investigator Madeline Miller. "Underwater missions requiring humans typically do so because they involve some sort of manipulation a robot can't do, like repairing infrastructure or deactivating a mine. Even ROVs are challenging to work with underwater in very skilled manipulation tasks because the manipulators themselves aren't agile enough."

Beyond their superior dexterity, humans excel at recognizing objects underwater. But humans working underwater can't perform complex computations or move very quickly, especially if they are carrying heavy equipment; robots have an edge over humans in processing power, high-speed mobility, and endurance. To combine these strengths, Miller and her team are developing hardware and algorithms for underwater navigation and perception — two key capabilities for effective human-robot teaming.

As Miller explains, divers may only have a compass and fin-kick counts to guide them. With few landmarks and potentially murky conditions caused by a lack of light at depth or the presence of biological matter in the water column, they can easily become disoriented and lost. For robots to help divers navigate, they need to perceive their environment. However, in the presence of darkness and turbidity, optical sensors (cameras) cannot generate images, while acoustic sensors (sonar) generate images that lack color and only show the shapes and shadows of objects in the scene. The historical lack of large, labeled sonar image datasets has hindered training of underwater perception algorithms. Even if data were available, the dynamic ocean can obscure the true nature of objects, confusing artificial intelligence. For instance, a downed aircraft broken into multiple pieces, or a tire covered in an overgrowth of mussels, may no longer resemble an aircraft or tire, respectively.

"Ultimately, we want to devise solutions for navigation and perception in expeditionary environments," Miller says. "For the missions we're thinking about, there is limited or no opportunity to map out the area in advance. For the harbor entry mission, maybe you have a satellite map but no underwater map, for example."

On the navigation side, Miller's team picked up on work started by the MIT Marine Robotics Group, led by John Leonard, to develop diver-AUV teaming algorithms. With their navigation algorithms, Leonard's group ran simulations under optimal conditions and performed field testing in calm waters using human-paddled kayaks as proxies for both divers and AUVs. Miller's team then integrated these algorithms into a mission-relevant AUV and began testing them under more realistic ocean conditions, initially with a support boat acting as a diver surrogate, and then with actual divers.

"We quickly learned that you need more sensing capabilities on the diver when you factor in ocean currents," Miller explains. "With the algorithms demonstrated by MIT, the vehicle only needed to calculate the distance, or range, to the diver at regular intervals to solve the optimization problem of estimating the positions of both the vehicle and diver over time. But with the real ocean forces pushing everything around, this optimization problem blows up quickly."

On the perception side, Miller's team has been developing an AI classifier that can process both optical and sonar data mid-mission and solicit human input for any objects classified with uncertainty.

"The idea is for the classifier to pass along some information — say, a bounding box around an image — to the diver and indicate, "I think this is a tire, but I'm not sure. What do you think?" Then, the diver can respond, "Yes, you've got it right, or no, look over here in the image to improve your classification," Miller says.

This feedback loop requires an underwater acoustic modem to support diver-AUV communication. State-of-the-art data rates in underwater acoustic communications would require tens of minutes to send an uncompressed image from the AUV to the diver. So, one aspect the team is investigating is how to compress information into a minimum amount to be useful, working within the constraints of the low bandwidth and high latency of underwater communications and the low size, weight, and power of the commercial off-the-shelf (COTS) hardware they're using. For their prototype system, the team procured mostly COTS sensors and built a sensor payload that would easily integrate into an AUV routinely employed by the U.S. Navy, with the goal of facilitating technology transition. Beyond sonar and optical sensors, the payload features an acoustic modem for ranging to the diver and several data processing and compute boards.

Miller's team has tested the sensor-equipped AUV and algorithms around coastal New England — including in the open ocean near Portsmouth, New Hampshire, with the University of New Hampshire's (UNH) Gulf Surveyor and Gulf Challenger coastal research vessels as diver surrogates, and on the Boston-area Charles River, with an MIT Sailing Pavilion skiff as the surrogate.

"The UNH boats are well-equipped and can access realistic ocean conditions. But pretending to be a diver with a large boat is hard. With the skiff, we can move more slowly and get the relative motion in tune with how a diver and AUV would navigate together."

Last summer, the team started testing equipment with human divers at Michigan Technological University's Great Lakes Research Center. Although the divers lacked an interface to feed back information to the AUV, each swam holding the team's tube-shaped prototype tablet, dubbed a "tube-let." The tube-let was equipped with a pressure and depth sensor, inertial measurement unit (to track relative motion), and ranging modem — all necessary components for the navigation algorithms to solve the optimization problem.

"A challenge during testing was coordinating the motion of the diver and vehicle, because they don't yet collaborate," Miller says. "Once the divers go underwater, there is no communication with the team on the surface. So, you have to plan where to put the diver and vehicle so they don't collide."

The team also worked on the perception problem. The water clarity of the Great Lakes at that time of year allowed for underwater imaging with an optical sensor. Caroline Keenan, a Lincoln Scholars Program PhD student jointly working in the laboratory's Advanced Undersea Systems and Technology Group and Leonard's research group at MIT, took the opportunity to advance her work on knowledge transfer from optical sensors to sonar sensors. She is exploring whether optical classifiers can train sonar classifiers to recognize objects for which sonar data doesn't exist. The motivation is to reduce the human operator load associated with labeling sonar data and training sonar classifiers.

With the internally funded research program coming to an end, Miller's team is now seeking external sponsorship to refine and transition the technology to military or commercial partners.

"The modern world runs on undersea telecommunication and power cables, which are vulnerable to attack by disruptive actors. The undersea domain is becoming increasingly contested as more nations develop and advance the capabilities of autonomous maritime systems. Maintaining global economic security and U.S. strategic advantage in the undersea domain will require leveraging and combining the best of AI and human capabilities," Miller says.

Like the fibers that bundle together to form biological muscles, these fibers can be arranged in different configurations to meet the demands of a given task. Unlike conventional robotic actuation systems, they are compliant enough to interface comfortably with the human body and operate silently without motors, external pumps, or other bulky supporting hardware.

The new electrofluidic fiber muscles — electrically driven actuators built in fiber format — are described in a recent paper published in Science Robotics. The work is led by Media Lab PhD candidate Ozgun Kilic Afsar; Vito Cacucciolo, a professor at the Politecnico di Bari; and four co-authors.

The new system brings together two technologies, Afsar explains. One is a fluidically driven artificial muscle known as a thin McKibben actuator, and the other is a miniaturized solid-state pump based on electrohydrodynamics (EHD), which can generate pressure inside a sealed fluid compartment without moving parts or an external fluid supply.

Until now, most fluid-driven soft actuators have relied on external “heavy, bulky, oftentimes noisy hydraulic infrastructure,” Afsar says, “which makes them difficult to integrate into systems where mobility or compact, lightweight design is important.” This has created a fundamental bottleneck in the practical use of fluidic actuators in real-world applications.

The key to breaking through that bottleneck was the use of integrated pumps based on electrohydrodynamic principles. These millimeter-scale, electrically driven pumps generate pressure and flow by injecting charge into a dielectric fluid, creating ions that drag the fluid along with them. Weighing just a few grams each and not much thicker than a toothpick, they can be fabricated continuously and scaled easily. “We integrated these fiber pumps into a closed fluidic circuit with the thin McKibben actuators,” Afsar says, noting that this was not a simple task given the different dynamics of the two components.

A key design strategy was to pair these fibers in what are known as antagonistic configurations. Cacucciolo explains that this is where “one muscle contracts while another elongates,” as when you bend your arm and your biceps contract while your triceps stretch. In their system, a millimeter-scale fiber pump sits between two similarly scaled McKibben actuators, driving fluid into one actuator to contract it while simultaneously relaxing the other.

“This is very much reminiscent of how biological muscles are configured and organized,” Afsar says. “We didn’t choose this configuration simply for the sake of biomimicry, but because we needed a way to store the fluid within the muscle design.” The need for an external reservoir open to the atmosphere has been one of the main factors limiting the practical use of EHD pumps in robotic systems outside the lab. By pairing two McKibben fibers in line, with a fiber pump between them to form a closed circuit, the team eliminated that need entirely.

Another key finding was that the muscle fibers needed to be pre-pressurized, rather than simply filled. “There is a minimum internal system pressure that the system can tolerate,” Afsar says, “below which the pump can degrade or temporarily stop working.” This happens because of cavitation, in which vapor bubbles form when the pressure at the pump inlet drops below the vapor pressure of the liquid, eventually leading to dielectric breakdown.

To prevent cavitation, they applied a “bias” pressure from the outset so that the pressure at the fiber pump inlet never falls below the liquid’s vapor pressure. The magnitude of this bias pressure can be adjusted depending on the application. “To achieve the maximum contraction the muscle can generate, we found there is a specific bias pressure range that is optimal,” she says. “If you want to configure the system for faster response, you might increase that bias pressure, though with some reduction in maximum contraction.”

Cacucciolo adds that most of today’s robotic limbs and hands are built around electric servo motors, whose configuration differs fundamentally from that of natural muscles. Servo motors generate rotational motion on a shaft that must be converted into linear movement, whereas muscle fibers naturally contract and extend linearly, as do these electrofluidic fibers. 

“Most robotic arms and humanoid robots are designed around the servo motors that drive them,” he says. “That creates integration constraints, because servo motors are hard to package densely and tend to concentrate mass near the joints they drive. By contrast, artificial muscles in fiber form can be packed tightly inside a robot or exoskeleton and distributed throughout the structure, rather than concentrated near a joint.”

These electrofluidic muscles may be especially useful for wearable applications, such as exoskeletons that help a person lift heavier loads or assistive devices that restore or augment dexterity. But the underlying principles could also apply more broadly. “Our findings extend to fluid-driven robotic systems in general,” Cacucciolo says. “Wherever fluidic actuators are used, or where engineers want to replace external pumps with internal ones, these design principles could apply across a wide range of fluid-driven robotic systems.”

This work “presents a major advancement in fiber-format soft actuation,” which “addresses several long-standing hurdles in the field, particularly regarding portability and power density,” says Herbert Shea, a professor in the Soft Transducers Laboratory at Ecole Polytechnique Federale de Lausanne in Switzerland, who was not associated with this research. “The lack of moving parts in the pump makes these muscles silent, a major advantage for prosthetic devices and assistive clothing,” he says.

Shea adds that “this high-quality and rigorous work bridges the gap between fundamental fluid dynamics and practical robotic applications. The authors provide a complete system-level solution — characterizing the individual components, developing a predictive physical model, and validating it through a range of demonstrators.”

In addition to Afsar and Cacucciolo, the team also included Gabriele Pupillo and Gennaro Vitucci at Politecnico di Bari and Wedyan Babatain and Professor Hiroshi Ishii at the MIT Media Lab. The work was supported by the European Research Council and the Media Lab’s multi-sponsored consortium.

Now, MIT engineers have designed an ultrasound wristband that precisely tracks a wearer’s hand movements in real-time. The wristband produces ultrasound images of the wrist’s muscles, tendons, and ligaments as the hand moves, and is paired with an artificial intelligence algorithm that continuously translates the images into the corresponding positions of the five fingers and palm.

The researchers can train the wristband to learn a wearer’s hand motions, which the device can communicate in real-time to a robot or a virtual environment.

In demonstrations, the team has shown that a person wearing the wristband can wirelessly control a robotic hand. As the person gestures or points, the robot does the same. In a sort of wireless marionette interaction, the wearer can manipulate the robot to play a simple tune on the piano and shoot a small basketball into a desktop hoop. With the same wristband, a wearer can also manipulate objects on a computer screen, for instance pinching their fingers together to enlarge and minimize a virtual object.

The team is using the wristband to gather hand motion data from many more users with different hand sizes, finger shapes, and gestures. They envision building a large dataset of hand motions that can be plumbed, for instance, to train humanoid robots in dexterity tasks, such as performing certain surgical procedures. The ultrasound band could also be used to grasp, manipulate, and interact with objects in video games, design applications, or other virtual settings.

“We think this work has immediate impact in potentially replacing hand tracking techniques with wearable ultrasound bands in virtual and augmented reality,” says Xuanhe Zhao, the Uncas and Helen Whitaker Professor of Mechanical Engineering at MIT. “It could also provide huge amounts of training data for dexterous humanoid robots.”

Zhao, Gengxi Lu, and their colleagues present the wristband’s new design in a paper appearing today in Nature Electronics. Their MIT co-authors are former postdocs Xiaoyu Chen, Shucong Li, and Bolei Deng; graduate students SeongHyeon Kim and Dian Li; postdocs Shu Wang and Runze Li; and Anantha Chandrakasan, MIT provost and the Vannevar Bush Professor of Electrical Engineering and Computer Science. Other co-authors are graduate students Yushun Zheng and Junhang Zhang, Baoqiang Liu, Chen Gong, and Professor Qifa Zhou from the University of Southern California.

Seeing strings

There are currently a number of approaches to capturing and mimicking human hand dexterity in robots. Some approaches use cameras to record a person’s hand movements as they manipulate objects or perform tasks. Others involve having a person wear a glove with sensors, which records the person’s hand movements and transmits the data to a receiving robot. But erecting a complex camera system for different applications is impractical and prone to visual obstacles. And sensor-laden gloves could limit a person’s natural hand motions and sensations.

A third approach uses the electrical signals from muscles in the wrist or forearm that scientists then correlate with specific hand movements. Researchers have made significant advances in this approach, however these signals are easily affected by noise in the environment. They are also not sensitive enough to distinguish subtle changes in movements. For instance, they may discern whether a thumb and index finger are pinched together or pulled apart, but not much of the in-between path.

Zhao’s team wondered whether ultrasound imaging might capture more dexterous and continuous hand movements. His group has been developing various forms of ultrasound stickers — miniaturized versions of the transducers used in doctor’s offices that are paired with hydrogel material that can safely stick to skin.

In their new study, the team incorporated the ultrasound sticker design into a wearable wristband to continuously image the muscles and tendons in the wrist.

“The tendons and muscles in your wrist are like strings pulling on puppets, which are your fingers,” Lu says. “So the idea is: Each time you take a picture of the state of the strings, you’ll know the state of the hand.”

Mapping manipulation

The team designed a wristband with an ultrasound sticker that is the size of a smartwatch, and added onboard electronics that are about as small as a cellphone. They attached the wristband to a volunteer’s wrist and confirmed that the device produced clear and continuous images of the wrist as the volunteer moved their fingers in various gestures.

The challenge then was to relate the black and white ultrasound images of the wrist to specific positions of the hand. As it turns out, the fingers and thumb are capable of 22 degrees of freedom, or different ways of extending or angling. The researchers found that they could identify specific regions in their ultrasound images of the wrist that correlate to each of these 22 degrees of freedom. For instance, changes in one region relate to thumb extension, while changes in another region correlate with movements of the index finger.

To establish these connections, a volunteer wearing the wristband would move their hand in various positions while the researchers recorded the gestures with multiple cameras surrounding the volunteer. By matching changes in certain regions of the ultrasound images with hand positions recorded by the cameras, the team could label wrist image regions with the corresponding degree of freedom in the hand. But to do this translation continuously, and in real-time, would be an impossible task for humans.

So, the team turned to artificial intelligence. They used an AI algorithm that can be trained to recognize image patterns and correlate them with specific labels and, in this case, the hand’s various degrees of freedom. The researchers trained the algorithm with ultrasound images that they meticulously labeled, annotating the image regions associated with a specific degree of freedom. They tested the algorithm on a new set of ultrasound images and found it correctly predicted the corresponding hand gestures.

Once the researchers successfully paired the AI algorithm with the wristband, they tested the device on more volunteers. For the new study, eight volunteers with different hand and wrist sizes wore the wristband while they formed various hand gestures and grasps, including making the signs for all 26 letters in American Sign Language. They also held objects such as a tennis ball, a plastic bottle, a pair of scissors, and a pencil. In each case, the wristband precisely tracked and predicted the position of the hand.

To demonstrate potential applications, the team developed a simple computer program that they wirelessly paired with the wristband. As a wearer went through the motions of pinching and grasping, the gestures corresponded to zooming in and out on an object on the computer screen, and virtually moving and manipulating it in a smooth and continuous fashion.

The researchers also tested the wristband as a wireless controller of a simple commercial robotic hand. While wearing the wristband, a volunteer went through the motions of playing a keyboard. The robot in turn mimicked the motions in real-time to play a simple tune on a piano. The same robot was also able to mimic a person’s finger taps to play a desktop basketball game.

Zhao is planning to further miniaturize the wristband’s hardware, as well as train the AI software on many more gestures and movements from volunteers with wider ranging hand sizes and shapes. Ultimately, the team is building toward a wearable hand tracker that can be worn by anyone, to wirelessly manipulate humanoid robots or virtual objects with high dexterity.

“We believe this is the most advanced way to track dexterous hand motion, through wearable imaging of the wrist,” Zhao says. “We think these wearable ultrasound bands can provide intuitive and versatile controls for virtual reality and robotic hands.”

This research was supported, in part, by MIT, the U.S. National Institutes of Health, the U.S. National Science Foundation, the U.S. Department of Defense, and Singapore National Research Foundation through the Singapore-MIT Alliance for Research and Technology.

At MIT’s Center for Bits and Atoms (CBA) in late January, about a dozen students explored that possibility during an Independent Activities Period (IAP) workshop on Raman spectroscopy, a technique that uses laser light to “fingerprint” materials. The session even featured a robotic dog equipped with sensing equipment, demonstrating how chemical analysis can be done remotely.

The workshop, led by MIT postdoc Lamyaa Almehmadi in collaboration with the CBA, introduced participants to a powerful technique now used by law enforcement and first responders to identify narcotics and explosives, by gemologists to authenticate precious stones, and pharmaceutical companies to verify raw materials and ensure product quality. CBA graduate researcher Jiaming Liu co-hosted, delivering lectures, demonstrating Raman equipment, and contributing to the curriculum and hands-on demonstrations.

“It can open up new possibilities for innovation across many fields,” said Almehmadi, an analytical chemist in the Department of Materials Science and Engineering (DMSE). After attendees learned the fundamentals, she encouraged them to think creatively about new applications: “My hope is to inspire all of you to think about doing something with Raman spectroscopy that no one has done before.”

Fingerprinting materials

Participants brought items to class to analyze using handheld devices, which fire laser light and measure how it bounces back. The resulting pattern behaves like a molecular fingerprint, identifying the materials in the item — whether it’s a paper clip, a piece of tree bark, or a mixing bowl.

Workshop attendee Sarah Ciriello, an administrative assistant at DMSE who brought a stone she found at the beach, was taken aback by the results. The Raman device suggested a 39 percent probability that the sample contained concrete-like material, with the remaining readings matching synthetic compounds — blurring the line between natural and manufactured materials.

“It’s man-made — I was surprised,” Ciriello said.

Developed in 1928 by Indian scientist C.V. Raman, who later won the Nobel Prize in Physics, Raman spectroscopy was groundbreaking because it used visible light to probe materials without destroying them, a major advantage over other techniques at the time, such as chromatography or mass spectrometry. But for decades, the Raman signal — the light scattered back from a sample — was weak, and the instruments were big and bulky, limiting its practical use.

Advances in lasers, computing power, and miniaturized optics have transformed Raman spectroscopy into a portable tool. Today’s handheld devices can instantly compare a sample’s molecular fingerprint against vast digital libraries, allowing users to identify thousands of materials in seconds. Because it doesn’t destroy the sample, Raman is especially useful in fields that require preserving materials — such as law enforcement, where evidence must remain intact, and art restoration.

Almehmadi’s own research focuses on advancing Raman spectroscopy by developing highly sensitive, semiconductor-based sensors that make portable chemical analysis possible, with applications ranging from medical diagnostics to forensic and environmental monitoring.

“Raman can be used to analyze any material,” Almehmadi says. “That’s why I decided to introduce it to students from diverse backgrounds.”

IAP classes are open to students and staff across MIT, and the Raman workshop reflected that range — from administrative staff to graduate and undergraduate students and postdocs in departments and labs including DMSE, the Department of Mechanical Engineering, the Media Lab, and the Broad Institute.

Walking the robot dog

A crowd-pleasing element in the workshop was the integration of a robot dog that belongs to the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). The demonstration highlighted how Raman technology can be used in dangerous environments, such as crime scenes or toxic industrial sites.

The handheld device was secured to the robot using tape, and Almehmadi showed how she could navigate the dog to a plastic bag filled with a white powder — baking soda.

But in a real-world scenario, “How can we know if it is baking soda or not?” she says. “So we just shined the light, and then the instrument told us what it was.”

Participants used a Wi-Fi app on their phones to view the results and a small remote controller to operate the robotic dog themselves.

“I loved the robot dog,” Ciriello says. “I was able to control it a bit, but it was challenging because the gauge was really sensitive.”

Michael Kitcher, a postdoc in DMSE, also praises the robot demonstration.

“Given that we just duct taped the device onto the dog — it was cool to see it actually worked,” he says.

Looking ahead

Kitcher, who researches magnetic materials for electronic applications, joined the workshop to learn more about Raman spectroscopy, which he had read about but never used. He was impressed by its versatility — in addition to the beach stone and baking soda, the device identified materials in a contact lens, cosmetics, and even a diamond.

Although it struggled to analyze a piece of chocolate he brought — other signals from the chocolate interfered — Kitcher sees strong potential for his own research. One area he’s interested in is unconventional magnetic materials, such as altermagnets, with unusual magnetic behavior that researchers hope to better understand and control for more energy-efficient electronics.

“Over the last couple of years, people have been trying to get a better sense of why these materials behave the way they do — how we can control this unconventional magnetic order,” he says. Raman spectroscopy can probe the vibrations of atoms in a material, helping researchers detect patterns in the crystal structure that underlie unusual magnetic behaviors. By understanding these vibrations, scientists could unlock material design rules that enable ultra-fast, low-energy computing.

Hands-on workshops like this — that inspire innovative future applications — Almehmadi says, are at the heart of an MIT education.

“I’ve always learned best by doing,” she says. “Lectures and reading are important, but real understanding comes from hands-on experience.”

Not vendors. Not investors. Not even many researchers.

Because admitting the problem means admitting that the current trajectory of AI hype—bigger models, more data, more “intelligence”—rests on a foundation that is already cracking.

This week, we confront the crisis head‑on.

1. The Illusion of Infinite Data

For years, the industry behaved as if the internet were an endless reservoir of pristine training material. But the reality is stark:

• High‑quality, human‑generated text is finite.
• Much of the web is duplicated, spam‑ridden, or low‑signal.
• The best datasets have already been scraped—multiple times.

A 2022 study from Epoch AI estimated that the supply of high‑quality language data could be exhausted by 2026–2032, depending on consumption rates. That window is closing fast.

The myth of infinite data was convenient. It justified the “just scale it” era. But the numbers no longer support that fantasy.

2. When Quantity Masquerades as Quality

The industry’s response to data scarcity has been predictable: Use more data, even if it’s worse.

This has led to:

• Synthetic data loops (models training on their own outputs)
• Massive inclusion of low‑quality web text
• Relaxed filtering standards
• Increased reliance on user‑generated content (which is noisy by design)

The problem? Models trained on degraded data don’t just plateau—they drift, hallucinate, and amplify errors.

Synthetic data in particular creates a recursive collapse: Models generate data → that data trains new models → the signal decays with each generation.

It’s the AI equivalent of photocopying a photocopy until the image becomes noise.

3. The Corporate Incentive to Stay Quiet

Why isn’t this crisis openly acknowledged?

Because the incentives run in the opposite direction:

• Vendors want to sell bigger models.
• Investors want to believe in exponential growth.
• Enterprises want to believe they’re buying “intelligence,” not statistical mimicry.
• Researchers want to publish breakthroughs, not bottlenecks.

Admitting data scarcity would force a shift from “scale solves everything” to “we need new paradigms.” And paradigm shifts are expensive.

4. The Rise of Synthetic Data: Solution or Mirage?

Synthetic data is marketed as the savior of AI scaling. But the truth is more nuanced.

Strengths:

• Cheap
• Fast
• Infinite
• Useful for narrow, structured tasks

Weaknesses:

• Lacks true novelty
• Reinforces model biases
• Degrades signal‑to‑noise ratio
• Risks “model collapse” when used at scale

A 2023 paper from Stanford and Rice researchers warned that synthetic data can cause irreversible performance degradation if not carefully controlled.

Synthetic data is a tool—not a replacement for human‑generated knowledge.

5. The Coming Divide: Data‑Rich vs. Data‑Poor AI

We are entering a bifurcated AI landscape:

Tier 1: Data‑Rich Models

These are built by organizations with:

• Exclusive licensing deals
• Proprietary datasets
• Partnerships with publishers, platforms, and content owners
• The capital to acquire or generate high‑quality human data

These models will continue to improve.

Tier 2: Data‑Poor Models

These rely on:

• Public web data
• Synthetic data
• Open‑source scrapes
• Crowdsourced or low‑quality corpora

These models will stagnate or regress.

The divide will not be about compute. It will be about who controls the last reservoirs of clean human knowledge.

6. The Ethical and Legal Storm Brewing

The data crisis intersects with a second, equally volatile issue: copyright and consent.

As publishers, artists, and platforms push back, the supply of legally usable training data shrinks further.

Recent lawsuits—from authors, news organizations, and image creators—signal a future where:

• High‑quality data becomes paywalled
• Licensing becomes mandatory
• Training costs rise dramatically
• Open‑source models face existential constraints

The era of “scrape now, apologize later” is ending.

7. What Comes Next: The Post‑Scarcity Strategy

The industry must pivot from “more data” to better data.

This means:

1. Curated, domain‑specific datasets

Precision over volume.

2. Human‑in‑the‑loop reinforcement

Not cheap annotation farms—expert‑level refinement.

3. Data provenance and traceability

Knowing where data came from and how it was used.

4. Hybrid architectures

Models that combine symbolic reasoning, retrieval systems, and neural networks.

5. Ethical, compensated data partnerships

A sustainable ecosystem where creators are part of the value chain.

The next generation of AI will not be defined by scale. It will be defined by data stewardship.

8. The Reality Check

The data quality crisis is not a footnote—it is the defining constraint of the next decade of AI development.

Ignoring it is easy. Admitting it is uncomfortable. Solving it is essential.

The organizations that confront this reality now will lead the next era of AI. Those that cling to the illusion of infinite data will be left behind.

Key References (with links)

1. Epoch AI — “Will We Run Out of Data? Limits of LLM Scaling Based on Human‑Generated Data” (2024)

Direct link: https://epochai.org/blog/will-we-run-out-of-data

Summary: Peer‑reviewed analysis estimating ~300T tokens of usable human text and projecting exhaustion between 2026–2032.

2. Associated Press / CityNews — “AI ‘Gold Rush’ for Chatbot Training Data Could Run Out of Human‑Written Text” (2024)

Direct link: https://www.citynews.ca/halifax/ai-gold-rush-for-chatbot-training-data-could-run-out-of-human-written-text-2-8637894

Summary: AP‑reported global analysis confirming that public human‑written text may be depleted between 2026–2032, with industry scrambling for high‑quality sources.

3. Forbes — “AI May Be Running Out of Data, Stanford Report Warns” (2026)

Direct link: https://www.forbes.com/sites/joemckendrick/2026/04/14/ai-may-be-running-out-of-data-stanford-report-warns/

Summary: Coverage of the 2026 Stanford AI Index Report, highlighting industry‑wide concerns about “peak data,” limits of synthetic data, and sustainability of scaling laws.

Conceived, written and published by AI Quantum Intelligence with the help of AI models.

For the past decade, AI progress has marched to a familiar rhythm: bigger models, larger datasets, more GPUs, more power. Scaling laws became the industry’s compass, and compute became the currency of innovation. But by early 2026, the cracks in that paradigm were impossible to ignore. Training costs ballooned. Energy demands surged. Frontier models approached physical, economic, and thermodynamic limits.

The industry hit what many quietly called the scaling wall.

Now, for the first time, a credible path beyond that wall is emerging—not through incremental GPU gains or clever sparsity tricks, but through quantum‑accelerated optimization. This isn’t the sci‑fi dream of fully universal quantum computers replacing classical systems. It’s something more immediate, more practical, and potentially more disruptive.

Quantum optimization is positioning itself as the first real break from the bottlenecks that have defined AI’s trajectory for years.

The Scaling Wall: A Problem No One Can Ignore

The scaling wall isn’t a single constraint—it's a convergence of several:

• Training costs doubling every 6–9 months

• Memory bandwidth limits choking model parallelism

• Diminishing returns from brute‑force parameter growth

• Energy ceilings at hyperscale data centers

• Latency constraints for real‑time inference

Even with next‑gen accelerators, the industry is running out of room. The physics of classical compute simply doesn’t bend fast enough.

This is where quantum enters—not as a replacement, but as a pressure valve for the most computationally punishing parts of AI.

Quantum Optimization: The Missing Accelerator

Quantum optimization focuses on a specific class of problems that dominate AI workloads:

• Large‑scale matrix factorization

• Combinatorial search

• Model routing and mixture‑of‑experts scheduling

• Hyperparameter and architecture optimization

• Reinforcement learning policy search

• Sparse attention routing

• Multi‑objective optimization for training efficiency

These tasks are notoriously expensive on classical hardware. But quantum systems—especially annealers, photonic processors, and early gate-based devices—excel at exploring vast solution spaces in parallel.

The result: orders‑of‑magnitude speedups in the optimization loops that govern training, inference, and model design.

This isn’t hypothetical. In Q1 2026 alone:

• Quantum‑assisted MoE routing reduced training time by 30–40% in early enterprise pilots.

• Hybrid quantum‑classical solvers cut reinforcement learning search costs by up to 70%.

• Quantum annealing demonstrated superior scaling on large combinatorial optimization tasks relevant to model compression and architecture search.

These aren’t full‑stack quantum models. They’re quantum‑accelerated classical models—and that distinction matters.

Why This Breaks the Scaling Wall

The scaling wall exists because classical compute hits diminishing returns. Quantum optimization breaks that cycle by attacking the hardest parts of AI workloads:

1. Faster Training Without Bigger Clusters

Quantum solvers reduce the number of iterations needed to converge on optimal weights, architectures, or routing patterns. Fewer iterations = less compute = lower cost.

2. Better Models Without More Parameters

Quantum‑accelerated search can uncover architectures that classical methods miss—enabling leaps in performance without parameter inflation.

3. Energy Efficiency Gains That Actually Matter

Quantum systems, especially photonic and annealing‑based designs, can solve certain optimization tasks with dramatically lower energy budgets.

4. New Regimes of Model Design

Quantum‑accelerated architecture search opens the door to model families that would be computationally unreachable today.

This is the first time in years that AI progress has a path forward that doesn’t rely on simply stacking more GPUs.

Redefining “Frontier Models”

Today, “frontier model” is shorthand for “the biggest model money can train.” Quantum acceleration changes that definition entirely.

Frontier models of the quantum‑accelerated era will be defined by:

• Optimization depth, not parameter count

• Search efficiency, not brute‑force scaling

• Hybrid compute architectures, not monolithic GPU clusters

• Quantum‑assisted reasoning, not just larger transformers

• Energy‑aware intelligence, not energy‑indifferent growth

A frontier model in 2027 may have fewer parameters than a 2025 model—yet outperform it because quantum‑accelerated optimization found a better architecture, better routing, or better training trajectory.

This is a paradigm shift: Frontier no longer means “bigger.” It means “better optimized.”

The Hybrid Future: Quantum as a Co‑Processor for Intelligence

The most realistic near-term architecture is hybrid:

• Classical GPUs/TPUs handle dense linear algebra

• Quantum accelerators handle optimization, search, routing, and compression

• Photonic interconnects bridge the two worlds

• Distributed orchestration systems schedule workloads across both domains

This hybrid model mirrors how GPUs once entered the data center — first as niche accelerators, then as essential infrastructure.

Quantum is following the same trajectory, but faster.

What This Means for the Industry

1. Training Costs Will Fall — Dramatically

Quantum‑accelerated optimization could cut training budgets by 30–60% within the next three years.

2. Smaller Labs Will Re‑Enter the Frontier Race

If optimization becomes the differentiator, not raw compute, innovation becomes more accessible.

3. Model Design Will Become a Quantum‑Native Discipline

Just as deep learning created new engineering roles, quantum‑accelerated AI will create new optimization‑centric specializations.

4. The AI Arms Race Will Shift From “More Compute” to “Smarter Compute”

Efficiency becomes the new battleground.

The Breakthrough We’ve Been Waiting For

For years, the AI industry has been sprinting toward a wall—faster, harder, with ever‑larger budgets. Quantum optimization doesn’t just slow the collision. It opens a door in the wall.

A door to:

• More efficient intelligence

• More accessible innovation

• More sustainable compute

• More powerful models

• And a fundamentally new definition of what “frontier” means

Quantum‑accelerated AI isn’t the future of AI. It’s the first real escape route from the limits of the present.

Written and published by AI Quantum Intelligence with the help of AI models.

So let’s start with regulation. The government is not just looking on. It’s not just dipping its toes in the water. It’s jumping in and trying to make waves. Check out the following articles to see what I mean.

The argument is heating up. Safety and control are now on the table. And by control, I mean who actually owns the keys to these systems.

But as with everything in life, this isn’t just about politics. This is about whether innovation can survive regulation without being completely destroyed in the process.

Some argue you need guardrails. Others argue that guardrails kill innovation. I’m reminded of the time they put a speed limit on the Autobahn. It’s safer. But not everyone was happy about it.

Then there’s the very thorny issue of content. AI-generated books are now hitting the shelves. And sometimes nobody even notices. Could you imagine reading a book and getting to the halfway point and realizing that the book wasn’t even written by a human?

Welcome to the brave new world. Check out the following articles to see what I mean. Authenticity and ownership are now in play.

And it gets personal. Because readers love an author’s “voice.” That slight mistake in a sentence that makes it more memorable. Can AI do that? Yes. Should it? Well, that’s a different story. A story people aren’t afraid to tell. Loudly.

And then there’s business. Companies aren’t just playing with AI. They’re monitoring it. They’re measuring it. They’re trying to figure out how to use it every day without losing control. Banks. Tech companies. You name it.

They’re all asking themselves the same question: How can we use AI without losing control? Check out the following articles to see what I mean. Enterprise adoption is now coming with strings attached.

But AI isn’t just about screens. It’s about robots. It’s about automation. It’s about machines that don’t just think. They act. And AI is increasingly being let out of its box. It’s being used in factories. In warehouses. Even on our streets.

And when that happens, the stakes get a lot higher. Because when AI screws up on a screen, you can just hit the reset button. But when AI screws up in real life? Well, that’s a different story altogether. Check out the following articles to see what I mean. The stakes just got real.

So what does it all mean?

Well. We’re somewhere in the middle. We’re excited. We’re nervous. And we’re trying to figure out the rules of the road as we go. AI isn’t going to slow down.

But it is going to keep maturing. And like anything in its teenage years, it’s going to trip and fall before it finds its feet.

One thing is for sure. The debate has shifted. We’re no longer asking, “What can AI do?” We’re asking, “What should AI do?” And to be honest, that’s a much tougher question to answer.

If you skim through today’s news, like the report on White House efforts to curb dangerous advanced AI, you’ll get a sense of what is going on. The government, bankers, and AI leaders are all in urgent talks over something.

Why are they meeting with such urgency? Several current state-of-the-art AI models are not just able to write letters or make pictures, they are able to write software, find flaws in security, and leave people a little worried.

What’s surprising about this is that this is not something that is happening somewhere in the future. It is happening right now. Someone said that everything was happening “faster than we expected”, which is another way of saying we might not be acting fast enough.

But, let’s step back for a minute. This was not a sudden surprise. If you were following the evolution of the technology or something like the current debate over the proper ways to control and ethically use AI, then you’d know that each new milestone has generated a “hold up, let’s wait” response. And, yet, the reaction has never been strong enough.

What sets this apart is that the atmosphere has gotten tense. It’s not hopeful and anxious; it’s fearful. To make this clear: if AI can uncover security vulnerabilities without help in key systems, then it’s not just an efficiency, it’s a threat. That is my view and I know those in charge are fearful of that.

Meanwhile, tech companies are not standing still. They are working fast on improving their AI. Well, why wouldn’t they? The money is great. As the headlines about the race for AI dominance show, countries and companies are treating AI like it’s something new and it will be a disaster if they’re late.

But, there is this odd unease that is not talked about: What if the machines get too smart to contain? Not the “AI is going to kill us all” version, but the non-alarms and yet even more frightening version.

Devices making decisions we can’t grasp, tools that can be weaponized faster than we can stop them. It is like if we gave our citizens brand-new super-cars, but there were no new roads to handle them, or any way to stop them.

It’s not America. Countries all over are grappling with the same dilemma. In the European Union, leaders are attempting to introduce a new regulation as they try to implement the EU AI Act. Different approach, same question: How do you use the best tool, and not have it get out of hand?

To me, that’s where we are now. The excitement is not gone; the anxiety is just beginning. It was the early days of the internet, no one knew where it was going, but everyone thought it was a huge change. Just, maybe now, it feels more serious.

So what’s left to do? It looks like we will have to find a way to walk the line between innovation and caution, balancing both sides without falling in a hole. From what we can tell from all these White House meetings, it seems like those in power already see how delicate that balance is.

It turns out you don’t need to dream big to make this conversation a reality. It’s now on the agenda of OpenAI.

OpenAI released a proposal suggesting AI doesn’t merely increase productivity. It also offers a way of restructuring the economy. Shorter working weeks. New forms of public wealth. Taxes on work done by AI.

And this might be what is truly disturbing and compelling about this report: If the robots are doing all the work, what are people doing?

The report advocates for a four-day workweek, which is already under trial in countries around the world. The results have been good, higher productivity, greater satisfaction among workers, fewer burnouts.

We all know how we feel on Friday afternoon when we try to get things done. The broader picture of that experiment, shows that the four-day workweek could be a reality after all.

Not because AI rose up and destroyed their workplaces, but because it assumed many of their responsibilities. And AI doesn’t draw a salary.

So far this year, more than 50,000 tech workers have lost their jobs, and employers say AI tools are making it easier to cut staff.

While layoffs began last year, companies have accelerated the process in 2026 by relying on AI to replace workers in roles such as software testing and customer service, according to employers.

The trend shows no signs of slowing down as companies overhaul their operations to incorporate AI, which can perform tasks without rest or complaint.

There’s another element to this story. At least one laid off engineer told me, “I helped train the AI that replaced me.”

Ironic? Yes. According to company executives, it’s something else: the inevitable march of progress.

This is not a one-off thing. A lot of big tech companies are embracing automation.

The No. 1 thing is companies are trying to adopt AI, and they’re using AI to automate jobs … [Companies are] literally looking at where they can replace people with AI models.

Then there are the people who don’t think this is a huge deal. Like some economists, who say that this is just another industrial revolution, only this time it’s happening really fast and involves lots of computers.

But those new jobs require very different skills, which makes it hard for employers and pundits to tell laid-off workers that they should simply “reskill.”

I talked to a few tech employees this week and they’re feeling both thrilled and terrified about AI. Here’s one developer: “What AI can do is amazing. But what’s kind of terrifying is how fast it’s making us unnecessary.”

The word that jumps out there is “unnecessary.” For a long time, working in tech was seen as a sure thing. Now it doesn’t seem that way at all.

Which brings us to the obvious question: Now what?

It’s way too soon to tell. For now, we’re left with a sense of wonder mixed with unease. AI isn’t a villain, but it is disrupting lots of people’s lives.

Maybe the issue isn’t that machines are taking our jobs. Maybe the issue is that we weren’t prepared for how fast they’d do it.

If this is the start of a long story, that’s scary. But it’s also exciting.

Buckle up.

AI is quietly changing the way software is built and it’s a massive deal. Instead of software applications that users must learn to navigate, we are witnessing the birth of applications that users can talk to.

Once you notice it, it’s impossible to ignore.  Software applications are moving from dashboards and processes to conversational interfaces. In essence, users don’t want tools, they want results, and AI is now making it possible to deliver them.

That is, if you want to generate a report that takes 5 tabs to do, you could simply type “generate a summary of our quarterly results” and that’s it. It’s handy, but it is also somewhat weird. It is as if software went from something that you use to something that works for you. It is as if it is a coworker.

But it is not just this article that is saying that. I looked at a bunch of other articles that discuss what the current state of the industry is, and there is a trend towards automating entire business processes, and letting AI take the reigns.

Companies are optimizing for speed, for efficiency, for scalability, and perhaps for lack of knowledge about what’s going on in their companies. I don’t know.

I am torn. On the one hand, I like that idea. I would rather do less busywork. On the other hand, I feel like we’re going to lose knowledge about what we are doing. It is as if everyone using GPS makes us all terrible at navigating.

In addition, there are far-reaching economic consequences. AI isn’t just changing the way we interact with software, it’s changing the way we work. That’s great, in principle. Except that now the tasks that enabled that productivity may be handled by the AI, too.

Oh, and almost no one is talking about the psychological effects of all of this. We’re not just changing what we use, we’re changing how we behave. Instead of needing to understand something, we only need to understand how to ask about it.

This is a massive shift. Yes, it makes things easier, more accessible… but it also makes us far more dependent on things we don’t understand.

I came across a more general treatment of how AI is affecting society that sort of touches on this tradeoff between efficiency and agency, speed and understanding.Everything is complicated. It always is.

So yes. Software is improving. Getting faster. Simpler.

But the thought that keeps echoing in my brain is this: when we make everything as simple as asking a question… do we stop caring about the answers?

Because once you’ve gotten used to software that will simply listen to you… well. There’s just no going back.

The post Asia’s $78 Billion AI and quantum inflection point draws global tech leaders to Singapore  appeared first on IoT Now News - How to run an IoT enabled business.

The post NVIDIA AI Ecosystem expands as Marvell joins forces through NVLink Fusion appeared first on IoT Now News - How to run an IoT enabled business.

The post IBM and ETH Zurich join forces to shape the future of algorithms for the AI and quantum era appeared first on IoT Now News - How to run an IoT enabled business.

The post SEALSQ and IC’Alps achieve key common criteria certification steps appeared first on IoT Now News - How to run an IoT enabled business.

The post UK automotive’s EV crossroads: pressure, pushback and the race to net zero appeared first on IoT Now News - How to run an IoT enabled business.

The post Soracom expands professional services to North America appeared first on IoT Now News - How to run an IoT enabled business.

The post When AI and Energy Collide appeared first on IoT Now News - How to run an IoT enabled business.

The post IoT Now Contract Win List – March 2026 appeared first on IoT Now News - How to run an IoT enabled business.

The post EMQX Enterprise 6.2 introduces native agent discovery and governance for AI and IoT systems appeared first on IoT Now News - How to run an IoT enabled business.

The post Has connected intelligence for resource-agnostic IoT arrived? appeared first on IoT Now News - How to run an IoT enabled business.

“The unique resources of MIT.nano enable not just the foundational research of academia, but the translation of that research into commercial innovations through startups,” says START.nano Program Manager Joyce Wu SM ’00, PhD ’07. “The START.nano accelerator supports early-stage companies from MIT and beyond with the tools and network they need for success.”

Launched in 2021, START.nano aims to increase the survival rate of hard-tech startups by easing their journey from the lab to the real world. In addition to receiving access to MIT.nano’s laboratories, program participants are invited to present at startup exhibits at MIT conferences, and in exclusive events including the newly launched PITCH.nano competition.

“For an early-stage startup working at the frontier of superconductor discovery, the combination of infrastructure and community has been irreplaceable,” says Jason Gibson, CEO and co-founder of Quantum Formatics. “START.nano isn’t just a resource,” adds Cynthia Liao MBA ’24, CEO and co-founder of Vertical Semiconductor. “It’s a strategic advantage that accelerates our roadmap, allowing us to iterate quickly to meet customer needs and strengthen our competitive edge.”

Although an MIT affiliation is not required, five of the 16 companies in the new cohort are led by MIT alumni, and an additional three have MIT affiliation. In total, 49 percent of the startups in START.nano are founded by MIT graduates.

Here are the intended impacts of the 16 new START.nano companies:

Acorn Genetics is developing a "smartphone of sequencing," launching the power of genetic analysis out of slow, centralized labs and into the hands of consumers for fast, portable, and affordable sequencing.

Addis Energy leverages oil, gas, and geothermal drilling technologies to unlock the chemical potential of iron-rich rocks. By injecting engineered fluids, they harness the earth’s natural energy to produce ammonia that is both abundant and cost-effective.

Augmend Health uses virtual reality and AI to deliver clinical data intelligence services for specialty care that turns incomplete documentation into revenue, compliance, and better treatment decisions.

Brightlight Photonics is building high-performance laser infrastructure at chip scale, integrating Titanium:Sapphire gain to deliver broadband, high-power, low-noise optical sources for advanced photonic systems.

Cahira Technologies is creating the new paradigm of brain-computer symbiosis for treating intractable diseases and human augmentation through autonomous, nonsurgical neural implants.

Copernic Catalysts is leveraging computational modeling to develop and commercialize transformational catalysts for low-cost and sustainable production of bulk chemicals and e-fuels.

Daqus Energy is unlocking high-energy lithium-ion batteries using critical metal-free organic cathodes.

Electrified Thermal Solutions is reinventing the firebrick to electrify industrial heat.

Guardion is making analytical instruments, chemical detectors, and radiation detectors more sensitive, portable, and easier to scale with nanomaterial-based ion detectors.

Mantel Capture is designing carbon capture materials to operate at the high temperatures found inside boilers, kilns, and furnaces — enabling highly efficient carbon capture that has not been possible until now.

nOhm Devices is developing highly-efficient cryogenic electronics for quantum computers and sensors.

Quantum Formatics is speeding discovery of the world’s next superconductors using proprietary AI.

Qunett is building the foundational hardware stack for deployable quantum networks to power the next era of global connectivity.

Rheyo is developing new ways to make dental care more effective, efficient, and easy through advanced materials and technology.

Vertical Semiconductor is commercializing high-voltage, high-density, high-efficiency vertical GaN (gallium nitride) to power the next era of compute.

VioNano Innovations is developing specialty material solutions that reduce variability and improve precision in semiconductor manufacturing, allowing chipmakers to build even smaller, faster, and more cost-effective chips.

START.nano now comprises over 32 companies and 11 graduates — ventures that have moved beyond the prototyping stages, and some into commercialization. See the full list here.

MIT researchers have now developed a system that boosts the performance of storage devices by handling three major sources of variability simultaneously. Their approach delivers significant speed improvements over traditional methods that tackle only one source of variability at a time.

The system uses a two-tier architecture, with a central controller that makes big-picture decisions about which tasks each storage device performs, and local controllers for each machine that rapidly reroute data if that device is struggling.

The method, which can adapt in real-time to shifting workloads, does not require specialized hardware. When the researchers tested this system on realistic tasks like AI model training and image compression, it nearly doubled the performance delivered by traditional approaches. By intelligently balancing the workloads of multiple storage devices, the system can increase overall data center efficiency.

“There is a tendency to want to throw more resources at a problem to solve it, but that is not sustainable in many ways. We want to be able to maximize the longevity of these very expensive and carbon-intensive resources,” says Gohar Chaudhry, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this technique. “With our adaptive software solution, you can still squeeze a lot of performance out of your existing devices before you need to throw them away and buy new ones.”

Chaudhry is joined on the paper by Ankit Bhardwaj, an assistant professor at Tufts University; Zhenyuan Ruan PhD ’24; and senior author Adam Belay, an associate professor of EECS and a member of the MIT Computer Science and Artificial Intelligence Laboratory. The research will be presented at the USENIX Symposium on Networked Systems Design and Implementation.

Leveraging untapped performance

Solid-state drives (SSDs) are high-performance digital storage devices that allow applications to read and write data. For instance, an SSD can store vast datasets and rapidly send data to a processor for machine-learning model training.   

Pooling multiple SSDs together so many applications can share them improves efficiency, since not every application needs to use the entire capacity of an SSD at a given time. But not all SSDs perform equally, and the slowest device can limit the overall performance of the pool.

These inefficiencies arise from variability in SSD hardware and the tasks they perform.

To utilize this untapped SSD performance, the researchers developed Sandook, a software-based system that tackles three major forms of performance-hampering variability simultaneously. “Sandook” is an Urdu word that means “box,” to signify “storage.”

One type of variability is caused by differences in the age, amount of wear, and capacity of SSDs that may have been purchased at different times from multiple vendors.

The second type of variability is due to the mismatch between read and write operations occurring on the same SSD. To write new data to the device, the SSD must erase some existing data. This process can slow down data reads, or retrievals, happening at the same time.

The third source of variability is garbage collection, a process of gathering and removing outdated data to free up space. This process, which slows SSD operations, is triggered at random intervals that a data center operator cannot control.

“I can’t assume all SSDs will behave identically through my entire deployment cycle. Even if I give them all the same workload, some of them will be stragglers, which hurts the net throughput I can achieve,” Chaudhry explains.

Plan globally, react locally

To handle all three sources of variability, Sandook utilizes a two-tier structure. A global schedular optimizes the distribution of tasks for the overall pool, while faster schedulers on each SSD react to urgent events and shift operations away from congested devices.

The system overcomes delays from read-write interference by rotating which SSDs an application can use for reads and writes. This reduces the chance reads and writes happen simultaneously on the same machine.

Sandook also profiles the typical performance of each SSD. It uses this information to detect when garbage collection is likely slowing operations down. Once detected, Sandook reduces the workload on that SSD by diverting some tasks until garbage collection is finished.

“If that SSD is doing garbage collection and can’t handle the same workload anymore, I want to give it a smaller workload and slowly ramp things back up. We want to find the sweet spot where it is still doing some work, and tap into that performance,” Chaudhry says.

The SSD profiles also allow Sandook’s global controller to assign workloads in a weighted fashion that considers the characteristics and capacity of each device.

Because the global controller sees the overall picture and the local controllers react on the fly, Sandook can simultaneously manage forms of variability that happen over different time scales. For instance, delays from garbage collection occur suddenly, while latency caused by wear and tear builds up over many months.

The researchers tested Sandook on a pool of 10 SSDs and evaluated the system on four tasks: running a database, training a machine-learning model, compressing images, and storing user data. Sandook boosted the throughput of each application between 12 and 94 percent when compared to static methods, and improved the overall utilization of SSD capacity by 23 percent.

The system enabled SSDs to achieve 95 percent of their theoretical maximum performance, without the need for specialized hardware or application-specific updates.

“Our dynamic solution can unlock more performance for all the SSDs and really push them to the limit. Every bit of capacity you can save really counts at this scale,” Chaudhry says.

In the future, the researchers want to incorporate new protocols available on the latest SSDs that give operators more control over data placement. They also want to leverage the predictability in AI workloads to increase the efficiency of SSD operations.

“Flash storage is a powerful technology that underpins modern datacenter applications, but sharing this resource across workloads with widely varying performance demands remains an outstanding challenge. This work moves the needle meaningfully forward with an elegant and practical solution ready for deployment, bringing flash storage closer to its full potential in production clouds,” says Josh Fried, a software engineer at Google and incoming assistant professor at the University of Pennsylvania, who was not involved with this work.

This research was funded, in part, by the National Science Foundation, the U.S. Defense Advanced Research Projects Agency, and the Semiconductor Research Corporation.

“Nuclear energy has been a tremendous part of our nation’s energy infrastructure for the past 60 years, and the number of people who maintain that infrastructure is incredibly small,” says Price, an MIT assistant professor in the Department of Nuclear Science and Engineering (NSE), as well as the Atlantic Richfield Career Development Professor in Energy Studies. “By becoming a nuclear engineer, you become one of a select number of people responsible for carbon-free energy generation in the United States.” 

That was a mission he was eager to take part in, and the goals he set for himself were far from modest: He wanted to help design and usher in a new class of nuclear reactors, building on the safety, economics, and reliability of the existing nuclear fleet.

Price has never wavered from this objective, and he’s only found encouragement along the way. The nuclear engineering community, he says, “is small, close-knit, and very welcoming. Once you get into it, most people are not inclined to do anything else.”

Illuminating the relationships between physical processes

In his first research project as an undergraduate at the University of Illinois Urbana at Champaign, Price studied the safety of the steel and concrete casks used to store spent reactor fuel rods after they’ve cooled off in tanks of water, typically for several years. His analysis indicated that this storage method was quite safe, although the question as to what should ultimately be done with these fuel casks, in terms of long-term disposal, remains open in this country.

After starting graduate studies at the University of Michigan in 2020, Price took up a different line of research that he’s still engaged in today. That area of study, called multiphysics modeling, involves looking at various physical processes going on in the core of a nuclear reactor to see how they interact — an alternative to studying these processes one at a time.

One key process, neutronics, concerns how neutrons buzz around in the reactor core causing nuclear fission, which is what generates the power. A second process, called thermal hydraulics, involves cooling the reactor to extract the heat generated by neutrons. A multiphysics simulation, analyzing how these two processes interact, could show how the heat carried away as the reactor produces power affects the behavior of neutrons, because the hotter the fuel is, the less likely it is to cause fission.

“If you ever want to change your power level, or do anything with the reactor, the temperature of the fuel is a critical input that you need to know,” says Price. “Multiphysics modeling allows us to correlate the fission neutronics processes with a thermal property, temperature. That, in turn, can help us predict how the reactor will behave under different conditions.”

Multiphysics modeling for light water reactors, which are the ones operating today with capacities on the order of 1,000 megawatts, are pretty well established, Prices says. But methods for modeling advanced reactors — small modular reactors (SMRs with capacities ranging from around 20 to 300 MW) and microreactors (rated at 1 to 20 MW) — are far less advanced. Only a very small number of these reactors are operating today, but Price is focusing his efforts on them because of their potential to produce power more cheaply and more safely, along with their greater flexibility in power and size.   

Although multiphysics simulations have supplied the nuclear community with a wealth of information, they can require supercomputers to solve, or find approximate solutions to, coupled and extremely difficult nonlinear equations. In the hopes of greatly reducing the computational burden, Price is actively exploring artificial intelligence approaches that could provide similar answers while bypassing those burdensome equations altogether. That has been a central theme of his research agenda since he joined the MIT faculty in September 2025.

A crucial role for artificial intelligence

What artificial intelligence and machine-learning methods, in particular, are good at is finding patterns concealed within data, such as correlations between variables critical to the functioning of a nuclear plant. For example, Price says, “if you tell me the power level of your reactor, it [AI] could tell you what the fuel temperature is and even tell you the 3-dimensional temperature distribution in your core.” And if this can be done without solving any complicated differential equations, computational costs could be greatly reduced.

Price is investigating several applications where AI may be especially useful, such as helping with the design of novel kinds of reactors. “We could then rely on the safety frameworks developed over the past 50 years to carry out a safety analysis of the proposed design,” he says. “In this way, AI will not be directly interfacing with anything that is safety-critical.” As he sees it, AI’s role would be to augment established procedures, rather than replacing them, helping to fill in existing gaps in knowledge.

When a machine-learning model is given a sufficient amount of data to learn from, it can help us better understand the relationship between key physical processes — again without having to solve nonlinear differential equations. 

“By really pinning down those relationships, we can make better design decisions in the early stages,” Price says. “And when that technology is developed and deployed, AI can help us make more intelligent control decisions that will enable us to operate our reactors in a safer and more economical way.”

Giving back to the community that nurtured him

Simply put, one of his chief goals is to bring the benefits of AI to the nuclear industry, and he views the possibilities as vast and largely untapped. Price also believes that he is well-positioned as a professor at MIT to bring us closer to the nuclear future that he envisions. As he sees it, he’s working not only to develop the next generation of reactors, but also to help prepare the next generation of leaders in the field.

Price became acquainted with some prospective members of that “next generation” in a design course he co-taught last fall with Curtis Smith, the KEPCO Professor of the Practice of Nuclear Science and Engineering. For Price, that introduction lasted just a few months, but it was long enough for him to discover that MIT students are exceptionally motivated, hard-working, and capable. Not surprisingly, those happen to be the same qualities he’s hoping to find in the students that join his research team.

Price vividly recalls the support he received when taking his first, tentative steps in this field. Now that he’s moved up the ranks from undergraduate to professor, and acquired a substantial body of knowledge along the way, he wants his students “to experience that same feeling that I had upon entering the field.” Beyond his specific goals for improving the design and operation of nuclear reactors, Price says, “I hope to perpetuate the same fun and healthy environment that made me love nuclear engineering in the first place.”

But while these AI-driven outputs may be technically optimal, are they fair? What if a low-cost power distribution strategy leaves disadvantaged neighborhoods more vulnerable to outages than higher-income areas?

To help stakeholders quickly pinpoint potential ethical dilemmas before deployment, MIT researchers developed an automated evaluation method that balances the interplay between measurable outcomes, like cost or reliability, and qualitative or subjective values, such as fairness.   

The system separates objective evaluations from user-defined human values, using a large language model (LLM) as a proxy for humans to capture and incorporate stakeholder preferences. 

The adaptive framework selects the best scenarios for further evaluation, streamlining a process that typically requires costly and time-consuming manual effort. These test cases can show situations where autonomous systems align well with human values, as well as scenarios that unexpectedly fall short of ethical criteria.

“We can insert a lot of rules and guardrails into AI systems, but those safeguards can only prevent the things we can imagine happening. It is not enough to say, ‘Let’s just use AI because it has been trained on this information.’ We wanted to develop a more systematic way to discover the unknown unknowns and have a way to predict them before anything bad happens,” says senior author Chuchu Fan, an associate professor in the MIT Department of Aeronautics and Astronautics (AeroAstro) and a principal investigator in the MIT Laboratory for Information and Decision Systems (LIDS).

Fan is joined on the paper by lead author Anjali Parashar, a mechanical engineering graduate student; Yingke Li, an AeroAstro postdoc; and others at MIT and Saab. The research will be presented at the International Conference on Learning Representations.

Evaluating ethics

In a large system like a power grid, evaluating the ethical alignment of an AI model’s recommendations in a way that considers all objectives is especially difficult.

Most testing frameworks rely on pre-collected data, but labeled data on subjective ethical criteria are often hard to come by. In addition, because ethical values and AI systems are both constantly evolving, static evaluation methods based on written codes or regulatory documents require frequent updates.

Fan and her team approached this problem from a different perspective. Drawing on their prior work evaluating robotic systems, they developed an experimental design framework to identify the most informative scenarios, which human stakeholders would then evaluate more closely.

Their two-part system, called Scalable Experimental Design for System-level Ethical Testing (SEED-SET), incorporates quantitative metrics and ethical criteria. It can identify scenarios that effectively meet measurable requirements and align well with human values, and vice versa.   

“We don’t want to spend all our resources on random evaluations. So, it is very important to guide the framework toward the test cases we care the most about,” Li says.

Importantly, SEED-SET does not need pre-existing evaluation data, and it adapts to multiple objectives.

For instance, a power grid may have several user groups, including a large rural community and a data center. While both groups may want low-cost and reliable power, each group’s priority from an ethical perspective may vary widely.

These ethical criteria may not be well-specified, so they can’t be measured analytically.

The power grid operator wants to find the most cost-effective strategy that best meets the subjective ethical preferences of all stakeholders.

SEED-SET tackles this challenge by splitting the problem into two, following a hierarchical structure. An objective model considers how the system performs on tangible metrics like cost. Then a subjective model that considers stakeholder judgements, like perceived fairness, builds on the objective evaluation.

“The objective part of our approach is tied to the AI system, while the subjective part is tied to the users who are evaluating it. By decomposing the preferences in a hierarchical fashion, we can generate the desired scenarios with fewer evaluations,” Parashar says.

Encoding subjectivity

To perform the subjective assessment, the system uses an LLM as a proxy for human evaluators. The researchers encode the preferences of each user group into a natural language prompt for the model.

The LLM uses these instructions to compare two scenarios, selecting the preferred design based on the ethical criteria.

“After seeing hundreds or thousands of scenarios, a human evaluator can suffer from fatigue and become inconsistent in their evaluations, so we use an LLM-based strategy instead,” Parashar explains.

SEED-SET uses the selected scenario to simulate the overall system (in this case, a power distribution strategy). These simulation results guide its search for the next best candidate scenario to test.

In the end, SEED-SET intelligently selects the most representative scenarios that either meet or are not aligned with objective metrics and ethical criteria. In this way, users can analyze the performance of the AI system and adjust its strategy.

For instance, SEED-SET can pinpoint cases of power distribution that prioritize higher-income areas during periods of peak demand, leaving underprivileged neighborhoods more prone to outages.

To test SEED-SET, the researchers evaluated realistic autonomous systems, like an AI-driven power grid and an urban traffic routing system. They measured how well the generated scenarios aligned with ethical criteria.

The system generated more than twice as many optimal test cases as the baseline strategies in the same amount of time, while uncovering many scenarios other approaches overlooked.

“As we shifted the user preferences, the set of scenarios SEED-SET generated changed drastically. This tells us the evaluation strategy responds well to the preferences of the user,” Parashar says.

To measure how useful SEED-SET would be in practice, the researchers will need to conduct a user study to see if the scenarios it generates help with real decision-making.

In addition to running such a study, the researchers plan to explore the use of more efficient models that can scale up to larger problems with more criteria, such as evaluating LLM decision-making.

This research was funded, in part, by the U.S. Defense Advanced Research Projects Agency.

But previews generated by most 3D-printing software focus on function rather than aesthetics. A printed object may end up with a different color, texture, or shading than the user expected, resulting in multiple reprints that waste time, effort, and material.

To help users envision how a fabricated object will look, researchers from MIT and elsewhere developed an easy-to-use preview tool that puts appearance first.

Users upload a screenshot of the object from their 3D-printing software, along with a single image of the print material. From these inputs, the system automatically generates a rendering of how the fabricated object is likely to look.

The artificial intelligence-powered system, called VisiPrint, is designed to work with a range of 3D-printing software and can handle any material example. It considers not only the color of the material, but also gloss, translucency, and how nuances of the fabrication process affect the object’s appearance.

Such aesthetics-focused previews could be especially useful in areas like dentistry, by helping clinicians ensure temporary crowns and bridges match the appearance of a patient’s teeth, or in architecture, to aid designers in assessing the visual impact of models.

“3D printing can be a very wasteful process. Some studies estimate that as much as a third of the material used goes straight to the landfill, often from prototypes the user ends of discarding. To make 3D printing more sustainable, we want to reduce the number of tries it takes to get the prototype you want. The user shouldn’t have to try out every printing material they have before they settle on a design,” says Maxine Perroni-Scharf, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on VisiPrint.

She is joined on the paper by Faraz Faruqi, a fellow EECS graduate student; Raul Hernandez, an MIT undergraduate; SooYeon Ahn, a graduate student at the Gwangju Institute of Science and Technology; Szymon Rusinkiewicz, a professor of computer science at Princeton University; William Freeman, the Thomas and Gerd Perkins Professor of EECS at MIT and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Stefanie Mueller, an associate professor of EECS and Mechanical Engineering at MIT, and a member of CSAIL. The research will be presented at the ACM CHI Conference on Human Factors in Computing Systems.

Accurate aesthetics

The researchers focused on fused deposition modeling (FDM), the most common type of 3D printing. In FDM, print material filament is melted and then squirted through a nozzle to fabricate an object one layer at a time.

Generating accurate aesthetic previews is challenging because the melting and extrusion process can change the appearance of a material, as can the height of each deposited layer and the path the nozzle follows during fabrication.

VisiPrint uses two AI models that work together to overcome those challenges.

The VisiPrint preview is based on two inputs: a screenshot of the digital design from a user’s 3D-printing software (called “slicer” software), and an image of the print material, which can be taken from an online source or captured from a printed sample.

From these inputs, a computer vision model extracts features from the material sample that are important for the object’s appearance.

It feeds those features to a generative AI model that computes the geometry and structure of the object, while incorporating the so-called “slicing” pattern the nozzle will follow as it extrudes each layer.

The key to the researchers’ approach is a special conditioning method. This involves carefully adjusting the inner workings of the model to guide it, so it follows the slicing pattern and obeys the constraints of the 3D-printing process.

Their conditioning method utilizes a depth map that preserves the shape and shading of the object, along with a map of the edges that reflects the internal contours and structural boundaries.

“If you don’t have the right balance of these two things, you could use up with bad geometry or an incorrect slicing pattern. We had to be careful to combine them in the right way,” Perroni-Scharf says.

A user-focused system

The team also produced an easy-to-use interface where one can upload the required images and evaluate the preview.

The VisiPrint interface enables more advanced makers to adjust multiple settings, such as the influence of certain colors on the final appearance.

In the end, the aesthetic preview is intended to complement the functional preview generated by slicer software, since VisiPrint does not estimate printability, mechanical feasibility, or likelihood of failure.

To evaluate VisiPrint, the researchers conducted a user study that asked participants to compare the system to other approaches. Nearly all participants said it provided better overall appearance as well as more textural similarity with printed objects.

In addition, the VisiPrint preview process took about a minute on average, which was more than twice as fast as any competing method.

“VisiPrint really shined when compared to other AI interfaces. If you give a more general AI model the same screenshots, it might randomly change the shape or use the wrong slicing pattern because it had no direct conditioning,” she says.

In the future, the researchers want to address artifacts that can occur when model previews have extremely fine details. They also want to add features that allow users to optimize parts of the printing process beyond color of the material.

“It is important to think about the way that we fabricate objects. We need to continue striving to develop methods that reduce waste. To that end, this marriage of AI with the physical making process is an exciting area of future work,” Perroni-Scharf says.

“‘What you see is what you get’ has been the main thing that made desktop publishing ‘happen’ in the 1980s, as it allowed users to get what they wanted at first try. It is time to get WYSIWYG for 3D printing as well. VisiPrint is a great step in this direction,” says Patrick Baudisch, a professor of computer science at the Hasso Plattner Institute, who was not involved with this work.

This research was funded, in part, by an MIT Morningside Academy for Design Fellowship and an MIT MathWorks Fellowship.