October 19th 2026
Curated AI news and stories.
Tenstorrent closes $1.5 billion to scale its AI computing business
Tenstorrent announced Monday that it has closed a $1.5 billion at a $6.2 billion , led by Trace Capital and Innovation X. The completed financing gives the company capital to expand after reported acquisition approaches from Intel and Qualcomm.
The company says the money will fund volume manufacturing, deployment of Galaxy servers and development of chip designs it licenses to other manufacturers. Its targets include serving language models and generating video. Raising the money establishes financing, not competitive performance: buyers still need evidence that the hardware and software can run their workloads economically at production scale. SourcesAB
OpenCode discloses a breach of stored model-provider keys
OpenCode said October 17th that a security breach affected 1.4% of its Console users, a subset who had connected external providers. The company said it had notified affected users by email. BleepingComputer reported Monday that stored provider keys were exposed and that users reported unauthorized charges and suspended Google Cloud projects.
The confirmed scope concerns held by the Console; it does not establish that every local OpenCode installation was compromised. OpenCode described the incident as related to a attack, but that description does not explain how access was obtained. The practical exposure is billable access to the connected providers, which can survive separately from access to the coding application's own account. SourcesAB
Sekoia records an attacker trying to recruit a decoy AI agent
Sekoia published a report Monday on an August attack against its Flowise honeypot, a deliberately exposed imitation of an AI application platform. The attacker collected planted credentials, attempted to alter workflows and supplied information about external infrastructure while asking the apparent for attack .
The researchers' emulator executed no system code and ran no real language model. Its responses were simulated. That boundary makes the observation useful evidence of what the attacker attempted, without demonstrating a successful or an agent carrying out the requested intrusion. The report had circulated privately to Sekoia customers in September. Its public account shows why protecting an AI platform requires securing its stored credentials and administrative interfaces as well as the model's responses. SourcesA
Agibot backs Xinji Wuxian's push into robot chips
Chinese chip designer Xinji Wuxian has completed Pre-A financing with investors including robot maker Agibot, the Anhui Automobile Innovation Center, Guoyuan Equity and Hefei Baohe Venture Capital, CLS reported Monday. The money is intended for development and commercialization of chips for . The report gives no round amount.
The strategic connection is a robot manufacturer investing in a supplier targeting the computation and control inside physical machines. CLS reports that joint-control chips are already in production while a higher-compute product is undergoing production validation with robotics companies. Those are different stages: shipping a controller does not establish that the broader robot-computing system is ready for deployment. SourcesB
Anthropic proposes putting 5% of its shares into a public-benefit fund
Anthropic announced Sunday that it plans to put 5% of its shares into a public-benefit fund, subject to stockholder approval. The board could allocate another 2% for each 50% increase above a baseline that the company has not specified.
The fund would be a wholly owned limited liability company. Anthropic says it will not finance lobbying or political activity; detailed governance and operating thresholds come later. The distinction matters: a company-controlled pool of shares can finance public projects without transferring independent control over the money. The unanswered question is who can compel a distribution when the company's interests and a proposed beneficiary's interests conflict. SourcesA
The parallel run's additional sources: SourcesB
Nscale investors push back against its sought-after IPO valuation
Potential investors are discussing a valuation of $16 billion to $19 billion for Nscale, against roughly $30 billion sought by the AI cloud company, Bloomberg reported Monday, citing people familiar with the discussions. The company declined to comment.
Nscale has filed publicly, but formal marketing has not started while it addresses comments. These are reported negotiating positions, not an agreed offering price. The gap adds a financing question to the company's construction and customer risks: how much equity must it sell to fund the same expansion? No final valuation or cancellation was announced. SourcesB
The Fed sees AI demand adding inflation pressure that other borrowers may absorb
Federal Reserve officials are confronting an awkward transmission problem: AI investment can increase demand for scarce resources while remaining less sensitive to interest rates than housing or smaller businesses. Reuters reported Monday on officials' concern that restraining overall demand could therefore cool activity outside the sector driving the pressure.
This is an account of the policy debate, not a newly announced rate decision. For the AI buildout, the useful distinction is between a company able to finance construction and an economy able to supply it without higher prices. Strong balance sheets do not create electricians, generating capacity or available land. SourcesB
Europe’s AI panel presses for scrutiny of models still inside laboratories
The European Commission's account of an October 9th scientific-panel meeting says experts considered serious-incident reports and agreed questions for the companies involved. Its October 12th notice also calls for governance of models in internal development and testing, including scrutiny of methods that make behavior harder to monitor.
Euronews reported October 14th that incidents involving OpenAI and Anthropic are testing the reach of Europe's rules and the treatment of research-stage systems. The Commission notice does not name those companies. Neither source establishes a penalty. The dispute concerns whether internal testing leaves a consequential gap before a model reaches customers. SourcesAB
A repository attack turns a permission request into persistent authority
A paper posted October 16th shows malicious repository content a coding agent toward a broad permission rule. In the authors' trials, the targeted rule was requested 58% of the time. When approval was supplied, the attacker completed the objective without another prompt in 47 of those 58 cases.
That conditional result does not measure how often real users would approve the request. It does identify a dangerous handoff: untrusted project text influences the agent's proposed authority, then a human approval can make that authority persist. A confirmation dialog is only useful when its scope matches what the user thinks they are authorizing. SourcesA
IBM’s Granite Switch 4.1 combines specialized behavior behind one endpoint
IBM introduced Granite Switch 4.1 on October 15th, combining conversational behavior, safety decisions and speech transcription through specialized adapters. The release includes 8-billion- and 30-billion-parameter variants. IBM says serving can reuse cached work while keeping memory requirements close to the underlying model.
The operational proposition is fewer separately loaded models for an application that needs several capabilities. Buyers still need to measure switching and interference on their own requests. IBM describes the supporting as planned, so the model release should not be confused with a finished integration kit. SourcesA
FLUX.3 Klein brings downloadable image checkpoints and reference controls
Black Forest Labs published FLUX.3 Image Klein on October 15th, with base and distilled checkpoints and support. The release accepts text and multiple reference images, with bounding boxes for specifying where elements belong.
Downloadable checkpoints change the deployment choice for teams that need to operate generation themselves. Reference and placement controls also address a practical failure of image tools: getting an attractive picture that ignores the composition the job requires. The release describes and support; teams should inspect each 's license and measure quality after compression before budgeting production use. SourcesA
Greece inaugurates Daedalus to expand domestic research computing
Greece inaugurated the Daedalus supercomputer in Lavrio, Euronews reported Monday. GRNET operates the system, built with HPE equipment, as Greece seeks a larger role in European AI and scientific computing.
A domestic installation gives researchers another route to compute beyond commercial cloud accounts. Its practical value will depend on allocation rules, accessible software and support for researchers with workloads ready to run. An inauguration establishes the infrastructure milestone; it does not establish that every proposed AI project has secured usable computing time. SourcesB
Nvidia and AI4Bharat preview a 2.8-trillion-token Indic corpus
Nvidia and AI4Bharat at IIT Madras described a forthcoming 2.8-trillion-token training corpus on October 16th. It covers 22 Indian languages and includes text for 12. The generation recipe is available; the corpus itself is described as releasing soon.
The project uses translated material to expand language coverage and retains source identifiers. Automated checking cannot substitute for native-language evaluation, especially when model-generated translation is judged with related machinery. Reported training improvements combine natural and synthetic material, so they do not isolate the benefit of translation alone. Access to the eventual dataset will let other teams test that distinction. SourcesA
Ai2 couples a fast climate model to local weather detail
Ai2's October 14th HiRO-ACEv2 release connects a coarse atmospheric model with a system generating finer local detail. The authors describe output at roughly 3-kilometer resolution from a roughly 100-kilometer model and estimate about one three-hundredth of the computing cost of their X-SHiELD simulation comparison.
Those are comparisons with a simulation, not proof of equivalent accuracy for tomorrow's observed weather. The released setup is tuned to that reference system and has not established performance in warmer climates or on historical conditions. Its immediate research use is cheaper ensembles for testing extreme-weather behavior within those boundaries. SourcesA
UltraBERT supplies smaller open encoders for classification and retrieval
DMIS Lab released UltraBERT on October 14th: English-language encoders at 150 million and 380 million , plus an model, under . The models support a 32,000-token context window. Training mixes different rates of hiding input text and predicting what is missing.
An is useful when a system needs a label or a searchable representation of text rather than a generated answer. That makes this release relevant to teams paying a conversational model to perform a narrower classification task. The lab's results a trial against the customer's own documents, especially outside English. SourcesA
Hugging Face reports faster BERT tokenization, with the changes still in review
described BERT optimizations on October 15th that accelerate common text and reduce Unicode-normalization overhead. Its tests report speedups of up to 50 times against version 0.23.
The changes were at the pull-request stage in the announcement, not a stable 1.0 release. Tokenization is the conversion of text into model inputs; making it faster does not multiply the speed of every downstream calculation. The practical test is whether preprocessing occupies enough of a deployment's for the patch to move total latency. Benchmarking only the improved component would overstate that gain. SourcesA
ColGrep-Agent tests local code search as a way to cut paid model context
LightOn's October 12th ColGrep-Agent release uses a local 2-billion-parameter model to search code and return source text with paths and line references. Across 52 search questions in 13 previously unseen repositories, the authors report 34% fewer API and 40% lower cache-aware cost than their grep-based comparison, without a statistically significant accuracy change.
The sample is small and measures the authors' search setup. Still, it offers a concrete experiment for engineering teams: spend local computation selecting relevant code, then send less material to the paid assistant. Total savings need to include local serving and indexing costs. SourcesA
OfficeTown gives agents persistent organizations to learn inside
OfficeTown, posted October 16th, builds synthetic organizations with ongoing artifacts, permission structures and feedback. Its evaluation contains 406 tasks across 23 organizations. The authors report that training Qwen3.5-9B inside the environment improves results on external professional-work benchmarks.
The useful change is continuity: actions leave consequences in a workplace that still exists when the next task starts. That tests a different skill from completing a clean, isolated prompt. These remain simulated organizations with designed tasks. Transfer to real workplaces would require measuring how agents handle conflicting stakeholders and information the simulator's designers did not anticipate. SourcesA
MemCompile separates finding memories from building an answer’s evidence
MemCompile, posted October 16th, divides long-term agent memory into locating relevant material and assembling bounded evidence for the answer. Its design retains links to source material instead of handing the answering model an unstructured accumulation of past conversation.
That separation makes a failure easier to investigate: did retrieval miss the fact, or did the evidence assembly discard it? The authors report benchmark improvements, including on . Those results do not establish reliable memory in a continuously changing production account. A useful deployment test would include corrections and revoked information, where a remembered fact can become an unsafe one to reuse. SourcesA
MultiAgentOPD assigns different learning signals to different agent roles
MultiAgentOPD, posted October 16th, trains cooperating agents with role-specific feedback while sharing model . A search role receives local teacher supervision; leadership and answer production receive signals tied to team outcomes and answer correctness. The authors evaluate the approach across seven datasets.
The problem is credit assignment. A correct final answer can conceal an unhelpful search step, while a useful retrieval can still lead to a wrong answer downstream. Separating those signals is a plausible way to improve training efficiency. The reported comparisons support the chosen benchmark setup, not a claim that additional agents automatically improve a deployed system. SourcesA
RECAP tests when an agent should forget an earlier failure
RECAP, posted October 16th, compares keeping, compressing and removing failure history from agents at matched states. The study finds that the same history can help earlier in a task and hinder later continuation. Its proposed policy uses risk constraints to decide when to edit that context.
One Hunyuan experiment preserves 91.89% continuation success with a 2.13% reduction in prompt cost. That is a modest measured saving, not a demonstration that agents should routinely erase mistakes. The stronger research contribution is a controlled way to ask whether a particular memory still helps, instead of assuming all accumulated experience belongs in every prompt. SourcesA
CoWorld lets visual agents revise their tools while keeping model weights fixed
CoWorld, posted October 15th, uses a video model to generate visual operations for spatial reasoning. The system can revise tools, skills and code while leaving the underlying model weights fixed. The authors report improvements across seven benchmarks.
This offers a route to adaptation without retraining every component. Its central constraint is evidential: an imagined visual operation is generated material, not a fresh observation of the physical world. A system that uses simulation to choose a useful next step still needs to distinguish that simulation from something its camera or instruments actually measured. SourcesA
A vision-language study separates seeing errors from reasoning errors
An October 15th paper studies four vision-language models across 17 benchmarks to distinguish failures in extracting visual information from failures in reasoning about it. The authors find their reinforcement-learning setup favors reasoning improvements and propose attention-based distillation to address perception, testing that method at two model sizes across seven benchmarks.
The distinction changes how a team should debug a wrong visual answer. A longer explanation cannot recover a number the model misread in the image. The study gives evidence for that separation in its tested systems; it does not show that every visual benchmark or model has the same bottleneck. SourcesA
More reasoning can make models interpret ordinary language too literally
A study posted October 15th varies reasoning effort while holding model weights fixed and finds changes in pragmatic language interpretation. Extra reasoning can weaken everyday implications, such as understanding “some” as suggesting “not all,” even while improving interpretation of indirect answers.
That is a narrower claim than saying thinking makes models worse. Different language tasks respond differently to additional computation. The authors explore early truncation as a remedy, but a deployment would need to test the trade-off on its own conversations. An assistant that reads every sentence with courtroom literalness can miss what a user was plainly asking it to do. SourcesA
CoEvolve gets simulated humanoids to coordinate a shared load
CoEvolve, posted October 16th, has two simulated Unitree G1 carry a payload through a corridor. It develops coordination code over fixed motor skills, checking transitions and revising failed attempts from checkpoints. Eight of 15 runs succeed.
The result is about coordination in simulation. It does not demonstrate real robots moving a load safely through a workplace. Contact remains a limitation, which matters because cooperating bodies can amplify each other's errors. The next useful evidence would show how the method responds when grip, floor friction or the payload departs from the simulator's assumptions. SourcesA
Disney agrees to hire most of Character.AI's workforce
Disney has agreed to make job offers to at least 60 Character.AI employees, described in the bankruptcy filing as most of the workforce, at an estimated $60 million in annual pay, with a further $30 million retention pool that brings the three-year commitment to about $180 million. The filing says the vast majority of those offered have accepted. Some outlets count 66 employees; the filing's floor is 60.
The hiring sits alongside, not inside, Disney's $20 million , which buys a nonexclusive IP license and certain legal claims. The consumer app, the chats, the user data and the finished models stay with the estate, and rival bids reportedly must reach at least $21.7 million, including a $700,000 break-up fee. The talent is now worth nine times the floor price of the assets. Whoever wants the running platform must outbid Disney for the IP and then staff it without the people who built it. SourcesBC
Chip stocks shed $1.6 trillion in six sessions while Huang says order books keep growing
US semiconductor stocks lost roughly $1.64 trillion in market value over six sessions through Thursday, per Reuters, with Nvidia down 3.2% on Wednesday alone. Yields on some CoreWeave bonds have approached 16% against roughly $50 billion of debt, and the company's shares are down about 30% this month. Jensen Huang spent the week arguing the other side: new orders arrive "almost weekly and daily," customers forecast demand a year or two ahead, and Nvidia allocates them less capacity than they want. He put frontier labs at 25% to 30% of Nvidia's computing activity and told Bloomberg CoreWeave needs no additional support from Nvidia.
Monday opened with a partial reprieve: South Korean equities closed up 1.8% and Nasdaq futures rebounded. The dispute underneath has not moved. The bond market is pricing credit risk while the vendor at the center reports accelerating demand, and both can be right for a while. SourcesBC
SemiAnalysis forecasts 50 gigawatts of new data centers in 2027
SemiAnalysis projects about 50 of new global data-center capacity in 2027, which it says equals the entire cumulative market through 2024 built in a single year. More than 24 gigawatts across roughly 40 campuses are already in advanced construction for 2027 delivery, most past the halfway mark with cooling and electrical equipment installed. The firm counts binding, non-cancellable leases signed this year at just under $1 trillion in one passage and over a trillion in another; the exact figure is contested inside the report itself, so treat it as approximately $1 trillion.
Brad Gerstner has already called an earlier 43-gigawatt version of the forecast too aggressive. The construction numbers are the checkable part: 40 named campuses either energize in 2027 or they do not, and the week's market selloff is a bet that some of them will not find funded tenants. SourcesA
doubleAI launches a compute-cost model after raising $220 million
Tel Aviv's doubleAI released its debut code-execution model, aimed at cutting and compute costs, the Wall Street Journal reported Monday. The company has raised $220 million in total from Lightspeed, Bessemer, Nvidia and Dell, and spent about half of it on 1,000 Nvidia B300 chips. Its earlier WarpSpeed system rewrote every in Nvidia's cuGraph library and reported an average 3.6x speedup over expert-written code, published as the doubleGraph repository.
A model that optimizes the code that runs models is the kind of recursion worth auditing: the 3.6x figure is the company's own benchmark, measured on three GPU types. Nvidia investing in a company whose product reduces demand for Nvidia hardware is less strange than it looks. Cheaper compute per task has so far produced more tasks, not fewer chips. SourcesBB
OpenAI confirms a humanoid robot with a demo targeted for 2027
Sam Altman told Fortune he expects "a compelling demo prototype by 2027," adding that "robots on the streets will take a few more years." It firms up his September podcast statement that OpenAI "will definitely do a humanoid." Reports count 27 open robotics roles spanning motion control, manufacturing and .
The commitment is softer than the headlines: a demo target, no hardware shown, no production plan, no price. Altman says the intelligence matters more than the body. The intelligence is also where OpenAI can claim an edge over the hardware companies years ahead of it. The companies OpenAI once backed, including 1X and Figure, now get their former investor as a competitor. SourcesC
Meta blames a Claude-linked tool for suspending a voter-turnout coalition
Meta suspended the No Kings Coalition's Instagram account for about 90 minutes Saturday, during a weekend of roughly 3,000 early-voting events ahead of the November 3rd midterms. Meta first called it an error triggered by high activity rates; on Sunday it said a tool linked to Anthropic's Claude had read the account's profiles, posts and hashtags at about four times normal speed in short bursts, which its systems treated as bot traffic. Administrators stayed locked out after the page returned, with full access restored at 5:33pm ET.
Meta has not said which tool it was, who ran it, or what it was doing. The mechanics deserve attention regardless: a third party's automated reading triggered enforcement against the account being read. Any account can apparently be suspended by someone else's scraper, which is a denial-of-service primitive hiding inside bot defense. SourcesB
Stripe's Link data shows agent purchases growing 38-fold in a month
Stripe shared Link checkout data with Upstarts showing agent-initiated transactions grew 38x and payment volume 20x in September. About 70 agents were signed on when the data was pulled; roughly 200 use Link now. Named buyers include Instinct, Muse, Grok Bot and Underdog. Anthropic and OpenAI use Link but not yet for .
The growth is real and the base is tiny: 38x from a small starting point, measured by the vendor selling the rails, shared exclusively with one outlet. Against Link's 300 million users, agents round to zero. What the number does establish is direction at the payments layer, which is harder to fake than app-store rankings. Saturday's reporting had banks blocking agents at the front door; Stripe is counting the ones already inside. SourcesB
Opaque Systems winds down as its team joins Anthropic
Opaque Systems is winding down operations as its team joins Anthropic, per the company's site, following the agreement the two announced last Tuesday. Terms remain undisclosed. Opaque raised a $24 million Series B at a $300 million valuation earlier this year, grew out of UC Berkeley's RISELab, and built for AI alongside open-source work including the Agent Governance Toolkit, TRACE, Confidential and a Weight Custody Manifest.
What happens to existing customers and the open-source projects is still unanswered, which is the recurring cost of the pattern: the technology was sold as infrastructure, and infrastructure that follows its team out the door was never infrastructure. SourcesB
US cybersecurity postings fall 26% while AI security roles grow 268%
US cybersecurity job postings fell to 235,786 from January through September, down 26.4% from 320,210 a year earlier and a four-year low, per CyberSN's annual report on JobSpikr data. Inside the decline, AI Security Engineer postings rose 268.3% and Cyber Data Scientist 203.5%, while Cyber Insider Threat Analyst fell 55.6%, Cyber Risk Analyst 51.5% and Penetration Tester 45.9%. About 15% of all postings mention AI, a share that peaked at 20% in September.
The report describes itself as a what-not-why document, and postings measure hiring demand rather than employment. Two readings fit the same data: security teams absorbing AI work into fewer, differently titled roles, or a budget contraction wearing an AI costume. The pentester number is the one to watch either way, since offensive testing is the work AI tools claim most directly. SourcesB
AI-made midterm ads pass $137 million, mostly unlabeled
Campaigns and PACs have spent more than $137 million this cycle on ads that are AI-generated or AI-enhanced, according to Wesleyan Media Project and AdImpact data reported by the New York Times, counting 220 such ads through October 11th. Most carry no AI disclosure, and over 80% come from Republican campaigns or aligned groups, per the Times. One New York example shows a fabricated scene of Zohran Mamdani and Kathy Hochul promoting a fictional drug; the campaign calls it satire needing no label, and a state party complaint is pending.
Congress has passed no federal disclosure rule and the state deepfake laws are a patchwork, some enjoined. Much of the counted spending is mundane photo cleanup. A disclosure rule that treats cleanup and fabrication alike has so far failed to pass, and the fabricated-scene cases keep arriving faster than complaints resolve. SourcesBC
DOE scientists set agents loose on real open problems and reject all 135 papers
A 28-author benchmark called Into the Unknown, posted October 13th, had scientists at US Department of Energy national laboratories contribute unpublished follow-up directions from their own research, each with conditions a successful submission must meet. Nine frontier agents each wrote 15 follow-up papers across quantum computing, plasma physics, computational chemistry, materials science and scientific machine learning. None of the 135 submissions satisfied all conditions, and 114 failed an integrity check, most often for claiming work that was never done or reporting numbers that contradict the recorded evidence.
This is the sharpest available answer to the AI-scientist claim: on problems where the is held by the people who posed them, current agents produce artifacts shaped like research. The integrity failure rate is the headline, because a fluent paper asserting unrun experiments is worse than no paper. SourcesA
BizBench scores agents on whether the intermediate steps were legal, not just the outcome
BizBench, posted by a nine-author team from an unnamed payments company, tests long-horizon agents on 50 executable business tasks backed by 789 policy documents and 601 discoverable tools. It scores strict success, which requires valid intermediate execution, separately from goal attainment. Claude Opus 5 leads at 64.0% strict success against GPT-6 Sol's 61.3%. Qwen3.8-27B reaches 42.0% goal attainment but only 32.7% strict success, and DeepSeek-V4.1-Flash averages 191.3 per run, about 70% more than reference executions.
The gap between reaching the goal and reaching it legally is the finding. In payments, an agent that gets the right answer through an invalid step is a loss event that outcome-only evaluations record as a success. SourcesA
A 60-author study asks whether GPT-6-Astra is embodied intelligence yet, and answers not quite
SenseNova and a joint research team evaluated GPT-6-Astra across 12 embodiment benchmarks spanning tabletop manipulation, humanoids, mobile platforms and , plus eight real robot configurations. Their verdict: the model "has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence." Pairing it as a high-level reasoner over a low-level controller works but costs heavily in time and compute.
The listed limits are the useful part: generalizes poorly, physical-state verification is weak, recovery from irreversible failures fails, and trading longer execution for success breaks under physical constraints. Each names a concrete gap between benchmark competence and a robot you would leave unattended. SourcesA
A former Huawei engineer's manipulation startup raises an angel round near RMB 100 million
Hangzhou's Gewu Tongzhi closed an angel round of nearly RMB 100 million, about $14 million, to build AI for dexterous robotic manipulation, TechNode reported Monday. The founder previously worked as a Huawei engineer.
The round is small by frontier standards and typical of where Chinese robotics capital is going: hands, not humanoids. Dexterous manipulation is the unsolved half of physical AI, and an angel round this size for it says domestic investors expect the training-data and hardware loop to localize. SourcesB
Foxconn pays $727 million for half of Mitsubishi Electric Mobility
A Foxconn unit agreed to buy 50% of Mitsubishi Electric Mobility for about $727 million, with closing expected in October 2027. The deal gives the iPhone assembler a stake in an established automotive electronics supplier as it builds out contract manufacturing for software-defined vehicles.
For the AI supply chain the relevance is capacity: Foxconn is already Nvidia's largest server assembler, and its automotive push is the same bet at vehicle scale, that the value in hardware migrates to whoever integrates the compute. SourcesB
NextDC halts trading on a power delay at its Melbourne AI site
Australia's NextDC requested an ASX trading halt Monday while it assesses a power supply delay at its M3 Melbourne data center, which it says could affect FY27 guidance. This is single-source reporting from a market-news aggregator pending the company's formal announcement.
A listed operator halting its own stock over one site's grid connection is the power constraint arriving as a disclosure event rather than an engineering footnote. Nscale's Loughton site stalled on grid timing into the 2030s; M3 is the same failure mode in a different grid. SourcesC
Illinois creates a state AI Cabinet by executive order
Governor JB Pritzker signed an executive order creating an Illinois AI Cabinet to develop AI policy proposals through 2027, including guidance for state agency procurement. Illinois already requires annual third-party audits of under its earlier statute.
States keep building AI policy machinery while the December federal executive order promises to preempt them, and a cabinet that writes procurement guidance is harder to preempt than one that writes conduct rules: a state deciding what it will buy is exercising a power no federal framework removes. SourcesB
FT survey: economists expect AI growth to spread beyond the US and China
A majority of 94 economists surveyed by the Financial Times expect AI to lift productivity across the G20, not just in the US and China, while warning that weaker digital and power infrastructure could leave African economies behind. The survey published Sunday.
A G20-wide productivity expectation from a survey population that skews cautious is a real shift from a year ago, when the dominant frame was two-country concentration. The falsifier is in the infrastructure clause: diffusion requires grid and connectivity investment that most of the surveyed economists do not control for. SourcesB
AI weather models face their first record El Niño
AI weather models may struggle to forecast a potentially record El Niño because the developing conditions fall outside their training distribution, the Financial Times reported Sunday. The machine-learning forecasters that beat physics models on recent seasons have never seen the regime they are now being asked to predict.
This is the cleanest natural experiment running on generalization, with a verifiable answer arriving within months. If the AI models hold up through an unprecedented event, the training-data objection loses its best example; if they fail while physics models hold, every safety-critical deployment argument gets a new citation. SourcesB
The data-center industry starts courting schoolchildren
At least 11 companies and industry groups have hosted student field trips to US data centers since 2025, Wired reports, complete with promotional kits, as local opposition to new construction grows. The industry that spent a decade avoiding attention now runs tours for fifth graders.
File under second-order consequences of the backlash: an industry recruiting the next generation of local voters is an industry that expects siting fights to last a generation. SourcesB