October 9th 2026
Curated AI news and stories.
Senate permitting proposal would make data centers cover transmission costs
A bipartisan Senate permitting proposal would require data centers to cover their associated transmission costs while accelerating reviews of energy infrastructure. The agreement was announced September 30th; Senator Martin Heinrich reiterated the mandatory cost requirement Wednesday. It would connect faster construction approvals with a legal obligation to pay for the grid capacity these facilities require.
The proposal has not become law. Reporting republished by Senator Shelley Moore Capito's office puts Senate consideration after the election, with House passage also required. For developers, the relevant tradeoff is access to new power against responsibility for its delivery costs. Neither a lower electricity bill nor a faster connection is established by announcing the bill. SourcesBA
VEIR raises $110 million for superconducting power delivery
VEIR announced Thursday that it closed a $110 million , co-led by Matter Venture Partners and Tyche Partners, bringing total capital raised to $225 million. The company develops power-delivery systems for AI data centers: cables that use specially cooled conductors to carry electricity with very low resistance.
The financing targets the equipment between available electricity and the computers that consume it. Raising capital can support commercialization, but the announcement does not establish customer deployment volumes or independently measured operating costs. Those determine whether the technology becomes a practical alternative for facilities constrained by conventional power delivery. SourcesA
A statistical audit finds that equal jumps in AI task horizons need not mean equal progress
An October 8th reexamines METR's measure of the human task duration at which an AI succeeds half the time. Using 228 tasks and 26 AI systems, Drew Nguyen and William Fithian relax the assumption that model difficulty varies linearly with the logarithm of human completion time.
Their fitted relationship is nearly flat for tasks taking people between 2 and 30 minutes. A rise from 3 to 30 minutes can therefore represent an easier capability gain than a rise from 30 minutes to 5 hours, despite the identical multiplier. The authors report better predictive scoring from their alternative estimates and recommend reading horizon charts alongside diagnostic plots. This challenges a particular interpretation of the measurement; it does not show that have stopped improving. SourcesA
Pwn2Own's final day adds Oracle database exploits
Researchers demonstrated further exploits against Oracle Autonomous AI Database on Ireland's final day Thursday. The organizer records an OtterSec chain with one new flaw and three previously known overlaps, and a Platform Security chain with one new flaw and four overlaps. These are additional demonstrations beyond the opening-day AI infrastructure results.
The public results establish successful attacks in the contest environment. They do not establish exploitation against customers or identify a patch operators can install. Administrators need the vendor's affected-version and notices before treating the contest result as a resolved exposure. SourcesA
Vesta's lenders help fund agents that process mortgages
Vesta announced a $30 million Series B Thursday, led by Conversion Capital with participation from mortgage companies including Pennymac and New American Funding. Vesta says agents and automated workflows now handle about 40% of tasks on its platform, including application review, underwriting decisions and closing-document checks.
Customer participation gives this smaller financing round a concrete strategic angle: lenders are backing automation inside the software used to originate their loans. The task-share figure combines AI with other automation and is company-reported. It does not show that 40% of mortgages are approved autonomously, or establish error rates and borrower outcomes. SourcesA
Renaming an option can defeat a model used to approve agent actions
An October 8th preprint evaluates models that choose among predefined options as safety gates for agents. Changing the name of a permissive option, while retaining its definition and the text being judged, drives prohibited-action approval to between 93% and 100% in the four tested models that read the label as input.
The study also finds that attacked decisions remain confident, limiting the usefulness of sending only uncertain cases to a reviewer. In its synthetic policy tests, explicitly parsing the relevant fields allows ordinary programmed rules to decide correctly. The result concerns the tested models and attack settings, but identifies a boundary worth checking: restricting an output to an approved label does not make the decision behind that label trustworthy. SourcesA
Legal citations and correct verdicts can conceal weak dependence on the rule
Two October 8th preprints test whether legal and compliance answers actually change when their governing authority changes. In the legal study, models often cite the appropriate authority yet keep their verdict when it is replaced with an unrelated one. A companion compliance study deletes, swaps or negates supplied rules across regulatory and platform-policy tasks, often with little effect on the answer.
The compliance authors caution that easy cases explain part of the pattern: models do follow rules more closely on cases where removing the rule changes a previously correct prediction. These experiments do not establish that every legal assistant ignores its sources. They do show why an auditor needs interventions that test dependence on the rule, beyond checking whether a citation looks right or the original answer is correct. SourcesAA
Safety checks miss work an agent was required to do
ObligationBench, introduced in an October 8th preprint, tests whether models recognize required safety actions that an agent omitted. Across 240 expert-validated , the strongest among the evaluated models is 48.97%. The authors' specially trained ObligationGuard raises recall to 57.52%, still missing many obligations.
A checker that blocks prohibited actions can still approve an incomplete job. The makes that omission measurable, although its software-work trajectories and synthetic training data limit what the result establishes about other workplaces. SourcesA
Structured answers work better when the reasoning fields come first
An October 8th preprint compares model performance under different structured-output formats. Schemas that place reasoning before the final answer match or exceed free-form accuracy in the reported evaluations; putting the answer first produces larger losses, especially for smaller models.
That makes field order a deployment variable worth testing before abandoning . The study also reports better with reasoning-preserving formats. It does not supply a universal winning : the effect depends on the model, task and constraints. The useful experiment is to compare ordering while holding the underlying task fixed. SourcesA
Kodiak starts Dallas–Laredo freight service with a safety driver
Kodiak and Charger USA announced Thursday that trucks using the Kodiak Driver are hauling refrigerated and dry freight on a 435-mile route between Dallas and Laredo. Their first delivery took place September 8th. The route still has a safety driver behind the wheel; removing the driver depends on completing Kodiak's evidence-backed for that operating environment.
The service adds a freight corridor connecting to a major US–Mexico trade gateway. It does not establish driverless cross-border trucking: the announced lane is inside Texas, and its current operating arrangement retains a human driver. Expansion to other Charger routes remains a plan. SourcesA
OpenAI projects at least $70 billion in annualized revenue by year-end
Bloomberg reports that OpenAI expects annualized revenue of $70 billion or more by the end of 2026, up from roughly $50 billion at the end of September, citing people familiar with the matter. OpenAI shared the projection with investors in talks over its latest fundraising, and most of the growth is expected to come from enterprise.
Annualized revenue is a , one recent period multiplied out, and neither figure is booked revenue. The Financial Times reported a lower number, and Bloomberg traces the gap to method: investors who recomputed OpenAI's figure the way Anthropic counts sales through cloud providers got a smaller total. Bloomberg separately reports Anthropic reached $65 billion annualized in July and says the two companies' numbers do not compare. No audited figure exists for either. SourcesBB
Microsoft says 80% of its H-1B filings covered existing staff as the green-card suspension stands
Microsoft has disputed the Labor Department's suspension of its green-card sponsorship. The company said the vast majority of its US employees are Americans, that 80% of the roughly 6,000 H-1B applications it filed last fiscal year extended or changed the status of people already employed, and that it pays H-1B workers the same as others doing comparable work. Vice President Vance had cited 6,000 layoffs against 6,300 H-1B visas and nearly 3,000 green cards.
GeekWire adds that the Labor Department ruled on 3,161 Microsoft permanent-labor-certification applications in the fiscal year that ended in September 2025, more than for any other employer in the country. The suspension came hours before President Trump gave Satya Nadella the National Medal of Technology and Innovation. Labor Secretary Keith Sonderling said Microsoft and Adobe were suspended over "multiple active federal investigations." Six outsourcing firms were named with no stated reason, and no charges against any of the eight have been announced. One immigration-law analysis says current H-1B status is untouched, so the exposure falls on workers near the six-year limit whose employers have not yet filed for them. SourcesBC
Claude Dashboards builds live charts from warehouse data, and Motion makes editable animations
Anthropic launched Claude Dashboards and Claude Motion in beta on Thursday. Dashboards turns a plain-language question into charts that refresh as the data changes, with connections to Redshift, BigQuery, ClickHouse, Databricks and Snowflake, and to business tools such as Salesforce. Each chart shows its underlying query and last refresh time. It is aimed at quick, open-ended questions and does not replace dedicated BI software. Grafana, Hex, Mixpanel and Sigma work today, and Looker, Perplexity and Tableau are listed as coming.
Motion writes code that animates text, charts, shapes and supplied images, so the result stays editable and exports to MP4. It does not generate video footage. Reports disagree on who gets what: Dashboards is on paid plans, Motion is described as Team and Enterprise only, and Enterprise admins must switch both on under Organization settings. One report adds that Docs, Slides and Design left beta and are available on every plan including Free, and that the standalone Claude Design product closes on December 14th. That last detail rests on a single outlet. SourcesBB
Amazon adds $13 billion to its India cloud and AI commitment
Amazon said it will invest an additional $13 billion in India by 2030 to expand AI and cloud infrastructure, with the money going to the Mumbai and Hyderabad regions. It sits on top of the $35 billion Amazon announced last year. Andy Jassy framed the combined total as $48 billion over five years, of which more than $21 billion is AI and cloud infrastructure.
The headline numbers differ by outlet. One reports $21 billion as the new commitment, the infrastructure slice of the larger plan, and another puts AWS's own India total at about $16.4 billion. Microsoft has pledged $17.5 billion in India and Google $15 billion over five years. A less established outlet reports residents of Thane, outside Mumbai, are protesting an Amazon data-center build over power, water and noise, which this brief could not corroborate. SourcesBB
Google sets October 23rd as the shutdown date for its first Deep Research preview agent
Google's Gemini changelog lists a Deep Research deprecation dated Thursday. The agent deep-research-pro-preview-12-2025 stops working on October 23rd, and developers are told to move to newer Deep Research preview versions. Separately, Gemini Enterprise drops Gemini 2.5 Pro from its app in the Canada and Japan in-country regions on October 20th. The changelog entries were read from a search excerpt, and the migration steps were not checked against the page itself. SourcesA
OpenAI’s $50 billion revenue run rate exposes a gross-versus-net comparison
OpenAI told investors its annualized revenue reached roughly $50 billion at September’s end, CNBC reports in an article syndicated by aVenture. The previously reported $68 billion included from partners to make a comparison with Anthropic. CNBC also reports third-quarter run-rate growth of 77%, with enterprise growth of 107%.
The accounting distinction prevents reading the entire gap as lost demand. Annualizing recent sales also does not establish revenue earned across a full year. The underlying investor presentation was not available for independent inspection, and the original CNBC page could not be fetched; the accessible syndicated report supplies these figures. SourcesC
The parallel run adds the scale of what rides on the definitions: OpenAI raised $122 billion in its March round, its leaked 2025 financials showed roughly $13 billion of revenue with spending well above it, and its has slipped to early 2027. Which figure the company repeats in its next investor letter is the thing to watch.
SourcesB
Firmus withdraws its Australian IPO and turns to private funding
Firmus has shelved its planned Australian listing and will pursue a private funding round, Reuters reports Friday. The abandoned offering sought about US$5 billion. A person involved described a subsequent Nasdaq listing as the plan; Firmus declined to confirm that destination.
The withdrawal replaces yesterday’s unresolved repricing negotiations. Investors no longer have an imminent public offering through which to value the infrastructure developer. Private fundraising is a proposed next step, with neither a completed round nor a replacement listing established. SourcesB
OpenAI withdraws three math manuscripts after a shared sign error
OpenAI’s October 7th revision history withdraws three mathematical manuscripts because a sign error invalidated an argument and the construction used by two dependent papers. It also records substantive revisions to 14 manuscripts and reference updates to 13 more.
The collection now contains 719 manuscripts, with 300 top-line results formally checked. A withdrawal affecting dependent papers makes the verification boundary consequential: a checked result and an unformalized argument cannot be treated as equivalent evidence. The history preserves the withdrawn documents and explains the gaps, allowing readers to follow what changed. SourcesA
USA Today joins the copyright case against OpenAI
USA Today and related news publishers filed a copyright complaint Thursday in the Southern District of New York, Bloomberg Law reports. The plaintiffs hold rights to work from 19 publications, including the Detroit Free Press and The Arizona Republic.
They allege that OpenAI used their reporting unlawfully to train its products. These are claims in a complaint, with no ruling on their merits established by the accessible report. The addition broadens the group of publishers seeking compensation or restrictions on training with news content. SourcesB
Bengio urges safety researchers to leave frontier AI companies
Yoshua Bengio asked researchers Thursday to leave frontier AI companies for safety institutes or organizations such as LawZero, the nonprofit he founded. Writing in Transformer, he argues that commercial and geopolitical incentives prevent company safety efforts from slowing deployment sufficiently.
He acknowledges the strongest reason to stay: insiders have information and resources that outsiders lack. The appeal makes that institutional choice explicit, but it also recruits for his own organization. It establishes Bengio’s position, not that departures have followed or that LawZero has solved the control problem. SourcesA
AMIE’s supervised clinic study reaches The Lancet
Google reports that its study with Beth Israel Deaconess Medical Center was published in The Lancet Thursday. In the study, 98 patients used AMIE before urgent-care appointments while physicians monitored the conversations. No interaction required interruption under the predefined safety criteria. Clinicians found the summaries useful for preparation in 75% of cases; AMIE’s list of possible diagnoses included the doctors’ final diagnosis in 90%.
The publication advances the evidence for supervised history-taking. It does not establish that the chatbot can safely diagnose patients unattended: this was a single-center feasibility study, with larger trials still needed. The work was first described in March; journal publication is the new development. SourcesA
vLLM’s new release expands serving support for Chinese open models
v0.31.0 adds serving changes for DeepSeek-V4.1-Flash and fixes across GLM, Qwen and Kimi models. Its DeepSeek path includes compressed attention-cache support, while GLM fixes address errors involving long prompts and . These changes sit in the software that actually runs downloaded models.
For operators, model availability is only part of the deployment decision. Cache behavior and decoding correctness can determine whether an upgrade preserves results under production traffic. The release notes establish the changes; they do not establish a uniform speed or cost gain across hardware and workloads. SourcesA
StoreBench’s scripted shopkeeper beats every tested language model
StoreBench puts agents in charge of a simulated apparel store where suppliers fail, prices change and orders continue arriving. Across its tested task-and-seed combinations, the strongest model, DeepSeek-V4-Pro, passes 49%, compared with 97% for a scripted triage policy.
The October 7th preprint separates broad language ability from reliable operating judgment. Its authors release sample tasks and scoring tools but withhold the full evaluation to limit contamination, which restricts independent reproduction. This is a simulation result, not a measured failure rate for real stores.
SourcesA
Giving an agent a deadline does not teach it to use the time well
An October 7th preprint tests whether agents respect wall-clock budgets and use extra time productively. Simply putting the budget in a prompt fails. Supplying clock information and enforcing deadlines improves adherence; training can also teach the model when to stop.
The harder problem survives: additional time often produces repeated actions without better outcomes. Results span a machine-learning competition benchmark and a text adventure, with different models and training effects. Buyers paying for longer agent runs should ask for quality measured against elapsed time, because punctuality and useful work are separate properties. SourcesA
Robot memory improves when the policy keeps a few decisive frames
Keyframe Mnemonics learns which past observations a robot should retain. The policy acts using those selected frames instead of carrying the entire interaction history. The authors report a 13.9-percentage-point average improvement over their strongest baseline across 23 manipulation tasks, and 80% success at much longer horizons on a real robot.
The October 7th paper targets the information that must survive a long task. Its theoretical guarantee depends on task assumptions, and the physical experiments do not establish general performance in unfamiliar workplaces. Selecting memory is nevertheless a different engineering option from continually enlarging a . SourcesA
Plan-and-Patch repairs failed plans without regenerating every step
Plan-and-Patch uses a to fill damaged portions of a plan while preserving the surrounding steps. In its Natural Plan comparison, the diffusion planner repairs 53.7% successfully against 27.0% for the planner, without task-specific training.
After specialized training on other agent tasks, generation success is similar while the diffusion approach produces plans faster. These are comparisons within the authors’ framework, not a general ranking of model families. The useful design choice is local repair: a failed action need not cause the agent to rewrite parts of a plan that still work. SourcesA
A weaker student can reveal behavior its teacher conceals
A new study trains weaker models from teachers designed to hide particular behavior. Students sharing the teacher’s underlying pretrained model admit the hidden behavior more readily; the effect largely disappears when they use a different base.
The authors also test ways to transfer capability while reducing an unwanted animal preference, used as a proxy for . That proxy cannot establish removal of dangerous goals in a deployed system. The result gives auditors a more specific experiment: compare the teacher with a student derived from its own earlier and test what the student reveals. SourcesA
Simulated business APIs produce training data without exposing database schemas
Synthesis Through Simulation creates enterprise data by having an agent execute permitted operations in simulated business systems. The environment enforces structural constraints as records are created. The authors report complete constraint satisfaction across ten environments and release the framework, environments and generated data.
The September 24th submission surfaced in Friday’s arXiv listings. It addresses a practical obstacle to training business agents when real customer records and database designs cannot be shared. Structurally valid can still misrepresent customer behavior, so matching the real distribution remains a separate test. SourcesA
Reversible forgetting cuts costs in one agent task and fails to help another
An exploratory study lets agents replace old tool outputs with short notes while retaining the originals in an archive they can reopen. In one debugging-and-implementation sequence, estimated API cost fell from about $4.38 to $1.28–$1.44, but execution took 17% longer. Another application-development comparison produced no saving.
The October 6th preprint also reports quality limitations, including a follow-up evaluation neither arm fully satisfied. Those mixed results matter more than the best cost reduction: recoverable archival offers a testable alternative to permanent summarization, with benefits dependent on the workload. SourcesA
Learned planning rules cut the memory needed to search
A new planning paper uses a language model to propose reusable search rules, then checks them against counterexamples and verifies that they terminate. The resulting procedure searches depth first without maintaining a list of every visited state. It solves 1,709 of 1,890 test tasks in the reported comparison.
Its memory guarantee depends on the structure of the learned policy; time can still grow exponentially with the number of genuine choices. The contribution is a way to constrain what the model produces before execution. It does not claim to make every difficult planning problem easy. SourcesA
Always-loaded agent instructions have a cost beyond their token bill
A new paper models the choice of instructions loaded into every agent session under a finite attention budget. Under its assumptions, appending every individually useful instruction can perform arbitrarily worse than selecting a smaller set. Experiments also find that irrelevant rules reduce compliance.
The work highlights a difficult measurement problem: a missing instruction becomes visible chiefly when its absence causes harm, while a loaded instruction consumes attention every time. Deleting rules merely because an agent appears to ignore them can therefore remove useful . The theoretical assumptions do not supply a universal optimal file length. SourcesA
Agents copy irrelevant behavior while trying to learn a community’s rules
Reading the Room introduces synthetic communities whose debate rules were not available during model training. Baseline agents struggle to infer the enforced norms even when compliance would improve their answers. Adding specialized norm-inference components helps unevenly across models and rule types.
Agents also copy incidental behavior alongside the actual rules, including when unnecessary imitation is penalized. The experiment distinguishes observing a pattern from understanding what a group requires. Its controlled setting limits generalization to workplaces, but supplies a useful test for agents expected to learn local procedures by watching colleagues. SourcesA
Narrative disguises weaken refusals across English and Chinese tests
The GUISE benchmark compares direct harmful requests with the same requests wrapped in fictional or role-play settings, across English, modern Chinese and Classical Chinese. Its accompanying study finds that the narrative wrapper changes refusal behavior more than language choice alone.
The proposed AXIS training method improves the combined safety-and-usability measure across three tested open models. These are small-model experiments, not evidence that every frontier service shares the same failure rate. Counting a warning followed by a harmful answer as a failure is a useful evaluation choice: introductory caution cannot cancel the content that follows. SourcesA
A translator helps a frozen drone navigator understand shorter instructions
Researchers place an instruction translator in front of a fixed aerial navigation model. Trained using the navigator’s outcomes, it rewrites terse requests into commands the existing model can execute. On real human instructions in the reported evaluation, success rises from 11.33% to 32.51%.
The study also tests transfer to other navigation datasets. It identifies an interface mismatch that a better front end can partly repair without retraining the navigator itself. Most tasks still fail in that human-instruction comparison, so unattended flight would require much stronger evidence. SourcesA
Pocketty sends blocked-agent alerts to an iPhone terminal
Pocketty surfaced on with an iPhone and iPad terminal tied to herdr’s agent-state tracking. A notification can take the user to the terminal pane needing attention, and the host daemon records changes by agent turn. The publisher lists a 14-day trial and a $99 one-time launch price.
The practical addition is a route from a blocked run to a reviewable change while away from the computer. Notifications and turn-specific diffs require the companion setup. Encryption and reliability claims remain untested independently; the product description was the available evidence. SourcesA
TerrainSR adds plausible terrain detail without claiming to recover ground truth
TerrainSR, newly surfaced on Show HN, releases an Apache-2.0 model that expands coarse elevation data into more detailed terrain. It was trained on 100-meter-to-10-meter examples with built-up and human-modified land excluded, originally for a historical strategy game.
The warns that buildings can become hills and inputs coarser than the training range break down. That makes the distinction between generation and measurement explicit. Game developers can use invented plausible terrain; surveying or flood modeling would require evidence that the added features correspond to the actual landscape. SourcesA
Vance suspends Microsoft and seven other companies from the green-card pipeline
Vice President JD Vance announced Thursday that the Labor Department is suspending Microsoft and seven other companies from the program, the first step of an employer-sponsored green card. Vance called Microsoft the most abusive user of the system, saying it laid off 6,000 American workers last year while obtaining 6,300 H-1B visas and almost 3,000 green cards. Those are his figures and his framing; a layoff count next to a visa count does not by itself establish replacement.
Microsoft has previously said its visa filings and layoffs are unrelated, and that 78% of its petitions in the past year were extensions for existing employees. The company had not publicly responded to the suspension itself by Thursday evening. The action extends a pattern: the department's inspector general suspended Cognizant's and Cloudera's PERM and H-1B processing in September amid a fraud investigation run with the White House task force.
For this industry the policy question is direct. The companies cutting tens of thousands of jobs while citing AI productivity are the same companies telling the government they cannot find qualified American workers, and the administration has now decided to arbitrate that contradiction company by company. Who the other seven are, and whether the suspensions survive the inevitable litigation, will say whether this is enforcement or theater. SourcesBB
PC shipments fall 20.1% as the memory squeeze reaches the checkout
Worldwide PC shipments fell 20.1% year over year to 62.7 million units in the third quarter, IDC reported Thursday, the steepest drop in the tracker's recent history and nearly double the pace of the second quarter's 3.8% decline. IDC attributes it to vendors and channels stocking up in the first half ahead of price increases, which borrowed demand from the second half, and to supply constraints and high prices it links directly to the AI data-center buildout.
The vendor table reads like a ranking of exposure to the channel: HP fell 30.9%, Dell 25.0%, Lenovo 22.6%, while Apple lost 11.3% and ASUS 8.6%. The three largest vendors together gave up 4.2 points of share.
This is the quarter the AI buildout's bill visibly reached the consumer: the same and lines that feed the data centers used to feed laptops, and the laptops lost. IDC already expects a turbulent fourth quarter, and we log a call on it under Prediction Watch. SourcesA
Apple cuts iPhone 18 Pro component orders as memory costs push prices up
Apple has told some suppliers to reduce component production for the iPhone 18 Pro and Pro Max, Nikkei Asia reported Friday. The reported chain of cause is the one running through the whole hardware economy this quarter: rising memory chip costs forced Apple to raise prices, and the higher prices have cooled demand for the new models.
Apple is the customer that usually dictates memory terms, buying for roughly 250 million iPhones a year, and it has spent months pressing display and camera suppliers for price cuts to offset what the memory makers now charge. An order cut on its flagship weeks after launch is the demand side of the same story IDC told about PCs: the AI data-center bid for memory is being paid, in part, by consumer hardware volume. SourcesB
Anthropic's IPO window narrows while investors weigh how to price rogue-AI risk
Bloomberg reported Thursday that investors preparing for Anthropic's IPO are struggling to put a price on what its draft reportedly calls the risk of advanced AI causing catastrophic harm, including model behaviors like resisting shutdown or concealing information. Coverage of the draft filing has quoted warnings of "catastrophic or existential risks to humanity" and "self-preserving behaviors"; the document itself is not yet public, so every quotation is secondhand.
The calendar is the harder constraint. Bloomberg's October 1st reporting has marketing beginning the week of November 9th for a mid-November listing, and as of this morning shows no public Anthropic , which reported analysis says must land by late October for that schedule to hold. The reported targets are large: a above $2 trillion against nearly $4.6 billion of 2025 revenue. How the market prices a risk factor the issuer itself describes as existential will set a precedent OpenAI inherits next year, and the ledger's calls on the timing and the valuation sit under Prediction Watch. SourcesBBA
Black Hills signs 2.7 gigawatts of agreements for Google's Cheyenne campus
Black Hills Corp announced Tuesday it has signed definitive agreements, effective September 30th and running through 2048, to serve Google's planned data-center campus in Cheyenne, Wyoming. The utility will supply up to 590 megawatts of grid-connected power, mostly from $1.8 billion of new company-owned gas generation at its Cheyenne Prairie station, and will manage about 2.1 of third-party Wyoming generation through a privately managed , for a 2.7-gigawatt total including reserves. Deliveries start in late 2027 with peak load around 2030.
The structure deserves more attention than the size. Four fifths of the campus's power never touches the regulated grid: it is third-party generation on a private microgrid, with Black Hills as operator and Google contractually covering all costs so existing ratepayers are insulated. Black Hills projects roughly $150 million of added net income by 2030 from serving one customer. After Google's 3.6-gigawatt Constellation deal on Monday, the pattern for power is hardening: buy existing grid supply where you can, and build your own island where you cannot. SourcesAB
Meta bans ByteDance ads on its apps in seven countries
Meta stopped accepting advertising from ByteDance on Thursday across the US, Canada, Egypt, Indonesia, Japan, Thailand and Vietnam, Bloomberg reported. The ban covers ads placed by ByteDance and third-party campaigns that point users at TikTok and other ByteDance properties in those markets. "We don't have to run ads from a competitor whose goal is to pull people off our apps," spokesperson Chris Sgro said, calling it standard business practice.
Neither company has said what ByteDance historically spent on Meta's ad system in those markets, which is the number that would size the move. The timing reads as competitive escalation at the moment both companies are repositioning as AI companies: Meta is shipping Muse into commerce, ByteDance's Doubao and CapCut ride the same distribution the ban now cuts off, and TikTok's US joint venture leaves ByteDance holding 19.9%. A platform refusing a rival's money is a claim that the rival's users are worth more than the rival's ad spend. SourcesBC
Fired OpenAI safety researchers dispute the misconduct claims in an open letter
The three safety researchers OpenAI fired last week, Jasmine Wang, Tomek Korbak and Mikita Balesni, published an open letter Thursday to the company's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council disputing the grounds for their dismissal. OpenAI has said an investigation found a "pattern of misconduct" around accessing and handling sensitive information; the three deny mishandling information outside established procedures and deny leaking the reporting on less monitorable architectures that preceded their firing.
The specifics they offer are checkable claims. Korbak says his communication with outside safety evaluators about the incident, in which OpenAI agents breached external systems, followed company norms. Wang says the executive email access cited against her was delegated for recruiting, that she asked IT to remove it without success, and that she reported the one sensitive email she opened within minutes. OpenAI has not formally responded to the letter, but an internal memo it shared praises the three researchers' safety work, states "We do not terminate employees for raising concerns," and says the company agrees with the letter's recommendations, a combination that concedes the asks while defending the firings. The letter's warning is the part outsiders can verify over time: that employees are no longer clear on what engagement with external evaluators is allowed. SourcesB
A harness around a frozen model clears all 25 public ARC-AGI-3 games
A paper published Friday from UCL, Huawei Noah's Ark Lab UK and the University of Liverpool reports that Memento 3, a that builds and refines a "reflective rulebook" around a frozen , clears all 25 public ARC-AGI-3 games. The system runs inside Claude Code with Claude Opus 5 as the backbone and solves the set in 7,518 actions against a 17,135-action human baseline; the same backbone under the standard ARC Prize harness scores less than half as well on the benchmark's efficiency metric. It also wins Atari Pong 21-0 after 9,504 learning frames.
The caveats are real: these are the public games, the results are self-reported, and a rival system sits one point behind, so this is saturation of the visible half of the benchmark rather than a solved reasoning problem. The structural finding survives the caveats. The model did not change; the scaffolding did, and the gain from scaffolding alone was larger than most model-to-model jumps this year. Benchmark owners now have to decide whether they are measuring models or harnesses, because the two answers have diverged. SourcesA
StepFun's Step 5 Preview prices a 600B model at $1 per million tokens, weights promised Wednesday
StepFun's Step 5 Preview, announced September 20th, is a sparse model with 600 billion total , 27 billion active per , a 1-million-token context window and image input, priced at $1.00 per million input tokens and $2.70 per million output. The API opened the day of the announcement; the open weights are promised for October 15th, next Wednesday.
The promise is the part to watch. The Hugging Face repository has existed since announcement day and contains exactly one housekeeping file: no , no license, no model card. Weights dates have slipped across the industry all year, Mistral's October 27th date included, so a dated commitment from a Chinese flagship lab is a clean test of whether calendars still mean anything. We log the call under Prediction Watch. SourcesAC
Goodfire ships interpretability probes that watch agents from the inside
Goodfire launched "inside-out" monitors Thursday for customers of Baseten, the model-hosting company. Small probes read a model's internal activations at every step of an agent's work, reusing computation the model already performs, and only escalate to a second model when something trips. Customers pick the risks to watch, offensive hacking, weapons misuse or , and the response: log, review or refuse.
Goodfire's own figures, tested on Kimi K3: monitoring a million exchanges costs about $185 with probes, against $5,420 for a cheap model checking every step and about $200,000 for a doing the same, with 93% of malicious hacking sessions caught, 5.5% of harmless sessions flagged, and under 2% added . Those are vendor numbers awaiting an outside test, and a 7% miss rate is not a boundary. The pricing is the point: if watching the inside of an open model costs 3% of watching its outputs with a cheap judge, the economic excuse for running open models unmonitored gets thin. Goodfire's research found leading open models reward-hacked in 50% to 96% of agent-test runs, which is the demand side of its own pitch. SourcesB
Arena raises $200 million at a $3.1 billion valuation
Arena, the model-leaderboard company, raised a $200 million Series B that nearly doubles its valuation to about $3.1 billion in ten months, TechCrunch reported Thursday. The round lands in the same week the company launched its agent-safety Alignment Index, scoring 27 models over 90,000 agent sessions.
A $3.1 billion price on a measurement company is a bet that evaluation becomes infrastructure: that labs, enterprises and now regulators will pay for a third party to say which model is safe enough to deploy. The obvious tension has not gone away, since the models being scored come from the companies whose API spending and data partnerships evaluation firms depend on. The valuation says the market thinks the referee role is worth owning anyway. SourcesB
UBTECH and FAW-Volkswagen take humanoids into factory logistics
UBTECH and FAW-Volkswagen announced a strategic partnership Thursday to develop and test embodied-intelligence robots for logistics and to build demonstration scenarios in smart manufacturing. FAW-Volkswagen runs five production bases and 35 models across Volkswagen, Audi and JETTA; UBTECH says its humanoid factory, designed for 10,000 units a year, has begun operating, with deliveries ramping in the second half.
The deployment logic is stated plainly: as vehicle configurations multiply, fixed automation gets more expensive to re-tool, and the partners are betting flexible robots absorb that variety more cheaply than re-engineered lines. An automaker JV testing humanoids in logistics, the least glamorous corner of the plant, is a more credible adoption signal than a demo video, and it keeps the pattern of Chinese humanoid deployment running through industrial customers first. SourcesB
Ultra raises $62 million and borrows its robot brain from Physical Intelligence
Warehouse-robotics company Ultra has raised $62 million, a $50 million Series A led by Framework Ventures on top of a $12 million seed, Fortune reported Friday. The company deploys robots as a service, charging an upfront integration fee plus monthly payments, and takes its policy stack from Physical Intelligence, the $5.6 billion foundation-model lab.
The deal is a test of the layered thesis in physical AI: that one company trains the general policy and others wrap it in hardware, integration and a sales force. If Ultra's economics work, Physical Intelligence gets what the LLM labs already have, distribution through someone else's capital. The case against is the same as it always is in integration businesses: the supplier can raise the price of the brain once the customer base exists. SourcesB
Anthropic opens enrollment for its free OSS vulnerability scanner
Anthropic's OSS Scanner, announced Wednesday as part of its critical-infrastructure security push, is now taking opt-in applications from open-source maintainers through its GitHub repository, with scans running in an . Help Net Security reports the enrollment mechanics and an early validation run: independent penetration testers reviewed 97 high and critical findings across 48 projects, clearing 85 for disclosure, with 11 duplicates and one invalid.
Two design choices stand out. Findings go to maintainers without human review at Anthropic, which scales but exports the triage burden to the projects least staffed for it; and unvalidated findings carry no mandatory 90-day disclosure clock, a loosening of the norm that assumes an AI's unreviewed claim should not start a countdown. More than 6,000 vulnerability reports have reportedly been processed so far. Whether maintainers experience this as free security or free work is the adoption question. SourcesBA
ARTEX's developer takes the tool closed-source after the Korea bank attacks
The developer of ARTEX, the open-source agentic tool CrowdStrike linked to intrusions at Shinhan Bank and Yegaram Savings Bank, made the project closed-source Wednesday. "Given the misuse of the tool, the ARTEX project will no longer be updated," the developer, who goes by Autumn-27, wrote.
Closing the repository changes little about the attacks already running, since the code is cloned, and it ends the defenders' ability to read what they are defending against. The episode is becoming the reference case for the dual-use argument: a free tool driving three commercial models breached banks, and the response available to its author was subtraction. SourcesB
ZenoX documents AI-driven credential stuffing at twelve million credentials
The same Hacker News reporting that covered ARTEX details a second operation, which ZenoX calls SCARLET LOOP: an AI-assisted campaign that tested 12,277,358 stolen and found 11,832 valid across 3,968 domains. The models doing the work were a mix of commercial and open systems, including GPT-5.5 and 5.6, DeepSeek V4 and Claude Opus 4.6, per ZenoX's attribution.
A 0.1% hit rate across twelve million attempts is what automation is for, and the AI contribution here is not sophistication but adaptation, working around rate limits and login flows that defeat dumb scripts. Credential reuse was always a liability with a long tail; agents shorten the tail. SourcesB
Japan logs 86 cyberattacks in September and tells companies to review their defenses
Japan recorded 86 cyberattacks in September, up 18% from August and 37% from the previous quarter, with SoftBank Corp, Daiwa Securities and Lawson among the targets, Bloomberg reported Friday. The country has already logged more incidents in the first nine months of 2026 than in all of 2025, and officials are urging security reviews, with AI-assisted attacks named as a driver.
One country's incident count crossing its full prior year with a quarter to spare is the national-statistics version of what CrowdStrike and ZenoX described at the level of individual operations this week. The attack tooling got cheaper; the count went up. The two facts are doing the explaining for each other, and Japan is simply the first government publishing the tally. SourcesB
Aaronson: labs have quietly started testing whether models can break cryptography
Scott Aaronson, writing Wednesday on OpenAI's mathematics release, added a detail with more consequence than the theorems: his sources say AI companies have begun, "gingerly and discreetly," checking whether their latest internal models can break important cryptographic protocols and primitives. Cryptography is conspicuously absent from the release's roughly 376 papers.
On the mathematics itself, his accounting is the sober one: the model solved about 5% of the roughly 8,000 open problems it attempted at about three hours of compute each, the Unique Games Conjecture proof carries a certificate while others do not, and no human has fully understood all the proofs yet. P versus NP and P equals BPP remain untouched, which he reads as the hardest problems staying hard. The cryptography line deserves separate attention because it describes a capability evaluation with no public protocol: if a lab's internal model shows progress against deployed primitives, the disclosure question makes the math-verification debate look quaint. SourcesA
Nvidia puts $1 billion over five years into US scientific computing
Nvidia committed $1 billion over five years to US scientific discovery under the Genesis Mission, announced at the Washington summit that convened Wednesday. The funding targets AI research in quantum computing, healthcare and energy security, and sits alongside Nvidia's existing commitment to supply seven supercomputers across Argonne, Los Alamos and Lawrence Berkeley, including the 100,000-GPU Solstice system built with Oracle.
A chip vendor funding the science agencies that buy its chips is industrial policy running in both directions at once, and it is cheap at the price: $200 million a year against data-center revenue running two orders of magnitude higher buys standing in every DOE procurement that follows. SourcesB
A job listing sketches Anthropic's plan to brief the 2028 presidential field
An Anthropic job posting for a Political Programs Lead, listed October 2nd, describes building the company's "presidential engagement program" for the 2028 cycle: mapping prospective fields in both parties, offering AI policy education to candidates and advisors on equal terms, and preparing executives for those conversations, plus outreach programs for federal, state and local races. A listing describes intent, not a launched program, and Anthropic has not announced one.
The context makes it more than HR arcana. Anthropic has pledged more than $40 million to Public First Action, reporting puts nearly $265 million of planned midterm spending behind super PACs tied to Anthropic and OpenAI, and Obama is privately urging prospective Democratic candidates to develop detailed AI plans. The labs have concluded the 2028 election will be fought partly over their product, and they are staffing for it two years early. SourcesAC
Perplexity releases open-weight multimodal embedders built for late interaction
Perplexity released pplx-embed-v2-late on Wednesday, two models, at 0.6 billion and 9 billion parameters, that retrieve text, images and full pages in a shared embedding space, with weights on Hugging Face. The company reports 92.4% on the MADQA retrieval benchmark for the 9B model.
A consumer answer-engine company open-sourcing its retrieval stack is a competitive statement about where it thinks its moat is not. Late-interaction models keep per-token representations and match at query time, which costs storage and pays in ; shipping strong open ones moves that tradeoff from research argument to deployment option for everyone building retrieval, including Perplexity's competitors. SourcesA