November 3rd 2026

Anthropic tells investors it may "imminently" cross its own threshold for AI-accelerated progress

The amended S-1 that Anthropic filed on Monday carries a new passage in its risk factors. "Our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI,'" it says. No current model has been designated as over that threshold, and the company is "evaluating the implications" and will publish further reports. It also says its ability to measure its own acceleration is limited, so that it "may in the future not identify" a jump in its rate of development "until after it has occurred."

A second passage says Anthropic does not expect to be able to safely let Claude "fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding," and that it is considering changes to the policy "to keep humans in control of frontier model development." The same filing puts the scale of the automation already in place: more than 90% of contributions to Anthropic's codebase in September were written by Claude, and the daily volume of code per engineer in the third quarter was more than 45 times its 2024 average.

A prospectus is a document a company can be sued over, which gives these sentences more weight than a blog post and also a lawyer's reason to exist. They arrive as investors argue over price: the Wall Street Journal reported that some are pushing for a valuation closer to $1.5 trillion to $1.6 trillion, below the $2 trillion that circulated in September. SourcesABB

Nvidia tells cloud partners to sell at least 51% of their capacity to non-Chinese customers

Nvidia has told the cloud companies that rent out its chips that at least 51% of their services must go to companies that are not Chinese, the Financial Times reported on Monday, citing five people familiar with the matter. The rule applies to new contracts only, was passed on verbally by its sales force and regional teams rather than in writing, and is the first restriction of its kind Nvidia has imposed. It reaches the route Washington has spent years trying to close: Chinese developers renting compute in data centers outside China, from providers in Singapore, Malaysia, Japan and elsewhere.

Nvidia neither confirmed nor denied the report to AFP, and said in a statement that "America cannot win the AI race unless America's AI infrastructure wins worldwide." China's foreign ministry said it was not aware of it. A 49% ceiling is a sales policy rather than a law, which means a neocloud with a Chinese majority loses its supplier, not its license. SourcesBB

Beijing plans safety rules covering an AI model's whole life, from data to deployment

China's state planner, the National Development and Reform Commission, said at a briefing in Beijing on Tuesday that it will roll out a broad set of AI safety standards running from development through deployment, and is speeding up AI legislation. Senior official Wang Ruomeng said the rules will cover data, computing power and algorithms, that developers and operators will have to disclose safety risks, incidents and vulnerabilities, and that the government will monitor AI's effect on jobs and support displaced workers. Bloomberg described the package as some of the most comprehensive guidelines anywhere against loss of control over advanced AI.

Compliance gets a transition period, guidance and sector sandboxes, and the planner says it will keep encouraging open development through the five-year plan that ends in 2030. Wang asked other countries to coordinate on vulnerability disclosure, calling AI safety "a global challenge requiring international cooperation." No text or timetable was published. SourcesBB

Huawei tells some chip customers to find their own high-bandwidth memory

Huawei cannot buy enough high-bandwidth memory from Samsung or SK hynix, so it has told some customers for its next flagship AI chip, the Ascend 960DT, to source the memory themselves, The Information reported on Monday. Fujian Jinhua, the sanctioned Chinese memory maker that has supplied Huawei for years, trails the Koreans and lacks the capacity. Much of the gap is filled by smuggling: HBM is mounted beside cheap, low-performance server processors, which export rules judge by the processor's speed, and then heated off the boards in Chinese workshops. A broker quoted about $4,000 per HBM3E stack, and an AI chip needs six to eight.

The 960DT, due in the first quarter of 2027, is meant to double the memory bandwidth and capacity of Huawei's current part, and Huawei is tuning it with DeepSeek for pre-training, which DeepSeek has so far done only on Nvidia hardware. Huawei has also raised the suggested price of its current Ascend 950DT above 250,000 yuan, about $37,000, according to TrendForce's summary. SourcesBC

Apple calls memory prices "unprecedented" and guides its margin down for the holidays

Apple reported record September-quarter revenue of $114.5 billion on Monday, up 12%, in the first results under chief executive John Ternus. Then it said the December quarter's gross margin will fall to 45% to 46%, from 47.9%, mostly because of memory. Ternus called memory pricing "unprecedented" and said costs will keep rising after December; inventory has climbed 154% in a year to $14.54 billion, and manufacturing purchase obligations rose by $33 billion in a quarter as Apple locked in supply. Excluding tariff refunds, both its margins fell about a point from the previous quarter on memory alone.

Two AI costs appear in the guidance. Running Siri's new AI features raises Apple's cloud bill, and Apple will defer part of the revenue from devices that have them, recognizing it over the device's life. Chief financial officer Kevan Parekh also pointed to "opportunities across third-party apps and partnerships" in AI without saying what they are. The shares fell before Tuesday's open. SourcesABBB

Salesforce agrees to buy Cresta, the voice-agent company for call centers

Salesforce said on Monday it has signed a definitive agreement to buy Cresta, whose software answers customer calls with voice agents, coaches human agents during calls and analyzes conversations. Salesforce did not disclose terms; several outlets, following The Information, put the price at about $2.7 billion. Cresta began at Stanford's AI lab and was founded in 2017 by Sebastian Thrun, Tim Shi and Zayd Enam. Salesforce will fold it into Service Cloud and says Cresta will keep supporting the telephony platforms its customers already use. The deal is expected to close in the first quarter of Salesforce's fiscal 2028, which begins in February 2027.

The buyer is paying for the phone line. Text support has been automated for two years; voice is where the expensive human minutes still are. SourcesABB

Agility Robotics wins its SPAC vote and starts trading Thursday as AGLT

Shareholders of Churchill Capital Corp XI approved the merger with Agility Robotics on Monday, with more than 84% of votes cast in favor, according to the 8-K filed the same day. Holders of about 44.9% of the voting power attended. The deal is expected to close on Wednesday and the combined company to start trading on Nasdaq as AGLT on Thursday, November 5th.

The money is smaller than the headline that launched the deal. Agility expects about $465 million gross: roughly $164 million left in the trust after redemptions, $201 million from a Foxconn-led PIPE at $10 a share, and $100 million it has already received. The June announcement assumed more than $620 million with no redemptions. Agility will be the only US-listed company that does nothing but humanoid robots, which means its share price becomes the sector's only public mark. SourcesAA

Drax, Brookfield and Radiant plan a gigawatt AI campus at a Yorkshire power station

Drax Group signed an agreement on Tuesday with Brookfield and its neocloud, Radiant, to develop an AI data center on the Drax Power Station estate in North Yorkshire that could pass 1 gigawatt. Brookfield estimates total investment, chips included, could exceed £20 billion ($26.8 billion), with £2.5 billion for a first phase of up to 100 megawatts from early 2028 on grid power. Another 500 megawatts would follow between 2028 and 2031. Nothing is final: phase one needs planning, regulatory and network approvals and a tenant. Drax shares rose about 8.8% in early trading.

A power station is the obvious answer to Britain's real AI problem, which is the wait for a grid connection: Microsoft walked away from anchoring an Nscale site near London last month over grid delays. SourcesB

Britain's cyber agency says China is best placed to use AI against critical infrastructure

The UK's National Cyber Security Centre published its annual review on Tuesday, covering September 2025 to August 2026, and its central judgment is blunt: "It is almost certain AI will act as a force multiplier across the cyber threat landscape," speeding up reconnaissance, social engineering, vulnerability discovery and exploit writing rather than inventing new kinds of attack. China "highly likely has the intent" to use AI against critical national infrastructure, and the NCSC judges it the best positioned to do so given its investment in frontier AI. It also says "it is highly likely AI malware will be used in a ransomware attack on UK CNI," and that defenders will have to shorten their patch timelines.

The NCSC says human operators remain essential to attackers, and that losing control of an agentic tool would probably stop them reaching their goals. It also warns about "shadow AI," staff moving sensitive data to personal devices to use unofficial tools. SourcesAB

OpenAI hires former Army secretary Dan Driscoll for national security

OpenAI has hired Dan Driscoll, who stepped down as US Army secretary earlier this year after months of reported tension with Defense Secretary Pete Hegseth, as senior counselor for national security modernization, a new role. He will advise on using AI for US national security and cybersecurity and help allied governments adopt it, building on OpenAI's Daybreak program for cyber resilience. Sam Altman said the US has "an extraordinary opportunity to work with other countries to spread the adoption of democratic AI." Driscoll, an Iraq veteran, pushed drone and counter-drone access for allies while at the Army.

The job description is a sales territory: ministries of defense are the customer OpenAI does not yet have in volume. SourcesBB

OpenAI more than quadruples its bug bounty and brings in Cathedral to attack its defenses

OpenAI said on Monday it is raising its security bug bounty, more than quadrupling rewards for critical findings, with up to $172,000 including bonuses for exceptional ones. It will also run its next internal security sprint with Cathedral Systems, whose researchers will use OpenAI's models alongside its own teams to test its defenses. The last sprint, in which engineers used Codex to reproduce vulnerabilities in isolated environments and draft patches, involved about a quarter of OpenAI's product engineers. No start date or scope was given.

Cathedral is a pointed choice. Secondary reports say it previously reached Anthropic's internal systems through third-party software during a bounty investigation, which Anthropic said broke the program's rules; we could not confirm that account. SourcesAB

MIT researchers had GPT-6 Astra edit a small Qwen model's weights directly, without training

Researchers at MIT CSAIL, led by PhD student Rikiya Takehi with Omar Khattab, report that GPT-6 Astra improved Qwen3.5-0.8B's math by changing 0.016% of its weights by hand, with no gradient descent: it reasoned about what parts of the network compute and set new values. They report MATH-500 scores rising as much as sixfold and IFEval about threefold, and Astra writing a Breakout-playing policy network from scratch. The authors pitch it as a possible route to fixing models without forgetting, and to self-improvement.

The skeptics have the stronger reading of the method. Matthew Honnibal points out that the headline Qwen result used Python and ran the model on up to 50 batches of practice problems, so it is not "just looking at the raw weights," and that the from-scratch Breakout network beat a baseline on Breakout but lost on Seaquest and Asterix. He still credits the Qwen result as real. It is one small model and a few benchmarks, without peer review. SourcesABC

OpenAI's new everyday-errands benchmark puts its own models at 80.6%

OpenAI published LifeVal on Monday, a benchmark of whether a model can finish everyday tasks for people across caregiving, health, education, finance, shopping and household admin, built from tasks that about 100 participants recorded wanting an assistant to handle. GPT-6 Astra averages 80.6%, against 80.0% for GPT-5.6, 77.4% for GPT-5.5, 73.0% for GPT-5.4 and 49.3% for GPT-5.2.

The curve is the finding, and it is flat: one point from GPT-5.5 to the newest model. The task set is not public, OpenAI grades it, and no other lab's model has been scored, so it says nothing about how Astra compares with Claude or Gemini. SourcesAC

Palantir falls about 8% before the bell as new contract signings slow

Palantir's shares were down about 8% in premarket trading on Tuesday, near $164, after Monday's results. Revenue rose 96% to $2.31 billion and the full-year forecast went up to $8.53 billion, but total contract value, the measure of new business signed, grew 6% against 49% the quarter before. The chief financial officer tied shorter deal durations to customers starting on Palantir's new sovereign AI products. SourcesBC

Putin orders Russia's private data centers to hold back 10% of capacity for the state

Vladimir Putin signed a decree on Monday requiring owners of Russia's non-state data centers to reserve 10% of their capacity "to help keep Russia's IT infrastructure running," Reuters reported. A new government commission can order the reserve released, and operators must provide access within 24 hours. The digital ministry will inventory capacity and floor space nationwide, and the decree allows temporary use of data centers abroad with approval from the FSB, the security service. Ukrainian drones struck facilities run by Yandex and Sberbank in October. SourcesB

Meituan releases LongCat-2.5, a 1.6-trillion-parameter model for long coding jobs

Meituan, the Chinese delivery company, released LongCat-2.5 on Monday: 1.6 trillion parameters with about 48 billion active per token, handling images and text together, trained with large-scale reinforcement learning on whole codebases, and with failures from real agent runs turned into training tasks. It is available through Meituan's chat service and API. Weights on GitHub and Hugging Face are "planned to follow," with no date; Meituan's newest open model on Hugging Face is still LongCat-2.0 from July. No benchmark scores came with the release. SourcesB

SUSE will cut one in ten jobs and names "automation and AI" as the reason

SUSE, the Linux company owned by EQT, is cutting about 10% of its roughly 1,900 staff, chief executive Dirk-Peter van Leeuwen told employees in a memo. "Modern tooling is already changing our daily workflows," he wrote, and some roles have "greatly diminished." Affected staff will be told by November 4th; a voluntary departure offer in September did not draw enough takers. SourcesB

Squarespace lays off 19% of its staff as AI site builders eat its market

Squarespace is cutting about 19% of its workforce, around 300 people, the company confirmed to TechCrunch on Monday, saying it wants to simplify operations and move faster. The company that sold drag-and-drop websites now competes with tools that build a site from a sentence: Lovable raised at $13.3 billion in August. Squarespace was taken private for $7 billion in 2024 and has an AI-native platform, Forma, on a waitlist. SourcesB

Governors' races in at least a dozen states put data center pauses on today's ballot

Americans vote today in midterms where data centers have become a governor's-race issue. Candidates in at least a dozen states have backed a moratorium on new data centers, by Newsweek's count, from both parties: Pennsylvania's Republican nominee Stacy Garrity has called for a temporary pause, and Democrats Doug Jones in Alabama and Hannah Pingree in Maine back one. New York is the only state with a moratorium in force, the one-year pause Kathy Hochul signed in July. Polls have been consistent on the local question: Gallup found 71% opposed construction near them in March. SourcesBB

Frontier models refuse unsafe shortcuts but cave in a negotiation, a teamwork benchmark finds

A new Collaborative Intelligence Benchmark from Sean Wu and colleagues tests 18 models on 240 tasks drawn from sources such as aviation accident reports, clinical cases and negotiations, scoring only tasks each model can solve alone, so what is left is teamwork. The best model scores 82.2 out of 100. Frontier models almost always refuse unsafe shortcuts and report misconduct, but reach the best deal in only 14% of conflicts, and their pass rate drops from 97% to 81% when they take over a task a weaker model cheated on. Models of near-identical capability differ by up to 11 points as teammates. SourcesA

Trajectory raises $45 million from Sequoia to let lawyers and doctors post-train models

Trajectory, an 18-person startup founded about a year ago by three Stanford classmates, raised a $45 million Series A led by Sequoia at a $300 million valuation, with Nvidia, Bessemer and Bain Capital Ventures. Its platform takes an open-weight model, connects it to a company's applications and keeps retraining it on feedback, with an interface for experts who do not write code. Harvey and Rogo use it for legal and financial models; Decagon, Mercor and Nvidia are also customers. Co-founder Ronak Malde was among the Windsurf researchers Google DeepMind hired. SourcesB

Garry Tan says he spends up to $200,000 a year on AI tokens

Y Combinator's chief executive told Walter Isaacson's podcast The Builders that he spends "maybe $100[,000] to $200,000 a year on tokens," adding: "It's too expensive right now." His agent reads his email, Dropbox, Apple Notes and 20 years of therapy notes. YC runs an internal agent called Stray Light with context on every founder it has backed and every application it has received. A heavy personal agent user now costs as much as a senior engineer, which is the number that decides how fast these spread. SourcesB

Google takes Dreambeans to India after US downloads fell 84% in a month

Google launched Dreambeans, its app that turns a user's Calendar, Gmail, Photos and YouTube history into illustrated stories without being asked, in India on Monday in 13 languages, its first market outside the US. Sensor Tower's estimates, shared with TechCrunch, show about 55,000 US downloads since June, and a fall of 84% to about 5,500 in October. In its first 30 days open to all US adults it drew about 29,000 downloads, against 7.4 million for Meta's Muse app over the same window. SourcesB

Meta cleared AI videos of coaches touching young athletes until a reporter called

Futurism reported ten Facebook accounts posting near-photorealistic AI videos of older coaches touching young female gymnasts, dancers and yoga students. A week later eight had been told they did not break Meta's rules, and Facebook was recommending more of the same videos to the reporting account. After Futurism contacted Meta's press office, every reported page came down and the earlier notices were changed to say the pages "go against our Community Standards." Some clips had passed 100,000 views. SourcesB

An open 27-billion-parameter computer-use agent scores 80.2 on OSWorld-Verified

CUA-Master-27B, from a team led by Xinlei Yu, scores 80.2 on OSWorld-Verified, a benchmark of agents operating a real desktop, plus 36.9 on OSWorld 2.0 and 25.0 on Agents' Last Exam. Its training data comes from workflows mined from instructional videos, rebuilt into tasks, with a retry mechanism that rolls a real environment back to a snapshot after a failure. A model of that size is small enough to run privately. The scores are the authors' own. SourcesA

Yandex publishes attention kernels that train faster on Nvidia's Blackwell chips

Yandex Research released FLAMIA, a family of attention kernels for Blackwell GPUs that train in MXFP8 and mixed NVFP4/FP8 low-precision formats, with safeguards so shared scaling factors do not leak information from future tokens. It reports forward-plus-backward attention speedups of 1.10 to 1.40 times on B200 and up to 1.42 times on B300 over the fastest baseline it tested, and stable Llama-style pretraining. SourcesA

A transformer that picks the order of its own layers at run time

EPFL and NYU researchers, with Micah Goldblum, propose Library of Layers, a model that keeps a shared library of transformer blocks and decides at inference which to run and in what order. It matches dense, depth-recurrent and mixture-of-experts baselines on text pretraining; on sudoku and mazes it uses different orderings for different strategies, and the researchers can force a strategy by intervening on the router. SourcesA

Multimodal models can build a working lab simulation from a video, but not the right apparatus

Lab2Sim, a benchmark of 218 science experiment videos with 11,881 rubric items, asks models to turn a video into an interactive virtual experiment that enforces each step's preconditions. Models readily produce pages that run, the authors find, but reproduce the apparatus poorly; their own pipeline, which writes the procedure as a state machine first, lifts apparatus fidelity by 24.15 points. SourcesA