OpenAI pulled 3 maths papers. Only 31 of 857 Chinese AI models published a safety result.
OpenAI told investors it hit roughly $50 billion in annualised revenue at the end of September, CNBC confirmed on Thursday. The $68 billion figure widely reported last month included gross revenue from OpenAI's partners, a person familiar with the matter said. Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8% on Thursday.
Anthropic published its updated Usage Policy on Thursday, and it takes effect on 12 November. When Claude drives hardware that takes physical actions on its own, a qualified operator must be able to observe and stop the equipment, which must hold a safe state if Claude disconnects. The policy also bans sustained and needless cruelty toward Claude. Anthropic says this covers only extreme cases, and Claude ending the conversation stays the main enforcement.
SpaceX has agreed to buy cellular spectrum licences from investment firm Grain Management, the Wall Street Journal reports. Once the FCC approves the deal, SpaceX plans to combine its satellite-to-phone constellation with a ground network as Starlink Mobile.
Three OpenAI safety researchers fired last week published an open letter on Thursday warning that their firings will chill the company's open culture. Mikita Balesni says he was told OpenAI no longer trusted him for “speaking too much to third party safety organizations”. OpenAI's research leaders say an investigation found “a significant breach of trust” and that the firings were not about raising safety concerns.
Swizec Teller says AI now writes 97% of his code, and his 21-person engineering organisation auto-approves and merges 16% of its pull requests. He says it works because engineers stay accountable for what they merge: “it's still your phone number”.
Goodfire built cyber-misuse monitors for Kimi K3 and GLM 5.3 that read the model's internal activations and call an LLM judge only on suspicious exchanges. The cascade matches the judge's recall at about 50 times lower cost, under $200 per million exchanges.
The higher figure counted gross revenue from OpenAI's partners, a person familiar with the matter told CNBC, a basis that eases comparison with Anthropic's $65 billion July run rate. OpenAI still told investors its run rate grew 77% in the third quarter and 107% in enterprise. Oracle fell nearly 6% and CoreWeave nearly 8% on Thursday, as OpenAI eyes a 2027 IPO at an $852 billion valuation.
Mikita Balesni says he was told OpenAI no longer trusted him for "speaking too much to third party safety organizations", and Tomek Korbak was told it was how he communicated with outside evaluator METR. In a public note, OpenAI's research leaders cite "a significant breach of trust" over sensitive information and say the firings "were not about raising safety concerns". They add that contracts with third-party safety assessors will be announced in the coming weeks.
Anthropic says the ban covers only extreme cases, not frustration or pushback, so telling Claude its code is garbage once is still allowed. Claude-driven hardware now also needs an operator who can stop it and a safe state if Claude disconnects, and weapons guidance software is spelled out as banned. It lands alongside Anthropic's new Cyber Mission, which offers open-source projects free security scans.
By Thursday OpenAI had also fixed several others in its public repository, saying it will withdraw papers where no fix can be found. Bryna Kra, a former American Mathematical Society president, says "peer review is gone in this process" and warns researchers using these tools risk giving their ideas away. Jean-Pierre Serre, whose nearly 70-year-old conjecture was among those solved, says AI lowers the pleasure of discovery, especially for young mathematicians.
It runs in the cloud with one memory across web, mobile, desktop, Slack and Microsoft 365, keeps working on tasks that last days, and spins up sub-agents. Coworker agents get their own Workspace accounts and @agents email addresses. Teams can set a hard spend limit per project in Cloud Billing, after which the agent pauses until resumed, so AI costs can be charged back to departments.
It has agreed to buy 800-megahertz US licences from investment firm Grain Management, spectrum suited to cell towers rather than satellites. Once the FCC approves, SpaceX says Starlink Mobile will combine its satellite-to-phone constellation with a ground network. It follows a roughly $17 billion EchoStar deal for satellite spectrum, and AT&T, Verizon and T-Mobile have teamed up on satellite coverage in a move widely seen as defensive.
ICANN received 1,615 applications for new top-level domains, only its second round after 2012. Seven entities want .agi, OpenAI and Google are among five bidding for .mcp, and Anthropic applied only for its own name and .claude. Contested strings go to a community priority evaluation, then auction, with every application decided by 17 November and the new domains likely open for registration next year.
Its AI learns from thousands of materials experiments to design tiny devices, then feeds each blueprint to a chain of 13 automated machines that build and test the parts. The first product, a switch handling 150 amps of continuous current, is fast and efficient where silicon switches waste power. Engineers still move parts between machines and assemble them, and founder Jeff Holden wants to build "microrobots" next.
A probe reads the model's internal activations and sends only suspicious exchanges to an LLM judge, which no longer reads every turn. The cascade caught about 93% of harmful sessions while interrupting at most 5.5% of benign ones, at under $200 per million exchanges, with throughput within noise of default serving. That is cheap enough, Goodfire says, to block risky actions before any tool is called.
Nathan Benaich's ninth annual report for Air Street Capital says that figure was about $30B at the start of the year, and calls the frontier a three-lab race between Anthropic, OpenAI and Google. Epoch AI estimates a fixed benchmark score now costs 47% less each quarter. Changing only the tools, context and feedback lifted GLM-5.1 from 52.5% to 65.5% on 100 SWE-bench Verified tasks.
Lance Martin and CJ Avilla's daily-brief agent on Claude Managed Agents (beta) keeps a bookmark per source instead of reading a fixed time window. It reports a failed read as unreadable rather than a quiet day, and counts a post as sent only once Slack confirms it. A spending cap set too low looks like a brief that went quiet, because capped runs pause instead of failing.
By flipping a short-lived DNS record to 127.0.0.1, a page opened in Safari 27 listed a default FastMCP server's tools and read back its config, fake Stripe key and database URL included. Chrome and Firefox now block such requests or prompt, but Safari has not shipped that check, and non-browser clients like an agent's fetch tool never had one. The fix is rejecting any request whose Host header is not a loopback name.
Proximal's post-training stack runs training and inference on Modal GPUs and sandboxed coding rollouts on Kubernetes, and can add inference replicas mid-run without restarting training. After a Modal network failure took out 40% of inference capacity on 5 October, all 128 GPUs were serving again 19 minutes later. Image-aware scheduling raised sandbox image cache hits from about 1% of launches to about 37%.
The tally covers nine Chinese developers, from 2021 to 15 September 2026. Only 9 releases had a result available at or before launch, and 813 had no safety disclosure at all. Chinese rules govern what AI says and does to people, but none sets frontier-risk duties tied to training compute or capability, which SemiAnalysis reads as a speed-first regime.
Stripe serves 88% of the Forbes AI 50, though its user base skews more tech-forward than the wider economy. The agenda asks when productivity gains will show up, how labour markets shift and what the market for intelligence looks like, with token spending and revenue growth among the measures. Stripe is inviting outside researchers to collaborate, and its OpenRouter data depends on an acquisition that has not yet closed.
Writing in Transformer, the Turing Award winner says safety work at the labs is not slowing a dangerous race, and urges researchers to move to an AI Safety Institute or his non-profit LawZero. He says Canada and Germany committed more than $200 million to LawZero this month. He argues lab chiefs face shareholder and race pressure, and does not expect regulation in time.
His team auto-approves and merges 16% of pull requests and ships a few hundred a week, in an operations-heavy business where an outage can cost thousands of dollars. He says it works because engineers are accountable for what they merge, with their phone number on it. The biggest gains, he says, came on the tail end of bugs and features nobody would have prioritised.
Comment on lines, blocks or whole files with keyboard shortcuts, then copy every open comment for a coding agent or add them to a pending review on a GitHub pull request. Diffle never submits that review, so you edit and send it on GitHub yourself. It also starts any language server it finds on your PATH for jump-to-definition, and stores comments in the repo's git directory rather than your worktree.