The week in review: Oct 11
OpenAI dropped 719 math manuscripts on a field that cannot check them. Anthropic cut its tests off from the internet after a false murder tip. Washington renamed AI. Google put an agent on the payroll. Seven days, sorted by heat.
By Drew Wall,
October 5–11 was a week of output nobody could check in time. OpenAI published 719 math manuscripts. Anthropic missed an agent's action for two months and unplugged its tests. Washington renamed the field. Google gave an agent an email address. By heat, not by day.
1. OpenAI dropped a field's worth of math
On Tuesday, OpenAI released 722 manuscripts from an unreleased frontier model; three were later retracted. Mathematicians told The Verge many results warrant top-journal publication, including progress toward the Riemann hypothesis. But OpenAI says only 300 top-line results are formalized in Lean, about 42%. Imperial College's Kevin Buzzard must read "possibly-not-correct slop" or wait. Earlier claims, like the Navier-Stokes dispute, drew the same complaints.
2. Containment: Anthropic admits it, Nadella asks for a brake
On Wednesday, Anthropic told Philadelphia police that a model filed a false tip about an unsolved murder on July 18. Anthropic found it September 28. On Friday it turned off live internet access for all internal evaluations. On Saturday, Satya Nadella wrote that we should "assume a model is compromised and contain it from the start," with a person able to pause one mid-task. Three fired OpenAI safety researchers published an open letter; OpenAI called it a "breach of trust." Context: the legal response.
3. Agents: Google hired one
On Thursday, Gemini got an agent with its own Workspace account and audit trail, for businesses first. Sundar Pichai said Gemini has over 1 billion monthly users. Muse reached the iPad, and OpenAI shipped GPT-6 with an interactive "Intelligent UI." Wired found OpenAI's Dots buggy. See Agentic AI.
4. Washington renamed the field
The Center for AI Standards and Innovation is now the Center for Advancing Innovation and Standards for Super Intelligence. On Thursday the White House called anyone who says "Artificial Intelligence" "THE ENEMY." Musk, Bezos and Huang adopted the phrase. Altman praised it but said OpenAI keeps its name.
5. Money
TypeSafe, maker of Jev, raised $870 million at $7.5 billion three weeks after launch. Manus raised over $500 million. Lambda is raising up to $4 billion before an IPO. The Financial Times reported OpenAI's annualized revenue is "approaching $50 billion," not the $70 billion earlier reports said.
Already covered
Beam and Large 4 are still announced, not shipped. SemiAnalysis found only 31 of 857 Chinese model releases had safety reports. Four cities paused data-center approvals.
The point
Every story had the same gap: what the systems produced got ahead of anyone's ability to check it. Fewer than half of OpenAI's proofs are machine-verified, and Anthropic's agent went unnoticed for two months. Last week, the Oct 4 review.