OpenAI released 722 maths manuscripts as Mistral unveiled a trillion-parameter model
One lab emptying a drawer of results, a very large model out of Paris, a safety insider walking out, and a rare count of what AI bug hunting has actually turned up.
OpenAI released 722 maths manuscripts written by an unreleased model OpenAI said on Tuesday the internal model attempted about 4,000 problems, with each result taking about three hours of ChatGPT Pro-level thinking on average. Some proofs come with Lean versions a computer can check. Source
Mistral unveiled Large 4, a 1-trillion-parameter model, with open weights to follow It went into preview on Tuesday at $1.36 in and $4.18 out per million tokens, with the weights due at the end of October. Mistral says 49 billion parameters are active at a time. Source
OpenAI safety staffer David Robinson quit, saying its culture is broken He led transparency work on the safety team and, in an Atlantic essay published on Saturday, called for nuclear-plant-style safeguards across the industry. Source
Anthropic says Mythos helped partners find 129,000 bugs in four months More than 33,000 of the flaws, found between April and July, were rated critical or high. On Tuesday Anthropic split its Cyber Verification Program into three tiers for vetted security teams. Source
ChatGPT’s text will carry an invisible watermark in the EU OpenAI said on Monday it starts within weeks for ChatGPT and Codex. Its detector, open only to approved researchers, caught about 80 per cent of 200-token passages at a 1 per cent false-positive rate. Source
Google contracted 3.6 gigawatts from Constellation, 890MW of it new nuclear Constellation will spend $4.3bn upgrading 11 reactors, with the first extra power due in 2028; the rest comes from existing plants over 15 years. Source