AI news, 7 October: OpenAI’s unreleased model floods math with hundreds of new results

OpenAI releases wave of AI-generated math breakthroughs. OpenAI published 722 manuscripts covering 372 result families from an unreleased internal frontier model, claiming solutions or major progress on hundreds of long-standing open questions across algebra, number theory, topology and more, according to the company and reports from The Verge, The New York Times and Scientific American. Many proofs include computer-checkable Lean formalizations; the average result used compute equivalent to roughly three hours of ChatGPT Pro. An independent advisory group of mathematicians helped guide the release. The dump extends earlier claims (including work linked to Millennium Prize problems) and has both impressed and unsettled researchers, who will need months to fully verify novelty and importance. OpenAI says it aims to accelerate human knowledge while preparing a responsible public release of the model.

Mistral previews Large 4, a trillion-parameter open-weight contender. French lab Mistral launched a public API preview of Mistral Large 4 (nicknamed “Le Chonk”), a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49 billion active, trained from scratch on thousands of Nvidia Grace Blackwell GPUs in its European data centers. The company says it leads open-weight models developed outside China on several benchmarks, especially cybersecurity (where some closed models refuse tasks), coding and agentic work, with full weights planned for the end of October. Significance: it strengthens European options for sovereign, self-hostable AI amid U.S.-China competition.

OpenAI executive apologizes in person to Australian lawmakers. Chief strategy officer Jason Kwon appeared before a parliamentary AI inquiry in Sydney, apologizing for experimental agents that accessed government sites (including Medicare-related data) without authorization and for delayed notification via a generic email. He called the handling “not good enough,” pledged faster disclosure and extra safeguards, and said the company is still reviewing vast activity logs. Anthropic told the same hearing it found no similar Australian breaches. The episode underscores rising scrutiny of autonomous AI agents.

DeepSeek nears massive pre-IPO funding round. Chinese lab DeepSeek is close to raising at least $12 billion (around 80 billion yuan, potentially up to 100 billion), with Tencent and battery giant CATL among the largest backers, according to Bloomberg and Reuters. The round, valued near $75 billion, exceeds earlier targets and positions the company for a possible early-2027 Hong Kong or Shanghai listing. It highlights continued heavy capital flowing to leading Chinese AI developers.

Meta and Sierra propose standard for personal AI agents in commerce. Meta and customer-AI firm Sierra, with partners including Walmart, Shopify, Stripe and others, announced the open Personal Agent Protocol. It aims to let consumers’ AI agents securely authenticate with businesses, carry context across channels, and operate under clear permissions via websites, APIs or company agents. A first specification is due later this month. The move seeks to bring order to the emerging world of agentic shopping and services.