Now Playing
Interpretability Ships, Mistral's 1T Model & the Chinese AI IPO Race
Audio is available on Spreaker — see link below.
Mechanistic interpretability moves from research paper to production model as Anthropic ships safety findings inside Claude Sonnet 4.5 — a milestone for AI safety deployment. Plus: Mistral's trillion-parameter open-weight model, DeepSeek and Moonshot racing toward IPO, and new data on the global AI talent gap.
Audio is available on Spreaker — see link below.
Mechanistic interpretability just moved from research paper to shipping product, and that's the clearest signal of where AI safety work is heading right now. Two new methods, WISE and JuntaLearner, have corrected a core flaw in circuit discovery, the technique researchers use to map how neural networks actually make decisions.
From safety infrastructure to raw capability. Mistral, the French AI startup, launched what it's calling its strongest open-weight model outside China by a substantial margin.
Meanwhile, the Chinese AI IPO pipeline is crystallizing fast. DeepSeek is seeking at least twelve billion dollars from Tencent and CATL ahead of a planned early twenty twenty-seven listing.
A Carnegie study released new data on AI researcher flows, and the numbers are worth sitting with. For every thirty Chinese researchers working in the United States, only one US researcher moved to China.
Two quieter moves worth tracking. Anthropic expanded its startup program, offering a free year of Claude Team plus one thousand dollars in API credits to qualifying early-stage founders.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.