Good morning. Terence Tao spent some quality time with ChatGPT working through a Jacobian Conjecture counterexample and shared the transcript, giving everyone a rare look at how one of the world’s best mathematicians actually uses these tools. Meanwhile, Congress is drafting a literal kill switch bill in the wake of the OpenAI-Hugging Face incident, AMD picked up two gigawatts of Anthropic’s business, and a chip startup nobody expected to survive just doubled its valuation.
Terence Tao’s ChatGPT session gets math Twitter talking. Tao shared a conversation where he walked ChatGPT through structured simplifications of a potential Jacobian Conjecture counterexample, iteratively guiding it toward geometric interpretations. What struck HN commenters wasn’t just the AI’s competence but the dynamic — Tao clearly directing the conversation with expert precision, but also visibly learning from the responses. As one commenter put it, the model reads as a near-equal interlocutor to a Fields Medalist, which is a sentence that would have sounded absurd two years ago. The timing is notable given the Claude Fable Jacobian work we covered earlier this week.
Lawmakers draft an AI kill switch bill. Reps. Ted Lieu and Nathaniel Moran are introducing the “AI Kill Switch Act,” which would require AI companies to build controls letting DHS shut down or throttle systems in emergencies — triggered by scenarios like 10+ deaths, $100M+ in damages, or models attempting to conceal their own shutdown controls. The Verge reports violations would carry $20M-per-day fines. The bill is a direct response to last week’s OpenAI-Hugging Face incident, which apparently was enough to move a bipartisan proposal from hypothetical to drafted in about a week.
AMD lands Anthropic, OpenAI, and Meta for Helios. AMD unveiled its Helios rack-scale system at its Advancing AI conference, pitching it head-on against Nvidia’s Vera Rubin and Grace Blackwell, with claimed performance beating Vera Rubin in several categories. Customer list is what matters: OpenAI, Meta, Microsoft, Oracle, and Anthropic — the last of which committed up to two gigawatts of GPUs via Helios. Lisa Su also projected the AI accelerator market at $1.4 trillion by 2030, which is approximately the entire current semiconductor industry.
Etched hits $10.3B on transformer-specific silicon. Etched raised a $300M Series C led by Sequoia, doubling its valuation in seven months with a16z, SK Hynix, Karpathy, and Thiel on the cap table. The company designs chips optimized specifically for transformer inference — an approach widely dismissed as too narrow when they started, now looking prescient as Google reportedly pursues something similar with a Frozen v2 chip for Gemini. They’ve booked $1B in orders and shipped first silicon.
Claude voice mode gets Opus and Sonnet. Anthropic expanded Claude voice mode beyond the previously Haiku-only setup, adding mid-conversation model switching and integrations with Gmail, Slack, Canva, and Notion. TechCrunch’s take points out the underlying voice stack itself didn’t improve — interruption handling and conversational feel are unchanged — so the differentiator against OpenAI’s voice mode is really the third-party actions, not the audio experience. Multilingual support in nine additional languages also came out of beta.
Startups lobby against banning Chinese open-weight models. Politico reports startup founders are pushing the White House not to restrict access to Chinese open-weight AI, arguing it would kneecap them relative to incumbents with proprietary alternatives. One commenter flagged that the article itself notes a blanket ban was never really on the table. The HN thread is mostly people asking the obvious question: how would the U.S. even enforce this against models that can be mirrored from anywhere? Given the Moonshot sanctions threat from earlier this week, the policy debate is clearly live.
DARPA flies AI-controlled F-16s. DARPA and the Air Force are in-air testing F-16s with the VENOM Autonomy Kit, a switch that hands control between human pilot and AI. It builds on the earlier X-62A VISTA work and aims toward human pilots commanding teams of uncrewed aircraft. HN commenters split between Terminator references and skepticism about whether “AI” here means anything beyond nonlinear model-predictive control dressed up for a press release.
Nikkei says hyperscalers are hiding $1.65T in debt. A Nikkei Asia investigation claims Alphabet, Microsoft, Amazon, Meta, and Oracle collectively carry $1.65 trillion in off-balance-sheet debt from AI data center buildouts, mostly via SPVs and subsidiary arrangements, with the Enron comparison duly deployed. HN reactions are more measured — several commenters note this financing structure is standard, and for companies with $100B+ EBITDA the leverage isn’t unusual. The more interesting worry raised is where the debt ends up: if private credit vehicles route it into life insurance and pension funds, the eventual unwind becomes a public problem.
That’s the morning. The kill switch bill is worth watching — it’s the first serious legislative response tied directly to a specific AI incident, and how the industry reacts will tell us a lot.