Good morning. Washington is making its presence felt today — Anthropic is drafting its own AI safety bill while DHS boots DeepSeek off federal systems, even as DeepSeek open-sources the inference stack running its production fleet. Google and Amazon are also making coordinated moves against Nvidia’s grip on PyTorch, which may matter more in the long run than any single policy announcement.
Anthropic drafts a bill that would let courts block dangerous models. Anthropic has been quietly circulating proposed legislation on Capitol Hill that would let courts halt AI model releases if independent evaluators find a “foreseeable, imminent risk” of 50+ deaths or $1B+ in property damage tied to bioweapons, cyberattacks, or models evading developer controls. Enforcement would run through the Commerce Department and D.C. judges, with no federal agency empowered to pull a model unilaterally. The catch: evaluators would be paid by the labs themselves, with only a weak clause saying payment can’t be contingent on outcomes.
DHS orders DeepSeek out of federal systems — same week DeepSeek open-sources its inference stack. The Department of Homeland Security told civilian agencies to remove DeepSeek products, including third-party apps built on its models, citing data-access concerns. HN sleuths noted the exclusion order was actually signed in July and only made public now. On cue, DeepSeek dropped DeepInfer, the framework powering V4.1-Flash in production — complete with Huawei Ascend 950 support and reported numbers of 155 trillion tokens/day per cluster at $38/GPU/hour. Commenters called the Huawei support a direct shot at Nvidia and noted DeepSeek has become one of the few labs shipping serious infrastructure without a press release.
Google and Amazon both go after Nvidia’s PyTorch moat. Google open-sourced TorchTPU, a native PyTorch backend for TPUs with custom ATen kernels, XLA caching, and torch.compile() integration — days after Amazon shipped a similar project for Trainium. Since roughly 85% of AI apps run on PyTorch, and PyTorch has historically meant CUDA, coordinated first-party support from both hyperscalers is a real shift. HN developers who remember the pain of coaxing PyTorch onto TPUs via XLA were relieved; several also wrote TensorFlow’s obituary in the comments.
OpenAI buys Orchestra Bio to deepen its life sciences bet. OpenAI acquired Orchestra Bio, a four-year-old a16z-backed biotech startup, folding its team into the Rosalind platform already used by Moderna and Novo Nordisk. The addition brings clinical trial scheduling and drug-discovery operations tooling. It’s another data point for the pattern of frontier labs absorbing vertical startups rather than letting them scale alongside — a choice founders in applied-AI categories are increasingly being handed.
Anthropic’s cloud deals are getting awkward. A leaked S-1 shows Anthropic struggling to secure enough capacity from AWS, Google, and Microsoft as all three now view it as a competitor. The filing projects $750B+ in cloud contracts against an $80.7B first-half net loss on $16.3B in revenue, with Q3 revenue hitting $16.8B (up 45% QoQ). A November IPO is reportedly the target, with Anthropic increasingly leaning on second-tier data center operators to fill the gaps.
Mistral’s Voxtral 3 is good, closed, and not quite what people wanted. Mistral launched Voxtral 3, a speech-to-text and TTS family with a voice agent SDK covering 50+ languages, hosted in the EU or self-deployable. HN reaction was lukewarm: practitioners pointed out that WER and voice quality are largely solved, and the real problems are latency and interruption handling. Several commenters also noted Mistral’s drift away from open weights — “Has mistral just stopped doing open source?” was a representative question.
Meta’s Muse leaks user conversations via DNS. Researchers found that Meta’s AI agent Muse exposed thousands of shared conversations — including loan applications, business documents, and employee records — because the sharing URLs put non-secret identifiers in subdomains, making them trivially enumerable via passive DNS. Meta suspended the sharing feature. Zuckerberg had publicly described Muse as “built from the ground up for privacy and security”; HN mostly responded with gallows humor.
SlopCodeBench finds “xhigh” effort often loses to “medium.” The October SlopCodeBench update puts GPT-6 Astra on top at 16.3% strict pass, with Opus 5.5, Fable 5.1, SOL 6/6.1, and GLM 5.3 clustered close behind. The more useful finding: cranking effort to “xhigh” cost 2-4x more while actually performing worse than “medium” on several models, including Opus 5.5. Commenters validated the pattern from daily use — higher effort often means more defensive code and more unrequested scaffolding, not better answers. One user noted Astra tends to produce “code golf” that’s technically correct but nearly unreadable.
That’s the brief. Watch whether Anthropic’s bill picks up any actual sponsors — the evaluator-funding question is the one worth tracking.