Nvidia's $12.9B agreement to buy Hugging Face puts the open-model commons under the compute vendor that sells the silicon it runs on, while Mistral's record €21B-valuation raise argues the opposite case for sovereign, open-weight AI in Europe. Underneath the money, the day's technical thread is verification breaking down everywhere — agents that cannot use testing techniques, an AI detector desk-rejecting 178 NeurIPS papers, and 512-bit CA keys falling to a desktop GPU in 32 hours.
- Nvidia acquires Hugging Face for $12.9B —
- Factoring the RSA keys of a 1990s certificate authority —
- Mistral raises €3B at a €21B+ valuation —
- How well do coding agents actually use testing techniques? —
- OpenAI's chief scientist: no lab has solved alignment well enough to keep scaling at full speed —
- git.kernel.org spends more CPU serving scrapers than serving developers —
- NeurIPS desk-rejected 178 papers on AI-detector scores, with no appeal —
- "Google Jail": new domains can't get indexed past the front page —
- 89.6% of European companies using a CDN are behind Cloudflare —
- Broadcom pulled public VDDK downloads, breaking VMware exit tooling —
- Jellyfin 12.0 drops the "10." prefix, moves to .NET 10 —
- D2's proprietary layout engine TALA is now open source —
- LLM-guided program evolution breaks 10 Packomania records for $28 —
- MiniCPM5-2B tops every open model under 4B parameters —
- Currency charts flag downside risk for chip stocks —
- Task-aware quant claims 99% of BF16 reasoning on Qwen3.8-27B at 15% the size —
- Australia will require platforms to offer a non-algorithmic feed —
- llm 0.35 adds GPT-6 Astra support —
- Chip paper roundup: HBF for LLM inference, 2nm M3D SRAM, GaN superjunctions —
- Northland upgrades Intel to Outperform on foundry and server-CPU shortage —
- Yen hits a six-month high, and the move may not be done —
- Morgan Stanley: AI adopters, not AI builders, are the next leaders —
- Custom llama.cpp build for one or two 7900 XTX cards —
