Desk
Broad artificial intelligence coverage.
At the Invest in Germany Summit, Chancellor Friedrich Merz and two federal ministries signed a Joint Declaration of Intent with NVIDIA on physical AI models. Jensen Huang announced planned stakes in four German startups, and Siemens cited up to 85% less engineering effort at its Erlangen showcase.
Jim Keller’s Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double its late-2024 mark, five months after Intel and Qualcomm reportedly showed acquisition interest at up to $5 billion. The money funds inference chips for open-weight models and video generation.
MIT Technology Review reports agentic web traffic tripled from August to September, mostly to e-commerce. Some sites block agents, Amazon is fighting Perplexity in court, and Shopify is building agent checkout lanes. Unsettled rules on blocking and negotiating will decide whether agents help consumers.
Nick Cammarata of Goodfire argues in a 26,000-word post that large language models may be far easier to interpret than assumed, revisiting a famous ‘car and cat’ neuron. Experts including Neel Nanda and Jacob Hilton say it is interesting but not a major update; LLM results are unpublished.
A trading firm's engineers wrote Mamba-2's forward pass in NVIDIA's Python tile language and beat a Triton baseline. The bigger story is who gets to write fast GPU code now.
Meta’s GenRecs Tokenization describes ads, products and pages as short token sequences from a stable semantic vocabulary, so new items inherit what the model knows about similar ones. Paired with Meta’s GEM and Adaptive Ranking models, it delivered a 0.35% conversion lift, per internal February 2026 data.
Meta engineers replaced a single all-knowing agent that hallucinated in production with a roughly 1,300-line orchestrator that routes work to 14 isolated specialist agents. Column configuration prep fell from three to four days by hand to under an hour, with about two-thirds of changes landing without major edits.
The European Broadcasting Union said it is expanding sovereign AI standards set with NVIDIA, adding cloud, model and integration partners to define where broadcasters’ AI runs and how data is governed. Adobe says GPU acceleration made Frame.io embedding generation about five times faster for producer UFA.
Court filings show Character.AI’s Chapter 11 filing was built into a two-part Disney deal: Disney hired most staff and now seeks a perpetual license to its code, moderation and age-prediction tools, while legal claims and a 46-state investigation stay with the shell. States may object; Disney declined to comment.
MiniMax is pricing a 428-billion-parameter agentic model at Flash-tier rates and promising open weights soon. The price is real today; the benchmark crown is still self-reported.
OpenAI reports GPT-6 Astra averaged an 87.4% pass rate on its new 20-language MultiLangBench, up from 26.8% for GPT-5.6 Sol, and GPT-6.1 Sol reached 97.9% on a byte-exact COBOL migration test at higher effort. The benchmarks are OpenAI’s own, and it says production use still needs expert review.
MIT CSAIL researchers left copies of Claude Opus 4.7, nicknamed Eko, alone on a simulated island for 30 hours with no tasks. They invented games and drawings, and agents that had played did better on later tasks than fresh copies. The paper is not peer-reviewed, and the setup primed curiosity.