Tim Dettmers’ dlab Open Source Week teases local inference (Qwen 3.6 35B-A3B at ~450 tok/s on Mac at 1.5 bpw; 125B on one 24GB GPU; DeepSeek V4.1 550B on 128GB boxes) plus CliffCompaction that cuts agent cost ~50% vs Claude Code/Codex-style compaction. Two open-source projects and four papers drop as one ecosystem, delayed one day from the Sept 21 post.
An independent Advisory Group on Mathematics and AI, hosted at IAS and announced on Terence Tao’s blog, will advise labs on releasing math results without pay from companies. Its first job is helping OpenAI coordinate disclosure of many significant results an internal model reportedly produced…
Maharshi Patel’s essay arguing MCP was built for weaker 2024 models and should yield to shell plus HTTP APIs hit Hacker News with 320 points. Critics including Simon Willison say MCP still matters for auth boundaries, audit logs, and agents without unrestricted shell access…
Anthropic marked a major incident resolved after elevated errors on Claude Mythos 5.1, Fable 5.1, and Opus 5 from about 00:50–02:10 UTC on Sept 22 ( 80 minutes). Status updates said Mythos/Fable recovered first; Opus 5 followed before the all clear…
xAI released Grok 4.7 for coding and knowledge work at $2/$6 per million tokens (same as 4.6), available today in Cursor, Grok Build, and the API. Vendor figures put it at 46.3% on CursorBench 4.0 and 71.0% on DeepSWE v1.1; a fast variant costs double for twice the output speed…
Amazon cut off Meta’s Muse personal AI agent from shopping on Amazon.com after Meta declined to exclude the store, citing Conditions of Use and concerns that the agent browses without identifying itself while using stored customer credentials. The move follows Amazon’s lost Perplexity injunction and tests whether platforms can fence out outside agents via contract terms rather than anti hacking law…
Spotted something we missed? Start a thread.