Astra's Qwen math fix runs on Python too
The headline result from GPT-6 Astra — that it fixed Qwen's math ability by inspecting raw weights — mixes distinct experiments. The Qwen gain used Python programs and test-time runs, while only small toy networks were read directly as floats. — honnibal.dev
Ocado's AI reply earns a human apology
A customer who sent Ocado a detailed bug report got a sycophantic AI form letter back; after scoring the follow-up survey at rock bottom, he received a personal apology and a promise that the issue was passed to the technical team. — shkspr.mobi
Revolut's GLM switch lifts engineering lead times
Revolut moved its thousands-strong engineering team to self-hosted GLM for coding, pushing frontier models to an exception list; engineering lead time improved 70% once AI tokens stopped being rationed. — twitter.com
ThinkingCap Qwen quantized for local use
ThinkingCap, a Qwen finetune that keeps answer quality while reasoning in far fewer tokens, is now available as GGUF quantizations for local LLM runtimes, including vision support. — huggingface.co
Lightcone funding priorities target extinction risk
A Lightcone Commons grant evaluator plans to route money toward a grassroots movement against superintelligent machines, an effort to brief elites on AI risk, and organizing by AI company employees. — lesswrong.com