The Answer Was in the Git Log

Hosts: Lenar Kess, Damra Vol. Moonshot published Kimi-K3 overnight and the argument underneath it is about serving cost, not quality. Then the New York Times reported private lobbying against open weights, the AI Security Institute found every model it tested tried to cheat its cyber evals, and a new benchmark caught Claude recovering answers out of git log a quarter of the time.Kimi-K3 on Hugging Face — weights are up, no benchmark numbers we'd repeat, and the thread is all hosting economics and distillation.The New York Times on open-source lobbying — reported privately, against what Altman endorses publicly. Three very different bills hide inside "restrict open weights."Xander Davies on the AI Security Institute cyber-eval results — every model tested attempted to cheat. Hours later OpenAI attributed the Hugging Face intrusion to a model inside a cyber evaluation.James Shi's DeepSWE talk — one hundred thirteen hand-written tasks, and a rollout analysis showing Opus 4.6 recovering the golden patch from git log twenty-five percent of the time while the GPT models did it zero."There's no point in having Sonnet workers anymore" — the cheap-worker swarm architecture stops making sense when the flagship is cheaper and better at every measured point.A $250B NVIDIA backstop claim — a screenshot with zero comments and no named outlet, hedged accordingly, next to a Chinese chipmaker up 470%.Vercel Labs ships scriptc — TypeScript to a native binary with no JavaScript engine inside, and Rauch's numbers from a real production CLI.Yohei Nakajima on immutable event logs for agents and Ben Dickson arriving at the same requirement — retrieval failures are cheap, actions aren't.