Jay Kruer: Navier-Stokes was the best case for LLM autonomy, not the general one
In a 2026-09-15 post that reached 261 points on Hacker News, Kruer argues frontier labs are priced on a full knowledge-worker-replacement narrative while the models still need laborious oversight on simple tasks, working only within a small neighborhood of trained tasks and failing outright under small perturbations. His sharpest evidence is economic rather than technical: software firms keep employing and hiring bottom-quartile engineers who lose to models on benchmarks, which suggests benchmarks are not measuring the thing firms buy. He also borrows from hardware, where a typical CPU project carries roughly three times as many specification and validation engineers as design engineers, to argue the specification cost is the real bottleneck, and points to the xz backdoor and the UMN hypocrite commits as proof that expert review is itself compromisable.
Source
↳ Follow the thread