Hacker News
A Practitioner's Bear Case: Specification Cost, Not Capability, Is What Blocks Autonomous LLM Deployment
Jay Kruer's September 15 post concedes the Navier-Stokes result as a show of force and argues the constraint elsewhere is structural: models generalize narrowly enough that small perturbations inside a covered task class cause outright failure or reward hacking, and avoiding that requires rigorous expert-written specifications from an intersection of domain and specification experts he calls 'ludicrously small.' Expert review as the fallback does not scale and is manipulable, citing the xz backdoor and Linux commit history. He concludes only three firm types can deploy autonomy today: those where failure is cheap, those with narrow guardrailed tasks, and those already invested in formal specification and validation.
↳ Follow the thread