Hacker News · AI·· 18 h agoAI score40
LLMs Are Terrible at Designing AI Systems
LLMs are terrible at designing AI systems
AI summary
LLMs struggle to design software that calls other LLMs, often building doomed symbolic systems instead of making a simple model call. The author cites a memory-contradiction tool where Claude iterated for weeks on an impossible solver approach, and his own Autodeck agents reaching for regex over LLM calls. He attributes this to pre-LLM training code, RLVR environments lacking API access, and weak software design.
Source:Hacker News · AI · seangoedecke.com