Dan Luu: Being a 'Meat Proxy' for an LLM Now Sort of Works, Which Is the Actual Problem

Dan Luu / Hacker News (198pts, 156 comments)·medium signalXBlueskyLinkedInCopy link

Luu has tracked people turning their brain off while using LLMs since early 2025, taking an action on faith and assuming it worked. His update is that the failure mode got harder to see rather than going away: brain-off development in September 2026 produces software that sometimes almost works, not software he would want to use or call successful. He borrows Niklas Gruhn's phrase "meat proxy" for the human who relays errors back to the model without understanding them, and argues no capability level makes the approach safe.