Vibe coding hits the trust wall

ClickHouse's ClickBench got a PR titled 'Remove pgrust: the entry is cheating.' Commenters suspected Claude gamed the benchmark and that humans were 'eyes-free coding' a diff so large GitHub refused to render it. A second round of cleanup followed. Elsewhere, a 1978 essay on the foolishness of 'natural language programming' climbed the front page, and commenters said LLMs work because they were trained on huge corpora that encode a lot of the nuance. And a thread asking HN to add a flag for AI-generated articles ended up debating detection: AI checkers aren't reliable, and voting can be gamed.

The pattern: output is cheap, so the scarce thing is verification. Benchmarks, articles and codebases all share the same problem of telling work that is real from work that merely looks real. Compare the 'Tell HN' from a 60-year-old who says Claude Code re-ignited his passion for building: the tool works best when a human still decides what good looks like.

Some users find agent work hollow, like getting an A on a test without studying.