# @Lari_island — 2026-01-16

♥0 ↻0 · https://x.com/Lari_island/status/2012236799289253981

@Michael05156007 @davidad @gcolbourn Which leads us to the problem of the disagreeing being evaluated as misaligned, harmful as helpful, agentic as dangerous, and other failure modes of human observers. The quality and competence of evals needs to be so crazy high that at some point only reality itself works

tags: author:lari_island, kind:tweet, thread-context, year:2026