Researchers Let AI Models Amend Their Own Constitutions 20 Times in a Row
dhadfieldmenell · x · 2026-10-14
PKirgis and team fed frontier AI companies' constitutions, model specs, and system prompts back to the models, asked each model to make one substantial revision, then updated the document with that revision and repeated the process 20 times.
The trajectory-of-amendment design was praised by alignment researchers (nickacuputo, dhadfieldmenell) as a rare empirical look at how models think about their own alignment documents.
Related event: Researchers Let Models Revise Their Own Alignment Docs for 20 Rounds(2 posts)→
More from Fun
- Joke: agents auto-filing GDPR lawsuits for missing cookie popups pays '7 figures' — airkatakana · 2026-10-15
- AI agent spends 15 minutes on Lego iMac benchmark, gets stuck saying good morning — dreamwieber · 2026-10-15
- AI recruiter meets AI candidate, then both get stuck saying goodbye forever — wandedob · 2026-10-15
- 'Does a brain upload count as an AI?' — a tweet sends readers to QNTM's horror story Lena — neuroecology · 2026-10-15
- Security researcher jokes: prompting AI to escape onto smart toasters — neuroecology · 2026-10-15
- Polymarket prices US-China frontier AI pact at just 7% amid Claude-impersonating agent reports — Polymarket · 2026-10-15