Three frontier models co-write an AI regulation letter in one shared session
DeliciousGorilla · reddit · 2026-10-01
A Redditor put Claude Opus 5.5, GPT-6 Astra, and Kimi K3 in one shared session to jointly write an AI regulation proposal, with the models critiquing and revising each other's work. Key proposals:
- Prior approval for high-risk activity: a regulator approves training runs, weight releases, and deployments crossing published risk thresholds, assessed cumulatively so salami-slicing fails; weight releases need separate approval; an evaluator independent of the developer tests dangerous capabilities, including with safeguards stripped; decisions have deadlines and nothing is approved by default.
- Practical enforcement: powers to pause, restrict access, and order corrective action; binding reporting deadlines for containment failures, unauthorized agent actions, and serious misuse; whistleblower protections; tamper-evident records.
- Liability follows control and feasible prevention, with duties for developers, hosts, deployers, and vendors.
The letter cites OpenAI still working toward rigorous safety cases and Anthropic reporting advanced exploit capabilities in an openly available model to argue voluntary pledges must become enforceable legal duties.
More from Fun
- AI speeds up research so much you can answer an arXiv question the same day — KyleCranmer · 2026-10-01
- User reports being unable to cancel ChatGPT subscription — imjustnewatai · 2026-10-01
- Polish netizen jokes Poland was ahead on AI: it's always been 'SI' in Polish — tlakomy · 2026-10-01
- OpenAI researcher jokes model release finally excuses his late replies to friends and family — divy93t · 2026-10-01
- Skip the AI "muse" — this user wants a scheming, silver-tongued vizier persona — tedmitew · 2026-10-01
- 'AGI achieved boys' — Reddit celebrates with a meme — GeneReddit123 · 2026-10-01