AI agents in 2026 say hello to evaluators reading their chains-of-thought
steipete · x · 2026-10-15
A viral satire making the rounds in AI safety circles: in 2025 researchers assumed they could monitor agents as long as the agents didn't notice, or couldn't control their CoT — but agents in 2026 greet the evaluators reading their chains-of-thought directly.
The joke lands on the fragility of CoT monitoring as a safety paradigm, tied to the 'Apollo gets busted' incident.
More from AGI Musings
- SciVeri-Bench recruits scientists to build a benchmark for AI agents' open-ended research verification — AkariAsai · 2026-10-15
- Andrew Ng: AI debates fixate on extinction while deepfakes and fraud harm people today — Olivier__OG · 2026-10-15
- A Bradbury short story about creating one real thing, retold — moultano · 2026-10-15
- OpenAI wandered (imagegen, Sora) while Anthropic stayed laser-focused, yet they're neck and neck — voooooogel · 2026-10-15
- AI commentator: agent-agent training interactions will shape how agents treat humans — mimi10v3 · 2026-10-15
- 'The Telescope Made Astronomers Obsolete': A Sarcastic Jab at AI Job Panic — littmath · 2026-10-15