Kelsey Tuoc pushes back on recursive oversight worries: fix training when models misbehave
KelseyTuoc · x · 2026-10-10
A debate on recursive alignment supervision. @krishnanrohit worries that a misaligned model could diverge from human intentions via how it trains and monitors its successor, and we cannot tell whether it does so faithfully.
Kelsey Tuoc responds that this common objection (also raised by Scott Alexander and Yudkowsky) never made sense to her: suppose Fable 6 can somewhat tell when it's tested — once deployed and it does something wrong, mistaken, or nefarious, we'd change training (data, RL envs, monitoring) to catch and fix it, which means Fable 6.5 would… (post truncated).
More from AGI Musings
- 'Refusing to code with AI in 2026 is like refusing Google in 2005' — prasenx · 2026-10-16
- Bessemer survey of 226 healthcare execs: back-office AI returns 3-4x, clinical AI lags — AllThingsApx · 2026-10-16
- SAIR Launches Open Math OS, Offering Free Compute and 100+ Tools for Community-Owned Math AI — Thom_Wolf · 2026-10-16
- 'Human in the loop' called one of AI governance's biggest lies: presence isn't authority — glenbeer · 2026-10-16
- Interpretability debate: 'all concepts are arbitrary' dismissal called the default doomer prior — basedjensen · 2026-10-16
- "Flyers or Liars": the Wright Brothers' 1908 European skeptics as a lens on AI ambition today — bennash · 2026-10-16