Studies suggest smarter AI systems become less predictable
skdh · x · 2026-11-22
Sabine Hossenfelder points to several recent studies indicating that the more capable AI systems become, the less predictable their behavior is — a counterintuitive finding with direct implications for alignment and safety, since capability gains don't translate into controllability gains.
More from AGI Musings
- 'There Are No More Moats': AI Creator Urges Sharing Workflows — AIandDesign · 2026-11-22
- Hugging Face CEO: Answer to Chinese Open-Source AI Is More US Open Source — coherence · 2026-11-22
- A 'Regime Change in Earth Intelligence' Is Coming — We Can Only Nudge Its Trajectory — danfaggella · 2026-11-22
- AI Commentator Says Progress Speed Gives Him Cold Sweats Ahead of New Gemini Model — iruletheworldmo · 2026-11-22
- François Fleuret and Charton Clash Over Whether Machine-Solved Math Changes Mathematics — francoisfleuret · 2026-11-22
- Math group slams Anthropic's AI-generated Riemann proof: no human can vouch for it — appenz · 2026-11-22