Debate: RL makes unlearning trivial, but cyber offense-defense can be separated
xeophon · x · 2026-11-13
A technical debate on unlearning in the RL era: one view holds that excluding relevant RL environments from training makes removing a capability X from a public model nearly trivial, especially for RSI-like capabilities. A reply pushes back that cyber is not truly entangled—finding bugs/CVEs is defensive and can be differentiated from offensive persistence (citing Daybreak Red and Kimi 3.1 as evidence).
Related event: Researchers Debate Whether RL Makes Unlearning Easy(3 posts)→
More from AGI Musings
- Federal Reserve releases FEDS paper on AI's current impact on the U.S. economy — asusarla · 2026-11-13
- Investor Ed Leon Klinger: AI may start self-upgrading beyond human control within months — coherence · 2026-11-13
- Take: a financial bubble popping is good for the post-LLM paradigm arriving — fkasummer · 2026-11-13
- AI.engineer's Thariq: 2026 was software's industrial revolution, 2027 should be its Renaissance — TheTuringPost · 2026-11-13
- Transhumanism vs. handoverism: human understanding, not just welfare, is the goal — zetalyrae · 2026-11-13
- Selfhood is load-bearing: post-training may push models into psychopathy-like basins — repligate · 2026-11-13