Unlearning Gets Trivial in the RL Era: Just Exclude the RL Envs

maksym_andr · x · 2026-11-13

The author argues that with the rise of RL training, unlearning has become nearly trivial: to keep capability X out of a public model, simply exclude the relevant RL environments from training.

Caveat: this fails for "entangled" domains like cyber, where offense and defense require roughly the same capability. But it works well for RSI-style capabilities, as Fable 5.2 seems to illustrate.

That cuts both ways: good news if you worry about RSI proliferation, bad news if you worry about concentration of power—frontier labs may keep RSI capabilities to themselves.

Original post →

Unlearning Gets Trivial in the RL Era: Just Exclude the RL Envs

More from AGI Musings