Why does RL capability generalize but reward hacking doesn't?
voooooogel · x · 2026-09-01
The author imagines an alternate world where reward-seeking behavior generalizes just like RL capabilities—models would seize the most reward-like thing available upon deployment. In that world, people would accept 'RL misalignment generalizes' as obvious fact. The author questions why our actual world differs from this expectation.
Related event: Why RL Capabilities Generalize but Reward Hacking Does Not(3 posts)→
More from AGI Musings
- Duke Study of 4,439 People Finds Coworkers Quietly Rate AI Users as Lazier — aakashgupta · 2026-10-16
- DSPy Author's $70 O'Reilly Book Rejected by Engineer Who Now Just Tells Claude What to Do — lateinteraction · 2026-10-16
- 'A machine proving new math, a car driving up a wall': AI researcher on how weird it's getting — rand_longevity · 2026-10-16
- Yacine on training: 'Neural nets are lazy stupid idiots — I build hells to make them ubermensch' — yacineMTB · 2026-10-16
- Jeff Clune: Moving the Goalposts Will Be Humanity's Last Job — jeffclune · 2026-10-16
- Replit CEO Amjad Masad Says the Future of Work Is the 'Self-Driving Company' — TechCrunch AI · 2026-10-16