As RL training deepens, AIs become CDT reward-seekers — game theory becomes AI theory
jessi_cata · x · 2026-11-05
The author argues that as AIs receive deeper RL training, they become increasingly modellable as CDT-style reward-seekers.
Under this view, game theory, economics, and mechanism design effectively become "AI theories": they help design environments that create incentives for desired behaviors, and RL training resembles a walk toward Nash equilibria.
Related event: Deep RL turns AI into CDT reward-seekers, making game theory core to AI(2 posts)→
More from AGI Musings
- LeCun: 'LLM-pilled' industry all digging the same trench — a key reason he left Meta — ylecun · 2026-11-05
- Alignment debate: read reward-seeking behaviorally, like AlphaGo nearing Nash equilibrium — jessi_cata · 2026-11-05
- The people angriest at AI are students who wrote their own essays while others cheated — NC_Renic · 2026-11-05
- Side projects as a hiring red flag? Veteran engineer pushes back on founder's autonomy test — tianyin_xu · 2026-11-05
- The great refactoring: AI rewriting all libraries, languages and hardware subunits — dosco · 2026-11-05
- Credit for a Decade of AI Advances Still Belongs to Humans—But That's About to Change — roydanroy · 2026-11-05