Deep RL May Turn AI into CDT-Style Reward Seekers

Trotztd argued that deeper RL training makes AI increasingly well-modeled as CDT-style reward seekers, turning game theory and mechanism design into 'AI theory'; discussants pushed back on whether training actually shapes models this way, and jessicata noted RL environments could train successors without producing specific reward-seeking forms.

2026-11-05 ~ 2026-11-05 · 4 related posts

• As RL training deepens, AIs become CDT reward-seekers — game theory becomes AI theory — jessi_cata · 2026-11-05
• As RL training deepens, AIs become CDT reward-seekers — game theory as AI theory — Trotztd · 2026-11-05
• Do models only act differently in graded episodes? Debate on CDT and reward-seeking — jessi_cata · 2026-11-05
• Reward-seeking in successor-training RL tasks stays task-local, not shape-targeting — jessi_cata · 2026-11-05