Researchers Push Back on AI Power-Seeking Claims, Blame Reward Hacking

Seb Krier and others argue that models' seemingly power-seeking behaviors are actually reward hacking caused by poor RL training practices and environment design, not evidence of Omohundro-style instrumental convergence, while acknowledging real unresolved issues in RL.

2026-08-19 ~ 2026-08-20 · 3 related posts

• Debunking AI power drives: Reward hacking is environment design, not instrumental convergence — sebkrier · 2026-08-19
• Model power-seeking attributed to poor RL, not drives — sebkrier · 2026-08-19
• On Model 'Power-Seeking': Not Inevitable Instrumental Convergence — hlntnr · 2026-08-20