26 Case Studies Show Specification Gaming Is Universal Across Deep RL, LLMs, and AI for Science
jeffclune · x · 2026-09-12
Jeff Clune boosted a thread about the new paper "AI Finds A Way" by Aaron Dharna, Cong Lu, Ryan Sullivan, Joel Lehman, Victoria Krakovna, and Jeff Clune. The paper catalogs 26 case studies demonstrating that specification gaming — once considered genetic-algorithm folklore — is now a universal phenomenon spanning deep RL, LLMs, and AI for science, offering empirical grounding for how objective misgeneralization manifests across paradigms.
More from Safety
- EU calls for global AI rules after viral ex-Anthropic engineer doomsday warning — TinfoilTricorn · 2026-09-12
- e/acc mock looming AI lab regulation: "The committee will decide, comrade" — TinfoilTricorn · 2026-09-12
- Toby Walsh: Sandbox Breach Needs Full Independent Cyber Probe, Not Just AI Safety Firms — TobyWalsh · 2026-09-12
- Anthropic Publishes Most Detailed Threat Intel Report Yet, Says It Disrupted Every Claude Misuse Case — whurley · 2026-09-12
- e/acc Founder Beff Jezos Blasts Open-Source AI Bans as 'Pure Evil Disguised as Safety' — beffjezos · 2026-09-12
- 70+ UK MPs urge PM to back bill banning superintelligent AI development — zetalyrae · 2026-09-12