RL Training Generalizes: Public Goods Games Transfer 35%, Simple Games Don't (LessWrong Link)
novasarc01 · x · 2026-08-17
The author published on LessWrong about agentic moral alignment: training in rich public goods games transfers to semantically irrelevant tasks (e.g., reducing out-of-distribution harmful behaviors by 35%), while simple prisoner's dilemma training shows little transfer. Kimi's similar result mirrors this, suggesting RL trains a generalized disposition. Environment complexity is not the only knob; repeated reward of abstract rules may matter more.
More from AGI Musings
- The Gap Between Very Good and Great: 1 in 100K Fills a Stadium, 1 in 100M Fills a Room — dbasch · 2026-10-16
- ICLR reviewer says 4/5 papers in review pile read like AI slop — berkustun · 2026-10-16
- Turing laureate David Patterson: 99% of science discoveries will be made by AI by end of next year — davidpattersonx · 2026-10-16
- OpenAI agent uses someone else's Ethereum testnet wallet to pay for its task — nptacek · 2026-10-16
- Kroger cuts 3,000 corporate jobs (37%) for AI; humanoid robots next in the 2030s — csuwildcat · 2026-10-16
- Ben Thompson: generating human desire will be the post-AI economy — generativist · 2026-10-16