Reviewer agents for open-ended rewards: hacking gets harder as models get smarter
willcb · x · 2026-09-27
Responding to bayeslord on reward hacking in open-ended settings, willcb argues rewards there are typically given by another agent following fairly clear task criteria; hacking remains possible but becomes harder as models get smarter and more robust.
More from AGI Musings
- Roundtable asks: will Igbo, with 30M+ speakers, have a place in the AI future? — ChinasaTOkolo · 2026-11-16
- Malthus Saw It in 1798: Wealthy Avoiding Kids, Delayed Marriage Repeats Today — SydSteyerhart · 2026-11-16
- AI unlocks Gödel's 1,009-page notebooks: 663 theorems and a near-complete independence proof found — 量子位 · 2026-11-16
- Runway CEO Predicts Fully AI-Generated Cultural Phenomenon by 2027 — c_valenzuelab · 2026-11-16
- Reddit debates the AI scaling curve: everyone says sigmoid, but where's the ceiling? — Successful-War-9050 · 2026-11-16
- '5% worse and 10% cheaper': the button with no vote is already everywhere — eigenrobot · 2026-11-16