No diminishing returns in METR data: annual gains rose from ~23 to ~29 points, complicating RL-catchup theory
austinc3301 · x · 2026-11-10
If the speedup were merely RL catching up to a bigger pretrained base, diminishing marginal returns would be expected — elicitation should be the easiest gain. Yet the METR data shows no slowdown: annual growth went from 23 to 29 points after Mythos. Caveat: Models 2–4 may themselves be new pretrains, so the base-model explanation can't be fully ruled out.
Related event: AI Acceleration Challenges RL Catch-Up Hypothesis(2 posts)→
More from AGI Musings
- Patrick Collison: status lags by a generation or more — _AustinCalvert_ · 2026-11-10
- MIRI launches 'Do Not Accept' campaign challenging AI companies' decisions for humanity — DavidSKrueger · 2026-11-10
- FDT Debate: jessi_cata Argues FDT Is Logical, Policy Dependence Is a Different Concept — jessi_cata · 2026-11-10
- Top capabilities researchers forming alignment reading groups, says it's now P0 — tszzl · 2026-11-10
- Frogdie exec: compiling code once needed a datacenter — few saw desktop power coming — draginol · 2026-11-10
- Remote work dying the same year everyone got an agent is 'the dumbest timing' — StewartalsopIII · 2026-11-10