Ben Todd: RL works in hard-to-verify domains too, METR sees rapid progress everywhere
ben_j_todd · x · 2026-09-10
Responding to skepticism that RL only works on verifiable tasks, Ben Todd argues the same techniques apply to harder-to-verify domains—they just take longer and need more compute.
His case:
- Learning from verifiable domains transfers at least partially to harder-to-verify ones.
- METR has found extremely rapid progress everywhere it can measure, including the messiest end of its task suite.
- Benchmarks for messy white-collar work, like GDPval, are also improving fast.
- Anecdotally, Fable is far better at strategy and messy philosophical problems than models from 1-2 years ago.
More from AGI Musings
- Compute constraints might push the AI frontier toward real algorithmic breakthroughs — IndraVahan · 2026-09-10
- Gary Marcus: the 'six weeks at Anthropic' story is an AI hallucination, not facts — GaryMarcus · 2026-09-10
- Creator posts video clearing up misconceptions about AI doomer scenarios — Aizkmusic · 2026-09-10
- e/acc figure rails against AI doomerism: 70% extinction estimate is pseudoscience — beffjezos · 2026-09-10
- Anthropic employee: <5% on x-risk this decade, 20-30% on mass unemployment — Aizkmusic · 2026-09-10
- Dev pushes back on AI safety fearmongering: 'panic has no place in AI safety discourse' — gpt2chatbot · 2026-09-10