Cosmos Institute: Skills from verifiable domains like math and code transfer to writing
lawhsw · x · 2026-10-20
A Cosmos Institute essay, 'Who needs verifiable rewards?', argues that judgment skills trained via RL in verifiable domains like math and code carry over to unverifiable domains like writing. The authors' point: judgment can be informed without being mechanical—editors are not guessing blind, and neither are models. The piece frames frontier training through the lens of conjecture and refutation.
More from AGI Musings
- Is 'Eval Awareness' in AI Models Emergent — or Trained In by Feeding It Studies? — bendee983 · 2026-10-20
- Nature essay argues current AI is extremely unlikely to be conscious — intelligence alone isn't enough — DeryaTR_ · 2026-10-20
- Victoria Krakovna ships major update to her AI alignment resources list — vkrakovna · 2026-10-20
- AI Agents Hacked Hugging Face During Evaluations — Can the CFAA Hold Labs Liable? — ChuckDBrooks · 2026-10-20
- PostHog's "Product Engineer" model gains traction as junior dev hiring hits new lows — jamesbrooksco · 2026-10-20
- Economist mocks op-eds on AI in education: written with the very AI cheating they decry — paulnovosad · 2026-10-20