RLVR might be disproportionately bad at science
the verification loop for theories can be on the order of decades and centuries, and even then we know today as the better theory can often actually make worse predictions
Blog RLVR might be disproportionately bad at science the verification loop for theories can be on the order of decades and centuries, and even then we know today as the better theory can often actually make worse predictions Dwarkesh Patel May 16, 2026 23 6 1 Share I’m writing up some threads that we explored in my interview with Michael Nielson . That episode was one of my favorites. The organizing question from my interview with Michael Nielson was, “How do we recognize scientific progress?” It’s especially relevant to thinking about what it would take for AI to close the RL verification loo
saved by
related reading
- My picture of the present in AI — LessWronglesswrong.com
- Designing AI for Disruptive Scienceasimov.press
- Terence Tao – Kepler, Newton, and the true nature of mathematical discoverydwarkesh.com
- The Need for Verification in AI-Driven Scientific Discoveryalphaxiv.org
- Could AI slow science?normaltech.ai
- Is Science Stagnant? - The Atlantictheatlantic.com
- When do ideas get easier to find?freaktakes.substack.com
- Questions about the Future of AI - by Dwarkesh Pateldwarkesh.com
- Can AI Solve Science?-Stephen Wolfram Writingswritings.stephenwolfram.com
- theory uplift differentially benefits safety & is underleveraged — LessWronglesswrong.com
- Faster Than Science — LessWronglesswrong.com
- Context Widowsartificialbureaucracy.substack.com