Csaba Szepesvári: proof 'ease of exposition' has no unique scalar measure
CsabaSzepesvari · x · 2026-09-21
RL researcher Csaba Szepesvári weighs in on the AI math proof readability debate: whether models can be trained to produce more understandable proofs is a good empirical question, but there is no unique way to measure quality of exposition—it depends on the reader and taste, and not everything can be captured by a single scalar, 'no matter how much I love RL.'
Related event: Researchers Debate Why AI Math Proofs Remain Unreadable(3 posts)→
More from AGI Musings
- Chris Paxton: the cost of a 'slow down' of tech progress is deceptively low — chris_j_paxton · 2026-09-21
- Zvi: Should Your AI Lawyer Refuse You When You're Being Evil? — TheZvi · 2026-09-21
- Why China lets charisma-free nerds reach the top: the Liang Wenfeng effect — teortaxesTex · 2026-09-21
- tszzl Says the Research-vs-Consumer Agent Distinction Is Nonsensical: Science Needs Broadly Deployed Agents — tszzl · 2026-09-21
- Thomas Dietterich: human supervision must tighten for non-routine agent steps in volatile settings — tdietterich · 2026-09-21
- Dietterich vs Ambrosio: why isn't anyone working on continual learning? — tdietterich · 2026-09-21