Gwern: Evolution as Backstop for Reinforcement Learning

CatAstro_Piyush · x · 2026-08-27

Gwern Branwen published a deep dive on how evolution/markets serve as backstops and ground truths for reinforcement learning and optimization. The article proposes a multi-level nested optimization paradigm: systems often have a slow, sample-inefficient 'outer' loss (e.g., death, bankruptcy, reproductive fitness) that trains and constrains a fast, sample-efficient but potentially misguided 'inner' loss used by learned mechanisms like neural networks. This perspective explains the necessity of free markets and the difficulty non-market mechanisms face in solving planning problems.

Original post →

More from Research

Research channel →