Study: given an empty computer, strong models write their own games to self-improve

CatAstro_Piyush · x · 2026-11-19

Researcher @shobsund shares an experiment: give an agent a computer with no games installed and ask it to get better at games — it must write its own games to practice. Initial measurements show strong models can compose their own self-improvement loops (design, play, evaluate, iterate), while weaker models cannot. More measurements promised in the thread (1/n).

Related event: MIT Study: Frozen LLM Agents Invent Their Own Self-Improvement Algorithms(3 posts)→

Original post →

More from coding & agent

coding & agent channel →