250 agents train a single algorithm: 'program learning' recreates world maps efficiently

willdepue · x · 2026-11-17

Developer cdngdev describes a map-generation model trained not via classic deep learning but a process he calls "program learning": a team of 250+ agents works through a global dataset of places, tuning a single algorithm to recreate them, with grader agents reviewing for hillclimbing.

The resulting algorithm combines distilled classification models and rules-based components, enabling far more efficient generation than fully learned approaches, with some tradeoffs — an echo of symbolic AI in the modern era.

Original post →

More from Fun

Fun channel →