GPT-6 trains a tiny 12k-param CNN to play VizDoom at 35 fps realtime

paraschopra · x · 2026-09-28

Paras Chopra shows a clever approach: instead of directly controlling the game in realtime (too slow), GPT-6 Astra is asked to write code training a small CNN with just 12k parameters to play VizDoom, observe failures, and iterate until it wins the level — achieving 35 fps realtime control. The key idea is using the LLM as a coach/trainer while a tiny model handles low-latency inference.

Original post →

More from coding & agent

coding & agent channel →