gpt-6VIDEO
GPT-6 Astra trains small CNN to play VizDoom at 35 fps
Rather than controlling the game in real time, the author had Astra write code to train a 12k-parameter CNN, observe failures and iterate until it could win the level at 35 fps.
Paras Chopra
@paraschopraArchived 2026.09.29Originally posted 2026.09.28
Paras Chopra @paraschopraOriginally posted 2026.09.28
This is GPT-6 Astra playing VizDoom at 35 fps. How? Isn't the model slow to do any realtime control? Instead of playing the game directly, I asked it to train a small CNN (12k parameters) + code to play the game, observe failures and iterate until it could win this level.
x.com/paraschopra/status/2104414175049269633Archived 2026.09.29 10:16