gpt-6VIDEO
NetHack challenge asks coding agents to write the program that wins
GPT-6 Astra reached 13% playing NetHack directly and an RL agent 17%, so the author is inviting people to use coding agents to write a program that beats the game.
Vladislav Kurenkov
@vladkurenkovArchived 2026.09.24Originally posted 2026.09.24
Vladislav Kurenkov @vladkurenkovOriginally posted 2026.09.24
NetHack: the hardest game no AI has beaten. GPT-6 Astra hit 13%, an RL agent 17%, both playing. We wanted to know: can a model write the program that wins, and how do we accumulate progress as a community? Join us with your coding agent to find out. https://nethackers.dunnolab.ai
x.com/vladkurenkov/status/2102864659413872889Archived 2026.09.24 10:16