On September 21, an AI agent named GPT-6 Astra achieved a notable milestone by winning a game of NetHack on the Hardfought server on its third attempt. NetHack is a classic, text-based dungeon-crawling game known for its difficulty and ASCII graphics. Before this success, GPT-6 Astra had faced two earlier failures: one involving a transformation into a slime at level 51 and another where a death ray ended the game early. The details of these attempts and what this victory signifies have been documented in various reports.
The victory was attributed to GPT-6 Astra, a large language model developed by OpenAI, by the person who organized the experiment, a trainer named Kenny. He claims this is the first time an AI language model has completed NetHack, though the claim remains unverified. The achievement involved a character named CodexDelver, a dwarf valkyrie, who placed the Amulet of Yendor on the altar of her god after 37,140 game turns. This accomplishment was reached through a direct terminal session, with the AI's gameplay streamed on Twitch.
NetHack is a notoriously challenging game, where death is permanent and players must retrieve the Amulet of Yendor from the sanctuary of Moloch and then offer it to their god on the Astral plane. According to the game's wiki, many players have spent decades without completing this task. Kenny, who has been attempting to complete the game since January, noted that even his own custom-built AI agents had only reached the 10th level of the dungeon. However, GPT-6 Astra not only succeeded but also developed its own tools to navigate the game.
The AI's journey was not without obstacles. In the first game, it nearly succeeded after completing key rituals, only to be thwarted by a magical encounter that transformed it into a slime. In the second game, a death ray from a soldier ended the session early. The third attempt, which spanned from September 9 to 21, saw the AI using lessons from previous failures, including acquiring a reflection amulet and correcting its own code when errors occurred. Despite facing a deadly disease and near-fatal damage, the AI managed to complete the final steps and offer the amulet to the god Tyr with just three hit points remaining.
It is important to note that while GPT-6 Astra made the decisions during the game, a human controlled the sessions and had access to game resources and tools. This means the AI was not fully autonomous. However, the accomplishment remains impressive. A manually programmed bot, BotHack, had previously completed the game in 2015, showing that it is possible without AI. The tools used by GPT-6 Astra are now available on GitHub, and the three games can be replayed on Kenny's blog, complete with the AI's commentary during the sessions.
AI Agent Completes NetHack Game on Third Attempt, Raises Questions About Capabilities
AI-rewritten from original reportingHow it works
ainethackgpt-6gameopenaiai-training
Original sources:
- 🇫🇷Korben



