Late in 2021, a competition was held at the NeurIPS conference to see if any team could build an agent capable of completing the well-known roguelike game NetHack. It also aimed to compare approaches using neural networks to those using symbolic logic. Ultimately, there were no complete runs of the game by any agent submitted, and symbolic algorithms significantly outperformed neural networks on the challenge's scoring criteria.

Machine learning, especially approaches based on neural networks, have made significant progress since then. Before the end of 2024, will anyone build an agent, neural network-based or otherwise, that successfully completes NetHack at least once?

Resolves to NO at midnight EST on December 31, 2024. Resolves to YES if, at any time before then, someone posts a link to evidence that this has been accomplished.