This is a follow-up to https://manifold.markets/NLeseul/will-an-artificial-agent-ascend-in.

While a program called BotHack did successfully ascend in 2014, it made use of a strategy called "pudding farming" which allows it to amass all items necessary for ascension in a reliable way. More recent versions have made changes to the game rules so that players can no longer benefit from pudding farming and need to use more creative strategies to gather items.

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.

As the authors of a report on the competition's results observe: "In the past, these bots have often made progress by exploiting bugs or weaknesses in the earlier versions of the game that have since been removed by the game developers. To the best of our knowledge, no bot has ever ascended in the most recent version of the game, NetHack 3.6.6."

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 3.6.6 or later 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.