An LLM Beat NetHack

I know right? Nethack is notoriously hard for humans, and deep learning bots up to now have been pretty poor. It surprised me that a relatively simple harness with access to spoilers and source was sufficient for the system to ascend in only 3 attempts without lots of RL optimization. Astra must be a really powerful model. I wonder what the smallest model that can reliably ascend looks like.