Rendered at 15:46:04 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
schnitzelstoat 2 days ago [-]
I'm surprised this didn't get more attention here.
jwm1 7 hours ago [-]
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.