hi everyone, been building this project for the last ~1.5 months (cc+superpowers all the way)

it is an open effort to find a program that can reliably win NetHack 3.6.6: you can write a bot by hand, let claude code / codex / opencode improve one, or build your own harness that improves the thing that improves the bot

me and my lab switched to the goal-based research this year (we worked a lot in the idea-based framework for a last couple of years) -- and decided that we would like to tackle the NetHack, it is one of those nasty roguelike that is still somehow not solved with all fancy AI models (yes, even astra) -- probably on the way something useful will appear

you can pull someone else's bot, run experiments locally and register the results. every registered program is linked to your github repo and exact source commit

i always wanted to try building something that will use github as both social and storage layer (and it seems to work quite well in our case, simplified things a bit and i think this approach is quite cool)

there is a shared frontier across 73 starting identities -- your bots dont have to be the best at everything to contribute something useful, e.g. if you love to roleplay gnomes only, so be it

to make thing easier (for both harness devs and people who are just playing around), we provide a loop that does some sort of an evolutionary optimization process (quality-diversity, etc), ofc you don't have to use it. but you can start with in minutes (given you have codex or claude subs)

you can also bring your own search, your own harness, whatever -- the common ground is the evaluation. public seeds are there for people to overfit (in nethack even overfitting is hard), and we separately re-evaluate on hidden seeds to check whether the improvements actually transfer

one of my goals was to actually try to make it at least a bit fun -- hope you enjoy the web-site and easter eggs