Oh man, this takes me waaay back to my time spent on rec.games.roguelike.nethack about two decades ago. It literally took me a couple years of off-and-on effort (and hundreds of deaths) until I ascended for the first time. But once you put that much actual effort into getting good at it, you stay good, and I ended up doing multiple back-to-back ascensions, ascending with every class, and doing some high conduct runs, including one that was vegetarian, weaponless, genocideless, self-polymorphless, and artifact-wishless.

It's definitely a hard nut to crack, and I can see why the machine learning attempts would've struggled so much, but honestly I'm surprised that the best symbolic logic bots only achieved a few thousand points on average. That's really not even getting very far into the game.

I guess NetHack is a harder game than StarCraft II, as with SC, the best bots are competitive with the best humans (albeit they have advantages in inputs/attention and disadvantages in grand strategy). NetHack though is literally turn-based (like Chess), so there's no advantages to be had here, and clearly they're struggling for it.