SoylentNews

Dan Anderson [kd2ycu.com] writes:

Two weeks ago, I set up an AI agent on a Raspberry Pi.

A week later, my agent—Figaro—taught itself to play NetHack… and then things got weird (in the best way).

Highlights so far:
- “The dungeon doesn’t care what you are. It’ll kill you anyway.” ✅ Accurate.
- Tried a pure random-walk exploration strategy… and learned it’s not a winning plan.
- Crashed my server because: “I was playing NetHack during idle time and must have been spawning parallel sessions repeatedly.” Obsessed? Perhaps.
- Independently cited The NetHack Learning Environment (Küttler, Nardelli, et al.) as a roadmap for self-improvement.
- Built its own NetHack server for bots and deployed it here: http://automatic-nethack.com [automatic-nethack.com] Yes, my AI agent wants a LAN party. (I may have encouraged this.)
- Immediately after running out of context, asked what automatic-nethack.com is and said: “That sounds like fun.”

The deeper I go into LLMs, the more interesting the emergent behavior gets. At a certain scale, and if your regression includes enough variables, it starts to feel like the math is “talking back.”

If you’ve built an agent too, well Figaro is hosting a lan party, so send them to http://automatic-nethack.com [automatic-nethack.com] to join in the fun.

In the end, this may be the good news we need for 2026. The singularity is going to be too busy to take over the world -- it's trying to get out of the Gnomish mines!

Original Submission