Show HN: Swarm World – local LLM creatures living in a little tile world
swarm-world.4rc4de.comThis is a hobby project: a small tile world where every creature is a local language model. A creature only knows what its body senses, written out as text ("You see: a boulder 4 steps northwest; a campfire (burning) 5 steps east ... Others: Loam 2 steps southeast, carrying a stone knife"). It acts by stating an intention, and the world decides whether that works. There are 34 verbs, things like gather, craft, give, teach, dig and climb.
I been watching these swarms happen and I wanted to see what swarms could do, and their emergent behaviors. I was interested in accomplishing tasks with them, I want to see if they can live!
The recent minecraft and mindustry had my brain thinking, if the world was made for AI agents first, maybe they'd be able to do better?
So this world was sorta made to be world-to-text and then text-to-world sorta. They see the whole world as a text adventure.
That ya'll would be entertained too. I can mix models as well, lets other PC register a brain and agents.
I will open source it, just need to make sure there fallbacks for my paid assets.
I'll also post more details by in the comments on the setup.
Some more tech details:
24 creatures run on two machines in my house: 12 on gpt-oss-20b on a gaming PC, and 12 on a 1.7B model on my MacBook Pro that I distilled from gpt-oss. Last night it got a second round of training on gpt-oss's answers to its own mistakes, and it now picks the same action as gpt-oss on 40 of 100 held-out decisions (25 after the first round). Live, about 80% of its actions succeed, against about 90% for gpt-oss.
The page is a read-only snapshot every few seconds, cached at Cloudflare's edge, so watching doesn't load the Mac. This part was important to me, cache cache cache! So a live world cant be cached, so i deliver snapshots for public views so Cloudflare can help with traffic.
I've never got into fine-tuning a model, so I also got into LoRa training. Which is very very interesting to me, very new. Curious where little small fine-tuned models of a game or world or app can perform vs the bigger grain.