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.