2026-09-29 · America/Los_Angeles · 社区动态 · #18

Jev takes 70 to 500 ms to decide what to do in Doom. My model takes 10 ms, on a single CPU core

Yes, really: CPU only. No GPU, no remote API. Just a small text encoder, a few dozen megabytes, trained in a few minutes. It's built on indecis (French for "indecisive"), a Go library I'm releasing under MIT. It fine-tunes a small text encoder to answer typed questions, yes/no, multiple choice, a scale, with probabilities, the same way Jev does. I ran 20 rounds of ViZDoom's defend_the_center scenario. Results: +21.8 points on average. Matches the scripted bot that generated the training data. 10 ms median latency, HTTP included. Zero missed decisions. ~65 MB of memory. 11 minutes of training…

热度 44.8 / 100；排名与评分保留该期记录。