AI Improves AI on Its Own — Right on the @ONcompute🧐 An autonomous research agent, inspired by Karpaty’s autoresearch, has been added to the official tutorials. The local Qwen3-14B runs on the same H200 as the training: it reads the experiment history, formulates hypotheses, rewrites train. py - and keeps only the improvements. Up to 200 iterations, without a single API key. Autonomous ML research on decentralized hardware. It’s no longer science fiction.