Stanford professor uses Jev for a human-free LLM training loop
The system trains on Opus 5.5, with Jev reviewing results to decide the direction of the next run, giving more compute to good configurations and discarding bad training runs.
a Stanford professor built a Jev system that can train LLMs without any human researchers and sold it for $13,780 the system running on Opus 5.5 continuously trains different versions of a language model, evaluates their performance, and decides what should change in the next training cycle. Jev looks at the training results and decides which direction the next run should take. a configuration that improves the model gets more compute. a bad training run gets discarded. and promising changes trigger new experiments automatically. the system essentially turns LLM training into a self-improving feedback loop. I found a breakdown of the architecture they used, and the way Jev decides how the next training cycle should run is surprisingly clever. going to post how the system works next, so stay tuned if you want to see what Jev Engineering looks like inside an LLM training pipeline. would you let an AI decide how your LLM should be trained?
This post quotes another post: x.com/me_barnyx/status/2101976764779999381