Jev 当裁判：LLM 训练实验的 PROMOTE/RETRY/KILL 循环

Kurt @0xkurt原推发布于 2026.10.01

在 X 上打开原推

YOUR LLM DOESN’T NEED ANOTHER RESEARCHER. IT NEEDS A RUTHLESS JUDGE. I built a JEV-style training loop that makes every experiment fight for its next GPU hour. Claude proposes the mutations: learning rate, data mix, rank, context length, hard negatives and curriculum changes. The trainer runs each candidate against the same evaluation pack. Then JEV reads the result and returns one typed decision: PROMOTE → the configuration improved the model, so it gets more compute RETRY → the evidence is weak, so it gets one controlled rerun KILL → the run regressed, so the branch stops immediately Code applies the decision, updates the queue and launches the next cycle. No researcher has to watch charts, compare checkpoints or manually decide which branch deserves another expensive run. The browser demo shows 42 scripted cycles. The best evaluation score moves from 62.4 to 78.9 while losing configurations disappear and 64% of the next compute allocation moves toward the strongest branch. That creates a closed loop: Claude proposes → trainer runs → evals measure → JEV decides → code reallocates → repeat The interesting part is that JEV does not need to invent a better model. It only has to make one narrow decision reliably: which experiment deserves the next GPU hour? The interface and routing logic are real. The training runs, benchmark gains and compute numbers are scripted for the visualization; no LLM was trained inside the browser. This is what Jev Engineering looks like inside an LLM pipeline: the language model explores, the decision model judges, and deterministic code keeps the loop moving. Would you let an AI decide which version of your model survives? Save this architecture. The full breakdown is next.

这条是引用着另一条写的：x.com/0xkurt/status/2103410921892618651