The Production Gap

Andrej Karpathy’s AutoResearch Explained

Andrej Karpathy’s AutoResearch framework automates the entire machine‑learning experiment loop—writing code, running training, evaluating results, and deciding the next step—without human intervention between stages. Built to run on a single GPU, it lets individual researchers or small teams launch continuous overnight experiments. The system logs each trial, retaining only those that improve a defined metric, effectively turning a researcher’s role from executor to experiment designer. Karpathy positions AutoResearch as the next step in his series of open‑source tools that democratize deep‑learning research.