Last Friday, Andrej Karpathy open-sourced a 630-line Python script and went to bed. By morning, an AI agent running on a single GPU had completed roughly 100 complete LLM training runs, each lasting exactly five minutes, autonomously modifying the neural network architecture, the optimizer, the hyperparameters, evaluating the results, keeping improvements, discarding failures, and moving on to the next experiment. No foundation model. No API calls to a frontier lab. Just data, a training loop, and an agent that doesn't sleep. Within 48 hours, the post had 8.6 million views . The repo hit 8,000 GitHub stars . Shopify CEO Tobi Lutke cloned it before bed on Saturday, pointed it at his own data, and woke up to a smaller model that outperformed a larger one he'd configured manually. A 19% improvement in validation scores. From a model trained from scratch. On his data. Overnight. Most of the commentary has focused on the "AI doing research while you sleep" angle, and that's reasonable. It's a compelling image.…