Wow. I checked whether it was April Fools’ Day before I read it.
The “it” is autoresearch, Andrej Karpathy’s new open-source project, which VentureBeat covered this week under a headline promising you can run hundreds of AI experiments a night. The piece is careful to note that the whole thing is, by Karpathy’s own admission, a simple 630-line script, on GitHub under an MIT license. The ambition and the file size are wildly out of proportion, which is usually the mark of either a joke or a good idea.
The idea is this. You give an AI agent a small but real LLM training setup — the README calls for a single NVIDIA GPU, tested on an H100 — and let it experiment on its own overnight. It modifies the training code, trains for five minutes, checks whether the result improved, keeps the change or discards it, and repeats. In VentureBeat’s telling, the agent reads its own source, forms a hypothesis such as changing a learning rate or an architecture depth, runs the experiment, and keeps the change only if validation loss improves. The human’s job shrinks to editing a markdown file that sets the research direction. Karpathy’s stated goal is to “engineer your agents to make the fastest research progress indefinitely and without any of your own involvement.”
A friend also sent me this one last night. We’re going to put it to the test today.
