Karpathy's Autoresearch slashes eCLIP mean rank

This is a strong proof of concept for agentic research, but the real story is scoping: once the task is bounded by a single metric, a single file, and a hard time budget, the agent can do useful work. It looks less like an autonomous scientist and more like a very fast ablation engine that still needs a human to set the question.

• –`program.md` is the real control surface, effectively acting like a lightweight operating system for the agent.
• –The sandbox and permission lock mattered as much as the model, because they kept the loop safe and reviewable.
• –The biggest gain came from a temperature clamp bug fix, which says a lot about how much low-hanging fruit still hides in research code.
• –Hyperparameter tuning delivered more value than architectural changes, which is exactly the kind of search current agents are good at.
• –Once the exploration moved into moonshot ideas, success dropped sharply, showing the ceiling of today’s autonomous research loops.