AI scientist that runs experiments while you sleep

Want an AI agent to run repeated experiments without checking every attempt yourself? This brain helps you design an experiment loop with a clear goal, a fixed evaluation, a time or compute budget, and rules for keeping or discarding changes. Explore what makes a result trustworthy before giving the agent more freedom.

• What would a first experiment loop look like for my project?
• How do I choose a metric and decide which changes to keep?
• The score improved, but the result got worse. How should I check the evaluation?

Describe what you want to improve, how you measure it, and what the agent may change. You can discuss a bounded first experiment, including adaptations beyond machine-learning training.

An independent, unofficial guide based on Andrej Karpathy’s public autoresearch code, writing, and interviews. Original experiments, reported results, and proposed adaptations are distinguished. Not affiliated with or endorsed by Karpathy.