Autoresearch: Run ML Experiments on Autopilot with Git-Backed Rollback

🤖 Autoresearch: ML Experiments That Run Themselves While You Sleep
#

How much time do you spend adjusting a hyperparameter, waiting for training to finish, and checking if it improved? Autoresearch automates that entire cycle.

✨ How Does It Work?
#

It’s an open-source framework by Andrej Karpathy that implements an autonomous experimentation loop:

🔀 The AI agent proposes a change to the model or experiment.
💾 It makes a git commit before each attempt (automatic snapshot).
⏱️ Trains for 5 minutes and measures the target metric.
✅ If improved → the change stays. ❌ If not → automatic rollback to the last good state.
🔁 Repeats with no human confirmation needed.

🎯 Key Benefits
#

🗂️ Full git history – every experiment is tracked and reversible. 📋 Structured results log – survives crashes and records every attempt. 🌙 No supervision needed – explore the solution space while doing something else.

💡 In Simple Terms
#

Imagine hiring an assistant who works through the night testing variants of your model, saves each attempt in git, and shows you in the morning what worked and what didn’t. That’s Autoresearch: ML research on autopilot.