AI improved itself for two days

Andrej Karpathy, one of the founders of OpenAI, walked away from his computer for two days and let an AI agent work with nobody watching. It ran 700 experiments on how a smaller AI model gets trained, and it kept the 20 improvements that actually worked.

Nothing broke while he was gone.

The reason it worked matters more to your business than the result. Karpathy built four guardrails before he walked away.

The agent could score its own work, so a good change stayed and a bad one got binned. Every change could be undone, so a bad idea never damaged anything. It worked in tiny steps, one small improvement at a time. And it stayed inside strict boundaries, free to experiment in one part of the system and locked out of the rest.

Now look at how most owners hand work to AI. A vague job, no way to check the output, no undo, no limits. Then the first mess convinces them it can't be trusted.

The owners getting real work out of AI while they sleep built the fence before they let it loose.

Write those four rules down before you hand AI its next unsupervised job.