Andrej Karpathy (one of the godfathers of AI) dropped an open-source project called Autoresearch, and it’s going viral for a reason. In this video I break it down in the clearest way possible: it’s basically a super-nerd robot intern that you give a goal, and then it runs experiments in a loop on a GPU—editing code/settings, running short training runs, reading metrics, keeping winners, discarding losers—while you sleep. You define what “better” means (more clicks, cheaper leads, higher sales, better model score), and the loop just grinds. Then I go into the fun part: how you actually make money with Autoresearch loops. I share 10 concrete ideas—from niche “agent-in-a-box” products (Amazon listing experiments, email sequence tuning, SaaS pricing optimization) to always-on A/B testing for landing pages and ads, research-as-a-service, and agencies that can run 100× more tests than anyone else. I also cover non-business implications (I’m especially excited about what this could mean for science/medicine) and Karpathy’s companion project AgentHub, which is basically “GitHub for agents.” Finally, I show how to get started fast: I used Claude Code as my install buddy, and because I’m on a MacBook M1, I ran it on a rented NVIDIA GPU—Google Colab is the easiest on-ramp (switch runtime to a T4 GPU and follow the commands).