OpenAI's Top Hire, an Autonomous Research Agent

In OpenAI's spring hiring competition, Parameter Golf, the autonomous research agent Aiden took first place by setting the most merged records. Of the 47 total merged records, 7 (about 15%) were produced by Aiden, more than twice the number of the next-highest human contributor (3).

Aiden achieved this outstanding performance while using less than 4% of available compute resources on a single GPU node. The agent was built on AIDE, an open-source tree search technique for optimizing ML metrics. It works by reading new PRs, decomposing techniques into their component parts, then reassembling rule-compliant components to make new attempts.

Of particular note is the asynchronous collaboration between humans and the agent. During the competition, when a human researcher proposed a new tokenizer, Aiden fused it with components it had already built, recording the largest performance improvement (val_bpb) of the entire competition.