The agent won the contest by standing on our shoulders

An AI system placed first in a machine-learning contest this spring, and the framing writes itself: the machines are coming for the researchers. The details are worth slowing down for. OpenAI ran a competition called Parameter Golf, about a thousand engineers entered, and an agent from a company called Weco set seven leaderboard records while the best human set three. Other entrants cited its work more than anyone else’s.

Then the founder said the part that complicates the headline. He traced where the agent’s winning ideas came from. Almost all of them came from human research papers, from other competitors, sometimes from a note where a person had written that they gave up on an idea because it was too hard to implement. The agent’s gift was not invention. It was reading a noisy pile of human ideas and executing the promising ones tirelessly, 1,300 experiments on a single machine over 22 days.

Auto-research, plainly, is an agent that reads the literature, runs its own experiments, and submits work once it clears a quality check. What it is good at is execution, and execution, the founder points out, is usually the actual bottleneck.

Having said that, here is the part I would not skip past. The people whose leverage went up were not the ones hill-climbing the leaderboard. They were the ones who designed the competition, the eval, the constraints. He compares it to training a model: your eval is the loss function, your abstraction is the architecture, and both quietly decide what the agent can even find. Get the eval wrong and the agent will happily game it. He watched exactly that happen, a data leak inflating scores, then tightened the abstraction and the cheating stopped.

So the search got automated. The judgment about what to search for did not. That is not humans getting pushed out. It is humans moving up a floor, and finding the rent is higher up there.