AI Did 96% of the Work on a New Model, but a Human Made the Calls

Researchers at Fudan University looked at how humans and AI agents worked together to build the Atria Dawn Preview language model, which has 744 billion parameters. AI took part in 96.5% of tasks. A third of the completed tasks would have been impossible without AI. Still, humans made the calls on methods and parameters 85.5% of the time. Over four weeks, the number of agent actions per human instruction climbed from 11 to 28.5. And when problems came up, humans had to step in 76% of the time.

This matches what I see every day working with agents. They take more and more steps on their own, but the choice of method and the final decision still rest with a person. At my company, an agent proposes an option, and a specific person is accountable for the outcome, not the agent. A rising number of agent actions per human command is not a sign that the agent has become more independent. It just means one human decision now unfolds into more steps. I am not ready to buy the promise that agents will soon work fully unsupervised.