Study finds humans still steer AI‑assisted model development
Fudan University researchers analyzed 700 task logs from building a 744-billion-parameter model and found humans set the goals while AI carried out most of the execution.
A team at Fudan University analyzed more than 700 task logs from 56 participants building Atria Dawn Preview, a 744-billion-parameter mixture-of-experts language model. Humans made 93.4% of the decisions about a task's goals and scope, the team found. AI agents were involved in 96.5% of the tasks.
Humans set methods and parameters 85.5% of the time versus 9.2% for AI, the researchers said. The median ratio of agent actions to human inputs climbed from 11 to 28.5 over four weeks. The team attributed that rise to growing task complexity, not to agents working with less oversight.
About a third of AI-assisted tasks, 151 of 455, were rated impossible without AI help, the study found. That suggests the tools expanded what the team could attempt, rather than simply speeding up existing work. When problems came up, humans intervened in 76% of cases, mostly by supplying context or diagnosing the issue.
The researchers summarized their finding as "AI proposes, humans choose," according to The Decoder, which reported on the study. The paper is based on Fudan's own project rather than an outside audit, and its conclusions have not yet been tested against other teams' AI-assisted research.