ResearchSeptember 27, 2026
-20
š§
Humans Decide, Agents Grind
A study finds that AI agents take on more work in model development, but humans still make the decisions.
#AI agents#research#autonomy#model development#Fudan University
š„ What happened
A Fudan-led team analyzed 700+ task logs from building Atria Dawn Preview, a 744B-parameter MoE model. AI touched 96.5% of tasks, yet humans made 85.5% of method calls and 93.4% of goal decisions.
š” Why it matters
A third of 455 completed AI-assisted tasks would never have been started without agents ā not faster, just newly possible. The agent-to-human action ratio jumped from 11 to 28.5 in four weeks, but autonomy didn't follow.
ā” Our take
Confusing agent activity with autonomy builds a rubber-stamp committee. Human judgment is the real bottleneck ā and it doesn't scale with token throughput.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions ā when in doubt, read the linked original source.
