AI agents do more of the work in model development, but humans still make the decisions
decoder ·
A study by Fudan University researchers analyzing the development of the Atria Dawn Preview model reveals that while AI agents perform the majority of execution tasks, humans retain decisive control over project direction. The analysis of over 700 task logs showed that the ratio of agent actions to human inputs increased significantly, yet this reflected efficiency rather than autonomous decision-making. Approximately one-third of AI-assisted tasks were deemed infeasible without artificial intelligence, indicating that AI enables work that would otherwise not be attempted. Humans made 85.5 percent of decisions regarding methods and parameters, consistently selecting from options proposed by the agents. When problems arose, human intervention primarily involved providing context or diagnosing issues rather than performing manual labor. Despite this high level of automation, the authors warn of a 'rubber-stamp' risk where oversight becomes difficult due to the length of agent workflows. These findings contrast with industry trends at companies like Anthropic and OpenAI, suggesting that frontier models can handle engineering but still struggle with critical judgment calls.
decoder