A recent preprint argues that the landscape of AI development has shifted dramatically: AI systems now write the majority of code within the companies that build them, a change from even a year ago. The authors suggest that as more of the AI research and development pipeline becomes automated, the pace of progress could accelerate in a way that is fundamentally different from incremental improvements.
The paper explores the possibility of a radical acceleration, often termed an intelligence explosion, where AI systems improve themselves in a self-reinforcing loop. While the abstract does not provide specific predictions or timelines, it frames this as a scenario that warrants serious consideration. The authors appear to agree that the trend toward automation in AI R&D is real and accelerating, but the abstract alone does not reveal any dissenting views or alternative scenarios.
The core concern is not just speed but control. If AI systems are driving their own development, the ability of humans to steer or intervene may diminish. The paper stops short of offering solutions, but it positions the question of controllability as central to any discussion of automated AI research. For now, the analysis serves as a cautionary note about the potential consequences of handing the research process to the machines themselves.