One of the most persistent promises in AI is that models will eventually accelerate their own development by conducting research autonomously. A July preprint from Princeton researcher Sayash Kapoor, titled "Can AI agents conduct open-ended AI research?," argues that promise remains far from realized. "I don't think full automation of open-ended research is on the horizon right now," Kapoor wrote, a view he expanded on in a subsequent Nature commentary.
Kapoor's analysis draws a sharp distinction between two very different capabilities that often get lumped together. AI systems have gotten genuinely good at improving specific, well-defined algorithms — tweaking a model architecture, optimizing a training pipeline, finding a faster implementation of a known technique. What they have not demonstrated is the ability to independently frame open-ended research questions, design experiments to test genuinely novel hypotheses, and interpret ambiguous results the way human researchers do.
The distinction matters beyond academic debate. Predictions about explosive AI progress often assume that AI research automation will create a feedback loop, with AI systems improving themselves faster than humans could improve them manually. If Kapoor is right that this kind of open-ended automation isn't close, that particular path to rapid self-improvement looks less imminent than some industry commentary suggests — even as narrower forms of AI-assisted research continue to show real, measurable gains.
