In AI research, the prevailing belief is that more effort—more researchers, papers, and funding—leads to faster progress. Technological growth theories, however, locate the difficulty of discovery in the structure of the idea space. By analyzing 36 years of computer science data—7.2 million papers, 73,089 peer reviews, and 3.7 million researchers' topic shifts—we find that neither sheer scale nor idea space alone explains breakthroughs.
After the advent of the Transformer, the field’s attention narrowed sharply even as its overall size kept growing. Over half of the researchers who shifted topics moved into AI, especially those in mid‑career, while reviewer confidence dropped by a full point on a 5‑point scale. Each successive breakthrough—from AlexNet to large language models—raised the field’s carrying capacity for new ideas but was followed by a steeper exhaustion of those ideas. Larger breakthroughs capture attention more completely, depleting the diversity available to successors. A similar cycle appears in structural biology after cryo‑EM and AlphaFold.
Simulations show that well‑intentioned policies such as reciprocal reviewing empower prolific authors and further narrow what the community accepts. We therefore call on institutions to protect diverse attention in AI research rather than reward concentration on a single winning paradigm.
Review: The work highlights that AI progress cannot rely solely on resource accumulation; fostering a breadth of research perspectives is essential to avoid systematic depletion of innovative capacity.