Stronger artificial intelligence breakthroughs hasten their own exhaustion

[arXiv] Beichen Lu, Honglin Bao, James A. Evans and other authors published a paper on arXiv arguing that breakthroughs in artificial intelligence depend not only on the scale of investment or the space for ideas, but also on diverse research attention, and that collective crowding depletes this diversity. The study analyzed 7.2 million papers, 73,089 peer reviews, and the topic migrations of 3.7 million researchers in computer science over 36 years. The paper says that after the emergence of Transformer, the field expanded in scale but its attention narrowed markedly; more than half of the researchers who switched topics moved into AI, especially mid-career researchers, and reviewer confidence fell by a full point on a five-point scale. The authors say that from AlexNet to large language models, each breakthrough raised the ceiling on the ideas the field could accommodate, but the subsequent creative exhaustion was also steeper; structural biology showed the same cycle after cryo-electron microscopy and AlphaFold. Simulations show that widely adopted policies such as reciprocal reviewing would enhance the influence of highly productive authors and further narrow what the field accepts. The authors call for establishing mechanisms to protect the diversity of attention in AI rather than rewarding concentrated investment in the winning paradigm.