All news

Stanford AI100: the data and compute scaling era runs into limits

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

According to AI100, the next leap in SI will depend not simply on even larger clusters but on more efficient algorithms and ways of reasoning. At the same time, society will have to negotiate who controls a technology that has already entered everyday life.

The international panel behind Stanford AI100 has released its third report, "AI Goes Mainstream: The Scaling Era and the Rise of Generative AI", its first since 2021. The experts set out 12 findings on the past five years. Growth in data and compute drove the main advances in generative SI, but is now constrained by the shortage of high-quality data and the cost of compute; frontier development is shifting toward reasoning and additional inference-time compute. The report documents research timelines shrinking from years to months or weeks, unprecedented infrastructure investment and pressure on energy and freshwater supplies, and the concentration of frontier development among a small number of US and Chinese players. Other findings address cyber risks, military use, the loss of automatic trust in photographic, audio and video evidence, the public's limited influence over SI deployment, and fragmented international governance. The authors stress that few predicted the technology's current state even five years ago and do not present the report as a precise forecast for the next five.