A research note from METR asks how much large language models (LLMs) have changed the overall rate at which new knowledge appears. The authors examine public data for signs of a change in slope, treating January 2026 as a possible turning point. They define discovery broadly as any advance in public knowledge, including new findings or more efficient methods. Their main observations rest on public records alone; they note that internal work inside companies may be moving faster and remaining unpublished.
Reports of software vulnerabilities have risen sharply. Counts for projects such as cURL, OpenSSL, Firefox and Microsoft products all climbed markedly in 2026 compared with 2025. Aggregate databases of known flaws also show acceleration. In some cases a sizable share of the new reports carry marks indicating artificial intelligence helped find them. Lists that track only flaws already used in real attacks show milder growth.
Mathematical activity has increased as well, though measurement is harder. Submissions to arXiv, the main open repository for research papers, roughly doubled in certain branches of mathematics within months. A few long-standing open problems from classic lists were solved or advanced with the help of language models in 2026.
Looking at algorithmic improvements and possible explanations
By contrast, records of algorithmic optimizations show no clear acceleration. Time series for several well-tracked problems, including image classification speed, compression, and chess engines, continue at roughly their previous pace even when language models contributed some of the recent records. This lack of a sharp public rise is surprising given announcements of models that improve optimization results with modest computing effort.
The authors list possible reasons why some fields show stronger public effects than others. These include differences in how much computing power people devote to searching, how well the models handle each type of task, how freely results are published, and how easy the data are to track. Many published examples claim the models needed little specialized human guidance, yet the authors caution that expertise may still hide in the choice of problems or in checking answers. Overall the public picture is mixed: clear acceleration in vulnerability reports, moderate signs in mathematics, and little change yet in optimization records.