AI Trend Tracker

Research momentum counts the papers submitted to arXiv’s AI-related categories each week whose title or abstract mentions a topic, as a share of all AI papers that week. Using share rather than raw counts keeps conference deadlines, which lift every topic at once, from looking like a trend.

Discussion momentum counts Hacker News stories with the topic in the title. These numbers are small and jumpy, so read them as a second opinion, not a verdict.

A topic is called rising or cooling when its paper share moved by at least 10% between the two four-week windows. Topics are defined by keyword lists, so a paper can count toward several topics, and work that uses unfamiliar wording is missed. Small topics swing more easily, so the headline only names topics averaging at least 25 papers a week.

The map places a topic by two changes over the same four-week windows: its share of AI papers (across) and its Hacker News stories (up). A topic is placed only when it averages at least 25 papers and 5 stories a week, because smaller counts swing too easily. A change beyond the edge of the map is drawn on the edge. The week in five lines is written by fixed rules from the same numbers, not by an editor or a model.

Economy and society themes count Hacker News stories whose title contains AI and a theme word, such as “layoffs” or “data center”, as a share of all stories with AI in the title. Each story counts once per theme. The counts are small, often 10 to 30 stories a week, so a theme is called rising or cooling only when its share moved by at least 15%. Hacker News readers are mostly technologists, so this shows what that community is discussing, not public opinion.

Closed models released only through an API do not appear in the model lists, which cover open models on Hugging Face. They show up in the news and in the lab announcements instead.

The indicators each rest on one source and one rule. Job postings are Indeed’s own count of US postings asking for AI skills. Federal documents are those whose text includes the phrase “artificial intelligence”. A model release is the first post announcing a new version number in a lab’s main line (GPT, Gemini or Claude); variants of a version already announced, such as a smaller or faster edition, are not counted again, and labs other than these three are not covered.

From the analysts lists the newest pieces from a fixed set of ten newsletters and research institutions, at most two from each, from the past 30 days. It is a reading list chosen by the editor, not a ranking. Publishers that cover other subjects are filtered to pieces that mention AI.