Knowledge graph
Every headline in the feed is reduced to its keywords. A node is a keyword; an edge means two keywords turn up together far more often than chance would predict. Each node keeps only its strongest few links, which is what pulls the graph into distinct clusters instead of one dense mat. Darker shades are the bigger clusters; click any node to read the articles behind it.
Not enough data yet
The graph needs a few articles before topics start connecting. Lower the thresholds, or come back after the next poll.
Keywords are extracted with a stopword filter, so ordinary English and newsroom filler ("says", "report", "according") never become nodes. Phrases are preferred over single words where both occur, which is why you see language model rather than language and model as separate topics. Links are scored by normalised pointwise mutual information rather than raw co-occurrence: in an AI feed almost everything shares an article with ai, so counting alone draws a hub joined to the whole graph and hides the pairings that mean something. Clusters come from label propagation over the surviving links, and "rising" compares a topic's share of the newer half of the feed against the older half.