Explore the AI landscape.

This prototype samples up to 512 recent, publicly visible signals, using the first 2,000 characters of each post. Filters apply to this sample, rather than the entire archive. Dates use when we observed a signal, not necessarily when it was published.

Clusters are discovered from similarities between posts, without a predefined topic list. Labels summarize phrases shared by their visible members. Jev independently scores membership in each discovered cluster; overlapping membership and unassigned signals are allowed. Discovery uses TF-IDF text similarity, so paraphrases and different languages can form separate groups, and labels can be rough. Assignments require a model score of at least 0.75; these scores are not a verified accuracy measure. Names come from a starting vocabulary and conservative text patterns for startup cohorts, announcements, named tools and project links. Detected names and domains can be ambiguous; this is not verified company identification. Distinctive ordering favors names whose posts concentrate in the selected context, so a single-post connection can appear above a popular name. Most shared orders by the number of shared posts. Link-only mentions do not count toward established company names. Products and companies are separate nodes. Lines show shared posts, not verified commercial relationships. Counts measure posts in this collection, not market share, growth or importance.

Cluster analysis is generated in a shared background job. Viewing, filtering and zooming make no model calls and consume no personal semantic search credits. Reloading reads the latest completed analysis; new or changed signals stay unassigned until a successful refresh. Open source posts for full context.