Early-Signal Detector

How it works

A free, transparent tracker that reads the long tail of the technical web every day, detects ideas that are accelerating before they reach mainstream attention, and keeps a public, timestamped record of its own calls.

Every day

  1. Ingest. New arXiv papers in AI, machine learning and related computer-science categories, new GitHub repositories that gather stars, and Hacker News stories with traction.
  2. Extract. Claude reads each abstract, repository description and post and names the specific techniques, tools and ideas in it.
  3. Merge. Different names for the same idea (an acronym, a longer phrase, a paper's own name) are grouped into one concept. Candidates come from shared words and acronyms; Claude confirms each merge.
  4. Score. Plain statistics, no language model: acceleration (growth of growth), novelty (first-seen date), independence (distinct authors, owners and posters; one lab posting ten papers is not a trend), spread across sources, and propagation from papers to code to discussion. Large established topics are left out.
  5. Explain. For the top candidates, a stronger Claude model writes a short explanation from the sources and filters spam, promotional bursts and established topics.

Every week

The top 10 concepts are frozen into the ledger, together with a baseline of the 10 most-mentioned concepts. Each call is later checked automatically against a fixed definition of “went mainstream”, and the scorecard reports hit rate and lead time for both.

Caveats

Rising attention in research is not a forecast of commercial success, and nothing here is financial advice. Language models already know which past topics became big, so ranking is left to statistics, and any backtest would be contaminated; the forward record is the real evidence. Names can split, merge or change over time: merged concepts keep their ledger entries and follow the surviving name.