← Back to Signal

Signal methodology

Every number on the Signal page is computed deterministically — no manual entry, no placeholder, no LLM judgment call. This page explains how, in plain language. (Established 2026-10-01; before that date the four composite tiles, Startup Radar, and the directory/entity scores were honestly-labeled-nowhere placeholder figures.)

Data sources

Everything here comes from free, no-budget sources: ai-tracker's daily snapshots of Hugging Face, OpenRouter, and Ollama activity (the same tracker behind the Hugging Face Observatory and OpenRouter Observatory dashboards below); GitHub's public API for a curated watchlist of open-source AI/ML repositories; and machine.net's own published News articles and entity timelines, each of which now carries a few extra structured fields (a real publish date, and for some categories a deal size or talent-signal tag) specifically so these numbers can be computed rather than guessed.

AI Momentum Index

A blend of day-over-day activity growth across three live feeds: Hugging Face downloads among trending-ranked models and datasets only (so long-entrenched repos like a five-year-old sentence embedding model don't drown out real movement), OpenRouter's total paid-usage token volume, and Ollama's total pulls. The index is 100 × (1 + the average of those three growth rates), comparing the latest snapshot to the oldest one within a trailing window that targets 7 days and will extend toward 30 as more daily snapshots accumulate.

GitHub Pulse

The sum of commits in the most recent complete week across roughly 20 major open-source AI/ML repositories (llama.cpp, vLLM, Transformers, PyTorch, LangChain, Ollama, ComfyUI, and others), read directly from GitHub's own weekly commit-activity statistics for each repo. The delta is the week-over-week change in that same figure. Commit volume is naturally choppier week to week than a usage metric like downloads or tokens, so don't read too much into any single week's swing.

AI Capital Flow

The sum of disclosed deal sizes across every Funding & M&A article machine.net has published in the trailing 30 days. This is explicitly a machine.net-coverage figure, not an estimate of total AI funding activity everywhere — it only counts deals the site has researched and written about.

Hiring Heatmap

A net talent signal built from every Talent & Organizations article in the trailing 30 days, each tagged at write time as a hire, a layoff, a departure, or neutral. Hires push the index above 50, layoffs and departures push it below — a layoff counts for more than a single executive departure. Like Capital Flow, this reflects machine.net's own coverage, not a comprehensive labor-market measure.

Startup Radar

A curated list of roughly 90 known AI-industry company names is matched against the title and summary of every published article. Each mention scores points that decay with age, with a bonus when that article also discloses a deal size naming that company. The five highest-scoring companies are shown, normalized so the top company reads 100. A company has to actually appear in machine.net's own coverage to show up here — this isn't pulled from an external funding database.

Entity Signal Scores

For the site's full Signal Page profiles (14 so far), the score starts at a neutral 50 and moves up or down based on each timeline entry's tagged sentiment (positive, negative, or neutral), weighted more heavily the more recent the entry. A string of positive developments pulls the score up; a recent negative one pulls it down even if older entries were positive.

How "delta" works

Every index keeps an append-only history log of its own computed values. The delta shown is the change versus the most recent prior entry in that log — not a fixed calendar comparison — which is what lets deltas be real from the very first computed run instead of requiring months of backfilled history. A metric computed for the first time shows "new" rather than a misleading percentage.

Known limitations

AI Capital Flow and Hiring Heatmap only see what machine.net itself has published, not the full market. Startup Radar's company list is curated and will miss a real company that isn't on it yet. The AI Momentum Index's trailing window is still shorter than its 7-30 day target whileai-tracker's snapshot history builds up. 14 companies have full entity profiles so far; the rest of the Signal Pages directory illustrates scope rather than reflecting real scores. The P(doom) category's AI Risk Index still has no real data source and isn't built yet. None of this is hidden — it's tracked as open work in the project's content pipeline doc.

Computed daily by scripts/compute_signal.py in the machine.net repo. Full formula writeups live in the project's content/signal-methodology.md doc.