How the numbers on this map are sourced, what the visual language means, and what we refuse to fake.
Two gates, both required. A sourced valuation of at least $100M — a public market cap, or the best public mark for a private company — and at least one mappable relationship. A company with no edges is dead weight on a relationships map, so we don't add a node until we can also add what it connects to. Products and subsidiaries are never valued separately: the legal entity carries the number, and the product appears as a division (Google DeepMind under Alphabet, Weights & Biases under CoreWeave).
Public companies' market caps refresh every weekday from a price feed (Yahoo Finance), converted to USD, with a sanity band that rejects implausible moves rather than silently applying them. Live-flagged valuations are never hand-edited.
Edges are typed — thirteen kinds, from routes to and rents from down to fabricated by, and critically the distinction between invests in (equity) and commits capacity to (spend commitments), which is what the circular-financing picture depends on. Edges added since August 2026 carry their own confidence level and a citation URL — click an edge to open its source. Uncited edges render faded.
The dataset lives in a git repository; every change is a reviewed, attributable commit. The "Changed" panel diffs today against a baseline roughly a week back — repricings, new companies, acquisitions, new edges — straight from that history.
Headlines are a separate layer from facts. Once a day a keyless script pulls one Google News feed per company (Google News aggregates Reuters, Bloomberg, the FT, TechCrunch, CNBC, company newsrooms and the press wires, so one feed per company covers the outlets we would otherwise poll one by one), keeps the last seven days, and classifies each headline by keyword into a kind: raise, acquire, ipo, partner, launch, other. A headline counts only if the company's name is in it; a short list of stock-churn outlets (MarketBeat, GuruFocus, Simply Wall St, Motley Fool, Zacks, TipRanks, Seeking Alpha, Investing.com, 24/7 Wall St) is dropped outright because they publish automated share-trade notices and ticker commentary, not company news. That is the whole editorial hand: no model, no summarising, no ranking beyond newest-first within a kind.
The industry-wide headlines come from Techmeme, the human-edited tech news front page, filtered to AI. Its feed links to its own story pages, so at ingest each item is resolved to the original article and the outlet that wrote it; Techmeme is kept only as the credit. The daily digest's "Today in AI" picks from that stream, written in our own words, linking to the outlet.
On the map, a dot on a bubble means an event you have not seen since your last visit (a round, a deal, a listing), and it clears when you open the company. The dossier lists the headlines with their outlet. None of it changes a number: a headline is a lead, and the map's marks move only after a sourced data pass verifies the lead against a primary report.
Some things we deliberately do not paper over: companies whose only public marks are stale keep their old date visibly; companies with no defensible mark say unsourced; a company whose relationships we cannot yet map stays flagged rather than getting an invented edge. If you see a gap, that is the dataset telling the truth about itself.
The dataset is served as plain JSON: entities.json (companies, valuations, sources), relations.json (typed edges with citations), changes.json (the current diff). If you use it, a link back to stateofai.world is appreciated. Corrections welcome — every number has a source attached, so bring yours.
Built and maintained by Han.