Who is betting on what in AI, and against whom.
Hundreds of labs work on AI, but they disagree about only a handful of things. This site sorts them by those disagreements rather than by product. Six camps, each defined by its answer to four questions, with the people and labs in each, the money behind them, and the moves between them.
It is written for two readers. Researchers and students choosing a direction, who need to see the live scientific arguments and what would settle each one. And anyone who thinks AI means chatbots, who needs one page showing there are five other serious bets with Turing laureates and billions behind them.
A camp is its answers to four questions. What does the model learn from? When does learning happen, once before deployment or continually after it? What is the internal representation? Where does it act? The fault lines page shows all six sets of answers side by side. Product categories, chatbots, robots, video, cut across the camps; the questions are what the camps argue about.
Every claim that is not trivial carries a source: a URL and the date it was published. Money figures cite the announcing article. Signature ideas cite the paper or essay. A person's camp membership rests on a public statement, a paper, or an affiliation, never on inference. When credible sources disagree, the claim stays, marked as contested, with both sides cited. A build check refuses data that does not meet this rule, and the number of unsourced claims is zero.
Sources go stale. A dead link is a bug; report it and it gets replaced or the claim comes out.
Public professional facts only: name, roles with organization, title and dates, public papers and statements, honors, and a portrait from Wikimedia Commons with its license. Nothing else. No contact details, no private information, no characterizations beyond what the person has said in public. Anyone listed who would rather not be can ask to be removed.
Round sizes and valuations are what the cited article reported. A valuation is marked pre-money, post-money, or reported, the last meaning the figure came from the press rather than the company. Rounds that were only reported as talks are shown but excluded from totals, as are listings and mergers. Investor names are as written in the announcement; the investors page merges spellings into one entity. None of this is advice.
Artificial Atlas is maintained by a small group of researchers who would rather the sources got the attention, and by anyone who sends a correction. The data is open under CC BY-SA 4.0 and the code under MIT; both live in the repository, along with the schemas and a JSON export of everything on the site.
Every page has a "Suggest a correction" link in the footer that opens an issue with the page already filled in. A pull request against the YAML is even better. The changes page lists every edit to the data with its date.