What’s Out There

How this works

There is far more data in the world than anyone can keep in their head. This site is a way to build a feel for it: 1,000 data sets, each boiled down to a one-page card, shown one at a time. Say whether each one interests you, and the feed adjusts.

What is on a card

What the data set is and what one row represents; who publishes it; what it costs, how it is licensed, and whether you need to sign up or apply; what it covers and how often it is updated; a quality rating from 1 to 5 with specific caveats; other data sets it joins or pairs well with; and a few questions it could answer.

Where the catalog comes from

629 cards were hand-picked by topic to cover the sources worth knowing in each field, including 129 that are paid, partly paid, or available only by application. The other 371 were harvested from public catalogs (data.gov, city and state open-data portals, Hugging Face, the Registry of Open Data on AWS, Zenodo, Harvard Dataverse, data.gov.uk, and open.canada.ca) to sample the long tail.

The cards were written by AI (Claude) and have not been reviewed line by line by a person. Each one says how far it was checked: for 517, the provider’s own page was fetched and read when the card was written; 365 rely on the catalog’s listing; and 118 could not be fetched and are written from background knowledge. Treat license and cost details as a starting point and confirm them at the source. Catalog built 2026-11-16.

How the feed learns

Every card is converted into an embedding: a list of 256 numbers that places it near cards with similar descriptions. Your votes train a small statistical model (a Bayesian logistic regression) on those numbers, plus a few plain facts such as whether the data is paywalled. For every card you have not seen, the model produces a probability that you would be interested and a measure of how unsure it is.

The next card is chosen by sampling from that uncertainty, so early on, when the model knows little, the feed ranges widely; as you vote, it leans toward what you like while still testing its assumptions. About one card in ten is a pure wildcard, and the feed avoids showing the same topic twice in a row. The “match” percentage on a card is the model’s estimate, not a promise. The map shows the same embeddings flattened to two dimensions.

Your votes

Votes are kept in this browser, with no account. Where the site has vote storage switched on, a copy is also saved under a random ID so you can carry your votes to another device from the Saved page. You can clear them there at any time.