The pages
- Discoveries: the catalog of AI-enabled scientific results, with charts, filters and a list by year.
- Models: model releases from leading labs and the model families named in discoveries.
- Labs and families: one profile per organization or model family, with its timeline, discoveries and releases.
- Vuln discoveries: public vulnerability discoveries with documented AI involvement, disclosure context and fix evidence. This is separate from the scientific catalog and is not a live patch inventory.
- Dataset and API: download the catalog as JSON or CSV, schemas and a read-only static API.
- Contributors: who maintains the catalog, how to contribute and how contributors are credited.
- Methodology: what is included, how entries are verified and how relevance is scored.
Start with the timeline
The blue banner on the Discoveries page shows one mark per catalog entry, placed by date; height represents editorial relevance. Hover over a mark to preview it, or select it to open the entry. Where several entries share a date or appear close together, choose from the nearby entries; their actual dates are preserved. With the keyboard, focus the chart, use the left and right arrows, then press Enter. The Models scatter plot also supports keyboard exploration and a chooser for nearby releases.
Narrow the catalog
- Fields: the buttons under “Explore by field” filter by science field; the number is the discovery count.
- Charts: click a row (field or organization) or a year in “By the numbers” to filter the list; click it again to clear. Aggregate charts describe the full catalog, even when the list is filtered.
- Filter bar: search text, organization, year, model family, minimum relevance and sort order. “Clear filters” resets everything.
- Copy filter link: every filter is stored in the address, so you can share or bookmark the exact view.
Read an entry
- Each entry gives the date, the organization, a summary with the evidence limits, the AI technology and an editorial relevance score from 20 to 100 in 13 possible steps (open the source and editorial details to see impact, validation and novelty). It is not an evidence grade or a precise measure of scientific value.
- “Original source” opens the original paper, preprint or institutional announcement; “Paper” opens a separately linked paper when available. Figures are reported by sources and are not independently verified.
- “Source checked” means we checked the source page, not that the finding was replicated. The expandable details work with touch and keyboard, show the source type and editorial subscores, and are included when printing. A missing paper link does not mean no paper exists.
- The organization name and the family line lead to profiles; the family line also links to the Models page.
Read the annual chart
Choose catalog-entry counts or mean editorial relevance. The accompanying table shows both and the annual sample size. These figures describe this selected, incomplete collection, not worldwide discovery rates. The most recent catalog addition is not a scientific review date. The evidence and coverage tables show which entries have structured annotations and how primary source types vary by year; neither is an evidence-quality ranking.
Suggest a correction or discovery
“Suggest a correction” on an entry opens a GitHub draft with its ID and source. “Suggest a discovery” in the catalog and profile footers opens a separate form requiring an original source and announcement date. GitHub sign-in is required and submitted reports are public: do not include personal information. Every suggestion is reviewed before any catalog change; opening a form does not submit it.
Models and profiles
On Models, click a year or lab to filter releases, or a family to open its card. A family is linked to a discovery when the discovery names it, so the link identifies a family, not the exact version used. On Labs and families, pick any row to see that profile.
Share and print
- Copy link on an entry copies a link that shows a preview card on social networks and opens the entry.
- Use your browser’s Print or “Save as PDF” on Discoveries, Models or Vuln discoveries to include every matching entry or release, including rows not yet loaded on screen. Discovery and vulnerability evidence details are expanded for print; the page restores your reading state afterwards. Your current filters apply.
Good to know
- Light and dark themes follow your device settings.
- Coverage is manual and incomplete; absence of a result does not mean it did not happen.
- Model data comes from Epoch AI under CC BY 4.0.