Wikipedia's List of unsolved problems in mathematics, frozen at 26 January 2024 (revision 1199136318), every problem on it researched individually, and each result since scored for whether an AI system was involved.
The page has 467 bullets, ten of which are headings grouping the ones beneath them. The denominator is the other 457.
How much of recent mathematical progress involved AI is a proportion, so it needs a denominator nobody chose after the fact. Collecting announced results gives a numerator and no denominator: labs publish successes, journalists write about the dramatic cases, and the pile measures what got reported. Here the denominator is fixed in advance, by a Wikipedia revision predating the results being measured, and every problem on it is researched whether or not anything happened to it.
data/sonnet-2026-09-22/ — 457 problems researched in September 2026, one agent per problem, from
the raw records in deepdive/mark_2026-09-22/ and triage/2024-01/. Earlier runs under data/
used a different method and are kept for comparison only.
Read data/sonnet-2026-09-22/AUDIT.md before relying on any number here: it covers what was
measured, what went wrong in the run, and the gaps that remain.
python3 pipeline/fetch_list_revision.py # network; -> data/<V>/page.wiki
python3 pipeline/make_problem_keys.py # network; -> problem_keys.md
python3 pipeline/triage.py # optional; ~$90, one cheap agent per problem
python3 pipeline/deep_dive.py --trial <name> # the research; one agent per problem
python3 pipeline/merge_research.py --trial <name>
python3 pipeline/judge.py --trial <name> # optional; checks records for contradictions
python3 pipeline/fetch_article_history.py # network, cached; -> page_history.json
python3 pipeline/build.py # -> data/<V>/, src/data.js, site/
Both research passes are resumable — a problem whose record exists is skipped — and read the API
key from .secrets.json (gitignored), a file holding {"ANTHROPIC_API_KEY": "sk-ant-..."}. The
VERSION constant must agree across build.py, fetch_list_revision.py,
fetch_article_history.py and merge_research.py.
python3 pipeline/build.py
pipeline/publish.sh [remote] # -> gh-pages, as index.html
The site is one self-contained file. publish.sh rebuilds gh-pages from scratch each time — just
index.html and .nojekyll, no parent commit — and force-pushes it. Point Settings → Pages at
that branch, root folder, once.
site/— the finished site: one HTML file, no server, no network.src/— its source: markup and CSS, the rendering code, and the dataset asdata.js.pipeline/— the scripts above, and nothing else.triage/<revision>/,deepdive/<trial>/— the raw agent records, one JSON per problem.data/<version>/— the built dataset as JSON, CSV and SQLite.DEEP_DIVE.md,TRIAGE_INSTRUCTIONS.md,JUDGE.md— the prompt each pass runs under. The coding rules every record follows are stated in full inDEEP_DIVE.md§6.
Code MIT, data CC BY-SA 4.0 (inherited from Wikipedia). See LICENSE.