ai-discovery-data

The goal is to track rates of discovery over time across many domains and see whether there has been a recent acceleration; the collection supports LLMs' Contribution to Discovery. A series is included when it has a consistent definition, a usable time axis, and public, rebuildable data. Evidence about AI usage is useful context, but is not required.

Potential future series and cross-domain causal designs are tracked in the appendix of additional candidates.

The collection is browsable at tecunningham.github.io/ai-discovery-data, where each series page renders its folder's full write-up with the interactive chart inline — hover any mark for the underlying record, and on several charts click through to the original reference. The pages are built from the same vendored CSVs and documents by tools/build_docs.py; the PNGs in the table below remain the static record.

A companion cumulative index redraws every series in one shared format — a single step function of progress to date, declining toward zero where the series has a known denominator.

Series

Vulnerabilities

Series	Chart
curl vulnerability disclosures Metric: vulnerabilities disclosed per quarter, split by whether the finder credit carries an AI marker Coverage: 2000–2026, partial through 2026-06-24 Acceleration? 📈 accelerating — 36 disclosures through 2026-06-24 annualize to roughly 75 against 9 in 2025 and a 13.1/year mean over 2014–2023 Discussion · Data · Source · Interactive
Firefox vulnerability disclosures Metric: distinct CVEs per quarter, split by whether the reporter credit names an AI method, an AI-security employer, a fuzzer, or none of these; advisory–CVE mentions retained as a sensitivity count Coverage: 2016–2026, partial through 2026-08-04, the latest advisory in the snapshot Acceleration? 📈 accelerating — 342 distinct CVEs through 2026-08-04 against 210 in 2025; the part year alone is 1.6 times the 2025 full year Discussion · Data · Source · Interactive
All software: vulnerabilities known exploited Metric: CVEs added per quarter to CISA's Known Exploited Vulnerabilities catalogue Coverage: 2021–2026, from the catalogue's November 2021 launch, partial through 2026-08-10 Acceleration? ➡️ no acceleration — 178 additions through 2026-08-10 annualize to about 293 against 245 in 2025 and a 206/year mean over 2023–2025 Discussion · Data · Source · Interactive
Microsoft security-update CVEs Metric: CVEs issued by Microsoft's own CNA per month, dated by first publication in the Security Update Guide, split by whether an acknowledgment credit names an AI method, an AI-security employer, a fuzzer, or none of these Coverage: 2016–2026, partial through 2026-08-11; no February or March 2016 document exists upstream, so the first year is ten months Acceleration? 📈 accelerating — 1,927 CVEs through 2026-08-11 against 1,243 in 2025; the part year annualizes to about 2.5 times 2025 Discussion · Data · Source · Interactive
All software: vulnerabilities disclosed Metric: CVEs published per quarter in the US National Vulnerability Database Coverage: 2016–2026, partial through 2026-08-10 Acceleration? 📈 accelerating — 49,838 CVEs through 2026-08-10 annualize to about 82,000, roughly 1.6 times 2025's 49,972, after +32% growth into 2024 and +23% into 2025 Discussion · Data · Source · Interactive
OpenSSL vulnerability disclosures Metric: vulnerabilities disclosed per quarter, split by finder provenance: corroborated AI method, AI affiliation with method unverified, conventional or fuzzing credit, or no reporter credit Coverage: 2002–2026, partial through 2026-08-05 Acceleration? 📈 accelerating — 39 CVEs by 2026-08-05 against 6 in all of 2025; the largest prior full years were 35 in 2016 and 32 in 2015 Discussion · Data · Source · Interactive
OSS-Fuzz vulnerability discoveries Metric: vulnerability records published per quarter by an automated fuzzing programme Coverage: 2020–2026, partial through 2026-08-10 Acceleration? 📉 declining — 1,041 records in 2020 to 244 in 2025; 2026 annualizes to roughly 396 Discussion · Data · Source · Interactive
Open-source CVEs represented in OSV Metric: distinct CVE IDs linked to at least one active affected-package record in OSV, per quarter by earliest OSV publication date Coverage: 2016–2026, partial through 2026-08-10 Acceleration? 📈 accelerating — 21,321 distinct CVEs through 2026-08-10 annualize to about 35,100, 2.3 times 2025's 15,146 Discussion · Data · Source · Interactive

Open problems

Series	Chart
Erdős problems catalogue Metric: problems catalogued, statuses marked solved, and statements formalized in Lean, at monthly site snapshots; plus an imputed solution year per solved problem Coverage: thirteen monthly snapshots, 2025-08-31 to 2026-08-10; imputed solution years 1940–2026 Acceleration? ❓ inconclusive — 55 imputed resolutions in 2026 through 2026-08-10, against 33 in 2025 and a 5.9/year mean over 2000–2023 Discussion · Data · Source · Interactive
Top 10 Erdős problems Metric: dated resolutions per year across 12 scored rows Coverage: list posed 2026-04-16; dated resolutions 1975–2026; statuses read 2026-08-14 Acceleration? ❓ inconclusive — 1 resolution in 2026 against 3 in the 90 years since 1936; a series this small sets no rate Discussion · Data · Source · Interactive
FrontierMath Open Problems Metric: dated solution events on Epoch AI's pool of open research problems, placed by curator-assigned notability tier Coverage: benchmark announced 2026-02-26; pages read 2026-08-14, with recorded solves from 2026-03-23 to 2026-08-12 Acceleration? ⏳ too early — 6 dated solves between 2026-03-23 and 2026-08-12; the pool was announced 2026-02-26 and no prior-year rate exists Discussion · Data · Source · Interactive
Ben Green's 100 open problems Metric: dated resolutions per year across 101 scored rows Coverage: 2018–2026; statuses as the December 2025 revision records them, read 2026-08-13; dated resolutions 2019–2025 Acceleration? ➡️ no acceleration — 0 dated resolutions in 2026 against 3 in 2025 and a 1.9/year mean over 2019–2025 Discussion · Data · Source · Interactive
Hilbert's problems Metric: dated resolutions per year across 28 scored rows Coverage: list posed 1900; dated resolutions 1900–1998; statuses read 2026-08-14 Acceleration? ➡️ no acceleration — 0 resolutions in 2026 and 0 since 1998; 12 dated resolutions over 1900–1998 Discussion · Data · Source · Interactive
Landau's problems Metric: dated resolutions per year across 4 scored rows Coverage: list posed 1912; no dated resolution 1912–2026; statuses read 2026-08-14 Acceleration? ➡️ no acceleration — 0 resolutions in 2026; 0 dated resolutions over 1912–2025 Discussion · Data · Source · Interactive
Millennium Prize Problems Metric: dated resolutions per year across 7 scored rows Coverage: list posed 2000; one dated resolution, 2003; statuses read 2026-08-14 Acceleration? ➡️ no acceleration — 0 resolutions in 2026; 1 dated resolution (2003) over 2000–2025 Discussion · Data · Source · Interactive
Smale's problems Metric: dated resolutions per year across 19 scored rows Coverage: list posed 1998; dated resolutions 2002–2026; statuses read 2026-08-14 Acceleration? ❓ inconclusive — 1 resolution in 2026 against 4 over 2002–2016; a series of 5 events sets no rate Discussion · Data · Source · Interactive
Thurston's 24 questions Metric: dated resolutions per year across 24 scored rows Coverage: list posed 1982; dated resolutions 1993–2013; statuses read 2026-08-14 Acceleration? ➡️ no acceleration — 0 resolutions in 2026 and 0 since 2013; 22 dated resolutions over 1993–2013 Discussion · Data · Source · Interactive
The Open Problems Project Metric: dated resolutions per year across 78 scored rows Coverage: list begun 2001; dated resolutions 2000–2024; statuses read 2026-08-14 Acceleration? ➡️ no acceleration — 0 resolutions in 2026 and 0 since 2024; 17 dated resolutions over 2000–2024 Discussion · Data · Source · Interactive

Mathematical bounds and records

Series	Chart
Inventory of the AlphaEvolve problem set Metric: per problem, whether it has a live numeric record and how many dated prior works the paper cites Coverage: the 65 problems the paper numbers 6.1 to 6.65; cited works span 1852–2025; built 2026-07-26 Acceleration? ⚪ baseline — 65 problems inventoried, 31 with a live numeric record; built 2026-07-26 Discussion · Data · Source · Interactive
Finite construction records around AlphaEvolve Metric: cumulative record steps in five groups of finite construction and packing problems Coverage: 1949–2026, 22 record steps across the five groups; transcription current to 2026-08-12 Acceleration? ❓ inconclusive — 1 record step in 2026 against 9 in 2025 and a 0.2/year mean over 1949–2024 Discussion · Data · Source · Interactive
ANTEDB analytic-number-theory exponents Metric: cumulative slice-level record changes across 58 exponent slices in the three families $\mu$, $A$ and $\beta$ Coverage: 1920–2024 in the underlying literature; extracted from the database as of 2026-07-26 Acceleration? ➡️ no acceleration — 0 slice changes in 2025 or 2026 against 2 in 2024 and a 3.5/year mean over 1931–2024 Discussion · Data · Source · Interactive
Elliptic-curve rank records Metric: the largest rank exhibited for an elliptic curve over Q, counted as one step per record, split into a proved-lower-bound frontier and a frontier of ranks known exactly; a third table holds every curve on the ICARM leaderboard with its rank and its size, and a fourth dates the posts around the 2026 record Coverage: nineteen record steps, 1938 to 2026, dated by year; the leaderboard snapshot spans 2026-05-27 to 2026-08-20, read 2026-08-20 Acceleration? ❓ inconclusive — 1 record step in 2026 against 0 in 2025 and 1 in 2024; 3 steps over 2001–2026 against 14 over 1974–2000 Discussion · Data · Source · Interactive
Sphere-packing lower-bound ladder Metric: cumulative improvements to the asymptotic lower bound on sphere-packing density in high dimension Coverage: 1905–2025, eight recorded steps; statuses read 2026-08-14 Acceleration? 📈 accelerating — 4 steps over 2011–2025 (2.7/decade) against 4 over 1905–2010 (0.4/decade); 0 steps dated 2026 Discussion · Data · Source · Interactive
Sums-and-differences and autoconvolution constants Metric: best known lower bounds on two additive-combinatorics constants, $C_{6.44}$ and $C_{6.3}$ in the AlphaEvolve numbering Coverage: 2007–2025, twelve record steps across the two ladders; source bibliography read 2026-08-14 Acceleration? ❓ inconclusive — 0 record steps in 2026 against 7 in 2025; the other 5 fall in 2007 and 2010 Discussion · Data · Source · Interactive
Matrix-multiplication exponent ω Metric: best proved upper bound on the asymptotic exponent ω of n×n matrix multiplication; lower is better Coverage: 1969 to 2026, sixteen recorded steps; transcription current to 2026-08-18 Acceleration? ❓ inconclusive — 1 new bound in 2026 against 0 in 2025 and 2 in 2024; movement of 0.0025 over 2010–2026 against 0.4319 over 1969–1990 Discussion · Data · Source · Interactive

Algorithms

Series	Chart
CIFAR-10 speedrun Metric: seconds of training to 94% test accuracy on CIFAR-10 on a single A100, per claimed record Coverage: 2018–2026; the plotted series runs 2022-12-29 to a claim of 2026-07-09, with acknowledgment last checked 2026-07-28 Acceleration? 📉 declining — yearly improvement factor 1.09 in 2026 (through 2026-07-09, claim included) against 1.3 in 2025 and 2.4 in 2024 Discussion · Data · Source · Interactive
CVRPLIB X-instance record frontier Metric: better best-known objectives and later optimality proofs recorded for a fixed cohort of 100 CVRP X instances, one event per posting Coverage: 2015–2026, 289 event rows posted through 2026-07-04 Acceleration? 📉 declining — 3 events in 2026 against 3 in 2025; 264 of the 267 better- objective events were posted 2015–2021 Discussion · Data · Source · Interactive
ECDSA.fail secp256k1 point-addition circuit Metric: best validated score (average executed Toffoli count × peak qubit width) for a reversible secp256k1 point-addition circuit; lower is better Coverage: 2026-05-30 to 2026-08-10, 433 accepted records Acceleration? ⏳ too early — first record 2026-05-30, so no prior-year rate exists; the 2026 series is a 7.3× fall over 72 days Discussion · Data · Source · Interactive
Hutter Prize compression: enwik9 Metric: total size in bytes of decompressor plus archive for a fixed 1 GB text corpus, under a CPU-time and memory cap, per awarded record Coverage: 2019 baseline to 2026; the prize moved to enwik9 on 2020-02-21; prize site read 2026-07-28, benchmark page's own update dated 2026-07-08 Acceleration? ➡️ no acceleration — 0 awarded records in 2026 (one pending claim of 2026-06-26) against 2 in 2024 and 4 over 2021–2024; the uncapped comparator is unchanged since 2023-10-23 Discussion · Data · Source · Interactive
Gurobi mixed-integer programming speed Metric: cumulative vendor-reported MILP speedup across releases, every version rerun on one machine Coverage: releases 10 through 13, announced 2022-11-14 to 2025-11-18, baselined at version 9.5; transcription current to 2026-08-10 Acceleration? ➡️ no acceleration — no 2026 release exists (series ends 2025-11-18); the 2025 release gained 0.6% against 13.1% in 2024 and a cumulative 1.40× over 2022–2025 Discussion · Data · Source · Interactive
MIPLIB 2017 solution frontier Metric: better feasible incumbents, first feasible solutions and optimality updates announced in MIPLIB 2017 solufile releases Coverage: 2019-08-26 through 2026-01-26, 28 releases with explicit solution counts Acceleration? ➡️ no acceleration — 40 announced updates in the single 2026 release against 13 in 2025 and a 90.1/year mean over 2019–2025 Discussion · Data · Source · Interactive
modded-nanogpt training speedrun Metric: minutes of training to a fixed target validation loss, per accepted record Coverage: 2024-05-28 to 2026-07-17, all 89 records listed in the repository README Acceleration? ➡️ no acceleration — the standing record fell 1.5× in 2026 (33 records through 2026-07-17) against 1.9× in 2025 (39 records) and 12.6× in 2024 (17 records) Discussion · Data · Source · Interactive
Stockfish development builds on fixed hardware Metric: Elo relative to Stockfish 15, from 20,000 games per build on one fixed machine and time control Coverage: 2013-04-30 to 2026-07-26, 2,542 tested development builds Acceleration? ➡️ no acceleration — 14 Elo through 2026-07-26 (annualizing to about 24 Elo/year) against 32 Elo in 2025 and a 51 Elo/year mean over 2013–2026 Discussion · Data · Source · Interactive

Outside the three domains

Series	Chart
Integer factorization records Metric: cryptanalysis; decimal digits in the largest hard semiprime factored, as a running maximum over dated records Coverage: 1991-04 to 2020-02, confirmed unmoved as of 2026-08-10 Acceleration? ➡️ no acceleration — 0 records in 2026 against 0 in 2025 and a 0.4/year mean over 1991–2025; the standing record is 250 digits, set 2020-02-28 Discussion · Data · Source · Interactive
arXiv submissions Metric: research output; preprints submitted to arXiv per month Coverage: 1991-07 to 2026-08, monthly, the last month partial at the 2026-08-10 fetch Acceleration? 📈 accelerating — a 28,450 submissions/month mean over 2026-01 to 2026-07 against monthly means of 23,707 in 2025 and 20,336 in 2024 Discussion · Data · Source · Interactive
DOI records deposited with Crossref Metric: formal publishing volume; DOI records deposited with Crossref per year, by created date Coverage: 2010 to 2026, annual, the last year partial through 2026-08-10 Acceleration? ➡️ no acceleration — 2026 annualizes to roughly 13.3 million records against 12.80 million in 2025 and an 8.63 million/year mean over 2010–2025 Discussion · Data · Source · Interactive
Git pushes to GitHub Metric: code output; git pushes to GitHub per quarter, summed over economies Coverage: 2020-Q1 to 2026-Q1, quarterly; fetched 2026-08-10 Acceleration? 📈 accelerating — 319.8 million pushes in 2026-Q1 against 246.8 million in 2025-Q4 and a 2025 quarterly mean of 212.2 million Discussion · Data · Source · Interactive

Validation

37 problems holding 87 figures and 63 data files. 23 refetch from upstream and 14 are maintained by hand and say so. 37 recompute their prose arithmetic. No failing cells.

How to read the series

Each row shows at most one primary graph, preferring the time-series view when one exists. It links to the folder that draws it, where the full-size figure and any supplementary diagnostics sit beside the data and documentation. The verdict asks only whether the series shows an acceleration in the rate of discovery, not whether AI contributed:

📈 accelerating · 📉 declining · ➡️ no acceleration · ❓ inconclusive · ⏳ too early · ⚪ baseline

Attribution is deliberately not an admission test. The first-stage question is whether output under a stable inclusion rule bends upward in the agent era. Finder credits, where they exist, help investigate a mechanism; where they do not, the time series still supplies evidence about the claimed acceleration. Neither case identifies causation by itself.

Open-problem ledgers are separated from mathematical bounds and records because their instruments differ. The former show dated resolution events; the latter track changes in numerical quantities.

The final group sits outside the three worked domains. Integer factorization is a cheap-verification control, while the output-volume series are contrast cases whose curves can bend without measuring discovery.

What validation checks

The repository checks that every chart can be traced to a public source and rebuilt from it. Each validation column is one kind of thing that can go missing:

✅ passes · ❌ fails · ✍️ maintained by hand, and the document says so · ➖ not run

The Arithmetic column exists because prose does not move when a CSV does. A refetch changes a number and leaves the sentence quoting it behind, stating a figure the data no longer supports, and nothing about the files looks wrong. A folder check.py recomputes each printed figure and asserts the document contains it, so the document stays the place the number lives while the CSV stays the thing that decides it. The status table above counts the folders that do this; the rest print numbers no check reads, and their ➖ says so rather than claiming a pass.

Reproduction runs every figure.py and compares the result with what is committed. It restores the original bytes afterwards, so a stale figure is reported rather than quietly staged. make check skips this slower step; make check-figures runs it, and make index runs it before regenerating the two tables above.

What the numbers are and are not

Four conventions run through every series here, and reading a chart without them will mislead you.

Attribution is optional, and acceleration is not attribution. A series is included when its events are selected consistently enough to compare over time. An upward bend is a signal to investigate alongside external evidence, not an estimate of AI's causal share. Conversely, a series does not become informative merely because a few events name a model.

A finder credit is a floor, not a measurement. Where a project records who found a vulnerability, this data classifies a report as AI-credited only when the credit string explicitly names an AI system, an AI-security firm, or an agent. A researcher who used a model and did not say so counts as human. So every AI share here is a lower bound by an unknown margin.

A disclosure is not a discovery, and a status change is not a solution. Vulnerability series count what got published, on the date it got published. The Erdős catalogue records the date a status was edited, which is not the date a problem was solved.

Records are lumpy with no AI in them. Half of all algorithm families never improve at all [@sherry2021fast], solver records jump every few years, and a century-scale exponent can sit still for eighty years and then move by hand. A staircase inside the agent era is not by itself an AI signature, and a flat stretch is not by itself an exhausted frontier.

Layout

One folder holds everything about each problem:

problems/cyber-curl/
  README.md                  what the problem is and what the chart supports
  curl-by-year.csv           the series, vendored from a public source
  curl-finders.csv           who was credited with each find
  fetch.py                   rebuilds those CSVs from curl's vuln.json
  figure.py                  draws the PNGs from the adjacent CSVs
  chart_spec.py              declares the docs page's interactive charts
  check.py                   recomputes the numbers the document states
  discovery-cyber-curl.png   committed, never hand-edited

Every README.md follows the reference format defined in FORMAT.md, which tools/check.py enforces.

A folder is self-contained except for generic helpers. Cross-series comparison happens in the prose rather than in a composite chart.

Reproducing

Figures are built in one digest-pinned Linux/amd64 container, both locally and in CI. Install Docker Desktop, OrbStack or another Docker-compatible runtime; the host's Python, matplotlib and fonts are deliberately not used:

make figure-image               # optional warm-up; later targets build it too
make figures                    # redraw every PNG in the pinned renderer
make figure PROBLEM=cyber-curl  # redraw one folder in the pinned renderer
make check                      # fast host-side data/document/source checks
make check-figures              # containerized redraw and byte comparison
make index                      # rewrite the generated README/CUMULATIVE tables
make docs                       # rebuild the interactive pages in docs/

Do not run a figure.py directly. The shared save helper rejects PNG writes outside the canonical container and points back to the corresponding Make command. make index is containerized too because it performs the full figure check before rewriting the generated README tables.

CI runs the same make check-figures target on every push and pull request. A second workflow, freshness.yml, runs weekly, refetches every automatable series, and fails if any vendored CSV no longer matches its upstream — the one failure mode that is invisible from inside the repository, since a stale series passes every other check. It checks the documented URLs in the same run.

The renderer pins the Python base image by digest, forces linux/amd64, and installs the exact versions in requirements.txt. Each PNG's Software metadata records the Python, matplotlib and FreeType versions plus its generator path. Local checks and CI therefore compare bytes produced by the same rendering ABI rather than merely similar Python environments.

Rebuilding data is a separate networked step:

make fetch                          # run every automatable fetcher
make fetch-one PROBLEM=cyber-curl   # run one folder's fetcher

Refetching can leave the repository failing its own check, by design. Every chart is drawn as of one date, AS_OF_DATE in lib/dates.py, which is where the shaded era ends and where a series that stops early is understood to stop. tools/check.py fails when any vendored row is newer than that date, because a figure drawn to an older horizon than its data is a figure that quietly omits rows. So a successful make fetch that pulls in newer data is followed by bumping AS_OF_DATE and rerunning make index. make fetch prints a reminder to that effect.

Some sources are prose pages rather than feeds, so their rows are transcribed by hand with source URLs recorded in the CSV. Every folder without a fetch.py states in its Method section how its data is maintained — hand-scored status ledgers, transcriptions from a paper, a series confirmed unmoved as of a stated date — and tools/check.py fails a folder that does neither.

problems/math-alphaevolve-records/fetch.py also writes the sums-and-differences slice into its sibling folder so those datasets cannot drift apart.

Who reads this

The blog at tecunningham.github.io renders the argument these series support. It reads the CSVs here directly rather than holding copies, so a number that goes stale in its prose fails its audit rather than quietly disagreeing with the data. It looks a CSV up by filename, which is why filenames are unique across folders and tools/check.py enforces it. Its figures are the GitHub-hosted PNGs in this repository, embedded by URL, so it holds no copies of those either: what is committed here is what the blog shows.

Provenance and licence

Every CSV records where its rows came from, either in a per-row source column or in the header of the fetch script that built it. The underlying facts belong to their publishers — the curl project, Mozilla, OpenSSL, NIST, CISA, OSV, Google OSS-Fuzz, the Erdős problems community, ANTEDB, Google DeepMind, the Hutter Prize, nextchessmove.com, and the speedrun leaderboards — and are collected here under the terms each publisher offers. The aggregation, classification, and arithmetic are this repository's, and are the part that can be wrong.