8 Oct 2026 · MathPeer-reviewed
A top journal publishes AI-made math
Science published DeepMind's AlphaProof Nexus on 8 Oct. It settled 9 Erdős problems.
Science · PNASPrime Signal · Updated 29 Oct 2026
Singularity Tracker
Tracking the transition in the age of superintelligence.
Singularity Index
52/100▲ 1since Aug
Verified only
29/100▲ 1since Aug
How far AI has come across ten fields, from no role at 0 to driving each field's progress at 100. The 23-point gap is progress that no outsider has checked yet.
The Index
Mathematics leads at 85. Energy trails at 30.
Ten fields, ten milestones each. Brighter squares were checked more thoroughly. Open a field for its milestones and sources.
ReplicatedPeer-reviewedPreprintClaimedPartialNot yet
01 Mathematics 3Underway 8570 verified
Bottom line
Journals now publish AI-made proofs. The biggest 2026 claims, including Navier–Stokes, are not refereed yet.
Holding it back
Peer review can't keep pace with how fast AI produces results.
Next milestone · 9
AI develops genuinely new theory
Time to check a result
Seconds with a proof checker (Lean); months for peer review
- Stage 1 · Loading
- 01AI matches the best humans at competition mathIMO gold under official grading, then perfect scores in 2026ReplicatedJul 2025
- 02AI writes machine-checked proofs of known results at research scaleStrong Prime Number Theorem formalized in Lean in three weeksReplicatedSep 2025
- Stage 2 · Initiated
- 03AI solves a genuinely open problem, with a machine-checked proofErdős problem #728, Lean proof · arXiv:2601.07421Replicated7 Jan 2026
- 04AI disproves a famous conjecture, confirmed by senior mathematiciansUnit distance conjecture (1946), checked by nine mathematicians · arXiv:2605.20695 · OpenAIReplicated20 May 2026
- 05A second lab's AI disproves a major conjectureJacobian conjecture false in dimension 3. Reviewed by Tao, rechecked in Isabelle · Tao · Isabelle AFPReplicated21 Jul 2026
- Stage 3 · Underway
- 06A batch of open problems solved at once, with public machine-checked proofsOpenAI: ten results in Lean (Aug), then 721 manuscripts, about 42% with Lean proofs (Oct). No outside audit yet · Lean repo · openai/mathPreprint1 Aug 2026
- 07AI makes a major advance on a first-rank problemZeta zeros proven on the critical line: 41.6% to 67.2%. A simpler human proof followed · Anthropic · arXiv:2608.13637 · arXiv:2609.02882Replicated10 Aug 2026
- 08A journal publishes an AI-led result after peer reviewNewFirst AI-proved papers in journals (Axiom Math). Science published AlphaProof Nexus on 8 Oct 2026 · Science · Indag. Math. · J. Combin. Theory APeer-reviewedMay 2026
- Stage 4 · Recursive
- 09AI develops genuinely new theoryWatch how experts judge OpenAI's local-smoothing proof · OpenAINot yetExpected 2028 to 2031
- 10AI is the main prover on a Millennium Prize ProblemNewNavier–Stokes, with a Lean proof. Not refereed; no Clay ruling · OpenAIPartial · Preprint8 Sep 2026
02 AI / Recursion 3Underway 6010 verified
Bottom line
AI now speeds up the work of building AI. It doesn't yet invent new AI on its own.
Holding it back
Research taste: AI ideas that score well in the lab rarely hold up in production.
Next milestone · 7
An AI research result holds up in production
Time to check a result
Seconds to days: tests, benchmarks, production trials
- Stage 1 · Loading
- 01AI agents solve over half of real coding tasks (SWE-bench Verified)o3 reported 71.7% (Dec 2024). Latest: 81.4%, Claude Opus 4.6 (Feb 2026). Lab-reportedClaimedDec 2024
- 02An AI-discovered learning algorithm beats the best human one, in a journalDiscoRL · NaturePeer-reviewed22 Oct 2025
- Stage 2 · Initiated
- 03AI-found optimizations run in production at a frontier labAlphaEvolve recovered 0.7% of Google's compute · DeepMindClaimed14 May 2025
- 04AI improves the chips or code it runs onA circuit change in a TPU; 32.5% faster attention kernel · DeepMindClaimed14 May 2025
- 05A model cuts the cost of serving itselfGPT-5.6 Sol: 20% lower serving cost, done with Codex · OpenAIClaimed29 Jul 2026
- Stage 3 · Underway
- 06AI writes most of the code at a frontier lab (over 75%)Over 80% at Anthropic · AnthropicClaimedMay 2026
- 07An AI research result holds up in productionA 0.97 lab score shrank to noise in production · AnthropicNot yetExpected 2028 to 2030
- 08A frontier lab shows AI has at least doubled its research speedAnthropic said no as of Apr 2026Not yetExpected 2028 to 2031
- Stage 4 · Recursive
- 09A frontier model designed end to end by AINot yetExpected 2030 to 2035
- 10AI designs and builds its own successorNot yetExpected 2032 to 2040+
03 Physics 2Initiated 5530 verified
Bottom line
AI finds real results where physics can be simulated. New fundamental physics needs new experiments, decades away.
Holding it back
Experiments take years to decades to build.
Next milestone · 6
AI matches expert physicists on a research-grade test
Time to check a result
Years to decades: build a new detector or collider
- Stage 1 · Loading
- 01AI rediscovers known physical laws from raw dataNewton's second law, energy conservation and more, through Mar 2026PreprintApr 2025
- 02AI runs in the control loop of a frontier instrumentLIGO: 30 to 100x less control noise · SciencePeer-reviewed4 Sep 2025
- Stage 2 · Initiated
- 03An AI-designed instrument beats the human designUrania detector designs · Phys. Rev. XPeer-reviewed14 Apr 2025
- 04AI corrects an accepted physical model, confirmed by experimentDusty plasma: two standing assumptions overturned · PNASReplicatedJul 2025
- 05AI finds new mathematical objects in physics equationsNew fluid singularities (Sep 2025). Plasma counterexamples, preprint (Sep 2026) · DeepMind · arXiv:2609.26742 · UMDPreprint18 Sep 2025
- Stage 3 · Underway
- 06AI matches expert physicists on a research-grade testCritPt 32.3%, up from 5.7% in 11 months · Artificial AnalysisNot yetExpected 2027 to 2029
- 07AI does publishable research end to end, above grad-student levelA paper its own author rated second-year-grad level · arXiv:2601.02484Partial · PreprintJan 2026
- 08AI predicts new fundamental physics, later confirmed by experimentNot yetExpected 2030s
- Stage 4 · Recursive
- 09AI produces a new fundamental law the field adoptsNot yetNo credible window
- 10AI contributes to a confirmed result beyond the Standard ModelWaits on new collidersNot yetExpected 2040s
04 Materials & Chemistry 2Initiated 5040 verified
Bottom line
AI predicts many new materials. Few have been made and confirmed by others, and some big claims were walked back.
Holding it back
It's hard to confirm what a lab actually made, and training data is messy.
Next milestone · 6
An autonomous lab runs a multi-week discovery campaign with no humans
Time to check a result
Weeks to months to make and test; years to scale up
- Stage 1 · Loading
- 01AI-curated materials databases are standard infrastructureMaterials Project: 650,000 usersReplicated2023
- 02AI generates stable crystal candidates at scaleNewGNoME, 2.2M structures. Disputed: 83,000+ removed; duplication preprint Oct 2026 · arXiv:2610.13885Peer-reviewed29 Nov 2023
- Stage 2 · Initiated
- 03An AI-predicted material is made and its target property measuredMatterGen: 158 GPa against a 200 GPa target · NaturePeer-reviewed16 Jan 2025
- 04An outside group makes an AI-predicted compound and confirms itMnFeCo4Si2 · arXiv:2603.22748Preprint24 Mar 2026
- 05AI predicts a new superconductor, then it's made and confirmedYRu3B2 and LuRu3B2, confirmed four ways at Rice · RicePeer-reviewed17 Jun 2026
- Stage 3 · Underway
- 06An autonomous lab runs a multi-week discovery campaign with no humansHumans still move samples · arXiv:2609.30133Not yetExpected 2029 to 2033
- 07An AI-discovered material ships in a commercial productDiscovered Materials claims it; no data published · TechCrunchNot yetExpected 2028 to 2032
- 08An AI-designed catalyst is used in industryBest public result: 20 millilitres · Nature CommunicationsNot yetExpected 2029 to 2034
- Stage 4 · Recursive
- 09Discovery to qualified material in under three yearsNot yetExpected 2033+
- 10AI finds a room-temperature superconductorAI finds so far top out at 6.5 K · arXiv:2604.23758Not yetAfter 2035, maybe never
05 Health & Bio 2Initiated 5040 verified
Bottom line
AI-designed drugs are in trials, one in Phase III. None is approved yet.
Holding it back
Clinical trials take years, and AI can't shorten them much.
Next milestone · 6
A regulator approves a generative-AI-designed drug
Time to check a result
3 to 10 years: lab work, three trial phases, then a regulator
- Stage 1 · Loading
- 01AI protein structure prediction is standard research infrastructureAlphaFold database opens. Independently benchmarked; 2024 Nobel PrizeReplicatedJul 2021
- 02AI improves a clinical outcome in a randomized trialMASAI, The Lancet, 100,000+ women: 9% more cancers foundPeer-reviewed2025
- Stage 2 · Initiated
- 03The first AI-originated drug enters Phase IExscientia DSP-1181ClaimedJan 2020
- 04AI designs new proteins with atomic accuracy5 Nov 2025 · NaturePeer-reviewed5 Nov 2025
- 05A generative-AI drug reaches Phase IIINewRentosertib, 320 patients; first patient dosed 10 Sep 2026 · Insilico · Insilico · Nature MedicinePeer-reviewed7 Jul 2026
- Stage 3 · Underway
- 06A regulator approves a generative-AI-designed drugNone on the FDA's 2026 list · FDANot yetExpected around 2030
- 07AI beats doctors on real patient outcomes, in a prospective trialNot yetExpected 2029 to 2033
- 08A virtual cell model reliably predicts how cells respondNot yetExpected 2028 to 2031
- Stage 4 · Recursive
- 09A disease durably cured by an AI-designed therapyNot yetExpected 2032 to 2035+
- 10AI runs drug discovery, trial design and readout end to endNot yetExpected 2035+
06 Economy & Labor 2Initiated 5020 verified
Bottom line
About 1 in 4 US firms use AI. It doesn't show up yet in productivity or jobs data.
Holding it back
New technology takes years to spread and show up in national statistics.
Next milestone · 6
Productivity grows above trend for three years
Time to check a result
5 to 8 years: adoption spreads, then shows up in national data
- Stage 1 · Loading
- 01At least 20% of US firms use AINewCensus survey: 21.8% (two weeks to 26 Jul 2026). Latest: 24.6% (two weeks to 4 Oct) · CensusReplicated26 Jul 2026
- 02A frontier AI lab passes $50B annual revenueAnthropic past $65B. Company figureClaimedJul 2026
- Stage 2 · Initiated
- 03AI matches experts on half of real work tasksGDPval, 44 occupations, blind gradingPreprint25 Sep 2025
- 04A job effect shows up in payroll dataWorkers aged 22 to 25 in AI-exposed jobs 19% below trend. Working paper · StanfordPreprintJun 2026
- 05AI becomes the top stated reason for announced job cutsFive straight months, Mar to Jul 2026, though small against 158.9M jobsReplicatedMar 2026
- Stage 3 · Underway
- 06Productivity grows above trend for three years0.8% in 2025 · BLSNot yetGoing the wrong way
- 07Job losses show up broadly in national dataNot yetNot seen through Q1 2026
- 08AI agent spending reaches material scaleNewx402: about $54M since launch, self-reported · Coinbase · ChainalysisNot yetMeasurable, not material
- Stage 4 · Recursive
- 09A developed economy grows 4%+ in a yearNot yetExpected 2031+
- 10Labor's share of income falls measurablyNot yetExpected 2031+
07 Robotics 2Initiated 5010 verified
Bottom line
Robotaxis work at scale. Robots that handle objects are still unreliable: 46% success picking items off the floor.
Holding it back
Too little training data, and reliability far below what paid work needs.
Next milestone · 6
Humanoids work real shifts at customer sites, at scale, without remote control
Time to check a result
Trials run daily; proving reliability takes years
- Stage 1 · Loading
- 01Robot AI models transfer to new robot bodies with little dataGemini Robotics 2: under 200 examples · DeepMindClaimed30 Jul 2026
- 02Full-size humanoids cost under $20,000Unitree R1 from $4,900. Public pricesReplicatedJul 2025
- Stage 2 · Initiated
- 03A robotaxi service gives 250,000+ paid driverless rides a weekWaymo: 400,000+ a week. Company figureClaimedFeb 2026
- 04A driverless fleet proves safer than humans over 100M milesWaymo: 220.6M miles, 94% fewer serious-injury crashes. Company data · WaymoClaimedMar 2026
- 05A robot runs a full commercial shift over 95% autonomously96.4% autonomyClaimedFeb 2026
- Stage 3 · Underway
- 06Humanoids work real shifts at customer sites, at scale, without remote controlLargest so far: 3 to 10 unitsNot yetExpected 2028 to 2031
- 0710,000 humanoids in commercial useTeslaNot yetExpected 2028 to 2030
- 08Robots handle new objects and tasks with over 80% successToday: floor pick 45.7%, dustpan 32% · DeepMindNot yetExpected 2028 to 2030
- Stage 4 · Recursive
- 09Physical labor automated at economic scaleNot yetExpected 2032 to 2035+
- 10A robot performs a full surgery on its ownAll 3.1M da Vinci procedures in 2025 were human-runNot yetExpected 2030+
08 Neuro & Interface 2Initiated 4535 verified
Bottom line
Brain implants let paralyzed people talk and type at home. All are still experimental.
Holding it back
Low bandwidth, and no way to write information into the brain.
Next milestone · 5
AI-driven brain mapping yields a new neuroscience finding
Time to check a result
1 to 3 years: surgery plus a long clinical study
- Stage 1 · Loading
- 01AI makes high-bandwidth brain decoding practicalSpeech decoding from implants in Nature. Now 56 words a minute sustained · Nature MedicinePeer-reviewedAug 2023
- 02Brain-implant cursor control matches able-bodied speedOver 10 bits per secondClaimedJan 2026
- Stage 2 · Initiated
- 03A brain implant is used at home, unsupervised, for years3,800+ hours. One participant · Nature MedicinePeer-reviewed15 Jun 2026
- 04AI decodes imagined speechCell, four participantsPeer-reviewed14 Aug 2025
- 05AI-driven brain mapping yields a new neuroscience findingFly connectome: AI aligned the data, humans did the sciencePartial · Peer-reviewed8 Jun 2026
- Stage 3 · Underway
- 06A brain implant restores conversation-speed speech (140+ words a minute)Not yetExpected 2028 to 2031
- 07FDA approves a permanent high-channel brain implantParadromics trial expanded Oct 2026 · ParadromicsNot yetExpected around 2030
- 081,000 patients implantedAbout 150 todayNot yetExpected 2029 to 2032
- Stage 4 · Recursive
- 09Imagined speech decoded without surgeryNot yetExpected 2032+, maybe never
- 10Writing information, not just sensation, into a brainNot yetNo credible window
09 Chips & Compute 2Initiated 4525 verified
Bottom line
AI improves parts of chips that ship today. Humans still design and sign off on the whole chip.
Holding it back
Memory and power supply, and human sign-off on verification.
Next milestone · 3
An AI-designed chip layout ships in production
Time to check a result
Simulation in days; three to six months to make the chip
- Stage 1 · Loading
- 01AI optimizers are standard in commercial chip design toolsNewSynopsys: hundreds of AI-enabled tapeouts (Mar 2024). Cadence: over 1,000 (Feb 2026). Unaudited · Synopsys · CadenceClaimedMar 2024
- 02The cost of a fixed AI capability falls over 100x in under two years$37.50 to $0.12 per million tokens, Mar 2023 to Dec 2024 · Epoch AIReplicatedDec 2024
- Stage 2 · Initiated
- 03An AI-designed chip layout ships in productionAlphaChip layouts in Google TPUs since 2020. Peer-reviewed but disputed · DeepMindPartial · Peer-reviewedJun 2021
- 04An AI-written circuit change ships in a production acceleratorAlphaEvolve · DeepMindClaimed14 May 2025
- 05An AI-designed radio-frequency block beats human designs when builtDesigned in minutes rather than weeks · Nature CommunicationsPeer-reviewed30 Dec 2024
- Stage 3 · Underway
- 06Chip verification signed off with no human reviewSemiconductor EngineeringNot yetExpected 2029 to 2032
- 07A full production chip designed with no human interventionSemiconductor EngineeringNot yetExpected 2030 to 2034
- 08An AI-originated chip architecture beats the human one it replacesNot yetExpected 2031 to 2035
- Stage 4 · Recursive
- 09AI designs, verifies and builds its own next acceleratorNot yetExpected 2035+
- 10AI runs the compute supply chain end to endNot yetExpected 2035+
10 Energy 1Loading 3010 verified
Bottom line
AI helps run plasma experiments and grid queues. It hasn't changed how much energy gets built.
Holding it back
Power plants and grid hookups take 4 to 15 years.
Next milestone · 4
AI controls a net-energy fusion device
Time to check a result
4 to 15 years: build, connect to the grid, switch on
- Stage 1 · Loading
- 01AI controls the plasma in a real fusion reactorTCV tokamak, NaturePeer-reviewed16 Feb 2022
- 02AI speeds up an energy bottleneck at utility scalePJM: 811 grid applications sorted in under an hourClaimedApr 2026
- Stage 2 · Initiated
- 03An AI-found battery material works in a real cellMicrosoft and PNNL: 70% less lithium, lab prototypePreprint9 Jan 2024
- 04AI controls a net-energy fusion deviceSPARC not running yet · CFSNot yetExpected 2027 to 2029
- 05An AI-found design improves a working energy assetNot yetExpected 2028 to 2031
- Stage 3 · Underway
- 06An AI-found battery or solar chemistry in mass productionNot yetExpected 2029 to 2033
- 07AI makes binding grid decisionsTexas' 2026 audit uses human engineers · ERCOTNot yetExpected 2029 to 2032
- 08Fusion power on the grid, with AI in the controlsNot yetExpected 2030 to 2034
- Stage 4 · Recursive
- 09Energy stops limiting computeTexas alone has 474 GW of large-load requests · ERCOTNot yetExpected 2032+
- 10AI-designed energy tech powers AI itselfNot yetAfter 2035
The stages
Seven of ten fields are at Stage 2. None has reached Stage 4.
Every field climbs the same five stages, from AI as a tool to AI running the field. The milestones in the Index are the rungs. A field moves up a stage when it reaches that stage's milestones and passes its test, so points alone don't set it.
None now
0Pending
Before milestone 1AI plays no measured role.
The test
None. Every field starts here.
1 field
1Loading
Milestones 1 to 2AI is a tool. It speeds up work that people lead.
The test
A measured speed-up in real use, not a demo.
Example
AI controls the plasma in a real fusion reactorEnergy · Feb 2022 · Peer-reviewed
7 fields
2Initiated
Milestones 3 to 5AI finds something new, and outsiders confirm it.
The test
A new result credited to AI, checked by someone with a reason to disagree.
Example
An AI-predicted material is made and its target property measuredMaterials · Jan 2025 · Peer-reviewed
2 fields
3Underway
Milestones 6 to 8New AI results become routine. They stop being news.
The test
New results from at least two independent groups, and a visible change in the field's pace.
Example
AI makes a major advance on a first-rank problemMath · Aug 2026 · Replicated
No field yet
4Recursive
Milestones 9 to 10AI runs the field. It picks the problems, does the work and checks it, and each result speeds up the next. In AI itself, that means AI building better AI.
The test
AI carries the field's frontier work end to end, with no human sign-off on the key decisions.
Watch for
AI runs drug discovery, trial design and readout end to endBio · Expected 2035+
What 100 means. Every field has reached all ten milestones and stands at Stage 4. That is the last of the three tests in our definition of the Singularity.
Since 2020
From 12 to 52 in under two years.
The Index by the date each milestone was reached.
Singularity IndexVerified onlyGap
The verification gap
Outsiders have checked 30 of the 54 milestones reached.
By who checked them
13Replicated17Peer-reviewed9Preprint15Claimed
Strongest checkWeakest check
Gatekeepers signed off: 1 of 3
- YesMath journals
Accepting AI-proved papers since May 2026, starting with Axiom Math.
- Not yetDrug regulators
No AI-designed drug approved by the FDA or EMA. First decision to watch: zasocitinib, due Q1 2027, though it was designed with physics-based software rather than generative AI. Rentosertib's Phase III began dosing on 10 Sep.
- Not yetStandards bodies
No standards body has adopted a result that AI originated.
The pattern
Fast to check, first to move.
Fields where a machine can check an answer in seconds are furthest along. Fields that wait on trials, experiments or construction are bunched in the middle.
- Fastest to checkScore
- MathematicsSeconds
- AI / RecursionHours
- Materials & ChemistryWeeks
- Chips & ComputeMonths
- RoboticsA year
- Neuro & InterfaceYears
- Health & BioYears
- Economy & LaborYears
- EnergyYears
- PhysicsDecadesPhysics scores well on work that is quick to check: instruments, models and lab tests. Its milestones that need new fundamental experiments are all still ahead.
Latest
Five stories since August. Outsiders checked two.
Newest first. Each item links to its primary source.
What to watch
Five early signals. Two have new readings since August.
These numbers change before a field's stage does.
5.6 : 1As of Apr 2026
Reliability gap
The length of task AI finishes half the time, divided by the length it finishes 8 times in 10. Past 10 to 1, AI is getting faster than it is getting reliable.
METR data24.6%Up from 21.8% (26 Jul 2026)
US firms using AI
Share of US businesses that used AI in the last two weeks, from the Census Bureau. The survey changed in Nov 2025, so older readings don't compare.
Census10 / 49Up from 4 / 49 (Aug 2026)
Open math problems solved
Problems on Epoch AI's list of open math problems that AI has solved, alone or with people. Fewer than 15 by Aug 2027 would mean progress has stalled.
Epoch AI0.8%As of 2025
US productivity growth
How much more the US economy gets from the same work and capital each year. Two years above 1.5% would be the first sign of AI lifting the economy. Next reading Mar 2027.
BLS474 GWAs of Jun 2026
Texas large-load power requests
Power that data centers and other large users have asked the Texas grid to supply. Compare it with power actually switched on: a wide gap points to overbuilding.
ERCOTOur definition
The Singularity is when superintelligent AI improves itself and drives invention in every field.
We'll call it when all three tests below are met and outsiders have checked the evidence. Each test reads straight from the data and updates every week.
Test 1First signs
Superintelligence Score
How far AI has surpassed human intelligence and ability, across eight abilities we can measure against people.
57.3/100all evidence
41.7/100verified only
Ability1 · Matches experts2 · Surpasses the best3 · Surpasses whole fieldsScore
Math100100 verified
Science8383 verified
Coding7533 verified
Engineering5033 verified
Physical skill5033 verified
Professional work500 verified
Forecasting3333 verified
Medicine1717 verified
ReplicatedPeer-reviewedPreprintClaimedPartialNot yet
Each ability faces three steps against people: 1, as good as skilled professionals; 2, beats the best person, or the best human design; 3, does what took whole teams, or what a whole field couldn't. Step 3 counts three times as much as step 1, so the score only reaches 100 when AI goes past whole fields in every ability. This scores abilities, not fields: AI's progress in each field is the Index.
Test 2First signs
Builds better AI, faster
Recursive self-improvement. AI improves the code, chips and models it runs on, and each new version arrives sooner.
Read from the AI / Recursion row of the Index. Met when outsiders confirm milestone 10: AI designs and builds its own successor.
Test 3First signs
Invents in every field
AI makes progress, inventions and innovations in all ten fields, not just a few.
9/10fields with an AI discovery outsiders confirmed
0/10fields at Stage 4
Met when all ten fields reach Stage 4 and outsiders have checked every milestone.
No test met yet
Where the idea comes from
The idea
In 1965 I. J. Good described a machine that could far surpass any person's thinking: the last invention we'd need to make. Vernor Vinge named what comes after it the Singularity. Ray Kurzweil puts it at 2045.
Superintelligence
Nick Bostrom's 1998 definition: much smarter than the best human minds in nearly every field. Google DeepMind's 2026 version: more capable than large organizations of people. The top two steps of our score follow them.
Already here?
Sam Altman wrote in June 2025 that the takeoff has started. Alex Wissner-Gross's The Innermost Loop and Peter Diamandis's Moonshots say we're living in it. We check that claim against the record.
How it arrives
Field by field, not on a single day. The pattern shows which fields go first.
Open data
100 milestones and 84 sources. Free to use.
Every number on this page comes from a small open dataset. Download it, query it, build on it. Credit Prime Signal and link back.
README.mdSchema, scoring rules and example queries. Start here.5 KBmilestones.csvAll 100 milestones: status, who checked it, date reached, evidence, sources.100 rows · 15 KBability_steps.csvThe Superintelligence Score: eight abilities, three steps each against people.3 KBfields.csvThe ten fields: stage, time to check a result, bottom line.10 rows · 3 KBsources.csvEvery primary source, by type.84 rows · 7 KBstories.csvThe update feed.9 rows · 1 KBmetrics.csvReadings for the numbers we watch.8 rows · 647 Btracker.jsonEverything above in one file, with scores and history.100 KBtracker.dbSQLite database with score views. Open it with any SQLite tool.88 KB
Query it with SQLite
sqlite3 tracker.db
SELECT name, score, verified_score
FROM field_scores
ORDER BY score - verified_score DESC;- License
- CC BY 4.0. Free to use and share with credit.
- Cite as
- Prime Signal, Singularity Tracker, updated 29 Oct 2026. primesignal.org/singularity-tracker.html
- JSON
- data/tracker/tracker.json
How it works
What we track, and how.
We score only what's on the record.
The ten fields
Mathematics
Competition math, machine-checked proofs, open problems, journal acceptance.
AI / Recursion
Coding agents, AI-found algorithms, AI improving the chips and code it runs on.
Physics
AI in instruments, rediscovered and corrected laws, predictions confirmed by experiment.
Chips & Compute
AI in chip design tools, AI-designed circuit blocks, autonomous chip design.
Health & Bio
Protein design, AI in clinical care, AI-designed drugs in trials and approvals.
Materials & Chemistry
AI-predicted materials, lab synthesis, autonomous labs, materials in products.
Robotics
Robot AI models, humanoid cost, robotaxis, real-world deployments.
Economy & Labor
Business adoption, AI revenue, job effects, productivity, AI agent payments.
Neuro & Interface
Brain-computer interfaces, speech decoding, approvals, patient counts.
Energy
AI in fusion control, grid planning, new energy materials.
The score
An achieved milestone is worth 10 points and a partial one 5, out of 100. The verified score counts only peer-reviewed or replicated results. The Index is the average of the ten fields.
Who checked it
- Replicated. Reproduced or rechecked by outsiders with no stake in it.
- Peer-reviewed. Accepted by a journal after peer review.
- Preprint. Published openly, not yet reviewed.
- Claimed. The lab or company says so.
Rules we follow
- A lab's claim alone never moves a field up a stage. Someone with a reason to disagree has to check it.
- Finding an answer that already exists is not discovery.
- Requested power capacity is not built capacity.
- A demo is not a deployment.
Sources and method
Sources. Every milestone links to its primary source: papers, company pages, regulator lists, grid filings and government data. Journal citations are checked against Crossref, and big claims are read on the claimant's own page.
The curve. The history uses the date each milestone was reached and its current check level. The milestones were chosen in 2026, so earlier years count only what this ladder measures. Dates given as a year sit at mid-year.
Independence. Ratings read the public evidence; they are not an endorsement. Prime Signal has no financial ties to the labs or companies named.
Earlier editions
- v1.0, 20 Aug 2026. First edition. Ten fields, 100 milestones.
One email a week: what moved, and what got checked.
Get the weekly update ↗