How to read this site

One tracker, two lenses. The diffusion lens follows Narayanan and Kapoor's AI as Normal Technology: five stages a technology passes through (the paper calls them stocks), a flow between each (valves), and normal and fast ranges with thresholds that would falsify the argument. The capture lens measures who keeps the surplus, layer by layer. They are the same question asked from two ends, so they share one data store, one indicator object and one discipline, adapted from the AI 2027 tracker.

Three layers, strictly separated

Observation (raw, sourced, dated) → Derived (a formula over observations, defined in the semantic layer) → Indicator (proxies, status, confidence, evidence, counterevidence). No derived number exists without a traceable path to observation ids, and the site never computes a number: every figure was exported with the ids behind it.

Evidence tiers

The tier is stored; the letter grade is derived for display. A = tier 1, or tier 4 audited or government statistics. B = tiers 2–3, tier 4 company statements (8-K), tier 6 with a transparent method. C = tier 5, or tier 6 estimates. D = tier 7 or single-source. Tier 7 evidence never moves a status above emerging. Audited is reserved for tier 4 and a 10-K; a run-rate reported by the press or by the company is never audited, whatever its tier.

Status vocabularies

Flow indicators (diffusion): consistent with normal · faster than normal · slower than normal · emerging · not yet measurable, each against a normal band, a fast band and an optional falsifying band with a written rationale. Capture indicators: concentrating · dispersing · stable · unclear, from a trend over N periods with a dead-band. Every indicator also carries a timing tag: leading (moves before capture shows in filings), coincident (moves with it) or lagging (confirms it after the fact). Predictions: confirmed · ahead · on track · behind · emerging · not yet testable. Not yet measurable means the thing cannot be measured yet; an indicator whose connector has not landed is simply unpublished.

Bands

Each diffusion indicator names the single series or derived metric its bands apply to (shown as “applied to” on the indicator page), a normal band anchored on the pace of an earlier general-purpose technology at the same age, a fast band anchored on the pace the AI 2027 scenario needs, and where one exists a falsifying threshold. The rationale is written next to the numbers. Bands change only through a reviewed pull request that states why; the evaluator never moves them.

Confidence, 0–95, independent of status

90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

How we frame value capture

The capture lens asks who keeps the value AI creates. Five ideas organise the reading, and none of them supplies a number. The value stick: a transaction creates value between what a buyer would pay and what a supplier would accept, and each firm keeps the slice it can defend against rivals, buyers and suppliers (Brandenburger and Stuart, 1996). Schumpeterian profits: innovators have kept only a small share of the surplus their innovations create, and most of it reaches users (Nordhaus, 2004). Complementary assets: when imitation is easy, the owners of distribution, manufacturing, data and customer relationships capture the profits rather than the inventor (Teece, 1986). Installation and deployment: a technological revolution's installation period runs through irruption, when the technology finds its first uses, and a frenzy financed by speculative capital; it ends in a turning point, after which synergy and then maturity spread the technology through the wider economy (Perez, 2002). The site places its readings by a published rule on spending per dollar of AI revenue and on deals where a seller funds its buyer; when the finer phase cannot be told apart, it shows only the half. Kinds of rent: scarcity rents on a constrained input, scale and network rents, switching-cost rents and regulatory rents, which migrate as bottlenecks move. The Ask model is given the same framing and may use it to organise an answer, never as the source of a figure.

Discipline, enforced in code

• Every observation carries a URL that resolved at ingestion, its HTTP status, a content hash, the retrieval time, the tier, an audited/company-stated/reported/estimated flag, the extraction method and the raw snippet.
• Figures quoted from documents enter through a manual connector that fetches the page and asserts the snippet is a verbatim substring of it. Nothing is typed into an indicator.
• No single-source status change unless the source is a benchmark, a model release or an official filing. Non-empty counterevidence before anything publishes.
• Status lives only in status events, each with a reason, evidence ids and an author. The changelog is generated from them. Observations are superseded, never deleted; disputed figures carry their dispute text everywhere they appear.
• The evaluator proposes; it writes its own reason only for a first reading inside a band, and a band crossing waits for a human. Bands and formulas change only through a reviewed pull request.
• Rebuilt nightly at 06:17 UTC into a pull request that merges itself when the consistency checks, the tests and the site build all pass on exactly what it built; a failure holds it for review. What merges that way was decided by a published rule: observations that passed their source's checks, and first readings inside a band. A band crossing still waits for a human-written reason, and the card says so meanwhile.
• Fetched HTML passes through a prompt-injection scrub; anything addressed to an AI agent is logged and never stored.

Query layer

The Ask button and the console on /query run against the same store the site is exported from, on a separate read-only service whose database cannot read files or the network and cannot see pending or superseded rows. A question goes to a model with nine tools: read-only SQL over the semantic layer, a metric lookup, an indicator lookup, status changes since a date, a keyword search over observation snippets and the public research notes, a one-off trend fit, the labour-tracker concordance, the crosswalk, and an entity lookup. Every number in an answer must be followed by the citation token of the record it came from; a post-check extracts the numbers and verifies each against that record's value, confidence bounds, band edges or verbatim snippet. Research notes found by search carry no citation, so a number found only in a note cannot pass. An answer that fails is revised once; if it still fails the question is asked afresh once, and otherwise the answer is shown as blocked with the unverified numbers marked. The service cannot write observations or statuses and has a daily spend cap. For each answer it keeps the time, outcome, cost, number of lookups, the ids of the records cited, the model and the prompt version, never the question or the answer, and the nightly copies those records into the repository.

X posts

The tracker makes no request to X: its robots.txt disallows the embed endpoint, and the API is paid. A post by someone the tracker reads enters only when a page it already fetches carries the post's text (an article quoting it, or a static embed), with that page as the source, the text as the verbatim snippet and the X link taken from the page. Every post is tier 7 and recorded as a watchlist row; when a post points to a tier 1–6 artifact, the artifact is fetched and entered in its own right. Posts never move a status, and the site shows their text with a link, never X's script.

Colour

Three colours carry meaning and nothing else does. Violet and teal mark direction (faster or concentrating; slower or dispersing) and always ship with an icon and a word. Orange marks how tightly an input binds, darker for tighter. Nothing is red or green: fast is not good, and concentrating is not good.

Credits

Method after the AI 2027 tracker (independent, not affiliated).

Data credits, generated from the source registry: METR, Measuring AI Ability to Complete Long Tasks (Horizon v1.1); SEC EDGAR, XBRL company facts API; METR, Time Horizon 1.1 (29 Jan 2026); Epoch AI, 'Data on AI Companies' (CC BY 4.0), epoch.ai/data (CC BY 4.0); U.S. Bureau of Labor Statistics, Productivity and Costs (PRS85006092/93, PRS85006173), Total Factor Productivity (MPU4910012); Bick, Blandin & Deming, Real-Time Population Survey, via FRED; Brynjolfsson, Collis, Eggers, Kazinnik & Nguyen, 'What is Generative AI Worth?' (2026); MIT NANDA, 'The GenAI Divide: State of AI in Business 2025'; U.S. Census Bureau, Business Trends and Outlook Survey; Menlo Ventures, '2025: The State of Generative AI in the Enterprise'; ICONIQ, 'State of AI: The Builder's Economy' (2026), a survey of over 300 software executives building AI products; HSBC Insights, 'The billions of AI consumer surplus' (10 Aug 2026); Fortune, 'MIT report: 95% of generative AI pilots at companies are failing' (18 Aug 2025); FRED Blog, 'Does generative AI save time at work?' (27 Aug 2026); Bick, Blandin & Deming, Real-Time Population Survey, retrieved from FRED; Federal Reserve Board, Bureau of Economic Analysis and Bureau of Labor Statistics series, retrieved from FRED; Canaries Dashboard, a project of the Stanford Digital Economy Lab and ADP Research; Stanford Digital Economy Lab and ADP Research, AI Economic Indicators: Canaries Dashboard; Revelio Labs, AI Labor Market Tracker; Hyman, von Wachter et al. (2026), Tracking AI-Related Job Loss Using Unemployment Insurance Claims Data in California, California Policy Lab; California Employment Development Department, AI and the economy; Anthropic Economic Index, open dataset on Hugging Face (CC BY 4.0) and the 2026 reports (CC BY 4.0 (dataset)); Federal Reserve Bank of New York, The Labor Market for Recent College Graduates; SEC EDGAR, inline XBRL filings (segment disclosures); SEC EDGAR archives; NVIDIA Newsroom; Official Microsoft Blog; Microsoft Investor Relations; Anthropic newsroom; About Amazon; CNBC; TechCrunch; Comet; Independent Insurance Agents & Brokers of America, IndependentAgent.com; Claims Journal; Hunton Andrews Kurth, Insurance Recovery Blog; Chaucer Group; AIUC; UiPath; KPMG LLP; Insurance Journal; GeekWire; Cloudflare Radar, via the Cloudflare blog; SiliconANGLE (relaying the WSJ report); Investing.com via Yahoo Finance; CDS data ICE Data Services; Ramp AI Index, Ramp Economics Lab; OpenAI; METR; Anthropic, Claude Fable 5.1 & Claude Mythos 5.1 System Card (1 Sep 2026); Sacra estimates (free company page); The AI Insider; CoreWeave (company announcement); Leonis Capital, The Thesis; Federal Reserve Board, FEDS Notes; TIME; Stanford Institute for Economic Policy Research; MultiState; Transparency Coalition.AI; Mercor (company announcement); South China Morning Post; Andreessen Horowitz; Stanford HAI, AI Index Report 2025; Jamin Ball, Clouded Judgement; OpenRouter list prices; SEC EDGAR Form D data sets; Epoch AI, Notable AI Models (CC BY 4.0) (CC BY 4.0); Epoch AI, Machine Learning Hardware (CC BY 4.0) (CC BY 4.0); Epoch AI, LLM inference price trends (CC BY 4.0) (CC BY 4.0); Epoch AI, Benchmarking Hub and Epoch Capabilities Index (CC BY 4.0) (CC BY 4.0); Epoch AI, AI Chip Sales (CC BY 4.0) (CC BY 4.0); Epoch AI, Data on AI Chip Components (CC BY 4.0) (CC BY 4.0); Epoch AI, AI Data Centers (CC BY 4.0) (CC BY 4.0); Epoch AI (CC BY 4.0) (CC BY 4.0); FDA, Artificial Intelligence-Enabled Medical Devices list; NCSL, Artificial Intelligence 2025 Legislation; rl-list.com; Our World in Data, AI investment (CC BY 4.0), from the Stanford AI Index (CC BY 4.0 (OWID; underlying data AI Index / Quid)); UN World Population Prospects (2024), V-Dem (2026) and the International Labour Organization, processed by Our World in Data (OWID charts CC BY 4.0; UN World Population Prospects CC BY 3.0 IGO; V-Dem CC BY-SA 4.0; ILO CC BY 4.0); Google DeepMind; arXiv; Fireworks AI; Applied Compute; Joy Larkin, cleverhack.com; Crunchbase News; PE Hub; SEC EDGAR; Baseten; Thinking Machines Lab; Cursor (Anysphere); Harvey; Engram; Decagon; Google Research; NBER; MIT Sloan; People Matters; Automation Anywhere; AI Futures Project; AI 2027 Tracker; The Budget Lab at Yale; Humlum & Vestergaard; Chicago Booth; CEPR VoxEU; WashU Olin Business School; PwC 2026 Global AI Jobs Barometer; Outsource Accelerator; Dealroom; Dwarkesh Podcast; The Motley Fool's transcripts of company earnings calls; FreightWaves; En-Contact; sennder (company announcement); Shield Technology Partners (company announcement); MSP Summit; Exigent Technologies; UnifyCX (company announcement); Beacon Software (company announcement); Corum Group; Latham & Watkins; TV Tech; Cyclingnews; Vimeo; Variety; Skift Meetings; Tractive; Bending Spoons (company announcement); Circeus (company announcement); BusinessCloud; Kaufman Hall; Law Society Gazette; Tech.eu; 2X (company announcement); Rutland Herald; CPA Practice Advisor; Current (formerly Crete Professionals Alliance) (company announcement); Property Industry Eye; The Negotiator; Alpera Partners (company announcement); actuEL EC brief republished on lexia-conseil.fr; Munich Startup; Brazil Journal; Iterative Health (company announcement); Loop (company announcement); Metropolis (company announcement); Oosto; Electrek; FDA 510(k) database; STAT News; Innolitics; PYMNTS (relaying Bloomberg); OpenRouter; HAL: Holistic Agent Leaderboard, AI Agent Reliability Tracker, Princeton University; Sierra, tau2-bench leaderboard; GetDeploying, GPU rental price history (CC BY 4.0) (CC BY 4.0); StatCounter Global Stats (https://gs.statcounter.com), CC BY-SA 3.0 (CC BY-SA 3.0); Pew Research Center; Microsoft Source; SoftBank Group; US Census Bureau, Statistics of US Businesses; Artificial Analysis; Stripe newsroom; NVIDIA Blog; AMD blog; Calcalist Ctech; Crusoe newsroom; d-Matrix announcements; Emmi AI news; Nicolas Bustamante's blog (Fintool founder); Git AI blog; Google blog (The Keyword); Gretel blog; HKEXnews (listed-company filings); Humanloop website; Koyeb blog; Limitless website; Manus blog; Mendral blog; Nebius newsroom; NextSlide website; Nscale press releases; The OpenAI Deployment Company newsroom; Poolside blog; Silicon Republic; Stilla blog; Suno blog; Workshop Labs blog; World Labs blog; AWS Contact Center Blog; Business Insider; Cirrus Labs website; Jams website; The Media Leader FR (with AFP); Weights website; AWS Big Data Blog; Forbes; LawSites (Bob Ambrogi); Tech Startups; Glean (company announcement); Sierra (company announcement); Newcomer; Lovable (company announcement); ElevenLabs (company announcement); Legora (company announcement); RuntimeWire; Clay (company announcement); 24/7 Wall St. (relaying The Information); Runpod; The Next Web; Gartner; S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning; CIO Dive; CIO.com; Lenovo CIO Playbook, with research by IDC; McKinsey Global Survey, 'The state of AI'; Upstarts Media; Abridge; Cision PR Newswire; GlobeNewswire (Notified); Hebbia; Together AI; Modal Labs; Cognition; EvenUp; Parloa; Ambience Healthcare; Hippocratic AI; Exa; Parallel Web Systems; Factory; LangChain; Browserbase; E2B; Rogo; Crescendo; Artisan; Mastra; Institutional Investor LLC; Yahoo Inc.; Kadan Capital; Lindsey Li, Josh Hechtman, David Cowan & Brielee Lu, Bessemer Venture Partners; Rishi Maheshwari, Dallas VC; Thomas Krane & William Blackwell, Insight Partners; neolabs.fyi; BetaKit; Deepgram (company announcement); Midjourney (company announcement); Rebellions (company announcement); AI Roll-up Nexus (ai-rollup.fyi); Caritas Venture Co. (itself a transformation partner on the map); Suggested by Alex Brogan (Dragonfly); research sweep 28 Sep 2026; Slow Variables research sweep, 28 Sep 2026; Trajectory (Ronak Malde, Arjun Karanam, Michael Elabd); Joy Larkin, cleverhack.com trackers and Awesome-AI-Market-Maps.

How the charts read

Series marks are one slate, at three lightnesses, told apart by labels and dashes; the colours are reserved for meaning: violet for faster and concentrating, teal for slower and dispersing, orange for how tightly an input binds. Text is always ink: a colour appears only on a mark, a glyph or a border. Colour never carries a verdict on its own, so every status also has an icon and a word, and no chart uses red and green. The three colours were checked together for colour-vision separation on the white chart ground, for normal, protan, deutan and tritan vision. Hatching marks what is not firmly known. On the capture charts that is an estimated layer, hatched rather than tinted. On the tightness scorecard every score already rests partly on estimates, which its grade says, so hatching there marks a score whose confidence is low. Every chart has one vertical scale, names its sources beneath it, and repeats its numbers as text: each one links to the row it came from, and each series has a table and a CSV.

Tightness scores

The migrating bottleneck scores twenty-three inputs to AI, from minerals to money, on how hard each is to get. The scale runs from nought to a hundred and is read in five words: slack from 0, easing from 20, moderate from 40, tight from 60, severe from 80. It belongs to that page only. It is a different thing from the speed ranges above, and it never moves an indicator's status.

Gauges. Each input is read by a few gauges. A gauge takes one figure from the store, such as how fast the supply of chip packaging is growing, and turns it into points on a scale set by hand: a list of readings and the points each is worth, with straight lines between them and a flat line beyond either end. A reading off the end of its scale is marked as such. An input's tightness is the average of its gauges' points, weighted, shown as a whole number. Every scale, with the reasoning behind it, is listed below; they were set by judgement and change only in public, with a written reason.

No data, no score. A gauge is dropped when it has no reading or its reading is older than its age limit. An input is withheld, never shown at fifty, when none of its gauges has a usable reading, when the one it depends on is missing, or when less than 35% of its intended evidence is fed. Each withheld input says why, and the reasons are kept apart: nobody publishes a series; one is published and this site does not read it yet; the gauge available does not measure what it claims; or the readings are too old or too few today.

Confidence. Beside every score is a confidence from nought to a hundred. It is worked out, not set by hand, and so it differs from the confidence on indicator pages. It rests on three things, multiplied together, each raised to a weight:

• Coverage (0.28). How much of the intended evidence has a usable reading.
• Freshness (0.2). Freshness halves each time a reading ages by three quarters of its gauge's age limit, so a reading at its limit still counts for about two fifths and one past it is dropped.
• Source (0.29). An editorial table, not a measurement: it turns the site's evidence grade for a gauge's weakest input into a factor between nought and one. An audited annual filing counts in full; a quarterly filing, which is a company's own unaudited statement, counts for nine tenths, and a gauge takes the grade of its weakest row, so a total over four quarters is graded as a quarterly filing. A respected group's estimate, which is what most of these gauges rest on, counts for three quarters.

Because the three are multiplied, one weak factor drags the whole figure down. Coverage and source quality count most, and about equally, because a score resting on a small part of the intended evidence, or on a weak source, is the easiest to misread; freshness counts least. The prototype also weighed the width of each source's error range and the agreement between news reports; this store carries neither, so those are left out and the rest are rescaled.

The result cannot exceed the input's ceiling, which is set by hand and explained below. A bar is hatched when its confidence is low. With these weights that happens only where an input's ceiling is set low, which is done on purpose for scores that rest on a stand-in. A score whose confidence fell under 20 would be withheld. Most of these gauges share one source, so their confidence moves together, and a gap of under ten points between two inputs means nothing.

Age limits. A reading of a shortage is about now, and a quarter more than about three quarters old no longer says what is short now. So no gauge is read once its newest reading passes that age, even where the publisher is slow. Epoch's chip-components file arrives with its newest complete quarter already more than two quarters old, and its chip-sales file is read only from a quarter in which every designer reported; when neither has a newer complete quarter, those gauges are withheld and say so, rather than being read from an older quarter.

The bottleneck map

The bottleneck map lays the chain of inputs AI is made from, and the frictions Arvind Narayanan and Sayash Kapoor describe beyond it, against the stages of diffusion. On that page a bottleneck is whatever sets the pace, a scarce input upstream or a slow institution downstream. Each row prints today's reading once: an input's tightness score, or, for a friction, how many of its published indicators read faster than normal, normal or neither. A cell carries marks, never a colour: the row acts on that stage; nothing this site reads measures it; a named writer, or this site, expects it to bind there, one mark each. A claim about what happens on a row that does not say the row will bind, such as who gains or loses, is listed under the row and marks no cell. Where a stage is this site's placement rather than the author's, the cell says so.

Signposts

Some readings are dated records of what AI can do, such as the best score on each capability test, or how fast agents grow more accurate against how fast they grow more reliable. They carry no faster-or-slower status, because nobody has yet defended a normal pace for them. They are tested only through the claims on the outlook page.

Claims about what happens next

What happens from here sets out named writers' positions on the questions the outlook turns on. Each position names who holds it, or says it is this site's own, or this site carrying an author's argument further than they did, and each names a rival position. Each claim that has a test is read every night against one reading and one line; a claim with no test reads can't be tested yet. It is holding when the reading meets the line, failing its test when it does not, and can't be tested yet when there is no reading, the reading is too old, or the claim is about reaching a line by a date that has not come and the reading is still short of it. When the rival position expects the same reading, the claim reads “both sides expect this” instead of counting as a win for either side. A claim never changes an indicator's status. The predictions ledger, which scores dated forecasts people made in their own words, is kept apart.

Where the board marks a forecast too early to tell, a model's lean may sit beside it: one of five words, from likely true to likely false, with a short reason. It is a judgement, not a reading. It never changes a status word or a tally, it is dropped once a reading settles the forecast, and the query service does not see it. One named model made the leans in groups of related forecasts, extending a trend where a reading had one and saying so; where it had no reading to rest on, the lean says it rests on the model's general knowledge. Separate reviewing runs of the same model then changed some and withdrew one; no person reviewed each. The reasons type no figure. A lean is not the confidence number on a forecast's card: confidence is how firm the evidence behind a status is, and a lean is whether the model expects the forecast to come true. The model is made by Anthropic, which some of these forecasts concern.

Forecasts on the singularity timeline

The singularity timeline places dated forecasts of AI milestones, from essays of the 1960s to this year's revisions, on one row per milestone. Each forecast is a prediction in the ledger, written in this site's words, credited to its author and linked to its source; a book is reworded and credited, and a public work may carry one short exact quote. A mark sits at the year forecast: the most likely year where one was stated, otherwise the last year of the window, with a bar for a stated range; when it was said shows on hover, in the table and through the year slider. Calls that give only odds, or that bet against a date, are listed, not placed, and a claim about a step toward a milestone is drawn lighter and kept out of the forecaster table. The axis squeezes the decades before 2020 and after 2050 so the years around now get most of the width, and marks the places it changes.

Each forecast carries one of the board's five words, from a status this site wrote with a reason: a window that has not opened reads too early to tell; a window that closed without the milestone reads slower than said; a forecaster's or a lab's own grade never settles one. Each milestone lists the readings this site takes that bear on it, and the four worlds for 2036 are read through the cells of the scenario grid, so a world reads consistent with tonight's readings until one of its signposts fails. Novels and stories sit on a last row, unscored: they are imagined futures, not forecasts.

The Singularity Atlas

The atlas gathers what named writers expect AI to change in eight parts of life, from money and work to health and an ordinary day. Each expectation is one writer's view in this site's words, credited and linked to a source this site has fetched, and placed by era (now and on the way, the first decade after AI can do most of what people do, and the long run) and by which of the four worlds it assumes; a view that holds whichever way things go counts in every world. They were drawn from the singularity canon by reading each work, and every line was checked against its source for who holds it, its conditions and its dates. Expectations are not scored and do not sit on the predictions board: most name no date and no test.

The map counts sourced works, not writers and not odds: two lines from one work count once, and a darker cell means more works expect change there, never that change is more likely. The tint steps at fixed counts, so each world's map reads the same way. Beside each part of life sit the public series this site reads for it, each shown with the year of its latest reading, since several are annual; culture has none. The plates are illustrations made in v0, Vercel's AI design tool, for the site's owner in September 2026, recorded by the fingerprint of each original file; they illustrate and are not evidence.

The automatability census

The automatability census asks which US knowledge work passes a structural hand-over screen, task by task: work whose shape would let it be handed to software. It is not a claim about what AI can do, and not a forecast. It was built in a separate project from the Labor Department's list of the tasks in each occupation (O*NET), the Bureau of Labor Statistics' wage bills by occupation and industry, and Anthropic's Economic Index of how people use its models. A task passes when at least two of three AI models each find that whether it was done right can be told within hours, a check that already exists settles it, and a single failure is cheap; work that needs a body and sign-offs a person is accountable for are left out. The rule was frozen before any score in this version existed, though it was shaped after seeing earlier usage data, so tests against that data are post hoc; no person judged individual tasks.

This site imports each version of the census as published and never recomputes a verdict or share. It refuses a draft, checks every file against the census's own list of fingerprints, checks that every table adds up to the same totals and that the published scores reproduce every verdict, fetches every source a deal card links to, and loads the tables into the query console. Census figures are a judged snapshot, not observations, so they carry a reference to the file and row they come from instead of observation ids. Every total that passes appears beside the part all three models pass, and each model's own total is shown too, because they disagree a great deal. Passing is not saving: the one modelled figure, a saving that assumes a cost cut and how easily one kind of work stands in for another, sits beside the headline labelled as modelled, and the query service will not cite it. Whether a business keeps a saving was not measured.

The list of trades on the census page sets the census beside what buyers do. The roll-ups are the ones this site files under AI roll-ups, placed in a trade and a country by one public directory read once; their purchases come from a dated search of the larger buyers only, each row quoted from a page this site fetched. A firm a buyer announces as joining its group is listed in those words, since no stake is stated. The roll-ups are counted worldwide and the census payroll is American. Where the census's reasoning and the buyers part ways, the page says so and settles nothing.

Futures: imagined technologies and who would profit

Futures lays out technologies imagined in science fiction, from Technovelgy's glossary of science-fiction inventions as compiled in Not Boring's Sci-Fi Idea Bank, plus dated forecasts from the books and essays of the singularity reading list behind the singularity timeline. Names, works, authors and years are the glossary's; every description is this site's own rewording, checked when it was written so that no line repeats a run of five words from the original. Whether an idea has been built, and when, is as the idea bank marks it: a real thing that existed before the story reads as already existing and is left out of the time it took to arrive, and a date given only vaguely is shown as unclear rather than guessed.

Everything else was judged by AI models, not people. For each idea, three models each chose its kind of technology, the decade a working, commercial version is most likely, and what must happen first; the answer at least two chose is kept, and where all three differed, the answer of the strongest model decides and the row says so. How the profit would work was judged by the strongest model alone, told to judge each market as if the thing already existed and was on sale, because a first attempt without that instruction mostly treated anything not yet built as hard to copy. It answered whether the thing is a technology at all, whether it would be sold, built once for its owner, or banned, what kind of profit it could earn, whether a rival could copy it within a few years, whether reaching customers needs assets few hold, and who holds them.

Two things are then worked out by fixed rules, never by a model. Where the profit ends up follows David Teece, who studied why inventors so often lose the profits of their inventions: a maker that can stop copying keeps them; when copying is easy and customers can only be reached through assets someone else owns, that owner collects; when both are easy, competition passes the gain to users. How large the profit is (thin, moderate, fat or monopoly-like, from a small or brief margin to one rivals can barely touch) follows its kind and how long it lasts, stepping up for profits that grow with the number of users or come from a licence, after David Ricardo on rent going to whatever cannot be expanded and Eliyahu Goldratt on shortages moving once relieved. Nothing that is not a technology, would not be sold, needs physics we have no evidence for, or could not be judged gets a result. Wrong answers remain; reviewers checked samples against the sources, and reviewed corrections carry their reason on the row. The forecasts from the canon were each checked against the passage they come from by three reviewers; a quote appears only when the forecaster wrote the work. The illustrations were made with AI image generation from prompts written from this site's own lines: the first set by the site's owner in ChatGPT, the rest with Qwen Image, generated and checked one by one by an AI agent. They were reviewed for text, likeness and accuracy, and one that failed is not shown. They are not evidence.

The profit rubric above was the first judgement of profit, and it now shows only on the few ideas the site could not place in a kind. Every other idea carries a second judgement, made in October 2026 by Claude, a model made by Anthropic, and never a reading. The model was asked to judge each idea as if it worked the way its story promises, whether or not it looks possible, and to answer in words: how large a prize it would be (vast, large, small or no prize), who would pay (households, businesses, a government or nobody), and which seller would keep what is left once buyers have had their gain (no seller in particular, the maker, existing firms or permit holders), with a reason for the prize and a reason for who is paid, and whether it is fairly sure. One run judged each batch of ideas of one kind; a separate run reviewed every batch and changed answers; a further run gave near-identical ideas within a kind one answer; a last run did the same across kinds. No person read each one. The counts on the Futures pages count these judgements and say so, and the site's Ask reads them and is told to say whose judgement they are; the owner chose that for this section. One rule was set afterwards across every kind, and the ideas it covers were placed by reading each description, then checked by a separate run: anything done to a person without their say, to hurt them, break their mind or override their will or body, has no prize, whether the person is a captive, a worker or a whole population. A weapon or a restraint that only stops or holds someone keeps its prize, and so do confinement, anything the person chooses for themselves, and a device that only reads or watches; where a description does not say who chooses, the person is assumed to. The weapons kind also gives no prize to poisons and gases.

Reuse, citation and corrections

Code is MIT licensed. The operator's own contributions to the compiled dataset (observations as recorded, derived values, status events) are CC BY 4.0; each observation also carries its upstream source and licence, which govern that row, and third-party quotations, reworded source content and third-party selections are not covered (see the legal notice). Cite the site by URL and the observation ids you rely on. Corrections: open an issue or pull request on the repository; every change to a number or a status leaves a dated event in the changelog.

A model's judgement where there is no reading

Where the site has no reading, or holds its verdict back, a model's judgement may sit beside the empty slot: an input the tightness scorecard cannot score, a barrier nothing here measures, an expectation in the atlas, an indicator held at emerging or unpublished. Each is a word and a short reason, marked with a dotted rule and the words “a model's judgement, not a reading”. It is never a figure, a score, a status or a source. Nothing reads it: no status, tally, metric or answer from Ask uses it, and it is removed when a reading fills the slot. One named model wrote them, from the site's readings where it had any and otherwise from its own general knowledge; separate reviewing runs of the same model then changed some and withdrew others, and no person read each one. Where the model's word runs against a reading on the same row, the reason says so first. In the atlas a verdict says whether the thing a writer expects actually happens, by the writer's date or in the era it is filed under; a claim about what could happen is judged on whether it does, a recommendation, a minority chance or a story the writer did not offer as a forecast gets no verdict, and nor does a claim whose own horizon is centuries. On the scorecard the word uses the measured inputs' scale, and for a research or tooling gap it means how far the gap slows builders, not that a supply has run short. The model is made by Anthropic, which some of these slots concern.

A model's reading of companies' own pages

The data market is built differently from the rest of the site. Model builders do not publish what they buy, so the section reads the public pages of the firms that sell to them: job listings, research posts, product pages and terms. A model took notes from each page and grouped them into signals; a second run of the same model checked each vendor against each signal and withdrew the pairings the vendor's own page did not support. Every page cited was fetched under this site's usual rules, and nothing was read from social media. What a page says is the company's own account of itself: the site reports that the page says it, and does not vouch that it is true. A count beside a signal is a count of vendors whose pages show it, as a model sorted them, and no person checked each one. Vendors whose evidence of growth is weak are counted apart. No status or tally elsewhere on the site reads these records. Ask may quote a price, or a passage a dossier cites, word for word from the fetched page; it cannot read the signals, the vendor descriptions, the levels or the judgements, because those are a model's reading.

Not measured, and why

• Return on capital against its cost, by layer. Invested capital is not filed by segment, the labs publish neither capital employed nor a cost of capital, and any cost of capital would be our assumption laid on theirs.
• Markups. The marginal cost of a token is not public, and list prices are not transaction prices.
• Abilities no public, dated test reads. Reasoning about cause and effect, memory that lasts from one session to the next, whether a model knows when it is wrong, how far training on generated data can go, whether models do what their makers intend, and sudden jumps in ability. The outlook makes no claim about them until a test does; the nearest readings are named where they exist, such as the predictability score inside the reliability index kept by HAL, Princeton's tracker of AI agents, and a benchmark of learning inside one task, which is not memory across sessions.
• Anything that needs paid data. Lab and app gross margins, GPU rental prices, private valuations between rounds and market-wide round data are paywalled; each affected indicator says so on /indicators, and the paid sources are listed on /sources.

Privacy

The site sets no cookies and keeps nothing in your browser beyond the tab's session; what it and its hosts record is set out in the privacy section of the legal notice.

Disclaimers

Nothing here is advice, figures come without warranty, and parts of the site are written or judged by AI models: the terms are in the legal notice. A status is an editorial judgement made under published rules: the maintainer chooses the bands and direction rules, writes down the rationale, and changes them only through a reviewed pull request.