✳ AI Field Notes · 40 cards

You’re earlier than you think.

One argument in forty cards: AI adoption is microscopic, capability is compounding, and the skills gap is the opportunity. Every number below is sourced and linked — no vibes, no invented stats. Save them, share them, argue with them.

How these are made: an image model paints the backdrop (labeled “Art: AI backdrop” on the card) — but every dot, bar, and number is drawn by deterministic code from the sourced figure. AI draws the world; code draws the data. The receipts for all 40 cards are at the bottom of this page.

I · Adoption is microscopic

Each dot is 3.3 million people. — 2,500 dots = 8.28 billion humans. Colour = deepest AI use, 2026.
No. 01 Each dot is 3.3 million people.
1 in 9 humans. Once a week. — 900M weekly ChatGPT users vs. everyone alive, Feb 2026.
No. 02 1 in 9 humans. Once a week.
50 million pay. 4.6 billion could. — Paid AI subscribers (OpenAI, Feb 2026) vs. mobile-internet users on Earth.
No. 03 50 million pay. 4.6 billion could.
71% use it. 1 in 3 scaled it. — Of 100 organizations: those regularly using gen AI vs. those scaled past pilot, 2025.
No. 04 71% use it. 1 in 3 scaled it.
Half quit within a year. Paid users stay. — ChatGPT users still active 12 months in — free vs. paid. (est., Yipit/a16z)
No. 05 Half quit within a year. Paid users stay.

II · Capability is compounding

Intelligence fell off a cliff. — Price per 1M input tokens, frontier-grade → today.
No. 06 Intelligence fell off a cliff.
100 million users, in two months. — Time for each technology to reach 100M users.
No. 07 100 million users, in two months.
2,048 tokens became 2 million. — Frontier context windows, 2020 → 2024. Roughly a thousandfold in four years.
No. 08 2,048 tokens became 2 million.
4.4% → 71.7% in twelve months. — Share of real GitHub engineering tasks AI agents could resolve (SWE-bench).
No. 09 4.4% → 71.7% in twelve months.
No medal. Silver. Gold. — AI at the International Mathematical Olympiad, 2023 → 2025.
No. 10 No medal. Silver. Gold.

III · The skills gap is the opportunity

AI postings up. Everything else down. — Job postings, year over year, from ~1 billion ads analyzed.
No. 11 AI postings up. Everything else down.
The AI premium: +56% on wages. — Wage premium for jobs requiring AI skills. It more than doubled in one year.
No. 12 The AI premium: +56% on wages.
Only 1 in 3 get AI training at work. — Workers whose employer offers AI-skills training, 2025. The rest are on their own.
No. 13 Only 1 in 3 get AI training at work.
What the trials actually found. — Randomized and controlled studies of AI at work:
No. 14 What the trials actually found.
80% of devs use AI. 1 in 10 pays for it. — Of ~47M developers worldwide: AI-tool users vs. paid GitHub Copilot subscribers.
No. 15 80% of devs use AI. 1 in 10 pays for it.

IV · Every platform shift looked this early

1995: under 1% were online. — Share of humanity using the internet, then and now.
No. 16 1995: under 1% were online.
2008: one phone in nine was smart. — Smartphone share of mobile handsets: sales share 2008 → installed share 2025.
No. 17 2008: one phone in nine was smart.
63¢ of every VC dollar goes to AI. — AI's share of US venture capital invested (trailing 12 months).
No. 18 63¢ of every VC dollar goes to AI.
$1.7B → $37B. — Enterprise spending on generative AI — now over 6% of all enterprise software, three years after ChatGPT.
No. 19 $1.7B → $37B.
Faster than the PC. Faster than the internet. — ~40% of US adults used gen AI within two years of launch — the fastest measured adoption of any technology.
No. 20 Faster than the PC. Faster than the internet.

Field extras

400M. 800M. 900M. — ChatGPT weekly active users, Feb 2025 → Feb 2026.
No. 21 400M. 800M. 900M.
23% use it at work. 9% every workday. — Employed US adults using generative AI on the job, Aug 2024.
No. 22 23% use it at work. 9% every workday.
Speed records are cheap. — Threads reached 100M in a record 5 days. ChatGPT took 2 months — then kept climbing to 900M weekly.
No. 23 Speed records are cheap.
4.7 million pay to code with AI. — GitHub Copilot paid subscribers, Jan 2026 — up 75% in a year.
No. 24 4.7 million pay to code with AI.
$19B on apps. $18B beneath. — 2025 enterprise gen-AI spend: application layer vs. infrastructure (incl. $12.5B in model APIs).
No. 25 $19B on apps. $18B beneath.
2 in 100 households pay for AI. — US households with a paid gen-AI subscription, 2026 — up 155% in a year.
No. 26 2 in 100 households pay for AI.
1 in 4 humans has never been online. — 26% of the world — roughly 2.2 billion people — was still offline in 2025.
No. 27 1 in 4 humans has never been online.
The race has barely started. — Consumer AI usage: ChatGPT ≈4× Gemini, ≈8× Claude. (est., Yipit/a16z)
No. 28 The race has barely started.
33% → 65% → 71%. — Organizations regularly using generative AI in at least one function.
No. 29 33% → 65% → 71%.
At expert parity. — MMLU knowledge benchmark: models vs. the measured human range, 2020 → 2024.
No. 30 At expert parity.

V · Proof you can do this

Beginners gain the most. — Support agents with an AI assistant: +14% output overall — but +34% for the newest workers (n=5,179).
No. 31 Beginners gain the most.
41% were given nothing at all. — US workers whose employer provided no AI tools, training, or guidance — none (n=1,020, May 2026).
No. 32 41% were given nothing at all.
Not just faster. Better. — Professionals writing with ChatGPT: output quality +18%, time −40% (preregistered RCT, n=453).
No. 36 Not just faster. Better.
Double the learning. Less time. — Harvard physics RCT: a purpose-built AI tutor produced over 2× the median learning gains of active-lecture class — in ~49 min vs 60.
No. 39 Double the learning. Less time.
America's fastest-growing job: AI engineer. — LinkedIn Jobs on the Rise — #1 in 2025 and again in 2026. Four of the top five US roles are AI roles.
No. 40 America's fastest-growing job: AI engineer.

VI · The size of the bet

42% abandoned most of their AI projects. — Companies scrapping the majority of their AI initiatives before production — up from 17% a year earlier. The average org kills 46% of its PoCs.
No. 33 42% abandoned most of their AI projects.
$110B in one round. — OpenAI's Feb 2026 raise at a $730B valuation — announced alongside 900M weekly users.
No. 34 $110B in one round.
20 million developers walked this path. — Cumulative GitHub Copilot users by mid-2025 — 4.7M now pay for it.
No. 35 20 million developers walked this path.
$500M revenue. Three years old. — Cursor, the AI code editor: 1M+ daily users and ~$500M annualized revenue by mid-2025. (est., TechCrunch-reported)
No. 37 $500M revenue. Three years old.
9.7 trillion became 3.2 quadrillion. — Tokens processed across Google's surfaces, monthly — ~330× in two years. (Google-disclosed; includes reasoning + auto-triggered AI.)
No. 38 9.7 trillion became 3.2 quadrillion.

The receipts

Every card’s figures with their sources. Estimates are labeled “est.” on the card itself. If a number here stops matching its source, tell us and we’ll fix the card.

The best time to learn was 2022.
The second best is now.

The workshop takes you from these charts to actually building — prompts, agents, and a working website.

Start the free AI Workshop

✳ Data snapshot: July 2026 — every figure as published at its linked source on the day we drew it. The numbers will move. That’s the point.