We gather what the future learns from.

Cicada Intelligence gathers the data the future of AI is trained on, of every kind, with consent, from people and places everywhere.

Scroll

Every model ever made learned from us.

Intelligence is not born. It is taught.

But the written record is running out.

Epoch AI estimates that the stock of public human text could be fully used between 2026 and 2032.

The next leap needs data that does not exist yet.

So we gather it.

The film

Enough

The film is being cut.

A minute on what people know, and what it could become. Music: "Wildflowers" by Scott Buckley, CC BY 4.0. Footage: Pexels.

What becomes possible

What the future is waiting to learn.

Every leap in AI so far stood on data someone gathered first. These are the leaps still waiting, and the data each one needs.

Mathematics01 of 13

Mathematics machines can check and extend.

Most mathematics lives as prose. When the reasoning behind each proof is recorded, and not only the proof, machines that verify and extend it come within reach.

Already trueIn 2024, AlphaProof reached silver medal standard at the International Mathematical Olympiad. It first had to translate about 80 million statements into formal language, because too few existed. Nature, 2025.

Quantum computing02 of 13

Quantum machines that correct themselves.

Every quantum processor fails in its own way, and simulated noise is only an approximation. When real machines record how they actually fail, error correction fast enough to make them useful comes closer.

Already trueIn 2024, AlphaQubit, trained on simulations and then on data from a real processor, made fewer errors than the leading decoders on Google's Sycamore experiments. Nature, 2024.

Encryption and trust03 of 13

Software that proves it is secure.

Codes are rarely broken as mathematics. They are broken as code. When the ways real systems fail are written down and made checkable, software that proves its security, instead of promising it, becomes possible.

Already trueIn January 2026, AISLE reported that its AI system found all 12 flaws fixed in that month's OpenSSL release. Some had been there since 1998.

Medicine and biology04 of 13

Medicine for every body.

Medicine learns from measured lives, and most lives have never been measured. When the people in the data reflect the people in the world, care that works for everyone comes within reach.

The gapMore than eight in ten people in genome wide association studies are of European ancestry. GWAS Diversity Monitor, 2021.

Materials and energy05 of 13

New matter, designed on purpose.

Hundreds of thousands of materials have been predicted, and only a few hundred made. When failed attempts are recorded beside the successes, materials designed for batteries, clean energy and medicine, rather than found by luck, come within reach.

Already trueIn 2023, GNoME predicted about 380,000 stable new materials. An autonomous lab made 41 of the 58 it tried, in 17 days. Nature, 2023.

Climate and the Earth06 of 13

Warnings that reach everyone.

The sky is measured far better than the ground, and the ground least where the risk is highest. When the ground is measured as well as the sky, warnings for the places with the fewest instruments come within reach.

Already trueIn 2024, a flood model trained on 5,680 river gauges began giving five day warnings in basins with no gauges at all. It now serves more than 80 countries. Nature, 2024.

Robotics and physical AI07 of 13

Machines that can lend a hand.

Text can be gathered from the web. Skill cannot: every grip, correction and recovery has to be shown. When skilled work is recorded from the hands outward, machines that help with physical work come within reach.

The gapThe largest open collection of robot data is estimated to be about 200,000 times smaller than the text language models learn from. Survey of vision language action models, 2025.

Every language08 of 13

Every language, as it is spoken.

About seven thousand languages are alive today, and most are missing from AI. When a language is recorded the way people actually speak it, technology that serves its speakers, instead of skipping them, becomes possible.

Already trueIn 2023, Meta's speech models reached 1,107 languages, largely by learning from readings of religious texts, because for most of those languages little other recorded speech existed.

Food09 of 13

Advice before a harvest is lost.

Small farms grow much of the world's food, and are almost absent from its data. When fields are recorded where food actually grows, advice that arrives in time comes within reach.

Already trueIn 2020, a phone app trained on images of cassava diagnosed crop disease in the field about as well as experts, with no connection. Frontiers in Plant Science, 2020.

Learning10 of 13

A patient teacher for every learner.

AI tutors learn from answers, the output of teaching, and not from teaching itself. When good teaching is recorded as it happens, patient one to one instruction for every learner comes closer.

Already trueIn 2024, a randomised trial with 900 tutors and 1,800 students found that live AI suggestions raised topic mastery by 4 percentage points, and by 9 with the lowest rated tutors. Tutor CoPilot, 2024.

Care and accessibility11 of 13

Technology that hears everyone.

Speech recognition still fails many of the people who need it most. When the voices current systems miss are recorded with care, technology that answers to everyone who speaks comes within reach.

Already trueIn 2024, 151 hours of speech recorded with people living with Parkinson's cut recognition errors from 36 to 24 percent. Speech Accessibility Project, 2024.

Exploration12 of 13

Discovery at the speed of the sky.

Telescopes now see more than anyone can look at. When expert judgment about rare events is recorded beside the observations, discovery as fast as the sky changes becomes possible.

Already trueIn February 2026, the Rubin Observatory sent 800,000 alerts on its first night. It is built for about 7 million a night.

Protecting the human future13 of 13

A future we can trust.

Safety data is made, not found. When human judgment about hard cases is recorded widely, across cultures, AI that can be measured, corrected and trusted comes within reach.

Already trueIn 2022, a model trained on human judgments was preferred to one more than 100 times its size. InstructGPT, 2022.

Every result named here belongs to the people who achieved it. What comes next is still to be gathered.

What we gather

Every kind of data the future needs.

What people do, say, write, see, know and judge. Gathered with consent, wherever it lives.

How we hold it

How we hold it

  1. Nothing is gathered without consent.
  2. Privacy is part of the design, and personal details are removed wherever they are not needed.
  3. We follow the data protection law of every place we gather data.

The mark

Thirteen or seventeen years in the dark.

Periodical cicadas spend thirteen or seventeen years underground, unseen. Then, within a few weeks, they emerge together by the billions.

Our mark is the centre of the prime spiral. Write the counting numbers in a square spiral and fill in the primes: its two top corners are 17 and 13. Gather enough numbers, and lines appear that nobody drew.

Then, all at once.

Build what comes next with us.