Slow down AI development.
That is the whole ask. We want time — for the evaluation, oversight, and safety work meant to keep frontier AI in check to catch up with a capability curve that has compounded at roughly 4–5× a year for over a decade. The case for this is not a mood. It is a measurement, and every line of it below can be checked.
There is a version of the argument for slowing down AI that runs on vibes — on unease, on science fiction, on the sense that this is all moving too fast. That version is easy to dismiss, and it gets dismissed daily.
This campaign makes the other version. It contains no predictions we cannot source, no statistics we invented, and no claim that rests on trusting us. This page carries the strongest two charts, the statement, and the ledger; the full brief — the benchmarks, the forecasts, the incident record, the expert surveys, and every source — is on the Learn page.
Two things on this site are interface rather than evidence, and are marked as such wherever they appear: the signature ledger below and the video board on the gallery page. This is a prototype build with nothing behind it, so the ledger opens at zero and the video cards are samples. We would rather show you an empty counter than a plausible one.
A billion-fold increase in thirteen years
The most reliable single indicator of AI capability is the amount of computation poured into training a system. Since the deep learning era began around 2012, the compute used in the largest training runs has grown by roughly 4–5× per year, every year, through three hardware generations, one pandemic, and one chip shortage. The safety and oversight work meant to keep pace has not. The full comparison is in the brief.
Training compute for frontier AI systems has grown about 4–5× every year since 2012
Estimated floating-point operations used in the final training run. Note the vertical axis: each gridline is one hundred times the one below it.
Approximate increase in the computation used for the largest AI training run between AlexNet in 2012 and the frontier systems of 2025. If a 2012-era training run were a one-second event, the 2025 equivalent would run for nineteen years.
Derived: the ratio of Epoch AI’s estimates for AlexNet (≈4.7×1017 FLOP) and the largest 2025 run (≈3×1026 FLOP, est.). Order of magnitude, not a precise figure.“We want time.” Two fair questions about that sentence.
Who is “we”?
The public — not just the experts. 72% of US voters would rather AI development be slowed down than sped up; 8% want it faster. The researchers building these systems are worried too, but the majority asking for time is ordinary people, polled repeatedly, across pollsters and years. The chart below is that majority.
How much time?
Until safety demonstrably keeps pace. In the statement's own words: development “should proceed no faster than our ability to understand, evaluate, and control” these systems. That is not a calendar date and not forever — it is a condition, measured by capability against oversight, and it can be met.
Slowing down is not a fringe position. It is the majority one.
Public opinion on AI is often reported as confused or volatile. On the specific question of pace, it is neither. Across pollsters, framings and years, large majorities of the American public say they would rather AI development went slower, and they do not trust the companies building it to regulate themselves.
This is worth stating plainly, because the perception runs the other way: people who want a slowdown routinely believe they are in a small minority arguing against the public mood. They are not. They are the public mood, without a mechanism to express it.
On pace, on risk, and on self-regulation, the majority is not close
Share of respondents by answer. US adults or registered voters, 2023.
The gap this campaign exists to close. A large majority wants a slower, more carefully governed rollout. A near-unanimous expert community agrees the risks are serious enough to name alongside pandemics and nuclear war. And there is still no binding requirement in US law that a frontier system be evaluated before it is released. Majorities do not become policy by existing. They become policy by being counted.
We want time. Add your name to the count.
This is a public record of people who have read the evidence and concluded that frontier AI should not outpace our ability to understand and control it. It exists so that the majority in the polling above has a name, a number, and a document that can be handed to a legislator.
Development of AI systems more capable than today's frontier models should proceed no faster than our ability to understand, evaluate, and control them.
We call on governments to establish binding oversight of frontier AI development — registration of the largest training runs, independent evaluation before deployment, and the authority to require a pause — and on developers to accept it.
prototype — not a live count
The ledger has not opened yet. On a live build this number is read straight from the signature table and nowhere else — which is the only way a page like this one is allowed to have a number on it.
Most recent signatories prototype — no stored signatures
Empty. On a live build this shows the occupation and country of the most recent names, as each person chose to record them — never an email, never a full name. We have left it blank rather than fill it with people who do not exist.