Vertigo

There is a strange arithmetic to this moment. Most of humanity has barely touched artificial intelligence. Most of those who have met it know only a polite and somewhat clumsy chatbot. And a very small number of people are watching the curve go vertical beneath their feet. The gap between what these systems can do and what most people believe they can do is no longer a misunderstanding between two camps. It has become a funnel, wide at the mouth, astonishingly narrow at the neck, and narrowing still.

I live inside that strangeness every day. We are all in the same house tonight. We are simply standing on different floors, and from some of those floors the view has begun to make people dizzy.

The funnel

Put rough numbers on it and the shape becomes unmistakable. Andrej Karpathy recently sketched them on X, and they match what I see around me. Roughly six billion people, three quarters of humanity, have barely touched these systems at all. One to two billion use a free chatbot now and then, as a better search engine or a patient writing aid, and come away unimpressed. Perhaps twenty million use the most capable models professionally in code and mathematics, and have watched work that once took weeks or months be completed by an agent on request. And at the very neck of the funnel sit a few thousand people with access to the frontier systems inside the laboratories, where swarms of agents labour for weeks on vast projects and human review has begun to fall behind.

Three forces give it this shape. The first is ambition. A question that can be answered in a paragraph barely stirs these systems, and their power only becomes visible to someone who brings large, difficult problems they genuinely care about. The second is that capability is uneven. It rises highest where work is digital, verifiable and economically valuable, because that is precisely what the models are trained to master, and so code and mathematics have run far ahead of everything else. The third is plain access, the difference between the free tier, the paid tier and the private rooms of the laboratories.

All of this is happening at once. The public meets a clumsy chatbot, most professionals notice a modest lift, and a thin sliver of people feel the vertigo of a curve going vertical.

Fear grows in the dark

Numbers like these miss what I find most important, which is the temperature on each floor.

At the wide mouth of the funnel, among those who use these systems rarely or never, I see fear growing faster than understanding. It is not hysteria, and it is certainly not stupidity. It is the oldest human response to an unknown quantity. A thing one has never touched, of which one hears only that it will take one’s job, write one’s children’s homework and perhaps one day outwit us all, is bound to look larger in the dark than it does in the light. Fear, in this case, is simply what distance from a technology feels like from the inside.

That changes how I read the funnel. It is not only a map of capability. It is a map of anxiety, and the two run in opposite directions.

The view from the narrow end

An essay about vantage points owes the reader its own, so let me say where I stand. I work with these systems constantly. How I use them depends on the case in front of me, but the honest description is that I use them a great deal, for nearly everything that can be expressed in text, code or numbers. By any reasonable definition I am a power user.

What surprises me is not the gap between my habits and those of an older generation. It is the gap between me and people of my own age, many of them technically fluent, who still use the most remarkable tools of the century as a slightly cleverer search box. They are not wrong to be cautious. They are simply standing one floor below a view they would very much enjoy.

I find that observation more exhilarating than worrying. The potential has not yet been harvested. For the most part it has not even been noticed.

A continent on the lower floors

If the funnel sorts people, it sorts continents too, and here my agreement turns uncomfortable. Europe stands structurally too low. The frontier models come from the United States and from China. What Europe offers by comparison is, with a few honourable exceptions, closer to a toy than to a rival, and the infrastructure on which serious models are trained and served lags further behind still.

For a continent in this position, access is anything but abstract. In June, Anthropic suspended access to two of its newest models for almost three weeks in order to comply with American export controls, and restored it only after Washington lifted them. Whatever one thinks of the policy, the lesson is unsentimental. The upper floors of the funnel belong to whoever owns the building, and the terms of the lease can be rewritten from abroad.

This is why I build what I build. With Cubitics we are putting compute infrastructure on European ground, and with Starlex we are working on European foundation models. I will not pretend that either will close the gap by next spring. Infrastructure is poured in concrete and counted in megawatts, and none of it arrives overnight. But the honest conclusion from a structural deficit is not resignation. It is acceleration. We have to move faster precisely because we are starting lower.

Starting behind is not an argument for patience. It is the only argument for speed that has ever been needed.

The tide reaches every shore

Uneven capability is sometimes read as an accusation, as though these systems were snobbish about which professions they deign to help. I see nothing of the kind. What we are watching is simply the order in which the fruit is picked. Code and mathematics hang lowest, because their results can be checked automatically and their value is easy to count, and so they were harvested first.

Every other field will follow as the methods mature, including the ones that look furthest away today. The trades seem to stand entirely outside this story, the electrician, the carpenter, the roofer with weather on his face. I would not bet on that. The quotes and the paperwork, the planning of a job, the diagnosis of a fault from a photograph, the ordering of parts, all of it is already within reach, and the remainder is a question of time rather than of principle.

The jagged edge is not a wall. It is a tide line, and the tide is coming in.

The locksmith’s nightmare

There is one floor of this house I know better than the others, because I spent years on it professionally, breaking into systems for a living long before I began building them. Of everything now happening at the narrow end of the funnel, nothing concerns me more than agents that find vulnerabilities and run attacks and defences at machine speed.

For as long as I have worked in security, the cheapest way into an organisation ran through its people. Phishing and social engineering were the attacker’s lever of choice, not because they were elegant but because technical vulnerabilities were hard to find. Uncovering a genuine flaw in well-maintained software demanded rare skill and long patience, and that scarcity quietly formed part of every company’s defence.

That scarcity is ending. Technical vulnerabilities are becoming as discoverable as human ones. Open-source intelligence, the patient assembly of everything an organisation reveals about itself, can now be gathered and turned into an attack in a fraction of the time it once required. And the old lever has not been retired. It has been sharpened. Phishing messages are better written, better researched and faster to produce than at any point in history. Both doors now stand open at once, and both swing faster than before.

The defensive habits of the last twenty years were built for an adversary who was slow, scarce and human. That adversary is disappearing. Cybersecurity does not need another incremental product. It needs a radically new approach, one that takes for granted that the attacker moves at machine speed and asks, from first principles, what defence must look like under that assumption.

A word to those who sign the cheques

Executives deserve a particular word, and I can offer it with some standing, because I am one of them and owe my peers the courtesy of candour. Many of us spend our days talking to people rather than working with these systems, and so we understand what is happening intellectually, but only at second hand.

Reading about this technology is not the same as working with it, any more than reading about the sea is the same as sailing. Leaders who do not take the trouble to work their way into it themselves, at the keyboard, on a real problem of real size, will be left behind, and their companies will be left behind with them. Not in some distant decade. Within the next few planning cycles.

Coda

The clumsy chatbot, the modest lift and the vertical curve all exist side by side, and that is the peculiarity of this moment. Yet every floor of the funnel suffers from the same affliction, and its name is distance. Distance makes the public afraid, executives complacent and a continent dependent.

Distance, unlike most of the problems of our age, has a known cure. One closes it by stepping forward.

So climb a floor. Take one large problem you genuinely care about and hand it to the best system you can reach. Then stand at the window for a moment. The vertigo passes. The view does not.

---

Sources

1. Andrej Karpathy, post on X, 11 October 2026
2. Anthropic, statement on access to Fable and Mythos