A Chicago Mayor, a Crash Historian and an AI Researcher Walk Into a Room

Trillions are being spent building the most powerful technology in human history. We have put it in nearly everyone's pocket. And 95% of companies are getting nothing measurable back from it. My non-technical 12-year-old can build a working app with it in an afternoon, and a hospital down the road still carries patient records between buildings in a paper folder.

This is not really a story about AI, but about every powerful technology we've ever invented, and the long, awkward gap between what it can do and what we actually usefully do with it.

AI is just the latest example, and with AI the stakes are so much higher: the cost if we don't manage to embed it effectively in our jobs, lives, education systems, societies; the potential gain if we do.

I spent a weekend here in New York with some of the people thinking hardest about the state of AI right now, and what it means for you and me.

At the FT Weekend Festival on Saturday, one could, in a single day, brush shoulders with Rahm Emanuel, Simon Schama, Harlan Coben, Paul Krugman and Meredith Whitney, and with senior leaders and researchers from Google and Anthropic. A room full of people who think deeply for a living. I was there to speak, on the Tech & Futures stage, where Richard Sanderson and I argued that the lasting value in this revolution will sit with the people who learn to use and direct the technology, not simply with the technology itself. Between sessions I did the more useful thing and talked to the people who had come to listen: the investors, the founders, the politicians and the parents in the audience.

The speakers disagreed on plenty, which makes for engaging listening. But each of their very different areas of expertise shone a light on the question I'd walked in thinking about: what all this AI stuff means for our portfolios, for our jobs, and for our children's futures.

What we know to be true:

So if we put those together:

If the technology is cheap, abundant and already in people's hands, and the returns still are not coming, then the technology can't be what is holding us back. MIT arrived where the evidence keeps pointing: the cause is human, not technological; what its authors call the learning gap - the failure to weave these tools into how people actually think and work.

We still need the technology; there is no capability without it. But it is now the cheap and abundant part, and in every previous wave the value has gone not to the abundant thing but to the scarce one beside it. Cheap electricity did not make the great fortunes; the factories and assembly lines built to run on it did.

The economist Carlota Perez has shown this is the rule, not the exception: every technological revolution arrives in two acts, a frenzied installation phase when the money pours into the infrastructure, and a later deployment phase when the value actually shows up, in the hands of whoever learns to use it.

The world is investing trillions in AI infrastructure. By comparison, investment in the capability required to use these systems effectively barely registers.

That gap is what this piece is about, and it runs through all three questions about where the risks and opportunities are for our portfolios, our jobs, and our children's futures.

The Coming Wave

Meredith Whitney, the analyst who saw the 2008 banking crisis coming, and Liaquat Ahamed, who won a Pulitzer for his history of the bankers behind the Great Depression, took the long view: 150 years of financial crises, back to 1873. Their key distinction: ordinary market wobbles come and go. The crises that leave lasting scars are the ones built on debt and mass default, like 2008, because debt destroys capital, and capital does not grow back quickly.

Then they turned to AI, and brought in the railway boom of the 1870s. America poured something like 5% of its national income, year after year, into laying track. The railways were real and they remade the country. They were also wildly overbuilt, and more than 120 American railways had gone bust by 1875 . Today's equivalent - they say - is the trillion dollars a year now pouring into AI data centres.

One of Whitney's main worries is the chips. The expensive heart of a data centre is its processors, and Nvidia has shifted from a roughly two-year product cycle to annual AI-platform releases, compressing the useful life of cutting-edge hardware. Jensen Huang, its own chief executive, joked that once the new ones arrive, you couldn't give the old ones away. So a chip bought today is outclassed within two or three years, yet the companies are spreading its cost across their accounts over five or six. In plain terms, the kit may be obsolete long before it has paid for itself. Build a data centre now, and much of it could be a stranded, depreciating asset before it has earned back what it cost. That is what Whitney means by a faith-based market: enormous sums spent on the faith that the returns will arrive before the hardware is junk.

As I write, this faith is being tested in public. The labs themselves are heading for the stock market: Anthropic has filed confidentially, OpenAI has followed suit. For the first time in years, comparisons with the dot-com era are no longer confined to sceptics. For the oldest reason: that at some point a business has to actually make money.

Ahamed drew from history: in every boom like this, the durable value did not stay with the people who built the infrastructure. It went, later and somewhere else, to whoever worked out what to run on it. The assembly line came after the railway.

The $15.7 trillion that AI is meant to add to the global economy is not materialising yet, for one simple reason: owning extraordinarily powerful technology is not the same as using it well. An F1 car parked in my driveway does not make me a champion driver, just as a great paintbrush does not a Michaelangelo make.

Why our children still need their times tables & functions

This was the most head-spinning conversation of my weekend. If the nature of intelligence is your kind of thing, come with me. If not, skip the greyed section below to the so what for our children.

Blaise Agüera y Arcas, a vice president at Google, gave an interview that got my neurons firing. He sees technology as a real science, not an engineering trick or a creation myth, and his argument runs like this: go back to our cave-dwelling ancestors and intelligence really was individual; each person was a generalist who could more or less do everything the group needed like hunting and gathering, animal rearing, cooking, shelter building, etc. Then came technology, and with it complexity, and with complexity, specialisation. None of us, now, could build a flush toilet or reach the moon alone. Our intelligence stopped being individual a very long time ago. It became collective, distributed across people, tools, and everything we have ever written down. It's social and shared.

Talking with him afterwards, we got onto our own children, and what this means for how they learn. We agreed that individual mastery still matters, precisely because it is the foundation higher thinking is built on. A child still needs her times tables in the age of the calculator, but not because the battery might die. It is that without that foundational fluency settled in her own head, she is locked out of the higher thinking that depends on it; she cannot move easily through ratio, or probability, or compound growth.

We don't teach children to navigate by the stars any more, or to darn socks. The definition of "necessary skills" does change over time; the uncomfortable position we're in now is that the skills people need are changing faster than we can teach them with our old methods, or accept in our nostalgic minds. The need is for people to learn how to self-learn, unlearn and relearn at pace, with special attention to the foundations: numeracy, literacy, critical thinking, curiosity and collaboration. In a world where intelligence is shared with both other humans and machines, why is learning still so focused on individual mastery and not shared knowledge creation and intelligence? The irony is a happy one: AI, the very thing forcing this question, is also the best instrument we have ever had for answering it, provided it's shaped well and effectively - the latter being the difference between educational destruction and educational flourishing.

Why is any child still behind?

If Blaise was the most radical, Rahm Emanuel was the most furious, and I shared every degree of it. He argues that the deep troubles of America, and of the western democratic world, the stalled economies, the politics that curdles, the sense of things coming apart, trace back to one root cause, and that it is an economic one: skills and education. A country that does not build the capability of its people cannot grow, cannot compete, and cannot hold itself together in effective democracy. Fix that, and we begin to fix the rest. Fail at it, and nothing else will save us.

What enraged him, and I must say I share his fury, is that this is not a mystery we are still trying to solve; we know what works, we have the evidence. What we lack is the will and the energy to act. American children are running around two grade years behind, with reading and maths scores at their worst in decades, and, he lamented, we are choosing, year after year, not to act.

We have known for forty years what the most powerful lever in education is. In 1984 Benjamin Bloom showed that a child given one-to-one tutoring outperformed roughly 98 per cent of children taught in an ordinary classroom. He called it the two sigma problem, and the catch was always cost: you cannot give every child a personal tutor. That is the constraint AI has just loosened, and the early evidence is striking.

In 2024 the World Bank ran a randomised controlled trial in Edo State, Nigeria. Around 800 secondary-school students were given a six-week after-school programme that paired a generative-AI tutor with a teacher in the room. The result, in a paper titled From Chalkboards to Chatbots, was a learning gain of about 0.3 of a standard deviation, roughly one and a half to two years of ordinary schooling compressed into six weeks, an effect that beat about 80 per cent of education interventions ever rigorously tested. The largest gains of all went to the girls who had started furthest behind.

At Harvard the same year, the physicist Gregory Kestin and his colleague Kelly Miller built an AI tutor for an undergraduate class and published the result under a title that did not hedge: AI Tutoring Outperforms Active Learning. Their students learned more than twice as much, in less time, than students taught the same material in a well-run, in-person class, and enjoyed it more.

Let's consider what that means. In a world where a well-built AI tutor can do that. In a world where the greatest store of knowledge ever assembled sits in every child's pocket, free, in any language, at any hour. In a world where a patient, tireless, expert tutor costs about the price of a streaming subscription. In that world, why is any child two years behind? Why is any child behind at all? Why aren't we all flourishing and learning more than we ever have, more enlightened, inspired, and capable than we've ever been?

One answer is that we are paying the price now for disconnecting learning from purpose. So many children don't understand why they're learning what they're learning. Learning seems disconnected and irrelevant to the rest of the world they live in.

Educational inequality used to be about access; the new inequality, I'm convinced, is motivation.

And the other answer is that we have not taught anyone, not the children, not the teachers, not the institutions, how to use these advanced technologies well and effectively. And this is where good use and bad use of AI part company.

Same tool, opposite outcomes

A scalpel in a sculptor's hand carves a masterpiece. The same blade, in a killer's hand, takes a life. A scalpel owned but never lifted does nothing at all. Three very different outcomes, one identical object, and the only variable is the human holding it.

AI is no different, and nowhere is it clearer than with our children. ChatGPT can be a cheating machine or a learning machine. Claude can be the thing a student outsources their thinking to, or the thing that sharpens it by arguing back. Gemini can help you interrogate a financial model, or churn out a worse one you don't really understand. ChatGPT can turn out glowing AI slop, or higher level thinking shared with the world when it otherwise wouldn't have been. Same tools, but opposite outcomes. The variable every time is the judgment, purpose and capability of the person holding it.

The studies bear this out, almost too neatly. The Harvard tutor that doubled learning was not raw ChatGPT; Kestin and his team had shaped it with care, telling it to give one hint at a time, never to hand over the whole answer, to stay brief, and feeding it the worked solutions so it would not invent them. Configured that way, it transformed the class. Configured carelessly, the same underlying model does the reverse.

In a trial of nearly a thousand high-school students in Turkey, published this year in the Proceedings of the National Academy of Sciences under the title "Generative AI without guardrails can harm learning", researchers gave one group plain ChatGPT and another a version with teaching safeguards built in. Both looked as though they were racing ahead while the tool was in front of them. Then it was taken away and they sat an exam alone, and the plain-ChatGPT group scored seventeen per cent worse than classmates who had never touched it. They had been leaning on a crutch. The safeguarded version wiped that damage out. Same model; one set-up taught, the other quietly de-skilled. The whole difference was how a human educator had configured it.

This is the point I keep coming back to with the leadership teams and boards I work with, and on the MBA programmes where I teach: stop dwelling just on which model to buy, and start asking who will shape it, and how. What the research and the data tell us is that the model is rarely what decides whether you win or lose with AI; the configuration is where that's decided.

Who gets to decide what good is

On the screen behind Saffron Huang, the researcher from Anthropic, were three words to describe her role: AI for Good. With a gaping hole in the middle of them. Who decides what good is?

This is the part of the weekend that made me sit up, because I have strong feelings about it. When it comes to my children, I decide who comes into their lives. I check the front door before I open it and welcome someone into our house. I vet the adults, as I know you do for your own children. So the idea that a company in San Francisco, however thoughtful its researchers, should decide how an AI speaks to my daughter, when it should be patient and when it should be firm, how encouraging versus challenging, and what "good" means, does not sit easily with me. It has the flavour of the stranger in the supermarket who tells you off for handling your own child the wrong way. With respect: who asked you?

Of course, we all know and accept that there are things an AI should simply never do, a kind of ten commandments for this technology: thou shalt not encourage a person to harm themselves, thou shalt not pose as a therapist, thou shalt not coach someone through building a weapon. We absolutely need to set that baseline and hold it hard. But above that floor, what good looks like, it feels to me, is not the job of Anthropic to decide, for Sam Altman to decide, or for the Google gods to decide.

For this I come back to the car analogy. Ford builds the vehicle and fits the sat nav. Ford does not choose your destination, and Ford does not teach you to drive. That's the job of a driving instructor: someone who understands how people learn, who is patient and clear, and who has nothing to do with the manufacturer. We have been comfortable with that division of labour for a century. In the noise of the AI revolution we have lost it, asking the people who build the engine to also choose our destination and teach the whole country to drive, then wondering why it feels wrong, and why we keep arriving somewhere none of us chose to go.

Just as with driving, we need to become competent at pointing our tools where we want them to go. Driving badly has consequences; driving well can take you places safely you could never have gone on foot.

The process of learning how to point our machines where we want them to go is configuration, and it is the thing almost nobody has been taught yet, and the thing that needs teaching if we are to wield these immensely powerful tools well, and for economic and societal good.

The labs have started to notice and tried to help, although like Ford coming up with pre-set satnavs, they're not the natural qualified folks to do this. In the past year Claude, ChatGPT and Gemini have all added "learning" modes, settings meant to make the tool coach a student through a problem rather than hand over the answer. It is the right instinct; it is the Harvard finding turned into a button. But this setting is still a company deciding what good looks like, at scale, for students as a whole. It is also still mass education with a generalised approach, in a world where we finally have the tools to tailor finely to a person's needs.

Four very different rooms, one conclusion

None of the speakers I've mentioned here stood up to make this argument; it wasn't any of their theses. But I took what each of them said back to the world I work in every day, and this is what it added up to for me.

Ultimately, the technology is no longer the scarce thing. The scarce thing is human capability: the judgment to know which problem to point it at, how to shape it, and when to overrule it.

And we're not building that capability anywhere near fast enough.

The chart below: what the technology can do has shot upward; what we actually use of it crawls along the bottom, and on current form it is starting to slip backwards. Over 90% of AI use today is at the very basic, surface-level. 95% of corporate AI projects fail to return anything measurable. The gap between those two lines is not a technology problem. It is the value we're leaving on the table by not addressing the human challenge of embedding technology into our jobs, lives, education systems and societies, and it widens with every wave.

So, practically, taking the three questions we opened with:

What we gain, and what we throw away

From the small amount of good embedding we've managed so far, that low dotted line on the chart above, the results from effective embedding of powerful technology are already extraordinary:

This is what the technology does in the hands of people who know how to point it. Every development in this paragraph happened because a domain expert grabbed hold of a powerful piece of technology and pointed it, moulded it, and made a masterpiece out of it.

Now if we set that against the cost of not knowing how. As I write, England and Wales have just recorded the highest number of prisoners released by mistake on record, 262 in a single year, more than double the year before. The reason, the Justice Secretary told Parliament, is that every prisoner's sentence is still worked out on paper, across courts and prisons whose systems do not talk to one another.

In the NHS, doctors still print out patient records and carry them between buildings in folders; the last government set aside £3.4 billion simply to drag the health service's productivity into the digital age.

Disaster relief is routinely uncoordinated.

Resources are mis-allocated.

In much of the world, a child's chance of ever having a good teacher is still a lottery.

None of this is because the tools do not exist. They do. We simply haven't built the human capability to use them.

The numbers are so large, the effects so monumental, and the technology so powerful, that we don't need a revolution to move them. We need a few percentage points. One, even.

Lift corporate AI from the 5% of projects that currently work towards even a third of them, or nudge everyday use a few tenths of a per cent above the surface level it mostly sits at, and we're talking about hundreds of billions of dollars in value, and a real dent in problems we have been told for decades are simply the way things are. The gap in the chart is the prize, and to me that is worth spending time on.

Closing the gap, and the one skill that lasts

When Louise and I started harking on about the AI capability gap just about two years ago, we were fairly lone voices. Now every man and his dog is opening an AI training business, having spotted the same gap. It's good: the gap is real and enormous (and we feel mighty validated). But most of what's rushing in to fill it is the old model in new clothes: a course, a certificate, a one-day workshop, delivered once and ticked off. A curriculum fixed today is out of date by the time it is delivered, in a field that moves this fast.

What's happening is that we're using our old, archaic methods to address a problem of supersonic pace and scale. It is a bit like a business running on paper letters instead of email in 2026, and wondering why it can't seem to keep up with things.

So what is actually needed to close this large, menacing and wasteful gap?

It's the gift AI has given us: capability-building that is highly tailored to the person and moves at the same speed as the technology. It also needs to be independent. The driving instructor does not work for Ford, and does not only teach you to drive Fords. The whole point of them is that they are on your side, across whatever you end up driving, from a VW Polo to the family van to a Maserati, and that thing will be different at different stages of your life.

That is the gap we built Atheni.ai to fill. We are not selling people an AI model, so we have no reason to push one; we start with the people, not a platform; and the AI is configured to teach them how to embed their whole stack in their particular workflow, moving between Copilot and Claude and ChatGPT and Gemini in a single afternoon, guiding them to the use cases, tools, techniques and learning they need to become properly good at this in days and weeks, not years. The reason Atheni can do that is because it's using AI to gather those developments that are happening at supersonic pace, using AI to translate those changes to this person's needs, and using AI to create custom content dynamically, continuously, day in and day out.

It is also why we can say something most of this market cannot. Across every programme we've run, in further education, financial services, leisure, manufacturing, we have produced that coveted measurable return, and over 90% of participants reach daily, guided AI use, sophisticated, strategic and compliant, inside ninety days. That shows up in the work itself, not in a certificate filed away.

The reason we built Atheni.ai is because we looked at that chart, above, and thought: wouldn't it be great to see what amazing things might happen when that dotted line moves even just a few points up?

The hard part now is us.

If we take only one thing from a weekend of very clever people talking to FT journalists, each other and the public, let's take this: building this technology was extraordinarily hard, and the people who did it deserve their due. But the hard part now is not more technology. It is us: whether we build the judgment to know what to keep, what to hand over, and whose hand is on the dial. That's a human challenge, and one it feels we simply must rise to.

There is a smaller thing I hope my own children take from watching this work, and I think it's important for anyone reading. When you see a problem, you are allowed to try to fix it. You do not need permission, and you must not be stopped by the fear of getting it wrong in front of people who know more than you. Knowing is not the whole job. The other half is translation: taking the big ideas the brilliant people circle and turning them into the plain, practical "so what" the rest of us can use on a Monday morning, in a real classroom or a real office. That translation is also the work.