AI Has Already Won. Now the Question Is Whether We Do Too

Late August, fan running, me cleaning up a couple of issues on my projects while scrolling through LinkedIn, where for weeks now the AI takes have been piling up from “it’s all over” to “it’s all a bubble.” I read, I took notes, I argued with a few people in my head, and I realized I have been postponing a post of my own for months. Not a technical one, the kind with commands to copy and paste. A post where I say what I actually think about AI, without discounts, without that “we’ll see, maybe, who knows” tone we all use now so we do not have to take a position.

My thesis in one line: AI has already won. The real question now is another one: making sure humanity wins too. All of it, not just a tiny slice.

I have been chewing on this for a while, and on a couple of points I have to be honest even with myself, because I am not entirely sure all the way down. But let’s go in order.

Let’s stop telling ourselves fairy tales

There comes a moment, in every big change, when pretending stops being an option. With AI, we are there. The evidence is on the table: models write code, analyze documents, handle conversations, and they do it better every three months. You can find a thousand flaws in them, and they exist, but pure denialism - “it’s just a statistical parrot,” “it’s a bubble,” “in a year people will talk about it like NFTs” - has become pub talk, not a technical position.

The problem with this resistance is not even that it is wrong in itself. It is that it blocks governance. If half of the debate is taken up by people denying the phenomenon, nobody starts thinking about how to steer it. And an unstoppable phenomenon that nobody steers only moves in one direction: chaotically.

Imagine a river in flood. You can deny that it is raining, you can build levees and channels, or you can stand there while the water picks its own path. The first two are choices. The third is what happens when you refuse to choose.

The plateau that never arrives

Every two months, as predictably as utility bills, somebody announces the plateau: “models have stopped improving, the party is over.” Then a new frontier model comes out, benchmarks move upward, and the plateau gets postponed to the next two-month cycle.

In my view, we do not need a brand-new revolution for models to keep improving. What we already have, pushed harder, is enough:

• Classic scaling: bigger models, more data, more compute. The chips get faster, the training farms get bigger. It is not over, it has simply become more expensive.
• Reinforcement learning: the reinforcement stage is producing results that would have sounded like science fiction two years ago.
• Architectures that keep evolving: beyond context windows, a continuous stream of experience, a kind of residue from previous sessions. In practice, memory that does not reset every time you close the chat.

I ran head-first into that last point with Loom: persistent memory is still the piece missing most in everyday LLM use, and when it becomes native inside the models themselves, the jump will be obvious even to people who currently use them only to write emails.

And I will allow myself one small jab: some insiders criticize LLMs in public while quietly jumping on the wagon to make money from them, so they can stay “ideologically pure” while cashing in. I will not name names, but if you spend any time on LinkedIn you know exactly who I mean.

A body? Not needed. Better without one

There is a widespread belief, often attached to Yann LeCun’s name, that to be truly intelligent an AI would need a body and sensory input. No hands and eyes, no real intelligence.

It may be a fashionable idea, but Descartes would have rejected it four centuries ago: the body is a machine, the mind does the real work, and nerves are just tubes. In modern terms: everything we experience happens in the mind - neurotransmitters, brain processing, connections. Eyes are a webcam, arms and legs are keyboard and mouse: you can detach them, and the computer keeps processing. The body is not what makes intelligence possible.

What really matters is something else, and it is what LeCun means when he talks about “world models”: for certain kinds of intelligence - the practical kind, the one that knows a full glass spills when you tilt it - you need to have interacted with an environment and learned its rules. A child does that by touching, falling, moving objects around. But the environment does not have to be physical, and the body does not have to be made of flesh: a simulation, a stream of experience, a rich enough digital world can do the same job. A body is one possible way of learning the rules of the world, not the only one.

And that gets me to the point I care about most: not giving AI a body is a precaution. As long as intelligence lives in a data center, there is still a switch. Somebody can turn it off. The day that same intelligence runs across a network of autonomous robots moving through the world, the switch is gone.

It is the difference between a dog barking behind a gate and one running loose in the yard. The first one stays there until you open it. The second one, if it changes its mind, is your problem.

So not only is a body unnecessary: it is precisely the thing we should avoid giving it. As long as there is still a switch, we should keep it.

The two illusions about work

This is the part that hurts most, especially for people in my profession. When it comes to work, we keep telling ourselves two comforting stories, and both of them need to be dismantled.

Illusion number one: “more software means more work”

The idea goes like this: if AI makes software easier to write, more software will be written, so more people will be needed. It sounds logical. But it has a hole in it: people will build software for themselves. The salesperson who wants a sales dashboard, the office clerk who wants to automate invoices, the shop owner who wants a management tool: today they ask a technician, tomorrow they will ask a model. And the model will build it.

It is not that software will grow and require more developers. It is that software will grow without developers.

There is a difference between spinning up a sales dashboard and maintaining a payment system that has to stay compliant with regulations. But “disposable software” really will be built by the people who use it. The kind that has to last, stay secure, and answer to someone will, for a while, still need somebody accountable for it. For a few years. Not forever.

Illusion number two: “we will all move into conceptual roles”

“Fine, the machine writes the code, we will handle the ideas.” Lovely. Except there are very few “idea people.” Most people do not have that kind of ability and - something people rarely say out loud - many do not even want it. Some people are perfectly happy doing a clearly defined job well, and there is nothing wrong with that.

The problem is geometric: a pyramid gets narrower as you go up. If everybody moves one step higher, there is still not enough room at the top.

The market is already stiffening

And the effects are already visible. Fewer hires, less turnover. Companies are not laying people off in giant waves, but they think twice before hiring: the fear is that the new person may become redundant before they even finish onboarding. And people who already have a job hold onto it more tightly, switch less willingly, accept compromises they would have rejected before. A wise and completely normal response, to be clear. And that trend has not reversed.

It is not just software: one summer inside a company

This summer, in the company where I work, we needed about 50% fewer call center operators than we did last year. Not because we sold less, but because a large part of the work was automated and handled by models.

And more processes will be automated in the future. Supplier price research, competitor price analysis, recalculating selling prices every time purchasing costs change: all the work that today still means human effort, spreadsheets on screen, people comparing one window with another, will be handled almost entirely by AI agents. Not “assisted.” Handled. Somebody will supervise, but they will not really do the work anymore.

None of those were programming jobs. They were normal jobs, done by normal people, and that is exactly the kind of work that supports the widest layer of the pyramid.

The real problem is not income

The debate around “what do we do when AI does everything” almost always ends up on universal basic income, taxes on robots, who pays for what. My feeling is that, sooner or later, that part gets solved: some form of material support will be found.

The real problem is psychological resilience.

Try to picture the transition phase: people at home, without a role, without a reason to get up at the same time every morning, sitting in front of the TV for months before politics manages to find a solution. That is not poverty. It is emptiness. And for a human being, emptiness is often worse than poverty.

That dark phase is almost certain, and very few people are seriously thinking about how to get through it. We talk constantly about regulating models and almost never about what happens inside people’s heads when work disappears.

And the answer cannot be only political: it is also cultural and, if you want, personal. People who have a project of their own, an activity they do because they care about it and not because somebody pays them, will get through that phase better than people who built their entire identity on a badge and a timetable. That is one reason I keep writing open source code nobody asked me for: it is not just passion, it is also a small insurance policy for the mind.

The cultural awakening, and politics lagging behind

In 2022, anyone writing about AI outside technical circles was treated like a crank: nerd stuff, science fiction, “come on, they are just chatbots.” I remember it well, because I was in the middle of that too. Today philosophers and public intellectuals talk about it constantly. The first layer - the cultural one - has been crossed. The second one is still missing: politics.

I do have one criticism of the humanities crowd, though: technical competence is often missing. Some of them genuinely seem convinced that an LLM is “manufactured” by manually inserting ideas into it, as if you were programming a thermostat. That is not how it works. It is a training process, and what emerges from it was not written line by line by anybody. Still, technical competence aside, some of them are making wiser arguments than many AI company CEOs. Which, if you think about it, is both reassuring and deeply worrying.

On the scale of the great revolutions

I am convinced AI is a discontinuity on the scale of the great revolutions in history: the agricultural one, which made us sedentary; the scientific one, which changed how we understand the world; the industrial one, which transformed labor and power. We know all three changed where we lived, how we worked, and who got to rule, and that it took decades - when things went well - before society found a new equilibrium. The difference is that those moved at the pace of generations. This one moves at the pace of quarters. If it is the same order of magnitude but ten times faster, chaos is not a possibility. It is the default plan.

The minimum we can do is stop denying it. The maximum is to guide it. In between there is still room to work - as long as they let us, and after that too.