The guest house in Tannsjö is ten square meters. I built it myself, all wood, winter-insulated, a man cave really — the kind of place where the rest of the world stops mattering the moment you close the door. That night a tempest was moving through. I could hear it arriving in waves, the howling rising and falling, the small windows shaking in their frames. Inside, I felt protected. Held, even.

I was reading Freeman Dyson when I found the story about Fermi. Dyson had rushed in to show Fermi a graph. His team’s new theory matched Fermi’s experimental data beautifully, almost perfectly. Fermi glanced at it and dismissed it.
“How many adjustable parameters did you use?”
“Four,” Dyson said.
Fermi quoted von Neumann: With four parameters I can fit an elephant. With five I can make him wiggle his trunk.
Years later, Dyson called this the best bad news he ever received. It saved him from wasting years on a dead theory.

I put the iPad down and looked up at the wooden ceiling. The storm pressed against the windows. I picked the iPad back up and read the passage again.

I thought about the greatest inventions in human history. Fire and Language are the Big Two. They set everything else in motion. After them come Agriculture, the Scientific Revolution, the Printing Press, the Industrial Revolution. Each one didn’t just change what people did. It changed how people did think and what they could become.

In my lifetime, two inventions belong on that list. The Internet and AI.
I built my career on the Internet revolution. My son is building his on the AI revolution.

I thought about Fermi’s elephant again.

Modern AI systems have millions of parameters, sometimes billions. By Fermi’s logic, they should be able to fit anything — elephant, trunk, and all. They should be useless. They should memorize without understanding.

And yet they translate languages. They recognize faces. They make predictions that hold up. Something real was happening inside them, not just mimicry.

I thought about how I study. When I want to truly understand something, I don’t go back to the same examples. I work new problems. If I can solve those too, then I have built something — a mental model. And almost always, solving new problems changes the model. Cracks appear. I revise. I test again. Over many repetitions, the model gets closer to the truth.

That is the same test they use for AI. Train it on part of the data. Test it on data it has never seen. If it only memorized, it fails the moment the questions change.

Still in bed, the storm still moving through in waves, I thought: with enough data, cheating becomes impossible. A handful of data points and you can fit anything, fake anything. But when the sea of examples is large enough, memorizing stops working. The machine is forced to find the actual signal — the structure underneath — or it fails. The same way I was forced to actually understand mathematics, or I couldn’t solve the problems I hadn’t seen before.

I saw Fermi’s elephant standing there on my bedside table. An icon of the old science.
Before computers, before AI, science could only study nature in isolation. Reality had to be simplified down to idealized systems — a frictionless surface, a perfect sphere, a closed system.

Not because scientists were lazy, but because those were the only elephants small enough to fit inside four parameters. Science drew its boundaries around what it could manage.

Those boundaries are gone now. What comes after them, we are only beginning to find out. Not just in science. In everything.

I lay there listening to the storm. My son is building his career in this. I thought about how AI would change our thinking. And what creatures we would become. I turned off the light.