More is different: my useless PhD just became relevant

For the better part of twenty years, my PhD was the most beautiful useless thing I owned.

It surfaced at dinner parties, occasionally, when someone made the mistake of asking what I had studied. I would watch the eyes glaze over somewhere around the third sentence, and we would both agree, silently and with some relief, to talk about something else instead. A doctorate in the kind of complex-systems modelling that has no obvious application to anything is a conversational cul-de-sac; a line on a CV that makes people assume you are either very clever or faintly unemployable, and often both.

Here, roughly, is what it was about.

Take a unit so stupid it barely deserves the name. It obeys three rules and three rules only: it eats B, it is eaten by C, and it leaves A alone. That is the entire intelligence of the thing. Rock beats scissors, scissors beats paper, paper beats rock; cyclic dominance, in the jargon. No memory. No strategy. No idea that anything exists beyond its immediate neighbours. Then you take a few million of these idiots, scatter them across a grid, and let them eat one another for a few million rounds.

And here is the part that kept me up at night, in the good way.

What comes out is not noise. What comes out is order: gorgeous, rotating spiral waves, structure nested inside structure, patterns that hold their logic at whatever scale you choose to zoom into. Nothing in the recipe mentions spirals. Nobody wrote "form a spiral" into the rules; there is no line of code, no instruction, no plan anywhere that contains the word. The spirals are simply there, the moment the population is large enough, as though several million brainless dots had quietly agreed on a shape none of them was ever told about.

Let me get it out there: what I spent those years watching is the single most unsettling idea I have ever worked with, and it has a name: emergence. Behaviour that is nowhere in the parts, yet unavoidable in the whole.

The word is honest about itself. Emerge comes from the Latin emergere, to rise out of, to come up out of something you were submerged in. That is exactly the sensation of watching it happen: order rising out of the soup, uninvited, fully formed.

And it is not a fringe idea, nor a recent one. In 1972, the physicist Philip Anderson, who would go on to a Nobel, wrote a short and now famous essay with a title that says the whole thing in three words: More Is Different. His point was deceptively simple. A large amount of some simple thing is not merely more of that thing; past a certain scale, it becomes something else entirely, governed by laws you could never have read off the behaviour of a single part. Water is not a wet molecule. A brain is not a clever neuron. The whole knows things the pieces never encoded. That is the territory my useless PhD lived in.

Then the word turned up in my feed

I had made my peace with the irrelevance. I had moved on to organisations, to people, to projects that fail at two hundred miles an hour. And then, over the last couple of years, I started seeing my old word everywhere, in a field I had nothing to do with.

In 2022, a group of researchers led by Jason Wei published a paper titled, almost cheekily, Emergent Abilities of Large Language Models. The claim was that as you scale these models up, certain abilities that are simply absent in the smaller versions appear, more or less suddenly, in the larger ones: multi-step arithmetic, following instructions in languages the small models flailed at, chains of reasoning nobody had explicitly trained. Below some size, nothing. Above it, the ability is just there. And, tellingly, they reached straight back to Anderson for their definition. Same word. Same 1972 essay. Quantitative change tipping into qualitative change. More is different.

I will admit I felt a small, petty thrill. My cul-de-sac had an exit after all.

The honest part, before the exciting part

Now, the voice in my head that spent years in seminar rooms is obliged to say something inconvenient, because the story is not as clean as the headlines made it.

In 2023, a team from Stanford published a sharp reply with an even cheekier title: Are Emergent Abilities of Large Language Models a Mirage? Their argument was that the dramatic, cliff-edge jumps were partly an artefact of how we measure. Score a task all-or-nothing, and gradual underlying progress looks like a sudden miracle; score it more gently, and the cliff flattens into a ramp. The claim landed hard enough that a US congressional committee cheerfully wrote that emergence had been "debunked".

It had not. Two things are worth keeping in view. First, the original authors had already conceded, in print, that harsh metrics can dress up steady gains as sudden leaps; the critique was anticipated, not fatal. Second, and more importantly, the smooth, well-behaved metrics were found after the capabilities showed up, never before. Which leaves the genuinely frightening question entirely untouched: can anyone tell you, in advance, what a system of the next size up will be able to do? The answer remains a confident, well-funded no.

And here is where the physicist in me gets stubborn. Anderson's emergence was never a claim about the sharpness of the jump. A cliff and a ramp are an argument about the shape of a graph. The thing that matters, the thing that has frightened me since I was twenty-six, is the novelty: a capability that was not in the recipe, that you could not have deduced from the parts, arriving simply because the system got big enough. Whether it arrives in a leap or a long climb is a footnote. More is different either way.

Why this is worse than what I already believed

A while back I wrote that planning, for the most part, does not work; that complex systems are chaotic, and that no amount of detail about the starting conditions lets you forecast where such a system will end up. Heisenberg, Poincaré, the stubbornly unforecastable weather. I stand by every word.

Emergence is the more disturbing cousin of that idea, and I had rather hoped to keep it on a lattice where it belonged.

Chaos tells you that you cannot predict the path a complex system will take. Emergence tells you that you cannot predict its repertoire: not merely where it will go, but what kinds of things it will turn out to be able to do at all. Not the trajectory; the inventory. With my spirals, that was a delight, because the worst an unexpected pattern could do was sit there being beautiful and crash my simulation overnight. We are now building systems vastly larger and vastly less legible than anything I ever simulated, wiring them into the middle of everything, and discovering their repertoire the way I discovered my spirals: by running the thing and seeing what shows up.

So, after twenty years, my useless PhD is finally, unmistakably relevant. The dinner-party cul-de-sac opened onto the main road of the decade.

I am, I confess, quite thrilled about this.

I am also, for the first time since I left the lab, a little afraid of my own subject.

Excited. And scared. Mostly scared.