More Is Different

A single neuron is a switch. A single artificial neuron is a multiply and an add. Neither one contains anything you could point at and call intelligence, and yet wire enough of them together and you get systems that reason, plan, and argue with you about your own code. The intelligence isn’t in the parts. It lives in the interactions, and I’ve come to think emergence is the most underrated idea in the entire argument about where AI is going.

The physicist Philip Anderson gave this its slogan in 1972, in a four-page paper in Science called “More Is Different.” His target was reductionism, or rather one particular overreach of it. Yes, everything reduces to particles obeying a small set of laws. No, that does not mean you can start from those laws and rebuild chemistry, then biology, then a mind. “The ability to reduce everything to simple fundamental laws,” he wrote, “does not imply the ability to start from those laws and reconstruct the universe.” At each level of scale, new behavior appears that you could not have read off the level below. Psychology is not applied biology. Biology is not applied chemistry. The whole is doing something the parts were not.

That word, emergence, gets thrown around loosely, so let me be concrete about what it actually looks like, because the examples are the argument. An ant knows almost nothing. It follows a few chemical rules about pheromone trails. No ant holds a map. And yet the colony routes around obstacles, balances its foragers against its nurses, and finds the shortest path to food, solving problems no individual ant could state, let alone solve. A starling flock wheels and folds across the sky with no leader and no choreography, each bird tracking maybe seven neighbors, and out of that local rule comes a global shape that looks designed. Your brain runs on cells that individually do something close to nothing, and somewhere in the interactions of eighty-six billion of them is the thing reading this sentence.

The pattern is always the same. Simple units, local rules, and behavior at scale that is not present, not even hinted at, in any single unit. You cannot find the flock in the bird. You cannot find the colony in the ant. And I don’t think you can find intelligence in a neuron, biological or artificial.

Which is exactly what makes large language models worth staring at. They produce capabilities nobody wrote down. Nobody sat at a keyboard and implemented three-digit arithmetic, or in-context learning, or the ability to hold a chain of reasoning across a dozen steps. Those behaviors showed up when the networks got big enough, and they were not there in the smaller versions. The engineers specified an architecture, a training objective, and a very large pile of data. What came out was more than what went in. That is emergence, the same shape as the flock and the colony, running now in a data center.

Here is the objection I have to take seriously, because it’s the one that keeps “emergence” from being a real explanation. Sometimes the word is a way of not explaining anything. “It emerges” can be a scientific-sounding shrug, the modern version of saying a phenomenon is due to its nature. If I can’t say anything about why intelligence arises from these networks and not from a hurricane or a stock market, both of which are also enormous systems of interacting parts, then I haven’t explained the intelligence. I’ve just given the mystery a longer name. And the honest state of things is that we cannot say, precisely, what separates the networks that think from the networks that merely churn. Almost no networks produce anything like intelligence. These do. The line between them is exactly the thing nobody can draw yet.

So I want to hold two claims at once, and resist collapsing them into a slogan. The first: intelligence is emergent, a property of networks at scale, not a substance stored in their components, and this is as true of the brain as it is of the model. The second: naming something as emergent is the beginning of the work, not the end of it. Anderson’s point was never that complexity is magic. It was that complexity has its own laws, laws you have to discover at each level, because they don’t hand themselves to you from the level below.

That’s the part I can’t stop turning over. If intelligence is what happens when you wire up enough simple parts and let them interact at scale, and it appears without anyone designing it, then the interesting question stops being how do we build it. It becomes what were the conditions that made it inevitable, and whether we’re anywhere close to understanding them or just close enough to keep tripping over them by accident. That’s the argument I want to make next, and it’s a stronger claim than this one, so it deserves its own post.

There’s a separate fight about whether these emergent abilities are even real or just an artifact of how we measure them, sharp jumps on a chart that turn into smooth slopes the moment you change the metric. That fight gets its own post, and it’s a genuinely different question from this one. Whether a capability’s arrival looks abrupt or gradual on a graph is about measurement. Whether intelligence is the kind of thing that arises from networks given scale and feedback is about what intelligence is. This series is about the second question. It’s the one I find I can’t put down.