Excerpt from Chapter Four

The Evolution of a Large Language Model

Suppose that it’s a couple of million years ago, back when our ape ancestors were pretty smart but not as smart as us. Suppose that one of these apes is the lucky recipient of a genetic mutation that strengthens the connections among certain neurons in a way that improves his ability to use language.

For example:

Suppose that when a female ape says to him, “Do you want to have sex?” he says, “Sure! My place or yours?” And suppose that a male ape that lacked the new gene for selectively strengthened neuronal connections would, in the same situation, just stare at the female blankly for a few seconds and then go take a nap. Clearly, the new gene that enhanced linguistic ability would have a better chance of getting into future generations than the alternative gene that resided in the nap-taking ape.

I concede that, in all of human evolutionary history, no scenario like this ever played out in exactly this way. Still, this scenario illustrates two things that did play out and that help explain why humans are blessed with a biological infrastructure for using language. First, the evolution of that infrastructure involved genes that selectively strengthened neuronal connections. Second, the genes that strengthened these connections were preserved by natural selection because the use of language helped our ancestors get their genes into subsequent generations.

For example, warning offspring about various kinds of dangers could keep them alive long enough to reproduce. Strategizing with fellow hunter-gatherers about how to track and kill big game could also pay off genetically. So could wooing a mate, though the wooing may often have been subtler than in the example above.

All of which brings us to the neural networks that house large language models. Though a large language model and a human brain are different in many ways, they have at least three things in common: (1) Both involve lots of interconnected nodes that are called neurons; (2) both have been shaped by an evolutionary process (even if, in the case of the LLM, the evolution is often referred to as “training”); and (3) in both cases the evolution worked, in part, like this: Mutations that change the strength of neuronal connections in ways that enhance the ability to use language are preserved, and these changes can add up to powerful new linguistic capabilities.

Before explaining how the evolution works in the case of large language models, I should reiterate that in many ways a neural network is not like a human brain. For starters, the “neurons” aren’t physically distinct things you can reach out and touch, and the “connections” between them aren’t either. It’s true that neural networks can involve thousands of microchips that are interconnected, but it’s not true that each of those microchips represents a single neuron. A neural network is in some ways more like a simulation of a physical artificial brain than like an actual physical artificial brain.

Still, this simulated artificial brain works exactly the way the physical artificial brain would work if somebody took the trouble to build it. So if you want to understand how a neural network works, you can start by imagining this physical artificial brain. Here’s how that would look:

Excerpted from The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning by Robert Wright. Copyright © 2026 by Robert Wright. Published by Simon & Schuster.