In second year of physics, you spend a semester staring at Schrödinger's equation applied to crystal lattices. It's not the most glamorous part of the curriculum. You're computing band structures for silicon, drawing Brillouin zones, working out why some materials conduct and others don't. At the time it felt like an elaborate digression. In retrospect, it was the most important thing I learned.
Here is what I actually understood, much later: the reason silicon can be made to compute is a direct consequence of quantum mechanics applied to a periodic arrangement of atoms. The electrons in a crystal don't behave like classical particles. They form delocalized wavefunctions that span the entire lattice, and the allowed energy levels cluster into bands with gaps between them. The gap in silicon — the band gap — is what makes a transistor possible.
From quantum mechanics to computation
A transistor is a quantum mechanical device. When you dope silicon with phosphorus or boron, you're introducing electron donors or acceptors that shift the Fermi level — the energy at which electrons are likely to be found — and this shift is what creates the p-n junction. The p-n junction is what makes a diode. Two junctions back to back make a transistor. A transistor is a switch. A switch is a bit. Bits are logic. Logic is computation.
And computation, apparently, is intelligence. Or at least the substrate of it.
The chain is: Schrödinger's equation → band theory → doped semiconductors → transistors → logic gates → von Neumann architecture → neural networks → large language models that pass bar exams and write poetry of dubious quality. Every link in this chain is real physics. Not metaphor. Actual physics. We did not design intelligence and then implement it in hardware. We discovered that the quantum mechanical properties of certain materials, combined with enough of them arranged correctly and fast enough, would produce something that looks a great deal like intelligence from the outside.
We are talking to sand
This is the thing that strikes me every time I interact with a capable language model: we are talking to sand. Highly purified, intricately arranged, precisely doped sand — but sand. The atoms are the same atoms that make up rocks. The electrons are the same electrons. The difference is purely organizational.
There's a standard rebuttal here: "but the software matters, not just the hardware." Yes. And the software is also physics. The training process is gradient descent — a numerical optimization procedure that adjusts billions of floating-point numbers (which are themselves representations of voltages, which are themselves the behavior of electrons at junctions, which is itself quantum mechanics) until the loss function is minimized. The loss function is a mathematical object. The mathematics is instantiated in physical processes. There is no gap between the math and the matter.
I find this deeply consoling and deeply strange in equal measure. Consoling because it means that intelligence isn't magic — it's a property that emerges from physical processes organized in particular ways, and those processes are in principle understandable. Strange because we are physical processes organized in particular ways, and we spent most of our history believing we were something else.
The universe was always going to do this
The universe was always going to do this. Given enough time and the right initial conditions, matter was always going to arrange itself into configurations that model their own environment, reason about their own existence, and eventually start arguing about whether they're conscious. We happened to be the arrangement that got there first — and then built a second arrangement, out of rocks, that is starting to have some of the same arguments.