The Inevitability Thesis

The comfortable story about AI is that a handful of brilliant people are inventing intelligence, architecture by architecture, breakthrough by breakthrough. It flatters everyone involved and it fits how we like to think progress works. I’ve come to believe the uncomfortable version is closer to the truth, and it’s worth stating plainly, because a hedged version of this claim isn’t worth arguing about.

Here it is. Given sufficient scale and feedback, the emergence of intelligence may be close to inevitable. Architectures aren’t so much invented as discovered. We are not building minds. We’re assembling the conditions under which minds become the thing that reliably falls out, the way crystals fall out of a cooling solution whether or not anyone is watching.

Call it the inevitability thesis. It’s a bet, not a proof, but the evidence points the same direction from four separate places, and it’s the convergence that makes me take it seriously rather than any single leg.

The first is mathematical. There’s a result called the Strong Lottery Ticket Hypothesis, proved for basic networks by Eran Malach and colleagues in 2020, that a large enough randomly initialized network already contains, hidden inside it, a subnetwork that computes the function you want, before any training happens at all. Training doesn’t build the solution. It prunes away everything that isn’t the solution. The capability was present in the noise, and learning is the process of finding it. If that’s right, and it’s since been extended to the architectures that actually matter, then in some sense the trained network was always there. Scale just makes it likely to be in there somewhere.

The second is physical. Complex systems of interacting parts tend to organize themselves toward a critical boundary, the edge between order and chaos, and that boundary happens to be where computation works best. Beggs and Plenz found the fingerprint of this in actual cortex in 2003, activity that propagates in scale-free avalanches exactly as a system poised at criticality would predict, with no external hand tuning it there. The brain didn’t get designed to sit at that edge. It fell toward it, because that’s what these systems do. Evolution found brains without an engineer, which is the existence proof that intelligence can be reached by a process with no foresight at all, only scale and feedback and time.

The third is the pattern of the field itself. Independent research lines that started from completely different places keep arriving at the same math. In 2024 Tri Dao and Albert Gu proved that Transformers and state-space models like Mamba, which look nothing alike on paper, are two views of the same underlying operation, connected through a class of structured matrices. RWKV, coming at it from the direction of recurrent networks, lands in the same neighborhood. When separate teams chasing separate intuitions converge on equivalent solutions, the parsimonious reading is that they’re not each inventing something. They’re circling a destination that was already there, an attractor in the space of architectures, and their different paths are just different routes to the same place.

The fourth is biology, and it sets the scale of what’s still ahead. When David Beniaguev, Idan Segev, and Michael London asked in 2021 how complex a single cortical neuron really is, they trained artificial networks to imitate one, and it took a deep network of five to eight layers, on the order of a thousand artificial neurons, to reproduce what one biological neuron does. Evolution compressed all of that into a single cell. That number is proof of something specific: architectures far more efficient than anything we’ve built exist and are findable by optimization alone, because optimization already found them once, without a designer, running on chemistry.

Put the four together and the picture is coherent. The solution is latent in scale. Systems self-organize toward the regime where computation happens. Different search paths converge on the same answers. And biology proves there’s enormous headroom above where we are, reachable by a blind process. If all of that holds, then the honest way to describe what the labs are doing is not designing intelligence but cultivating it. We’re gardeners, not sculptors. You don’t carve a plant. You create the conditions, the soil and water and light, and the growing is done by the thing itself, following rules you didn’t write.

Now the part where I have to be careful, because a thesis this clean gets seductive, and seductive theses rot into slogans.

The largest hole is the one my own last post pointed at. Almost no networks produce intelligence. A hurricane is a vast system of interacting parts at scale, and it does not think. A stock market has feedback everywhere and produces panics, not minds. So “scale plus feedback yields intelligence” is plainly incomplete. There’s some set of constraints that separates the networks that get somewhere from the overwhelming majority that don’t, and I cannot tell you what those constraints are. Nobody can, precisely. Which means the inevitability thesis, stated honestly, has a hole in the middle exactly where it most wants to be solid. It says intelligence is inevitable given the right conditions, and then can’t fully specify the conditions. That’s a weaker claim than it sounds when you first say it out loud, and I’d rather admit that than smuggle it past you.

The second problem is that “inevitable” quietly changes meaning depending on how hard you lean on it. There’s a modest version: intelligence is reachable by undirected optimization, not a miracle requiring a spark of genius. I’m confident in that one. Then there’s a strong version: given scale and feedback, intelligence is more or less guaranteed to appear, on something like a schedule. That one is a much bigger bet, and the honest evidence for it is thinner than the confidence with which people, including me on a good day, tend to assert it.

So I’ll hold the line I can actually defend. Intelligence looks like something discovered rather than authored, latent in scale, approached by convergent paths, proven reachable by a blind process that already did it once in carbon. That’s a strong claim and I’ll stand on it. What I won’t pretend is that it comes with a guarantee or a timetable, because the missing piece, the constraints that decide which networks wake up and which just churn, is the piece we understand least and need most.

Which leaves the question that the whole thesis hands you and can’t answer. If intelligence is discovered rather than built, latent in networks we’re only beginning to know how to grow, then what else is latent in there, sitting on the far side of a scale we haven’t reached, waiting to be found by a search that isn’t aiming at it? The next four posts take the four pillars one at a time, because each one is more interesting, and more contestable, up close than it is in a list.