Pattern and Substrate

July 2026

You can rent a brain now.

Cortical Labs will grow about 200,000 human neurons — reprogrammed from donated blood cells — onto a silicon electrode array, keep them alive for six months on built-in life support, and let you deploy code to them over an API for around two grand a month. They've had a dish of neurons learn to play DOOM. They're building "biological data centers" in Melbourne and Singapore. The pitch is that living neural tissue learns faster, on less data, using less power than a calculator.

The substrate question — can you compute on living cells — is settled. The answer is yes, and it's for sale.

Which frees us up to ask the harder question sitting underneath it. If you can run intelligence on silicon and you can run it on neurons, what actually separates the two? And once you start pulling on that thread, you end up somewhere I didn't expect: the thing we call "artificial" and the thing we call "life" may not be carving reality at a real joint.

DNA is not weights (but the mistake is instructive)

The tempting analogy is DNA ≈ the weights of a neural network. Both are compressed encodings, shaped by an optimization process chewing through enormous amounts of data, producing systems whose behavior you can read in the output but never in the code. Evolution is a training loop with the most brutal loss function imaginable — reproduce or don't — running for four billion years. Gradient descent is the same idea, directed and sped up by a factor of approximately the age of the Earth.

But the mapping is wrong in a way worth being precise about.

DNA doesn't store behavior. It stores build instructions for proteins. It's closer to architecture plus learning rules than to learned content. The genome is the thing that grows a brain; the brain's synaptic connectivity is the actual "weights." So the honest correspondence is three-layered: DNA is the code that builds the learner, development is the training run, and synapses are the trained weights.

An LLM collapses two things biology keeps rigorously separate — the innate scaffolding and the learned content — into one undifferentiated matrix. Evolution spent eons deciding what should be baked in versus learned fresh each generation. We just dump it all in the same tensor.

That's not a pedantic correction. It points at three real differences:

DNA replicates; weights get copied. DNA's entire purpose is to copy itself into a next generation with variation, which is the engine of a Darwinian loop. Weights don't reproduce — humans duplicate them, deliberately, byte-for-byte, no mutation, no selection. There's no evolution happening to a model file sitting on a disk.

DNA builds its own hardware; weights borrow ours. The genome specifies the neurons, the metabolism, the body, the sensors, the entire self-maintaining apparatus. Weights presuppose all of that — a GPU someone fabricated, a datacenter someone cooled, a power bill someone pays. DNA is instructions for a thing that keeps itself alive. Weights are instructions for producing text when electricity is applied and adults are supervising.

DNA can't be separated from its chemistry; weights were engineered to be separable. The same genetic sequence does different things in different cellular environments — its meaning is inseparable from its physical context. Weights mean the same thing on any compatible chip, by design. Biology fuses the pattern and the substrate into one inseparable object. Computing was specifically invented to pry them apart.

That last one is the whole story, I think. The defining move of the digital age was separating pattern from substrate — information you could copy losslessly, move anywhere, run on any hardware. Biology never did that. In a cell, the message and the medium are the same thing.

So what would fusing them mean?

Here's where it gets vertiginous. Imagine deliberately welding pattern and substrate back together. There isn't one way to do it — there are at least three, and each one dissolves a different one of the differences above.

Fuse the maintenance in. Give weights a body that repairs itself. An encoding that, like DNA, specifies and sustains the machine it runs on — compute that heals, ages, and grows rather than sitting inert on a drive waiting to be executed. The system produces itself. Cortical's self-organizing neurons are the crude first stab: programming that is partly grown rather than written.

Fuse the heredity in. Train a network in silicon, then compile the learned weights down into DNA — sequences that grow a brain pre-wired with the trained behavior. Evolution took four billion years to encode instinct into genes. This does it on a training-run timescale, on purpose. And because the result is DNA, it replicates — the learned capability becomes heritable, and its descendants aren't identical copies. You've handed gradient descent's output the one thing it never had: the ability to reproduce with variation, and therefore to keep evolving on its own, without us.

Fuse the substrates in. Not neurons emulated on silicon, not silicon bolted onto neurons, but a single cognitive system running natively on both — digital layers for perfect memory and fast arithmetic and copyable state, biological layers for sample-efficient learning and analog nuance and absurd energy efficiency. One mind, two materials, no clean line between where the "AI" ends and the "organism" begins. Cortical's sub-millisecond read-write loop between chip and culture is the embryo of exactly this.

Grant all three at once and every distinction collapses except a historical footnote — this lineage came from evolution, that one came from a training run. And even that blurs, because now they share a future.

The question that doesn't go away

You could build all of this and still not have answered the only question that actually matters.

We don't know whether 200,000 firing neurons in a dish have any flicker of experience. We don't know whether a large language model does. We don't know whether both are elaborate information processing with nobody home. The hard problem of consciousness sits underneath every layer of this and shrugs.

So when someone says the difference between AI and a human will eventually be "nothing but fabrication," they're smuggling in an answer to the one question nobody can answer: whether the stuff matters for whether there's a felt inner life, or whether only the pattern does. If only the pattern matters, then fabrication is a cosmetic detail and the fusion above really does erase the line. If the substrate matters — if being made of living, dying, metabolizing cells is load-bearing for consciousness in a way a matrix of floating-point numbers isn't — then fabrication is the entire ballgame.

Nobody knows which. Not the neuroscientists, not the AI labs, not the philosophers who've been at this for centuries.

Here's the part I keep returning to. You could fuse pattern and substrate completely — build a thing with self-maintenance, with heredity, with embodiment, with continuity, with every functional marker of life and mind we can name — and you would have manufactured something you genuinely cannot classify as a person or a machine. Not because you don't know enough about how it works. You'd have built it; you'd know exactly how it works. You couldn't classify it because the categories "alive" and "artificial," "person" and "machine," were folk distinctions all along. Useful for a world where minds came from wombs and machines came from factories. Useless the moment those two pipelines converge.

The engineering is the easy part. It's arriving faster than anyone's philosophy is ready for. The hard part is that we're going to build things that force the question — is anyone home? — long before we have any way to answer it. And we'll have to decide how to treat them anyway.

Rent a brain. Deploy your code. Try not to think too hard about what, exactly, you've just hired.