What if you don't need a brain to think?

Last week I wrote about a memristor where the material itself behaves like a neuron. Nine devices, no software, the physics did the learning. A few of you had things to say about that.

Douglas Scroggs , who's been building a bistable lattice simulator, showed me that he ran 14 experiments proving symmetric coupling can only do spatial filtering. You need asymmetric, activity-dependent coupling for associative recall. I'd been building with STDP and homeostatic scaling for months without being able to articulate why those choices mattered. Douglas gave me the vocabulary in one comment.

Brian Jones told me about Daimon, his cognitive architecture in Zig. No transformers, no LLMs. Hebbian learning, spreading activation, predictive coding. His system independently converged on an 800ms processing loop, almost exactly the 780ms relaxation time of the memristor I'd written about. Two completely different substrates arriving at the same timescale. That's either a coincidence or it isn't.

And Xuejing Zhou wrote something that stuck with me: intelligence might emerge from "collective interactions among uncountable tiny models." I replied at the time that this maps well to biology. No single neuron is intelligent, it comes from billions of simple units interacting locally. That's one of the ideas behind our work at the Center for Applied Complexity & Intelligence: intelligence as something that emerges from complex adaptive systems, not something you program into a single architecture.

I believed that when I wrote it. But I didn't have a good example of just how far down the "simple units" can go.

Then I watched a talk by Josh Bongard .

The frog skin experiment

Bongard is a computer scientist at the University of Vermont. He works with Michael Levin, a biologist at Tufts who has spent years studying how cells communicate and make collective decisions. Together they've been building what they call xenobots. The recipe is straightforward, if you happen to have a microsurgery lab. You take embryonic cells from the African clawed frog, Xenopus laevis. Specifically skin cells. You separate them from the embryo and put them in a dish. About 2,000 of them.

What happens next is the part that messes with your assumptions.

If that sounds like a tumor, you're not wrong to notice the resemblance. Cells outside their normal context, forming their own structure, growing into something the original organism never intended. The comparison is worth taking seriously, because cancer turns out to be more interesting than "cells going rogue." The atavism theory of cancer (Lineweaver, Davies, Vincent, 2021) argues that tumors aren't random chaos. They're a reversion to ancient unicellular programs. Cancer cells systematically over-express genes that are older than multicellularity, 600+ million years old. They reactivate ancestral stress responses, error-prone DNA repair from bacteria, ancient mobility programs. It's not a breakdown into noise. It's a coordinated step backward in evolutionary time.

And here's the part that complicates the simple story: cancer cells cooperate with each other. Tumor cell clusters migrate collectively during metastasis, with leader cells creating invasion tracks and follower cells trailing behind. They share growth signals, divide labor, collectively remodel their environment. Circulating tumor clusters are far more dangerous than single cells precisely because they maintain cell-to-cell contacts and coordinate. A 2023 paper in Evolutionary Applications called it "the paradox of cooperation among selfish cancer cells." Cells that defected from the organism's cooperation then build their own cooperative structure.

Levin frames it this way: cancer is what happens when cells disconnect from the bioelectric communication network that maintains the body's large-scale goals. They stop listening to the tissue-level signals that say "you are part of a liver" and revert to pursuing their own cellular-level goals. His lab has shown that you can actually suppress cancer phenotypes by forcing bioelectrical connections back onto tumor cells, overriding the single-cell agenda with the collective one.

So both tumors and xenobots are cells outside their original context, building new organization. The difference is the direction. Cancer cells revert to ancient selfish programs and then rebuild just enough cooperation to serve those selfish goals. Xenobot cells find genuinely new cooperative organization that nobody programmed and nobody predicted. The cells move around, figure out where they belong relative to each other. If you poke a hole in the blob with a needle, the surrounding cells migrate into the wound and close it. Self-healing, without any immune system, without any central coordination. Tumors don't do that. They grow, they invade, they cooperate for survival. But they don't repair.

But that's not the finding that got to me.

Thomas Varley, a postdoc in Bongard's lab, put these xenobots under a fluorescence microscope and recorded what individual cells were doing over time. Each cell produces calcium signals, brief flashes of calcium ion concentration visible as white pulses under the right imaging. Calcium signaling is the same mechanism neurons use to communicate. It's ancient, predating nervous systems by hundreds of millions of years. Varley recorded the calcium time series from every cell in the xenobot. Then he did something simple: he computed the pairwise correlation between every cell's signal and every other cell's signal. Which cells flash together? Which cells are independent? Is there any structure in there, or is it just 2,000 cells doing their own thing?

There was structure.

The cells had organized themselves into distinct communities, groups whose calcium signals co-varied, rising and falling together. Cell 47 and cell 1,203 on opposite sides of the xenobot, flashing in sync, while their immediate neighbors did something completely different. The communities weren't spatially contiguous. They were spread throughout the body, interleaved, with no obvious physical explanation for why those particular cells were coordinated. Varley ran the standard controls. He shuffled the time series data randomly and checked whether the permuted data produced the same community structure. It didn't. The organization was genuine, not an artifact of the analysis. Here's the part I keep turning over. The community structure Varley found looks like fMRI functional connectivity maps of human brains. The same kind of analysis, pairwise correlation of time-varying signals plus community detection algorithms, applied to oxygen-level data from brain scanners produces the same sort of result. Distributed, non-contiguous communities of co-active regions. In brains, we call these "resting state networks" and we associate them with cognition.

In a blob of frog skin, I'm not sure what to call them.

What the cells are actually doing

It's worth being specific about what these cells are and aren't. They're not neurons. They don't have synapses or axons or dendrites. They don't fire action potentials. They're epithelial cells, the stuff that makes up the outer layer of your skin. But cells talk to each other. All cells do. They release small molecules that diffuse through the surrounding fluid. They have gap junctions, tiny protein channels that directly connect the interior of one cell to the interior of its neighbor, allowing ions and small molecules to flow between them. They sense the chemical environment around them and change their behavior in response.

Neurons are just cells that evolved to be really good at a specific version of this. Faster signaling (milliseconds vs. seconds), longer range (axons can be a meter long), and a clean binary output (the action potential). But the basic machinery, sensing the environment, processing it, responding, is in every cell. That's Levin's core argument, and the xenobot data backs it up. The calcium signals in the xenobot aren't as fast or as clean as neural signals. They propagate through diffusion and gap junctions, not axons. The timescale is seconds, not milliseconds. But the organizational principle is the same: cells forming functional communities that share information across the body.

Why I care about this

In June last year I wrote a piece called "The Goroutine Epiphany" where I'd gotten excited about modeling neurons as independent software agents, each running in its own thread, communicating through message passing. I wrote: "Maybe intelligence isn't something you calculate. Maybe intelligence is something you grow." I was talking about software. Go-routines in Go, later actors in Rust. Each one a little agent with its own state, its own timeline. I built a small network, made the neurons drunk by messing with their GABA/glutamate balance, and watched the network degrade in patterns that matched real neuroscience literature on alcohol impairment. Fine motor control went first, gross motor function last. I hadn't programmed that. It emerged from the dynamics.

Since then I've been following a trail of papers and experiments that keep saying the same thing from different angles. Prof. Dr. Bernd Ulmann 's team at anabrid ran a neural network on their Model 1 analog computer, a 4-node oscillator network (nodes with homogeneous frequencies and damping, all amplitude-coupled, with no feedback connections) with 16 trainable parameters. Transferred from digital training, the analog version initially scored 28% because the readout expected floating-point precision. When they retrained just the readout layer to work with the noisy, drifty analog signals, it matched the digital version's accuracy. Six parameters. The analog dynamics preserved enough information that only the interface needed fixing, not the computation. Ludovico Minati built a physical network of 59 transistor oscillators wired with the topology of a real neuronal culture, symmetric connections everywhere, and directed information flow emerged spontaneously. Transfer entropy analysis revealed causal structure that simple correlation missed entirely. Then the memristor paper from Li et al. Nine perovskite devices with nickel atoms scattered on graphene, solving the traveling salesman problem faster than simulated annealing. The material dynamics were the algorithm.

And now frog skin cells organizing themselves into brain-like functional communities without anyone telling them to.

The same principle keeps showing up in places it shouldn't. Analog circuits, transistor networks, engineered memristors, biological skin cells. Give agents local interactions and time, and functional organization shows up. This has been a thread through everything we do at the Center for Applied Complexity & Intelligence, exploring this across substrates, from dataflow processor architectures to biological neuron simulations. The xenobot result hit different though, because nobody designed anything. Nobody optimized anything. The cells just did it.

The question I can't get rid of

If skin cells can do this, what are neurons actually for?

I don't mean that dismissively. Neurons are extraordinary. A single cortical neuron integrates thousands of inputs across a dendritic tree, maintains multiple timescales of plasticity simultaneously, and fires with millisecond precision. The fruit fly connectome, 139,000 neurons and 50 million synapses, produces navigation, learning, and social behavior from a brain smaller than a pinhead. But maybe neurons aren't the cause of intelligence. Maybe they're the substrate that evolution found most efficient for it. An optimization over a more general principle that was there all along, in every cell, from the beginning. Dennis Noble, the Oxford physiologist, has been making a version of this argument for decades. He wrote "The Music of Life" partly as a response to the gene-centric view of biology. His point: you can't find agency in any single component. Not in the gene, not in the protein, not in the cell. It's in the pattern of interactions. Levin takes this further and says every cell is an agent. It senses, processes, and acts. The brain is what happens when you put a lot of agents together and give them good communication channels.

Daniel Dennett would approach it differently. He would say the question of whether the xenobot is "thinking" is less interesting than whether it's useful to describe the xenobot as if it's thinking. His intentional stance: when predicting a system's behavior, if the intentional description (it wants, it prefers, it's trying to) is more accurate and more compact than the mechanistic description (cell 47's calcium concentration exceeded 400 nanomolar due to IP3 receptor activation following paracrine ATP signaling from...), then the intentional description is the right one to use. Not as a metaphor. As the more accurate model.

I wrote about philosophical zombies in April last year, arguing that AI systems act conscious without being conscious. Frog skin xenobots flip that question. Here's a system with no pretense of consciousness, no language, no behavior we'd call intelligent. And yet the internal organization looks like a brain's. Maybe consciousness and intelligence aren't binary things you have or don't have. Maybe they're more like temperature, a spectrum where even skin cells register something above zero.

I genuinely don't know. But I keep thinking about Xuejing's comment. "Collective interactions among uncountable tiny models." She was talking about AI architectures. It turns out biology got there first, with equipment simpler than anyone expected.

One more thing

@Mike W . and I where discussing on my memristor post last week: "the answers were in the dirt, not in the abstraction." That was in a thread about how Turing spent his last years on morphogenesis, not computation theory. Darwin's final book was about earthworms.

I keep coming back to that. The people who understood intelligence best always ended up studying the physical substrate. Not the math, not the theory. The dirt.

Maybe that's where the next answers are too.

Paper reference

"Identification of brain-like complex information architectures in embryonic tissue of Xenopus laevis organoids"

Varley, T.F., Pai, V.P., Grasso, C., Lunshof, J., Levin, M., & Bongard, J. (2025)

Communicative & Integrative Biology 18(1).

DOI: https://doi.org/10.1080/19420889.2025.2568307

Also discussed: Bongard's talk at https://youtu.be/kmofEynMNE8