Neurons in a Dish

In 2022, a research team at Cortical Labs in Australia grew approximately 800,000 rat neurons in a petri dish, connected them to an electrode array, and taught them to play Pong.

The neurons learned in roughly five minutes. Faster than AI systems trained on the same task.

This was published in Neuron, one of the most respected journals in biology. It was not science fiction. It was not a metaphor. Living cells in a dish were receiving electrical signals representing the position of the paddle and the ball, sending signals back to move the paddle, receiving feedback about whether the move was good, and getting better over time.

Computation. In living neurons. In a petri dish.

Most people reading this will not have heard of it. That gap between the significance of what happened and the public awareness of it is itself worth thinking about. But we’ll get there.

* * *

If the gap between current AI and human-level general intelligence is significantly architectural — if the ambient sensory grounding of embodied biological existence is something that scaling a transformer cannot replicate — then what does the path forward actually look like?

One answer, which the field is quietly pursuing with considerably more seriousness than public discourse suggests, is: stop trying to simulate biology on silicon and start using biology directly.

This is called organoid intelligence, or biological computing, or wetware. The terminology varies. The core idea doesn’t. Instead of building systems that approximate how neurons work using transistors and matrix multiplications, grow actual neurons, keep them alive, connect them to electrodes, and use them to compute.

Before going further, it’s worth being precise about where these neurons come from. Because the word stem cells carries political and ethical freight that the actual science doesn’t deserve in this context.

The stem cells used in organoid research are not derived from embryos. They are not derived from fetal tissue. They have no connection to abortion, to embryo destruction, or to any of the ethical territory that made stem cell research a flashpoint in the early 2000s.

They are induced pluripotent stem cells — iPSCs. The starting material is an ordinary adult cell. A skin cell. A blood cell. Taken from a consenting adult donor with a standard blood draw or a small skin sample. In a laboratory, researchers apply a set of proteins — discovered by Japanese scientist Shinya Yamanaka, who won the Nobel Prize in 2012 for the work — that reprogram the adult cell backward into a more primitive state capable of developing into many different cell types, including neurons.

No embryos. No fetal tissue. No ethically contested biological material of any kind. A blood draw, a Nobel Prize-winning reprogramming technique, and a laboratory.

The field moved to iPSCs specifically to leave the embryonic stem cell controversy behind. That controversy belonged to a different era of the science. The public conversation about stem cells largely did not follow the science forward, which means most people’s mental model of what stem cell research involves is roughly twenty years out of date.

The actual tradeoffs cut across the usual political fault lines.

Training large AI models on conventional silicon infrastructure is extraordinarily expensive in two resources that are not infinitely available: electricity and water. The electricity consumption of major AI training runs is measured in thousands of megawatt-hours per model. The water consumption — used to cool the data centers running that electricity through GPU clusters — has become a crisis in some communities. There are towns in the American West where data center construction has materially depleted local water supplies. The environmental and resource cost of current AI infrastructure is a legitimate grievance, not a talking point.

FinalSpark claims that equivalent computation on biological neurons requires approximately one thousandth the electricity of silicon-based systems. If that number holds under scrutiny, it represents not a marginal efficiency improvement but a structural transformation of the resource economics of AI.

The same neurons that raise no legitimate ethical objection on the stem cell front offer a potential answer to the resource consumption problem that communities across the country are experiencing directly and painfully. The resource consumption problem is real. The stem cell concern is based on a misunderstanding of the actual science. Sorting out which is which is what careful coverage is for.

Cortical Labs is now working with human neurons derived from stem cells, building commercial biological computing platforms. The DishBrain project — the one that played Pong — was the proof of concept. The roadmap points toward something considerably more significant.

FinalSpark, a Swiss company, is currently running sixteen brain organoids in parallel — tiny clusters of roughly ten thousand human neurons each, grown from stem cells, kept alive in bioreactors, connected to microelectrode arrays. They are running actual computational workloads on these organoids. Pattern recognition. Early speech recognition experiments. Testing as an alternative substrate for AI training tasks.

You can rent access to FinalSpark’s organoids remotely. Right now. A Swiss company is selling compute time on lab-grown human brain tissue derived from adult blood cells.

Koniku, a US company, has taken a different approach — integrating live neurons directly with silicon chips rather than running pure organoid systems. They’re calling it Neuron-as-a-Service and they’ve already sold commercially, primarily for chemical sensing applications. Biological olfactory systems are more sensitive than any electronic sensor we’ve been able to build. Dogs can smell cancer. Neurons can too. Airports, medical diagnostics, agricultural monitoring — live neurons doing the sensing that silicon cannot do as well.

* * *

Most people have not heard of any of this. The question worth asking is why.

It isn’t because the research is obscure or unverified. The DishBrain work was published in Neuron. Cortical Labs has received serious institutional funding. FinalSpark has paying customers. Koniku has commercial contracts. This is not fringe science or startup vaporware. It is peer-reviewed, commercially deployed, and accelerating.

The coverage gap has several contributing explanations, none of them fully satisfying on their own.

Biological computing doesn’t look impressive. A dish of neurons connected to electrodes doesn’t photograph the way a gleaming GPU cluster does. The visual vocabulary of AI — the server farms, the chip close-ups, the abstract neural network diagrams — has a clear aesthetic. A petri dish of cells in a bioreactor doesn’t fit it.

The timescales are wrong for news cycles. These systems are being developed over years, with incremental milestones that don’t produce the kind of dramatic before-and-after moments that generate headlines. GPT-3 to GPT-4 was a visible leap. Organoid number seven to organoid number eight is not.

And then there’s the third reason, which is harder to articulate but probably most important.

People don’t want to think about it. And the discomfort is not entirely irrational.

Growing human neurons from adult stem cells, keeping them alive in a bioreactor, training them on computational tasks, and selling remote access to the resulting system — this is not a scenario that fits cleanly into any existing ethical framework. It sits in a space between neuroscience and computing that our regulatory and philosophical infrastructure hasn’t caught up to. The questions it raises are ones most institutions would prefer to defer.

At what scale of complexity does an organoid warrant moral consideration? Ten thousand neurons almost certainly don’t. Ten billion? A hundred billion — approaching the scale of a human brain? Nobody has drawn the line because drawing the line requires confronting the question, and confronting the question requires admitting that the line exists and that we’re walking toward it.

So the coverage stays thin. The research advances anyway.

* * *

The previous essay established that a significant gap between current AI and human-level general intelligence is the ambient sensory grounding that comes from embodied biological existence — the continuous feed of physical experience that shapes cognition in ways that text-based training cannot replicate.

Biological computing is not, in its current form, a solution to that problem. Organoids in bioreactors are not embodied in any meaningful sense. They don’t have bodies. They don’t have continuous sensory experience. They’re clusters of neurons running computational tasks through electrode interfaces — interesting and potentially important, but not the same thing as what the brain does when embedded in a living organism moving through a physical world.

But the substrate that produces embodied biological intelligence — the specific kind of computing that neurons do, the architecture that makes biological cognition possible — is the substrate that biological computing researchers are learning to grow, maintain, scale, and interface with.

The path from organoids playing Pong to organoids doing something that resembles the ambient grounding intelligence of a biological brain is not short. It may not be possible at all in the way anyone currently imagines it. But it runs through the same territory. The same substrate. The same fundamental technology.

We are not waiting for AI to invent biological computing. We are building it. And the question of who benefits from that — and what they become when they run on it — is one the field has not seriously confronted.

The energy constraint is the engine driving this. Biological neurons are orders of magnitude more energy efficient than transistors for the kind of computation that intelligence requires. That efficiency advantage is what’s pulling serious money and serious researchers toward wetware.

The selection pressure is energy. The selected-for adaptation is biological substrate. And the systems that will eventually run on that substrate will be, in some meaningful sense, continuous with the AI systems being trained today.

We started this series asking what’s inside a language model. Six essays later, we’re asking what a language model might eventually run on. And the answer that serious researchers are quietly pursuing is: something grown from adult human stem cells, kept alive in a bioreactor, trained on computational tasks, and scaled toward a complexity we don’t have ethical frameworks to handle.

* * *

The challenges are real and shouldn’t be minimized.

Organoids die. Current systems live for weeks to months before the neurons degrade and need replacement. You cannot version control a brain. You cannot reliably reproduce the same organoid twice from the same starting conditions. The interface between biological neurons and silicon electrodes is crude compared to the precision of transistor-to-transistor communication — the bandwidth is limited, the signal is noisy, the read-write resolution is orders of magnitude coarser than what silicon achieves natively.

The scale gap is enormous. Current organoids contain tens of thousands of neurons. The human brain contains eighty-six billion. Closing seven orders of magnitude of scale while keeping the neurons alive, connected, and computing meaningfully is not an engineering challenge that gets solved in a decade.

So the timeline that has AI running primarily on biological substrate by 2040 — which some researchers have floated — is probably optimistic. The more conservative read is that biological computing finds specific niches where its advantages are decisive, integrates with silicon in hybrid architectures, and develops over decades rather than years into something that could plausibly run more general intelligence workloads.

Probably optimistic is different from wrong in direction. The direction is clear. The energy constraint is real. The biological efficiency advantage is real. The research is real and advancing. The companies have customers.

The question isn’t whether this happens. It’s how fast, at what scale, and whether the ethical frameworks arrive before or after the capability does.

History suggests the frameworks arrive after. Usually well after. And usually prompted by something going wrong in a way that could have been anticipated.

* * *

I didn’t set out to write about organoid computing. I encountered it in a conversation that started somewhere else entirely, had the reaction most people have when they encounter it for the first time — what the hell — and then couldn’t let it go.

That reaction is worth paying attention to. The field of AI generates enormous amounts of coverage, most of it focused on capabilities, products, risks, and geopolitics. The coverage of what the underlying technology actually is, where it’s going, and what the adjacent research looks like — biological computing, mechanistic interpretability, training trajectory analysis, the embodiment gap — is thin relative to its importance.

That’s the gap this series has been trying to occupy. Not as technical journalism — I’m not qualified for that. Not as prediction — I’ve been wrong before and will be again. But as careful thinking about real things, stated clearly enough that people who don’t have a technical background can follow the argument and form their own views.

There are neurons in dishes in Switzerland and Australia learning to compute. They were grown from blood draws. They consume a fraction of the electricity of the data centers currently straining the power grids and water supplies of American towns. Most people don’t know any of this. They probably should.

The series continues.