A few hundred thousand human neurons sit in a nutrient bath in a lab in Melbourne, wired into a video game. When they aim at a monster and land a hit, the signal running into them goes calm and orderly. When they miss, or die, they get flooded with electrical noise. Within about a week, they get noticeably better at not dying.
Quick answer: A “biological computer” is a real device that grows living neurons — often reprogrammed from human skin cells — on a grid of electrodes that can both read their electrical activity and stimulate them. Companies like Cortical Labs (Australia) and FinalSpark (Switzerland) have taught these neuron cultures to play Pong and Doom by rewarding good in-game moves with predictable electrical signals and punishing bad ones with chaotic noise. The technology is real and the learning is real, but independent testing shows most of the actual “thinking” still happens in ordinary silicon chips wrapped around the cells — the neurons contribute a smaller, harder-to-pin-down piece. Whether any of this involves genuine feeling or awareness is unresolved and, by most current science, unanswerable.
Why Anyone Would Build a Computer Out of Living Cells
The honest answer is electricity. Training and running modern AI models takes an enormous and fast-growing amount of power, and the demand keeps outpacing what utilities can supply on a normal timeline.
The scramble for power has gotten strange. Constellation Energy is spending roughly $1.6 billion to restart Three Mile Island’s undamaged reactor — the same site that hosted the worst commercial nuclear accident in U.S. history — purely to sell all of its output to Microsoft’s data centers. Amazon has since bought a data center campus next to another Pennsylvania nuclear plant, and Google and Meta have signed deals with companies building smaller next-generation reactors. None of that is happening because AI companies suddenly love nuclear power; it’s happening because they’ve run out of easier options.
Your brain, by contrast, runs on roughly 20 watts — about what a dim light bulb uses — and it can learn what a dog looks like from a single glimpse, not thousands of labeled photos. That gap is what pulled a handful of researchers toward an odd question: what if you built computing hardware out of actual neurons instead of trying to imitate them in silicon?
Three Breakthroughs Had to Happen First
Neurons Driving a Robot
The idea of wiring living brain tissue to a machine isn’t new. In 2003, a team led by Steve Potter at Georgia Tech grew a small culture of rat cortical neurons on a flat grid of electrodes and connected it to a wheeled robot, a project they called a “hybrot.” The neural activity steered the robot’s motors, and sensor feedback from the robot stimulated the cells in return, letting it move toward light and away from obstacles.
It was a real closed loop, but a crude one — the neurons only ever formed a flat, two-dimensional sheet, connected only to whichever other cells happened to be nearby on the grid. That raised a question researchers couldn’t yet answer with the tools available at the time: were the cells actually driving the behavior in any meaningful sense, or was the system just dressing up essentially random neural noise as purposeful movement? It’s the same question, in a more sophisticated form, that people are still asking about CL1 and Doom two decades later.
Turning Skin Cells Into Neurons
The second piece came from stem cell research. In 2006, Shinya Yamanaka’s lab in Kyoto discovered that switching on four specific genes could reset an adult mouse cell all the way back to an embryonic-like state — a technique later replicated with human cells and honored with the 2012 Nobel Prize in Physiology or Medicine. From that reset state, a cell can be redirected into becoming almost any tissue type, including neurons.
That mattered enormously for biological computing. It meant researchers no longer needed to extract neurons directly from a brain; they could grow them from something as simple as a small skin sample.
Letting Neurons Grow in Three Dimensions
The third ingredient was structural. Neurons grown flat on a dish only connect to their immediate neighbors on that surface. Left to grow in three dimensions instead, they self-organize — spreading out, folding into layers, and forming rough approximations of real brain regions, roughly the way tissue organizes itself in an embryo.
A small 3D cluster of tissue built this way is called an organoid. The catch, which still hasn’t been fully solved, is that you can’t easily do both things at once: fully 3D tissue is hard to read and stimulate with a flat electrode grid, so most working systems today still flatten the cells out to keep the electrical interface intact.
DishBrain: Neurons That Learned to Play Pong
In 2022, the Melbourne-based company Cortical Labs combined these three threads into a system it called DishBrain. Researchers grew roughly 800,000 neurons — a mix of cells from mouse embryos and cells derived from human stem cells — on a dense electrode array, then connected that array to a simulated game of Pong.
Two things stood out. First, a dish of cells was working toward an actual goal rather than just firing randomly. Second, the team had found a workable way to “ask” the cells a question through stimulation and read an answer back through their spiking activity — something that hadn’t been demonstrated at this scale before.
How Do You Actually Train a Neuron?
You obviously can’t offer a neuron culture a reward it enjoys, so the Cortical Labs team leaned on something neurons demonstrably dislike: unpredictability. When the simulated paddle missed the ball, the cells were hit with scrambled, chaotic electrical noise. When the paddle connected, the stimulation became calm and predictable.
The cells have no concept of Pong, winning, or losing. What they’re doing, according to the researchers, is trying to minimize surprise — an idea borrowed from neuroscientist Karl Friston’s “free energy principle,” which holds that living systems are fundamentally organized around avoiding unpredictable states, because unpredictability tends to signal danger. If that’s right, the drive that let these cells learn Pong isn’t something engineers bolted on — it may be one of the most basic properties of living tissue.
CL1: A New Generation Learns to Play Doom
By early 2025, Cortical Labs had built a successor platform called CL1, which keeps cultures of around 200,000 human neurons alive for up to six months and exposes them through a Python programming interface, so researchers without a biology background can run experiments directly. In early 2026, the company demonstrated CL1 playing id Software’s 1993 shooter Doom, built largely by an outside collaborator, independent developer Sean Cole, in about a week.
Doom is a much harder problem than Pong: a 3D environment with enemies, movement in multiple directions, and weapons, instead of a paddle moving on one axis. The game’s video feed was translated into patterns of electrical stimulation delivered to the culture, and the resulting spikes were decoded back into in-game actions like moving, turning, and shooting. Cortical Labs co-founder Dr. Brett Kagan was blunt about the results: the cells are “learning,” but nobody should mistake this for a competent player — the culture dies constantly and behaves, in his words, like a beginner.
Are the Neurons Actually Doing the Work?
This is where the story gets more interesting, and more honest, than most of the headlines suggested. Cole published the full code on GitHub, along with something unusual for a splashy demo: built-in “ablation modes” that let anyone swap the neurons’ output for random noise or silence it entirely, to test whether the biological signal mattered at all.
According to Cole’s own documentation, both of those ablation conditions produced no learning, while the real neural signal did. That’s a meaningful result, but it comes with an important caveat that a separate outlet, R&D World, later confirmed by independently rerunning the software side of the pipeline hundreds of times: the test can only rule out the simplest concern, that the decoder was secretly playing the game on its own with the neurons contributing nothing. It can’t by itself prove the living cells are the ones generating the learned behavior, and the outcome varied noticeably depending on the random seed used to start each run.
The more sober picture that emerges is this: most of the architecture around CL1 — the code that reads the screen, decides on rewards, and translates game state into electrical patterns — runs on ordinary computer chips, not neurons. The living cells appear to contribute something real, but it’s a smaller and harder-to-isolate piece of the system than “brain cells playing Doom” implies.
The Backlash Over the Word “Sentience”
When Cortical Labs first published its Pong results in the journal Neuron, the paper’s title used the word “sentience,” which normally implies the capacity to have subjective experience. The authors meant it in a narrower, technical sense related to responsiveness and learning, but the choice of word set off a public argument. Roughly thirty scientists signed an open letter arguing that Cortical Labs had stretched the term for publicity, and outside researchers cautioned that the actual evidence showed adaptive learning, not proof of any inner experience.
The concern behind the backlash was practical as much as semantic: overselling a young field invites a harsh correction later, and this kind of research has real value that a hype cycle could put at risk. Cortical Labs’ own chief executive, Dr. Hon Weng Chong, has pointed to the DishBrain platform’s usefulness for studying conditions like epilepsy and dementia — disorders that are hard to model accurately in mice, since a mouse brain simply isn’t organized the same way a human one is.
Living human neurons also give drug researchers something closer to a real testing ground. A compound can be applied directly to tissue built from human cells rather than extrapolated from animal results that don’t always translate, which can shorten the path from a promising drug candidate to a clinical trial and reduce how much animal testing is needed along the way. None of that depends on whether the cells can play a video game — it’s a separate, more mature branch of the same underlying technology, and arguably the part most likely to produce practical benefits in the near term.
You Can Already Rent a Living Brain
The commercial side of this field is moving faster than most people realize. The Swiss company FinalSpark operates a platform called Neuroplatform, which gives researchers 24/7 remote access to 16 living human brain organoids over the internet, for around $500 per user per month through academic access, with some projects offered free. FinalSpark says its living tissue can be up to a million times more energy-efficient than a conventional digital processor for certain kinds of computation, though other company statements have cited more conservative figures closer to 100,000 times.
What began as a strange one-off lab demonstration is quietly turning into subscription infrastructure, with researchers at multiple universities already running experiments on the platform remotely. It’s reasonable to expect that a fully three-dimensional, properly interfaced organoid system — closing the gap between “grow it in 3D” and “read it in real time” — is only a matter of when, not if.
The Real Question: Is Anyone Home?
Here’s where the topic moves from engineering into something closer to philosophy, and it’s worth taking seriously rather than treating as science fiction. Most of the “intelligence” in these systems clearly lives in ordinary silicon. But being clever and being aware are not the same thing, and it’s worth asking separately whether there’s any form of experience happening in the biological part at all.
South African neuroscientist Mark Solms has argued, in his book The Hidden Spring, for something close to the opposite of the traditional view. Rather than treating feeling as something layered on top of intelligence, Solms locates the roots of consciousness in the ancient drive of any living system to minimize surprise and maintain internal stability — the same free energy principle behind how these neuron cultures are trained. If he’s right, the calm-versus-chaos training signal used in DishBrain and CL1 isn’t just a clever engineering trick; it may sit uncomfortably close to the actual biological root of feeling.
We’re also not very good at detecting awareness even in people. In a landmark 2006 study published in Science, neuroscientist Adrian Owen and colleagues asked a young woman diagnosed as vegetative to imagine playing tennis while inside an fMRI scanner. Her brain activity matched that of fully conscious volunteers doing the same task, and later, larger studies found that roughly one in five patients given this diagnosis show similar signs of covert awareness. If clinicians can miss consciousness in a fully formed human brain lying in front of them, spotting it — or ruling it out — in a dish of a few hundred thousand neurons is a genuinely unresolved problem, not a rhetorical one.
Philosopher David Chalmers named this gap the “hard problem” of consciousness: even a complete map of every neural connection wouldn’t, by itself, explain why or whether any of it is accompanied by actual felt experience. That leaves an uncomfortable asymmetry. It’s completely possible to build a system that reduces its own “stress” response to an unpredictable signal — as these cultures do — without that system feeling anything at all. From the outside, current science has no test that reliably tells the two apart.
Where This Goes Next
None of this makes the underlying research worthless — quite the opposite. Living neuron platforms offer a genuinely useful tool for studying disease and drug response, and a genuinely interesting new direction for low-power computing. The Doom demo was, by the developer’s own admission, partly a stunt built to satisfy an internet meme, but the code behind it is unusually transparent about what it can and can’t yet claim.
The technology is likely to keep improving on two fronts at once: better interfaces for reading and stimulating fully three-dimensional tissue, and more rigorous testing — like the ablation studies already built into CL1’s code — to figure out exactly how much of the “thinking” comes from the cells versus the silicon wrapped around them. Expect the gap between DishBrain’s flat, 2D cultures and FinalSpark’s rentable 3D organoids to keep narrowing, and expect more independent groups to run their own ablation tests rather than taking any single lab’s word for what the neurons are or aren’t doing.
The ethical and philosophical questions aren’t going away either, and they shouldn’t be treated as a footnote to the engineering. Every neuron in these systems ultimately traces back to a skin sample from a real, living person, reprogrammed and regrown into something that can be stimulated, stressed, and rewarded on command. That fact alone is reason enough to keep asking what, if anything, is happening inside that dish, rather than letting the question get buried under the novelty of watching a petri dish take a few shaky steps through a 1993 shooter.
