In a lab in Melbourne, a sealed chamber the size of a shoebox keeps about 200,000 human brain cells alive in warm fluid.
The cells sit on a chip studded with electrodes. They have no eyes. They have never seen a screen.
And in early 2026, they were shown playing Doom.
Nobody was holding a controller. The cells were the controller.
Quick answer: Cortical Labs, an Australian biotech company, demonstrated its CL1 “biological computer” — living human neurons grown on a microelectrode chip — learning to play a version of Doom.
The game state is converted into electrical stimulation, the cells’ firing patterns are read back as in-game actions, and feedback shapes their behaviour over about a week.
The neurons are genuinely part of the loop, but the open-source code shows most of the heavy computation happens on ordinary silicon alongside them.
A CL1 unit costs around $35,000, or you can rent access through the cloud.
That’s the accurate version. The interesting part is everything underneath it.
What Cortical Labs actually demonstrated
Cortical Labs CTO David Hogan described how independent developer Sean Cole routed the game’s video feed into patterns of electrical stimulation, so that specific learned firing patterns made the in-game character shoot, and other patterns moved it.
The game used was Freedoom, a free clone running on the original Doom engine, and there is no screen involved from the cells’ perspective — an enemy on the left stimulates a left-side “sensory” zone of the electrode array.
Chief scientific officer Brett Kagan was blunt about the skill level.
He said the system is no esports champion, but insisted that the cells are learning, and that feedback for right and wrong actions still needs refining.
Kagan has also compared the play style to a complete beginner, and the Doom clip is a company video rather than a peer-reviewed result — a distinction worth holding onto for the rest of this piece.
The demo was built in roughly a week, and within that time the culture showed novice-level play: finding targets and firing with basic competence.
The hardware is the real news
Doom is the headline. The machine is the story.
The CL1 is Cortical Labs’ first commercial system, built around 59 electrodes arranged on a flat array, with denser contact and upgraded signal processing that cut response latency to sub-millisecond speeds.
The living tissue sits in a sealed chamber with an internal life-support system managing gas, temperature and waste filtration, and the neurons stay viable for up to six months in good conditions.
The cells aren’t harvested from brains.
They’re cultured from skin or blood cells taken from adult donors and reprogrammed — a process I’ll come back to, because it’s the single development that made all of this possible.
There’s also an operating system.
Cortical Labs built biOS, which sends and receives electrical stimuli through the electrode array, letting developers deploy code directly to the neuron layer.
That’s what the company means by calling the CL1 “code deployable.”
It was announced in March 2025 at $35,000 and shipped that summer.
Cole, notably, was not a biologist.
He had almost no biocomputing experience before the project and used the Python interface to get Freedoom running in about a week.
That accessibility is arguably a bigger milestone than the gameplay.
Why build a computer out of cells at all?
Electricity. That’s the entire pitch.
Your brain runs on roughly 20 watts — about the draw of a dim light bulb — while AI data centres are now competing with cities for power.
The demand has become severe enough that a shuttered reactor at Three Mile Island, site of the worst nuclear accident in US history, is being revived largely to feed them.
Constellation Energy announced in September 2024 that it would restart Unit 1, supplying 835 megawatts to grid operator PJM under a 20-year agreement with Microsoft.
The $1.6 billion restart of the renamed Crane Clean Energy Center was pulled forward from 2028 to 2027, backed by a $1 billion federal loan.
Against that backdrop, the biological numbers look absurd.
Cortical Labs CEO Hon Weng Chong told Bloomberg that each CL1 uses less power than a handheld calculator.
The company now runs a Melbourne facility of 120 CL1 units, each reported at about 30 watts, powering its Cortical Cloud service — with a second biological data centre planned in Singapore.
For comparison, a single AI training cluster can draw 100 megawatts.
The gap isn’t incremental. It’s several orders of magnitude.
But be careful with the direction of that comparison.
A CL1 drawing 30 watts is not doing anything close to what a GPU cluster does.
The energy argument is a statement about biological efficiency in principle, not a benchmark result.
Nobody has yet shown living neurons outperforming silicon on a useful task at any scale.
How do you teach a dish of cells anything?
You can’t reward a neuron with a treat.
So the researchers used something more fundamental: neurons appear to dislike unpredictability.
The DishBrain cultures were trained according to the free energy principle — the theory that neural systems act to reduce uncertainty — so missing the ball in Pong triggered random stimulation, while hitting it triggered ordered, predictable stimulation.
The cells have no concept of Pong or Doom.
They’re not trying to win. They’re trying to make the noise stop.
Whether that counts as learning in any meaningful sense is precisely what the scientific fight is about.
That fight is worth understanding.
The 2022 paper, published in Neuron by Kagan and colleagues including UCL’s Karl Friston, was titled “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.”
DishBrain used roughly 800,000 neurons; the team grew human cells from stem cells and mouse cells separately, and found the human cultures sustained longer rallies after the same training time.
(Worth correcting a common retelling: the human and mouse cells were separate cultures being compared, not a single hybrid dish.)
The word “sentience” detonated.
A formal response published in Neuron in March 2023 argued that the negative reaction from the scientific community was legitimate, objecting to the claim that a single layer of cortical neurons had displayed intelligent and sentient behaviour, and raising concerns about methodology, controls, unsupported terminology, and overselling of the work’s relevance.
Cortical Labs has since moderated its language.
The company’s use of “sentience” referred narrowly to the cells’ apparent preference for ordered over random stimulation, and it has since worked with some of its vocal early critics on an ethics framing study.
This matters commercially as well as philosophically.
Cortical Labs positions the CL1 as an ethically preferable alternative to animal testing that yields more relevant human data — a genuinely valuable use case that hype backlash could damage.
The part the press releases skip: what the code shows
Here’s where this story gets better than its coverage.
Cole published the project openly, and his own documentation is unusually candid.
In the repository README, he notes that his decoder — the conventional software translating neural activity into game actions — tends to start becoming a policy head, meaning the software may be learning to route around the neurons entirely.
He built in ablation modes so anyone can test whether the cells contribute at all.
The architecture splits three ways.
Cole’s CL1-side program acts as a hardware interface, receiving stimulation commands over UDP, applying them to the neurons, then returning spike counts to a separate training system — while the PPO reinforcement-learning agent, the CNN encoder and the reward loop all run on ordinary silicon elsewhere.
So: do the cells matter?
Cole’s own documentation reports a clear difference between the ablation conditions and real CL1 spikes — both random and zeroed spikes produce no learning — and states that the footage was captured using a zero-bias linear readout, meaning the action was a linear function of the cells’ output.
He also argues the cells aren’t a static input-output box: biological neurons carry internal state, so identical stimulation at different points in training produces different spike patterns, and performance kept improving even with encoder weights frozen.
Then someone checked.
R&D World cloned the repository and ran the experiment hundreds of times on GPU hardware, confirming that the structured signal matters and ruling out the simplest version of the concern — that the decoder is a standalone policy ignoring its input.
But they were explicit that a software-only run cannot confirm living neurons are the ones learning; that requires the full biological loop with the same ablation controls, and their results varied sharply by random seed.
Honest summary: the neurons are load-bearing, not decorative.
They are also not the brains of the operation.
Conventional hardware still handles game simulation, data acquisition, stimulation calculation and parts of the learning system — the CL1 replaces neither CPU nor GPU here.
What remains unproven is whether 200,000 human neurons could ever carry the policy rather than ride along.
How we got here
Three developments made the CL1 possible, and none of them are new.
Wiring cells to machines is old.
In 2008, a University of Reading team led by Kevin Warwick, Mark Hammond and Ben Whalley used hundreds of thousands of rat neurons on a 60-electrode array to steer a wheeled robot away from obstacles, hoping to study how brains store information.
Even then, observers pointed out that Steve Potter’s group at Georgia Tech had published closely related work back in 2002 — the “overhyped milestone” pattern has been following this field for two decades.
Reprogramming cells removed the ethical bottleneck.
Shinya Yamanaka’s demonstration that a small set of factors could return an adult cell to a stem-cell-like state — published in 2006 in mice and 2007 in human cells, and recognised with the 2012 Nobel Prize — meant researchers no longer needed brain tissue.
Skin or blood works.
That’s why the CL1 can honestly say human neurons without anyone donating a brain.
Three dimensions is the unsolved part.
Grown flat, neurons mostly connect to their immediate neighbours.
Grown in 3D, they self-organise into layered structures resembling early brain tissue — organoids.
But the CL1’s array is flat, so a spherical organoid would only touch it at the bottom.
To talk to the cells clearly, you give up the architecture that makes brains brain-like.
Can you actually get one?
Yes, if you have a lab budget.
The CL1 sells for around $35,000, with remote access available through the Cortical Cloud as an alternative to buying a unit.
The subscription model already exists elsewhere.
Swiss startup FinalSpark runs the Neuroplatform, which offers online access to 16 human brain organoids housed across four multi-electrode arrays, each array holding four organoids interfaced by eight electrodes.
Academic users pay $500 per month for access to four shared organoids, and selected projects get it free.
The organoids are roughly half a millimetre across, hold around 10,000 neurons each — about 160,000 system-wide — and the platform launched in Vevey in May 2024, with peer-reviewed documentation in Frontiers in Artificial Intelligence.
FinalSpark’s headline claim is the energy one.
The company says bioprocessors consume a million times less power than traditional digital processors, and its paper notes that training a model like GPT-3 required roughly 10 GWh — about 6,000 times what an average European citizen uses in a year.
Treat “a million times” as a theoretical efficiency comparison, not a measured like-for-like benchmark.
The limits nobody puts in the press release
Keeping cells alive is the bottleneck, not making them compute.
Cortical Labs describes feeding and cleaning the neurons as laborious and time-consuming, and swaps tubing every five to six months.
FinalSpark’s organoids were initially dying within hours; improved systems extended lifespans to around 100 days, against silicon chips that last decades.
There is no accepted benchmark showing a biological processor beating a conventional one at a real task.
There’s a working interface, genuine adaptation, and a maintenance bill.
The uncomfortable question, handled honestly
Intelligence and experience are different things.
A system can be capable and feel nothing, and — in principle — the reverse.
Neuropsychoanalyst Mark Solms has argued for years, most fully in The Hidden Spring, that affect isn’t a late add-on sitting atop cognition but something more basic, rooted in the drive to maintain the conditions for staying alive.
If that’s right, then a training method built on making instability stop is poking at something closer to the origins of feeling than to the origins of thought.
That’s a hypothesis, not a finding.
It’s also contested by researchers who locate consciousness firmly in large-scale cortical integration that 200,000 flat-grown cells simply don’t have.
What we can say with evidence is that we’re unreliable at detecting awareness even in humans.
A 2024 New England Journal of Medicine study of 241 patients who couldn’t respond to bedside commands found that one in four showed sustained, relevant cognitive responses on EEG or fMRI — tasks like imagining playing tennis, the paradigm Adrian Owen pioneered in 2006.
The same study found that over 60% of patients who were clearly conscious failed the scanner tests, suggesting the phenomenon may be more common than 25%.
If bedside tests miss conscious people lying in front of us, no one should feel confident ruling anything in or out in a dish.
That cuts against alarmism too: cells reducing their response to painful noise is exactly what non-sentient tissue does.
David Chalmers named this gap the hard problem in 1995, and it hasn’t moved.
Cortical Labs’ answer, for the record: Kagan rejects the interpretation that playing a video game demonstrates consciousness.
What to watch next
Three concrete things, not vibes.
First, whether anyone connects a true 3D organoid to a high-density interface without flattening it.
That’s the missing architecture.
Second, whether a biological processor posts a benchmark result on a task that matters.
Until then, the energy argument is a promise.
Third, whether the disease-modelling work — testing drugs on human neural tissue instead of mice — delivers.
That’s the application most likely to justify the field, and the one most exposed if the next Doom-style stunt overshoots again.
The cells in that Melbourne chamber began as skin from a living person.
They now spend their existence inside a loop of calm and chaos.
Whether that’s a tool, a model, or something we’ll be embarrassed about later is a question the evidence cannot currently settle — which is exactly why it’s worth asking now, while the answer is still cheap to act on.
(Article body: ~2,640 words)
9. FAQ
Is the CL1 really playing Doom by itself?
No.
The neurons are a genuine part of the learning loop — Cole’s ablation tests show that replacing their spikes with noise or zeros kills learning — but the game engine, the reinforcement-learning agent and the reward logic all run on conventional hardware.
Where do the human neurons come from?
Skin or blood cells from adult donors, reprogrammed into stem cells and then differentiated into neurons.
No brain tissue is harvested.
How much does a CL1 cost?
Around $35,000 per unit, announced at that price in March 2025, with cloud access as an alternative.
How long do the neurons live?
Up to about six months inside the CL1 under good conditions.
FinalSpark’s organoids are estimated at around 100 days.
Can the cells feel anything?
There’s no test that can answer this.
Cortical Labs’ chief scientist rejects the idea that game-playing demonstrates consciousness, and the 2022 use of “sentience” drew a formal published objection from other scientists.
Is this more energy efficient than a GPU?
Per unit, dramatically — but it isn’t doing comparable work.
No published benchmark shows a biological processor beating silicon on a useful task.
