Neuroscience · Biocomputing · The New Intelligence
Human neurons are learning to play video games faster than AI. Nobody programmed them. What this tells us about the nature of intelligence - and what comes next - may be one of the most consequential stories in science
In 2022, a group of researchers at a Melbourne laboratory announced that 800,000 human neurons growing in a petri dish had learned to play Pong. They had not been programmed. No reward function had been written. No parameters had been tuned. They simply adapted - and they adapted faster than state-of-the-art artificial intelligence. The paper was published in Neuron. The field it opened has no good name yet. The researchers are calling it organoid intelligence.
The DishBrain system, built by Cortical Labs in Melbourne, works by placing living neurons directly onto a multielectrode array - a chip covered in hundreds of tiny electrical contacts. The researchers used a mixture of mouse embryonic brain cells and human neurons derived from stem cells. The neurons did what neurons always do: they grew, formed connections, and began firing.
The Pong interface was simple but precise. Electrodes on the left or right of the array fired to indicate which side the ball was on. The frequency of the signal told the cells how far the ball was from the paddle. The collective electrical activity of the neural network controlled the paddle. When the paddle hit the ball, the cells received a stable, predictable burst of feedback. When it missed, they received chaotic, unpredictable noise.
The cells had no reason to prefer hitting over missing - except that the nervous system, as a general rule, strongly prefers predictability to chaos. Within five minutes, performance measurably improved. Within an hour, DishBrain was playing recognisable Pong. When the researchers later tested the same cells on an adapted version of the game with different rules, the cells adapted again. They then moved to DOOM. The cells adapted to that too.
In comparative trials using the CL1 - Cortical Labs' commercial biological computing platform - the neuronal cultures demonstrated significantly higher sample efficiency than state-of-the-art deep reinforcement learning algorithms. The cells needed less experience to reach the same level of performance. They learned, in the technical sense of that word, better than the AI systems built to replicate learning.
This is not a marginal result. Deep reinforcement learning - the family of algorithms behind AlphaGo, AlphaZero, and many current AI systems - typically requires millions of training examples to reach competent performance. DishBrain reached comparable performance in a fraction of that time, with no explicit training signal, on hardware consuming a fraction of the energy.
The theoretical framework that makes sense of this is the free energy principle, developed by neuroscientist Karl Friston: the idea that biological neural systems are fundamentally prediction engines that act to minimise surprise. The cells were not trying to win at Pong. They were trying to make their world more predictable - and hitting the ball with the paddle happened to be the most effective strategy for achieving that. Intelligence, on this account, is not a special property of sufficiently complex systems. It is what happens when any system tries to stay alive and coherent in an uncertain environment.
While Cortical Labs was building DishBrain, a parallel effort was underway at Johns Hopkins under Dr. Lena Smirnova. Rather than flat neural cultures on chips, Smirnova's team works with three-dimensional brain organoids: self-organising spheres of human brain cells, each the size of a needle tip and containing roughly 50,000 cells, grown from induced pluripotent stem cells.
These are not brains. They do not have the architecture, the blood supply, the sensory inputs, or the developmental history of a brain. But they do develop spontaneous neural activity, they form synaptic connections, and they respond to stimulation in ways that increasingly resemble the behaviour of real neural circuits. In 2025, Smirnova's lab published evidence that brain organoids show the building blocks necessary for basic learning and memory - specifically, the cellular and molecular machinery required for synaptic potentiation, the mechanism by which real brains encode experience.
The applications Smirnova's team is working toward are both practical and profound. In the near term, brain organoids grown from a patient's own cells can model that patient's specific neurology - allowing researchers to test drugs against a system that is biologically identical to the patient's brain, rather than against a mouse model or a cell line. For diseases like Alzheimer's, epilepsy, and autism spectrum disorder, where animal models are notoriously poor predictors of human outcomes, this is a potentially transformative research tool.
The longer-term ambition is biocomputing: using organoids as processing units in hybrid biological-silicon systems. The human brain processes information at approximately 20 watts. The AI systems we currently build to replicate some of its functions consume megawatts. If even a fraction of the brain's energy efficiency can be harnessed in synthetic form, the implications for computing at scale are difficult to overstate.
When the National Science Foundation launched its "Biocomputing through EnGINeering Organoid Intelligence" programme in 2024, it did something unusual: it required every research application to include a bioethicist as co-principal investigator, and it evaluated the ethics plan on equal footing with the scientific plan. This is not standard practice. It signals that the people funding this research understand that something qualitatively different is being attempted.
The question the ethicists are there to address is not whether DishBrain or a 50,000-cell organoid is conscious. The scientific consensus is that current systems are nowhere near the complexity required for consciousness as we understand it. The question is about trajectory. If a flat culture of 800,000 neurons can learn a task faster than a deep learning algorithm, and if 3D organoids are developing the cellular machinery for memory formation, the path from here to systems of much greater complexity - and much less certain moral status - is not a theoretical one. It is an engineering timeline.
We are not building minds. But we are building systems that share more properties with minds than any machine ever built before - and the gap is closing faster than the ethical frameworks are developing.
- Lisa Pedrosa
The leading researchers in the field are aware of this. A 2025 paper in Springer Nature's Current Stem Cell Reports called explicitly for an ethical-moral framework to govern organoid intelligence research before the field outpaces its ability to govern itself. STAT News reported in late 2025 that some pioneers in brain organoid research are concerned that inflated claims about biocomputing could trigger a backlash that derails research with genuine medical value before it reaches patients.
These are the early signs of a field finding its conscience - which is, historically, a better outcome than the alternative.
Three implications stand out from everything happening in this field, each significant in its own right.
The first is medical. Brain organoids grown from patient cells are already being used to test drugs for neurological conditions. The ability to test a compound against a system that is biologically identical to the patient's own neural tissue - before giving it to the patient - represents a fundamental shift in the logic of drug development. The majority of drugs that succeed in animal trials fail in human trials. A significant fraction of those failures are neurological, because rodent brains are structurally so different from human ones. Organoid testing does not eliminate this problem, but it reduces it substantially.
The second is computational. The energy gap between biological and silicon intelligence is enormous. If biocomputing systems can be developed that harness even a fraction of the efficiency of biological neural networks, the economics of AI infrastructure change completely. Current AI training runs consume electricity on a scale that strains national grids. A biological computing layer that handles certain classes of problem - particularly those involving rapid adaptation to novel environments - with the efficiency of living tissue would be a significant technological and economic development.
The third is philosophical. DishBrain did something that challenges a comfortable assumption: that intelligence, learning, and adaptation are properties of sufficiently sophisticated software running on hardware, and that the hardware itself is irrelevant. The cells that learned to play Pong were not running an algorithm. They were being themselves - doing what neurons do when faced with an uncertain and responsive environment. The implication is that what we call intelligence may be, at some level, a property of living matter itself rather than a property of computation. That distinction has consequences that reach well beyond neuroscience.
© 2026 Lisa Pedrosa · lisapedrosa.com
All articles cited to primary institutional or peer-reviewed sources
Buy me a coffee