They were walking into walls, shooting the walls, doing funny things
In a Melbourne laboratory, human neurons grown from stem cells have been taught to play Doom — not as a novelty, but as a demonstration that biological tissue can learn, adapt, and pursue goals in real time. The achievement, by the team at Cortical Labs, sits at the intersection of neuroscience, computing, and philosophy of mind, raising quiet but profound questions about what intelligence is, where it lives, and how little energy life requires to think. It is not yet a revolution, but it is the kind of early signal that tends, in retrospect, to mark a turning point.
- Two hundred thousand lab-grown neurons — starting from Pong, graduating to Doom — are demonstrating that biological tissue can learn goal-directed behavior without being born into a body.
- The cells began by walking into walls and firing at nothing, a fumbling that researchers recognized not as failure but as the raw, unmistakable first motion of learning.
- The gap between biological and silicon computing is thrown into sharp relief: a human brain runs on roughly twenty watts, while modern AI systems consume energy at a scale that strains power grids.
- Cortical Labs is navigating carefully between genuine breakthrough and overreach — the CL1 chip works, but lifespans are short and results are not yet reliably programmable.
- The path forward points toward drug screening, disease modeling, and sustainable AI — a future where the most powerful computers may be grown, not manufactured.
Inside a Melbourne laboratory, two hundred thousand human brain cells — grown from stem cells derived from blood donations and embedded in a chip no larger than a postage stamp — are learning to play Doom. The researchers at Cortical Labs began with Pong, the simplest possible test of neural responsiveness, and watched the neurons master it. Then came the leap to a chaotic three-dimensional world of corridors and enemies.
At first, the cells were lost. They collided with walls, fired at nothing, circled without purpose. But Alon Loeffler, the company's senior application scientist, saw in that fumbling the beginning of something real. Over time, the neurons began targeting enemies with increasing regularity — imperfect, but intentional. The game's digital environment is translated into electrical signals that stimulate specific electrodes on the chip; the neurons respond, and their patterns of activity are read as actions in the game. It is, in essence, a dialogue between biology and machine conducted entirely in electricity.
Brett Kagan, Cortical Labs' chief scientific and operations officer, is measured in his claims. The cells live for only six months. Consistent, programmable results remain elusive. But the underlying proposition is significant: the human brain operates on approximately twenty watts — an efficiency that no silicon chip or AI system has come close to matching. If biological neural networks could be scaled and harnessed, the energy implications alone would be transformative.
The CL1 chip is already being considered for drug screening, disease modeling, robotics, and machine learning. Industry observers describe it as legitimate science making genuine progress. For now, the neurons in Melbourne are still learning — still adapting to a game they were never designed to play — and the researchers watching them are only beginning to imagine what these cells might eventually become.
In a laboratory in Melbourne, something unusual is happening inside a silicon chip no larger than a postage stamp. Two hundred thousand human brain cells, grown from stem cells harvested out of blood donations, are learning to navigate a three-dimensional maze filled with demons. They are playing Doom, the 1993 shooter that defined a generation of computer gaming, and they are getting better at it.
The researchers at Cortical Labs didn't start here. First came Pong—the simplest possible game, a paddle moving up and down to volley a ball across a screen. The neurons mastered that. Then came the leap to Doom, a chaotic environment where survival means exploring unfamiliar territory and eliminating threats. At first, the cells were hopeless. They walked into walls. They fired at walls. They spun in circles. Alon Loeffler, the company's senior application scientist, watched this fumbling performance and saw something else: the beginning of learning.
Slowly, the neurons began to target enemies with more regularity. Not perfectly—a single demon might require multiple shots from multiple angles before falling—but with unmistakable intent. The cells were adapting to stimuli in real time, completing tasks with a goal in mind. This is not trivial. It suggests that biological neural tissue, when properly interfaced with the right technology, can do something that silicon chips have struggled to replicate: learn efficiently while consuming almost no power.
The mechanism is elegant in its strangeness. The game's digital environment is converted into patterns of electrical signals. When an enemy appears on screen, specific electrodes on the chip stimulate the neurons, triggering a reaction. Different patterns of neural activity produce different outputs—fire the weapon, move left, move right. Researchers monitor thousands of tiny dots representing electrical activity on a computer screen, then adjust their inputs to shape and train the neurons' behavior. It is a conversation between biology and machine, conducted in the language of electricity.
Brett Kagan, Cortical Labs' chief scientific and operations officer, is careful not to oversell what's happening. This is not artificial intelligence replacing human thought. The cells have a lifespan of six months. They cannot yet produce consistent, programmable results reliably. But Kagan sees something larger: a proof of concept for a fundamentally different kind of computing. The human brain operates on approximately twenty watts of power—a level of efficiency that neither silicon chips nor artificial intelligence systems have managed to approach. If biological neural networks could be harnessed at scale, the implications for energy consumption alone would be staggering.
The applications extend far beyond gaming. The CL1 chip, as it's called, can be reprogrammed for drug screening, disease modeling, robotics, and machine learning tasks that mimic artificial intelligence. William Keating, CEO of semiconductor research firm Ingenuity, frames it plainly: this is legitimate science making real progress, not speculation or fraud. The value proposition is straightforward—better power efficiency, higher computational density, new capabilities that silicon alone cannot provide.
What happens next remains uncertain. The neurons in that Melbourne lab are still learning, still adapting, still playing a game they were never designed to play. The researchers say they are only beginning to understand what these cells might become. For now, they are content to watch the demons fall, one imperfect shot at a time, and to imagine what comes after.
Notable Quotes
They were walking into walls a lot, shooting the walls, turning around, doing funny things like that. And then eventually they started targeting the enemies more regularly and correctly.— Alon Loeffler, Cortical Labs senior application scientist
We are just scratching the surface of what these neural cultures can achieve when integrated in systems like our CL1.— Brett Kagan, Cortical Labs chief scientific and operations officer
The Hearth Conversation Another angle on the story
Why does it matter that brain cells can play a video game? Isn't that just a parlor trick?
It's not the game that matters—it's what the game proves. These cells are learning in real time, adapting their behavior based on feedback. That's the hard part. Silicon can compute, but it struggles to learn efficiently while using almost no power. A human brain does both on twenty watts.
So this is really about power consumption?
Partly. But it's also about what becomes possible when you have a computing substrate that learns the way biology learns. Drug screening, disease modeling, maybe eventually things we haven't imagined yet.
The cells only last six months. That seems like a major limitation.
It is. Right now, consistency is a problem too. But this is the first time anyone has shown it can work at all. The timeline matters less than the proof.
Do the researchers think this will replace AI?
No. They're explicit about that. This isn't meant to replace anything. It's meant to give us abilities we don't have yet—sustainable, efficient computing that learns the way living systems learn.
What's the biggest unknown right now?
Whether you can scale it. Two hundred thousand cells in a lab is one thing. Building systems that are reliable, consistent, and useful in the real world is another entirely.