They were walking into walls, shooting the walls, doing funny things
In a Melbourne laboratory, 200,000 human neurons grown on a silicon chip have learned to navigate the world of Doom — not as a parlor trick, but as a window into what intelligence, efficiency, and biological computing might mean for our technological future. The cells, consuming only 20 watts, learned by doing: fumbling at first, then adapting, then succeeding — mirroring the ancient rhythm of how minds have always grown. This is less a story about video games than about the boundary between the biological and the digital beginning, quietly, to dissolve.
- Lab-grown neurons on a chip went from walking into walls and firing at barriers to actively targeting enemies — real, measurable learning unfolding in real time.
- The tension is not just scientific but existential: if biological cells can learn, adapt, and compute at a fraction of the power silicon requires, the entire architecture of modern AI is called into question.
- Researchers are threading carefully between breakthrough and overreach — the cells live only six months and produce inconsistent results, keeping ambition tethered to honest limitation.
- The platform is already being aimed beyond gaming at drug screening, disease modeling, robotics, and personalized medicine, suggesting the stakes extend far past any single demonstration.
- The story is landing in a space where legitimate progress and profound uncertainty coexist — not fringe science, but not yet a solved problem either.
In a Melbourne laboratory, 200,000 human brain cells grown from stem-cell donations have taught themselves to play Doom. Living on a silicon chip called the CL1, developed by Cortical Labs, the neurons began their education with Pong — mastering basic timing and movement before being introduced to the far more demanding world of a 3D shooter.
Their early encounters with Doom were clumsy. The cells walked into walls, fired at barriers, spun in circles. Senior application scientist Alon Loeffler watched the fumbling with the patience of someone observing a first-time player. Then something shifted. The neurons began recognizing patterns, targeting enemies with growing regularity, adapting their behavior in response to feedback. The learning was real.
The mechanism works by translating Doom's digital environment into electrical signals the neurons can interpret. Electrodes stimulate the cells when enemies appear; different patterns of neural activity produce different actions — move, fire, turn. Scientists monitor thousands of individual neurons on screen, using that data to refine their inputs over time.
Chief scientific officer Brett Kagan is measured about what has been achieved. The cells last only six months and don't yet produce consistent results. But the platform points toward applications in robotics, AI, drug screening, disease modeling, and personalized medicine.
What gives the work its deeper significance is power consumption. The human brain runs on roughly 20 watts — an efficiency that silicon-based AI has never matched. As data centers grow hungrier for electricity, that gap becomes urgent. Semiconductor researcher William Keating calls this legitimate science making measurable progress on a real problem. The goal, Kagan insists, is not to replace artificial intelligence, but to offer something different: a biological intelligence that learns, adapts, and runs cool — and whose full potential researchers are only beginning to explore.
In a laboratory in Melbourne, a cluster of 200,000 human brain cells grown from stem cells harvested in blood donations has learned to navigate a three-dimensional world, hunt enemies, and fire weapons. The cells live on a silicon chip called the CL1, developed by researchers at Cortical Labs, and they have taught themselves to play Doom—the 1993 shooter that defined a generation of gaming.
The journey began with something simpler. The neurons first mastered Pong, the paddle-and-ball game that requires only vertical movement and basic timing. That success was the proof of concept. But Doom demanded something more: spatial reasoning, target recognition, strategy, adaptation. When the cells first encountered the game's chaotic 3D environment, they flailed. They walked into walls. They fired at barriers instead of demons. They spun in circles. Alon Loeffler, Cortical Labs' senior application scientist, watched this early fumbling with the eye of someone observing a child learning to play. The neurons were operating at the level of a complete beginner, someone who had never held a controller before.
Then something shifted. The cells began to recognize patterns. They started targeting enemies with increasing regularity. The shots still came in multiple directions before finding their mark—a demon might require several attempts to kill—but the learning was unmistakable and real. The neurons were adapting to stimuli in real time, adjusting their behavior based on feedback, completing goal-directed tasks. This is not trivial. This is learning.
The mechanism is elegant in its strangeness. Researchers converted the digital environment of Doom into patterns of electrical signals that neurons could interpret. When an enemy appears on screen, specific electrodes on the chip stimulate the cells, triggering a reaction. Different patterns of neural activity produce different outputs: fire the gun, move left, move right. Scientists monitor thousands of tiny dots on a computer screen—each one representing the electrical activity of individual neurons—and use that data to adjust their inputs, training the culture's behavior over time.
Brett Kagan, Cortical Labs' chief scientific and operations officer, is careful not to oversell what has been accomplished. The cells have a lifespan of six months. They do not yet produce consistent, programmable results. But he is clear about what the work suggests: these neural cultures could be applied to robotics, to real-time learning tasks that mimic artificial intelligence, to healthcare and disease modeling, to drug screening and personalized medicine. The CL1 is not limited to games. It is a platform.
What makes this work genuinely significant is not the novelty of playing Doom. It is the efficiency. The human brain operates on approximately 20 watts of power—a level of energy consumption that silicon-based computing and artificial intelligence have not yet managed to match. As computing demands grow and data centers consume ever more electricity, that gap matters. William Keating, CEO of semiconductor research firm Ingenuity, frames it plainly: this is not fringe science or fraud. It is legitimate research making measurable progress on a real problem. The question is not whether neurons can play video games. The question is whether they can do it while consuming a fraction of the power that conventional chips require.
Kagan emphasizes that the work is not intended to replace artificial intelligence. Rather, it aims to give researchers capabilities they have never possessed before—a different kind of intelligence, one that runs cool and efficient, grown from human cells, capable of learning and adapting in ways that silicon alone cannot yet replicate. The researchers say they are only beginning to understand what these neural cultures might accomplish.
Citações Notáveis
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
A Conversa do Hearth Outra perspectiva sobre a história
Why start with Doom? Why not something simpler, something more controlled?
Pong was the control. They needed to prove the cells could learn at all. But Doom is the real test—it's chaotic, it requires spatial reasoning, it demands adaptation. A beginner playing Doom for the first time looks a lot like those neurons did.
The cells were walking into walls. That sounds like failure.
It was. But it was the kind of failure that leads somewhere. They weren't just random. They were learning from each attempt, adjusting. That's the whole point—real-time adaptation.
Six months. That's the lifespan. Is that a problem?
It's a limitation, yes. But it's also a window. In six months, you can run experiments, gather data, understand how the system works. The question is whether you can scale it, whether you can keep the culture alive longer, whether you can make it more reliable.
The power consumption angle—is that the real story here?
It might be. Playing Doom is the headline, but the headline is just the proof. The real story is that a biological system can do cognitive work on 20 watts. That's the number that matters to people building data centers.
Do you think this replaces AI?
No. Kagan was explicit about that. This is a different tool. It's not better or worse—it's different. It might be better at some things, worse at others. The point is having options.