The Computers Inside the Petri Dish: How Living Neurons Learned to Play Doom

Inside a Melbourne lab, a dish the size of a quarter holds a small clump of human neurons kept at body temperature for months. The dish has learned to play Pong. A second dish, in a different lab, has learned to play Doom. A third has guided a small robot through a maze.

These are organoids: small, lab-grown clusters of human cells, often neurons, that researchers can keep alive for months and stimulate electrically. They are not miniature brains in the everyday sense, and they are not conscious. But they can be programmed, in a loose sense, by the same kind of feedback loop that powers reinforcement learning in silicon. The wire version of this story is that brain cells play video games. The more useful version is the closed electrical feedback loop that teaches them, and what happens when living tissue becomes a substrate you can buy.

The mechanism is older than the headlines. In 2022, a team led by Brett Kagan at Cortical Labs published a paper in Neuron showing that in vitro neurons could learn to play Pong when correct hits were rewarded with brief, predictable electrical stimulation. Kagan and his co-authors called the result a form of sentience. The term drew coverage; the more durable finding was methodological. A dish of neurons, given a feedback signal it could predict, learned to play the game in roughly fifteen minutes ([Kagan et al., Neuron 2022](https://www.sciencedirect.com/science/article/pii/S0896627322008066)).

Cortical Labs has since turned the demo into infrastructure. Its CL1, shipped in 2025, packages the same kind of neuronal substrate on a chip and sells access to it as a research platform (corticalabs.com/cl1). Competitor FinalSpark, based in Switzerland, sells remote access to a similar biological-computing stack. IEEE Spectrum covered the commercial turn as the moment lab work became something a paying customer could plug into (IEEE Spectrum, "Biological Computer: Human Brain Cells on a Chip"). That is the relevance clock. A parallel computing track has moved from preprint to product.

The most striking demos are not the games. At the University of California San Diego, Alysson Muotri's lab has used organoids to steer small robots through obstacles and to study the effects of psychedelics on neural tissue. The lab presents the work as a way to model aspects of human neural development that animal systems cannot (Muotri Lab, UC San Diego Pediatrics). In a separate project, neurons-on-a-chip learned to play Doom. They did not recognize pixel patterns; they produced oscillatory signals that researchers mapped to in-game actions (Scientific American, "How Human Neurons on a Chip Learned to Play Doom"). A 2022 release from Monash University documented an earlier round of the Pong result (Monash, 2022). The throughline is closed-loop feedback. Give the neurons a stimulus, let them respond, reward the response that gets the system closer to the goal.

That loop is the part the wire copy skips. A dish that produces a brain-wave-like oscillation in response to a stimulus is not a computer the way a CPU is a computer. The neurons are stochastic, slow, and largely opaque to inspection. The "learn" and "sentience" language in the Kagan paper is author choice, not a community-wide claim, and the energy-efficiency numbers from CL1-class systems are vendor claims that have not been independently benchmarked. Organoids do produce repetitive oscillations that resemble the brain waves of a premature baby, and that signal is what researchers can tune. It is not, on its own, evidence of thought.

What it changes, narrowly, is the substrate menu. Reinforcement learning has, for a decade, meant gradient updates on silicon. A dish of neurons, given the right feedback signal, appears to learn in a similar way. Not because the neurons understand Pong, but because the closed loop shapes their activity toward the rewarded output. The pattern is the thing. Engineers can already use CL1 to study drug effects, neural disease, and low-energy information processing without an animal in the loop. The substrate is wet, the API is electrical, and the cost is unclear.

What it does not change is the silicon AI track. The Neuron paper, the Pong result, and the Doom demo are real, but they are also narrow. A single task, a single feedback signal, a substrate that must be fed, kept at body temperature, and replaced every few months. Biological computing is a parallel lane, not a replacement lane. The honest read is the one the Wired feature gestures at without quite saying. A small group of biologists is betting that living tissue deserves a seat at the table next to silicon, and the first product is already in customers' hands.

The watch item is whether anyone outside the field can reproduce the closed-loop learning result on a CL1 they did not help design. Cortical Labs says the platform is shipping. The next twelve months will tell whether it is the start of a substrate shift, or an interesting demo with a long road to a real workload.