DishBrain: brain cells in a dish that learned to play Pong

December 28, 2021

In December 2021 researchers at Cortical Labs, a startup in Melbourne, Australia, reported that living brain cells grown in a dish had learned to play the arcade game Pong. The system, called DishBrain, was not an artificial brain, and the game was not physical ping pong. It was a culture of roughly 800,000 to a million brain cells, taken from mouse embryos or grown from human stem cells. The cells sat on a microelectrode array, a chip that can both stimulate cells and record their electrical activity, connected to a simulated game in which a paddle returns a bouncing ball.

The results were first posted as a preprint on December 3, 2021 and, after peer review, published in Neuron in October 2022, with co-authors from Monash University, the University of Melbourne and University College London.

How the cells learned the game

The game itself ran on an ordinary computer. Pulses through eight sensory electrodes told the cells where the ball was: which electrodes fired marked its position, and faster pulses meant it was closer to the paddle. Activity in designated regions of the culture moved the paddle up or down, so the cells acted as the paddle.

When the paddle hit the ball, the neurons received a brief, predictable stimulus; when it missed, they got unpredictable, random stimulation. The design followed the free energy principle of the neuroscientist Karl Friston, a co-author of the paper, which holds that brains act to reduce uncertainty about their surroundings. To keep their world predictable, the cultures had to keep hitting the ball, and the researchers reported signs of learning within about five minutes of play.

From experiment to product: the CL1

Cortical Labs has since turned the research into hardware. In March 2025, at Mobile World Congress in Barcelona, it announced the CL1, which it calls the first code-deployable biological computer: lab-grown human neurons on a silicon chip, kept alive by built-in life support for up to six months. Priced at about $35,000 and aimed mainly at research labs, it was due to ship by the end of June 2025. Teams without a lab can use Cortical Cloud, a "wetware as a service" offering that runs their code on the company's own units.

Developers program the CL1 through a Python library, and a simulator lets them test code before it touches real cells. In early 2026 a CL1 culture of about 200,000 human neurons was shown playing Doom, using software an independent developer built on the Python API. Conventional computers still ran the game and translated it into stimulation patterns, and by Cortical Labs' own account the cells played like a complete beginner.

Biological and neuromorphic computing versus silicon AI

The appeal is efficiency. NIST estimates that the human brain performs the equivalent of an exaflop on about 20 watts, while the Frontier supercomputer needed some 20 megawatts to reach that speed. Cortical Labs argues that living neurons need less energy and less training data than conventional systems. Neuromorphic chips chase the same advantage in silicon: Intel's Loihi 2 imitates spiking neurons, and in April 2024 Intel unveiled Hala Point, a research system of 1,152 Loihi 2 chips with capacity for 1.15 billion artificial neurons.

The limits are clear: a culture lives for months, depends on sterile life support and so far handles simple games and lab experiments, not workloads that compete with GPUs.

The ethics debate, and what developers can do today

The paper's title said the neurons "exhibit sentience", and that wording drew criticism. In March 2023 a group of neuroscientists led by Fuat Balci argued in a response published in Neuron that the claims of intelligence and sentience went well beyond the data. Coverage of the 2026 Doom demonstration also noted that there is no evidence the cells are conscious or feel pain.

For developers and businesses, biological computing in 2026 is a research tool, not a production platform. Realistic uses are neuroscience and drug-discovery research and prototyping closed-loop code in the simulator before booking time on real neurons. Brain–computer interfaces, such as the Neuralink implant that let a monkey play video games, approach the same boundary between biology and software from the other side.