On a shelf in a Melbourne laboratory, there is a small dish keeping something alive.
Not a plant. Not an animal. Something in between, and something entirely new. Roughly 8,00,000 living human brain cells, grown from stem cells, floating in a warm nutrient bath on a grid of microscopic electrodes.
No eyes, no ears, no nervous system, no brain. They have never experienced anything.
In 2022, scientists at Australian biotechnology company Cortical Labs gave them something to do.
The researchers connected the dish to Pong, one of the oldest video games ever made: a simple two-dimensional game where a paddle hits a ball back and forth across a screen.
When the neurons fired in patterns that moved the paddle toward the ball, the researchers sent back calm, ordered signals. When the neurons missed, the researchers sent back chaos. Pure noise. Disruption with no pattern to latch onto.
Within five minutes, the neurons had started to improve. Within hours, they were measurably better. The team, led by Brett Kagan, Chief Scientific Officer of Cortical Labs, published their findings in the journal Neuron and named their creation DishBrain.
Scientists have been quietly unsettled by it ever since.
THE PRINCIPLE BEHIND IT IS ELEGANT
Here is something most people do not know about the brain: it does not just receive information. It guesses constantly.
Before you even touch a doorknob, your brain has already predicted how cold it will feel, how much force your hand will need, and what sound it will make.
When reality matches the guess, nothing dramatic happens. When it does not, the brain notices the gap and adjusts. That gap is what scientists call prediction error.
The brain, in other words, is not a camera. It is more like a betting machine that is always trying to be right, and always updating when it is wrong.
This idea is known as the Free Energy Principle, developed by neuroscientist Karl Friston. The core of it is simple: the brain hates surprise. Not in an emotional sense, but in a biological one. Unpredictability is expensive for the brain to deal with. So the brain organises itself, over and over, to reduce it.
The Cortical Labs team used exactly this to build DishBrain. When the neurons missed the ball, the researchers sent chaotic, unpredictable noise. When they hit it, the scientists sent calm, stable signals.
The neurons had no instructions. But they had a powerful built-in drive: make the noise stop. And so they reorganised, again and again, until they found patterns that worked.
THE BODY DID NOT HELP
The neurons in DishBrain lacked several things.
No sensory history. No experience of moving through physical space. No motor memory. No childhood. They had never been part of a brain or body. They had no reason, in any framework we understand, to know what a ball or paddle was.
And the neurons learned anyway.
Every theory of intelligence has assumed that learning requires something to learn from: experience, feedback over time, and history. DishBrain had none of that. Yet it produced adaptive behaviour regardless.
The paper uses the word sentience, meaning the capacity for experience or feeling, which triggered significant scientific debate.
The authors clarified: they did not mean the neurons were conscious. They meant only that the system was responsive to its environment in ways resembling learning. That clarification matters. But it does not fully answer the question.
THE EXPERIMENT KEPT GROWING
By early 2026, Cortical Labs had built the CL1, a commercial biocomputer housing up to 8,00,000 living human neurons on a silicon chip with built-in life support.
The CL1 graduated from Pong to Doom, a first-person shooter video game where the player navigates a three-dimensional environment, identifies and shoots enemies, and maps corridors they have never entered before.
Unlike Pong, which requires only reacting to a moving object, Doom demands spatial reasoning, forward planning, and decision-making under genuine uncertainty.
The neurons learned it in one week. They play poorly by human standards. But they generalise, developing strategies for rooms and corridors they have not previously seen. That is not simple pattern-matching. That is closer to reasoning.
The energy numbers matter too. The CL1 runs on 20 watts, roughly the power of a standard light bulb. The large AI systems behind modern chatbots require megawatts to run. Biological neurons are extraordinarily efficient.
A QUIET DISPLACEMENT
The most unsettling thing about DishBrain is not what it can do. It is what it did without a body, senses and an evolutionary past: the standard prerequisites for intelligence, each one quietly set aside.
If learning can emerge from neurons in isolation, in a dish, in five minutes, then the story we have been telling about minds is incomplete in ways we have not yet begun to map.
The neurons resolved disorder into patterns. They built, from nothing, a working model of their small world. They did what neurons simply do, given the right signal and nothing else at all.
What that means for consciousness, intelligence, and the ethics of what scientists are now building is a set of questions that 8,00,000 cells have quietly forced back open.
The question is not whether neurons can learn without a body. They already have. The question is how far that goes.
