Scientists Grew Brain Cells That Can Learn and Remember

The Brain in a Dish That Learned to Play Pong

In 2022, a team of researchers at Cortical Labs in Melbourne did something that made a lot of people uncomfortable. They took roughly 800,000 human brain cells, grown from stem cells into a tiny ball of tissue, placed them on a grid of electrodes, and showed them the video game Pong. The cells learned to play. Not well, not consistently, but they learned. When the ball came toward their virtual paddle, they figured out that hitting it was the goal. They got better over time. They got worse when the researchers fed them bad data. They behaved, in a crude and flickering way, like a brain.

That experiment was not part of the paper we are talking about today. But it is the reason you need to know about a new field called organoid intelligence, or OI. The paper, published in Frontiers in Science by Lena Smirnova and colleagues (Smirnova et al., 2023), does not report a single experiment. It lays out a roadmap. It says: We can now grow brain organoids that contain the molecular machinery of learning and memory. We can keep them alive for months. We can record their electrical activity at high resolution. And we can connect them to sensors, actuators, and computers. The question is no longer whether a dish of neurons can learn. It is whether we can build a genuine biological computer from human brain tissue, and what that would mean.

What Exactly Is a Brain Organoid?

Let us start with the thing itself. A brain organoid is not a whole brain. It is not even a mini brain. It is a three dimensional clump of neural tissue grown from human stem cells, typically induced pluripotent stem cells, which are adult cells reprogrammed to an embryonic state. Under the right conditions, these cells differentiate into neurons and glial cells and organize themselves into structures that resemble parts of the developing human brain. They form layers. They produce synapses. They generate electrical activity.

The early versions of these organoids were small, short lived, and simple. They lacked the supporting cells that real brains rely on, particularly glial cells like astrocytes and oligodendrocytes. They did not have myelin, the fatty insulation that speeds up signal transmission. They were more like a clump of neurons than a functional piece of brain tissue.

That has changed. Smirnova and her colleagues describe a new generation of organoids that are standardized, three dimensional, and myelinated. They have high cell density and enriched levels of glial cells. They express the genes known to be critical for learning and memory, including those involved in long term potentiation, the cellular mechanism by which synapses strengthen with use. The authors write that these organoids can now "recapitulate the molecular mechanisms of learning and memory formation" (Smirnova et al., 2023). That is a strong claim. Let us look at what it is based on.

How You Grow a Learning Machine

The paper is a review and a proposal, not a single experiment, but it synthesizes a decade of incremental advances. To understand what the authors are proposing, you need to understand the three technical pillars that make OI possible.

Pillar one: Better organoids

The first problem was that early organoids died too quickly and were too variable. You cannot do reproducible science on a system that falls apart after two weeks. The solution came from microfluidic perfusion systems, which are essentially tiny channels that pump nutrients and oxygen through the organoid, mimicking the blood supply a real brain gets. These systems keep organoids alive for months and allow researchers to deliver chemical signals in precise spatial and temporal patterns. That matters because learning in real brains depends on exactly timed chemical signals, like the release of dopamine when you get a reward. A static dish cannot do that. A perfused organoid can.

The second problem was cell diversity. Real brains are not just neurons. They have astrocytes that prune synapses, microglia that clean up debris, and oligodendrocytes that wrap axons in myelin. The new organoids include all of these. They are "enriched with glial cells" (Smirnova et al., 2023), which means they have the support staff a learning system needs.

Pillar two: High resolution recording

You cannot tell if a brain organoid is learning unless you can see what it is doing. The old method was to poke it with a single electrode and listen for spikes. That is like trying to understand a symphony by putting a microphone in one corner of the concert hall. You miss everything.

The authors describe novel three dimensional microelectrode arrays that can record from hundreds of points simultaneously, in three dimensions, at high resolution. These arrays allow researchers to map the spatiotemporal patterns of electrical activity across the organoid. When a real brain learns, it rewires itself. The patterns change. With these arrays, researchers can watch the rewiring happen in real time.

Pillar three: Stimulus response training

This is where it gets strange. The authors envision connecting brain organoids to real world sensors and output devices. A sensor detects something in the environment, a temperature change or a light pulse. That signal is converted into an electrical stimulation pattern and delivered to the organoid. The organoid responds. That response is read out, processed, and used to adjust the next stimulus. This is a feedback loop. It is how you train a network.

The authors call this "organoid computer interfaces" (Smirnova et al., 2023). They are not just talking about recording from organoids. They are talking about training them.

The Pong Experiment and What It Proved

The Cortical Labs experiment I mentioned earlier is not in this paper, but it is the proof of concept that makes the Smirnova paper urgent. That team, led by Brett Kagan, grew a layer of human neurons on a high density multielectrode array. They then connected the array to a simplified version of Pong. They delivered electrical stimulation to indicate the position of the ball. The neurons could respond by generating activity that moved the paddle. Over about five minutes of simulated gameplay, the neurons learned to hit the ball more often than chance. They did not know they were playing a game. But they learned.

The Smirnova paper takes that basic idea and extrapolates it into a full research program. The Pong experiment used a flat layer of neurons. Organoids are three dimensional, which means they have more connections and more complexity. The Pong experiment used a single organoid. The authors envision networks of organoids, connected to each other and to sensory organ organoids like retinal tissue. They call this "complex, networked interfaces" (Smirnova et al., 2023). Imagine a dozen brain organoids, each trained on a different task, all talking to each other. That is the vision.

Why This Is Not Just a Weird Science Project

You might be thinking: This is cool, but why do we need it? We have computers. They are fast. They can play Pong without needing to be fed oxygen.

The authors anticipate this objection. They argue that biological computing offers three advantages that silicon cannot match.

First, energy efficiency. The human brain runs on roughly 20 watts. A supercomputer that simulates a fraction of a brain uses megawatts. The authors claim that OI based systems could allow "greater energy and data efficiency" (Smirnova et al., 2023). A dish of neurons does not need a data center. It needs glucose and oxygen.

Second, continuous learning. Most machine learning systems are trained once and then frozen. They do not adapt in real time. A biological system can learn continuously, updating its connections as new data arrives. That is what the Pong experiment showed. The neurons got better as they played.

Third, speed. This one is counterintuitive because neurons fire in milliseconds while transistors switch in nanoseconds. But the brain compensates with massive parallelism. A single thought involves billions of neurons firing in concert. The authors argue that OI based systems could allow "faster decision making" (Smirnova et al., 2023) for certain types of problems, particularly those that require pattern recognition and adaptation.

The Medical Payoff Nobody Is Talking About

The headline is biocomputing, but the real promise may be medical. The authors are clear that one of the main motivations for this work is to understand disease.

Here is the problem: You cannot study a living human brain in detail. You can take slices from cadavers, but those are dead. You can do fMRI scans, but those are blurry. You can use animal models, but a mouse brain is not a human brain. The differences matter, especially for conditions like dementia, schizophrenia, and autism, which involve uniquely human aspects of cognition.

A brain organoid is not a human brain. But it is human tissue. It carries the same genes. It forms the same types of synapses. If you take stem cells from a patient with a genetic form of Alzheimer's disease and grow them into an organoid, that organoid will develop the same pathological hallmarks, the amyloid plaques and tau tangles, that the patient's brain has. You can then test drugs on it. You can watch the disease progress in real time. You can try to reverse it.

The authors specifically mention that OI research could "help elucidate the pathophysiology of devastating developmental and degenerative diseases (such as dementia)" and "aid the identification of novel therapeutic approaches" (Smirnova et al., 2023). This is not speculative. It is already happening. Organoids are being used to study Zika virus, autism, and Alzheimer's. The OI framework adds the ability to study learning and memory themselves, not just the static structure of the tissue.

What This Research Does Not Prove

I need to be careful here. The Smirnova paper is a roadmap, not a result. It describes what is possible, not what has been achieved. The authors are explicit about this. They call it "a collaborative program to implement the vision of a multidisciplinary field" (Smirnova et al., 2023). They are not claiming to have built a biological computer. They are saying: Here is how to do it.

There are open questions that the paper does not answer, because they have not been answered yet.

Can an organoid actually learn in a meaningful sense? The Pong experiment showed that a flat layer of neurons can learn a simple motor task. But that is a long way from the kind of learning that requires attention, memory, and abstraction. An organoid has no sensory organs. It has no body. It has no way to experience the world. It can only respond to electrical stimulation. Whether that is enough to produce anything we would recognize as intelligence is unknown.

How big does an organoid need to be? The human brain has roughly 86 billion neurons. The largest organoids have maybe a few million. That is a difference of four orders of magnitude. The authors do not claim that a single organoid can replicate human cognition. They envision networks of organoids, but nobody has built such a network yet.

What about ethics? The authors devote a section to this. They call for "an embedded ethics approach" that involves all stakeholders in an iterative process (Smirnova et al., 2023). But they do not resolve the hard questions. At what point does a dish of neurons deserve moral consideration? If an organoid can learn and remember, does it suffer? These are not abstract philosophical puzzles. They are practical questions that researchers will face if the field succeeds.

The Ethical Line Nobody Wants to Draw

The ethics section of the paper is worth reading carefully. The authors are not naive. They know that growing human brain tissue that can learn raises uncomfortable questions. They propose a framework for addressing these questions as the science develops, rather than trying to settle them in advance.

But there is a tension here. The whole point of OI is to create something that has properties of a mind. If you succeed, you have created something that might be conscious, or at least sentient. If you fail, you have wasted a lot of time and money. The ethical framework the authors propose is designed to handle success. It assumes that researchers will be able to detect signs of consciousness or suffering and adjust accordingly.

That assumption may be optimistic. We do not have a reliable test for consciousness in humans, let alone in a dish of neurons. We do not know what it would look like for an organoid to be in pain. The authors acknowledge this indirectly by calling for "the development of ethical guidelines" (Smirnova et al., 2023), but they do not specify what those guidelines should be.

What This Actually Means

The Smirnova paper is not a breakthrough. It is a blueprint. But blueprints matter when they point to something real. Here is what this means in practice.

• ▸Brain organoids are no longer just models of disease. They are potential computing substrates. The authors show that organoids can be grown with the molecular machinery of learning and memory. This changes the goal from studying the brain to building with it.

• ▸The key technical bottlenecks have been identified and solutions exist. Microfluidic perfusion, 3D electrode arrays, and stimulus response training are all demonstrated technologies. The challenge now is integration, not invention.

• ▸Biological computing offers real advantages over silicon for specific tasks. Energy efficiency, continuous learning, and massive parallelism are not marketing claims. They are physical properties of neural tissue. The question is whether those advantages can be harnessed at scale.

• ▸Medical applications are closer than computing applications. Organoids are already used to study disease. Adding the ability to monitor learning and memory in real time would give researchers a powerful tool for understanding and treating cognitive disorders.

• ▸The ethical questions are not hypothetical. If the roadmap works, researchers will eventually face decisions about what to do with a system that can learn, remember, and perhaps feel. The time to think about those decisions is now, not after the first organoid shows signs of distress.

The last point is the one that stays with me. The authors of this paper are not trying to build a smarter computer. They are trying to build a system that shares the fundamental properties of the human brain. If they succeed, we will have to decide what that system is. A tool. A model. A patient. A mind. The science will force the question. The answer is up to us.