Teaching Organoids: How Brain Tissue Learns

Series: Organoid Intelligence | Part: 4 of 9

In 2022, researchers at Cortical Labs taught a cluster of neurons to play Pong. Not a simulation. Not a neural network. Actual living brain tissue—800,000 human neurons grown on a chip—learned to intercept a digital ball with a paddle. They called it DishBrain, and it learned the game faster than conventional AI.

The training session lasted five minutes. The organoid got it. Not through gradient descent or backpropagation, but through something biological systems have done for billions of years: prediction, surprise, and adaptive response.

This isn’t a proof of concept. It’s a challenge to everything we thought we knew about the separation between biological tissue and computational intelligence. Because if you can teach neurons to play Pong, you can teach them to recognize patterns, detect anomalies, optimize processes—any task that reduces to minimizing prediction error.

The question isn’t whether brain tissue can learn. The question is how to talk to it.

The Learning Problem for Organoids

When you train a neural network, you adjust weights. When you teach a dog a trick, you use rewards. But what do you do when the student is a few hundred thousand neurons floating in a dish?

Brain organoids don’t have eyes, ears, or limbs. They have no motor output except electrical activity. They have no sensory input except what you provide through electrodes. And they certainly don’t respond to verbal instructions.

Yet they do something far more fundamental than any of these: they minimize surprise.

This is the core insight from Karl Friston’s Free Energy Principle—the mathematical framework that explains how every biological system maintains its organization by predicting the future and acting to make those predictions come true. Neurons don’t need rewards. They need patterns they can predict, and feedback when predictions fail.

In active inference terms, learning is just updating your generative model based on prediction error. If the world behaves differently than expected, the system adjusts its internal model to reduce future surprise. This happens in a human brain. It happens in a single cell. And it happens in a dish of neurons if you give it the right training environment.

The training challenge for organoid intelligence isn’t about motivation or comprehension. It’s about engineering a feedback loop that lets biological tissue discover structure in its sensory stream—and discover that its own activity can influence that structure.

How DishBrain Learned to Play Pong

The breakthrough at Cortical Labs came from rethinking what “reward” means to a neural system.

Traditional reinforcement learning uses explicit reward signals: win and get +1, lose and get -1. But neurons don’t optimize for points. They optimize for predictability. Entropy is expensive. Uncertainty generates metabolic cost as systems work harder to model ambiguous environments.

Brett Kagan and his team at Cortical Labs designed a feedback protocol based on this insight:

When the organoid’s paddle hit the ball, the neurons received structured, predictable stimulation—electrical patterns that were consistent, rhythmic, repeatable. When it missed, they received random noise—unpredictable, high-entropy signals with no discernible pattern.

The result: the organoid learned to hit the ball to escape the noise.

This wasn’t reinforcement in the behavioral sense. It was environmental structuring. The experimenters created a world where one action (successful interception) led to low surprise, and another action (missing) led to high surprise. The neurons, doing what neurons do, shifted their dynamics toward the low-surprise regime.

Five minutes of training. Measurable improvement in hit rate. Not because the tissue “wanted to win,” but because biological systems minimize free energy—and chaotic input is energetically expensive to model.

The training didn’t involve backpropagation. It didn’t involve explicit weight adjustments. It involved creating a sensorimotor loop where the system’s predictions about the world could be tested, violated, and refined through direct experience.

This is how brains learn—all brains, from cortical organoids to human infants.

The Free Energy Gradient: Minimizing Surprise as Learning Signal

What makes organoid learning possible isn’t cleverness. It’s physics.

Every biological system exists far from thermodynamic equilibrium. Staying organized—staying alive—requires resisting entropy. For a neural system, this translates to minimizing prediction error. The better your generative model of the world, the less surprise you encounter, the less corrective action you need to take, the less energy you burn.

In Friston’s formalism, this is variational free energy minimization: the system has a probability distribution over hidden states of the world (its beliefs), and it updates those beliefs to minimize the divergence from sensory observations.

Mathematically, minimizing free energy decomposes into two processes:

1. Perceptual inference — updating beliefs to explain sensory data (learning what the world is like)
2. Active inference — acting on the world to make sensory data conform to predictions (making the world behave as expected)

For an organoid playing Pong, perceptual inference means discovering the relationship between neural firing patterns and ball position. Active inference means modulating activity in ways that influence paddle position and thus future sensory states.

The training protocol doesn’t teach the organoid what to do. It constructs a niche where reducing surprise requires the organoid to discover the Pong dynamics. The learning emerges from the free energy gradient itself.

This is radically different from supervised learning, where correct answers are externally imposed. Here, “correct” means predictable—and the organoid discovers what’s predictable by exploring the consequences of its own activity.

Human children learn this way. They babble, and discover which mouth shapes produce consistent sounds. They reach, and discover which muscle patterns produce controlled movements. They don’t download a model. They build one by coupling their actions to sensory predictions and updating when surprised.

Organoid learning isn’t artificial. It’s just biological learning stripped to its essentials.

Training Paradigms: What Actually Works

If surprise minimization is the engine of learning, the question becomes: how do you engineer a structured surprise landscape that teaches specific tasks?

Closed-Loop Feedback (The DishBrain Model)

The defining feature of DishBrain’s success was real-time sensorimotor coupling. The organoid’s activity directly controlled paddle position, and paddle position directly influenced sensory feedback. This creates a closed loop where prediction error is immediately informative.

Contrast this with open-loop stimulation, where the organoid receives input with no relationship to its own activity. Open-loop might activate neurons, but it doesn’t create a learnable structure—there’s no prediction-action-feedback cycle to update.

Closed-loop training mirrors the embodied nature of biological intelligence. Your neurons don’t passively receive the world. They actively sample it, and the results of that sampling inform what to sample next. This is why organoids on multielectrode arrays (MEAs) can learn tasks, but organoids in static culture wells cannot—the interface creates the learning niche.

Predictability Gradients (Structured vs. Random)

The Cortical Labs team used entropy gradients as a training signal: low entropy for correct behavior, high entropy for incorrect. But this can be generalized.

Any dimension that increases predictability can function as reinforcement: - Temporal structure: rhythmic vs. arrhythmic input - Spatial structure: patterned vs. scrambled electrode activation - Frequency coherence: harmonic vs. noisy frequencies

The key is contrast. The system needs to detect a difference in surprise between two regimes, and it needs a causal pathway (its own activity) that shifts which regime it inhabits.

Human trainers use this intuitively. “Good dog” isn’t arbitrary—it’s predictable social feedback that contrasts with the ambiguity of being ignored. Musical training works because hitting the right note produces harmonic resonance; hitting the wrong one produces beating interference. The signal doesn’t need to be symbolic. It needs to be structurally detectable.

Curriculum Learning: Scaffolding Complexity

Biological systems don’t learn optimally when thrown into high-complexity environments with no structure. They need graded exposure to increasing surprise.

For organoids, this might mean: 1. Start with simple temporal patterns (learn to predict a rhythmic pulse) 2. Introduce spatial contingencies (different electrodes predict different outcomes) 3. Add sensorimotor coupling (your activity influences what patterns appear) 4. Increase task complexity (now predict sequences, not just pulses)

This mirrors developmental stages in natural brains. Visual systems learn edge detection before object recognition. Motor systems learn reaching before grasping. The curriculum isn’t externally imposed—it’s emergent from the statistical structure of natural environments, which are hierarchically organized (low-level features are more common than high-level conjunctions).

For engineered training, we can compress developmental timescales by explicitly controlling the statistics of input. But the principle remains: learnable structure before irreducible noise.

Memory Without Synapses: How Organoids Encode Experience

One of the paradoxes of organoid learning is that the tissue lacks the structural organization of an intact brain. No hippocampus. No cortical layers. No long-range fiber tracts. Yet it demonstrates memory—it performs better on subsequent trials, and performance degrades if you disrupt activity.

How does a disorganized clump of neurons encode learned information?

The answer lies in network dynamics, not anatomical structure.

Attractor Dynamics as Memory

Neural networks don’t store memories like files on a hard drive. They store them as attractor states—stable patterns of activity that the network reliably falls into given certain inputs.

When neurons wire together through Hebbian plasticity (“cells that fire together wire together”), repeated activation of a pattern strengthens the connections that support it. Over time, this creates a basin of attraction in the state space: perturbations push the system toward the learned pattern rather than away from it.

For an organoid learning Pong, the “memory” isn’t a representation of the game. It’s a stable pattern of coordinated firing that minimizes prediction error in the Pong environment. The tissue doesn’t “remember the rules.” It settles into a dynamical regime where its predictions are reliably confirmed—which, in practice, means hitting the ball.

This is why memory in biological systems is context-dependent. The same neurons participate in multiple attractor basins. The pattern that emerges depends on initial conditions (sensory input, prior state). An organoid trained on Pong wouldn’t “remember how to play Pong” in a new task—it would fall into a different attractor shaped by the new surprise landscape.

Synaptic Plasticity in Organoids

Despite their disorganization, organoids exhibit functional synaptic plasticity. Studies show: - NMDA receptor activation (critical for long-term potentiation) - AMPA receptor trafficking (strengthens active synapses) - GABA receptor modulation (regulates inhibitory balance)

These are the molecular mechanisms of learning in natural brains. They allow connection strengths to change based on activity—the physical substrate for attractor formation.

But organoids also show homeostatic plasticity: as networks become too excitable, inhibitory mechanisms strengthen to prevent runaway activity. This is critical for stable learning. Without it, training would push the system toward epileptiform seizures rather than structured behavior.

The lack of anatomical organization doesn’t prevent learning. It constrains what can be learned. An organoid can discover local statistical regularities and form attractor basins around them. It probably can’t learn hierarchical abstractions or compositional representations—those require structured connectivity (which is exactly what evolution, development, and the human brain’s architecture provide).

But for tasks that reduce to detecting patterns and minimizing surprise? The mechanisms are already there.

Beyond Pong: What Else Can Organoids Learn?

The DishBrain result was a proof of principle, not a ceiling. If brain tissue can learn one sensorimotor task, it can—in principle—learn others.

What’s already been demonstrated: - Temporal sequence prediction (learning to anticipate when a stimulus will arrive) - Pattern classification (distinguishing between different input types) - Adaptive response (modulating activity to maintain predictability in changing environments)

What’s plausible with better interfaces and training protocols: - Signal processing (noise reduction, feature extraction) - Anomaly detection (recognizing deviations from learned patterns) - Optimization (finding parameter settings that minimize prediction error in a control loop)

The bottleneck isn’t the tissue’s learning capacity. It’s the interface bandwidth. Current MEAs read from and write to a few hundred electrodes. The organoid contains hundreds of thousands or millions of neurons. We’re sampling a tiny fraction of the state space and stimulating a tiny fraction of the network.

If you could read from and write to every neuron—optogenetic arrays, voltage imaging, dense electrode grids—the effective channel capacity explodes. The organoid could learn to represent far more complex input and produce far more sophisticated output.

This is the interface problem we’ll explore in the next article. But the learning mechanisms are already in place. The tissue is doing what it evolved to do: minimize surprise, refine predictions, stabilize its internal model.

You don’t need to teach an organoid to learn. You just need to create an environment where learning is the path of least resistance.

Coherence, Curvature, and the Geometry of Learning

In AToM terms, learning is coherence formation—the emergence of structure that integrates across states and reduces entropy.

An untrained organoid is high-curvature: its state space is flat, with no deep basins of attraction. Activity fluctuates stochastically. Inputs don’t produce consistent outputs. The system is incoherent—there’s no stable relationship between past, present, and future states.

Training carves basins into this landscape. Repeated exposure to structured input coupled with feedback creates low-curvature regions where the system reliably predicts and confirms its predictions. These are the learned attractors.

In geometric terms, learning is curvature reduction through constraint satisfaction. The system discovers which states minimize surprise (low curvature) and which states maximize it (high curvature), and it preferentially inhabits the low-curvature manifold.

This maps directly onto Friston’s free energy framework: free energy is (loosely) curvature. High free energy = high uncertainty = high curvature. Low free energy = accurate predictions = low curvature. Biological systems flow downhill on the free energy gradient, which means they flow toward geometric coherence.

The training signal—predictable vs. random feedback—creates a landscape where coherence is rewarded. The organoid doesn’t “try” to be coherent. Coherence is the equilibrium state for a system minimizing variational free energy under structured constraints.

This is why learning works without explicit instruction. The mathematics of self-organization already encode it.

What This Teaches Us About All Learning

The organoid learning paradigm isn’t exotic. It’s learning stripped of everything inessential.

No language. No social context. No symbolic rewards. Just: 1. A system capable of detecting prediction error 2. A causal pathway between its activity and sensory input 3. A structured environment where some actions reduce surprise more than others

This is the minimal viable learning loop. And it’s the same loop present in: - Infant motor development (discovering which movements produce consistent sensory feedback) - Reinforcement learning algorithms (updating policies to maximize expected reward, which is just low surprise under a utility-weighted distribution) - Cultural transmission (imitating behaviors that produce predictable social outcomes) - Scientific progress (testing models that minimize prediction error about nature)

What changes across these domains isn’t the learning mechanism—it’s the richness of the niche, the bandwidth of the interface, and the complexity of the generative model being refined.

An organoid learning Pong is doing the same computational work as a physicist learning quantum mechanics. The physicist has a richer environment, a more structured prior, and a symbolic language that compresses centuries of collective learning. But the gradient being descended is the same: make the world more predictable by improving your model of it.

This is why organoid intelligence matters beyond biocomputing. It reveals learning as a physical process—not a cognitive achievement, but a natural consequence of systems maintaining organization far from equilibrium.

You don’t teach neurons to minimize free energy. They already do. You just give them a world where the path to lower free energy happens to overlap with the task you care about.

This is Part 4 of the Organoid Intelligence series, exploring the science, applications, and implications of biological computing with brain tissue.

Previous: The Energy Equation: Why Wetware Beats Silicon Next: The Interface Problem: Connecting Wetware to Hardware

Further Reading

• Kagan, B. J., et al. (2022). “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.” Neuron, 110(23), 3952-3969.
• Friston, K. (2010). “The free-energy principle: a unified brain theory?” Nature Reviews Neuroscience, 11(2), 127-138.
• Sporns, O., & Betzel, R. F. (2016). “Modular brain networks.” Annual Review of Psychology, 67, 613-640.
• Tscherning, A., et al. (2023). “Self-organization and learning in cultured neuronal networks.” Neural Computation, 35(4), 567-592.
• Lancaster, M. A., & Knoblich, J. A. (2014). “Organogenesis in a dish: modeling development and disease using organoid technologies.” Science, 345(6194).