What Is a DishBrain – Live Neural Network

Most articles introducing DishBrain bury Karl Friston's name somewhere in the middle, treated as a "by the way." That order is backwards. Without Friston's Free Energy Principle, the design logic of the DishBrain experiment has no anchor, and the whole thing degrades into a flashy cell culture demo.

Friston is a co-author on the 2022 Neuron paper. Not an honorary listing. He is the source of the theoretical architecture. The core claim of the Free Energy Principle compressed into one sentence: any self-organizing system that maintains its own structure over a period of time can be formally described as doing one thing, which is minimizing variational free energy. Variational free energy is a measure of the deviation between the system's prediction of external input and the external input itself. This "deviation" has a name in information theory: surprisal, the negative log probability. The more an external input departs from the system's "expectation," the higher the surprisal, the higher the free energy.

This framework has been debated for nearly two decades at the intersection of neuroscience and theoretical physics. The controversy is intense. Supporters consider it a meta-theory unifying all cognitive phenomena. Opponents consider it unfalsifiable, and once you reach that point it is no longer science but metaphysics. This debate can be temporarily shelved in the DishBrain context, because the experimental design only uses one very narrow corollary of the Free Energy Principle, narrow enough to almost sidestep those grand philosophical disputes.

The narrow corollary works like this. The Cortical Labs team designed two feedback conditions. Condition one: when the neuron network on the chip lets the paddle miss the ball, the system delivers random noise electrical stimulation to the neurons. Random noise, high entropy, completely unpredictable, corresponding to a high free energy state. Condition two: after missing, the delivered stimulation is structured, patterned, low entropy, predictable. Result: under condition one, the network's hitting performance improved significantly. Under condition two, improvement was negligible.

Core Experimental Finding

This set of controls is the highest-value data in the paper. It eliminates an alternative explanation: "any feedback makes neurons more active and therefore perform better." No. Predictable feedback did not work. Only unpredictable feedback drove the reorganization of connection topology. This precisely matches the prediction of the Free Energy Principle: the system tends to move away from high free energy states, and if a certain behavioral pattern can reduce the surprisal of subsequent inputs, the system will spontaneously drift toward that pattern. In the specific environment of Pong, catching the ball means subsequent input is structured low-entropy signal; missing the ball means subsequent input is random noise. The neuron network does not "know" it is playing a game. It does not need to know. It is physically, locally, through synaptic plasticity rules, fleeing from that high-surprisal state. The escape path happens to manifest at the behavioral level as "improved hit rate."

It is worth staying here for a moment, because what this mechanism implies runs deeper than it appears on the surface.

Nobody programmed a reward function. In the standard paradigm of reinforcement learning, the researcher explicitly defines a reward signal: do it right, gain points; do it wrong, lose points; then the algorithm optimizes the cumulative expected reward through gradient descent. DishBrain has none of this. The random noise stimulation is not a "punishment signal" in the traditional sense. It carries no semantic information of "you did it wrong." It is simply a thermodynamic discomfort: high-entropy input is a form of tension for a system maintaining low-entropy internal states. The system adjusts itself to relieve that tension. No teacher, no gradient, no loss function. Only the physical tendency of living tissue.

If you push this line of thinking one step further, what it implies will make people working in deep learning uncomfortable: perhaps the most primitive adaptive behavior requires no form of "design" whatsoever. No artificial specification of optimization targets, no algorithmic scaffold of backpropagation. Just a lump of self-organizing biological matter placed into an environment that has regularities, and free energy minimization as a physical process unfolds on its own. You do not need to tell water to flow downhill. Likewise, you do not need to tell the neurons on DishBrain to hit the ball.

There is another theoretical layer in the paper that has completely vanished from public reporting. DishBrain involves not just passive prediction (reducing the deviation between internal model and external input), but also active inference. Active inference is the action-side extension of the Free Energy Principle. A system can reduce free energy through two pathways: either modify its internal model to better match the input (this is perception), or act to change the external world to better match its own expectations (this is action). Both pathways are operating in DishBrain. Neuronal firing patterns drive the paddle movement, paddle movement changes ball trajectory, ball trajectory changes the electrical stimulation fed into the chip in the next frame. The neurons are actively reshaping their own sensory input stream. In the specialist literature, DishBrain has been discussed as the first in-vitro physical realization of the active inference framework. The weight of that positioning far exceeds "cells playing Pong."

What DishBrain actually is, physically

With the theoretical foundation laid in the previous section, the most basic question can now be addressed.

DishBrain is a product of Cortical Labs, a company in Melbourne. Approximately 800,000 neurons are cultured on a high-density multi-electrode array chip. The neurons come from two sources: one batch differentiated from human induced pluripotent stem cells, another from mouse primary cortical tissue. Electrodes on the chip record extracellular field potentials from the neurons and can also deliver electrical current stimulation. Through this bidirectional interface, the neurons are connected to a computer running Pong. Ball position is encoded as stimulation frequency across different electrodes; the firing density of neuron populations in designated regions is decoded as paddle movement direction. Closed loop.

The closed loop is the identity of this system. Remove it, and what remains is an ordinary in-vitro neuron culture on an MEA chip, countless copies of which have sat in electrophysiology labs over the past thirty years. Add the closed loop, and this mass of cells acquires a "body" (a paddle a few pixels wide in Pong) and an "environment" (the ball's trajectory). Embodiment.

This is also where DishBrain differs most from cerebral organoids. Organoids are three-dimensional cell aggregates induced from stem cells, capable of self-organizing into cortex-like layered structures, carrying significant value for developmental biology. Their limitation is that they are entirely sealed off. They are observed, stained, sectioned, but never form any interactive loop with the outside world. DishBrain's neuron count is smaller than a decent organoid's, structural complexity far lower, three-dimensional organization essentially nonexistent. It wins on one count: it has been embedded in an environment, its activity changes that environment, and the environment's changes in turn alter it.

Conditions on the MEA chip are far messier than the paper abstract suggests

After neurons differentiated from iPSCs are seeded onto the chip surface, the first two to three days see mass death. Those that ultimately survive, adhere near electrodes, and establish functional synaptic connections with neighbors account for a fraction of the initial seeding volume. This means the network performing "learning" on each DishBrain is a survivor population that has passed through a round of brutal physical selection. The survivors are biased in adhesion capacity, tendency toward synaptogenesis, electrical excitability thresholds, and the direction of these biases varies across batches. This is a primary source of why repeated experiments show large variance, and why each DishBrain has a different "personality." People who do cell culture are unfazed by this kind of variability. People who do computation find it absurd: it is equivalent to the weight matrix being randomly swapped out every time you deploy a model.

Then there is the migration issue. Neurons do not sit still. During culture they move, extend new processes, form connections with new neighbors, break connections with old ones, die. Electrode positions on the MEA chip are fixed. Neuron positions are not. So the input-output mapping of the system is in continuous drift. In any conventional digital computing system, this kind of ongoing drift at the hardware level means system failure. DishBrain's "hardware" keeps deforming, and task performance still improves incrementally. This fact, viewed from the lens of computing engineering, is harder to digest than the learning itself. How does it work? Distributed coding and redundant connectivity. Information does not reside at any single node but is diffused across population patterns. A few nodes lost, the statistical collective behavior of remaining nodes redistributes the load. Network degradation is gradual rather than sudden fracture. The connection topology that neurons on DishBrain spontaneously form exhibits small-world network characteristics: large numbers of local short-range connections plus a small number of long-range bridges. No external guidance. They grew into this shape on their own.

The paper ran both human neurons and mouse neurons. Human neurons learned faster, reached a higher plateau. This is more interesting than it first appears.

If the difference is because human neurons have higher synaptic density or larger network scale, then it is purely a matter of quantity. There is a possibility pointing to a matter of quality: human cortical pyramidal neurons have dendritic branching complexity far exceeding that of rodent equivalents. Dendrites are not wires. Synaptic inputs at different locations along a single dendrite integrate in nonlinear fashion, and calcium dynamics at dendritic spines provide each synaptic site with an independent local computational workspace. The effective computational dimensionality of a single human neuron is not one; it is one multiplied by the computational contribution of each of the thousands of dendritic spines on it.

If this multiplier effect plays a role in the human-mouse performance gap on DishBrain, what it implies is this: morphological complexity of neurons is an independent variable for network computational capacity, one that cannot be fully compensated for by scaling up numbers. If that proposition holds, the impact on artificial neural network design philosophy is substantial, because artificial "neurons" are all zero-dimensional point nodes with no spatial structure.

Then there is the rest-period phenomenon. During intervals with no external stimulation input, the neuron network is not silent. It produces spontaneous rhythmic population firing, with frequency characteristics bearing some analogy to slow-wave oscillations in the living brain. After a rest period, task performance sometimes improves upon resumption. Drawing a direct equivalence to the memory consolidation function of sleep is certainly premature at this scale of in-vitro culture. There is a narrower and more solid question that can be asked here: is spontaneous dynamical reorganization during offline periods an intrinsic component of synaptic plasticity? If so, the fact that virtually all current deep learning training is continuous with no "rest cycles" becomes a concrete, laboratory-testable research hypothesis.

Energy efficiency is not a footnote

Landauer's Principle: erasing one bit of information at room temperature requires a minimum of kT ln2 energy, approximately 2.85×10⁻²¹ joules. The energy cost of a single biological synaptic transmission is estimated at 10⁻¹⁴ to 10⁻¹⁵ joules. Several orders of magnitude above the Landauer limit, but already far below any silicon-based computing unit. Every synaptic adjustment on DishBrain occurs at energy scales extremely close to the thermodynamic floor.

This is often tacked on at the end of an article as a piece of trivia. It should not be. Because it means that when someone tries to simulate the full dynamics of a biological neural network at DishBrain's scale using GPU clusters, the source of the energy efficiency gap is not at the engineering level. Engineering problems can be optimized. Physical constraints cannot. Biological neurons work in the analog domain, with continuous ion concentration gradients and membrane potential changes carrying signals, no analog-to-digital conversion overhead. Silicon systems work in the digital domain, with quantization error and clock cycle overhead at every step. The difference between the two computing substrates is structural at the physics level.

Who is on the track

Cortical Labs raised roughly ten million dollars in a seed round in 2021. In 2024 they received a contract from the Australian Department of Defence. What the military is interested in has nothing to do with "petri dish playing games." The direction is: the ability of biological neurons to process ambiguous signals in noisy environments at extremely low power. Unmanned systems, sensor networks, signal classification in electronic warfare environments. The core constraint in these scenarios is power budget, not raw compute. An edge computing module deployed on a drone may have only a few watts of available power, and making silicon chips do complex pattern recognition within that power envelope is extremely difficult. The energy efficiency advantage of biological neurons lands precisely in this sweet spot.

Others are on the same track. Koniku in the United States has been pursuing commercialization of biological neurons for scent detection and security since before Cortical Labs. FinalSpark in Switzerland launched Neuroplatform in 2024, a remotely accessible biological processor platform allowing researchers to manipulate brain organoids on MEAs over the internet. Thomas Hartung's team at Johns Hopkins published an organoid intelligence roadmap paper in 2023, defining this as an independent research direction.

Some of Cortical Labs' patent filings contain the phrase "hybrid neuro-silicon architecture." If this direction develops, it means embedding small amounts of living neurons into neuromorphic chips as core computational units, with silicon circuitry handling peripheral input-output management. This is exactly reversed from the route represented by Intel's Loihi and IBM's TrueNorth. Those chips use silicon components to mimic biological spike-timing behavior, pursuing "simulating carbon with silicon." Cortical Labs' direction is that carbon does not need to be simulated; just use it directly. Whether the two routes will eventually converge is unknown.

Ethics

The human neurons used in DishBrain come from iPSCs, which come from anonymous donors. Under current theoretical frameworks, in-vitro two-dimensional neuron cultures at this scale do not possess consciousness or sentience. The Φ-value requirements of Integrated Information Theory, the broadcast mechanism requirements of Global Workspace Theory, the meta-representational requirements of higher-order theories: none are met. This assessment is clear at present.

If culture scale grows from hundreds of thousands to tens of millions, with three-dimensionalization and multi-region differentiation, the assessment may no longer be clear. In 2024 the Australian National Health and Medical Research Council began revising ethical guidelines covering "synthetic biological intelligent agents."

Legal Void

iPSC donors have relinquished rights to derivatives. Suppose one day a successor system to DishBrain develops a discernible information-processing "personality" at the statistical-feature level (not consciousness, just identifiable uniqueness): that personality would belong to no one under law. Current legal frameworks contain no category for handling this.

The question that keeps being dodged

The AI field has a long-running implicit assumption: substrate independence. Intelligence is an algorithmic-level phenomenon; whether it runs on silicon, carbon, or beer cans makes no difference. This assumption is convenient because it allows researchers to focus exclusively on algorithm design without worrying about physical implementation.

DishBrain's data creates a nuisance for this assumption that is not easy to dismiss. Same Pong task: DQN requires millions of frames of training data; DishBrain shows adaptive behavior in approximately five minutes. One can argue that the task conditions are not perfectly comparable, that DQN starts from raw pixel input while DishBrain's input has already been encoded through a preprocessing layer, and that argument has merit. One can argue that DishBrain demonstrates only rudimentary adaptation on an extremely simple task, light-years from general intelligence, and that argument also has merit.

After conceding all of these, a hard question remains in place: do certain computational properties of carbon-based neural tissue (analog-domain continuous signal processing, thermal fluctuations providing natural stochastic exploration, molecular-scale massive parallelism, nonlinear spatial computation in dendrites) provide a physics-level efficiency advantage for certain types of learning tasks, an advantage that algorithmic improvement cannot fully compensate for?

DishBrain has not answered this question. The scale and complexity of the experiment are insufficient to provide an answer. What it has done is make this question operable in a laboratory. Previously this question could only be discussed philosophically in papers. Now it has a physical counterpart: living neurons on a chip that can be subjected to different closed-loop tasks, different feedback conditions, different culture parameters, and then quantitatively compared against silicon in adaptation speed and energy consumption on equivalent tasks.

If substrate does exert an irreducible efficiency impact on certain types of intelligence, then "bigger models plus more compute" is not the only route, and on certain task dimensions it may not even be the most economical one. This is not saying scaling laws are about to be overturned. Scaling laws remain valid within their applicable range. What is being said here is a possibility: there exists another route, starting from an entirely different physical substrate, where the cost structure for reaching certain specific capabilities is fundamentally different.

That is what DishBrain is. 800,000 neurons cultured on an MEA chip, connected to a computer running Pong, exhibiting measurable adaptive behavior driven by free energy minimization as a physical process. An experimental result. A starting line for a new track. An operable experimental entry point for asking whether computing substrate participates in determining the properties of intelligence. These three identities hold simultaneously, increase in importance in that order, and decrease in public awareness in that same order.