The thesis · Primer

Biocomputing: the case for running code on living neurons

Biocomputing performs computation on living neural tissue, usually cultured brain organoids wired to a silicon microelectrode array that both stimulates the cells and records their firing. The substrate is biological; the interface is electronic.

For seventy years, "compute" has meant pushing electrons through fixed silicon logic. Biocomputing makes a different bet: that a self-organizing, three-dimensional culture of neurons, which rewires itself as it works, can be a more efficient substrate for the messy, adaptive problems that silicon handles badly. That bet is no longer purely theoretical. In 2022 a dish of cultured human and mouse neurons learned to play a simulated game of Pong inside five minutes of closed-loop training.1 The interesting question is not whether living tissue can compute. It is what it is good for, what it costs, and what we owe the tissue.

Macro view of a translucent brain organoid in culture, with faint cyan neural filaments against a black background.
A cultured brain organoid, the self-organizing tissue at the center of organoid intelligence research. Imaging is illustrative.

Why would anyone compute on neurons?

Two properties that silicon spends enormous effort to imitate come free in biological tissue: extreme energy economy per operation, and learning that happens in the hardware itself rather than in a separate training phase. The adult human brain runs on roughly 20 watts.6 A single large-model training run in a modern data center consumes that much in a fraction of a second. Neurons achieve their economy because computation, memory, and communication happen in the same place, in analog, without shuttling data across a bus.

Order-of-magnitude power comparison Bar chart comparing approximate continuous power for the human brain, a single cultured-neuron rack, and a GPU training rack, on a shared scale. Approximate continuous draw Human brain (tissue only) 20 W Cultured-neuron rack (incl. life support) 950 W GPU training rack 30000 W
Order-of-magnitude continuous power, plotted on a shared scale. The living tissue itself draws microwatts; the figure for a cultured-neuron rack is dominated by life support and data acquisition, not by the neurons. See the honest energy accounting below.

The second property is plasticity. A deep network learns by backpropagation during a distinct training phase, then freezes. Biological tissue never stops adapting: the same synapses that carry a signal also change their strength in response to it, continuously and locally. That is why a culture can be shaped by feedback in minutes rather than retrained over hours.

How does a microelectrode array talk to living tissue?

The bridge between a computer and a neuron is electrical, but the two sides speak different languages. Silicon moves information as electrons through wires. Neurons move it as ions, mainly sodium, potassium, calcium, and chloride, across the cell membrane through voltage-gated channels. The membrane voltage that results is described by the Hodgkin-Huxley formalism, the quantitative model of the action potential that still anchors the field.7

Cross-section of a microelectrode array interfacing cultured neural tissue A planar electrode array at the base, an electrical double layer at each electrode, neural tissue above with neurons and synapses, and bidirectional arrows showing stimulation downward and recording upward. Neural tissue (organoid) CMOS electrode array soma axon synapse double layer stimulate (uA) record (uV)
A high-density microelectrode array records extracellular voltage and delivers microampere stimulation through thousands of electrodes. At each electrode, an electrical double layer couples electronic charge to ionic charge in the medium. Stimulation depolarizes nearby membrane; recording captures the resulting field potentials.

A modern high-density array packs thousands of electrodes at pitches under 20 micrometers, close enough to address tissue at near-single-cell resolution. Each electrode does two jobs. It records the tiny extracellular voltages that leak from firing neurons, on the order of tens of microvolts, and it injects charge-balanced microampere pulses to drive firing. The conversion between the electronic and ionic worlds happens in a nanometers-thin electrical double layer at the metal surface, which behaves like a capacitor. Get that coupling wrong and you either fail to hear the cells or you electrolyze the medium and kill them.

How does the tissue actually learn?

Learning in a neural culture is not programmed. It is coaxed. The dominant local rule is spike-timing-dependent plasticity: when a presynaptic neuron reliably fires just before the neuron it connects to, that synapse strengthens; when it fires just after, the synapse weakens.5 The window that decides which way a synapse moves is only tens of milliseconds wide.

Spike-timing-dependent plasticity window A curve plotting synaptic weight change against the timing difference between pre- and post-synaptic spikes: when the presynaptic spike precedes the postsynaptic spike the connection strengthens; when it follows, the connection weakens. t(post) - t(pre) delta weight potentiation pre before post depression pre after post
Spike-timing-dependent plasticity. The sign and size of the synaptic weight change depend on the relative timing of pre- and post-synaptic spikes within a narrow window. Causal ordering (pre before post) potentiates; the reverse depresses.

To turn that local rule into useful behavior, researchers close a loop around the tissue. The leading account of why this works draws on the free energy principle, the idea that a self-organizing system acts to minimize the surprise of its sensory input.4 In practice, a culture is given structured, predictable stimulation when it produces the desired output and unstructured, high-entropy stimulation when it does not. The network reorganizes to make its world predictable again, which is to say, to perform the task.

Closed-loop training cycle for a biological neural network A five-stage loop: encode input as stimulation, deliver to the tissue, record the response, decode it, then deliver predictable feedback on success or unstructured high-entropy stimulation on error, returning to the start. 1. Encode task input 2. Stimulate spatial uA pattern 3. Record evoked spikes 4. Decode read out action 5. Feedback predictable = reward error -> unstructured stimulation (entropy)
The closed-loop training cycle. Task input is encoded as a spatial stimulation pattern, delivered to the tissue, and the evoked response is recorded and decoded into an action. Predictable feedback rewards a correct action; unstructured stimulation follows an error. This is the loop the DishBrain system used to train cultures on Pong.
The result that changed the conversation

In the 2022 DishBrain study, cultures of roughly 800,000 neurons learned to keep a ball in play in a simulated Pong environment, improving within a single session of closed-loop feedback.1 The headline is not that the dish was good at Pong. It is that goal-directed behavior emerged from a local timing rule and a feedback signal, with no model, no weights, and no training corpus.

The honest energy accounting

The most repeated claim in this field is also the most misleading: that biocomputing is radically more energy efficient than silicon. The tissue is. The system is not, at least not yet. A living neuron signals at the microwatt scale, but you cannot run one without a rack of equipment whose entire purpose is to keep it alive and to listen to it.

The neurons draw microwatts. The incubator, the pumps, and the digitizers draw the other 99.99 percent.

Independent operational estimates put a single cultured-neuron rack in the range of roughly 850 to 1000 watts under continuous operation. That figure is an engineering estimate, not a measured tissue property, and it is dominated by support systems rather than by computation: holding the culture at 37 degrees Celsius in a 5 percent carbon dioxide atmosphere, pumping oxygenated nutrient medium, and digitizing thousands of channels at 20 to 30 kilohertz each through low-noise amplifiers and acquisition servers. Anyone quoting the microwatt figure as the cost of biocomputing is quoting the cost of the neuron and hiding the cost of the laboratory around it.

Where a cultured-neuron rack spends its watts Bar chart breaking down the approximate power draw of a cultured-neuron rack across incubation, perfusion, data acquisition, shielding, and the neural tissue itself. Watts (of ~950 W total) Incubation (heat, CO2) 380 W Perfusion + microfluidics 210 W Amplifiers + ADC + servers 300 W Shielding + power conditioning 60 W Neural tissue 1 W
An illustrative breakdown of where a roughly 950 watt rack budget goes. The neural tissue is the thin bar at the bottom. The proportions, not the exact watts, are the point.

Power figures on this page are order-of-magnitude operational estimates and are labeled as such. Where a number can be cited to a primary source, it is.

Biocomputing vs silicon neuromorphic hardware

Both fields are reacting to the same problem, the energy and latency cost of moving data between memory and a processor, and both look to the brain for an answer. They diverge on how literally to take the analogy. Neuromorphic chips emulate spiking neurons in CMOS; biocomputing uses the neurons themselves. The trade is determinism and manufacturability against native plasticity and three-dimensional density.

Silicon neuromorphic vs biological substrate
PropertySilicon neuromorphicBiological substrate
Signal carrierElectrons in CMOSIons across membranes
LearningMostly off-chip, then frozenContinuous, in the hardware
Structure2D planar fabricationSelf-organized 3D tissue
ReproducibilityHigh, deterministicLow, each culture differs
LifespanYearsWeeks to months
MaturityCommercial silicon shippingEarly research platforms

The honest summary: silicon neuromorphic is an engineering discipline shipping products; biocomputing is a science establishing whether the substrate is worth the trouble. They are not yet competitors. A fuller, dated comparison of the available platforms lives on the platform comparison for the Cortical Labs CL1, where the specifications are tracked against primary sources.

What biocomputing is not

The term collects confusion, so it is worth drawing borders. Biocomputing in the sense used here is neither DNA computing, which stores and processes information in nucleic acid sequences rather than in firing cells, nor quantum computing, with which it shares nothing but a reputation for being futuristic. It is also not the cyberpunk "wetware" of fiction. The technical sense of wetware is narrow and unromantic: living neural tissue used as a computational medium. Anyone promising a brain in a jar that thinks like you is selling something. What exists is a culture that can be shaped, by feedback, to perform narrow tasks, and that remains hard to keep alive for more than a few months.

The open problems, stated plainly

Scaling is unsolved: cultures are small, every one is different, and they degrade in weeks to months. Readout is noisy and decoding is bespoke. And the field carries an ethical weight that silicon does not. The organoid intelligence research agenda treats these as first-class questions rather than footnotes.2

Unsettled, on purpose

If tissue derived from a human donor can learn, what was consented to, and by whom? Could a sufficiently complex culture have morally relevant experience, and how would we know? These are open questions, not settled positions, and this network presents them as such across every pillar. They are not a reason to stop the work; they are a reason to do it in the open.

Frequently asked questions

What is biocomputing in one sentence?

Biocomputing is computation performed on living neural tissue interfaced with a silicon microelectrode array that both stimulates and records the cells.

Is biocomputing the same as organoid intelligence?

Closely related. Organoid intelligence is the specific research program of using brain organoids as the computing substrate and studying what they can learn. Biocomputing is the broader category that includes it.

Has a living culture ever actually computed something?

Yes. In 2022, cultured neurons in the DishBrain system learned to play a simulated game of Pong through closed-loop electrical feedback, improving within a single session.

Is biocomputing really more energy efficient than silicon?

The tissue is, by orders of magnitude. The full system is not yet, because incubation, perfusion, and data acquisition dominate a rack budget of roughly 850 to 1000 watts. The microwatt figure describes the neuron, not the laboratory.

How does a neural culture learn without being programmed?

Through spike-timing-dependent plasticity, a local rule that strengthens or weakens each synapse based on the relative timing of spikes, combined with structured feedback delivered in a closed loop.

What are the main limits today?

Small culture size, poor reproducibility between cultures, lifespans of weeks to months, noisy readout, and unresolved bioethical questions about donor consent and moral status.

References

  1. Kagan BJ, et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron. 2022;110(23):3952-3969. doi:10.1016/j.neuron.2022.09.001. Accessed 2026-06-12.
  2. Smirnova L, et al. Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science. 2023;1:1017235. doi:10.3389/fsci.2023.1017235. Accessed 2026-06-12.
  3. Friston K. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. 2010;11(2):127-138. doi:10.1038/nrn2787. Accessed 2026-06-12.
  4. Bi GQ, Poo MM. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience. 1998;18(24):10464-10472. doi:10.1523/JNEUROSCI.18-24-10464.1998. Accessed 2026-06-12.
  5. Attwell D, Laughlin SB. An energy budget for signaling in the grey matter of the brain. Journal of Cerebral Blood Flow and Metabolism. 2001;21(10):1133-1145. doi:10.1097/00004647-200110000-00001. Accessed 2026-06-12.
  6. Hodgkin AL, Huxley AF. A quantitative description of membrane current and its application to conduction and excitation in nerve. Journal of Physiology. 1952;117(4):500-544. doi:10.1113/jphysiol.1952.sp004764. Accessed 2026-06-12.