Diff — Biological Computing

{{infobox

| title = Biological Computing

| subject = Wetware — computers made of living neurons, brain organoids and DNA, and the science of teaching life to compute

| span = 1906 Cajal's neuron doctrine → 2025 the first commercial biological computer

| first product = Cortical Labs CL1 — March 2025, ~800,000 human neurons on silicon

| rentable today = FinalSpark Neuroplatform — four brain organoids via Python API, ~$500/month

| why it matters = A human brain does ~10¹⁶ operations/second on ~20 watts; an exascale supercomputer needs ~20 megawatts

| open question = When does a dish of neurons become *someone* — and who decides?

| related = [[History of Physics]] · [[History of Mathematics]] · [[Neolithic Period]]

| image = bio-dishbrain-cells.jpg

| caption = Living cortical neurons growing across a multi-electrode array — the cells that learned to play Pong in the 2022 DishBrain experiment. The bright dots are individual neurons; the wiring underneath is silicon (Kagan et al., *Neuron*, CC BY 4.0).

}}

Somewhere in a Melbourne laboratory, on a silicon chip the size of a postage stamp, there is a small civilization of **human neurons**. They were once ordinary blood cells, donated by adults, reprogrammed backwards into stem cells, and then coaxed forward into brain cells. Now they live in a nutrient bath, kept warm and oxygenated, their chatter picked up by a grid of electrodes — and they are *working*. They play games. They respond to patterns. If you are a paying customer of Cortical Labs, they may even run your code.

The pitch is not science fiction. It is arithmetic. A human brain performs on the order of **10¹⁶ operations per second** — comparable to an exascale supercomputer — while running on about **20 watts**, roughly a light bulb. The machine in this article's own category, the exascale **Frontier** supercomputer, draws about **22 megawatts** — a million times more, to do a similar amount of work. Evolution spent four billion years optimizing wet, squishy, extremely efficient computation. Silicon, by comparison, is sixty years old and enormously wasteful. The question the field is asking is whether we can borrow what biology already knows.

This is the story of how a neuron became a computer — from the drawings of a Spanish anatomist to a dish of brain cells learning Pong — and of what might happen when the computers are alive.

# The discovery of the neuron

Physics began, arguably, with a man and a bathtub. Biological computing begins with two men who hated each other sharing a Nobel Prize.

In the 1870s **[[Camillo Golgi]]** invented a staining technique — the *reazione nera*, "black reaction" — that could dye a few whole neurons at random in a slice of brain tissue, rendering them visible under a microscope for the first time. **Santiago Ramón y Cajal** took Golgi's stain, improved it, and spent years drawing what he saw: not a continuous web, as Golgi believed, but *separate cells*, each touching others across tiny gaps, each apparently a one-way device taking in signals at one end and passing them out the other. Cajal's drawings — of Purkinje cells like elaborate corals, of pyramidal neurons like root systems — established the **neuron doctrine**: the brain is made of discrete cells, and information flows through them in a direction.[[cite: Cajal won the 1906 Nobel Prize in Physiology or Medicine *with* Golgi — who used his own Nobel lecture to argue against Cajal's conclusion. Golgi was wrong.]]

[[File:bio-cajal.jpg|right|300px|Cajal's drawing of a Purkinje cell from the human cerebellum — one neuron, hand-drawn in ink around 1899. The neuron doctrine started as pictures like this (public domain).]]

If a neuron is a one-way switch, then a brain is a machine — 86 billion switches wired together — and the question of whether neurons could *compute* becomes a matter of engineering rather than philosophy. Everything in this article follows from Cajal's ink.

# The equation of the spike

What kind of signal does a neuron actually send? In the 1930s two Cambridge physiologists, **Alan Hodgkin** and **Andrew Huxley**, found the ideal experimental animal: the squid, whose motor neurons include a **giant axon** a full millimetre thick — thick enough to thread electrodes inside. By the early 1950s they had measured exactly how the neuron's membrane lets sodium and potassium ions rush in and out in sequence, generating the electrical spike — the **action potential** — that travels down the axon to the next cell.

[[File:bio-axon.jpg|right|300px|Propagation of an action potential along a giant axon — the ionic wave Hodgkin and Huxley turned into mathematics (Qniemiec, CC BY-SA 4.0).]]

```fact The signal in one sentence

A neuron "fires" when its membrane voltage crosses a threshold: sodium channels open, ions flood in, voltage spikes, potassium channels restore the rest. All-or-nothing, ~1–100 times per second — and the *timing* and *pattern* of spikes, not their size, is the information.

```

# Life on the wire

By the 1970s, researchers could keep neurons alive outside a body — **cultured** in a dish, on nutrient medium — and the neurons did something remarkable without any instruction: they grew **axons and dendrites**, found each other, and formed networks. By the 1990s engineers could grow those cultures directly on **multi-electrode arrays** (MEAs) — chips patterned with dozens or hundreds of tiny electrodes — and for the first time read and write the traffic of a living network in real time.

[[File:bio-neurons.jpg|right|340px|A rat cortical neuron culture, deconvolved z-stack — the kind of network that grows, unbidden, when neurons are plated on a chip. Within days they are wiring themselves together and bursting with activity (ZEISS Microscopy, CC BY 2.0).]]

Criticality turned out to be the answer to "why is a brain so good at learning?" — a network at the edge of chaos has maximum dynamic range, maximum information capacity, and maximum sensitivity to small changes. The chaos is not a bug in the dish. **The chaos is the substrate.** Everything biological computing does depends on it.

# The robot with a rat brain

The Australian art collective SymbioticA took the idea further and stranger: **MEART — The Semi Living Artist** (2003–2007) grew rat neurons on an MEA in Atlanta and connected them, over the internet, to a robotic drawing arm exhibited in Perth. A dish of brain cells in Georgia made drawings in Australia. It was art, but it was also a working demonstration: cultured neurons + electrodes + a closed loop = a device that learns to act in the world.

# The other biocomputer

[[File:bio-dna.jpg|right|320px|DNA origami — DNA folded into designed shapes at nanometre scale. The same molecule stores data at ~455 exabytes per gram and can be used as computation substrate (Origamimonkey, CC BY-SA 3.0).]]

DNA computing never became practical for general-purpose work — the molecules are slow, error-prone, and single-use — but it founded **molecular computing** as a real discipline, gave us **DNA origami** (folding DNA into programmable nanostructures), and pointed to DNA's actual superpower: **storage**. A gram of DNA can theoretically hold ~455 exabytes — more than every data center on Earth, in a test tube. Microsoft and others have already stored files this way. The molecule that stores genomes can store anything.

# DishBrain: the dish that learned Pong

In October 2022, a small Melbourne startup called **Cortical Labs**, working with UCL's **Karl Friston** — the most-cited neuroscientist alive and architect of the **free energy principle** — published a paper that made the field real. They had built **DishBrain**: ~800,000 neurons, grown from human stem cells (and, in separate cultures, mouse brains), plated on a high-density MEA and connected to a computer running **Pong**.

[[File:bio-dishbrain-schematic.jpg|right|360px|The DishBrain closed loop — ball position in as stimulation, paddle movement out as activity, predictable vs. unpredictable feedback as reward and punishment (Kagan et al., *Neuron*, CC BY 4.0).]]

```fact How DishBrain's training actually works — the free energy principle

Karl Friston's **free energy principle** says any self-organizing system separated from its environment acts to minimize *surprise* — the gap between what it expects and what it senses. A dish of neurons is exactly such a system. Feed it predictable patterns when it does what you want and chaotic noise when it doesn't, and the network — seeking a predictable world — physically rewires itself to produce the rewarded behavior. Nobody writes the program. The neurons write it, out of plasticity.

```

The paper's title used the word *sentience* — deliberately, in a narrow technical sense meaning "responsive to sensory input in a goal-directed way," not consciousness — and it set off a firestorm of coverage anyway. The authors spent months clarifying that the dish was not a mind; it was a system that *learned*, which is a different and already extraordinary claim. But the word got out, and the question it raised — *how would we know?* — turned out to be the field's defining problem. (See **The question nobody can answer**, below.)

# The chaotic heart — why you can't program a dish

This is the part of the technology that makes every other kind of computing look tame, and it is worth understanding precisely.

A silicon neural network is a set of numbers. To train it, an algorithm — backpropagation — walks backward through the numbers and nudges each weight by a calculated amount. You always know exactly what changed and why. A **biological neural network** (BNN) has no weights to adjust. It has ~800,000 living cells, each making thousands of synapses, each synapse continuously strengthening or weakening according to local rules the cells evolved on their own.

The local rules are the deep part:

- **Hebbian plasticity** — "cells that fire together, wire together." If neuron A reliably helps fire neuron B, the synapse between them strengthens. The network literally carves the patterns it repeats into physical structure.

- **Spike-timing-dependent plasticity (STDP)** — the timing matters to milliseconds. If A fires *just before* B, the link strengthens (A predicted B); if A fires just *after*, it weakens. A dish of neurons is constantly, automatically learning **causality**.

- **Criticality** — as Beggs and Plenz showed, the network self-organizes to the edge of chaos. Too ordered and it can only repeat; too chaotic and nothing propagates. Living networks hold the knife-edge where information flows best — and they hold it on their own.

# CL1 — the product

In March 2025, at Mobile World Congress in Barcelona, Cortical Labs launched the **CL1** — the first commercial biological computer. The machine is shoebox-sized and self-contained: a chip carrying ~800,000 human neurons reprogrammed from adult donor blood or skin cells, a life-support system that feeds them, removes waste and keeps them alive for up to six months, a touchscreen, and no external computer required.[[cite: Cortical Labs CL1 launch, MWC Barcelona, March 2025; ABC Science; IEEE Spectrum]]

Karl Friston — who co-authored the DishBrain paper — offered the field's most quoted assessment: the CL1 "could be regarded as the first commercially available biomimetic computer, the ultimate in neuromorphic computing that uses real neurons… However, the real gift of this technology is not to computer science. Rather, it's an enabling technology that allows scientists to perform experiments on a little synthetic brain."[[cite: Karl Friston, quoted in IEEE Spectrum, March 2025]]

# The Swiss rent-a-brain

The details are genuinely strange. The organoids live over **100 days** in automated microfluidic incubators that feed them without human intervention; researchers stimulate and record through an eight-electrode interface per organoid; and FinalSpark's training trick is chemical — neurons are bathed in **dopamine** when they produce the desired response, mimicking the brain's own reward system. Over three years the platform has hosted **more than 1,000 organoids** and collected **18 terabytes** of neural data. Co-founder Fred Jordan states the goal plainly: "artificial intelligence for 100,000 times less energy."[[cite: Jordan, F. et al., *Frontiers in Artificial Intelligence* 7 (2024); FinalSpark statements]]

```fact The energy math, honestly stated

The frequently cited claim is that biological computing could be up to a **million times more energy-efficient** than digital hardware — because neurons compute at ~20W where silicon needs megawatts. The honest version: that number compares *training* large AI models to a brain's learning; nobody has yet built a biocomputer that actually out-performs a GPU at a useful task. The efficiency is real; the lead is not — yet.

```

# Organoid intelligence

[[File:bio-oi-mea.jpg|right|360px|The OI roadmap's vision — organoids interfaced with 3D microelectrode arrays for high-resolution stimulation and recording, "an EEG cap for organoids" (Smirnova, Caffo, Hartung et al., CC BY 4.0).]]

Hartung's team already built the prototype of that interface — a flexible shell densely covered in electrodes that wraps around an organoid like, in his words, "a kind of an EEG cap for organoids." The roadmap is explicit that the goal is not a brain in a vat but a computational substrate: organoids as biological hardware, trained by biofeedback and big data, connected to real-world sensors — and, eventually, to each other.

# What has already happened

Wetware is not a proposal; it has a track record. In rough order of what the dishes have already done:

- **Learned a game in five minutes** — DishBrain's Pong result, with human and mouse cultures, closed-loop only, controls confirming it was the structured feedback doing the work

- **Controlled a robot** — Potter's hybrot, two decades before the commercial era

- **Made art across an ocean** — MEART's Atlanta neurons drawing in Perth

- **Survived as a product** — the CL1 shipping at $35k with an SDK and a cloud service

- **Been rented by the month** — FinalSpark's platform, 1,000+ organoids, 18TB of data

- **Published an ethics framework** — the OI field's own roadmap includes the governance questions *before* the technology outruns them

# What might happen

Here the honest flag: everything below this line is extrapolation, and the field itself marks the difference carefully.

**Near-term (real roadmaps):** drug screening on human neurons instead of animals — the CL1's first pitch is "animal-free testing" with more relevant data; disease modeling for epilepsy, Parkinson's and autism on patient-derived neurons; robotics with wetware that adapts faster than any trained network; edge computing where milliwatts matter.

**Medium-term (extrapolation):** hybrid systems where organoids handle the messy, adaptive parts of a problem and silicon handles the bookkeeping; organoid brains learning in silico before being transplanted; the Johns Hopkins scaling goal of 10 million cells per unit; wetware data centers that grow their own processors.

**The arithmetic nobody can ignore:** if biological computing reaches even a fraction of its theoretical efficiency, the economics of computation change. A million-fold improvement in joules per operation is the difference between a megawatt data center and a light bulb — and between AI as a planetary energy drain and AI as a living process that costs less than the food it eats.

**And the uncomfortable one:** if these systems ever cross some threshold of complexity — the organoids keep getting bigger, the loops keep getting richer, the training keeps getting better — at some point we may have to ask whether the dish is not just computing but *experiencing*. Nobody currently knows where that line is. The DishBrain authors said "sentience" in a narrow sense and spent a year explaining they meant it; the field has not stopped using the word carefully since.

# The question nobody can answer

[[File:bio-oi-ethics.jpg|right|340px|The OI consortium's own preliminary ethics framework — published alongside the science, not after it: embedded ethics as part of the roadmap (Smirnova, Caffo, Hartung et al., CC BY 4.0).]]

The deeper problem is the one that makes this the strangest technology in the world: **we do not have a test for what a dish of neurons is.** Consciousness has no meter. We can measure spikes, and bursts, and learning — but whether there is anything it is *like* to be 800,000 cells in a nutrient bath is, for now, a question without an instrument. The dishes keep learning. The question keeps waiting.

# Quiz

- [ ] 1 — Who drew the neuron doctrine, and what prize did he share (with a man who disagreed with him)?

- [ ] 2 — What two-part system — predictable reward vs. unpredictable noise — taught DishBrain's neurons to play Pong?

- [ ] 3 — Roughly how many watts does a rack of CL1s draw, vs. the megawatts of an exascale computer?

- [ ] 4 — What does FinalSpark use as the "reward" that trains its organoids?

- [ ] 5 — What mathematical state — between order and chaos — do living neural cultures self-organize to?

```hide Show answers — click to check yourself

1 — **Santiago Ramón y Cajal**, who shared the 1906 Nobel with Golgi — who argued *against* the neuron doctrine in his own lecture.

2 — **The free energy principle** — predictable stimulation as reward, chaotic noise as punishment; the neurons minimize surprise by learning.

3 — **~850–1,000 watts** a rack — vs. ~22 megawatts for Frontier. A million-fold difference in joules per operation, roughly.

4 — **Dopamine** — the brain's own reward chemical, bathed over the organoids when they respond correctly.

5 — **Criticality** — the edge-of-chaos state where avalanches follow a power law; Beggs & Plenz measured it in 2003.

```

# Notes

1. The word *sentience* in the DishBrain paper's title is used in the narrow technical sense — responsive to sensory input — not consciousness; the authors spent a year clarifying it, and it's flagged here deliberately.

2. All CL1 specifications are Cortical Labs' published figures (price, neuron count, power, 6-month viability, biOS, SDK) — reported as the company's claims.

3. The "million times more efficient" figure is a field-level extrapolation comparing training energy, not a benchmark of deployed systems.

4. The DishBrain paper's rodent data used mouse cultures; the human-iPSC cultures were separate experiments within the same paper.

5. Adleman's DNA computation solved a seven-node Hamiltonian path — correct but small; it founded a field rather than solved a practical problem.

6. The energy comparison (brain ~20W vs Frontier ~22MW) is rough and deliberately so — the honest range is the order of magnitude, not the decimal.

# References

- Hodgkin & Huxley, "A quantitative description of membrane current…," *J. Physiol.* 117 (1952).

- Cortical Labs CL1 launch coverage — ABC Science, IEEE Spectrum, DCD (March 2025); FinalSpark Neuroplatform — finalspark.com/neuroplatform.

[[Category:Science]]

[[Category:Computing]]

[[Category:Emerging technology]]