---
title: "Organoid Array Computing: The Design Space of Organoid Intelligence"
description: "Jenn Leung, Chloe Loewith and Ivar Frisch survey what can currently be done with brain organoids wired to electrodes, then map the design space that opens if the present limits are overcome."
type: "reading-notes"
authors:
  - "Jenn Leung"
  - "Chloe Loewith"
  - "Ivar Frisch"
published: "2025-05-10"
original_url: "https://coginfra.antikythera.org/"
doi: "10.1162/ANTI.5CZG"
canonical_url: "https://antikythera.wiki/work/journal/coginfra/organoid-array-computing"
md_url: "https://antikythera.wiki/md/work/journal/coginfra/organoid-array-computing"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# Organoid Array Computing: The Design Space of Organoid Intelligence

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Jenn Leung, Chloe Loewith, Ivar Frisch
- **Published:** 2025-05-10
- **Kind:** Studio projects · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.5CZG
- **Original:** https://coginfra.antikythera.org/
- **Length:** 8.4k words
- **Part of:** [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra)
- **This page:** https://antikythera.wiki/work/journal/coginfra/organoid-array-computing
- **Notes generated:** 2026-08-23

## Summary

[Jenn Leung](https://antikythera.wiki/people/jenn-leung), [Chloe Loewith](https://antikythera.wiki/people/chloe-loewith) and [Ivar Frisch](https://antikythera.wiki/people/ivar-frisch) survey what can currently be done with brain organoids wired to electrodes, then map the design space that opens if the present limits are overcome. Their organizing claim is that cultured human neural tissue is a computer in the literal sense — a physical object implementing a computable function — with a property no artificial network has: its architecture changes while it works. From there they propose a layered speculative program. The first layer, organoid array computing, uses dozens of small organoids together as the untrained reservoir of a reservoir-computing system. The second gives individual organoids sensory specializations so an array can handle several modalities at once. The third imagines learned states being passed between organoid generations. Running through all of it is a wider thesis they call polycomputation: the electrodes are not the only thing computing, since the culture medium, the channel geometry of the housing and the tissue itself all shape the result.

## The argument

The paper starts from substrate. New materials have always reshaped what computing is, and stem cell research has produced one worth taking seriously: brain organoids, three-dimensional cultures grown from human pluripotent stem cells that develop functioning neural networks. If a computer is any physical object implementing a computable function, then brains and the organoids modeling them are computers, and the move is to stop treating computers as externalized brains and start treating cultured brains as computers. The practical motive is the hardware bottleneck facing current AI, which makes an alternative to neuromorphic silicon worth examining.

What organoids offer that artificial networks do not is architectural change. A network's topology is fixed once designed — layers, neurons and connections stay put, as true of continual and self-supervised learning as of ordinary supervised training, since none of those paradigms lets a model rewrite its own structure. Neurons in tissue form new connections as they learn. The design space that follows is an attempt to make that difference useful rather than merely notable.

The existing evidence is thin but real, and the authors report it without inflating it. Cortical Labs' DishBrain grew neurons on a multielectrode array and embodied them in a simulated game of *Pong*, reinforcing successful returns with a predictable feedback signal and answering misses with an unpredictable one; the system learned and was competitive with deep reinforcement learning agents. Alysson Muotri's group reintroduced the archaic NOVA1 variant into human stem cells to grow Neanderthal-like organoids with longer growth periods, a popcorn-like shape and fewer cortical connections, then connected them to robots. Brainoware, from Guo's team at Indiana University Bloomington, classified 240 clips of spoken Japanese vowels at 78 percent accuracy after training, below what conventional networks achieve. FinalSpark and Emulate already sell such devices.

Between that evidence and the proposal sit three obstacles. Organoids have no vasculature, so they develop necrotic centers and cannot grow large; the workarounds are engraftment into animal hosts, which raises the chimera problem, and bioprinting vascular structures with hydrogel bioinks. Microfluidic platforms are treated as static life support when they could be an active layer. And organoids are usually cultured alone, so the interregional traffic that makes a brain a brain is missing.

The proposal answers each. Microfluidics becomes a computational layer, via [Leroy Cronin](https://antikythera.wiki/people/leroy-cronin)'s chemputation and the claim that if liquids can compute, chemical gradients, media composition and dopaminergic stimulation are parameters of the machine rather than housekeeping. Interorganoid communication becomes assembloids — fused or cocultured organoids modeling traffic between brain regions — and, more strikingly, axon bundles grown through silicon elastomer microchannels, where two organoids perform a handshake of synchronized bursting and the geometry of the device alone directs how tissue develops. On that base sit the three scaffolding layers, of which only the first is worked out.

The tension the paper does not resolve sits inside that first layer. Reservoir computing is attractive here because it is substrate-independent and leaves the reservoir untrained, which is exactly what lets unengineerable tissue be used as it comes. But the untrained reservoir is precisely where the plasticity that motivated the whole argument gets set aside. The paper's case for organoids is that their topology develops; its concrete architecture asks them to sit still as a nonlinear medium while a linear classifier does the learning.

## Section by section

### Introduction

States the position: organoids interfaced with AI through multielectrode arrays instantiate new modes of computing, and the design space runs from organoids as software emerging from a biological substrate to organoids as hardware supporting higher-order processes. Polycomputation is introduced here as the frame — hardware, software, tissue and chemical substrates as one computational assembly coupling in-vitro biology to in-silico systems.

### Background on Organoids

Organoids are stem-cell-derived three-dimensional cultures modeling an organ's structure and function, beginning with Hans Clevers's cultures grown from patients' intestinal stem cells. Human brain organoids mimic aspects of cortical architecture and are used to study development, model disease and test drugs where working on living brains is not possible. They remain, in the phrase the authors quote, a minimal working model of some of the circuitry of a functioning brain.

The subsection on organoid intelligence defines the interface as tripartite: the organoid as the network, the multielectrode array for bidirectional stimulation and recording, with shank electrodes reaching deeper layers and mesh electrodes providing flexible contact, and the microfluidic platform supplying culture medium in place of a blood supply. The comparison with artificial networks that follows is the paper's clearest technical claim, and it is about topology rather than performance.

### Current Applications of Organoid Intelligence

Five cases, in ascending order of speculation: DishBrain, the Neanderthal organoids, Brainoware's speech classification, commercial bioprocessors, and FinalSpark's live monitoring platform, which the authors read as a first step toward integrating biological systems with digital infrastructure. The section is descriptive and cautious about what has actually been demonstrated.

### The Design Space of Organoid Intelligence

Explicitly speculative, framed as a way to identify current limits and imagine past them rather than to advocate a path. The theoretical warrant is Friston's free energy principle: if a brain works to minimize the gap between prediction and sensory input, organoids growing structures like optic cups can be read as demanding richer input, which the authors take as self-driven complexity in the substrate. That is a large inference from a developmental fact, and the paper does not defend it further.

Vascularization is the hard limit, with three responses — coculture with vascular cells, engraftment into rodent hosts, and bioprinting scaffolds that cells grow over — and the ethical note that chimeras may develop morally relevant qualities. The microfluidics subsection makes the polycomputational case concretely: culture media are shown to change neuronal maturation and plasticity, dopaminergic stimulation entrains activity elsewhere in an assembloid, and the authors ask for platforms with bidirectional traffic between electrical signals and chemical composition. The subsection on interorganoid communication is where the empirical results are most suggestive, since the axon-bundle work shows spatial constraint alone shaping development, and connected organoids showing higher short-term plasticity and more complex signaling than isolated ones.

### Scaffolding for Organoid Intelligence

Three layers. Layer one, organoid array computing, is specified: dozens of miniature organoids as nodes in a multichambered microfluidic device, each on its own high-density electrode array, used as the reservoir in a reservoir-computing system for speech recognition, with audio converted to spatiotemporal stimulation patterns and a linear classifier decoding the response. The design borrows DishBrain's finding that rate and place coding through eight stimulation electrodes matched or beat pixel-based input to reinforcement learning agents. The claimed gains are parallel task pooling, assembloid learning across coordinated organoids, and eventually personalized models grown from an individual's own cells.

Layer two proposes sensory specialization, with different organoids in an array guided toward different modalities by electrical, optogenetic or chemical stimulation and their outputs integrated, on the analogy of cortical regions. Layer three, intergenerational memory, imagines transferring a trained state from one organoid to its successors so each generation does not start over. Layers two and three are sketches; neither has a mechanism attached, and the paper says so.

### Implications

The ethical section is short and firm in both directions. Halting the field would be harmful given its medical value, and the emergence of morally relevant qualities requires criteria worked out in advance rather than after the fact. The authors endorse the letter "A Response to Claims of Emergent Intelligence and Sentience in a Dish," which criticizes premature use of "sentience" and "intelligence" for neurons in a dish, and apply it to their own field's language.

### Conclusion

The offering is a taxonomy rather than a result: parameters for organoid assemblies, devices, microfluidics, bioprinting and culture media that experimenters could work through. The closing claim is the one worth carrying: unlike systems with fixed hardware, an organoid array could grow, adapt and reorganize its physical structure in response to what it is asked to compute.

## Key concepts

- **[Organoid array computing](https://antikythera.wiki/terms/organoid-array-computing)** — dozens of small human brain organoids, each on its own high-density electrode array and linked by a multichambered microfluidic device, used together as the untrained reservoir of a reservoir-computing system, with only a linear readout trained.
- **[Polycomputation](https://antikythera.wiki/terms/polycomputation)** — the claim that a single assemblage computes at several scales and in several media at once, so that culture medium, channel geometry and tissue are all computational layers rather than support for one.
- **Organoid neural network** — the network that forms in living tissue as it learns, distinguished from an artificial network by its ability to change its own topology rather than only its weights.
- **Assembloid** — two or more region-specific organoids fused, cocultured or joined by grown axon bundles, so that interregional signaling can develop; connected pairs show synchronized bursting and greater short-term plasticity than isolated organoids.
- **Chemputation** — [Leroy Cronin](https://antikythera.wiki/people/leroy-cronin)'s programmable control of chemical synthesis, borrowed here to argue that the liquid environment of an organoid participates in computation rather than merely sustaining it.
- **Scaffolding** — the paper's method of layering: each speculative capability presupposes the one below, from chemical microenvironment through spatial conditioning and tissue assembly to multimodal processing and inherited memory.

## Connections

Within [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) this is the first paper under Organs Without Bodies, whose framing asks where information processing ends and cognition begins and treats neural tissue as a material with cognitive properties available for engineering. Its companion, [Cognition With and Beyond the Brain](https://antikythera.wiki/work/journal/coginfra/cognition-with-and-beyond-the-brain), works the same theme from philosophy rather than the bench, distributing cognition across assemblages of bodies and technical systems; both agree that cognition is not brain-shaped and disagree about where to look for it, one culturing the tissue and the other decentralizing it.

[Minimum Viable Interiority](https://antikythera.wiki/work/journal/coginfra/minimum-viable-interiority) is the sharpest internal counterpart. It asks what it would take to engineer an individual with genuine interiority, and defines interiority functionally by nested closure rather than by consciousness. That is exactly the question the ethical section of this paper defers, and the two together suggest the field needs criteria for morally relevant capacities that do not depend on anyone settling what sentience is.

[Xenophylum](https://antikythera.wiki/work/journal/coginfra/xenophylum) pushes the other way: where Leung, Loewith and Frisch take the most biomimetic route available by growing actual human cortex, that paper argues for engineering forms with no biological precedent. The pair marks a fault line in the compendium about whether biology is a template to copy or a constraint to escape. [Chronoseed](https://antikythera.wiki/work/journal/coginfra/chronoseed) shares the ambition to compute in biological matter but wants DNA for its inertness rather than tissue for its plasticity.

Outside the compendium, [Substrates Unbound](https://antikythera.wiki/work/journal/substrates) supplies the theory this paper works within, arguing that material substrates shape which functions are possible instead of passively carrying them and treating cultured neurons on electrode arrays as the clearest case; the argument for microfluidics as computation applies that position directly. [Modes of Cognition](https://antikythera.wiki/work/journal/modesofcognition) provides the corpus frame for cognition that is neither conscious nor human. [Organoid Nation](https://antikythera.wiki/work/journal/hemstacks/organoidnation) takes the same technology into scenario fiction, imagining a state built on biocomputing — the political dimension this paper's ethics section touches only as research oversight.


## Terms used

- [Organoid Array Computing](https://antikythera.wiki/terms/organoid-array-computing)
- [Polycomputation](https://antikythera.wiki/terms/polycomputation)
- [Synthetic Biological Intelligence](https://antikythera.wiki/terms/synthetic-biological-intelligence)