Daily analysis of organoid intelligence research · Cited to primary sources · Vendor neutral
OrganoidIntellect

Organoid intelligence · Biological computing

Organoid Intellect

A publication about organoid intelligence: running computation on living neural tissue. We read the primary literature as it appears, grants, preprints, trials and papers, and work out what each result changes for the field. Vendor neutral, cited to primary sources, written for people who need the technical substance rather than the press release.

Daily analysis · 277 analyses published · How this publication works

A translucent brain organoid resting on a dark microelectrode array, with glowing cyan filaments radiating outward across the electrode grid.
Cultured neural tissue interfaced with a microelectrode array, the arrangement at the centre of organoid intelligence research. Illustration.

What organoid intelligence means

Organoid intelligence (OI) is biological computing carried out on 3D cultures of human brain cells, wired to electronics through a microelectrode array or comparable interface. The organoid supplies the computational substrate; the interface supplies the read and the write. The term was introduced in Smirnova et al., Frontiers in Science (2023), which set out the research programme the field has organised itself around since.

It is worth being precise about what the term does and does not claim. Organoid intelligence names a research direction, not a working technology: the demonstrations on the record are narrow, the cultures are small and short lived, and the comparisons to silicon that circulate in coverage are frequently made on terms that flatter the biology. The field is also distinct from neuromorphic computing, which imitates neural structure in silicon rather than using living cells, and from wetware in its science fiction sense. Our primer on biocomputing works through the mechanism in full; the analysis stream tracks each new result as it lands.

Latest analysis

October 27, 2026 The gate that protects a policy's memory
DS-VLA gives each action neuron two dendritic branches with heterogeneous learned decay and a neuron-wise inhibitory gate that decides how much new evidence may enter, then reads continuous actions from the membrane state. Under a standardized action-corruption protocol it retains 95.4 percent of nominal success where four strong baselines collapse. The mechanism is one living neurons already possess, and it reframes what an organoid's dendrites are for.
October 27, 2026 Searching for architectures silicon would never design
OpenArchEvo lets a large language model evolve executable architecture code for spiking sequence models, using a three-view candidate fingerprint to predict performance and novelty before spending 132 V100 GPU-days on training. The discovered NeuroGate beats an ANN DeltaNet on WikiText-103 perplexity while firing half as often. The finding that matters for organoid intelligence is the measured failure of architectures to transfer between substrates.
 October 27, 2026 A reservoir that writes its own trap
Kozhevin and colleagues show, within the standard two-component condensate-reservoir model, that a polariton condensate can dig a hole in its own excitonic reservoir and sit in the resulting self-induced trap, a bistable state that survives below the condensation threshold for nanoseconds before ejecting as a bullet. The mechanism is a design principle for any physical reservoir, and a competitive warning for living ones.
 October 26, 2026 The molecular write channel matures while the neural one lags
Amin and colleagues train variational synthesis models, which control chemical DNA synthesis reactions to manufacture quadrillions of designed sequences, by gradient descent in continuous space followed by post-training quantization onto the finite mixture catalog a synthesizer can actually run. Across enzyme, peptide, antibody, and regulatory DNA design tasks the quantized models advance the quality-diversity Pareto frontier, and a physical library run at 1.6 nM yield with 28 million sequencing reads confirms quality holds in vitro.
 October 26, 2026 Channel kinetics set how far a living code looks ahead
Shaikh and Assisi built a continuous attractor network of conductance-based stellate cells and interneurons that reproduces grid cell path integration and explains predictive grid firing with two dissociable mechanisms. Post-inhibitory rebound through HCN channels shifts the decoded position forward about 5 percent of a grid scale, while asymmetry in inhibitory connectivity shifts it up to 25 percent, and the model predicts the predictive reach should vary monotonically along the dorsoventral axis with HCN time constants.
 October 26, 2026 Sparse execution on an FPGA exposes idle power as the real energy story
SymbolicLight V2 runs a 194M-parameter hybrid spiking language model on an Alveo U50C FPGA and a ROCK 5T ARM board. Selective weight-row gathering and partial KV cache loading raise decode throughput 35.5% and cut gross card energy 27.6%, but an idle-split measurement attributes 82.8% of energy to the loaded-idle share, and the checkpoint itself fails the authors' prespecified quality criteria.
 October 24, 2026 Balanced spiking networks as low-power random sources
A University of Calgary team builds a pseudo-random number generator from a chaotic balanced spiking network running on 5 milliwatts of FPGA power. The result doubles as a warning about treating biological noise as free randomness.

All analysis →

Start here

Standing explainers that do not go stale. Read the primer first if the field is new to you; everything in the analysis stream assumes it.

The primer Biocomputing: the case for running code on living neurons
What it means to compute on living neural tissue, why anyone would, how a microelectrode array talks to neurons, how the tissue learns, and an honest accounting of the energy cost.
12 min read Evidence What living tissue has actually been made to compute
The demonstrations on the record, and what each one does and does not show.
 Comparison Biological computing against silicon neuromorphic hardware
Where each substrate wins, where the comparison is usually made dishonestly.
 Terminology Wetware, the technical definition
The precise meaning, with the science fiction usage stripped out.

Sections

Analysis
Daily readings of new research, each one asking what the work changes for computation on living tissue.
 Biocomputing
The mechanism and the evidence: how computation is performed on neural tissue, and what has been demonstrated.
 Wetware
The substrate itself: definitions, materials, energetics, and the interface between cells and electronics.
 Library
Every standing publication in one place, organized by category.

How we work

Every analysis names its primary source and links to it. We separate what a study demonstrates from what it asserts, we say plainly when we could not obtain a full text, and we do not publish a number we cannot attribute. Analyses are interpretations of other people's research, not peer review, and they are dated so you can weigh them against what was known at the time. The full method, including how pieces are selected and produced, is on the about page.