The Brain Crumb That Felt the Alphabet

Three lumps of lab-grown neurons, a robot finger, and the strangest Braille lesson ever given. Nobody learned to read. Something more interesting happened instead.

A Lesson Delivered Over the Internet

Picture a robot arm in an engineering lab in Bristol, England. Strapped to its wrist is a fingertip of soft silicone, studded with a grid of tiny black dots and watched from the inside by a camera. The arm lowers the fingertip onto a 3D-printed plate, 23 by 14 centimeters, that holds all 26 letters of the Braille alphabet in three tidy rows. Then it slides, the way your index finger slides across a page, and the raised dots of the letter push into the silicone. ¹

Now picture the student. It is about half a millimeter across. It has no finger, no eye, no skull, no memory of yesterday, and no idea that the alphabet exists. It is a clump of human forebrain tissue grown from stem cells, sitting on eight steel electrodes in a perfused chamber operated by a Swiss company called FinalSpark, and the Bristol team never touched it. They reached it the way you reach your bank: through a Python interface, over the internet. ¹

That is the setup of a study published October 16 in Frontiers in Artificial Intelligence by Tianyi Liu, Hemma Philamore, and Benjamin Ward-Cherrier of the University of Bristol’s School of Engineering Mathematics and Technology. ¹ The headline result is that when the “feel” of each Braille letter was translated into electrical pulses and fed to these organoids, the cells answered with patterns distinct enough that a computer could tell which letter had been sent. ¹²

I want to tell you exactly how that worked, because the honest version of this story is stranger and more useful than the version that will be passed around with the words “mini brains learn to read.” They did not learn to read. The authors say so, plainly and more than once. What they did instead raises a question I find genuinely thrilling: what is a lump of living tissue good for, if not thinking?

First, Let Us Retire the Phrase “Mini Brain”

The press likes “mini brains.” I like accuracy, and also a good sense of scale, so here is the arithmetic.

The organoids in this study were cultured to roughly 500 micrometers in diameter and contain mature neurons along with astrocytes and oligodendrocytes, the support cells that keep neurons fed and insulated. ¹ The University of Bristol describes them as less than a millimeter across, lacking “several structures that a human brain has,” and containing about 10,000 neurons plus supporting cells. ² Under good conditions on FinalSpark’s platform they stay electrically active for up to roughly 100 days. ¹

Now the comparison. In 2009 a Brazilian team led by Frederico Azevedo and Suzana Herculano-Houzel counted the cells in four adult male human brains using a method called isotropic fractionation, which is a polite way of saying they dissolved the tissue into a uniform suspension of nuclei and counted samples of it. They arrived at an average of about 86 billion neurons per brain. ³ (The old textbook number of 100 billion, it turned out, had rested on estimates rather than a direct count.)

Divide one by the other and your head holds roughly 8.6 million times as many neurons as one of these organoids. If your brain were the population of Earth, the organoid would be a small town. Not a city. Not a country. A town with one stoplight. And it is a town missing most of its infrastructure: no blood supply, no sensory organs, no body to report to.

So I am going to call them what they are, which is brain crumbs. Remarkable crumbs, made of genuine human neurons that fire and form real synapses. But crumbs.

Why Anyone Would Wire a Crumb to a Robot

The motivation is energy, and it is spelled out in the paper’s first paragraph. Silicon chips designed to imitate the brain, such as IBM’s TrueNorth and Intel’s Loihi, already sip power in the milliwatt range. But the authors note that when engineers scale those designs toward cortex-level complexity, projected power climbs into the kilowatt to megawatt bracket, “whereas the human brain performs comparable computation on roughly 20 watts,” citing a 2022 Nature perspective by Adnan Mehonic and Anthony Kenyon. ¹⁴ Twenty watts is a dim incandescent bulb. Your brain runs on a dim bulb.

That gap is the founding fantasy of a field that now calls itself Organoid Intelligence, a term the Bristol paper credits to a 2023 roadmap by Lena Smirnova and colleagues at Johns Hopkins. ¹ The idea: instead of building ever-cleverer silicon imitations of neurons, borrow some actual neurons and figure out how to talk to them.

Talking to them turns out to be the hard part. The Bristol paper’s real subject is not Braille at all. It is translation. Robots speak in sensor data. Neurons speak in spikes. Somebody has to write the phrasebook.

Writing the Phrasebook

The fingertip in this experiment is a sensor called Evetac, and it is a delight. Inside its housing is an event camera, a kind of camera that does not take pictures in frames but reports only changes: each pixel fires off an “event” with an address, a polarity, and a microsecond timestamp when its brightness shifts. When the silicone skin is pressed and dragged, the markers move, the light changes, and the camera streams out a spray of events. Evetac can resolve vibrations up to 498 hertz and catches shear and incipient slip. ¹ It is, in a loose sense, already speaking in spikes.

The team pressed it against each of the 26 letters at five depths, from first contact down to 0.4 millimeters, ten times each, for 1,300 recordings. ¹

Then came the phrasebook. The organoid sits on only eight electrodes, so the team carved the sensor’s field of view into eight regions, one per electrode, in a 2 by 4 grid tuned to the direction of sliding. For each region they computed four features of the touch: how many events happened (intensity), how long contact lasted (duration), when the busiest 100-millisecond window occurred (timing), and how much the activity fluctuated (variation). Each feature was then mapped, linearly, onto one knob of the electrical stimulus. ¹

• Intensity became the number of pulses, from 4 to 10.
• Duration became the length of each pulse, from 50 to 300 microseconds.
• Timing became a trigger delay, from 0 to 4,000 microseconds.
• Variation became the pulse strength, from 4 to 20 microamps. ¹

Those ranges were not arbitrary. Before any Braille, the team spent a long stretch simply poking organoids and listening. They found that spike counts rose almost linearly with the number of pulses, that currents below 4 microamps “rarely evoke spikes,” and that even a 50-microsecond pulse could set off a clear burst. They found that a single pulse’s echo mostly dies away within 60 milliseconds, stretching to around 100 milliseconds with more pulses, which is why they listened for 500 milliseconds after each stimulus. And they found that trains of four or more pulses produced more stable, more distinguishable responses, so they started the scale at four. ¹

This is the part of science that never makes the headline and is, to my mind, the whole story: months of patient calibration so that you can say something to a culture dish without either whispering below its hearing or shouting it into a seizure.

What the Crumb Said Back

For each touch, the organoid’s answer was boiled down to a list of eight numbers: how many spikes each electrode recorded in the half second after stimulation. A standard machine learning tool called a support vector machine then tried to sort those lists into 26 letters. The team tried two other classifiers too, k-nearest neighbors and random forests, and chose the support vector machine because it scored best. ¹

The results:

• One organoid’s responses identified the correct letter about 61% of the time, according to the abstract, the discussion, and Bristol’s press release. ¹²
• Three organoids, with their eight-number answers stitched together into one 24-number list, reached 83%. ¹²
• Chance, with 26 letters, is 3.85%. ¹

So the crumbs were not guessing. Something about each letter survived the trip from silicone to spikes and back to silicon.

And there is a lovely, unsettling detail in the data. The paper reports that a harder press in one region of the fingertip “does not always elicit a stronger response from the electrode assigned to that region.” The authors conclude that “the organoid therefore imposes its own intrinsic mapping between input location and population response.” ¹ In plain English: the tissue rearranges what it is told. It does not simply echo the stimulus. It has, for lack of a better word, its own accent.

The Twist Buried in Section 3.3

Here is where I put down my coffee.

To find out whether the organoids were adding anything, the team ran a control that every biocomputing paper should run and not every one does. They took the exact same classifier and pointed it not at the organoid’s response but at the stimulation pattern itself, the 32 numbers describing what was sent into the tissue. No neurons involved. ¹

On clean data, that neuron-free baseline scored 82%. The three-organoid system scored 83%. ¹

Let that settle. On a clean signal, the living tissue did not make the computer smarter. A one-point difference is a tie. The authors are upfront about this: the stimulation pattern “already contains substantial class-discriminative information,” and the organoid stage “does not reduce” that information. ¹ The crumbs carried the message faithfully. They did not improve it.

So why is this paper interesting at all? Because of what happened when the team started breaking things.

The Crumbs Are Better in a Storm

To test robustness, the researchers corrupted four of the eight channels in four different ways: Gaussian noise, uniform noise, missing data (40% of selected values set to zero), and outliers (40% multiplied by three). ¹ Then they compared.

The neuron-free baseline, reading corrupted stimulation patterns, fell to 55%, 67%, 59%, and 65% across those four kinds of damage. The three-organoid system, with the equivalent proportion of its channels corrupted, held at 83%, 82%, 72%, and 68%. ¹ Under every type of corruption the team tested, the living system did better.

Why? Nobody knows yet, and the authors are careful to call their explanation a hypothesis. Their guess is that the three organoids are not identical. Differences in “network connectivity, developmental state, and ongoing network dynamics” mean each one transforms the same input in a slightly different way, so their combined answer carries overlapping but not redundant information. Damage one part and the rest can cover. ¹ The paper points to the neuroscience idea of degeneracy, in which different circuits can stand in for one another, and to the way human listeners can restore a missing phoneme from context. ¹

Liu put the double edge of this nicely in Bristol’s press release: “Different organoids may have slightly different network connections, developmental states, and ongoing activity, so the same input can produce similar but not identical responses.” And then, the engineer’s sigh: “At the same time, if every organoid behaves differently, it obviously makes it harder to build a predictable and standardised system.” ²

That tension is the whole field in miniature. The quirk that makes the tissue robust is the same quirk that makes it impossible to manufacture to spec.

I keep thinking of a telephone line in a thunderstorm. The Bristol result is roughly this: a wire made of human neurons carries your voice about as clearly as copper on a calm day, and somewhat more clearly when the weather turns. That is a real finding. But the wire is not having the conversation.

So, Can Brain Cells Read Now?

No. And on this point I want to quote the researchers directly, because they deserve credit for refusing the easy headline.

“We showed that tactile information could be encoded, processed through the organoids, and decoded, but we did not demonstrate learning or long-term changes in the organoids,” Liu said. ² The paper’s own future-work section says the experiment “should be interpreted as an encoding and decoding study rather than as a direct test of organoid learning or long-term plasticity.” ¹

This was an open-loop experiment. Signals go in, signals come out, and an ordinary computer does the recognizing. The organoid never got feedback. It never knew whether it had been “right.” It did not get better with practice, because practice was not on offer.

Compare that to the experiment that made this whole field famous. In 2022, Brett Kagan and colleagues at Cortical Labs reported in Neuron that cultured neurons wired into a simulated game of Pong improved their play when given structured feedback. ⁵ The Bristol paper cites that work as an example of a closed-loop system in which the network “had learned the task objective.” ¹ That is the line Bristol has not crossed yet, and knows it has not crossed.

A Small Note on Numbers, Because I Read the Whole Paper

One wrinkle for the fact-checkers. The abstract, the discussion, and the press release all give the single-organoid accuracy as 61%. ¹² But the results section reports that one organoid’s accuracy “remains stable at approximately 57%,” and that the three organoids scored 59%, 62%, and 67% when tested individually (an average of about 63%). ¹ These appear to be different runs or different organoids, and the paper does not fully reconcile them. The 83% for the trio is consistent throughout. None of this changes the story, but it is the kind of seam you only see if you read past the abstract, and I would rather show you the seam than hide it.

Why It Matters Anyway

The Bristol result lands in a busy year for what the industry likes to call wetware. In August, NUS Medicine, the data center company DayOne, and Cortical Labs unveiled a prototype in Singapore built around a 20-unit rack of Cortical Labs’ CL1 biological computing systems. ⁶ Racks of living neurons are, apparently, a product category now.

What a rack cannot do on its own is understand a robot’s fingertip. That is the gap the Bristol paper addresses. Its contribution, in the authors’ words, “is not the fixed Braille-specific mapping itself, but the general interface principle it provides for tactile-to-organoid communication.” ¹ Event-based sensors spit out enormous, asynchronous torrents of data. Low-density organoid platforms have a handful of input channels. Someone has to compress the torrent into something eight electrodes can carry without losing location, timing, and intensity. That is a translation problem, and the paper offers one tested answer, with its data posted publicly for anyone who wants to try to beat it. ¹

There are also honest limits. The organoids sat on just eight electrodes each, which the authors say “constrains our ability to detect subtle spatial activity patterns,” and they call for high-density arrays with thousands of sites. ¹ They tested one task, one encoding, and one simple readout. Nobody yet knows whether a fourth organoid would help or merely add noise; “the scaling beyond three organoids and its saturation point remain to be determined.” ¹

And the next steps are the exciting ones. “The next step is definitely to move beyond Braille,” Liu said. “We are interested in things such as textures, edges, and different kinds of touch, and eventually in closed-loop tasks where the organoid receives feedback from what it does. That would allow us to start asking whether the organoid can actually adapt through experience.” ² The paper goes further still, imagining a loop in which synthetic sensors drive an organoid and the organoid’s output modulates living nerves, a testbed for neuroprosthetics before anything goes into a person. ¹

The Ending I Did Not Expect

I went into this paper expecting a story about brain cells doing something human: reading. I came out with a story about brain cells doing something very unhuman and perhaps more valuable: carrying a signal through damage without being asked to.

That is not intelligence. It is closer to the quality we admire in a good old bridge, the kind built with more beams than the engineers strictly needed. Three small, idiosyncratic lumps of living tissue, each hearing the alphabet slightly differently, turned out to be harder to confuse together than a clean machine was alone.

Somewhere in a perfused chamber, on a platform run by a Swiss company, a crumb of human forebrain about the width of a grain of sand has now “felt” every letter from A to Z. It did not understand any of them. It did not need to. For the first time I can recall, I find myself admiring a thing precisely for what it failed to do, and for the careful people who wrote down exactly where the line was.

References

1. Liu, T., Philamore, H., & Ward-Cherrier, B. (2026). Encoding tactile stimuli for Braille recognition with organoids. Frontiers in Artificial Intelligence, 9, 1800219. Published Oct 16, 2026. https://doi.org/10.3389/frai.2026.1800219. Dataset: https://doi.org/10.5523/bris.19qmlywobvl3u22pzo67z3bc6l ↩
2. University of Bristol. (2026, Oct 16). Mini brain organoids “reading” Braille marks key step towards biohybrid computing [Press release]. https://www.bristol.ac.uk/news/2026/october/mini-brain-organoids-reading-braille-marks-step-towards-biohybrid-computin.html ↩
3. Azevedo, F. A. C., Carvalho, L. R. B., Grinberg, L. T., Farfel, J. M., Ferretti, R. E. L., Leite, R. E. P., Jacob Filho, W., Lent, R., & Herculano-Houzel, S. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532-541. https://doi.org/10.1002/cne.21974 ↩
4. Mehonic, A., & Kenyon, A. J. (2022). Brain-inspired computing needs a master plan. Nature, 604, 255-260. https://doi.org/10.1038/s41586-021-04362-w (cited in [1] for the roughly 20 W figure) ↩
5. Kagan, B. J., Kitchen, A. C., Tran, N. T., et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron, 110, 3952-3969. https://doi.org/10.1016/j.neuron.2022.09.001 ↩
6. NUS Medicine. (2026, Aug 17). NUS Medicine, DayOne and Cortical Labs unveil biological data center prototype in Singapore. https://medicine.nus.edu.sg/news/nus-medicine-dayone-and-cortical-labs-unveil-biological-data-center-prototype-in-singapore/ ↩