Around Halloween, I came across a post by Alex Ksendzovsky on X. The “we” in his post refers to a company called The Biological Computing Co., or TBC for short:
for four years, we've been brewing real brain cells on electrodes to understand how they compute. now, we've turned those discoveries into a video model anyone can use.
The post ended with “biocompute is real, commercial and this is just the beginning,” followed by “happy halloween.”
The word biocompute reminded me of an article I still owed myself. Last November, I wrote a piece on my WeChat public account called “What Will the Operating System of the Future Look Like?” At the end, I casually made myself a promise:
Later, I'll write about what I've learned about this thing called “biocomputing.” It has a direct connection to CPUs and RAM. There, I'm putting that promise in writing!
When I talked with Gemini about biocomputing in early November last year, I knew almost nothing about it. I was just curious, so I kept asking: “What exactly is biocomputing?” “So, has the first biocomputer been built yet?” “Why would a product like this use less energy?” “What would the user interface look like?” “Ordinary people make up most of the world. What would biocomputing mean for them?” …Then I got busy with other things. I already had an idea of how I wanted to write the piece, but never even got around to jotting it down. And just like that, I left that promise sitting there for almost a year.
Today, I'm picking those questions back up, one by one, and digging through the original papers, technical articles written by the companies themselves, industry reports, and even arguments researchers left on X… I'm thinking of making this a series, probably three pieces. This first one will tackle the question I was curious about at the very beginning: what exactly is biocomputing? And what conditions have to be in place for living cells to take part in a computation?
01 · Several Paths Under One Name
The single term “biocomputing” actually covers several very different lines of research.
The first brings living neurons directly into the computation. Researchers grow nerve cells on a chip covered with electrodes, send information in through electrical stimulation, and read the responses out of the cells' electrical activity. The cells themselves become part of the computation. That's how the 2022 DishBrain experiment got neurons to play ping-pong, and how a team at the University of California, Santa Cruz, got organoids to balance a virtual pole this February. The Swiss company FinalSpark took this a step further and turned experiments like these into a remote platform. The Australian company Cortical Labs went ahead and made a machine you can buy for your lab, called CL1. Then there's a team at Indiana University, which built a system called Brainoware in 2023. They placed a brain organoid on electrodes. Incoming signals set off complex electrical activity inside it, with a brief lingering aftereffect. Rather than specifying each step inside the organoid, the researchers trained an external readout layer to turn its activity into answers. The organoid itself could also change through stimulation-induced plasticity. This approach is called reservoir computing. They used it for proof-of-concept demonstrations of speech recognition and predicting the behavior of a chaotic equation.
The second path is the one TBC takes. The neurons are observed only in the lab. Researchers measure how they process information, draw out some principles, and write those into ordinary AI software. So when the “video model anyone can use” in Alex's post generates a video, the process doesn't pass through a single living cell. There's more to that story… I'll get into it in a later piece.
The third path began much earlier than either of the first two. It uses molecules. In 1994, computer scientist Leonard Adleman used DNA in a test tube to solve a problem: find a path through a graph with seven vertices that visits each vertex exactly once. His approach was to let DNA fragments representing different segments of a route join together in solution, forming all possible routes, then filter out the ones that failed to meet the conditions, step by step. The paper says the computation took about seven days of lab work. The magnetic bead separation step alone took a full day at the bench. More than thirty years later, this September, a paper in Nature presented a “Scaffolded DNA Computer” that ran ten programs. One added two 25-bit binary numbers. Small computations could finish in under a minute, and the same materials could be reused dozens of times.
Beyond these paths, there's another example that often comes up. In 2000, Japanese researcher Toshiyuki Nakagaki placed a slime mold called Physarum polycephalum in a maze, with a little food at the entrance and the exit. At first, its body spread throughout the maze. After a while, it contracted into the shortest route connecting the two food sources. It has no neurons, and “thinking” hardly seems the word for what it does. Yet it did solve a shortest-path problem. Here, the word “computation” stretches to include a change in the shape of a body.
02 · A Timeline, Side by Side
Lay these explorations and research projects out year by year, and someone might read them as one technology getting steadily more advanced. But they seem more like attempts to answer different questions, linked together by the same word.
John von Neumann was among the early thinkers to discuss the brain and the computer together. In 1958, his posthumous book The Computer and the Brain was published, based on a manuscript he had prepared for a lecture he never got to give. He made a comparison in the book: the human brain used roughly 10 watts. Spread across more than ten billion neurons, that came to about a billionth of a watt per neuron, while a vacuum tube used 5 to 10 watts and even a transistor used a fraction of a watt. By his estimates at the time, natural components had an advantage of eight or nine orders of magnitude in energy use, while artificial components had an advantage of four or five orders of magnitude in speed. We can no longer apply those numbers directly today. Still, the impression of the brain as “energy-efficient but slow” and the machine as “fast but power-hungry” has kept coming up ever since.
Over the following decades, the paths went their separate ways. Adleman computed with DNA in 1994; slime mold found its way through a maze in 2000. In 2013, Madeline Lancaster and colleagues at the Institute of Molecular Biotechnology of the Austrian Academy of Sciences published a method in Nature for growing cerebral organoids from stem cells. At the time, they were studying human brain development and microcephaly; it didn't have much to do with computers yet. Then, in 2022, DishBrain got a dish of neurons playing ping-pong. In 2023, researchers at Johns Hopkins University proposed a research roadmap for “organoid intelligence.” That same year brought Brainoware, as well as the debate around DishBrain. In 2024, FinalSpark published a paper on its remote platform. In March 2025, Cortical Labs launched CL1 in Barcelona, reportedly priced at about $35,000 per unit. And in 2026, the pole-balancing experiment, a research prototype built from 20 CL1 units at the National University of Singapore, the DNA computer, and TBC's video model all appeared over the course of the year.
03 · How Do You “Talk” to a Group of Neurons?
Of these paths, the first is the one I'm most curious about. How exactly does a dish of cells take part in a computation? With that question in mind, I went to read the original DishBrain paper.
The researchers grew neurons from the cortex of mouse embryos, or neurons differentiated from human stem cells, on a high-density microelectrode array. These electrodes could both stimulate the cells electrically and record the electrical signals the cells produced. That gave them channels for both “speaking” and “listening.”
Next came the ping-pong I mentioned at the beginning. More precisely, it was Pong, one of the earliest video games, with a paddle on each side of the screen hitting a ball back and forth. DishBrain used a simplified version, with a single paddle and a wall. A small area of the chip, with eight electrodes in total, was designated the “sensory region.” The ball's position and its distance from the paddle were translated into stimulation at different locations and frequencies across those electrodes. This step is called encoding. Two other areas were designated “motor regions.” Depending on which had stronger electrical activity, the paddle on the screen moved up or down. This step is called decoding.
Feedback was how the neurons were told how well they were doing. When the paddle hit the ball, all eight stimulation electrodes simultaneously delivered a brief, predictable burst of stimulation at a fixed frequency. When it missed, the neurons received four seconds of unpredictable stimulation, and the ball was then launched again in a random direction. This design drew on the free energy principle proposed by neuroscientist Karl Friston, one of the paper's authors. Roughly speaking, the idea is that a nervous system tends to make the input it receives more predictable. So if missing the ball always brings a burst of chaotic input, a group of neurons might gradually adjust its activity to hit a few more balls.
To show that the changes really came from this closed loop, the paper included several control groups: chips with culture medium but no cells; a “Rest” group in which cells still controlled the paddle but received no information about the ball's position; and an entirely computer-simulated group in which random noise drove the paddle. The researchers also compared conditions with and without feedback. The paper reported what it called “apparent learning” within five minutes. Cultures receiving feedback showed a clearer improvement in hitting the ball than the control groups.
04 · The Days After the Paper Came Out
The paper was titled “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.” I asked Grok to look through the discussions on X from that time. That was how I learned that researchers were already debating it within days of publication.
The first to speak up was Anil Seth, a neuroscientist who studies consciousness. In a thread, he said he greatly admired the work but was concerned about “exhibit sentience” in the title, because most people would read it as some minimal form of consciousness. He wrote, “There is no reason to suppose that … DishBrain experiences anything at all.” In an exchange with the paper's authors, philosopher David Chalmers also pointed out that most people read the title, not the definition the authors give in the paper.
A few days later, NeuroAI researcher Patrick Mineault criticized it from another angle. His concern was the evidence itself rather than the wording: “The effect size is tiny, the experimental controls are pretty barebones.” In his view, the controls used only “no stimulation.” A more meaningful approach would be to vary the noise or reward structure, which would do more to reveal what the cells were responding to. Brett Kagan of Cortical Labs, the paper's corresponding author, replied directly under his post: “There were five distinct control groups.” He also asked which essential controls Mineault thought they had missed. But how many control groups there are, and whether those controls can rule out a particular alternative explanation, are actually two different questions. Reading the paper makes that distinction clearer.
The following spring, a group of neuroscientists coauthored a critical letter in Neuron. Kagan's team responded in the same issue, saying they used sentience in a certain limited sense and had not intended to exaggerate the experiment's conclusions. Nicolas Rouleau, another researcher in the field, defended the paper on X, arguing that the letter seriously misunderstood the work.
Read those posts and the letter together, and what emerges is an experiment being pulled in several directions and examined from each one. Some people worried about what the public would take that word to mean. Some questioned whether the controls were enough to rule out other explanations. Then there was the authors' answer: there were, in fact, five control groups… These arguments never led to a neat, blunt verdict. But they did make the boundaries of “a dish of cells learned to hit a ball” a little clearer. What that sentence describes is a change in the cells' performance consistent with learning, within this closed loop and under these measures. Whether the cells “understood” the game, or experienced anything, remained unanswered by the experiment.
Later, in a 2024 preprint, Kagan's collaborator Adeel Razi and colleagues compared cultured neurons with several deep reinforcement learning algorithms. Under matched conditions of roughly 70 games and 20 minutes, the cultured neurons performed better. But the paper itself also said those 70 games were nowhere near enough for the algorithms. With longer training, all of them learned too. Energy use wasn't measured either. On that point, the paper left it at “difficult to directly compare energy consumption.”
05 · They Learned, but Couldn't Remember
The pole-balancing experiment published in Cell Reports this February used different cells and a different task. It also described “learning” in more detail.
The team at UC Santa Cruz grew three-dimensional cortical organoids from mouse stem cells and had them balance a virtual pole, rather like standing a ruler upright in the palm of your hand and keeping it from falling. The pole's angle was encoded as electrical stimulation and sent into the organoid. The organoid's electrical activity was then decoded into a force applied to the cart. Unlike DishBrain, the organoids received no training signals while balancing the pole. Only after a game ended, and only if performance across the recent games had not improved, would a reinforcement learning algorithm step in. It would decide which neurons to stimulate and send them a set of training signals. Ash Robbins, the paper's first author, compared the algorithm to a coach standing beside them, saying, “you're doing it wrong, tweak it a little bit in this way.”
In experiments that alternated between several training methods, 23.1% of training cycles reached the proficiency threshold when this coach selected the signals, compared with 4.5% with random signals and 2.3% with no signals. In a separate set of experiments that used only the coach throughout training, the figure reached 46.4%. These percentages refer to the share of training cycles that reached the proficiency threshold, not the share of organoids that learned to balance the pole.
But that progress couldn't survive a break. The experiment ran in cycles of 15 minutes of training followed by 45 minutes of rest. According to the university's news release, performance returned to baseline after the rest period. The little bit of progress they had just made was gone almost as soon as it appeared. David Haussler, a member of the team, suggested that more complex organoids, perhaps including several brain regions involved in learning, might be needed for them to retain what they learned the way animals do. He also made a point of saying that the team's goal was to advance brain research and the treatment of neurological diseases. They did not intend to replace robotic controllers or other computers with lab-grown animal brain tissue.
Beyond “not remembering,” there's also “every dish is different.” In their paper, FinalSpark looked back on four years of running the platform. They had replaced the organoids more than 250 times and tested over a thousand in total. At first, an organoid survived only a few hours. After improvements to the system that circulated the culture medium, the best cases lasted around 100 days. An organoid's spontaneous electrical activity also changes over its lifetime, and the minimum current needed to make it fire increases. With an ordinary computer, we take it for granted that the same input produces the same output every time. Living cells are different. The same dish of cells may respond differently today than it did yesterday. Swap in another dish, and it's a different story again.
06 · An Organoid Isn't a Little Brain in a Dish
When I see the words “brain organoid,” I picture a miniature brain, thinking away in a petri dish. But an organoid is a small clump of tissue that grows from stem cells in a three-dimensional culture. It can form certain cell types and layered structures resembling those of the brain, but it is usually only a few millimeters across. It has no blood vessels, no input from eyes, ears, or a body, and none of the full set of connections between regions of a complete brain. Lancaster originally grew them to observe early human brain development in a dish. Microcephaly, mentioned earlier, was one example: she wanted to study why patients' brains grew smaller.
Researchers themselves often bring up the gap between a dish and a brain. Last December, Rouleau wrote on X that neurons in monolayer cultures and organoids often fire all at once. These bursts of synchronized firing are typical in a dish but pathological in an animal's brain. “There, we call them seizures.” In his view, then, studying brains in a dish means finding a way to give these cells some form of “body” again: a way to receive ongoing input and feedback.
The cells come from different sources, too. DishBrain used neurons of both mouse and human origin. The pole-balancing experiment used organoids grown from mouse stem cells. CL1 reportedly uses human neurons derived from reprogrammed skin or blood cells from adult donors. Researchers and ethicists are still discussing how complex such tissue would have to become before we need to seriously consider whether it might have some form of experience. I'll leave that topic for the third piece in this series.
So, I now have some answers to that first question I was curious about a year ago. Biocomputing is a collection of attempts that take different forms. Some bring living cells directly into computation. Some only learn from the cells. Others bypass cells altogether and compute with molecules. On the path that involves living cells, one set of neurons grown in two-dimensional cultures showed apparent learning within minutes of closed-loop feedback. Another set of three-dimensional organoids, with a coach's help, made short-term progress at balancing a pole, but that progress didn't survive a 45-minute rest. It seems that getting a dish of living cells to take part in computation requires electrodes, encoding, decoding, feedback rules, and a whole set of conditions that keep it alive.
OK, now another question has popped up. If it's this much work to keep them happy, why have more and more people started studying them in recent years? Why have some even made a serious effort to turn them into products? In the next piece, I want to start with something as everyday as it gets: electricity.
Sources
Opening
1. Alex Ksendzovsky's launch post on X, October 30, 2026.
2. 是梦昕呀, “What Will the Operating System of the Future Look Like?”, 光合制梦, November 17, 2025.
Several Paths Under One Name
3. Cai, H., et al. Brain organoid reservoir computing for artificial intelligence. Nature Electronics, 2023.
4. Adleman, L. M. Molecular computation of solutions to combinatorial problems. Science, 1994.
5. Stérin, T., et al. A thermodynamically favoured molecular computer. Nature, 2026.
6. Nakagaki, T., et al. Maze-solving by an amoeboid organism. Nature, 2000.
A Timeline, Side by Side
7. von Neumann, J. The Computer and the Brain. Yale University Press, 1958.
8. Lancaster, M. A., et al. Cerebral organoids model human brain development and microcephaly. Nature, 2013.
9. Smirnova, L., et al. Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science, 2023.
10. Report on the launch of Cortical Labs' CL1. DatacenterDynamics, March 2025.
How Do You “Talk” to a Group of Neurons?
11. Kagan, B. J., et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron, 2022.
The Days After the Paper Came Out
12. Anil Seth's post on X, October 2022.
13. David Chalmers' post on X, October 2022.
14. Patrick Mineault's post on X, October 2022.
15. Brett Kagan's reply to Mineault, October 2022.
16. Balci, F., et al. A response to claims of emergent intelligence and sentience in a dish. Neuron, 2023.
17. Kagan, B. J., et al. Scientific communication and the semantics of sentience. Neuron, 2023.
18. Nicolas Rouleau's post on X, March 2023.
19. Khajehnejad, M., et al. Biological Neurons Compete with Deep Reinforcement Learning in Sample Efficiency in a Simulated Gameworld. arXiv preprint, 2024.
They Learned, but Couldn't Remember
20. Robbins, A., et al. Goal-directed learning in cortical organoids. Cell Reports, 2026.
21. Brain organoids can be trained to solve a goal-directed task. UC Santa Cruz News, February 2026.
22. Jordan, F. D., et al. Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence, 2024.
An Organoid Isn't a Little Brain in a Dish
23. Nicolas Rouleau's post on X, December 2025.