Live Neural Network

Can a Living Brain in a Petri Dish Do Math?

In 2022, in a lab in Melbourne, about 800,000 human neurons were planted on a chip. Five minutes later, they learned to play Pong. Nobody programmed them to do it; the cells figured it out on their own. To be precise, their hit rate started exceeding random chance. What really got people worked up was something else: these cells were "improving."

Not the Same "Neural Network" ChatGPT Uses

When most people hear "neural network," they think of code. Matrix multiplication running on Nvidia GPUs. Living neural networks have nothing to do with any of that. This is about extracting stem cells from human skin, growing them back into neurons in a lab, and letting these real, living, mortal brain cells process information.

These neurons spontaneously clump together into a sesame-seed-sized ball called a brain organoid. Inside there are synapses, electrical signals, and some even develop layering similar to early fetal cerebral cortex. Researchers press this clump onto a chip densely packed with electrodes. The electrodes do the translating: external signals become electrical stimulation fed to the cells, and the cells' firing patterns get read out as output.

Once this loop closes, you get a living computational unit. A silicon chip always returns the same output for the same input. This clump of cells won't. Feed it the same electrical stimulation five minutes later, and the response might be completely different. Traditional "programming" becomes impossible here. You can't tell it what to do. You can only try to get it to "figure things out" on its own.

Getting from Skin to Brain Cells Is Not Hard. What Comes After Is.

Stem Cell Reprogramming

No need to crack open anyone's skull. Yamanaka Shinya discovered in 2006 that if you scrape a few cells off skin and stuff in four proteins, they "revert" to a stem cell state. This discovery won the 2012 Nobel Prize. Differentiating stem cells into neurons is now routine in any decent lab. If you can't pull it off, that's on you.

Electrode Array Interface

Differentiated neurons spontaneously cluster into a half-millimeter ball, the brain organoid. Put it on a microelectrode array. In Potter's era, a board had a few dozen electrodes. DishBrain now uses thousands. Resolution improved by two orders of magnitude. Up to this point, everything is reasonably clear.

The Part Nobody's Figured Out

How do you make cells do something useful? DishBrain borrowed Friston's free energy principle. Ball caught: give predictable signals. Ball missed: give random noise. Cells naturally tend toward making their input predictable, so their firing patterns gradually shift toward "catching the ball." Can this approach generalize beyond Pong? No convincing evidence so far.

This Field Has a Very Short History, and Nobody Paid Attention for Most of It

From the earliest proof-of-concept to commercial products, less than 25 years. Most breakthroughs are concentrated in the last three or four.

2001

Steve Potter's Rat-Brain Robot

Georgia Tech. Rat brain cells spread on electrodes, controlling a small robot. A few dozen electrodes. Extremely limited capability. Almost zero attention from the computer science community. Potter spent the next twenty years doing related research. The recognition he received was completely disproportionate to how pioneering the work was.

2013

Brain Organoid Standardization

Lancaster's Nature paper. Without it, none of what followed would have happened. Then came nearly a decade of silence.

2022

DishBrain Learns to Play Pong

Brett Kagan and Cortical Labs. About 800,000 neurons, high-density electrode array, closed-loop feedback. 579 sessions of human cell data, 272 sessions of mouse cell data. Published in Neuron. Nature, Scientific American, NPR all covered it. "Organoid intelligence" broke into mainstream awareness.

2023

The Baltimore Declaration

Johns Hopkins' Hartung formally named the field. That same year, Indiana University released Brainoware, using brain organoids for speech recognition at 78% accuracy. That number is unremarkable for machine learning. For a system whose "processor" is a lump of living cells, it's worth thinking about.

2024

FinalSpark Goes Online

16 organoids connected to the internet. $500/month for remote experiments. Actual user experience reportedly "far more noise than signal."

2025

CL1 Goes Commercial

Cortical Labs released CL1. $35,000 per unit, power consumption near 1000 watts, neurons survive about six months. Primarily sold to drug R&D labs. The "using it as a general-purpose computer" narrative? Nobody is seriously making that case right now.

Everyone in This Field Worldwide Could Fit in a Single Conference Room

Three names keep showing up: one Australian company, one Swiss company, and one American university lab.

Melbourne

Cortical Labs

The team behind DishBrain and CL1. Brett Kagan is the most cited person in this field, and also one of the most level-headed. In every interview he emphasizes that DishBrain is nowhere close to a real brain, "no signs of consciousness whatsoever." CL1 is priced at $35,000. Buyers mainly use it to test how drugs affect real human neurons.

Switzerland

FinalSpark

Total team of six people plus three advisors. They don't sell hardware; they sell remote access. $500/month to run experiments on their Neuroplatform. Over four years they've burned through more than 1,000 organoids and accumulated 18TB of data. Ten-year goal: "biological cloud computing." Co-founder Fred Jordan likes an analogy: nobody imagined the smartphone when the transistor was invented in 1947. Whether the analogy holds is debatable. It's certainly compelling.

Baltimore

The Hartung Lab · Johns Hopkins University

Thomas Hartung coined the term "Organoid Intelligence." He doesn't run computation experiments himself. His role is more like the institutional architect of this field. He organized the Baltimore Declaration. He drove the ethics framework discussions. He convened the scholars. In an emerging field that doesn't yet have a formed academic community, having someone willing to do this kind of thankless organizational work is itself important.

The Media's Favorite "Consciousness" Angle Is Actually the Least Worrying Part

Every article, by the end, asks the same question: could these organoids be conscious? The largest organoids currently have only tens of thousands of cells. Cortical Labs themselves say "roughly cockroach-level." You probably don't lose sleep over the subjective experience of cockroaches.

The genuinely thorny issues don't make for good headlines. You extract stem cells from someone's skin, culture them into a brain organoid, hook up electrodes and run experiments. What rights does the "donor" have? Who owns the experimental data? If someday this lump of cells produces something commercially valuable, how are the profits split? When the person donated skin, they probably thought they were participating in a dermatology study.

"A million times more energy-efficient." "Replacing silicon chips." These claims appear over and over in pitch decks and press coverage. They're not technically wrong: the human brain really does run 86 billion neurons on just 20 watts. The problem is that between "the human brain is energy-efficient" and "our lump of tens of thousands of cells can compute more efficiently than Nvidia GPUs," there's an entire universe of distance. FinalSpark says they can store about 1 bit. One bit.

I'm inclined to think this direction is worth paying attention to. Not because it's going to replace anything tomorrow, but because if "computing with living biological tissue" actually works, it shakes a foundational assumption that nobody has questioned since Turing: that computation must be built on deterministic, controllable substrates. Living cells are inherently nondeterministic and uncontrollable. If nondeterminism itself can become a computational resource, that's the genuinely interesting thing.

Can a Living Brain in a Petri Dish Do Math?