Transcription
This is a simulation of Pong, the iconic 1970s arcade game where you use this paddle to stop this ball from getting past you. It's a pretty simple game for us humans to learn and, as of quite recently with advancements in AI, computers can learn it too. But today, Pong has a new kind of player, not a computer nor a human; a cluster of living human brain cells on top of a silicon chip can play Pong.
Yeah, researchers in a lab taught a tiny clump of neurons to, in a sense, play this video game. But it's more than just a game. Some believe merging biology with computing could upend our approach to artificial intelligence. AI is a massive, often volatile global market with people looking all over for a competitive edge. In 2025, private sector companies teamed up to invest more than $500 billion in AI infrastructure alone. But what if the building blocks of that infrastructure include living human brain cells, with the promises and challenges of our own biology?
Sounds like sci-fi, right? Well, we're kind of on our way. Science fiction talked about this. I love thinking about the Hollywood stuff. Are we really computers? Are computers really human? My life has been focused on how can you elicit intelligence from brain cells in a dish? I've got a pretty broad background. I did some work starting with cognitive psychology, neuropsychology, and then eventually moving into neuroscience.
Brett Kagan is the chief scientific officer at Cortical Labs in Melbourne, Australia. They made a name for themselves when they got their neurons to play Pong in 2022. They called it DishBrain. Yeah, let's just set up and see how it goes. Maybe we'll all be in for a surprise.
Quick timeout. Before we go any further, it might be worth understanding why companies like Cortical Labs are blending living neurons with synthetic systems. Compared to traditional machine learning, brains need far less training data to understand real-life environments and new situations. So what we're seeing here is we're investigating can neurons actually distinguish between different handwritten digits. And we wanna see kinda neurons behave in a way that says they recognize that all of the sixes look like sixes, all of the fours look like fours, and so on. Anything with biology can learn to navigate its environment incredibly quickly, incredibly efficiently. And we've had to learn to do that through evolution; otherwise, we've died.
The promise of far less data and energy consumption is the holy grail for AI advancement. Here's why. Supercomputers like this one can be a million times more powerful than the fastest laptop, but both supercomputers and AI run on the same basic technology: silicon chips. There's never been a focus like there is currently on the chip industry. People who've realized that chips are the heart of everything, that this fundamental change to the economy that's being promised in the form of AI depends upon semiconductors, depends upon their rapid development and depends upon an ever-increasing supply of them.
For over 50 years, we have been creating better chips by packing more and more transistors inside one integrated circuit. The transistors are on/off switches, the physical manifestation of ones and zeros, and we've gotten faster and faster by going smaller and smaller. Over the last six decades, we've been trying to double the amount of transistors on an integrated circuit every two years. This is the heralded Moore's Law, observed by this guy back in 1975. Today, a phone's processing chip contains billions of transistors, and those transistors are so small they're just a few nanometers apart in places. But as our good old Intel co-founder, Gordon Moore, observed, this steady march of progress also has an endpoint. You're creating a chip, a three-dimensional circuit by depositing layers of material on a disc of silicon. Some of those layers now are one atom thick. Guess what? When you get to one atom, you can't go any thinner. So we have a problem, and folks all over the world are looking for solutions to our processing limits.
After all, our demand for more computing likely isn't slowing down. It's not like suddenly companies and governments and organizations are gonna say, "You know what? We don't need more data." Everybody wants to collect more and more data, right? People liken it to the, you know, modern gold rush. I wonder whether DeepSeek has kind of changed. I mean, it feels like it's upended the race this week. DeepSeek is such a fundamental change to the economics of what's going on. DeepSeek's January 2025 effect on markets was a strong reminder of these economic incentives. Companies are racing to build more efficient AI models all around the world at times also fueling existing political tensions. The President said that he believes that this is a wake-up call to the American AI industry.
One longstanding breakthrough in AI comes from biology, and our ability to replicate it back to neurons. The state of data in computing right now is pretty exciting, right, there's a lot going on. A big part of that is trying to learn what we can from how neurons in the real brains and biological brains connect to each other and how they communicate, and then seeing if we can take what we've learned about the biology and then hardwire that into artificial systems, into computational systems. And that's where these ideas of artificial neural networks and deep learning came from, and the current status of these artificial neural networks and deep learning is pretty fantastic. But building these AI models and allowing masses of people to use them is costing a lot of energy. Already, there are parts of the world where there literally isn't enough power available via the grid to support these giant data centers. By 2034, data centers around the world are expected to consume roughly 1,580 terawatt hours yearly; that is about as much as is used by all of India.
To use the words of Jensen Huang, the CEO of Nvidia, there's a trillion dollars in infrastructure out there that needs to be modernized. The old way of doing things just simply doesn't stack up to the needs of modern computing. So neurons might help solve some problems, especially around energy. Our brains are much better and more efficient in similar tasks to AI. For example, a supercomputer must consume up to 40 megawatts of power to function. Our brains run on just 20 watts. That's the difference between a couple of LED light bulbs and thousands of homes. Our brains are also powerful, with about 86 billion neurons that form over 100 trillion connections to each other. And with cells, the distinction between hardware and software isn't quite as well distinct. Though dynamic and intertwined, the cells' actual structure can change and mutate too. Because of the way that computer chips are organized, they have limited logic functions. They tend to want to do things one thing after another. That's what they're good at. That's what they were designed to do. Human brain doesn't work that way. The human brain is kind of a lattice structure of interconnections where information flows in sort of different ways and different points. So the limitations of bio is that we're just at the very beginning of it.
DishBrain was our first prototype. We were essentially trying to ask the question, "Can you interact with a bunch of brain cells in a dish and get them to do anything useful at all?" Brett and his team started by growing a layer of 800,000 neurons on top of a silicon chip; that's about the size of a small bumblebee's brain. And since neurons in our brain naturally communicate using electrical impulses, the researchers could both send and receive impulses between the neurons and a computer hooked up to a silicon chip. To translate the Pong game into something this clump of neurons could sense and control, the chip is divided into three sections. There's one larger sensory section which sends electrical impulses to the neurons to represent aspects of the game, like where the ball is. There's also two motor sections. If neurons send impulses in this one, the paddle moves up. If neurons send impulses in the other, the paddle moves down. The chip acts as a communicator between the neurons and the game, but it still doesn't mean that these neurons would want to play the game. Why would a bunch of cells in a dish actually change their activity in any particular way? There's no Pong-playing gene that exists in brain cells to make them wanna play the game Pong. So the researchers try to give the neurons incentive. Every time the brain cells missed the ball with the paddle, the researchers stimulated them with chaotic and unpredictable bouts of electricity. On the other hand, when the neurons successfully hit the ball, they received predictable stimuli from the researchers. Within minutes, the neurons started hitting the Pong ball more consistently.
Why did it work? One possible explanation is a theory called the free energy principle. It suggests that almost all things in nature, even as small as the DishBrain, are constantly trying to perceive and understand their world and minimize surprise. If you can't predict what's going on in the world... If you can't say, "If I reach for a glass of water, I'll successfully pick the water up," it's very hard to survive. DishBrain was a proof of concept for this larger idea of biocomputing, also called organic computing, biological hardware, and, perhaps most vividly, wetware. DishBrain made a splash so big in fact that another team of scientists tried the same experiment using hydrogel in non-living, soft, flexible substance, similar to jello. The hydrogel also learned to play Pong. So the question became, what can neurons do that other substrates can't? How much computing abilities can they provide and, frankly, can they make any money now?
Vevey. It's a small city, it's peaceful. It's quite nice to do fundamental research here. Meet Fred Jordan. Fred's the CEO and founder of FinalSpark, a biotech company nestled in the quaint Swiss town of Vevey. FinalSpark is a deep tech startup with a team of just under 10 people. Deep tech startups don't focus on end-user services. Instead, they aim to solve complex problems by harnessing scientific discoveries. Like Cortical Labs, FinalSpark is also a biocomputing company. We have a way to have neurons connect together to create a small bowl that is called the brain organoid. FinalSpark relies on their own funds of just over one and a half million dollars. Cortical Labs in comparison raised around $10 million in 2023 from folks like In-Q-Tel and Horizon Ventures. FinalSpark says they have not accepted outside investment. They've used their own money towards research and development and made 16 brain organoids available for a subscription fee on a cloud computing network called the Neuroplatform. You can come and you will look at real-time information streamed 24/7 from our neurons here in Switzerland. This picture here shows a close-up version of what's inside the incubator. The important part for us is this activity. Each of these small graphs here represents the voltage for one second of one electrode. Now, people who are using the neural platform are able to also send some stimulation to some of these electrodes to which sometimes they will answer.
Researchers from all over the world can log on to the Neuroplatform to observe, study, try different things, some in the field of robotics. Others use the organoids as teaching tools in alternative computing university classes. Initially, we built this only for us, and then one university came, and then 9, and then 34, and now we have 200 people who want to access remotely to our brain organoids. This is totally new. There is no background at all. So there are more questions than facts about this way of investing, so I think this is the biggest challenge. Investors are always cautious fundamentally about deep tech, and it's even not only deep tech, it's a new deep tech. If you talk about quantum computing, for instance, it's deep tech, but it's an old deep tech. In the future, we'll have to live in a world where there will be some parts in computing which are living, and other which are just artificial.
It's all well and good for research purposes, but a future where biology and hardware become one isn't easy. Cells need to be fed, kept a comfortable 37 degrees Celsius, their waste cleaned. After a while, they die. You have to look at the environmental physics. A computer chip now is drawing multiple hundreds of watts. That throws off so much heat that you have to have these cooling systems in place and that's a tremendous problem for the computer industry. But imagine trying to bring in biological materials into that environment if you're gonna make a hybrid type technology. It's just not possible. You literally are going to start cooking the biological material. Any number of engineering problems exist that will need to be solved if we are to bring in this medium. All this is hard to scale. Investors who take this leap now aren't looking to turn a quick profit.
This is Cortical Lab's first commercial product: the CL1 unit. It's a processor with human brain cells in it. CL1s come with either neurons on a chip, organoids, or a type of neural material Cortical Labs calls bio-engineered intelligence. This box acts like a body for the cells, feeding them, disposing of their waste, filtrating, and pumping fluid around to keep them alive. The company's success hinges on whether or not they sell their CL1 units. Each one goes for around $35,000, and they're expected to launch in March 2025. Scaling is always a challenge just logistically. It's one thing to build a prototype in the lab; it's another to have hundreds of them rolling off a production line and have them be reliable and stable. I am not the pope. I can fail. Perhaps my imagination, my fantasy, is not big enough, but I would not invest into these companies from my side.
This is Thomas Hartung. He focuses on toxicology, think poisons, in an effort to improve public health. Thomas, how much time do you spend in here? Honestly, I come here when I look for my wife or some of the other people in our team. He spends a large amount of his time thinking about how we can use biocomputing and these brain organoids now. Our group was not the first to produce brain organoids. We used to call them mini brains at the time, but we were the first to standardize this. And we were very proud that when we showed this at one of these big science journalist events, the AAAs, we got on the front page of the Financial Times. I'm a crazy professor. I can do what I want. I'm trying to replace animal testing. Animal studies aren't always great at predicting how toxins or poisons affect humans, and although human brain organoids today deliver pretty modest computing tricks, their promise in the field of biomedicine is already remarkable. The potential for just basic neurology is enormous. It could be a very good strategy to accelerate drug development. It is a lengthy process to do an animal experiment, not only a costly one. The value of being one day earlier to the market for a pharmaceutical company is about $1 million.
We first started to work with brain organoids probably around 2015, 2016. The aim of the lab is to understand Parkinson's disease, and we were pretty unsatisfied with available animal models as well as more simple cell culture models. The connection between the biocomputing and the disease modeling is actually quite interesting because one of the features or characteristics that we see in neurogenerative diseases is dementia, so people start to forget things. The concept that we would like to explore is if I can train a brain organoid to learn certain tasks, can it also forget these tasks, and would it forget these tasks quicker or in a different way if it's in a disease context? Like, would an Alzheimer brain organoid forget things quicker than a healthy control organoid? There are no cures for neurodegenerative diseases, only symptom management. What if we could understand and get rid of the disease itself? If we can really understand the human brain, then we can potentially treat Alzheimer's, Lou Gehrig's disease, epilepsy, all kinds of brain-based disease that we really don't have good treatments for now. Many pharma companies have started actually to move away from these kind of diseases because it was just not commercially viable. Now with these new technologies coming up, I think there's kind of new hopes that we'd be really develop or be able to develop therapies for these diseases that burden our societies quite a lot.
Studies suggest that at some point in their lives, more than one-third of the entire human population will suffer from neurological conditions like Parkinson's, Alzheimer's, or dementia. By 2050, some project there will be an overall 22% increase in brain disorders, in part fueled by population growth and aging. The cost for Parkinson's in the U.S. alone, a bit more than $50 billion. This is direct and indirect costs, so really like the healthcare, but also what it does to the economy. At the moment, there's about 10 million Parkinson's patients in the world. Yearly in the U.S., there's about 90,000 people being diagnosed with PD, and since worldwide populations are aging, the likelihood to suffer from PD is increasing. The idea of using biological medium to either process or store information makes sense in a lot of ways, but there are those who are saying, "Actually, this is living tissue." Not only "How can we actually do this, but should we be doing this?" I think the biggest elephant in the room is would a brain organoid become self-aware? I'm not worried at all at the moment with these tiny, tiny little balls of cells, that they would suddenly develop something which I could consider consciousness. There's a lot of ethical issues in all biomedical research, and this starts if you're using stem cells, as we do. You have to imagine the donors of these cells are often still alive. Is your signature still valid because you couldn't imagine that somebody would produce a brain which is thinking? These are all interesting questions. And they really get hot if we assume at some point in the future these brain organs could suffer. They could experience pain, they could suffer from what they learn. So imagine this here, is this still acceptable? If they are self-conscious, even in a very simple sense, can I stop feeding them? Can I kill them? How much complexity you need to really call something intelligent? A lot of things start and take years in the lab before they become commercially viable. I think we'll get there one day. It's just too early. One way to look at this is that it's taken the semiconductor industry more than 50 years to get to this point, where they can supply the world of electronics. 50 years is quite a long time. They say they've been going at breakneck pace. The amount of money, the amount of effort, the amount of study, the amount of science that has gone into that will take a long time to replicate with another medium.
Whatever the future may bring, one thing is already clear: that brain cells will always deliver more than the sum of their parts, and that's neither boring nor binary.