Biocomputing on Human Neurons  ## Metadata - Author: [[The Changelog: Software Development, Open Source]] - Full Title: Biocomputing on Human Neurons - Category: #podcasts - URL: https://share.snipd.com/episode/497d2e9e-a486-43d5-ad66-e7ccd3e13680 ## Document Notes ## Highlights - **Remote Access To Human Neurons** - FinalSpark provides remote, browser-based access to real human neurons for experiments via a Python API. - They aim to use living neurons as energy-efficient processors to explore biocomputing possibilities. Transcript: Ewelina Kurtys So that means that our lab is available remotely. So everyone from all over the world can access our laboratory through the website browser. They can log in and they can write Python code to do experiments because everything is connected to real neurons. Because we are trying to build computers using living neurons. We want to use neurons as a processor because they are very energy efficient. So that's the reason. ([Time 0:04:34](https://share.snipd.com/snip/66e0002f-15e7-444d-8312-29d6569d34ca)) - **Massive Energy Efficiency Potential** - Neurons are estimated to be roughly one million times more energy efficient than silicon-based processors. - FinalSpark sees energy efficiency as a core motivation for pursuing biocomputing. Transcript: Kyle Galbraith So we recently had Greg Osuri on the show talking about the AI energy crisis and all these ways that we can potentially power this new compute demand, which is burgeoning. And it sounds like those ways were hard. And maybe if we figure this way out, it's way better. Exactly how much more energy efficient is it to compute on neurons versus silicon? Ewelina Kurtys So neurons are 1 million times more energy efficient. ([Time 0:05:33](https://share.snipd.com/snip/e7229020-e449-4f5b-ab3a-def6a08baef8)) - **Neurons Use A Different Code** - Neurons encode information in space and time rather than binary zeros and ones, so programming requires new paradigms. - This mismatch with digital logic makes biocomputing comparable to the challenge of quantum algorithms. Transcript: Ewelina Kurtys However, neurons work totally different. And that's why it's so difficult actually to build the computers, Because indeed, in the computer, you have zero and ones. And this is one of the reasons why actually they use so much energy. But the brain is encoding information totally differently in space and time. So when we have neurons in our head, it matters when and where in exact location they are active. And this is information. So this is totally different type of encoding. So no zero ones, actually. But there are a lot of ways how we can look at the activity of the brain. For example, how often you have spikes, or what are the time in between the spikes. So this electrical activity of the neurons. So we know for sure it's totally different. And that's why this project is so difficult because we have to learn totally. We have to figure out totally new way of programming, totally new approach. It's the same actually as in quantum computing. It's the same situation that you have totally different hardware which is working differently. So it's necessary to figure out a new way of writing algorithm. And that's why it's so difficult because actually someone has to come up with some idea, which would be totally, totally different. ([Time 0:10:24](https://share.snipd.com/snip/248a3927-e829-478a-8eb3-f39ec374940c)) - **Growing Mini-Brain Organoids** - FinalSpark grows neurosphere organoids ~0.5 mm with ~10,000 neurons over ~3 months and connects them to a remote platform. - These organoids currently support experiments but cannot yet process images, audio, or video. Transcript: Adam Stacoviak Me see if i understand this i'm grokking some of the stuff from what you're sharing and then also from your very awesome website final spark.com it says they transform stem cells into Mini brains that learn and adapt growing neurons in an orbital shaker which i have no idea what that sounds so cool yeah over the three month period. And these mini brains, organoids, not sure if that's a term y'all came up with or not, but that sounds cool too. Sounds like a match, right? 0.5 millimeters in size with about 10,000 neurons that function as real brain tissue. So you've found a way to take stem cells, grow them over a three month period. They have a, they used to have a half-life that was even shorter, like a few hours now that you can actually live a hundred days and they get connected to this neuro platform with 24 seven Access. So essentially the same way we treat a CPU in AWS, you're doing with stem cells, turn neurons, turn mini brains, turn or ornoids that can be compute platform. Ewelina Kurtys Yes, absolutely. Although now it's for experiments. So we cannot really process information the same way as in digital. But yes, it's available remotely. ([Time 0:12:10](https://share.snipd.com/snip/e5a09741-c812-4448-aa6c-864c74b60ed4)) - **Unstable Biology Is A Core Challenge** - Major roadblocks include unknown neural encoding and the inherent plasticity of living tissue, causing variable responses. - FinalSpark built an automated lab to run many repeatable experiments and discover reproducible control patterns. Transcript: Ewelina Kurtys No, it's very difficult, very challenging project. That's why we expected to build this real computers in around 10 years. Okay. So it's a bit challenge when we talk with investors potential because it's quite long-term project and it's very, very difficult because nobody knows how really neurons encode information. So this is the biggest challenge. So we know that neurons are active electrically, you know, they are spiking. Spike means that, you know, there is electrical activity. And we know quite a lot about this, how it happens. However, we cannot really translate this into some specific information. So for example, you have text or image, maybe about imaging, there is some understanding already in neuroscience. But you know, when you have, for example, words, text, it's hard to say how some word can translate to specific activity of neuron. So at the moment, as I said, many people, you know, do a lot of random experiments. ([Time 0:21:47](https://share.snipd.com/snip/16452a69-0805-464f-9203-7c8c127f59fb)) - **Programming With Electricity And Neurochemistry** - FinalSpark programs neurons via electrical stimulation and timed chemical feedback like dopamine and serotonin. - They use neurotransmitter release as reward signals to shape neural behavior in reinforcement-like loops. Transcript: Ewelina Kurtys Is actually a challenge. So what we do, we try to send them electrical signals because neurons are placed on the electrodes. You can see this on our website, finalspark.com. There is section live. You can see the readout. So we send them electrical signals and we measure how they respond. So how they change the activity, how they change electrical activity. Another thing how we try to compute neurons is also by sending them some chemical signals. At the moment, we can send them dopamine or serotonin. So programmatically, we can program in Python that neurons will get dopamine at some point. So this is why, so at the moment, this programming is not really to do some specific task as with computers, but actually to change the behavior of neurons. So this is actually the first step. So we want to be able to consistently control how neurons behave. ([Time 0:24:25](https://share.snipd.com/snip/eb469f58-ad69-4566-843b-f23b0ee694b5)) - **Storing One Bit In Neurons** - FinalSpark has reproducibly stored one bit of information by shifting a neurosphere's center of activity across electrodes. - The result demonstrates progress but highlights the early, incremental nature of current achievements. Transcript: Ewelina Kurtys It's not only so bad, uh, we were able to store one bit of information. So just to give you an idea about the stage at which we are, uh, we, we store one bit of information in neurons that was quite consistent and we were able to reproduce this, uh, many times. So we are happy. There is some kind of progress, uh, but yes, it's, it's very challenging to, to get some something. ([Time 0:26:04](https://share.snipd.com/snip/03cb8677-98fb-416c-812b-10f56e0ea276)) - **Speed And Memory Tradeoffs** - Neurons are slower than silicon and unsuitable for high-speed or large-memory tasks. - Their strength lies in energy-efficient solutions for complex, non-repetitive problems like generative AI. Transcript: Ewelina Kurtys Well, I can not, to be honest, I cannot tell you how much time it took. No, actually, I don't know, but I guess it's in seconds or milliseconds, but I don't know. I can say that in general, neurons are slow. And in general, neurons will be good for tasks, which don't have to be fast. Because also when we look at the human brain, and we look at computers, we can see that computers are very good in speed, in doing repetitive things very, very fast. And we will never be able to compete with the digital on that. ([Time 0:27:58](https://share.snipd.com/snip/2b474315-e059-4c30-9aa5-7001573dbd0c)) - **Reward-Based Training For Neurons** - Use neurotransmitters as feedback: deliver dopamine for reward and withhold it for no reward to reinforce desired neural responses. - Repeat electrical stimulation and reward loops to attempt consistent learning. Transcript: Ewelina Kurtys So we have UV lamp. And when the UV light is open, the dopamine is released because dopamine is closed chemically. It's encaged in the chemical so that it's not active. But when it sees the UV light, then it's released. So it's a way of releasing very, very quickly dopamine to the neurons. So this is how it works. So then it's connected to some controllers, which are connected to the lamp and then the lamp switch on. Kyle Galbraith Okay. Same thing for serotonin? Ewelina Kurtys Yes. But for serotonin, we have different wavelengths i don't remember which one it's not uv uh but then we have different wave wavelengths yes so basically so that they don't overlap so You can have both in the in the medium and we of course we plan to have them to have more of this and of course much more neurotransmitters but we are we started with dopamine now we added Adam Stacoviak Serotonin are those hormones are those? What is the proper terminology to call dopamine? No, they are neurotransmitters. Ewelina Kurtys Neurotransmitters, so you have several in the brain, and they affect learning. And actually, the whole idea of using them is because we use them for feedback. Because actually, the way how humans are learning is by feedback. So you have interaction with the environment, you get feedback, so things are going good or bad, and then you learn if you should do this or not. And the same way, actually, neurons are learning in vitro on the very basic level. Because, for example, when something good happened, then there is dopamine release and that reinforces the connection between neurons. So if they have done something good, then it kind of reinforced this behavior. It, but, and actually there are different opinions about how to, how to give a punishment or reward to neurons. But our idea is to give dopamine as a reward and no dopamine as a punishment. So that's a, that's used for the feedback loop. So for example, you stimulate a neurons with some electrical signals, you measure the behavior. For example, you want that they increase activity. So if they do this, you give them dopamine. If they don't do this, you do nothing. And then you send electrical signal again. And there is such a loop over and over. ([Time 0:30:10](https://share.snipd.com/snip/5109efad-9296-4168-b7a8-2901a5c6aca9)) - **Post-Silicon Heterogeneous Future** - Biocomputing will likely join a landscape of specialized post-silicon hardware rather than directly compete with quantum computing. - Diverse chips optimized for different tasks will coexist in future compute stacks. Transcript: Ewelina Kurtys Bioprocessor, very good. Or biocomputing, we call it biocomputing, bioprocessor. So no, I wouldn't say it's in competition because this is totally different mechanism, different things. So we know quantum computing actually is very fast and it can maybe be good for an encryption of information. So it is totally different type of task. So I'm not sure it will be in competition, but. Adam Stacoviak Okay. I was thinking more like one may win or one may actually prove to be fruitful in terms of viability. Ewelina Kurtys Well, that's kind of what I was thinking like. Actually, I think that the future will be that we will have very different type of hardware because generally you can see this kind of direction. Not only quantum, not only biocomputing, people are also working on many specific chips, also digital, which are optimized for some specific tasks. So I believe that we will have variety. So today we have mostly CPU, GPU, and in the future we will have hundreds or maybe of different chips, I believe so, which will be optimized on some specific task. ([Time 0:40:45](https://share.snipd.com/snip/ea8ab176-5b58-420d-a2b4-b299f966db37)) - **Early Demand From Academia And Industry** - Users include individual researchers, startups, and large R&D teams; universities use the platform for publishable research. - Industry clients are already paying for early remote access despite the technology being pre-production. Transcript: Kyle Galbraith This, like you said, you have these users all of a sudden, like which industries want this as a thing? Ewelina Kurtys I would say we have three types of users, individuals, fascinated engineers or, uh, um, small startups. Some of them want to do something related to biocomputing. Why they want to use our platform and big companies which have r&d teams very large companies which have r&d teams which want to do some project on cutting edge technology the same actually How people do with quantum they know that it doesn't work yet but they want to know what is going on they want to know how it works because they believe it will work in the future. Kyle Galbraith Okay. So that's cool. I mean, are they doing like, what kind of stuff are they trying to do? You don't have to give specific examples. Ewelina Kurtys No, actually this is confidential. What are doing our clients is confidential. Sure. But we have universities which are using our for free and they are going to publish. So actually that's why we chose them. We chose those who have the highest chance to publish. Sure. Kyle Galbraith That makes sense. Ewelina Kurtys And actually there will be some papers coming for what people are doing. So I hope everyone will be able to see, and we will be promoting this for sure. ([Time 0:43:34](https://share.snipd.com/snip/379c4398-9dea-4a72-a5ee-ab849a3fc0f0)) - **What The API Returns** - The API returns spike trains and spike shapes from electrodes, allowing users to analyze timing, frequency, and waveform features. - These outputs support signal processing and creative interpretations for experimenters. Transcript: Kyle Galbraith So we talked about what was at the neuron platform end, like, you know, a UV light turns on or some sort of electrode electrolyzes. What do I get back? Like I make a call. Is it like a one zero? Is it like a success fail? Is there more information coming back to me? Like, what do I get back at the other side so I can actually start mapping results or trying to make sense of it? Ewelina Kurtys Activity of neurons so this is what you can see also in our website on the live section so you get so the way how you can measure activity of neurons is a few different ways you can get a yes No response so this is spike trains so this kind of data you get just a dot and you know this was spike every time there was a spike you get a dot and there is already quite a lot you can analyze The patterns you know you can see if they're more active or less active this is actually the most common way how you collect the data and it's quite efficient also because you just have One dot one point on for each occurrence of the spike and very often it can be enough. But if you want to be more specific, you can also measure the shape of the spike because the spike means that the neuron will change the charge and this will always have a shape. And you can also analyze the shape of the spikes. So this is much more heavy data, but you can also get this. And of course, then you can have, you know, people try to have different way of interpreting the data. This is actually, it's a big room for creativity. For the moment, we look, for example, how late, what was the delay before we saw the signal, or we can see the distance between the signal, for example, how often a neuron is active. So we try to characterize all this, all these patterns on how they're active. So that's what you get. And then you can, yeah, you can do, there is a lot of signal processing, a lot of analysis of the data. Sure. And a lot of data. ([Time 0:44:58](https://share.snipd.com/snip/6f0daac8-6a30-4f1c-97c6-657f91eef767)) - **Centralize Bioprocessors For Stability** - Centralize biological servers to maintain strict physiological conditions like temperature, pH, sterility, and vibration control. - Running neurons in controlled data-center-like facilities improves stability and scalability over home setups. Transcript: Ewelina Kurtys Yes. So environment is very important. The neurons are very fragile and the environment has to be physiological. So the same as in our bodies. So there has to be physiological temperature. So there has to be, of course, always liquid around. So neurons are in the medium. So this is water with different substances, which keep them alive, which also feed them. And all this is very, very important, like pH, temperature, everything, even small vibration, everything is really important for the neurons to be stable. And it has to be very strict. Otherwise, the activity can change or they can die. And this is why also we believe in this bio servers, in the central servers idea, because we think that it will be easier to control these conditions of the neurons when they will be in Central server. So that's also the reason why we... ([Time 0:50:04](https://share.snipd.com/snip/1983792d-e500-45a9-b3ca-4403294c0543))