Биологическое вычисление: стартап ускоряет генерацию видео в 5 раз с помощью нейронных алгоритмов

My podcast guest this week Alex Ksendzovsky, co-founder of The Biological Computing Co. Alex and his team are doing something that sounds straight out of science fiction: they’re extracting computational principles from living neurons and translating them into software that runs on standard silicon. Alex and I chat about The Biological Computing Co. is bridging the gap between biological tissue and GPU-accelerated algorithms, explore how this approach is making foundational models faster and more efficient and where Alex sees this ground breaking technology headed in the future. Links for September 4, 2026 The Biological Computing Company Neural Optimizer (Compare open-source video models before and after TBC optimization) Blogs by the Biological Computing Company Click here to check out the Fish Fry Archive. Click here to subscribe to Fish Fry via Podbean Click here to get the Fish Fry RSS Feed Click here to subscribe to Fish Fry via Apple Podcasts Click here to subscribe to Fish Fry via Spotify Amelia’s Weekly Fish Fry – Episode 697 Publication date: September 4, 2026 Amelia Dalton: Hello there, everyone! Welcome to episode number 697 of this here electronic engineering podcast called Amelia’s Weekly Fish Fry, brought to you by EEJournal.com and written, produced, and hosted by yours truly, Amelia Dalton. In this week’s podcast, we’re diving into the future of AI infrastructure and why it might be more alive than you’d expect. [Laughter] We’ve heard all about AI scaling rapidly, but what if the next major leap in computational power doesn’t come from a new silicon process, but from biological principles? My guest this week is Alex Ksendzovsky, co-founder of the Biological Computing Company. Alex and his team are doing something that sounds straight out of science fiction. They’re extracting computational principles from living neurons and translating them into software that runs on standard silicon. In this week’s podcast, Alex and I break down how the Biological Computing Company is bridging the gap between biological tissue and GPU-accelerated algorithms. We also explore how this neuro-derived approach is making foundational models faster and more efficient, and where Alex sees this groundbreaking technology headed in the future. So, without further ado, please welcome Alex to Fish Fry. Hi, Alex. Thank you so much for joining me! Alex Ksendzovsky: Thank you very much for having me. I’m very excited. Amelia: Excellent! Okay, so Alex, let’s start with the fundamental shift your team is proposing. How does extracting computational principles from living neurons and translating them into software that runs on standard silicon change the way we should think about AI infrastructure? Alex: Yeah, absolutely. So, think about it this way. First of all, what we’re doing with our adapters is making AI algorithms and systems a little bit more biological. Our adapters are extracting neuroscience principles, as you mentioned, and data flows through them into the diffusion transformer, or any kind of transformer model, thus making the output a little bit more biological. What that does is, first, improve efficiency. It allows for much more efficient generation of video or generation of tokens, and so that’s a huge benefit. The second is performance. As I mentioned earlier, we are seeing significant improvements in things like how much video we can generate. The physics of the video is improved, for example, meaning the physical representation of objects. Those are things that your brain is very good at in terms of how you perceive the world. We’re able to apply that toward the models that we generate. Further on, in the context of real-time compute, we’re anticipating applying this toward things that current AI systems can’t do, such as continual learning. If you go to our blog, you’ll see our fourth blog post is about retraining of AI algorithms. A learning rule that we were able to extract from the neurons, applied toward a retraining task, allowed for a significant improvement in continual learning with a reduction in catastrophic forgetting. Going forward, with real-time compute and even more of the neuroscience principles we’re extracting, we can see significant advances in things like continual learning or out-of-distribution generalization. Amelia: All right. So, Alex, deriving computational principles from living tissues sounds like something straight out of science fiction. Talk to me about how this actually works. How do you get from neurons growing on an electrode array in a lab to algorithms running on a GPU? Alex: Absolutely. So, I’ll start from the beginning. We take brain cells. Those are brain cells either from rats, or we create them from human stem cells. We grow them on an array of electrodes. There are 4,996 electrodes, which gives us the resolution to be able to encode information, or send data into the biology, and decode the neural response, or how the neurons process that information. A toy example is a picture. A picture of a face, for example, has multiple features like eyebrows, a mouth, et cetera. We can create an electrical pattern that matches those features and send it into the brain cells via the electrodes in a stimulation pattern. That’s kind of step one: encoding information. And we do this across multiple modalities, in addition to images, including videos and language. Then, the neurons respond to that electrical stimulus. That neural response is very complex, rich, and dynamic. It includes where the neurons respond, so the space across the dish or the multi-electrode array, as we call it; when they respond, or temporally; and the patterns in which they respond. That’s the critical part. The pattern of neural response is how the information is processed. What that allows us to do is create a mathematical model of how neurons, or how the brain, processes an image or a language token, et cetera, and incorporate that as an adapter into a transformer model. Porting that over into, in our case, a diffusion transformer allows us to create longer videos that are higher quality and that abide by physics, for example, and things like that. Amelia: So, Alex, you said this approach makes generative models faster and more efficient. Where are you seeing the most significant gains? Is it primarily in inference speed, the quality of the generated output, or the cost and energy per generation? Alex: Absolutely. So, we’re seeing gains on both sides. Right now, we’re seeing our gains mostly in inference. I can talk about the training side a little bit in a second, but mostly we’re seeing gains in inference, and our first use case, as I mentioned, is in video generation. We’re seeing both performance gains and efficiency gains. On the performance side, we’re seeing an increase in the amount of video that you can generate. We’ve gone from a couple of seconds to minutes of video that you can generate. That video stays cohesive, and actually, one can interact with it in some of the models. In terms of interaction, we’re seeing improvements in action sensitivity. What I mean by that is, if you’re a little character in a Minecraft game that’s being generated, and you go forward, back, left, or right, the percentage of times that that action actually gets captured is improved. We’re seeing further improvements on the performance side in memory—things like the scene staying permanent, or object permanence. We’re seeing significant improvements there. And then, on the efficiency side, we look at this in terms of frames per second, but also cost per token on standard GPUs. We’re seeing at least a four- to five-times improvement in frames per second, or speed, and then a significant improvement in terms of cost per token or cost per frame to create the video. Amelia: So, moving from research to production is notoriously difficult, and building on top of biology adds a whole new set of variables. Alex, what are the biggest hurdles, both on the lab side, where you need neuronal cultures stable and productive enough to extract reliable principles, and on the software side, where those principles have to hold up as robust, scalable products in production? Alex: Yeah. You’ll be surprised, first, as to how robust the biology really is. In my career transition from academic research to now being in industry, it’s interesting how the focus starts to shift. When we first started the company four years ago, we really focused on some of the minutiae that are less interesting to academia, but are really important for maintaining our cultures, getting the best viable cultures—meaning neural cultures—and making sure they’re as production-ready as possible. So, we focused on that significantly. A number I’d like to share is that in my academic lab, when we were doing a similar process, we would get maybe 5% to 10% viability, meaning the amount of cells that stayed alive was about 5% to 10%. Now, it’s closer to 80% to 95%. So, from the robustness of the neural culture side, that’s certainly something that we’ve been able to really hone in on and tackle. I think the difficulty is really building something new constantly. We’ve developed a process, which I’m quite proud of, that is a very fast innovation cycle. What I mean by that is our neuroscientists and engineers are literally discovering something new every couple of weeks. A lot of what they discover can be a massive paper and can lead to a lot of efforts. So, I think the hard part on our side, given that we’re creating and discovering so much novel information, is determining what we hang on to. What is the thing that’s going to lead to an improved adapter or an improvement in our model? I think just creating too much and discovering too much—and going through what’s important—is the difficulty. Amelia: That makes sense. Now, Alex, looking at the next few years, where do you see the first real-world deployments? Are we talking about large-scale data centers and cloud serving, or will this eventually find a home in edge computing applications as well? Alex: Yeah. So, just to be clear, we are deploying in the next couple of months. Right now, the product that we are creating and deploying is software. Our software includes open-source models that have our biologically derived adapters. I think looking toward the future, the concept of what we call real-time biological computing comes into play, and I think that’s what you’re referring to. What that means is using the actual dish of neurons themselves as part of the compute circuit. So, rather than extracting neuroscience principles from them, it’s actually using the neurons themselves as the compute processor. That’s certainly, in my mind, from a commercial perspective, probably five to 10 years away. Now, we have tons of data suggesting that we can predict the next token and that we can do compute in real time, but it’s not a commercial product. That is something that we envision being deployed both at the edge and across data centers, where compute is distributed to biology and to standard silicon. But again, that’s probably five to 10 years away from a production perspective. Amelia: Fantastic! All right, Alex, it’s time for your off-the-cuff question. If you could have one meal right now—it doesn’t matter if it’s on the other side of the world and you need a passport to get there—what would you have? Alex: Anything with tomatoes. So, gazpacho, pan con tomate—anything with tomatoes, I’m there for it. So, that’s what I would have. Amelia: I love it! All right, well, this was super cool. Thank you so much for joining me, Alex. Alex: Thank you. Amelia: If you’d like even more information about the Biological Computing Company, I’ve included a couple of links below the player on this week’s Fish Fryin’ page on EEJournal.com and in the description for this week’s YouTube episode as well. Now, one of those links will direct you to a series of blogs that I found especially interesting, including Biological Neural Dynamics for Computer Vision, Using Neural Connectivity Patterns to Improve Image Reconstruction Quality, Bridging Biology and Silicon: The Neural Dynamics Adapter, and the blog that Alex mentioned in his interview called Beyond Weight: Recycling Algorithm Discovery for Sustained Plasticity in Neural Networks. Hey, have you checked out EE Journal on social media yet? Well, you should! You can find us at Facebook.com/EEJournal. If LinkedIn is more your thing, I dig it. You can follow us—or me—on LinkedIn. And we are also on Bluesky Social and Mastodon as well. And we have a YouTube channel, like I just mentioned: YouTube.com/EEJournal. Folks, it is chock-full of all kinds of techie videos, including our very popular Chalk Talk webcast series hosted by me. And, of course, you can subscribe to our EE Journal YouTube channel as well. Thank you, everyone, for tuning in. If you know of any cool new technology—or heck, you just want to chat—shoot me a line at Amelia, that’s A-M-E-L-I-A, at EEJournal.com, or post a comment on our forums on EE Journal. For the week of September 4th, 2026, I’m Amelia Dalton, and you’ve been fried!