生きたニューロンでAIは5倍速くなるか？ Biological Computing Company の挑戦

We have a robot strapped up to brain cells performing tests. It's a proof of concept. There's a ways to go. It is able to define or tell you when you raise your left hand versus your right hand and if it's consistent. Language and language models, it's a very small sliver, a very small abstraction of the amount of physical data and information in the world. Most information around us is not language. It's sensory information. It's multimodal. Moving down language, you talk about 3D spatial reasoning, cause and effect, you know, physical relationships. The whole thesis of the company is that all of these things, including the energy cost, including continual learning, including being able to create novel things, the brain does extremely well. If you look at language models, they're super performant. I mean, like, they're incredible, right? And you think about just models in general, things like continual learning still get to be desired, right? So there's this question of, like, generalizability, right? Like, are these models, can they actually create something outside of these different trends? Welcome, humans, to the Neuron AI Explained.

I'm Corey Knowles, joined as always by Grant Harvey. How's it going, Grant? Doing well, doing well. I'm very excited to have this chat because today we are talking about a company that is studying real brain cells to find better ways for AI to compute. So we want to understand what happens in the lab, how that becomes working software, and what people can actually use today. And to have this conversation, we are talking to Alex Zandowski, MD, PhD, CEO, and co-founder of the Biological Computing Company, and John Pomerantz, MD, MBA, and the COO and co-founder of the Biological Computing Company. Before we get started, I'd like to give a quick shout-out to the sponsor of today's video, MerCore, a leader in AI data production, evaluation, and model training. We'll hear more from them in a little bit. Also, please take just a quick second to like and subscribe before we get started. And with that out of the way, Alex, John, welcome to the Neuron. Thank you. Excited to hear. Excellent. It's so great to have you guys. We're big fans of what you do. It's been super interesting to watch and learn about, and we're excited to pick your brains today.

Great. No pun intended. No pun intended. I was thinking it, too. I was like, oh, this is great. You both came from neurosurgery, neuroscience. What made you decide that living neurons could help us build better AI? And what was the first result that made you take the idea seriously? Great question. I can start. I think, John, I usually have slightly different versions of the story, but we kind of convert to the middle. So, you know, for me, this dates back to college. So I was working in very first labs. I grew brain cells on electrodes. And at the time, it was, you know, for whatever reason, obvious to me that we should use that to predict the stock market. That didn't go well with my PI at the time. And so I ended up kind of moving on to something else. And throughout the course of my career, kind of had in the back of my mind this idea that using brain cells for compute, or at least to understand how brain cells compute, is going to be fundamentally important.

So fast forward to, and John can kind of lay out how we met one another through a lot of other work and neurosurgery. But ultimately in 2022, as we watched the convergence of multiple technologies, including biology, AI, computational tools, and the hardware to grow brain cells and electrodes, we realized that now you can actually do this crazy idea that we have in IAV 2006. And that was the inception of the company. Maybe, John, if you want to talk a little bit on your end and we can talk about then, like, the first result that to us was, whoa, this is actually real. Yeah, yeah. It feels like yesterday, but it's been a long time. Yeah. We're both neurosurgeons and neuroscientists, and we met a long time ago. Alex, the number keeps going up, honestly, with every time we talk to somebody. I think it was about 17 years ago that we met. and it was very obvious you know the version I like to tell it was very obvious that what bound us immediately was sort of this outsized ambition like how do we push the envelope

like into the frontier and in everything in life with research with you know with building companies with the former company together and as surgeons and as scientists what we realized like there was a huge huge mismatch between what was available, having access to the real brain and how it computes, versus the engineering tools in the real world. And so a little over four years ago, when we started this company, we wanted to be the first people in the world to commercialize this in a real way. How did we take this out of the lab, out of academia, out of drug discovery, out of biotech, into the real world to solve computational problems? And that's exactly what we did. We spent a couple of years in a small lab, a few people, building a communication layer, an actual language to communicate with what we believe is the best computer ever architected, invented, evolved over hundreds of millions of years, and how do we interact with it in a meaningful way?

How do we send information from this world into a neuron world? How do we understand how they're communicating with each other, and then how do we take those principles out and do something meaningful with it? Alice, you can tell the version of the early proof point. And I think at Early Proofpoint, you know, one of our first investors and investing all along the way is Jeff Dean, former head of Google AI. That's a legit guy. Yeah, yeah. And it was for sure as legit as they come. Yeah, for real. And we had an awesome conversation with him very early on, and he really helped us kind of define early tech milestones and how do we match that kind of commercially. And as scientists, we wanted to build this out in a scientifically rigorous way and really believe that the science is real. And the first time that we did, and then Jeff kind of bought in, that was the kind of aha moment that, like, we have to do this. Like, this has to be the thing that we do with our lives,

and it's honestly been exciting and uphill downhill, not how you put it, but like a lot of momentum, you know, since then. That's so awesome. So AI models are getting better fast, but how much better are they at doing actual professional work? That's what Mercore's AI Productivity Index, or APEX, is built to measure. Mercore is a leader in AI data production, evaluation, and model training, and APEX tracks how frontier models perform on long horizon tasks across fields like corporate law, consulting, investment banking, accounting, medicine, coding, and more. The progress is pretty wild. In January 2026, Frontier Models could complete just 25% of Mercore's agentic tests. Today, that number is nearly 70%. And Apex updates with every major Frontier Model release, so you can see how the newest models stack up, compare token costs, and explore results from third-party benchmarks like Browse Comp and Terminal Bench. See where the frontier actually stands.

Visit Mercor.com slash Apex, that's M-E-R-C-O-R dot com slash A-P-E-X, and explore the rankings today. Now back to our interview. So the proof point, yeah. So just before I get into that, here's what an array looks like. I don't know if I think this is recorded, so the audience will get to see this. Yeah. In the middle is a grid, and there's 4,096 electrodes. I mean, about 150,000 cells living on that, right? So when we started the company, we were like, okay, well, task number one, encode information, right? What information? How do you encode it? What does that actually mean, right? Taking a step back, just to, you know, pick your brains for a second, right? What's actually happening in your brain right now? You're looking at the pixels in front of you. Those pixels are getting converted by your retina into electrical signals, which propagate through the back of your brain, and then your neurons will process them and have a certain pattern of firing

based on what John looks like, what I look like, et cetera, right? When you have brain cells in a dish, you don't have the retina, right? And so when John was talking about creating the language layer to encoding information, it's how do you go from pixels that are in front of you, like my face, to some kind of electrical pattern that the brain cells can recognize. So from here to here, but without everything in between. Exactly, right. And how do we do it in a controlled way? So, like, how do we take this environment and control it in a way that we can understand the neural processing that happens? So the very first experiment was saying, okay, well, let's go back a little bit to the history of AI, and let's go on with, like, okay, what were the first data sets that showed deep learning was useful? MNIST, right? And as it turns out, it's also a very practical data set because it's spatially arranged. Meaning like you draw a five and my five versus your five is a little bit different because mine is a little crooked or whatever, right?

It's spatially arranged. What we have direct control over in this dish is how we spatially arrange information, right? So we have 4,096 electrodes. We can literally draw a number five onto those electrodes, right, by sending current to the electrodes that are in the shape of a five. And that's what we did. So we started with MNIST. We sent the biology zeros, ones, et cetera, et cetera, right? And we asked the question, which was, first of all, we were, like, happy if we got any kind of neural response in any sense, right? Yeah. But the idea was, can we take that neural response or how the neurons represent the shape of a five? Because it gets transformed in a completely different data set. It gets transformed into a time series of spikes. The idea is it's much more rich and dynamic and complex. Can we take that much more complex data set and do something useful with it? And by that, we mean going back to the original task, which was classified numbers.

And so the very first experiment we set up was we said, okay, let's take a bunch of enormous numbers. let's train a very simple classifier. So simple to, we're not doing like, you know, I know it's been B, you know, it's been B for many years now in terms of like classifying humanists, but we don't want it to be. We want to see is there an overhead we can, an improvement we can make on a very simple classifier. So take that classifier, train it against just the numbers themselves as one did in the, you know, late 90s or late 2000s, and compare it to the exact same parameter count, exact same classifier trained on the neural representations of that. So a vector that comes from the neurons that we train the classifier. And is there a difference? Is there a classification accuracy difference? As it turns out, not only were we able to classify immunists based on that neural representation, we actually were about 10% higher in classification accuracy than the classifier trained on the numbers. So what that suggested to us was there's some special features of this neural response that this convolutional neural network is grabbing onto and allowing us to improve classification accuracy.

Something special happening in this dynamic rich data set coming from the brain cells that's allowing off-the-shelf artificial neural networks to improve in classification. That was the aha moment That was the like oh crap there something here and we can work off of that Okay so I want to go back to something you said and I want to make sure I understanding it correctly So are you saying that the neurons on this chip or the cells on this chip, in the same way that computers, you know, they work with transistors, and, you know, that gets converted into binary, assembly code, all that stuff, that it's equivalent of binary or transistor, and you can correct me, is the spike data? Is that what you're saying? Is the neural spike data is kind of like the binary that gets converted into? I would think about it as it's an analog computer essentially, right? So I would think about the brain cells as there's a pattern that's associated, a spike pattern across space and time that, you know,

for an independent neural culture is associated with a shape of an image we send to it. And it's consistent every single time. The thing that defines what that pattern is is how the brain cells are connected to one another, right? And, again, we can talk about how the setup, they're different every single time, but the classifier pulls out those features that define, in space and time, a number, right? So that's point number one. Point number two, and this is actually what distinguishes it significantly from transistors from silicon, is they can actually learn. So the way the brain cells are connected to one another, which defines that neural response or that pattern of its attendant number, changes over time based on your teaching it. You can actually teach the biological numbers. When you're thinking about this in terms of applications, you're thinking about this in terms of, I want to build the best immunosclassifier in the world, and I want to use brain cells to do it. Well, you set it up like we did, which is, okay, you encode the number into the brain cells. and then transform the number from pixel space into neural space.

Neural space has a lot more features that you can then train an artificial neural network to grab onto those features, classify more accurately, and then you can, in a closed-loop way, punish or reward the biology to change the connectivity profile to the point where that final neural representation is even better to get higher classification accuracy. So that's kind of the way we think about it. But obviously, and I don't want to, John, I can see that face. I'm getting a little bit into the weeds right now. My filter is coming up. Don't worry. That's that face. But I do want to kind of bring you back to, like, what we've been able to build now, which is all of these experiments that we've done over the last four years have really given us insights. And this is what we're leveraging to build our current product, which is now out in the beta version of the demo. So, you know, all of these insights of how brain cells process pixels, images, videos, individual stem patterns, all of the different dynamics that happen in the dish is just insight into how the brain cells process information.

And all of those things can be mathematically modeled. And the kind of like we had like the scientific aha moment, which was the MNIST stuff. the business and product aha moment was, wow, we have a lot of this data and we can actually put this into code and apply this towards our first use case in video generation. Yeah. So when we did that, we saw this, what you guys probably saw in our demo, which is a video generation product that taps into these massive efficiency gains that biological networks have. Again, all in code, all in silicon still, and gives us five-fold reduction in latency time, 80% reduction in cost, and all these benefits. And this is, by the way, just like internally we're higher than this already, but like just the ground floor. Yeah. How is that benchmark against like a video model that runs on an NVIDIA GPU or something like that? So just to be clear, it is on an NVIDIA GPU.

Oh, okay. So it's taking the mathematical principles from how brain cells. So, like, the dish is just a research platform. Yeah, got it. We take these mathematical principles, apply them in code, and the five-fold efficiency gains are actually on GPUs. On GPUs. Direct apples to apples comparison. Yeah. I was kind of imagining it running on the brain cells. That is version, that is, you know, a goal of the company. The North Star is what we call real-time biological computing, and we have some very interesting announcements, not today, but at some very soon point in the future to talk about that. But right now we're productizing the insights from the dish. I just want to stay in the thread a little bit because Alex got very technical, and I want to make sure that we're like high level. Zoom it out. Kind of my role. Yeah, yeah. You know, at a high level, the brain processes information very differently than silicon does, right?

And so, you know, from the beginning, even before starting this company, this is something we started to notice, right? Having access to the real brain, recording from the real brain, understanding how the brain processes memory and learning and intuition and knowledge and things like that. And so at a very high level, what we are doing is we are tapping into those natural, evolutionarily built computational advantages of the brain, right? And so we can send information in, we can record how neurons respond, and then we can do something with those responses. And what Alex is talking about now is how do we make current AI workflows, current AI models a little bit more biological? Like, how do we improve the compute in those models and those architectures just a little bit to get some of the, even just a small fraction of the massive efficiency gains that the brain naturally has? Yeah. And what we know is that head-to-head, and it's, like, very hard to do this, but, like, purely head-to-head, if you were to do, like, the energy cost of neurons,

or, like, a synapse versus energy cost of, like, silicon, it's billions fold different. Yeah. Like, it's crazy, right? So this 5 to 10x that we're getting of faster and cheaper for video generation, like, that's literally the tip of the iceberg. And so the way that we think about it is, again, at a high level, it's not a single product that we're scaling. It's a whole discovery platform that we have. It's an engine. It's biology. It's like we're literally returning to biology to take these insights and make silicon-based processing better, better, faster, cheaper. So the more that we do, the more that we experiment, the more that we interact with this perfect computer, this dish of neurons, the more that we understand, the more that we learn, and the more that we can make these models better. So your agents can talk to each other, but they still can't think together. That's the gap OutShift by Cisco is closing. OutShift is Cisco's incubation engine, building the Internet of Cognition, shared intent, shared memory, and guardrails that allow agents to work together from anywhere.

Open source protocols and code, a system for intelligence that scales. Read the paper and experience the demo at outshift.com. That's O-U-T-S-H-I-F-T dot com. Check out the links in the description. Now back to our video. Not even just that. I feel like along the way you're learning an awful lot more about the brain and biology. Like, I think there's a lot of stuff you're going to uncover that will go back the other way in the platform. Not in the platform, but in terms of, like, benefit science, even, it seems like. Yeah, 100%. You know, we agree with you. You know, we're very product-driven because we're a company, right? Yeah, of course. But we do have collaborations in academia. You know, a lot of us, as you probably saw on our website, have come from academia and, you know, certainly appreciate that world. And so to that end, you know, we are actively, you know, have collaborations and pursuing collaborations with academic labs who want to ask those questions like, you know,

how do we use the insights that we're getting not just for product development, but for, like, to understand more about neural process. I 100% agree with you. A lot of the – we think of value creation through a lot of it is knowledge and discovery. It's not just conventional like, hey, we're going to build a product and have everybody in the world use it and monetize it. There is a lot, like you mentioned, that we're learning along the way. And from the beginning, we always wanted to balance the practicality of building products and commercializing, which we're doing. This is not a science experiment. This is happening right now in our lab. We're shipping literally within a few weeks with this sort of like deep exploratory spirit of where the company was founded, which is like, and going back to our original comment, like how do we push the limits? Like what are the limits of this platform? Like it's not, it's much more than the new generation, right? There's much more there. And so that's the exciting part. Well, and I'm sure you saw the fly brain map that went viral recently, like a couple weeks ago.

Everyone was using it, the digital representation of the fly brain, like for all sorts of wild and weird things. It got memed out pretty good for a while. It turned into a total meme. But famously, it's a really difficult challenge to create the same type of map for the human brain. So I have to think that as you're doing all these experiments and as you're learning, that you might have a pretty good shot here to be able to create something like that. I mean, what are your thoughts on that? I think those are two independent approaches towards a similar theme. So, like, and it's all about structure versus function, right? Yeah. What the connectome work does, which I think is awesome, right, you know, it gives us, like, what's the structure of a biological neural network as it is in a fly or, you know, up to the human brain, as you're suggesting, right? So what can we infer computationally from that structure? the bet that we're placing and again in the end it's going to be like understanding all this aggregate together

the bet that we're placing is rather than inferring just from structure it's what can we infer from the actual function of brain cells as they compute in time so like if you think about a connectome that's like frozen in time you have a single shot or single snapshot right and so there's a lot you can understand from it but I think it's missing a lot of terms in there are transmitter release etc all of the dynamics that happen in a living brain. Yeah. What we have access to is all of those dynamics. So, like, for example, the spikes, the neural spikes, that's how information is represented in a brain in time. There's a lot of processing. In fact, some would argue most processing happens temporally in the brain across time. And so I think the connect-com technology is amazing. I think that it has to be paired with neurophysiology. It has to be paired with some of the things that we're identifying. But, yeah, I mean, I think that as the tools get better and the technologies converge,

it's going to be super powerful and kind of like bridging that gap. Like the gap right now is five-fold efficiency gain that we're seeing just by observing and modeling some of the computational processing of living neurons to like the billion that John was talking about How do you close that gap It going to require understanding structure function or physiology et cetera Is that like monitoring you know activity over time and just doing enough time modeling to really understand it? Is it making it bigger? Like, what do you think that, you know, without giving away secret sauce, like, what's some of the ways that you can do that? Yeah, I mean, I think for, you know, for us, it's taking very discrete experiments, like, for example, all the way to as simple as, like, a pixel, right? Like, how is a pixel represented to brain cells spatially and temporally in many, many of these biological networks? And what's consistent across all of these biological networks? And if it's consistent, that potentially is a fundamental truth of neural processing, right?

Right. Now, I think, like, to go a layer further, it's like, okay, well, we have, like, this should brain cells, but, like, how does this compare to what some of the literature has shown happens in animal brains or in intact human brains, right? So, you know, John alluded to this earlier. We didn't really get into it, but our background as neurosurgeons are ones that implant electrodes to monitor for epilepsy. Again, it's the same. It's, you know, similar in a sense, like, the data we're collecting from patients with epilepsy when they're, you know, moving, doing tasks, et cetera, is the same data we're getting from this dish, right? It spikes across time, right? And so, yeah, I mean, so it gives you that kind of purview into it and allows you to really understand holistically in the very specific guided experiments that we have in the dish, plus what happens in the literature when people observe patients with electrodes when they see things, like what happens in the back of the brain, for example, and what's consisting across those two independent models. Right, right. Well, you know, I feel like, and I'm thinking of this from the angle of someone who's never heard the phrase biological computing.

And I feel like they might picture, you know, brain cells in a dish or something. And I'm wondering, at its simplest core, what is the easiest way to explain that connection? I think that, and, you know, one of the reasons we named ourselves the biological computing company is because we wanted to really be the company that defines what this means, right? And just like any other technology, I think it's got to be defined based on what it's applied towards, right? And so, again, when we started the company, it was like, how do you apply this to your brain cells? That was the first task, right? How do you apply this to your brain cells towards anything? And so I think it's a bit of an evolving idea. But right now, it's, you know, how do you take brain cells in addition to understand how – use that as a platform to understand neural processing and apply it towards a real-world use case? Yeah. That's evolving towards this additional world, which is, again, what we call real-time biological compute, which is, you know, actually using, like you're suggesting, brain cells in the compute circuit or for compute itself.

And, you know, that's not, like, that far off. I mean, again, from a product perspective, you know, to scale it, certainly it's not now, and it will be probably five to ten years before that happens. But even, you know, just behind me in our wet lab, we have a robot strapped up to brain cells performing tasks. Let's go. That was one of the things I wanted to ask you. Wow. We have it, right? So, again, like, it's a proof of concept, and there's, you know, there's a ways to go. But it is, you know, it is able to define or tell you when you raise your left hand versus your right hand, and if it's consistent, right? And so, like, it's solving a pretty fundamental, explicit core problem. So, like, again, that's happening. But in order to get to that point as a company, we have to be able to show utility in the market, which is what we're doing now, and then apply those findings as a foundational layer towards the next step, which is real-time biological computing. And I assume the ability to repeat it, make it economically viable, make it ensure consistency to some level that would get in there.

And I assume there's a lot of product and market angles you have to juggle, I assume, as well in this. And sorry for interrupting you, John. That was unintentional. No, no, that's all good. I think what I was going to say is actually kind of getting to your question, actually. So maybe we'll go there. All of this work builds on itself. Like, it's very mutually reinforcing, right? So the more that we are commercializing and productizing now with optimizing AI inference for video generation models, it's the same core technology, the same platform, the same sending information in recording and doing something with that information as output. The same thing applies towards this longer-term vision, which is what Alex is alluding to, which is real-time inference and neurons being part of the inference loop. And so going back to product and marketing business development, for sure, and this is all by design, by the way, right? This is like we are building, and we are 100% convinced of this, we are building a lifetime generationally defining company.

It's something that's going to outlive us and outlast us, and it's by design. But in order to get there, you kind of have to earn that right. You have to earn the right to build that future, right? Like, we had the chance and chose against going down this path of, like, a little bit like the quantum path, which is like, hey, give us a billion dollars in 10 years and maybe we'll give you something that will change the world, right? Yeah. Like, we wanted to de-risk that moonshot and say, how do we do it in an iterative, practical, scientifically rigorous way that is just continually building on itself? And what we're finding is it is a good thing to be able to plug into existing AI workflows from a product and marketing and business development standpoint, right? Like get buy-in, start to change the world in like iterative small steps, right? Yeah. And then all of that core technology, all that core R&D is actually leading towards this bigger future, which is what we're alluding to. That's awesome. So does your model have, like, a different architecture fundamentally than, like, let's say, like, the video models we're used to?

Like, are you just able to get these efficiency gains from the training? Or they're, like, diffusion or something. Yeah, how should we think about it in, like, the landscape that we know today? Yeah. Yeah, so it's a diffusion backbone. It's a DIT backbone. And that's open source. And so I think what's critical there is the fact that we can build this into, and we have, we've built this into Oasis, Cosmos 2.5, Cosmos 3, now other ones. And so like, you know, WAN and others. I mean, so we can do this across, you know, any DIT backbone. What it is, what the code is, again, it's an additional layer. That layer gets trained by videos, et cetera. And it plugs in somewhere in the diffusion stack. So, like, the somewhere is, you know, not me being vague. that's actually a lot of what our AI development team does is understand, like, where exactly in the transform architecture is the best and most optimal place to plug in our model that comes from the brain cells. Yeah. Another thing to note is the model that comes from the brain cells is not just one.

We have a leaderboard of 30 to 50 of these. We call them adapters. And we're constantly iterating time, right? So constantly redefining them, parameterizing them from the measurements from the brain cells, right? You know, comparing them against literature, et cetera, right? So, like, we're constantly optimizing and improving and seeing which ones improve our output, which in the case of video generation is, like, PI bench, feed bench, et cetera, right? Ultimately, user testing and quality, right? So that's the way we do it. And, again, I think just to underscore the importance is it's transferable across models. And what's nice for us, you know, probably call this luck, Like, what's lucky for us is open source, at least in VideoGen, have caught up to the closed source models. For sure, yeah. And we can get those weights, and we can incorporate our adapters. So, like, we've gotten our process to spin up an open source model, incorporate the adapters to, like, very little time. And so a new model comes up, we pull it out, and quality is even better, right?

And so we can leverage the gains that other people are getting and improve even further on theirs with our efficiency optimization strategy. That's awesome. So we already talked about the numbers. Five times faster generation, 80% lower inference cost. Does this enable you to run this with a lot less compute at this point? Yes. That's the idea. That's where the cost savings come from. So if you're paying per GPU hour, which we are, right, And so it's costing us, you know, fivefold, you know, again, internally more, less to run. And so we can provide that cost savings to customers. Awesome. Yeah. Yeah. That's great. Where you're either running less GPUs or getting way more for the same money, essentially. Exactly. And the other metric of latency, right? And so, like, we, you know, it's funny, John, again, because he's a scientist, but now we're in this creatives world, which is awesome. you know uh you know again this is why my sister created that that painting not me right it's like she's the creative in the least of my family and so like the painting behind john for listening

behind john the glass behind me which is thick um you know it's we're in this interesting world with creatives right and we're getting a sense of how they use uh these models to generate videos commercials etc right and so things like latency is extremely important like if you are an actual studio and you're generating 100 five-second clips just to choose the best one, and you're taking one to two minutes, and it's costing you like 50 cents to a dollar per clip, that's significant, right? Yeah. We've now shaved off a significant amount of cost and probably even more importantly, time. So now you have to wait 20 seconds for all those clips, right? And then you can get our ingredients. It gives you the freedom to build more, to build more quickly. you know someone we were recently talking to said like it achieves it allows you to achieve the flow state more quickly in creation which is which is an interesting concept from surgery like it resonated for me because when you're operating for a long time you get into this flow state you

kind of like come out eight hours later like whoa that was cool right similarly there that lives in the creative process which was really nice to hear the work contributing to that and and again And this all goes back to plugging into existing workflows, right? Like, nothing biological is shipping out of the lab, right? Like, for some people, like, the biology doesn't even matter, right? Like, that's the source. That's where all of these gains are directly derived from, is neuron-based processing. But for an end user, like, that kind of doesn't matter. What matters is what Alex was talking about, which is, like, what is unlocked by five times, I mean, by the way, internally we're showing ten times, but, like, five times faster, five times cheaper. Yeah. Like, it's not, like, it's time. Like, literally, it's, like, creative time. It's creative freedom. It's, like, being able to do more with less, right? It's, like, less wasted pre-production processing, for example. So, like, what's nice is that, like, there's real tangible economic impact right now, right,

that people can see directly right now. Wow. I have a question that may seem very basic, but how does one get brain cells? Yeah So you know it funny what available for purchase these days yeah um you know we are uh um we have an animal protocol we an accredited research lab you know i've been doing things like this for many years right and so you know you can buy it from um you can purchase stem cells you can purchase brain cells uh and they come to you and you spin them up and that's that's the easy way to do it You can harvest them from animals as well. And so all of those techniques are available. Yeah, so you all just did this deal with AWS here recently. Can you tell us kind of what that enables for you? Yeah, so we're really excited to partner with Amazon in a couple of different ways. So first of all, you know, we want people to use our model, right?

So it's going to be available on AWS Marketplace. Enterprise, customers, others can go and just use our mom, right? Bedrock, I assume, probably? SageMaker. It's going to be on SageMaker. SageMaker. Got it. Which is great, right? So that's an easy way. I think another extremely interesting collaboration and partnership with Amazon was via Tranium, right? to this idea that, like, you know, we were able to show that on standard GPUs, we have significant efficiency gain. But given that at base, right, Tranium is more efficient already than H100, for example, right, there's the question of, like, can you, with this very specialized architecture, can you incorporate our base plus, you know, one of our adapters is a recurring neural network, for example. So, like, can you incorporate that into or port that over onto our Tranium chip? Yeah. And we did that. So we did that. We showed that generation time or efficiency is on par.

Well, generation time is on par, but given the already more efficient nature of Tranium, now it's even more efficient in terms of generating video clips, right? Yeah. Those are pretty huge on lock, and so, you know, we're working towards eventually serving our models on Tranium to get even more efficiency out of it, right? I mean, I think, you know, showing more and more use cases on specialized hardware is going to be important outside of just, like, your standard transformers. And that's been a big part of our partnership with Amazon as well. I'd assume on Tranium, too, that you're getting to a point of, like, extremely low latency between diffusion tech, for starters, which is lightning fast in itself, combined with faster chips, combined with the biological approach. Where is the, like, over a longer term, where do you think that win matters? like especially in latency and cost? Like where do you think the – I mean at some point – and we already have some data to show this internally at like 4D videos.

At some point it's in real time, right? Yeah. At some point and you can start interacting with and generating videos in real time. And, you know, with our technology plus standard – plus a lot of optimization strategies, you know, plus on Trinium for sure that's already happening. And then all the applications that pin up with that. So that's where we're heading, and we're not very far from it, honestly, at this point. It's amazing. Yeah, I mentioned the robot earlier, so I've got to go back to that. Would there be a, you know, realistic use case for having real-time generation on board a robot, like kind of giving a robot vision or giving it real-time intelligence? Like what's the – because of the real-time – I'm trying to combine this thread of the real-time and robotics and think about what that could enable. There's an intermediary step that we didn't talk about in robotics, which is world models, right? Or like the policy that's trying to control the robot's actions

that are trained on some kind of simulation in the world. I don't want to dive too far into it, but the bottom line from our perspective is, I said this earlier, we've incorporated our technology into Cosmos 3. Cosmos 3 is, as you guys know, and videos will actually have world model and apply towards robotics. And so we, in our lab, have shown that we can actually take our adapter technology, improve the simulation of a world that a robotic arm lives in, and make it actually more accurate with a hallucination. And, again, same technology that improves video generation speed now ports over to a completely different, you know, additional customer base, which is physical AI, right? Yeah. And so that's going to be, you know, kind of the next step that we're working on, you know, in terms of our next market segment. But then that kind of perfectly segues into the, you know, in two to three years when we have a robot that's at least part of the compute is offloaded to real brain cells. And so the idea there is not necessarily, like, to give a robot vision or something super concrete.

the idea there is can we use neural processing to offload from an energy perspective processing that otherwise happens on board, right, at the edge. So that's the big picture there. That's the idea there. So it's not like a specialized thing that we're giving it. But it's more of a more efficient way to compute the same information. And representing the world more in the way that the world actually exists, right? And what I mean by that is like Like, language and, like, language models, it's a very small sliver, a very small abstraction of the amount of physical data and information in the world, right? Most information around us is not language, right? It's sensory information, right? It's multimodal. And, like, what computational tool has proved over and over again to be able to handle all this kind of multimodal different types of information with extreme energy efficiency? And it's the brain, and it's obvious. Yes. So moving beyond language, you talk about understanding the data the way that it actually is represented in the physical world,

and that's like 3D spatial reasoning, cause and effect, physical relationships. And there's so many things that you can unlock. Alex, I'm going to steal your word, but it's like intuition. I think a lot about that word. And intuition for what reason? And intuition for the reason, for the same reason that we're going back to this concept of flow and creativity and being able to do more. It's technology to help people do more, right? It's not to, like, make robots more human or give them, like, human traits. It's like, so what? The so what is, like, it's actually doing more in a world the way that the world actually exists. Yeah. Wow. Totally. do you have any intuition from all speaking of from all of your research that you've been doing and you know your your studying of the brain whether or not we we have what we need right now with all of the models you know the model different architectures that were that are out there um do

we have all of the pieces that we just need to put them together the right way or do you think we're still missing some kind of fundamental capabilities that, you know, the brain does that we just can't imagine that could get us closer to that, you know, 3D spatial model, like that perfect 3D spatial model. You could call it an intuition model if you want. Do you think we have what we need or are we still missing something? I think it kind of depends on what that thing is, right? It's kind of like, if you look at language models are super performant. I mean, like, they're incredible, right? And, you know, and I think, you know, no one will argue now that they're extremely useful and they, you know, don't, you know, maybe hallucinate sometimes, but not that much, or they're extremely performant, right? But if you look at them in the context of how much, like, data energy it takes to train them or how much it takes to use them, it is way far from perfect. there's a lot left to be desired

in terms of language models then you go into video generation so like you know obviously it's been trailing in terms of like allocation of transformer or DAT at least right it's been trailing so like they're becoming more and more performing every year but like you can generate maybe 5, 15, 30 seconds of video right? Yeah. In the context of like what makes world models special is like objects behind you disappear and change right so object permanence kind of like what John was talking about things like memory and, you know, these things don't, you know, aren't great, right? You think about just models in general, things like continual learning is still get to be desired, right? So there's this question of, like, generalizability, right? Like, are these models, can they actually create something outside of the data? So it depends on what is, right? If the what is doing research for something online, like, they're perfect. They're great. The amount of energy it takes is yet to be desired. There's other, you know, in terms of you want them to learn alongside you, they're not quite, they're not doing that yet, right? Yeah. And the whole thesis of the company is that all of these things, including the energy cost, including continual learning, including being able to create novel things, the brain does extremely well.

So I think my short answer to you is for sure there's a lot yet to be learned, right? And there's a lot yet to be applied. Our bet is it comes from the brain. And the only or at least the beginning of how to do that is by growing your brain in the dish and understanding at least this model right now. Hey there, humans. Quick heads up. The Neuron is going IRL in San Francisco on November 18th, starting at 4.30 p.m. If you're in town, come hang out with us. We'll have drinks, a live recording of the Neuron podcast, special guests, and a whole bunch of AI people together in one room. Huge thanks to Slack and Alumni Ventures for helping us bring the night to life. It's free, but space is limited. So scan the QR code on screen or hit the link in the description to RSVP and save your spot. We'll see you there. Wow. Totally. So one last question, and that is suppose a creator, a company, an individual wants to go try your model out.

Where can they do that? What can they expect? go to app.tbc.co that is where our model lives, that's a demo of it you can try it for free create some videos and the whole idea of this demo is to get user feedback as I mentioned we're going to be launching the marketplace soon and what we're doing is we're aggregating all of this feedback and we're tweaking the model so we want to make it more photorealistic or something else It's like we want feedback from people to tell us what to do, and so version one, as it comes out, is going to incorporate that feedback. That's awesome. That's awesome. Well, John, Alex, thank you so much for joining us. This has been fascinating. Yeah, so interesting what you all are doing. Thanks for joining us. Thanks, guys. Thanks so much. All right. Excellent. I'd like to give one more quick shout-out to our sponsor, today's video, Mercore. If you haven't yet, please take just a moment to like and subscribe. but that's all we have for today everyone

so farewell for now humans