Latest / The Adults In The AI Room / Earley AI Podcast - Episode 97: Biological Computing, Brain-Derived Algorithms, and the Future of AI Efficiency with Alex Ksendzovsky

Transcript

Earley AI Podcast - Episode 97: Biological Computing, Brain-Derived Algorithms, and the Future of AI Efficiency with Alex Ksendzovsky · The Adults In The AI Room

• 0:01
  SPEAKER_00: This is the Early AI Podcast, hosted by Stephanie, join us as detailed deep into the passions, expertise, and experiences of Thought Leaders and Captain House to talk about what's possible with artificial intelligence.
• 0:14
  SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science, your digital transformation journey with design and deployment of innovative technology solutions, as well as better.
• 0:26
  SPEAKER_00: Now enjoy the show.
• 0:28
  SPEAKER_01: Welcome to the Early AI Podcast.
• 0:30
  SPEAKER_01: My name is Seth Early.
• 0:32
  SPEAKER_01: I'm your host, and each episode we talk about how artificial intelligence and data are reshaping how we operate and how we do business and how we interact with customers and employees and colleagues.
• 0:42
  SPEAKER_01: Today we're going to be talking about something that's really exciting for me.
• 0:45
  SPEAKER_01: I have a background in science.
• 0:47
  SPEAKER_01: My undergraduate was biology and chemistry, and I've always been a science nerd.
• 0:52
  SPEAKER_01: And we're going to be talking about something that really sits at the frontier of AI, neuroscience, and computing, something that most people in the enterprise technology world have not yet encountered.
• 1:01
  SPEAKER_01: The brain consumes roughly, I've read here, 20 watts of power, less than a light bulb, a dim light bulb, and yet it outperforms some of the most powerful data centers on earth in certain ways, in certain aspects.
• 1:14
  SPEAKER_01: The gap is not a coincidence.
• 1:15
  SPEAKER_01: It's the result of hundreds of millions of years of evolution solving the same problems that we are now trying to solve in silicon.
• 1:22
  SPEAKER_01: And my guest today is asking whether we can close that gap, not just by drawing inspiration from biology, but actually using biology itself.
• 1:29
  SPEAKER_01: Alex Sandowski is my guest today.
• 1:32
  SPEAKER_01: He's CEO of the biological computing company, TBC.
• 1:35
  SPEAKER_01: Alex is a neurosurgeon and a neuroscientist who spent nearly two decades studying how the brain processes information, including implanting electrodes in the human brains to understand epilepsy and growing neurons in a dish to study them at the molecular level.
• 1:48
  SPEAKER_01: He co-founded TBC to turn those insights into a new generation of AI algorithms and ultimately computing architectures that are orders of magnitude more efficient than what we have today.
• 1:57
  SPEAKER_01: Alex, welcome to the show.
• 1:59
  SPEAKER_01: Awesome.
• 1:59
  SPEAKER_01: Thanks a lot for having me.
• 2:01
  SPEAKER_01: So I was on the phone with a colleague this morning and I asked how she was doing.
• 2:05
  SPEAKER_01: She said, absolutely fantastic.
• 2:07
  SPEAKER_01: And I said, Wow, I need some of that.
• 2:09
  SPEAKER_01: Give me the how do I do that?
• 2:10
  SPEAKER_01: And she says, Well, I just ignore all of my problems and then I over today.
• 2:13
  SPEAKER_01: So hopefully we can do that.
• 2:15
  SPEAKER_01: We'll have a fantastic day.
• 2:16
  SPEAKER_01: So let's start off with some of the common misconceptions about the AI and the brain.
• 2:20
  SPEAKER_01: When people hear that AI is inspired by the brain, most people think it means neural networks of basically how the brain works.
• 2:27
  SPEAKER_01: What is wrong with that assumption and how far has the field actually diverged from biology?
• 2:32
  SPEAKER_02: Yeah, that's a great question.
• 2:33
  SPEAKER_02: Um it started in the 1940s when the first AI algorithms, the McCullough's neuron, was actually based on how we understood uh neurons, the brain function at the time.
• 2:43
  SPEAKER_02: Uh but then moving forward to the 1980s when back propagation happened, which is an algorithm that's very non-biological, that's when you had a pretty massive divergence.
• 2:51
  SPEAKER_02: And so moving forward in history, past the 1980s into the 2000s and now into 2026, you have extremely performant AI systems, but they're being trained with brute force and they're extremely inefficient.
• 3:03
  SPEAKER_02: And so TB at TBC, we think the reason for this is because they become extremely non-biological.
• 3:09
  SPEAKER_02: At the same time, in the 1980s, you have this divergence in terms of neuroscience and AI.
• 3:14
  SPEAKER_02: And so we started to really understand how the brain works in the 1980s, 1990s.
• 3:19
  SPEAKER_02: We started being able to model the brain with uh brain cells growing on electrodes.
• 3:23
  SPEAKER_02: This is the labs that I was a part of, implanting electrodes into uh patients for epilepsy gave us insights into how the human brain works and similarly in animal models as well.
• 3:33
  SPEAKER_02: And so over that time, our understanding of neuroscience increased.
• 3:36
  SPEAKER_02: But again, those uh insights were not really being used for AI, although recently they have been.
• 3:42
  SPEAKER_01: And so when you when you talk about, you mentioned that backpropagation was a big divergence from biology.
• 3:47
  SPEAKER_01: What do we gain from that and what do we leave behind?
• 3:49
  SPEAKER_01: So we gained performance.
• 3:50
  SPEAKER_02: I mean, you know, that was the that's what led to the models that we have today, right?
• 3:55
  SPEAKER_02: And they're extremely performant.
• 3:56
  SPEAKER_02: Uh, I mean, you know, language models are used by by all people these days, and they're giving us extremely good outputs, but they're very energy inefficient.
• 4:04
  SPEAKER_02: And so we found ourselves as a society in the position where, you know, we need these tools because they have improved the lives of so many, but they're they take so much energy and they're so inefficient to use.
• 4:16
  SPEAKER_02: And so the idea behind TBC was okay, now that we understand neuroscience, can model it, and can actually derive principles from biological networks, which we'll talk about later, can we improve on these AI systems, make them a little bit more biological?
• 4:29
  SPEAKER_02: And in addition to increasing efficiency, improving efficiency, can we improve their performance as well?
• 4:35
  SPEAKER_01: And so talk about that.
• 4:36
  SPEAKER_01: You know, when you when you say inspiration and uh from the brain and empirically deriving principles from it, you know, what what what is that difference?
• 4:43
  SPEAKER_01: And and and again, when we when we say that the brain doesn't work, backpropagation isn't how we learn or how the brain learns.
• 4:50
  SPEAKER_01: Can you talk a little bit about those mechanisms and and uh how they're different and uh and what that uh inspiration versus uh application of the the principles would be?
• 5:00
  SPEAKER_02: Yeah, absolutely.
• 5:01
  SPEAKER_02: Um so thinking about uh inspiration versus being derived, and so this is a new idea.
• 5:07
  SPEAKER_02: Now, drawing inspiration from the brain, again, is not new.
• 5:10
  SPEAKER_02: It's been actually there's been a trend in the last several years.
• 5:13
  SPEAKER_02: There's several large companies that are drawing inspiration from the literature and asking questions, okay, well, you know, here are recordings from a human brain, from animal brain, like you know, how does that work and how does that lead to things like memory?
• 5:24
  SPEAKER_02: Um, and how can we use that to improve current AI systems?
• 5:28
  SPEAKER_02: Um now, the the problem with that is those are not directed experiments.
• 5:31
  SPEAKER_02: So, for example, you know, you have uh a hippocampus uh in an animal, for example, it's responsible for memory.
• 5:38
  SPEAKER_02: That's very different from how backpropagation works, for example, right?
• 5:41
  SPEAKER_01: And so the other just define backpropagation for uh a moment so we're all on the same page.
• 5:46
  SPEAKER_01: What's that?
• 5:47
  SPEAKER_01: Just define backpropagation so people understand what that principle is.
• 5:52
  SPEAKER_02: So backpropagation is is essentially training AI algorithms to define a certain loss.
• 5:57
  SPEAKER_02: And so you you decrease the loss and then it gives you a certain output and a certain prediction.
• 6:01
  SPEAKER_02: So it's a statistical framework rather than biological framework, which uses what's called heavian plasticity.
• 6:07
  SPEAKER_02: Which uses what?
• 6:07
  SPEAKER_02: I'm sorry, heavium plasticity.
• 6:09
  SPEAKER_02: The idea with heavian plasticity is when two neurons wire together, they fire together.
• 6:13
  SPEAKER_02: That's kind of the principle behind it.
• 6:16
  SPEAKER_02: So rather than just using statistical principles to connect neurons or regions of neurons, uh so that's how that works.
• 6:22
  SPEAKER_02: Yeah.
• 6:22
  SPEAKER_01: Yeah.
• 6:23
  SPEAKER_01: So backpropagation is basically looking at the the delta between what you want and what you what you uh expect for an output and what you have for an output, and then adjusting weights in that way.
• 6:32
  SPEAKER_01: And you're saying biological mechanisms uh change their wiring in in ways uh based on the patterns that uh neurons are perceiving or experiencing.
• 6:42
  SPEAKER_01: Exactly.
• 6:43
  SPEAKER_01: Correct.
• 6:43
  SPEAKER_01: So so talk a little bit more about uh what the distinction means in practice for you.
• 6:49
  SPEAKER_01: So so again, we're saying that you know we're we're running experiments on real neurons, and maybe talk a little about what your whole approach is.
• 6:58
  SPEAKER_01: Like walk us through what your research platform looks like, how you're actually using neurons and what you think.
• 7:05
  SPEAKER_01: Yeah.
• 7:05
  SPEAKER_02: Yeah, absolutely.
• 7:06
  SPEAKER_02: So we take brain cells.
• 7:07
  SPEAKER_02: Um, but there's two ways for us to get brain cells.
• 7:09
  SPEAKER_02: So one of which is um from rat cortical neurons, uh, the other which is we create them from stem cells.
• 7:14
  SPEAKER_02: We grow those brain cells on electrodes.
• 7:17
  SPEAKER_02: Um, and we have a high density of electrodes.
• 7:19
  SPEAKER_02: And so here, I have uh one of these arrays here you can see.
• 7:22
  SPEAKER_02: This currently does not have elect uh brain cells on it, but it's it's what it looks like.
• 7:27
  SPEAKER_02: And that gives us access to uh sending information through electrical current and receiving information from the biology as well via an electrical recording.
• 7:36
  SPEAKER_01: Now, really quickly, just I think this is so the brain, so so go hold that uh device up for uh again for a second.
• 7:41
  SPEAKER_01: So the the actual brain cells themselves are sitting in the middle of this uh device?
• 7:47
  SPEAKER_01: Exactly.
• 7:47
  SPEAKER_02: So you can see perfused with chemical, yeah.
• 7:50
  SPEAKER_02: You can see the grid there.
• 7:52
  SPEAKER_02: Yeah, yep, I see it.
• 7:53
  SPEAKER_02: And so no, there's certainly no brain cells on this one, right?
• 7:56
  SPEAKER_02: Uh but the there's 4,096 electrodes sitting here.
• 8:00
  SPEAKER_02: Okay.
• 8:01
  SPEAKER_02: And we have hundreds of thousands of neurons that are that grow on those electrodes, and so that gives us access to how they compute information.
• 8:08
  SPEAKER_01: Interesting.
• 8:09
  SPEAKER_01: Okay, keep going.
• 8:10
  SPEAKER_02: So, okay, so we have these brain cells growing on the electrodes, and again, I keep saying electrodes, and the reason that that's important is the way your brain is taking information now.
• 8:19
  SPEAKER_02: So, vision, for example, sound, the way that's being taken in is through electrical current.
• 8:24
  SPEAKER_02: So your eyes are converting the the light in front of you, your ears are converting the sound waves around you into electrical current that propagates through the brain, allows you to uh understand the world around you.
• 8:34
  SPEAKER_02: So when we don't have those senses in our biological cultures, we have to send information in a certain way.
• 8:40
  SPEAKER_02: So the way we do that is through electrical patterns.
• 8:43
  SPEAKER_02: So a big unlock for the company when we first started in 22 for the first couple of years was how do we uh take an image, for example, or how do we take a concept like a token, convert it into an electrical pattern, and send it into the dish of neurons?
• 8:55
  SPEAKER_02: Um, so that's encoding information.
• 8:57
  SPEAKER_02: So once we've encoded a piece of information, so let's say a picture of my face or a video, a video of multiple pictures together, frames together, once we've encoded that into the biological network, we can then can record.
• 9:07
  SPEAKER_02: And so we record with a very high resolution across all 4096 electrodes, and that allows us to record over time, so we have a time series of how these neurons process that image, for example.
• 9:19
  SPEAKER_02: So that's where we can build mathematical models.
• 9:21
  SPEAKER_02: So that's where we can say, okay, well, a pixel is spatially and temporally, so across space and time, defined by the brain cells in a certain way.
• 9:30
  SPEAKER_02: That's one example.
• 9:31
  SPEAKER_02: I'll get into the memory part in a second.
• 9:33
  SPEAKER_02: But just using that as an example, we say, okay, well, now we understand how brain cells structure spatial temporal information from a single pixel, for example.
• 9:41
  SPEAKER_02: Can we create a mathematical model from that?
• 9:44
  SPEAKER_01: And so that's so it's just perceiving what's there.
• 9:47
  SPEAKER_01: It's just you're you're you're providing these electrical impulses, and the neurons are reacting to those by doing what by by changing their wiring, by uh by actually uh making new neural connections?
• 10:01
  SPEAKER_01: What is what's actually happening at that point?
• 10:02
  SPEAKER_02: So that yeah.
• 10:03
  SPEAKER_02: So the neural connections are important in the context of uh deriving memory principles.
• 10:07
  SPEAKER_02: Okay.
• 10:08
  SPEAKER_02: Uh, but it's really how it's represented.
• 10:10
  SPEAKER_02: And so uh all our neural cultures are wired.
• 10:13
  SPEAKER_02: They're the we act they actually take a couple of weeks to perform to wire up.
• 10:18
  SPEAKER_02: And so uh that gives us a neural representation.
• 10:21
  SPEAKER_02: And so the things that we measure is uh what's called spikes.
• 10:25
  SPEAKER_02: So when neurons fire in relationship to an electrical stimulus, and the pattern of firing, the pattern of spikes, represents how those neurons, as you said, perceive an image or a pixel or a part of an image.
• 10:36
  SPEAKER_02: Okay.
• 10:37
  SPEAKER_02: Okay, and that's extremely powerful.
• 10:39
  SPEAKER_02: Just how neurons perceive information and the the perception information is also processing information that's extremely powerful and allows us a test bed to extract these principles as mathematical models and then apply them and test them in current AI systems.
• 10:56
  SPEAKER_02: And we've we've approached that through creating what we call adapters.
• 10:59
  SPEAKER_02: Uh, and these adapters are, we can talk about them a little bit, but are lightweight small parameter models or learning rules that can plug into current AI systems like diffusion transformers or other transformer models to see if they can improve the output and efficiency.
• 11:16
  SPEAKER_01: So, right now, when this first piece is it of this is just saying, let's put these signals in and let's take a snapshot of how these neurons are reacting.
• 11:25
  SPEAKER_02: Correct, but it's over many snapshots.
• 11:27
  SPEAKER_02: So it's not just a single image where our experiments are multiple days long uh and we encode multiple images.
• 11:32
  SPEAKER_01: Okay.
• 11:33
  SPEAKER_01: So a single image would take multiple readings to understand what that image is, but you're basically looking at how they're firing and how they're reacting to that and building a representation, a mathematical representation of what that perception is.
• 11:46
  SPEAKER_01: Exactly.
• 11:47
  SPEAKER_01: Correct.
• 11:47
  SPEAKER_01: Interesting.
• 11:48
  SPEAKER_02: Yeah, keep going.
• 11:49
  SPEAKER_02: Yeah.
• 11:49
  SPEAKER_02: And so once we built that mathematical representation across many of these plates that I've shown you, we then start to arrive at uh some common rules around how brain cells represent, and we'll focus on a single pixel just because it's kind of easier conceptually.
• 12:03
  SPEAKER_02: And so from there we can say, okay, well, if neurons represent uh pixels in a certain way, can we build an adapter?
• 12:10
  SPEAKER_02: And what we mean by adapter is like I said earlier, it's a certain number of parameters that it's essentially a transformation of data going through fusion transformers or going through the transformer model.
• 12:21
  SPEAKER_02: And so as kind of information in the form of a matrix flows through a transformer model in the context of video generation or image generation in a DIT, for example, can we augment that flow of information by building a mathematical model or an adapter that ultimately changes the matrix as it flows through the DIT, makes it look a little bit more biological, and then predicts the next frame or the next token in a more accurate way, for example, or in a more efficient way.
• 12:50
  SPEAKER_01: Okay.
• 12:51
  SPEAKER_01: Okay.
• 12:52
  SPEAKER_01: And so continue uh down the path.
• 12:54
  SPEAKER_01: So then we have this mathematical model, we're we're transforming it into what end?
• 13:00
  SPEAKER_01: What are we trying to get out of that model in this first piece of the process?
• 13:06
  SPEAKER_02: So what we're looking for is efficiency and performance.
• 13:08
  SPEAKER_02: So really what we measure ourselves again, and again, we started with how to encode information and we've developed this adapter product that plugs into current models.
• 13:16
  SPEAKER_02: And so, current models being video generation models, that's really our first proof of principle.
• 13:21
  SPEAKER_02: And so what we measure against is always a base model.
• 13:24
  SPEAKER_02: So we take an open source video generation model, we build adapters very specifically for that model.
• 13:30
  SPEAKER_02: And again, these are coming from derived principles from the brain cells.
• 13:33
  SPEAKER_02: And then we ask the question uh performance versus efficiency.
• 13:37
  SPEAKER_02: And what the way we define performance is there are evals and metrics out there to ask the question okay, if you're generating a video from a model, how accurate does it describe an actual world?
• 13:47
  SPEAKER_02: What's the resolution of the video?
• 13:49
  SPEAKER_02: In some of our first data, we looked at how much video can you create.
• 13:53
  SPEAKER_02: So we are really benchmarking ourselves against kind of state of the art and metrics within the the realm that we're operating in.
• 14:01
  SPEAKER_02: Okay.
• 14:02
  SPEAKER_02: And uh just really quickly, in terms of efficiency, we measure ourselves in terms of how many tokens or frames in the case of video generation can we create per second.
• 14:11
  SPEAKER_02: So frames per second leads towards cost reduction or or or a measure for cost, especially on current GPUs.
• 14:17
  SPEAKER_01: And and so so you're basically creating a new model representation of how video is processed and how video can be generated based on this core model that has been derived from these neurons, right?
• 14:33
  SPEAKER_01: Because they have solved the problem in a very different way than silicon, than we do in silicon, right?
• 14:38
  SPEAKER_01: It's using biological principles.
• 14:39
  SPEAKER_01: And now this new model is then applied to uh uh a problem of uh video generation or video processing, and then you're seeing efficiencies based on that biologically derived mathematical model.
• 14:53
  SPEAKER_02: Correct.
• 14:54
  SPEAKER_02: And just to just kind of drive one point home, it's an adapter that uh plugs into a current trist transformer.
• 15:01
  SPEAKER_02: So we're not so at this point, we're not completely changing the baseline architecture.
• 15:05
  SPEAKER_02: Um, and there's multiple places where our adapters can plug in uh across the DIT blocks, but um you know, the more we change it, eventually it becomes a completely a brand new architecture.
• 15:16
  SPEAKER_02: And so that's kind of a place where the company is heading.
• 15:19
  SPEAKER_02: But right now, our adapters are quite flexible.
• 15:21
  SPEAKER_02: They can plug into you know any transformer model, but most of our data comes from diffusion transformers.
• 15:26
  SPEAKER_02: And so it's not a complete changing of the algorithm, but it it incorporates into it and improves it.
• 15:32
  SPEAKER_01: And and so what are the results from that?
• 15:34
  SPEAKER_01: So so you have a traditional model, you have this biologically derived model.
• 15:38
  SPEAKER_01: What is the difference in performance?
• 15:40
  SPEAKER_01: Right.
• 15:41
  SPEAKER_01: So we're an efficiency, yeah.
• 15:43
  SPEAKER_02: What's so yeah, so we have been benchmarking this, and so you know, we're constantly uh creating new adapters.
• 15:49
  SPEAKER_02: And if you look at metrics like uh what's called PAI bench or Vbench, these are um five to six metrics looking at the motion of your video that you create, how what the back the background looks like, for example, aesthetic quality, image quality, and we're showing pretty significant improvements.
• 16:06
  SPEAKER_02: And so I hesitate to give direct numbers now.
• 16:08
  SPEAKER_02: I mean, I can tell you it's about you know 13%, for example, improvement in aesthetic quality, which just to give a frame of reference, all of the leaderboards, like if you look at the published literature, to get on the leaderboard, it's like a one to two percent improvement.
• 16:22
  SPEAKER_02: Okay.
• 16:23
  SPEAKER_02: Uh, and so you know, that's that's extremely significant.
• 16:26
  SPEAKER_02: So we're showing like upwards to 13 to 20 percent improvement in some of these metrics, which which is which is very large.
• 16:33
  SPEAKER_02: Another recent metric that we've seen, and so outside of Vbench and what's called PAI, you have uh subjective review or human preference ratings, and that's actually the gold standard for understanding how uh how well your video uh generator or your video generating algorithms create videos.
• 16:50
  SPEAKER_02: And so when we compared ours, our recent adapter against the base, 65% of the time the human raters chose our model.
• 16:59
  SPEAKER_02: Now, again, just looking at the literature to get on the ELO leaderboard, you know, usually it's like 53% of the time you have a significant improvement.
• 17:07
  SPEAKER_02: So again, human raters are are rating our uh videos as much better, a lot a very significant portion of the time.
• 17:13
  SPEAKER_02: And so these numbers are quite staggering in that uh the video that we generate is significantly better.
• 17:18
  SPEAKER_01: Interesting.
• 17:18
  SPEAKER_01: Okay.
• 17:19
  SPEAKER_01: And and so, and then the so that's the performance piece.
• 17:22
  SPEAKER_01: And what about the efficiency piece?
• 17:24
  SPEAKER_02: Yeah.
• 17:25
  SPEAKER_02: So on the efficiency side, we're about four to five per to four to five fold more efficient in terms of how many frames we can provide per second.
• 17:34
  SPEAKER_02: So we recently released a demo on our website.
• 17:38
  SPEAKER_02: I encourage the audience to go look at it.
• 17:40
  SPEAKER_02: And what the demo is, it's an interactive video model.
• 17:43
  SPEAKER_02: So what that means is you can actually play.
• 17:46
  SPEAKER_02: That's a Minecraft game, you can play it, and it's being generated in front of you.
• 17:49
  SPEAKER_02: And actually, a lot of those statistical metrics are are are shown.
• 17:53
  SPEAKER_02: And so we're about four to five times an improvement in the amount of cost.
• 17:56
  SPEAKER_02: And so that can be seen in the frames per second uh as the as the video is generated uh in front of the of the player.
• 18:02
  SPEAKER_01: So that's amazing.
• 18:03
  SPEAKER_01: So that's enormous when you start thinking about applying that efficiency across the spectrum of AI compute and and inference uh activities and and processes.
• 18:15
  SPEAKER_01: So that really sounds like a very, very significant impact on that.
• 18:18
  SPEAKER_01: And right now you're in a kind of a very narrow uh focused niche.
• 18:22
  SPEAKER_01: But talk a little bit about you know the roadmap to where you're going.
• 18:28
  SPEAKER_01: I I have a couple of other questions before I want to go down there.
• 18:31
  SPEAKER_01: So, you know, first of all, we're you started just to talk about learning and memory.
• 18:35
  SPEAKER_01: And and one of the fundamental limitations of current AI models is catastrophic forgetting, right?
• 18:40
  SPEAKER_01: The inability to keep learning without losing what was learned before.
• 18:43
  SPEAKER_01: How does bio how does biology solve that problem?
• 18:46
  SPEAKER_01: And then what have you been able to derive from neurons about continual learning?
• 18:50
  SPEAKER_02: Yeah, so that's uh our fourth blog post.
• 18:52
  SPEAKER_02: And the team can or or you know, your audience would love for them to go look at our blog because a lot of our science is kind of written out there.
• 18:59
  SPEAKER_02: So going back to our setup really quickly, just to kind of uh create a frame of reference for folks.
• 19:04
  SPEAKER_02: So we have information that goes into the biological network, the neurons are represented in a certain way, and we mathematically model them.
• 19:10
  SPEAKER_02: Okay.
• 19:11
  SPEAKER_02: But the neurons can also learn.
• 19:13
  SPEAKER_02: And what learning happens in a cultured setup or in our cultured neurons, what that means is the synapses actually change.
• 19:20
  SPEAKER_02: And we have a in our system, we we have what's called closed loop learning.
• 19:23
  SPEAKER_02: What that means is uh let's say we're trying to uh I'll set up an experiment in our minds.
• 19:28
  SPEAKER_02: Let's say we're trying to classify a number.
• 19:32
  SPEAKER_02: So MNIST is a pretty classic data set.
• 19:34
  SPEAKER_02: And let's say, okay, we want to build a classifier that grabs onto the features of a number and then classifies it.
• 19:40
  SPEAKER_02: That's like you know, classical computer vision.
• 19:42
  SPEAKER_02: And so in our early data, we did that.
• 19:44
  SPEAKER_02: And we said, okay, well, we're gonna encode a number into our biological network, take the response, and then build a classifier in that.
• 19:50
  SPEAKER_02: And as it turned out, it was high more performant, it was more accurate.
• 19:54
  SPEAKER_02: So going back to this, now we want to use this to actually train the biology.
• 19:58
  SPEAKER_02: And so, how do we do that?
• 19:59
  SPEAKER_02: So, when we take in our second training set, we take additional numbers, we then encode that into the biology, take the response, and anytime the classifier gets it wrong, we can within a couple of milliseconds send a second uh pattern of electrical stimulation that tells the neurons the answer was wrong.
• 20:15
  SPEAKER_02: Or we can send an electrical stimulation that tells the neurons the answer was right.
• 20:19
  SPEAKER_02: And so over these closed loop steps, you actually start changing the way the biological network is connected, and the representation for the neurons becomes more accurate and ultimately improves accuracy.
• 20:30
  SPEAKER_02: Now, the reason this closed loop setup is important in the context of memory is because it gives us an insight into how those synapses are actually being changed over time, and so over all of those closed loop steps.
• 20:41
  SPEAKER_02: And we can create mathematical models of that.
• 20:43
  SPEAKER_02: And so I mentioned heavy and plasticity earlier.
• 20:46
  SPEAKER_02: That's an example.
• 20:46
  SPEAKER_02: And so we can take statistical measurements and say, okay, well, you know, this is what the synapses look like from a neurophysiological perspective before training, after training, and we can start building rules around that.
• 20:57
  SPEAKER_02: So the way we did that in our fourth blog post is we derived a learning rule that more closely mimicked how the neurons change their structure or how they change uh their weights in response to this closed loop training, and ultimately found that when this learning rule was incorporated into our MNIST task for continued learning, we actually showed significant improvements over backpropagation.
• 21:21
  SPEAKER_02: And so, again, zooming out, big picture, the way we, you know, similar to us studying neural representations, we study how neurons change their the network over time, mathematically model that, build learning rules from that, and apply that towards continued learning tasks.
• 21:37
  unknown: Interesting.
• 21:37
  SPEAKER_01: Fascinating.
• 21:38
  SPEAKER_01: That really is incredible when you think about that.
• 21:40
  SPEAKER_01: So you're you're you're finding that just the the processing of the image provides a model, but now the learning from the processing of that image provides additional models for how that uh happens.
• 21:52
  SPEAKER_01: And then that leads to additional efficiencies and improvements in performance and and so on.
• 21:57
  SPEAKER_01: With that.
• 21:58
  SPEAKER_01: Yeah.
• 21:59
  SPEAKER_01: Very fair.
• 22:00
  SPEAKER_01: Very, very interesting.
• 22:01
  SPEAKER_01: So when you start thinking about the longer-term vision, right, it's beyond algorithms to hardware chips that more closely mimic the structure and connectivity of the brain.
• 22:09
  SPEAKER_01: What does that path look like?
• 22:11
  SPEAKER_01: And what has to be solved before it becomes real?
• 22:14
  SPEAKER_02: Yeah.
• 22:14
  SPEAKER_02: So the becomes real part.
• 22:16
  SPEAKER_02: So, you know, I posit that it's real now, right?
• 22:18
  SPEAKER_02: And so as I mentioned earlier, you know, we have a demo that's currently live on our website.
• 22:24
  SPEAKER_02: And what that is, is comparing a base model, video generation model, to a model that was enhanced by our neurons or it was enhanced by one of our adapters.
• 22:34
  SPEAKER_02: And so you can very clearly see that our model is a significant improvement over the base.
• 22:38
  SPEAKER_02: And so again, from that perspective, we're live.
• 22:42
  SPEAKER_02: From there, we plan to scale our first commercial product, which is our adapters for video generation.
• 22:47
  SPEAKER_02: So that's going to lead towards more open source models.
• 22:50
  SPEAKER_02: It's going to lead towards outside of video generation, other modalities, applying this towards language models, for example, especially in the context of continued learning.
• 22:58
  SPEAKER_02: And so that's kind of within the realm of our software product.
• 23:01
  SPEAKER_02: Going further, and as we kind of discussed a little bit earlier, right now the product is adapters, but that is leading towards novel architectures.
• 23:08
  SPEAKER_02: And so that is staying within the software realm, right?
• 23:11
  SPEAKER_02: As we grow significantly as a company, can we create new AI architectures?
• 23:17
  SPEAKER_02: So now this novel-based models that more closely mimic the brain from our experiments.
• 23:22
  SPEAKER_02: And so that's kind of the next step in software.
• 23:25
  SPEAKER_02: Then you think about hardware, right?
• 23:26
  SPEAKER_02: You say, okay, well, neuromorphic computing has been around for a while.
• 23:29
  SPEAKER_02: It hasn't really scaled.
• 23:31
  SPEAKER_02: You know, can we create hardware that more closely mimics the structure and function of our biological networks?
• 23:36
  SPEAKER_02: And so again, this is still using the biology as a framework to understand how neurons and to mimic or derive principles from.
• 23:46
  SPEAKER_02: And so hardware is a next step also.
• 23:47
  SPEAKER_02: And then finally, and this has kind of always been North Star of the company.
• 23:51
  SPEAKER_02: Finally, we'll get to a point where we're called real-time biological computing.
• 23:55
  SPEAKER_02: So that's kind of everyone's big interest in this field.
• 23:57
  SPEAKER_02: And that and the idea there is using the actual brain cells to perform compute in real time as part of the inference circuit, either at the edge or in data centers.
• 24:06
  SPEAKER_02: And so in the next five to 10 years, the goal is going to be actually incorporating the brain cells in the compute circuit and applying that towards data centers, for example, or at the edge.
• 24:16
  SPEAKER_02: Yeah.
• 24:17
  SPEAKER_01: So this is really, really fascinating.
• 24:20
  SPEAKER_01: When you start getting into this realm of this is starting to, you know, border on where science fiction is kind of uh, you know, taking our imaginations.
• 24:27
  SPEAKER_01: And now you start thinking, well, what is the implication of this?
• 24:30
  SPEAKER_01: Where can this go?
• 24:31
  SPEAKER_01: And you're actually talking about brain cells, potentially human brain cells that are derived from stem cells, right?
• 24:38
  SPEAKER_01: Skin stem cells, as you which you mentioned, and and and actually incorporating that biology and that that neurology and into a computing system.
• 24:47
  SPEAKER_01: That is kind of really, really fascinating.
• 24:50
  SPEAKER_01: And I imagine that that's probably going to freak some people out because it starts getting in the realm of potential creepiness.
• 24:56
  SPEAKER_01: And also, you know, this the you know, what what Ray Kroswell, you know, calls the singularity, right?
• 25:01
  SPEAKER_01: All of this kind of hype around where this is going.
• 25:04
  SPEAKER_01: But that's a fascinating, fascinating potential, right?
• 25:07
  SPEAKER_01: So talk a little bit more about that.
• 25:09
  SPEAKER_01: And what are people concerned with about that?
• 25:11
  SPEAKER_01: And what is the potential that that opens up for us?
• 25:13
  SPEAKER_02: Yeah.
• 25:14
  SPEAKER_02: So going back to our system, right?
• 25:16
  SPEAKER_02: So thinking about the brain cells that are on the dish, compared to our brains, for example, or animal brains that have a lot of emergent properties, they're actually quite different in in terms of the architecture in terms of the structure.
• 25:29
  SPEAKER_02: Uh, they're much smaller amounts.
• 25:31
  SPEAKER_02: So we have hundreds of thousands of neurons compared to folds of magnitudes larger, right?
• 25:36
  SPEAKER_02: The way they're connected to one another is is quite different.
• 25:39
  SPEAKER_02: And so it's enough for our purposes, and it's enough to comp to to pull out computational primitives and principles.
• 25:45
  SPEAKER_02: It's enough to actually perform computational tasks, but it's not enough to create this emergent property of sentience, consciousness, or even anything close, feeling pain or anything, anything close that animals have.
• 25:56
  SPEAKER_02: And so you know what's that today, today.
• 26:00
  SPEAKER_02: Today, yeah, but that's also uh that's true, but also you know, we are setting the guardrails ourselves, right?
• 26:06
  SPEAKER_02: And so at no point as a company do we have any interest in creating something even close to a three-dimensional structure of a human brain, even in the context of organoids, which um have a little bit more of a three-dimensional structure, like a brain, it's not even close.
• 26:21
  SPEAKER_02: And so again, uh, we are setting the guardrails, and as we build this, uh we we will maintain those guardrails.
• 26:26
  SPEAKER_02: And we've already worked with bioethicists in the field to really make sure that uh that it doesn't create a scary future.
• 26:34
  SPEAKER_01: Yeah.
• 26:35
  SPEAKER_01: Well, I think we have a scary future no matter what.
• 26:37
  SPEAKER_01: Even if you just take where AI is going in general, right?
• 26:40
  SPEAKER_01: And the models are getting more powerful and how those are uh emulating some very interesting, you know, human behaviors.
• 26:46
  SPEAKER_01: Uh obviously it's emulation.
• 26:48
  SPEAKER_01: So, but but I think that you know, you when when people kind of think about extrapolating, you know, when you talk about a hundred years from today, right?
• 26:55
  SPEAKER_01: You're you know, well, after we're gone or we're uploaded, right?
• 26:59
  SPEAKER_01: Where where is this going?
• 27:00
  SPEAKER_01: And you have guardrails, maybe others will not.
• 27:02
  SPEAKER_01: But anyway, that's getting completely in the realm of of speculation and and extrapolation beyond what we're talking about today.
• 27:10
  SPEAKER_01: You're we're really looking at, you know, this very focused application to improve these algorithms, and it makes a lot of sense, and then be able to incorporate that into actual computing infrastructure.
• 27:20
  SPEAKER_01: So, so one of the things that I was curious about, just at a biological level, what is the difference between a mouse neuron and a hu and a brain human brain neuron?
• 27:30
  SPEAKER_01: Human a mouse uh brain neuron and a human brain neuron.
• 27:33
  SPEAKER_02: Yeah, so so we're we're so in the context of what we're doing, we actually haven't seen much of a difference.
• 27:38
  SPEAKER_02: Okay.
• 27:38
  SPEAKER_02: Um, we are my sense is we are uh approaching this with such foundational levels of of compute, fundamental levels of compute, that we're not seeing a ton of differences, but we're actively understanding and studying that.
• 27:51
  SPEAKER_02: I think where uh our optimization strategies are coming from is really trying to understand the different types of neurons, meaning inhibitory, excitatory, what percentages of those that we find that gives us better results.
• 28:03
  SPEAKER_02: That's more of I think where we're gonna kind of our digging will lead to improved optimization rather than human versus mouse or sorry, human versus rat.
• 28:11
  SPEAKER_02: Um but yeah, we're actively studying that.
• 28:13
  SPEAKER_02: Like I said, right now we haven't seen a ton of differences.
• 28:15
  SPEAKER_01: You know, it's interesting when people talk about you know technology and and silicon uh in silico uh computing and start making analogies to the brain.
• 28:25
  SPEAKER_01: The human brain with three billion neurons and uh the the number of I I forget what the number of connections are, right?
• 28:32
  SPEAKER_01: But but the level of complexity of it one neuron can can connect to 10,000 other neurons, right?
• 28:38
  SPEAKER_01: And then you have uh neuro uh you have neurochemicals and and uh neurotransmitters.
• 28:43
  SPEAKER_01: There's a hundred different neurotransmitters and they are all uh analog, right?
• 28:48
  SPEAKER_01: They're gradients, right?
• 28:49
  SPEAKER_01: They can so so the fact that you have that level of complexity, it's so much more than anything we can build or conceive of today, right?
• 28:57
  SPEAKER_01: So the human brain is just orders and orders and orders of magnitude more complex.
• 29:00
  SPEAKER_01: I mean, I I imagine you know that better than than most people do, right?
• 29:04
  SPEAKER_01: What are your thoughts on that, just in in terms of the the combinatorial explosion?
• 29:09
  SPEAKER_02: Yeah, I I think you know it's interesting you say that because certainly in our adapter product now, like we're not capturing nearly as much the complexity there in terms of our neural representations or the way the neur our our cultured neurons represent information.
• 29:21
  SPEAKER_02: But you know, the fact that we're showing these massive gains in a very specific use case with just our first adapters, just making it a tiny, tiny bit more biological.
• 29:31
  unknown: Yeah.
• 29:32
  SPEAKER_02: These massive gains.
• 29:33
  SPEAKER_02: And I think it's a testament to that.
• 29:34
  SPEAKER_02: I think it's it's it's why we're we're able to see this and and really just capture a minuscule portion of that complexity of how the brain operates.
• 29:42
  SPEAKER_02: And so in the end, the more of these principles and primitives we can derive and apply to create adapters, novel architectures, I think the more improvements we'll get in terms of performance and efficiency.
• 29:52
  SPEAKER_01: Yeah, and and and again, we're a very long way from you know doing anything that that that gets into that that the science fiction realm, um, but you know, the principles are there.
• 30:03
  SPEAKER_01: Let's talk about cons energy consumption because that's one of the pressing infrastructure challenges of the next decade.
• 30:08
  SPEAKER_01: Where does where does biological computing fit into that?
• 30:10
  SPEAKER_01: And what's the order of magnitude of efficiency gain that is possible and achievable in this time frame, in this five or 10 year time frame?
• 30:17
  SPEAKER_01: What are your thoughts about that?
• 30:18
  SPEAKER_02: Yeah.
• 30:19
  SPEAKER_01: Maybe even near term, like the next couple of years versus the next year.
• 30:22
  SPEAKER_02: Yeah, so so there's two two things here.
• 30:24
  SPEAKER_02: Uh one is, and and and you can't really take performance away from efficiency, right?
• 30:28
  SPEAKER_02: It's kind of like if you want to increase performance without our technology, let's say you double the amount of parameters, that's gonna reduce your efficiency, right?
• 30:35
  SPEAKER_02: If you wanna, there's a lot of uh techniques that allow you to optimize current algorithms or current models, um, but they can significantly decrease performance.
• 30:44
  SPEAKER_02: What we've seen, which is interesting, is and first of all, just to give you a number, like I said, we're four to five X in terms of using our adapters, that's today.
• 30:51
  SPEAKER_02: Even today, even today.
• 30:53
  SPEAKER_02: And and by the way, uh we've we've also recently shown, and this is extremely important, that when you combine our adapter plus current optimization strategies, when you combine those two, they can be synergistic.
• 31:04
  SPEAKER_02: And so we can get four to five times efficiencies using our adapter, four to five times efficiencies using uh other optimization strategies.
• 31:11
  SPEAKER_02: And what what turn what it turns out, which is quite interesting, is our adapter, when you use those other optimization strategies, does not uh lead to a decrease in degradation of the quality.
• 31:21
  SPEAKER_02: So it maintains the quality of the video or the data that's being produced.
• 31:25
  SPEAKER_02: And so again, uh we are you know right now today 4 to 5x.
• 31:30
  SPEAKER_02: We're continue continuing to improve on that.
• 31:32
  SPEAKER_02: And so, you know, I I think we can achieve many, many fold within the next year, 20 fold within the next year, a hundredfold the year after, and then it keeps going.
• 31:41
  SPEAKER_02: Now, this is just in software, by the way.
• 31:43
  SPEAKER_02: I think that uh when using real-time biological computing, our uh our ceiling is extremely massive.
• 31:49
  SPEAKER_02: And so the brain now is millions of times more efficient than silicon.
• 31:53
  SPEAKER_02: And so, you know, going back to this example of like, you know, we if we can make software and then hardware just a little bit more biological, we can potentially achieve that.
• 32:02
  SPEAKER_02: And when biology is in the actual circuit itself, you know, our window of opportunity is massive.
• 32:08
  SPEAKER_01: Yeah, yeah.
• 32:09
  SPEAKER_01: I mean, I can't I can't imagine uh where that's going.
• 32:12
  SPEAKER_01: And and so and and and will this apply?
• 32:15
  SPEAKER_01: Imagine the answer is yes, right?
• 32:18
  SPEAKER_01: This is going to apply to, I mean, video is one area, but this is going to apply to any kind of large language model development, inference, any of those things.
• 32:27
  SPEAKER_01: You're really talking about those massive, massive improvements in efficiency.
• 32:30
  SPEAKER_01: And and who are you, who are your customers?
• 32:32
  SPEAKER_01: Who is who is does it doesn't does a corporate uh technology department uh need to know about this?
• 32:38
  SPEAKER_01: And can they purchase this in some way, shape, or form?
• 32:41
  SPEAKER_01: Or how how are how is this getting to the point of need in the market?
• 32:45
  SPEAKER_02: Yeah, so starting with our current use case with video generation.
• 32:48
  SPEAKER_02: So really, you know, the the customers are anyone who wants to generate videos.
• 32:52
  SPEAKER_02: So creatives, for example, marketing agencies, advertisement firms, you know, there are marketplaces where uh where folks go to access different models, open source models are a part of those marketplaces.
• 33:03
  SPEAKER_02: Um, you know, when it comes to things like continual learning, which is within the research realm of a lot of the uh current corporations, current AI companies, again, that's also something that we're pushing on.
• 33:14
  SPEAKER_02: And we're we're actively or would be happy to have partnerships to help do that as well.
• 33:18
  SPEAKER_02: And so as we scale the different modalities, we're pulling in more and more customers.
• 33:23
  SPEAKER_02: But yeah, I mean, in the immediate term, it's anyone that's interested in generating high quality video or lower cost.
• 33:29
  SPEAKER_01: Yeah.
• 33:30
  SPEAKER_01: And uh I imagine so those cost differentials are gonna get very significant from what you're telling me.
• 33:35
  SPEAKER_01: And that is going to be a very clear business case for using these models.
• 33:39
  SPEAKER_01: And I can imagine it's like, you know, water running downhill, right?
• 33:43
  SPEAKER_01: It's gonna take the path of least resistance and of course cost and performance are really the things that the leverage that people care about.
• 33:49
  SPEAKER_01: Uh, and that it's going to be meaningful for that.
• 33:51
  SPEAKER_01: So that's really fascinating when you think about the implication of that for these large language models and for data centers and for you know the massive amount of computing infrastructure that we're there that we're developing these days.
• 34:04
  SPEAKER_01: And again, if you had 20-fold improvement in uh in performance and efficiency, that's going to just go so far along with all the other areas of where we're getting incremental efficiency and and performance across the landscape.
• 34:18
  SPEAKER_01: So you talked a little bit about bioethicists and you know, using human brain cells in computing does you know raise questions that go beyond the engineering.
• 34:26
  SPEAKER_01: And what are those ethical and philosophical considerations that your work is surfacing?
• 34:30
  SPEAKER_01: And what do you think about those?
• 34:32
  SPEAKER_01: And you know, what has come up with with those discussions for those bioethicists?
• 34:35
  SPEAKER_02: Yeah, so the main framing is around the difference between our biological cultures and the emerging properties of an actual brain.
• 34:42
  SPEAKER_02: And so, you know, it really focuses on you know things like sentience, things like awareness.
• 34:47
  SPEAKER_02: Um, I think it is a yeah, again, those terms are still being defined in the in the literature.
• 34:53
  SPEAKER_02: Um, and you know, are there neurophysiological biomarkers for a sentience, for example, which there currently really aren't.
• 35:00
  SPEAKER_02: But but as we as a field start to define what it means to be sentient and or conscious or aware, you know, that allows us to build the uh the guardrails around what we're doing as well.
• 35:10
  SPEAKER_02: And so again, the difference is so huge right now that we're not super worried about it, but we are actively monitoring it.
• 35:17
  SPEAKER_02: But that it that's around that though, those are the main principles that we're operating around.
• 35:21
  SPEAKER_01: And where do you see this going in in 20 years, say 30 years?
• 35:25
  SPEAKER_01: I mean, again, 10 years, a year is a long high time horizon for our space, right?
• 35:31
  SPEAKER_01: For for AI.
• 35:33
  SPEAKER_01: But what can you extrapolate 10, 20 years or longer into the future?
• 35:38
  SPEAKER_01: What do you where do you think this is really going?
• 35:39
  SPEAKER_02: I think in 10 to 20 years, we're gonna have real-time biocompute.
• 35:42
  SPEAKER_02: I think we're gonna have robots that are operating with a neural culture that are making them a lot more performant.
• 35:49
  SPEAKER_02: Well, those are some things that we're trying to work around.
• 35:51
  SPEAKER_02: That's why NAS not production ready today, is we have to figure out how to immortalize the cells, how to maintain them, uh, feeding them and removing waste.
• 35:59
  SPEAKER_02: I mean, these are all things that, you know, this is why this is a 10 to 20 year problem, not a two-year problem, right?
• 36:04
  SPEAKER_02: Um, but I think they're gonna be, you know, at the edge, for example, robots that have neurocultures attached to them that can learn continuously in time with you, that much more performant.
• 36:13
  SPEAKER_02: You know, when it comes to data centers and incorporating real-time biocomputer data centers, again, massive efficiency gains out of uh distribution generalization is another problem we're trying to tackle as well.
• 36:23
  SPEAKER_02: That's something that neurocultures could potentially help with.
• 36:26
  SPEAKER_02: So, so yeah, so the idea is uh, you know, we'll have an AI environment that is much more biological, which allows it to be more performant and efficient.
• 36:34
  SPEAKER_01: Interesting.
• 36:35
  SPEAKER_01: So I think we could probably, I mean, there's other things that we could talk about because this is such a fascinating space.
• 36:42
  SPEAKER_01: What are what are some of the takeaways you'd like people to walk away with from this?
• 36:46
  SPEAKER_01: What do you want to leave people with?
• 36:48
  SPEAKER_02: Yeah, so I think one of the major takeaways is that this is not just in the realm of research at this point.
• 36:52
  SPEAKER_02: You know, we are production ready, we, you know, you can go to our website, play a demo of a model that has an adapter that came from brand cells.
• 37:00
  SPEAKER_02: We'll you can purchase videos from us in a couple of months that are much more uh higher quality uh and take less cost to make.
• 37:09
  SPEAKER_02: So I think the biggest takeaway in my mind is this is not something that's in the future, this is today.
• 37:14
  SPEAKER_02: Um, and the second takeaway in the context of real-time biocompute and these novel architectures we're getting from this is the AI future is gonna look much better when it's biological or more biological.
• 37:23
  SPEAKER_01: And when you uh you there's a recent uh CRA Computing Research Association dinner that talked about a demand rising 10,000 X just in five years.
• 37:34
  SPEAKER_01: In five years, 10,000 X.
• 37:36
  SPEAKER_01: Unbelievable.
• 37:37
  SPEAKER_01: And and when you're in the algorithmic bucket with hardware and biological compute later, but you know, again, the human brain runs in a less than a white uh a light bulb, right, and still outperforms these data centers.
• 37:51
  SPEAKER_01: You're talking about eventually not just hundreds of folds, but but hundreds of thousands of fold when you think about it.
• 37:58
  SPEAKER_02: Yeah, you know, that that dinner was a was a great, it was very interesting.
• 38:01
  SPEAKER_02: And, you know, we didn't just talk about what the problem was, right?
• 38:04
  SPEAKER_02: This 10,000 fold increase in energy demand.
• 38:07
  SPEAKER_02: We we talked about what the potential solutions are.
• 38:09
  SPEAKER_02: And so, and again, that's just the next five years.
• 38:11
  SPEAKER_02: And so, you know, 50-fold of that demand, uh, according to the group around the table, would come from uh novel architectures.
• 38:18
  SPEAKER_02: And so this, you know, software, for example, and this is where we live, and this is where we're already showing significant gains.
• 38:23
  SPEAKER_02: And so I think within five years for sure we'll contribute to that 50-fold improvement in efficiency from pure software.
• 38:29
  SPEAKER_02: The second bucket was 50-fold improvement in efficiency when it comes from hardware.
• 38:34
  SPEAKER_02: And so we're also pushing on that, as I mentioned earlier.
• 38:37
  SPEAKER_02: And, you know, we do intend to design hardware that's more closely mimics the biology.
• 38:42
  SPEAKER_02: So that's kind of where we're gonna help there.
• 38:44
  SPEAKER_02: And then the other fourfold from the group was novel ways to acquire energy through solar and wind, which is not a focus of our company, but certainly a lot of headweights being made, being made there as well.
• 38:55
  SPEAKER_02: So we're very happy to be part of this future that's going to help with this massive demand in compute, and we're pushing on two fronts.
• 39:02
  SPEAKER_01: Amazing.
• 39:03
  SPEAKER_01: And again, there's so much to to explore in this area, but I think you've really given us kind of a taste of what this means and just how practical and innovative that is today.
• 39:15
  SPEAKER_01: Where are you in terms of your competition?
• 39:17
  SPEAKER_01: Are are there other folks that are you know attacking this problem in similar ways?
• 39:21
  SPEAKER_01: Are you very unique in this in this area?
• 39:23
  SPEAKER_01: Tell me about the the landscape of the industry.
• 39:27
  SPEAKER_02: Yeah, I'd say we're very unique in the context of what we call applied biological computing.
• 39:31
  SPEAKER_02: You know, we're really, as far as I know, the only company out there that's really defining what to apply biocomputing to.
• 39:37
  SPEAKER_02: And so, you know, we're definitely the only company that's been building software from this.
• 39:41
  SPEAKER_02: This adapter concept came from us.
• 39:43
  SPEAKER_02: Um, but the space is starting to get a little hot, which is great.
• 39:46
  SPEAKER_02: You know, there are other companies that are interested in biocompute and so on the hardware side of things.
• 39:50
  SPEAKER_02: There are lots of academic labs that are starting to think about this.
• 39:54
  SPEAKER_02: DARPA has been thinking about this as well.
• 39:56
  SPEAKER_02: And so, you know, it's really nice for us to be leaders in this category and and kind of showing or or assisting in in growing the field in general.
• 40:06
  SPEAKER_02: Um we're we're proud of that.
• 40:07
  SPEAKER_02: I think that's that's great.
• 40:08
  SPEAKER_02: I think the more companies that come around, uh we see them as friendlies, and we're very happy that that's happening, and certainly academic labs as well.
• 40:15
  SPEAKER_02: And you know, we're in the process of building collaborations with some of these companies and academic labs to to help further progress of biocomputing.
• 40:22
  SPEAKER_01: Amazing.
• 40:23
  SPEAKER_01: Well, thank you for sharing your thoughts and your your uh your your innovation and your approaches and really giving us that glimpse into this frontier that that most people are not aware of, but is probably going to be the most consequential development in the history of computing.
• 40:40
  SPEAKER_01: I mean, I I I completely believe that, right?
• 40:43
  SPEAKER_01: Because that's where, you know, uh that's just where this needs to go and where it will go.
• 40:48
  SPEAKER_01: But I really appreciate you you sharing all this with us today.
• 40:51
  SPEAKER_01: Absolutely.
• 40:52
  SPEAKER_01: Thank you very much for having me out.
• 40:53
  SPEAKER_01: Yeah.
• 40:54
  SPEAKER_01: And thank you to our audience for tuning in.
• 40:56
  SPEAKER_01: Be sure to subscribe to the Early AI podcast uh for more conversations about how AI is shaping the future of business.
• 41:05
  SPEAKER_01: And uh thank you, Carolyn, for your hard work behind the scenes.
• 41:08
  SPEAKER_01: Uh, thanks, everyone.
• 41:09
  SPEAKER_01: See you next time.
• 41:11
  SPEAKER_00: Thank you for joining us on another deep dive into AI innovation.
• 41:15
  SPEAKER_00: Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world.
• 41:22
  SPEAKER_00: The Early AI podcast is sponsored by Early Information Science Invector.
• 41:27
  SPEAKER_00: To learn more, visit early.com at E A R L E Y dot com.
• 41:32
  SPEAKER_00: Thank you for tuning in.