A Brain Researcher's Bid to Make Digital Twins of Humans

But now imagine that if I had a way to create a model, a functional model, a digital twin of that system, then if I have enough resources and I buy enough GPUs or enough computers, I could run what would take 10,000 years, I can run it in an hour.

Simulation.

Because yeah, exactly. I can just download all the video on the internet and then just show it to this digital twin.

It's like a digital representation of the mouse's visual cortex.

Exactly.

Watching the movie. Exactly. Hello everyone, and welcome to another episode of Core Memory. Our guest today is Andres Tolias, brain researcher extraordinaire. I'll give you a proper intro here in a second. I just wanted to welcome people to our new podcasting studio. Which is in flux. It's gonna be even more glorious and beautiful, um, in the weeks to come. But thank you so much for joining us in person in our new, new digs. Yeah. Thank you.

Thank you for having me. And I'm very excited to be your first podcast.

You also have the, um, honor of being the first guest on a, when I read an ad for the podcast, we are some very kind people at, uh, E1 Ventures, a venture capital firm, have, have been, are are now sponsoring the show. Ah, and, uh, they're a local VC firm, so we're not doing, we're not doing, uh, supplements or mattresses like every other podcast. We're going pure capitalism. And, uh, and there you go. Um, everyone could check out E1 Ventures. And now, into the show we go. Um, Andres, I've known you for, I guess, a year, year and a half now. You're down the road at Stanford doing brain research. Before that, I think you did your equivalent of your undergrad studies at Cambridge and then, and then at least PhD stuff at MIT.

Yeah.

And you are in this area that has always fascinated me, which is, well, you do a lot, but one part of it is, is this idea of actually understanding how the brain works, trying to map it. And then now these days, doing a bit of pairing between neuroscience and AI research and trying to figure out where these fields meet together. Yeah, yeah.

So basically, you know, the brain is, you know, the example of an intelligent system that we have, you know, animals, humans. And, you know, we are primarily interested to understand how the brain works, what are the algorithms of intelligence. And it so happens that in the last 10 years or so, we have another form of intelligence that is also performing and doing very sophisticated tasks, artificial intelligence, that got early inspiration from the brain. You know, like there was many different attempts over the years, over the decades to build AI systems, you know, symbolic approaches, other approaches in computer science departments. But what has ultimately succeeded and what we have right now as AI, it's artificial neural networks. And what is very fascinating to us in neuroscience is that not only we can use the brain for inspiration to keep on hopefully developing better AI, safer AI, smarter AI, but also now we can use AI to help us understand the brain. So we can use AI for science. And what is particularly exciting is that we can build models of the brain that are in the same substrate as AI. So we can compare artificial intelligence with natural intelligence. So this is like tight loop from neuroscience to AI, AI to neuroscience. And we hope this loop will accelerate and enhance both fields in the next decade.

Yeah, I mean, and this is why I get so excited. We have like you said, you have this collision almost of these two things happening, which is new tools and abilities to understand the physical brain as we're racing to make an artificial intelligence. And to your point, I mean, they called these things neural nets from the outset. I used to feel like it was loose and vague at best in terms of representing the brain or being akin, but as time has gone on and we see the complexity of these neural networks, I mean, it feels more apt and realistic to me these days.

Yeah, definitely. So they are not implementing intelligence at what we call implementation level. They don't have like, you know, they're not built from, you know, the same complexity of neurons that are built in our brain or in animals' brains. They don't have ion channels. They don't have like the biological complexity. But they have, in many ways, they are built by units that essentially integrate information from other units. They have synapses that are plastic like our brains do. And what is particularly remarkable that has come out in the last few years is this concept of universality, that if you look at artificial neural network and you take different artificial neural networks with different architectures, you train them on similar tasks, they have a lot of similarities in the way they are representing information at this sort of information processing level. So you can think of it like if you can write down, let's say, an algorithm of how to compute something, you can implement it like with biological substrate or with silicon or, you know, maybe even other kinds of stuff. But the algorithm, the way it's representing information, is processing information. You can ask the question how similar it is. So what people have found in the last few years is that there's a lot of similarities within networks that have different architectures. And also we're finding now and others that there's similarities between biological networks, the way they present information, let's say about vision in particular, and these artificial neural networks. And there is some very nice work. One of the things is called the Platonic idea. So it's basically like It seems that as systems are becoming intelligent, they share a lot of similarities, even if it's almost like convergent evolution of some sort. But of course, there are also many differences, and that's where the excitement comes. Like, what are the similarities and what are the differences between these two systems?

So when you— given all this, when you started out your career, you know, I mean, AI As you know, I mean, it was really, had gone through these fallow periods where people did not think this neural net approach would work. I remember when I would, even in, I'd say like 2012, when it starts lighting up a bit, and then even 2015, 2016, when I would interview neuroscientists, you know, they were a bit skeptical about AI. They, yeah, there wasn't this huge overlap in the fields. I know that you've like embraced it and are doing now pioneering work uniting the physical brain and AI. But what was your path? Did you see any of this coming or you were just purely in the neuroscience world?

No, that's actually a very good question. When I started my PhD at MIT, AI for a lot of years, I would say until 12, 2015, maybe even, or even later, was like a bad term. Like, people use the word machine learning. It wasn't a very popular thing because of its overpromise over the years and failures and the winters, the so-called AI winters. And in particular, artificial neural networks were not popular at all for many, many years. It was like there was a few people like Geoff Hinton, Yann LeCun, Yoshua Bengio. If you academic-like labs that were— and even those were having a hard time getting funding sometimes— that were pioneering this field. There was other approaches, let's say the mainstream approaches in computer science or in computer vision was like other approaches. So, you know, of course, as being neuroscientist, you know, like my main interest always has been trying to understand the brain. You know, I'm a scientist, I want to understand the brain, but AI was, you know, as a way that, you know, if we understand the brain, one way to test our understanding is to build systems that mimic the biological behavior, right? Whether this is human or animal or whatever, right? So there was always been, as in neuroscience, there will always been an interest in artificial neural networks, but it wasn't like a mainstream thing from an AI or computer vision point of view. And then these two things came together, as you said, after 2010, 2012, and everybody then switched over very fast.

And embraced it.

Embraced it.

Do you know this guy? I think his name is Gary Lynch at UC Irvine.

Yeah, I know the name, but I, you know, yeah.

Okay, he's a neuroscientist I've interviewed over the years. There was this book about him called 101 Theory Drive.

Yeah, I don't know the book though.

It's like one of my favorite books of all time. I mean, this guy is, He let this reporter into his lab, just got to hang out for, I forget how long, a year or two. And Gary's one of these sort of eccentric researcher types. But I, you know, I, yeah, I just remember calling. I got to know him a bit and then I would call him about AI and he was just so skeptical. It's funny to me now to see how intertwined these two fields are. Okay. When you started out and you were doing your PhD work and then thinking what you wanted to do after, what was your driving question about the brain?

Yeah, so my driving question was trying to understand perception, you know, in particular like visual perception, perceptual decisions, memory, learning, visual learning, how are representations. But I was always mostly interested at what we call, at least my mainstream, although I work at other levels, has been at this algorithmic or presentational level. There's these very famous 3 levels of analysis that David Marr proposed that I think are still useful. There is this idea that to study an intelligent system, you have the behavioral or computational level. What are these goals? So this is, let's say for us, perception is very important or language or these are the things that we do. And then there is this algorithmic or representation level, which is how is information represented, transformed, manipulated in the brain. And then there is the implementation level, which is like, what are the biophysical mechanisms? You know, we have neurons, we have ion channels, genes, you know, the biological details. I was always interested to bridge this level. I mean, ultimately all levels, but in particular, since I was a graduate student at this representational algorithmic level, the perceptual level or the behavioral level. But I would say the reason that right now things are very exciting, and that was extremely frustrating for me as a graduate student and also as a postdoc over the years, is that it's only in the last 5 years or so that we have the technology to do large-scale data collections. At the level that at least I'm interested, which is the single neuron, single spike level, or looking at like fast time windows with calcium imaging and other stuff. Because ultimately, like, you know, science, if you look at the history of science, we need data. You know, without data, you know, we can come up with beautiful theories, beautiful hypotheses, but even Newton, you know, relied on data to fit his model, right? So data has always been the driving force. And because the brain is a very complex system, it's because of a, of a lot of it because of funding from the BRAIN Initiative that they decided to invest in neurotechnology also about 10 years ago at the time that the AI revolution was happening.

Right. The US government and I think the Europeans made these big bets, right?

Yeah. The Europeans made a different bet that was more a very like bottom-up modeling approach, like from the physical level to recreate intelligence that ended up not being very successful for reasons I was criticized from the beginning by many, and I think rightly so. But, but the United States government decided to invest in technology to build technology and also other institutions, private institutions, to build technology that will enable us to collect large-scale data. And if you think about AI now, it's a data-hungry enterprise, right? Like it's all relying on large data. So now we have large data in neuroscience and we have tools in AI to build models of the brain. And that's only possible because of these two things coming together.

So like, how did we, how did we even get data from the brain? I know we have to put a timeline on it. Call it like in the '80s, '90s, early 2000s.

So You know, it's, you know, I like to look at the history of a field to sort of understand, you know, the complications. So if you look at like, I would say in the last 80 years or so, we relied on measuring the activity of individual neurons. And if you look at, let's say, someone that did their PhD thesis in the '90s or in the '80s or even early 2000s, they were relying on like getting data from a few hundred neurons maybe, and then they would describe a system with just like an electrode. Yeah, with an electrode. Usually the classical thing, let's say Hubel and Wiesel, people that are famous, they got Nobel Prizes by understanding early computations in the visual system. They used a, you know, basically single electrode. So they would learn a single electrode, get close to one cell, recorded spikes or action potentials. This is the binary way that neurons communicate information. And then try and study these and they discover stuff that got them a Nobel Prize. And there's been a few Nobel Prizes like that over the years. The most recent one is the grid cells, place fields by the Mosers and Chonoke. So these were like mostly based on the analysis of individual neurons. And if we were lucky enough that even if you recorded from let's say a few hundred neurons, you were trying to, you were starting to understand something interpretable about the code. Because ultimately, if you think about it, what we're trying to do in neuroscience is there's the outside world and how we interact with it. And then there's the inside world, which is what's going on inside our brain. We're trying to find correspondences between these two systems. And there's no a priori reason why we would be able to even find any correspondence between when you record spikes from individual neurons when there's 100 billion of them how it will relate to something in the world. So that was sort of the paradigm, but it was very slow. And also it was a lot of the discoveries were happening by serendipity, by accidental discoveries. It wasn't like, and as much as we like to think, oh, hypothesis-driven science, we all have very smart people that have great ideas and we write papers that sort of, often these discoveries were happening accidentally. And if you read the history, the famous scientists will say, oh yeah, I was trying to do this and then, 'One day the experiment went wrong and I noticed this and then I got a Nobel Prize.' You know, so we're very data-limited.

So lucky.

Yeah.

And okay, so that, that's moving along and then I'm, you're going to have to help me out here. So I've done some reporting on a bit of the modern day and what this looks like with, with in particular this, the part that I've covered is called connectomics and I know you overlap with this, but the. You know, so I've been at Harvard at Jeff Lichtman's lab, and what they do is they take a mouse brain and then, um, or they take a tiny chunk of a human brain that they get from like a surgery. And, and then I've seen it. I mean, they have this, it's almost like a meat slicer. It slices, but it's, it's very precise. It slices these tiny, tiny slivers of the brain, and then they image these in 3 dimensions, and then they stitch back all the tiny slivers and create these beautiful images where you can, you can trace the neurons. Like, what else from the era you were describing, like, what are the main tools we're using now to get the most resolution?

Yes, exactly. So as I said, you know, there is— the brain is— it's a very interesting— neuroscience is a field of research. It's not a discipline like biochemistry or chemistry or You know, it's, it's an area of research that draws a lot of expertise and fields from philosophers to molecular biologists to people studying biophysics to— and if— and, and there's different levels to study. The, the connectomics level that you're describing is trying to understand, to describe the structure of the brain like a map. Like a map, yeah, at high resolution. Yeah. So you can And previously that field, it's basically anatomy. Classically in the past there was anatomy. If you go, let's say Cambridge University, it's a very old university, there's the anatomy department and the physiology department. The anatomists were studying structural stuff. For example, they were slicing the brain and looking using microscopy or other methods or staining. And the most famous father of that field is Ramón y Cajal, that discovered that the brain is built from individual neurons. So he was like looking at basically, he'd take a tissue and slice it and look at the, stain the neurons and look at them and describe them. But of course, the dream of everybody is trying to relate what you see structurally again with what has related to behavior, to cognition.

Sure.

And then there was another field, physiologists, who, and the father of that field is Adrian, uh, an Englishman at Cambridge. And he basically discovered, you know, people knew there was electricity in their brain and he discovered, again, accidentally, he thought it was an artifact that when neurons, the way neurons communicate with each other is through these action potentials or spikes. This is like almost like a bit-like code, like there's They're silent and all of a sudden there's a beep and they send a signal. And that opened the whole field of physiology in the last 100 years. So what I am talking about is that in addition to the developments in connectomics and anatomy and molecular tools to study the brain, like there's also other things like large-scale transcriptomic, single-cell transcriptomic analysis. So you take the neurons and you do, you see all the, you know, messenger RNA that they're expressing. You do large-scale transcriptomics and you find the cell types. There's been also a lot of development, exciting work done there. Then on the, you know, that's even at the lower level, like at the genetic level, let's say. Then there is the anatomical level or the connectomics level. You know, a lot of people are, develop techniques to do electron microscopy and now there's other methods like with light microscopy, expansion microscopy. But then the other level is, which I think is in many ways is the most exciting for me because it's the level that is closest to the computation and is the level where we can compare brains and machines directly right now, is the activity level. You know, because when you take a brain and you slice it, it's dead, right? So you're studying structure.

It's frozen.

Yeah, exactly. So it's not like computing anything when you're studying it. But now this technology to look at the computation live as it's happening in the brain, and then you're extracting the activity of the brain, and then you're trying to relate these activity patterns to the computation. And that's happening, you know, for example, if you go and you type a thing on ChatGPT or a large language model, or you use AI on your iPhone, is running a neural network that also generates activity. There are, as we said, individual neurons and they have activity and you can now study that activity that the artificial brain generates versus a real brain, compare them directly at this activation level or what we call the representation level. So that line comes from the school of physiology basically.

Okay. Yeah. So, yeah. Because as long as I've covered connectomics, I know it's fascinating, but there's always been naysayers that just say, well, I mean, that's great. Yes, it's useful to have this ultra-detailed picture of the brain. We'll learn more about the nature of the brain, but like, unless you see it working, we just don't have anything to correlate what we're seeing to. And I know that's really at the heart of what you've been doing is to get to this correlation. And I know recently, I think it was, I mean, we're recording this in April. I think it was, it was I know there was a Nature paper, but then I read a New York Times story from—

Yeah, it was a series of Nature papers that came out together back to back.

Okay, so, and then I'll try to set up some framework. You tell me where I'm going astray, but I think it was like in the 1980s, maybe the mid-'80s, that we did make this map of a worm brain. And so this was kind of like this first stab at connectomics. And then We've gone all this way until last year, we had a full map of the fly brain. It was kind of crazy that you go this huge length of time to go from the worm to the fly. And then the next missions have been the mouse and the human. And the thought is that getting a full map of a human brain is extremely daunting because of the amount of data and the incredible amount of wiring inside the complexity of the system. And then to to start doing stuff where you're, you're watching that brain, you know, it gets, it gets extra daunting as well. But so you guys, you just came out with this paper where I think it was on mice. Yeah. And, and you, you, you did what, the series of papers and you put electrodes into mice? But it was imaging, but it was all imaging. Yeah. This part. Okay.

It was imaging, but it was basically combining So what is unique, this is called the Microns Project. It was funded by IARPA as part of the BRAIN Initiative and then NIH too.

How does everyone, you guys always get these acronyms, Microns.

Yeah, so Microns, it was an acronym that came from IARPA when they formed the program. And IARPA basically created, they wanted to also push connectomics as a technology. So they created this program as a way to also put funding for connectomics, but also combining structure and function.

I just always wonder how much time gets spent as people shape the acronym for these.

Yeah, well, I don't know. The acronym wasn't made by us. It was made by the program officers at IARPA at the time.

I'm always impressed. Okay, so sorry, sorry. So tell me the— So IARPA is like the—

It's like DARPA.

Which is the R&D arm of the Defense Department, and IARPA does weird—

It's more of the intelligence department, but they also fund basic science and basic research, and this was part of the basic science funding. And I think when the Brain Initiative was announced, a bunch of organizations, NIH, IARPA, DARPA, they contributed to its mission. And this was an initial IARPA-funded project started by Jacob Vogelstein and David Markowitz at the time.

So the US government wants to fund basic research. I mean, they want to help us understand the nature of the brain. And then maybe they do interesting things with that knowledge at some point.

Exactly.

But we'll give them the— there's a purity on one level.

Yeah.

Yeah. To what they're doing.

Well, also, you know, talk about funding, like a lot of them, you know, when we say, you know, a lot of the stuff that we as society and companies, let's say the company are big tech here in the US and around the world benefit from is originally government funding, de-risking this stuff. You know, whether this is drug discoveries or it's usually like, you know, companies they cannot take too much of a risk because of their, you know, return on the investment or because of their mission. So often like the government, federal government research organizations in the US and around the world, like NIH and others, they fund basic research. And then, you know, people discovered, oh, you know, there is a non-coding or, you know, and then it reaches a stage that then in order to be commercialized, companies take over. So our whole society benefits from funding this basic science research. I think this is something that we need to sort of emphasize, that without that, we would stall scientific, biomedical, engineering progress, you know. No, you're right.

And there's like an enormous list of massive things that have come out of DARPA in particular, you know.

Yeah, exactly.

And it's in this particular project, the On that timeline that I was talking about, about worm, fly, mice. So we're not able to do the whole mouse brain yet because of the complexity and the amount of data. But you guys, if I'm not mistaken, you went after the visual cortex.

Yes, it was the visual cortex. And what was very unique about that experiment that was not true in the fly, in the previous one, or in other electron microscopy studies, let's say the human, was that for that same tissue that included visual cortex, it was a couple of visual cortical areas. We also, before, you know, the Allen Institute sliced the brain to do what Jeff Linkmann does, you know, at the Allen Institute, they did data from exoplanet at the Allen Institute and create this beautiful data that then Princeton, Sebastian Sagat, Princeton like segmented. We, back then I was still in Houston, in Texas at Baylor College of Medicine. We did high-resolution two-photon functional imaging. Okay. When the mouse was alive and it was watching movies and stuff like that. In fact, one of the movies we showed was Matrix and I think some people are fascinated about how we showed the Matrix part.

I want to talk about the movie part, but keep going for a second.

Yeah, so basically we, so we basically took this mouse It was live, he was watching movies, we're studying his brain, you know, activity in the brain as he was watching the movies. And then for that same brain, then the Allen Institute took it and did the high-resolution electron microscopy. And then we aligned this data together and that's very unique. This is the only— Yeah. Okay, wait.

So there's one group of living mice watching movies.

Yeah.

And then there's another where the brain did get frozen in time.

No, it's one mouse. It's the same mouse. Ultimately, it was one mouse. We imaged, let's say, a few mice. And then one of the mice that had the best data was selected because it was a very expensive experiment to slice it. Oh, I see. So it was exactly that mouse. We actually shipped the mouse live to Seattle from Houston.

It was like a prized mouse.

Yeah, it was a prized mouse. The mouse— had, you know, a private jet. That wasn't a private jet. He flew business class. No, no, but really, no, no, no, no, we, you know, we flew it like, you know, we took, you know, we flew to Seattle and then this mouse was, um, you know, basically they, they performed in that same mouse that we had the functional data.

I see it.

We, they performed the electron microscope. So then it was a matter of like computationally saying, this neuron that here is how its connectivity is, or here is where it is, the electromicroscopy volume, in the functional volume it's here. And then we kind of align this data together.

So you— okay, so you gathered all the data with electrodes while the—

with imaging? Yeah, but it was measuring activity while I was watching.

A movie, get this activity, and then the mouse did meet its demise for the last bit of—

Yeah. And then the mouse, yeah, ultimately when in order to do the anatomy and the connectivity, you know, the mouse, you have to like, you know, do the microscopy on the brain.

But then you— so, okay, but then you have these two datasets and you try to—

Exactly. Okay. So this is very unique in the fly, for example, the previous studies, electromicroscopy, there was just the anatomical data. There was no function.

Okay.

So it was very unique because now you can say these neurons or this neuron does this. How does its anatomy look like? What does it look like? Who is it talking to? How is it connected?

I see.

It's a very unique dataset because of this stitching.

Okay. And when— I mean, I'm making light of this, but I find it fascinating. And I've seen this in person. I went to Mackenzie Mathis's lab in Switzerland. I watched mice watching video games and movies. I mean, and they had, they had sort of a glass panel on their, on their, where their skull used to be to, to gather data. When you talk about a mouse watching a movie, I mean, and you were saying The Matrix, like to people who have no idea of this field, what does that even look like? You just have a mouse in a little thing.

Yeah, basically the mouse, you know. You know, the mouse is— and you can train it to also play video games, kind of, of one sort of that. That's probably something you see in MacKenzie's lab.

Almost like a Mario Kart.

Yeah, exactly. Like they're playing a game.

Yeah.

So in this case, the mouse was, you know, sitting comfortably on a— you know, mice like to run on a treadmill, so it was like on a treadmill running. And then there is, as you said, a little part of its You know, there's a little window on top that we can then put a microscope, a very powerful microscope that does not just look at the structure of the brain, but looks at the activity, particular calcium, you know, neurons have calcium like.

So it's kind of fluorescent.

Yeah, fluorescence. It's actually like GFP, like what Jennifer Lee, she has, but it's a specially modified version of that that is sensitive to the amount of activity. So, you know, if the neurons basically, they're silent, they're like dark, and if there is activity, they become more green. And we measure that. And there are ways to measure thousands of them at the same time. And the animal was— it didn't just watch one movie, it watched many different movies to sample the complexity of the world that it could experience.

Like it's watching like outdoor—

Yeah, like he watched like YouTube movies. You know, maybe next time we'll show you a few of your podcasts. Yeah, next time we'll show your show.

Yeah.

Actually, no, your show with the— Space Race would be good because mice are very sensitive to motion. So anything has a lot of action. For example, Mad Max. We chose a few movies that we knew were having a lot of motion energy in them that would be driving the activity a lot in mice.

You know, I'm so fascinated by this. So that's what you're looking for is the brain reacting.

Yeah, the brain reacting. You know, one thing that is fascinating is that, and this was not appreciated, And it's actually at the heart of the field of AI. Vision for us humans, it's very easy, right? Like it's very difficult for me to sit down and maybe solve some math problem or think about science or you would like have to like spend time setting up the business or setting the studio, how beautiful it is and blah, you know, all this stuff. But like vision is like, you know, you came, you recognized me and said, hey Andreas, how are you doing? You know, we hacked each other, right? But if you think about it, and people thought this is easy, right? We'll— and then there was this famous Dartmouth project in the '50s where they hired some undergrads, you know, to basically build an artificial visual system. And then when they tried, started doing it, they realized, oh my God, this is a really hard problem. And then over the years, from the '50s until now, we realized that in primates, and we're very visual creatures, like vision is our dominant sense, unlike mice and unlike other where there's people function and stuff. About 30 to 50% of our cortex, which is the most, you know, evolved part of our brain, is actually processing visual information. Okay. So it's a very difficult computational problem. And nowadays, you know, if AI excels, let's say at large language or even solving math problems, but it's still vision recognition problems that it lacks humans behind.

Just this complexity.

Because of the complexity. And it's a very difficult problem. Like, you know, we do it effortlessly because we've evolved to not think about it because we wouldn't survive in the savanna if we had to think about it. So when mice sit in front of the computer screen, they watch movies and these movies are composed of motion, objects, things moving around, and we're trying to understand How is it, how are they essentially extracting? Because if you think about it, what is the goal of vision? It's just photons hitting our retina. So you can think of it as some noisy photons hitting the retina and the brain has to infer or make a prediction of what is the causes of these photons. So it's not like, you know, we open our eyes and we, you know, we see what's there, but the brain has to compute it. It's not like it's on the camera.

Like a translation.

Exactly. It's not like, you know, you know, so basically It's like a noisy signal comes in, the brain has to figure out what has caused it, and that creates our person. And that's how we can also dream, right? We can, the brain can, without an input, generate things that look sort of realistic.

And when you're gathering this data, I mean, how specific does it have to be? Has somebody watched The Matrix and cataloged what's happening second by second and matching that to the data?

Yeah, yeah, basically the way you do it is the, You know, when you build this equipment to do the recordings, everything is synchronized. Like, so we, when we show something on the monitor, the data acquisition that we're acquiring the data with the imaging system, they have to be synchronized extremely precisely. So we know when, let's say, this scene appeared or this frame on the movie, it corresponded to this activity pattern. When it changed to this, it corresponded to this. So these things have to be perfectly synchronized. And the behavior of the animal. So for example, if this scene appeared and the animal started moving its eyes or it started running, we need to have everything measured very precisely to make these correlations between the stimulus, the behavior, and the activity in the brain.

And the mouse is kind of, okay, so it's on this treadmill, so it's a little bit sort of fixed in place with this screen in front of it and then a microscope.

Yeah.

Right over its head. And then it might watch— so this one special mouse, like how many movies did it watch?

He watched— basically this special mouse watched, you know, it was imaged for 2 to 3 hours a day for a couple of weeks. So he watched like a lot of different movies.

And like when you're calling out, I mean, you said Mad Max because of the motion. Yeah. Well, I mean, could you tell that it likes any kind of—

well, we don't know what it likes, you know. We never, you know, we know that the brain of the mouse likes certain movies versus others.

Okay.

And we kind of have an idea why, you know. Mice have a— unlike us, they have a low spatial resolution visual system. It's almost like, I think, they look at the world with a very blurred filter. So, and they're more, you know, if you think, you know, we We are very visually strong creatures. You know, basically vision is our strong sense, right? We see the world through our visual system mostly. And mice, you know, they use their vision for specific things like to avoid predators or, you know, and they're very, you know, as I said, they have a low spatial resolution so they cannot like discriminate details. But then they are more sensitive to movement. So movies that have a lot of movement and changes in spatial temporal dynamics, you know, the mouse will respond more to those movies. So this was one of the type of movies we showed, basically trying to extract these movies.

Okay, okay. Did the special mouse have a name?

It was called, you know, different people in the lab gave it different names. We used to call it the Platinum Mouse.

Okay.

I think the Allen Institute of Bristol called it the Mini Mouse. Mini.

And it was just special because it was responsive, it was on task.

And well, yeah, I mean, it was basically, you know, some mice will fall asleep, others are like, you know, you know, this mouse was like, it gave very good data, you know. It was, it was a very alert mouse, let's say, and it liked the movies maybe or something. So, and then it was also When the mice went to the Allen, there was a combination we had to collectively with the Allen Institute and us select one mouse that was the winner mouse that would go through their electron microscopy pipeline because that was a very heroic effort. It took many, many months to get the data. It was a very expensive project. So we had to select one mouse to get to this, as you've seen in Jeff's lab, to do this thin slicing.

Okay. Okay. And then I was reading this New York Times story about all this and last night, and now I've, I've, I'm trying to refresh myself on, on, um, some of the information in there, but like, what was our, what was our biggest takeaway from this experience? What did we learn?

Yeah. So there's a, there's a number of things. Okay. So I would categorize them into 3 main things. Okay, the first, and this was mostly work led by the Allen Institute, and it was on the anatomical side. He was trying to figure out, you know, the brain is composed of many cell types. And, you know, for example, our blood has red and white blood cells, and we know what they are sort of doing. The brain, we still are trying to figure out what is the logic behind the cell type diversity.

So, I mean, different types of neurons.

Different types of neurons. Yeah. Like there are neurons that, for example, we know there's major two classes. There's the excitatory ones that excite each other and the ones that inhibit each other. And then within each of these big families, there's many different types. And we knew this since the days of Cajal, like 100 years ago. They morphologically look different. You know, neurons have basically the area where they receive information, like their antenna is called the dendrite. And then they have the region where they send information out called the axon. And all of these neurons, if you look at them, they're like trees. You know, a redwood tree looks different than an olive tree or a banana tree. They look like that. You know, like you look at them, in fact, they do look like trees, a lot of them. And they look differently and you can classify them. So the electron microscopy allows you to do this morphological level characterization. And then one of the discoveries in this study was like the way, you know, they're special, the way they connect to each other. They're not just like randomly connected because you could think of them like they're different types, but maybe it's just random, the connectivity, right? So what we found, or what came out of this project, mostly work done at the Allen Institute, was that the special rules of who connects to whom based on their morphological characteristics. And of course, this was known before from other studies, but it learned new rules and developed a much more detailed let's say codebook if you want, of who is going to connect to whom at this morphological level. So this was one of the discoveries done at the anatomical level. And you need that large-scale data to do this. Another discovery that was part of the Pi-Gauge that was done actually with collaboration with University of Göttingen, my collaborator Alex Ecker, is these excitatory cells, you know, the cortex, the neocortex, has basically 6 layers and there's 5 of these layers, layer 2 to 6, they have excitatory cells. There's no excitatory cells in layer 1. And the idea is like, are these distinct types or is it more like a continuum? You know, because we don't have enough data, you know, you could think of it like if you are studying redwood trees and you go into a forest and you see some are very big and some are very small, but you say, oh, you know, there's two classes, big and small. But now it's like, well, let me like go into many parts of the country. I don't know, go to Muirwood, go to, you know, different places in California where you have redwood trees, and I'm just collect like hundreds of thousands of them. And then if I do some statistical analysis, maybe it's a continuum. Maybe it was my sampling that made me think. So what Alex and his students found was that if you look at this larger scale data, maybe it's more of a continuum. And that may change the lot because as I said before, you know, a lot of the systems like the heart, the liver, we know the types of cells and we understand the function somehow. But here we still don't even understand the logic or the categorization of these cells. So this was what he found was that maybe it's more of a continuum with these excitatory cells.

Okay.

And it still needs more research. In this direction. So, so this was another anatomical study. The other studies— and this is what is truly unique about this data, right? Again, this other data that did electron microscopy in different species, including humans, as you know, by Jeff. But this data had function, okay? And this is where, you know, the gold is, right? This is where the money is, I think, is the function. Because That's where the computation is happening. It's a live tissue that computes. So in this mouse, we had both structure and function. And for the first time, we could study, try and figure out what is the relationship between structure and function at the large scale. There's many heroic studies that were done before. A lot of them were done by doing very complex and heroic experiments where people would, let's say, go and use fine glass pipettes and patch 2 or 3 cells, study its function, and then at the same time to slice physiology and figure out if they connected or not. But here for the first time, we had both the anatomy and the physiology in the same mouse. So I'll tell you first what we discovered, then I'll tell you, if you're interested, how this discovery happened. Yeah. Which is by building this digital twin of the mouse.

No, I do want to get into that. Yeah.

So basically, if let's say I'm a neuron and I have some preference what I like in the world to fire, in vision you can think of its preference is broken down to what part of space I care, because that's called the receptive field of the neuron. So for example, if I'm looking straight at you, there may be a neuron that likes the upper left corner of my visual field. So if there's something there, it cares about, but if there's nothing there, it doesn't do anything. And then what is that thing that is there that it cares about? And just to make it simple, though mice don't see colors the way we do, let's say it likes red in the upper left visual field, but it doesn't like if green is there or yellow is there or some other color. It just likes something red. Okay. So the question is, if I am a neuron, who do I connect to? Do I connect to neurons that are similar to me, completely different than me? What is the rule? Like, who do I talk to? And there is, we know from previous studies, there's a very, you know, very well-known rule called like-to-like. If I like something, I'm more likely to connect with other people that like it. It's kind of like talking to people, like people have similar interests. They're more likely to talk to each other, right? So it's kind of the same rule. But what— and this was— no, but what this data allowed us to do was to ask the following question. There is a very famous rule in neuroscience called Peter's rule that says that if I'm a neuron, you have to imagine now that these neurons, as you've seen in Jeff's lab, It's like, think of it like you're in a forest and it's very dense. Each branch is like, it's much denser than any forest you can ever imagine.

The images are crazy.

Yeah, exactly. It's like, it's like each branch is like, and each, you know, leaf is like next to another leaf. And it's like, there's no, you cannot move through that forest. It's like, it's like a jungle, like, you know. And then the neurons, they send as I said, they have the antenna and then they have this broadcasting area called the axon. And these axons are very complex. And Peter's rule says you send an axon and then if the antenna of another neuron is close by, you flip a coin and you connect with it. It's a simple rule, right? You don't need a lot of complexity for this because let's say you need complexity, but you don't need to select who you talk to. Like based on what maybe they're doing or based on some more complicated rules. You sort of like randomly flip a coin.

There's no method.

There's no, yeah, it's like, and in some ways it's an important rule because maybe you create complex networks like that or maybe if you wanna build AI systems, maybe you'll use that rule or something, okay. So it is called Peter's rule. And what we found, because for the first time, we had this high-resolution anatomical data, we could ask the question, if I'm a neuron and I like red in the upper left visual field and I send my axon into an area, do I randomly flip the coin? Because in that area, they are also going to be neurons that happen to like the upper left visual field. But there's some variability, some— let's say most of them will be upper left visual field, upper, a left visual field with liking red. But there will be some exceptions. Some would like green, some will be a little bit further down, some would like yellow, some would like maybe red and green or a combination. So what we found was that this Peter's rule, this flipping of the coin, connects neurons based on their spatial relationships. And if anything, we even found a little bit of a repulsion, like If they had perfectly overlapping visual field, for a very small probability, they will almost avoid connecting. But the feature they like, whether they like red, green, I don't know, of course mice don't like this, but I'm just making this as an example. If they like this cup or the Coke, they would not do this flipping of the coin. They would most, there was a probability where they would select the one that, was more feature-like. And, you know, and this may have implications on how brains wire up. You know, we know that a lot of diseases, for example, or we think, for example, a lot of diseases like autism or schizophrenia or, you know, you know, basically mental disorders, right?

Like depression.

Yeah, depression. They may be wiring issues, right? And it may be these subtle changes in the wiring, these rules Okay, this is a hypothesis, speculation, but for example, rules like this, because this is like a simple rule, right? You say, I'm gonna connect neurons randomly based on what part of space they lie, but then I'm gonna be more selective in what features they like. Now you could think maybe that's like more complex rules like that or rules like that for other domains of cognition. So this is one of their studies that tries to understand the wiring of the brain and how it's related to function. At this level of precision. And this would not have been possible if we only had the anatomy, we couldn't do it. If we only had the function, we couldn't do it. It was only because this Microns project both put function and structure together to enable us to study this rule.

Okay, and this is fascinating. Um, I'm keeping an eye on time because I know you have meetings.

Yeah.

So, you know, but I'm going to go off what you said in the some, I think, interesting areas. You mentioned digital twins for a second, but okay, so here we are, we're on this moment where on one hand we've got, okay, so we have all these fascinating microscopy imaging techniques, right? We have, um, there's companies like, uh, E11 that are trying different techniques to sort of map out the brain with viruses and different things. We've got, we've got I don't know what, like 10 BCI startups putting electrodes into actual humans and just funneling out data. And then we've got obviously AI able to make sense of patterns of data and do all this enormous computation. So, okay. So, and we're, so we're kind of like learning a lot about the brain on one hand, making artificial brains over here. They kind of are complementary and overlap. When you talked about digital twin, I'm not 100% sure if this is where you were going, but I mean, I mean, obviously the holy, a holy grail would be, would be some kind of complete representation of how the brain works. And then some kind of union with computers, or maybe, I mean, if you're into that, right? Or is there some sort of bond between these intelligences. I mean, is that, is that like, is that where you were going a little bit, or is that— am I crazy? Is this where everyone—

no, no, no, I think in the future we'll be seeing things like that. But specifically what we did, yeah, you know what, let's say one important stepping stone is the following. If you, if you could, if you want to study a system Let's say this mouse, I said, I don't know, we did experiments. I don't know, let's say we show it 10 hours worth of movies or 20 hours worth of movies, but there's like millions hours worth of movies, right? And this mouse, you know, even if we did experiments, even if I did experiments every day of my life, there's so many experiments I can do because I'm only gonna live a certain amount of years. Right. Right. Now, and because, we're doing something in the physical world, we're limited by, you know, our lifetime ultimately, right? How many experiments we can do. But now imagine that if I had a way to create a model, a functional model, a digital twin of that system, then if I have enough resources and I buy enough GPUs or enough computers, I could run what would take 10,000 years, I can run it in a an hour simulation. Because yeah, exactly, I can just download all the video on the internet, or, you know, and then just show it to this digital twin.

It's like a, it's a digital representation of the mouse's visual cortex that's watching the movie.

Exactly. So this movie, basically what we created, and, and this used deep learning, it used similar architectures like ChatGPT, like large-scale data and neural networks. So what we did is We collected this data and then we trained a model, not to be a biophysical model like the Human Brain Project and the European project, which was criticized because of that level of analysis, but the functional model. The goal of this model was to say, can you predict, it was 75,000 neurons in this mouse, can you predict, how well can you predict A, all, each one of the 75,000 neurons, how will they respond to a new movie that I never showed to the model to train it with? But we did show it to the animal, a lot of them, so we knew the ground truth. So then we could ask the question, how well could this model predict? And once we built this, and it was built using a foundation model approach, Then we created a good enough model that then we could basically show it all kinds of stimuli. The mouse is dead now, but one could download its digital twin and still run experiments and then go back to the anatomy and maybe make a discovery. Of course, this is still in the digital twin. You have to be careful because like ChatGPT or other models can hallucinate, the digital twin can hallucinate and tell you stuff that may not be true.. So you have to develop hypotheses and then to maybe new experiments to test it. But this is the idea of the digital twin. And it's not, you know, we're not the first to use this term. I mean, it's used, for example, in the aircraft industry, you know, when, you know, they want to, engineers, when they want to optimize a system, often they will create a digital twin of it and they will run the optimization in the digital twin because they can do it cheaper and at scale.

Yeah. You mean when they take like the, all the electronics and just, just kind of lay— Yeah.

Or they create a simulation of the, of the, of the airplane or the wing of the airplane.

Yeah.

And then in the simulation they can say, okay, I'm gonna change the wing in this way.

Oh, sure. Yeah.

And see, do I get the, you know, so of course ultimately you have to test it in the, in the physical system or the biological system. And this process of collecting data Building digital twins and then testing the bug is very powerful. And this is one of the ways that we're seeing accelerated progress of AI being used in science.

Possibly incredibly stupid question, but how do you show a digital twin brain a movie? It's just seeing the code from the movie?

Yeah, it's basically the input to the digital twin is trained, you know, like it's a neural network. And right now, you know, if you open, let's say, your chat box, you just type code, right? But you can also download an image, right? And say, what do you see?

Yeah.

It's the same type of code.

It's looking for all the edges.

Yeah, if we download the code and we show it an image, but instead of predicting what is the image, that's what, let's say, large language model or a multimodal model does, predicts neural activity. Okay. So it's exactly the same. Idea, but you can think of it— its input is visual movies, is movies basically, and its output is neural activity instead of being, oh, this is a carrot or a cucumber or whatever.

All right, this might also sound stupid, but it's always kind of fascinating to me. I mean, so basically the brain is our most— I guess all our— most of our organs are important and you need them to be alive, but I mean, it, you know, it obviously shapes everything about our reality. And I think it's sort of safe to argue it's maybe the organ we know the least about.

Yeah, definitely. I would say it's safe to say, which is crazy, right?

I mean, but I guess this is because it's our most complex and hardest to study for, for not just because of its complexity, but because you can't, it's harder to get access to a lot of brains to, to do this kind of work. When you see what's happening in the world now with a lot of the technology we just mentioned, plus AI. I mean, what do you, what do you think? I mean, do you think we get to a much deeper fundamental understanding of how our mind works in, in what, 5, 10, 15, 100 years?

Yeah, that's a really good question. I think My hope, you know, is that I would say in the next, you know, 5 to 10 years, and maybe even earlier because, you know, there's early indication for that, we're going to start what we call cracking the neural code in a way that it will be an accelerated cracking of the code because we're already like know a lot of stuff about this code. And as I mentioned, a lot of people got Nobel Prizes because they understand, let's say, how edges are represented in the world, or maybe how spots of light in the retina, or how a rat is encoding his location in space, right? But most of the stuff we really don't know. You know, how is it, as I said again, the correspondences between the outside world, our actions, our feelings, our thoughts and what's happening inside the brain. And I think that in the next, I would say, 3 to 10 years, we're gonna start seeing an accelerated rate of discovery. And this often happens if you look at the history of science or engineering, you know, it's like you have incremental and then you have an accelerated, almost an exponential increase of knowledge. And I think a lot of it is going to be done at this information processing level. And it's gonna be a combination of large-scale data collection and AI. And it's gonna help us building these like digital twins, accelerate the progress of discovery and understanding the neural code. And why, what do I mean by the neural code? Just to be clear, as one of my colleagues, Sofia Sandberg, likes to call this like a Rosetta Stone where we're relating, it's like a translation where we can relate It's an image to an activity pattern. And if we can understand this translator, that means we understand something mechanistically about that code. It has to do with interpretability of that code. And it's also related to things like, you hear often this word that the brain, artificial neural networks are like a black box. We don't understand it. That's why we don't trust them. And there's AI safety and stuff that we've talked about before, right? So if the code is interpretable, for example, if you understand how the brain says the truth versus it lies, I'm just making this as an exaggerated example, but just to give you, and if you could build something like an artificial system, then you could not do jailbreaking, for example, because you'll be able to like see it, we humans or other AI systems, look at it and understand the inner workings of it. So basically this is a fundamental thing and I think that is one of the things that's gonna come. Now if that happens, then the next step, what you said is all these PMI companies, if you think about it, for us to build an interface to the brain, we can optimize that interface if we understand the code because it's like language. Like if I wanna like, if I go to a new country and I wanna communicate, If I understand, if I have a translator, sure, it'll be easier for me to communicate. Now, there are ways I can communicate, maybe by hand waving and doing all kinds of stuff, but you know, if I—

a little bit.

Exactly. If I go, let's say, to Tibet and I have someone that speaks Tibetan and also English, or Greek in my case, you know, I'll be able to like communicate better, right? So by cracking this code of communication because between behavior and perception and cognition and activity, then we can interface a device that's going to have, let's say, what I like to call the exocortex or artificial neural network, communicate with it, whether this is to write information for blindness or to read out information for, you know, driving a robotic arm, or ultimately maybe to enhance our cognition, right, to hook me up and give me like, I don't know, a very large bandwidth of memory that like I don't normally have in my brain. But I think there's, you know, I think ultimately, you know, if you look at the history of scientific progress and industrial progress, we always had big revolutions when we had a scientific understanding and not a brute force engineering approach, right? You know, and I think right now we may be more brute force engineering it, and that's why, you know, and I think we'll see a stronger impact. We have a true scientific understanding of intelligence, and I think that's where we'll see a lot of progress in the next 3 to 10 years.

All right, last question, just because I, I would ask you a million more. We didn't even get into like Neuropixels and stuff, which I really want to— people can go, go look that up on the internet, or we'll have you come back.

Yeah, yeah, I'll come back for that.

Okay, so if you're saying 3 to 10 years hopefully to get this new level of insight, develop some of this language, develop some of these bonds between our flesh brain and our computer brains. I mean, you know, you already alluded to it. I mean, my sense from you is you're kind of, you do see a cyborg type future and that you're not horrified. Is that true? What does the world look like in 25 years?

I think You know, it's hard for me to speculate what it will look like in 25 years because, you know, like there's all kinds of, you know, examples historically where we underestimate and overestimate at the same time technology. But I do think that, you know, I mean, in some ways we're already like having iPhone devices. I mean, if you think about it, the fact that we communicate with AI and ChatGPT now, or, you know, through our vision, I mean, our vision is a high camera input, is a BMI to drive a brain, right? So that already exists, the fact we have hands. So, you know, we've evolved to have, you know, interfaces to communicate to our neural network, and now we've built external devices like phones and computers, you know. So, so that's going to accelerate our way to do science, you know. For example, let's say You want to design this room and optimize it so people feel better. Or let's say every person that comes on the podcast, you optimize it for their happiness, right? Well, I mean, maybe, you know, you can have a computer help you do that. And in some ways that's a cyborg, right? Like you're, you have a human and a machine working side by side. And then maybe they're communicating through language or through a camera. Right now, whether we're gonna have a direct interface to the brain, that's also gonna happen in the future, I hope. But I—

You do hope.

Yeah, I hope, and I'm not worried about it, but I don't think it's something that necessarily will happen in the next 5 to 10 years. And the reason is because, you know, there's a lot of human, you know, unless, I mean, depends on the device. The more invasive a device is, it's a trade-off, right? If there is a non-invasive device that we can all try, and there are people trying these things, but they don't work very well right now in a way that is useful. You know, people try this EEG to, I don't know, control something, but you know, it's not like worth it to be carrying this thing on your head because, but you know, so if, but if it's something that is the more invasive it is, the more people are not going to adopt it. When I can just read it on my phone, right?

Like, why, why are you gonna have brain surgery?

Yeah, they're gonna have brain surgery because it's gonna take me a little bit faster to like, you know, read a book, you know, because I can have— if I want to do it faster, I'll have an agent read the book for me and then summarize it to me, or I'll ask it questions, right?

So, but if you could download kung fu— yeah, exactly, maybe. Yeah, if I could learn skills and, uh, So, okay, so you're not horrified by that future. And I, I'm not trying to put a date on things because I know all this stuff is really hard, but it sounded like you were saying not horrified by it, kind of like maybe in the next 10 years there's like a universe where that happens.

Yeah, I think we'll see in patients an accelerated form of brain-machine interfaces. I hope we do, you know, because It's been a lot of investment in new technologies, you know, to interface to the brain. And I hope that, you know, right now, let's say the state of the art in many ways that is more standardized and used, you know, it's like things, technologies from the '80s basically, including things like deep brain stimulation or the UTI, you know, we'll see new forms. There's been a lot of investment. Not just in industry, but also again, like DARPA, NIH has put a lot of investment in this. So there's a lot of technologies that I think they could be optimized and be deployed to human patients.

Yeah.

Like for blindness, for other kinds of prosthetics. So I hope that in the next 10 years we'll see a lot of that. But I think to get an invasive device inside our normal people's brain.

Yeah.

To give them these extra capabilities. I think that's gonna be longer. It won't, you know, unless there's something traumatic happens because, you know, like 10 years, I mean, if you have to think the last 10 years, we've still had these technologies and they still haven't been deployed at scale yet. But I think that's what we're gonna see next, you know, impatience at the beginning.

Okay. Can I sneak one last question?

Yeah, sure.

Yeah. There is a danger that this gets expansive, so try to keep it mindful of your time.

Okay.

But if we learn all the mechanics, if we learn more of the mechanics of the brain, oh God, I could be setting you up for trouble on this one. Do we find things like consciousness, you think? I mean, I know this could be like a really abstract thing. Like, do we just find the, yeah, like what level of mechanics of how the brain functions do you think we end up uncovering?

Yeah, I think that— I mean, you know, you know, we all— neuroscientists, we all like to think about consciousness, but frankly speaking, for me, I never really could. I mean, there are interesting experiments, and, and I, you know, even like stuff that when I was a postdoc, the person I was working with did this very cool experiments with things like bistable perception. I don't know if you know, like the Necker cube. Like, or the old young woman, like if you show an image, the same image, but your perceptual flips and recorded from neurons in the brain and found that some neurons were like flipping with the percept of the monkey, but others were like not. So maybe you can see, and then Francis Crick at the time was very excited about this. You know, so there's very cool experiments, but I do think that consciousness to me You know, I've never, you know, I think there's very interesting philosophical ideas.

I know, I'm sorry.

Yeah, no, but it's never, you know, and I'm not sure, for example, if you say what is the computational advantage of being conscious versus a zombie? I don't think there's even a good argument why we need consciousness to build intelligence. So maybe it will just emerge or, you know, it will be some sort of emergent property or maybe there's something very different about it, like what Penrose says that, something that we cannot right now with experimental study. I don't know, but I'm not as— I'm not worried about it. You know, I don't think it's a priority as a human species to understand consciousness. Yeah, I think it's our priority to understand intelligence, behavior, so then we can understand the brain and fix it for mental disorders and stuff like that. Then we can build AI systems are more aligned with us. They don't end up exter— killing us all eventually, you know. But I don't worry as much about consciousness right now. Yeah, yeah, I think about it, but I'm not worried about it.

I know, I was—

doesn't keep me up at night.

This would probably be another multi-hour episode if we fully explore that. Um, Andreas, well, we, we didn't talk much. You have this at Stanford, you have— you're working on this Enigma project. It's, it's a little bit, um, We'll come back to that in the future.

Yeah, we'll come back to that. We'll keep it enigmatic, right?

Yes, we'll have you back on the show, but you've got more research that you're pushing on and I mean, I'm so fascinated by your work and thank you so much for coming to visit and yeah, we'll update all our listeners on as you march along.

Yeah, yeah, thank you. I think it's been great, you know, seeing you finally in your new studio and it's an honor to be the first guest. Thank you so much for having me.

Maybe this is when the, when the AIs light up and they, they can appreciate this, this moment we were getting into all this stuff. Yeah. Um, so thank you very much. Thank you. And we will, uh, we'll keep everyone updated on your work.

Thank you.

Thank you. All right. Thanks, Andreas. Thanks. I hope everyone enjoyed that episode. It was sponsored by E1. Ventures. Very kind people. Turn to E1 Ventures for all your venture capital needs. The show was hosted by me, Ashlee Vance. Our guest today was Andreas Tolius. The show is also produced by me and David Nicholson. Our theme song is by James Mercer and John Sortland. Core Memory is edited by the percussive John Sortland. Short-lived. See you on the next episode.