The Next Step Toward Understanding The Nature Of Intelligence - EP 49 Sebastian Seung

What's very interesting about the fly connectome is that some of it already is readable, it's decipherable. And I like to quote, again, I'm originally a physicist, so there's a famous quote of Einstein, "Subtle is the Lord, but malicious he is not." Okay, and so what Einstein meant by that is that, you know, the universe is hard to comprehend, but it's not like God was evil and deliberately tried to make it hard to comprehend. But when you're working on the connectome and you're struggling to trace every path of every wire in the brain, you're thinking, man, God is really mean. But then when you start to look at the result in the fly brain, then you think, wow, God is, God is not so mean after all. Some of this is comprehensible. Hey folks, I'm Amar, Product and Design Lead at Google DeepMind. Have you ever wanted to build an app for yourself, your friends, or finally launch that side project you've been dreaming about? Now you can bring any idea to life, no coding background required, with Gemini 3 in Google AI Studio. It's called vibe coding, and we're making it dead simple. Just describe your app and Gemini will wire up the right models for you so you can focus on your creative vision. Head to ai.studio/build to create your first app.

Welcome to the Core Memory Podcast. This is Ashlee Vance, and I could not be more excited today. We're with Sebastian Sung. Sebastian, how are you?

I'm great. It's fantastic. I don't visit San Francisco area very often, but just driving the 101 and seeing the beach on one— ocean on one side and then the mountains on the other really makes your heart go to your route. Yeah, yeah.

But then you hit the traffic, and then, uh, that, that is, that is the downside. Um, well, I, I'm glad that we got you out here. This is a special treat. Um, I'm sure lots of people listening will know who you are already, but for those who don't, um, you are one of, if not the world's expert on the, the connectome, which we're going to explain to people, um, as, as we go through this. But you're, you're a professor, of computer science and neuroscience at Princeton and have this wonderfully distinguished career. You wrote a bestseller, bestselling book about the connectome last year. You gained a ton of attention, um, with a special issue in Nature for, for doing a full connectome of the fly brain. So this full map of the fly brain, and this was like a massive scientific milestone. And then you've also been in, in industry at different— you were the head of research at Samsung.

That's right.

Right. Yeah. And so you've had kind of this academic career and, um, and worked in, in industry. And then we're breaking some news on the podcast. You have a new venture, um, a startup. And, and so we're gonna dive into that. I love breaking news on the show.

Great to be here, Ashlee. I'm excited too.

Okay. Um, well, why don't we— okay, we're gonna be talking about the Connectome a lot, so I I feel like we should, we should define this for people. And, and I'm sure again, a lot of people listening will know what the connectome is, but I could not think of a better person on planet Earth to explain the basics of the connectome for those who, who, you know, might not be familiar with it.

Right. Well, as, as you, as everybody knows, a brain is a network of neurons that's been known since the night, really the end of the 19th century. But a connectome is a detailed map of the connections between the neurons in a brain. And so if you think about the flight maps in the backs of airline magazines, if you replace every city by a neuron and every flight between cities by a synaptic connection, that could be your mental image of a connectome, except that your, your map would be 100 billion cities and thousands of flights per city. So it definitely wouldn't fit in the back page of an airline magazine.

Quite complicated. Yeah, I mean, so this is a subject I got interested in, I want to say about 15 years ago, which is the first time I ran into you and your work. And back then there was a lot of excitement, but also doubters around how valuable this would be. So it often gets compared to the genome, um, in the sense of we're going to create this, this map of this very important part of the human body, in this case, the brain. We're going to show all the neurons, all the synapses, all these connections. I think we did the worm 40 years ago, right?

1986, it was published, the map of the worm nervous system.

Okay.

C. elegans.

And then we had this big, big, big gap.

It's about the connectome winter.

Yes. Yes. So, and then, and then the fly, we just happened last year. And, and then we've done chunks of like small chunks of human brain along the way. So, you know, there's been all this research that when I started digging in to this field. I mean, lots of people, yes, were excited. And then some people would say, this isn't really as valuable maybe as the genome. They were, they were kind of naysayers. They would say the brain that you're imaging is dead. It's frozen in time. You're not going to see any activity. We might get this map, but it's really not going to add to our knowledge of how the human brain works. I always found the argument a little Odd. It seems sort of obvious to me that we would learn something quite profound from doing this, but, but people were quite against spending money on this. And then, you know, you've really— you kind of stuck your neck out. I mean, you were, you were right, you wrote a book about this very early on, you know, not long after I dug into this and, and, um, and, and really did this pioneering work. All right, what do we do at Core Memory? We cover innovative, fast-moving forward-thinking companies, which is why Core Memory is sponsored by Brex, because Brex is the intelligent finance platform for many of these companies. 30,000 companies from startups to the world's largest corporations rely on Brex's technology for their finances. They've got smart corporate cards, high-yield business banking, and expense automation tools that are fantastic. I hate doing my expenses and Brex's AIs and software run right through those expenses, figure out where we're spending money, and take care of so much stuff for you so you don't have to waste your time on it yourself. Go to brex.com/corememory to learn more and just, you know, get with the program. Let's get going. Let's get out of this archaic finance software and move toward the future. CoreMemory and Brex.

You know, we started on this road. So there was a connectome winter after 1986. Uh, I guess getting the worm connectome was so difficult. It's just 300 neurons and thousands of connections. So it's very, very kind of tiny map, which would fit in the back page of an airline magazine. And yet that took over a dozen years of hard work to get. And so after that was a connectome winter when nobody wanted to work on it. So in the 2000s, people started getting interested again, mostly outsiders, not card-carrying neuroscientists, but people who came from physics, came from computer science and so on.

And you're a physicist.

Originally, I was trained as a physicist. Yeah. And I think, so one thing that you can't understand nowadays is how people thought, I think, in those days. And one big difference is the idea of connectionism. So today, everybody, since artificial intelligence is now driven by artificial neural nets, which wasn't the case then, um, everybody accepts the idea that the connections of a neural network are an incredibly important determinant of its function, right? That's what we train. We train the connections when we train AI, right?

100%.

And so if you believe that, then it seems kind of obvious that the map of connections in a brain ought to be important.

Yeah.

But in those days, people didn't—

were—

yeah, people, because AI was so puny and, and, uh, you know, awful in those days, uh, nobody accepted this idea that connections are really powerful. And I call that connectionism. It's a really old idea.

Okay. And, and to your point, I mean, some of this was, was, um, skepticism maybe around the idea, like around the winter, but then some of it is just, this work is incredibly hard to do. It's painstaking, it's expensive. You're taking a brain, you're getting these, you know, using what, like a diamond blade to cut these tiny slivers?

Right. It's a great story. So the fly, the fly connectome, uh, it actually started out in 2014, the quest to the fly connectome. A guy named Rick Feder, you could think of him as a virtuoso.

Okay.

He was able to cut a fly brain into 7,000 slices, um, each 1,000 times thinner than a human hair and not, not miss a single one. I mean, it's a superhuman kind of feat and Seems unbelievable.

Yeah.

Yeah. And it's a diamond knife, the world's sharpest knife, an edge that is atomically sharp. That's what has to be used to cut those kinds of slices.

And then you take the slices, you put them under an electron microscope, right? And then you get this wonderful high-resolution image, and then you have to stitch all these things back together. And then what you end up with is this mess of, of all these wires. You know, I mean, some people compare it to a bowl of spaghetti. It's almost, it's like more intimidating than that when you actually see, see the photos. So you, you get these images and then you have to now dig into like figuring something out and following these wires all the way through the spine. That's right.

Yeah, the stats are crazy. So the brain, your brain, um, if you think about the entangled branches of neurons, so it's a branched spaghetti of, of, uh, wires. Um, the total length is estimated to be millions of miles inside your head. And even inside a fruit fly brain, the total length of these wires is about a football field. Football field's worth. So although the connectome might sound like a glamorous concept, it's, it's, it's pain— it's sort of a painstaking, laborious kind of thing. You have to trace the path of every neuron's branch through the sequence of images that come from all of those slices. And, and the, you know, it would have taken maybe 100,000 years for a person to do that with the images from the fly brain. But luckily, We have artificial intelligence, right? AI does most of the tracing, which was part of the story here.

I mean, this was a big advance from previous efforts, was, was how far AI had come and being able to help out. So you have 140,000 neurons in the fly brain, and then you have, you've got, is it tens of millions?

Yeah, tens of millions of synapses.

Synapses. And so you're trying to, trying to label that, trying to follow these wires. And so the AI is, is going in and helping. Put different colors on these things, make this, make this map comprehensible.

Right. So AI has to find all the synapses. And then the tracing problem is like, imagine the brain is a gigantic 3D coloring book. You have to stay, stay inside the lines. So the AI does most of that job. Occasionally it makes mistakes. And so in the end, the final process, uh, humans Highly trained humans have to look at every neuron and fix the mistakes that the AI made. So it's not that AI replaces humans, uh, but it's a tremendous accelerator, right?

Yeah.

Yeah.

And I, uh, I want to get into the, what we learned about the fly in one second, but obviously people immediately have all been, well, we want to do a mouse next. We want to do a whole human next. Um, there, when you start talking about a whole human, I mean, the estimates usually get, this is a, a multi-billion dollar project possibly, and it could take, take many, many years, although there are several startups now that are doing different imaging techniques to try to help out, um, to gather all this. Do you, the, uh, well, maybe I'm jumping too far ahead, but no, it's a great question.

So people always ask, oh, how much money, how much time, like when do we get the human connectome? And so this is a sort of, uh, connectomics is a Moore's Law adjacent field, right? So we know that computer technology has been on this path of exponential progress, right? And we hope, I think many of us hope, that it will continue on this exponential path for a very long time. So if you think about the price of a connectome, let's say dropping by a factor of 2 every year, right? If a fly connectome was $1 million last year, then in 20 years, a human connectome would be $1 million. Yeah. And in 30 years, it would be $1,000.

Right?

Uh, on the other hand, if you want it earlier, you know, if you want a human connectome in 10 years, then it would be a billion dollars, right? So the sooner you want it, the more it's going to cost.

And okay. And there, and you said the fly work began in 2012?

Uh, 20, this particular fly began in 2014.

2014. And so it took about a decade.

Right.

And it was a race of sorts. I mean, there were multiple groups working on it.

That's right.

And you guys, I don't know if I'm upsetting anyone, I mean, you guys, you guys won, right?

Well, it's a, it's a strange, you know, So it's a— there's a lot of drama in the field of connectomics. Uh, one of the jokes we had was that, um, you know, in— as a scientist entering the field of connectomics was usually a career-destroying move. But, um, there were— so Howard Hughes Janelia, it's a research campus, they had, um, their flagship project was to get the fly connectome. And they had two teams, internal competition. This is a page actually straight out of the Samsung playbook. You know, if you really have an important goal, you have two inter— two teams competing inside the company, right? So they had two teams and Davi Bach's team, they finished imaging the fly brain first, but their data was judged to be inadequate. That it wasn't good enough and there was no way that AI could accelerate the process of getting the connectome based on those images. So Howard Hughes decided to abandon that dataset and invest tens of millions of dollars in developing a better imaging technique that would produce better data that, that AI could be applied to, to reconstruct the connectome. But we, Mala Murthy and I, my colleague at Princeton, we got the idea to fish that abandoned dataset out of the trash heap and apply AI, our own AI, to get the connectome. So it turned out to be a struggle between, I mean, there's many struggles that actually happened along this drama of the race, but you could think about it as better algorithms or better data. How do you invest your resources? And it turned out here that better algorithms won the day. We got, we got over the finish line first.

And you guys created a game along the way to get to sort of crowdsource doing, helping the AIs out, right? Well, or was that just human labeling?

No. So in, in, um, back in 2012 when, when the AI was really primitive, convolutional nets, we were, we were already using convolutional nets. We created EyeWire, which allowed, we gamified the human process of correcting the AI. But the AI has gotten so good now that it takes 3 months of full-time practice by a human to get good enough at correcting the errors of the AI.

Okay.

So we decided that was too hard to gamify. But we have, we have volunteers. We have, we have old-timers from Eyewire who are just really passionate about being, helping out with science. They're essentially amateur scientists. They helped out, but it's not gamification. I mean, you could think of science overall as being a kind of game, right? So, but it's not gamification in the kind of conventional sense.

Maybe this is a stupid question, 'cause we can just look at the images and see how accurate they are, but you know, there's so many wires.

Right.

Like, how do you know that the AI, you know, when I, if I do the equivalent of this with words, it's like the AIs are still making mistakes all the time. And if you make a mistake here, it's really throwing the map off in pretty bad ways. And so how do you know that they've, they've, again, I realize you can look at the picture and see if it's tracing the wire, but like, how often was it, how accurate are they? How often was it making mistakes that had to be fixed?

There's still mistakes.

Yeah.

Right. Just as in genomics, there's still mistakes in genomes. There's still mistakes left in the fly connectome. But the reason we, we can actually evaluate the accuracy by having humans look for more errors in a neuron. We can ask for humans to independently decide whether what the right answer is and see if they get the same answer. And then we can compare with images of neurons gotten by other methods, by light microscopy and genetics. And finally, just, there's just the, the science that comes out, right? If we, if we manage to get consistent results, scientific results out of the whole thing, that means that the error rate is low enough.

Okay. Which we are now. We're going to get into that. I mean, and I'll put some links in the post on this if you haven't seen it. I mean, the images are breathtaking and the complexity really comes through. I mean, there's both simplicity and complexity that comes through when you see how these neurons are arranged, but it's like so fun to— you can really explore this brain in strange ways. And I know I know we learned a ton of things. I mean, like I said, Nature did an entire special issue on this. People have, even after the Connectome was done, have been doing more work, more and more and more work, um, on this data. But I mean, just to give people a flavor of the biggest things that we immediately discovered from getting this map, can you give us just, just a taste?

Well, I should say that, uh, this map is used by every fly neurobiologist now. So. It's impossible to list all the things that it's being used for, but I can talk about my personal—

Yeah. Yeah. Yeah.

Yeah.

And I, and I mean, I was lucky enough to hang out with you a bit this weekend and just some of these images of like the eye and how the neurons are arranged is fascinating, but yeah, no, take us where, where you like to go.

My favorite part of the fly brain is the visual system. And I think most people have seen images of the eyes of insects. Right. And they have these facets that form a beautiful hexagonal lattice, like a perfect looking— it doesn't look biological. It looks like this, uh, um, mathematical kind of construction. And so the amazing thing about the fly visual system is that when you continue from the eye into the brain, it still has this lattice structure, this beautiful mathematical regularity. And, um, you know, I'm a, originally I'm a physicist and so I call it a brain crystal. It looks like it has a sort of crystalline regular order, unlike, you know, other parts of the brain, I guess, uh, look more like just tangled spaghetti, but in the fly visual system, it really looks like a mathematical object. Yeah.

There's all this order to it. And then I'm going to butcher this next part, but this weekend you were showing, I mean, some of the neurons, like a tree that is, is, or a plant that's chasing the light and is kind of like bending over really far. I mean, you have neurons that are part of the eye that's meant to kind of look to the right or sense an object to the right. They're bending in that direction, right? And we think that is like, well, so it's the form and function question.

So in the, you know, people would often question what can you learn from a dead brain about how a living brain functions? And it goes back to this old form and function idea, right? Architects, uh, uh, like Louis Sullivan argued that, uh, the form of a building, uh, somehow should be intimately related to the, its human function, right? And in the rest of biology, everybody accepts the idea that form and function should be related, but the mind is the last frontier, right? How could, how could something as ethereal and ephemeral and magical as the mind be related to some material structure. Right. And, but you see inside the fly brain very beautiful ways in which form and function are related. A layperson can just see them in a very visceral way. And so the example you're talking about, um, are these neurons inside the fly brain. Um, it's actually a family of neurons, family of neuron types. And, uh, unlike, uh, a tree with a symmetrical crown, you know, like a pine tree that's exactly symmetrical. They look like the windswept trees you might see near a beach where the branches go in one direction. Yeah. Uh, and in this case, there's 4 types and the branches go in left direction or the right direction or the up direction or the down direction. And interestingly, um, these neurons are already known to be detecting the, the direction of motion of a, of an object, right? So when something moves left or moves right, these neurons respond and the direction in which their branches point is basically the direction in which this, uh, visual stimulus moves. So it's like it, it went that way. This, they're saying it went that way. So, you know, my joke is that these neurons are pantomimers. They're trying to tell us what their function is. They're wildly gesticulating. They've been, they've been doing that for, you know, 200 million years and we couldn't see it until now.

Yeah. So, I mean, this happens over millions and millions of years. I mean, is it meant to like, is it, um, but it's not like actually putting, you're not sensing the object sort of faster, are you? Just because it's kind of actually closer to the—

No, it turns out that this, this branching, uh, this, uh, asymmetrical branching is related to connectivity.

Okay.

That the tips, the tips of the branches and the base, they are connected to different types of inputs. One is slow and one is fast. And that is what gives rise to direction selectivity.

Okay.

So it's really about— it's— there's the— it turns— the shape of the neurons is closely related to what they're connected to.

Okay. Was there any—

I mean, I'm sure there was a lot again, but another great example is the inside if you look inside the center of the fly brain at one particular kind of neuron, there's 50 of them, the EPG neurons, they literally form a donut. Yeah, it's basically, there's a donut in the middle of a fly brain. It's shocking when you look at it. And that donut is actually, when that fly is alive, activity in those donut neurons is encoding the direction in which the fly is heading.

It was like a compass almost.

Yeah. It's a, they're called compass neurons.

Yeah. Yeah.

So, you know, like when you use Google Maps that you see the arrow changing as you move around. Uh, and so that's, that's what exists inside as literally as a donut inside the fly brain.

Yeah. I mean, you were showing how they light up if you're going right or switching direction.

Yeah.

Yeah. Amazing. Were there, um, I'm sure there was a lot, but, um, were there things that just shocked you once you finally got this full map and started the things that were counterintuitive or that surprised you?

Well, the shocking thing, I guess, as a— it's kind of a meta, meta question, right? Which is, I was shocked by how comprehensible it is. So I think the people 20 years ago who argued that the connectome was not going to tell us very much, it's because they, they weren't denying that it was important, but they were afraid that it would be inscrutable. You might have it, but you wouldn't make it, be able to make much sense out of it. And, and, you know, a great example actually is the genome. If I give you the sequence of a gene and no one's ever studied anything like that gene before, nobody knows how to read that.

Yeah.

Nobody knows how to say what effect that has on, on, you know, on your body and, and, and mind and so on. Nobody knows. And the, The approach is profoundly atheoretical. You just break it. Yeah. You break it in a mouse and you see what happens to the mouse, right? Of course, later on, if you find a similar gene in a, in a rat and, and you've already figured out the function in the mouse, well then you can, you can say, oh, it's probably like that, but nobody knows how to read it. And what's very interesting about the fly connectome is that some of it already is readable. It's decipherable. And I like to quote, again, I'm originally a physicist, so there's a famous quote of Einstein, "Subtle is the Lord, but malicious he is not." Okay, and so what Einstein meant by that is that, you know, the universe is hard to comprehend, but it's not like God was evil and deliberately tried to make it hard to comprehend. But when you're working on the connectome and you're struggling to trace every path of every wire in the brain, you're thinking, man, God is really mean. But then when you start to look at the result in the fly brain, then you think, wow, God is, God is not so mean after all. Some of this is comprehensible.

You find some order. Yeah. Yeah. Okay. Well, there, you know, we could talk about the fly connectome, I think, for hours on end. And that I wouldn't mind doing that, but I want, I want to get to kind of where, you know, what you're building off from, from this work. And, and so, so all of this comes out in 2024. And, um, so you, you kind of had this last year to digest all of this and plot what's next. And so now you're going to do a startup that's, that's based, inspired by this work and sort of, and pushing it to the next level. Um, it's called Mazing. Um, you, I think you have a team of about 5 people at the moment, right? Because it's early days. You're like just, just getting going on this and we'll have a story about all this as well. Um, for people to read, but yeah, why don't you give us the, the, the, this is a big idea. Why don't you give us the basics of what you guys are trying to do?

Well, in my book Connectome, so the book came out in 2012. I tried to give a vision for the future of neuroscience and not just about the next year or two, but really about a kind of longer-term vision of what would would happen. And so I predicted that the connectome would be, um, become a very important part of neuroscience. And I also predicted what we would do with connectomes once we had them. And I call them the, the 4 Cs. So carving, um, codebreaking, comparing, and changing. Right? So carving means, um, if you chop up the connectome properly, you get the parts list of the brain, the cell types. And deciphering, right? Trying to understand what's written in the logic of the connections. And comparing is, of course, you know, what's different between a male and a female brain? What's different between your brain and my brain? What's different between a fly brain and a mosquito brain? So all kinds of comparisons we can do that will reveal something about how brains work and how they evolved. And then of course changing, right? Because every, in the end, every personal change is about changing your brain and likely about changing your connectome. And that kind of stuff is happening. Neuroscientists are learning how to change the brains of, you know, change the wiring of brains of flies and worms and so on. So this kind of thing is really important. But the most far-reaching impact of the connectome I covered in the last chapter of my book. Which is the idea of brain emulation, right? So there's an old dream, which is that, well, I should say there's a, there's a very old scientific discipline, which is about modeling the brain, modeling the brain, simulating the brain as a neural network on a digital computer. It goes back at least to the '50s. Yeah. And of course it's become, it was the basis of the modern artificial intelligence, right? But for the first time, now that we have a, a detailed map of neurons and synapses, we can make what people call an emulation of a brain, you know, unprecedented realism. The simulated neurons are in one-to-one correspondence with the neurons in a real brain. The simulated synapses are in one-to-one correspondence with the synapses in a real brain. So that's a brain emulation. And already in the special issue of Nature there and in another issue of Nature that was very similar in time, there were papers on simulating a fly brain using this approach, using the connectome to build the simulation. Okay.

Since day one, the Core Memory Podcast has been supported by the fine people at E1 Ventures. They are a young and ambitious VC firm in Silicon Valley. Investing in young and ambitious companies and people. Thank you so much to E1 Ventures for all your support.

Okay.

There's so much to get into here, and I feel like it's a big concept to wrap your head around at first. So you're talking about, we have this map, we know that this is what one fly's brain looks like. And so we're gonna translate this into software and not only are we gonna have the map in software, but now we're gonna try to unpack the mechanics. You've got the dynamics, the dynamics, just like you've got in a neural net. You've got these weights that are, are on a spectrum of, of, uh, connectedness between each other, between, between all these, these nodes in a neural network. That's what our neurons are doing. You know, it's not just like on or off. There's degrees of, of, of firing. There's degrees of strength between these connections. And so, so you guys start with this map, you put this in the software, And then you, I presume this is like now a process of experimentation to try to measure all of these connections and, and kind of codify that in software. Is that—

That's right. So, so in your neural net and AI, the connections are adjusted by a learning algorithm, right? We collect a lot of examples and then we have a learning algorithm which adjusts the strengths of the connections in the artificial neural net. But with the connectome, we have measurements of the connections in a real brain. So now let's take those measured connections and then stick them into the simulation. And if we run the simulation, then hopefully intelligence will emerge if we did the simulation right. Now you might ask, how do you measure the connection strengths?

I was going to ask that.

Right?

Right.

So, well, we can start with very naive means. In the case of the fly, we can count synapses. If we have neuron A and neuron B, there could be 1 synapse from A to B, 10 synapses, 100 synapses. There's a lot of dynamic range in the number of synapses in a fly brain. So that's our first estimate for the strength of a connection.

Okay. And okay. And then I, I want to get to more on how you guys are doing this, but let's do the example because I think it's tangible for people, um, of some of the work that did come out of the FlyConnectome with people trying to simulate small, small parts of this. So was it the, was it again back on the visual cortex where?

Well, no. So people have done not, not just small parts, but large, either whole brain simulations or large parts of the visual system. So one of the papers in this special issue of Nature gave a simulated taste to the fly brain by stimulating the taste receptor neurons. And then, um, got a simulated proboscis extension, right? Uh, because the motor neurons got activated for the proboscis.

So like in software, you're simulating, you're stimulating the taste neurons just by—

you just activate, you know, you know which neurons the taste neurons are and then you just activate them.

You're just sending some code. Yeah. These are activated and then you want to see how that message filters out and it should filter out. And if you've done this right, you should see this, this resulting action.

Yeah. So, so. The person who made the simulation, Phil Xu, I guess he didn't know in advance if it would work, but he ran the simulation and then it turned out that something plausible happened. Uh, certain neurons got activated and then, uh, his experimentalist collaborators, um, tested the real fly to see if those neurons would get activated. And you know, a lot of it was similar.

And so if you're saying quite a bit of work has been done here already and even whole brain, what are you guys doing that takes this to the next, next level?

Well, this is just the beginning, of course, right? So, um, there's a fly has many different behaviors. There's, you know, many different, a fly can see, it can hear, it can taste, it can smell. There's so many different sensory modalities. Uh, it can, you know, it can walk, it can fly, it can court, it can fight, uh, all kinds of, uh, behavior. So, um, verifying that these things work,, is actually not a one-time occurrence. It takes some time. And then the second problem is that of the information gap. So the connectome is an incredible amount of information, a great, uh, it's a game-changing constraint on such a brain emulation, but there's still an information gap, right? Not everything is measured. And so at this point in time, um, you know, the theorists, the computational people have to make plausible assumptions. And see if those assumptions give rise to behavior that matches experiments, right?

Yeah.

In order to cross that information gap, we have to make some assumptions.

And then, you know, I don't know how much your techniques here would be secret or not, but what is, what is the, what are the methods that you then use to unpack this?

Well, that might be a little technical, but, you know, I'll just give you an example. So of the kind of problem. Um, we said that we can estimate the connection strength by counting the number of synapses, but you might also know that some synapses are excitatory and others are inhibitory. Some make the neuron, the target neuron fire or become active and some try to turn it off, right? So is one excitatory synapse equal and opposite to one inhibitory synapse or are they more like apples and oranges? There's a missing parameter there, things like that.

Okay.

Uh, time scales. So some synapses might stay on for a long time. Some might stay on for a short time. We can't read those yet out of the connectome. So we have to make some educated guesses.

And this is what, this is just computationally intensive work and running just loads of experiments to, to push these connections, push these stimulus through and Well, there's various approaches.

So Srini Tiraga, his lab is applying backprop. Okay. It's a hybrid approach. So let's take backprop, which is the standard approach in AI, and then let's train this flybrain emulation to produce the proper, you know, motion detection of perception or whatever behavior you want. Let's train it to do properly and then tune these parameters, like the relative strength of an excitatory inhibitory synapse. We could tune that by backprop. So that's a hybrid of the bottom-up and top-down approaches.

Okay.

Yeah.

So the, you know, the second you and I started talking about this, I mean, my head went to, okay, like on one hand we can just learn an incredible amount about the mechanics of a brain that seems useful. You could start to learn what makes the human brain so efficient compared to an AI system. So that could, that could potentially inform AI work. It could, it could make our AI systems better, you could start to create such a realistic simulation of the human brain that we start getting into very interesting territory, like uploading minds and more philosophical sci-fi kind of questions. What is like the main— if you're doing a startup around this and you're shifting from academia, what is your main business quest on this, or do you even know yet?

Well, the ultimate goal, of course, is to enable and empower humanity to transcend biology, right? So human brain emulation is the ultimate goal of the quest. I would say that on the way, we want to slash the cost of AI, right? Compare the human brain, 10 watts. NVIDIA DGX-H100, 10 kilowatts. Right? So huge difference in power consumption. And a fly, a fruit fly brain, probably 1 microwatt. Like an amazing amount of computation for just a tiny, tiny amount of power. And, you know, slashing the cost of AI has a lot of important implications, right? It's the difference between an AI that's really centralized and requires a huge amount of investment and is controlled by a few corporations as opposed to AI everywhere, you know, intelligence too cheap to meter, right? Yeah. So, um, I think slashing the cost of AI is, uh, is one of the big motivations for, for us right now.

But you, in, when you say, you say it so clearly, like, you know, the end goal is to do a whole human brain emulation and, and that's what we want. I mean, for people not in this field, like, you know, why would that be so valuable? What would you do just with that, even outside of the—

Oh, with the human brain emulation?

Yeah. Yeah. Like, I mean, you, you know, clearly this is what you want to do and, I don't think most people don't get to think about this a lot. And so, so I don't think it's as obvious what you would do with that.

Um, well, many people think that, uh, human-level AI is economically valuable. So if you wanted to take a business slant to it, then certainly, um, if, if, if you think about the different approaches to human-level AI, AGI, right? So it seems that OpenAI and all these other companies with LLMs are on a really great track. Uh, but you could think about human brain emulation as a dark horse to AGI, right? If we could emulate a human brain, then, then by— it's a corollary that we're getting human-level intelligence, right? So that's this machine, right?

Okay. And then again, millions of questions on this. I mean, you're so uniquely positioned because you, we mentioned you were a physicist, but you did early pioneering work in the AI field, and then it's kind of incredible. And then you've brought all this neuroscience knowledge to bear on this. I mean, I see overlap a lot of times in these AI labs. There's neuroscientists all over the place, but man, I mean, you've, you've been at this for a while and have seen both these fields.

I've lived through two scientific revolutions. Yeah. Yeah.

Go, go in parallel. And then, you know, I'm old enough to, or I've covered this long enough to have remembered. I, when I used to talk to a lot of neuroscientists in the, like, you know, I'd call it the 2010s, 20— 2012, um, very kind of skeptical about AI and, and like, oh, people say it's a neural net, but it was like, we don't even understand how a human brain works. This thing doesn't emulate a brain because we, you know, we don't even know what's going on. I feel like over time that line, that has changed quite a bit. It does seem that, that with these neural nets, when you reach some level of complexity, very interesting Things start to happen and definitely we, you know, we see, um, it feels like more and more, um, parallels between, between these neural nets and what we know about the brain. And then we're at this, so this is a fascinating moment. I feel like I'm never very eloquent when I say this out loud. It's like almost like an intuitive feeling, but we're spending trillions of dollars building artificial brains. We're spending less trying to understand the human brain. And to me, in some sense, this work you're doing would be to level the playing field to some degree. I mean, it is, it's, it's kind of a race. Like, what do we learn first? Do we create an artificial intelligence that's, that's equal to or better than the human brain? Or, you know, do we understand how our own brain works at such a fine level and like harness that? And kind of which one happens first?

Well, you could think about it as a race, but I would actually say that they are simply two complementary quests and feeding off each other for humanity, right? You know, people liken the invention of AI as, um, you know, it's like the invention of fire, right? It's a— it's clearly a technological revolution for humanity. And if we step back and think about human history on the thousand-year and 100,000-year time scale, this is a significant turning point for humanity. But I would argue that it's the same for brain emulation, right? So the idea that our minds don't necessarily have to be tied to biology, but could somehow, we could somehow transcend, you know, into some other kind of substrate, could be silicon or whatever, that is a kind of science fiction idea. Um, but it's also, I think, part of, part, it's gonna be part of human history and it's complimentary, right? So the, the developing AI technology and developing brain emulation technology are two sides of the same coin.

Yeah. And okay. And how much do you, I mean, your dad was a quite famous philosopher. I think your mom was a very accomplished musician, right?

That's right.

So kind of coming from these, um, humanities side, um, you, that's right.

My father said, no son of mine will do humanities. And my mother said, no son of mine will do the arts. I had to do science.

You know, but when you start talking about possibly downloading a mind or where all this could go, I mean, how is that, is that something that actually calls to you or is that something you just think is interesting and obviously other people have, are quite pursuing this, you know, in my, in my book, I joke that you shouldn't say downloading.

Right. Because uploading is like going to heaven.

Yeah.

And downloading is like going in the wrong direction.

You wanna upload, but you know, like how, how into this are you? Do, is that really, is that part of what drives you or not so much?

Well, yeah. Okay. So I, I should say that you're trying to be immortal.

That's what I'm asking. Well, okay.

So So I want to talk about AI first.

Yeah.

And I think that our current discourse on AI, people are kind of screaming at each other and shouting at each other without really understanding the philosophical and theological currents of what they're saying. And that's what it fundamentally is. So, you know, people say that, you know, my joke is that every time you see a newspaper headline with AI in it, Just replace it with the Messiah. Dario says that AGI is coming. Dario says that the Messiah is coming.

Yeah, that feels right.

AGI will be here in 5 years. AGI is already here. The Messiah is already here, right? So this idea, um, and I first realized this, that I guess I went to this weird conference organized by Yann LeCun a long time ago about AI. I guess essentially he did it because of AI doomers. And I realized that my colleagues, their attitude was we are about to create God. Omniscient, all-powerful. And what will God do to us? Will God be a wicked Old Testament kind of figure and just smash us? Or will God be nice? And, you know, so it's this feeling and people project, right? So I think the people who people who have a sort of dark view, you know, they just think God's gonna be really mean. And people who have an optimistic view of life think that God's gonna be really nice. So it's hard to have a rational discussion about AI because really underneath it's about talking about your, your, your vision of humanity and of God.

And so you, so then you, you feel like you do have a uniquely rationalist view on this though? Or like, are you, are you one side or the other?

Then? Well, so I think it's great, you know, to, again, taking this idea further, you know, religions say that God created man, humans, right? And Feuerbach is the 19th century philosopher, and what he said was that man created God. God's a construction. Right? And what man did was he projected reason and love and will, these human qualities, he projected them onto this construct that is called God, right? And that's kind of what AI is today. We train, we take all this output of humans and we project it into this AI model, right? And then we regard it as God. And then, but Feuerbach said another thing, which is that this, the problem with religion is that it causes alienation because when you think about, oh, wisdom and love and, you know, power, these are God's qualities. You also have a tendency to debase humans. You say, you know, we are so, we're so bad. We lack these qualities, right? And so he said it led to a kind of alienation that worshiping God often meant debasing humanity. And that's also happening with AI, right? Yeah.

Right.

Oh, AI, you know, AI is going to kill, you know, it's going to take all our jobs. You know, we're so stupid now. It's this feeling that if AI is so great, humans must be losers. And so it's kind of recapitulating Feuerbach's idea about how religion can debase, in the worst case, can debase humans. So we have to think in the, he said we have to think in another way. We have to really think these are human qualities. And we should be thinking about how we can promote these great qualities in humanity, right? That's the right way to think about, I think, AI. How can we use AI to raise our consciousness, raise our human qualities to be the best that we can be? Okay.

But are you— so you're saying that you've kind of put some distance between these more rhetorical things and you see AI as a tool. You're not caught up in this.

Well, I don't want to elevate AI to this godlike thing that's going to make us all lose losers, you know?

Yeah.

Rather, we should think about how AI is going to enhance, uh, make us into better people. And one of the problems with technology now is it seems to bring out our worst. We have to figure out ways to help technology bring out our best.

Yeah.

So, okay.

And then, you know, I, I talk to lots of people in the, this overlap between AI and neuroscience, and there's several groups right now. I have to confess, I get the gist of what they're trying to do, but not the actual mechanics. And, and I can't quite tell this is like real or just hopeful, you know, but you have people who are like, okay, well, we're gonna, we're going to put some probes in brains. We're going to put so many probes in these brains to understand its mechanics, whether usually like a primate, say, and, and then we're going to somehow translate what we've learned from that brain into an AI system. And then, so hopefully it's like more aligned. With our actual brain and that will make it more aligned with our wants and needs and like humanness. So again, I get the gist. This seems like scientifically vague to me, or I don't know, you got to pursue these things, I suppose. But I'm hearing more and more about different groups trying to do this type of work. Does that make any sense to you that we could somehow infuse some portion of humanness from this, either a connectome or this BCI, brain-computer interface data we're getting into a neural net and that somehow helps alignment.

Well, I guess I think align— so yeah, so the idea of aligning AI with human needs is one way to think, but I want to get back then to, I gave this whole long digression about AI and the underlying religious and theological, you know, parts of the discourse. So what about mind uploading? And I would say that mind uploading is clearly the 21st century notion of religious quest, right? The idea of transcendence, the idea of, if you think about just the old-fashioned idea of heaven, heaven is all about leaving your body behind. Right? Or if you're a Zen Buddhist monk, it's, you know, you're trying to get control or, you know, conquer your own consciousness and achieve some kind of divinity by transcending the constraints of your normal mind. And so, mind uploading is really a religious quest in the end, I would say. And it is the quest to become like God. It's the quest to join God. Right? So you could think about it as not only being able to align AI, but be joined into some kind of, or, you know, progress to the next level of consciousness. Right? So once we are uploaded, then we will be free to change ourselves more for the AI and so on and so forth. So that's the way I think about it. You know, it's not that we are creating God, but we're going to become part of God.

So you do, you're into this, you're into this idea.

Well, I think I had no choice.

Yeah.

I think I had no choice, right? Because, you know, I was first exposed to these ideas. You know, I'm a scientist. I just want to know. And I felt that the connectome was the way to make progress in neuroscience, to do my part in transcending human ignorance and, you know, for us to know more about the universe. That was my original motivation. But then I became exposed to the back in the, in the, you know, in the 2000s, the people who, um, who had these ideas of, of transhumanism, right? And, and then I realized that, uh, to them, I am an instrument of their destiny. You know, I, I'm, I'm involved in this no matter whether I like it or not.

I was gonna say, because I, I hang out with, with some of these people.

Yeah.

And, you know, they are, are very passionate and, yeah, like religious about this. You don't strike me as that. Like, you've always struck me very grounded and you're pursuing your work. And so you do, you see yourself as a vessel.

Well, but I think many scientists are driven by some kind of religious— all right, so we do science because we want to be in the presence of awe-inspiring mystery, right? And we want to be there. We want to be We wanna be one with that. That's the big motivation, at least for, I don't know, for me, I don't know if ever, yeah, some people wanna, they do it for other reasons, but at least for me, there is this mystical side that drives a, you know, you can't be mystical in your daily work 'cause you actually have to discover something, but there's a certain drive or motivation behind it, right? So, but one of the things, if you think about the history of ideas, the scientific worldview has been regarded as antithetical to the religious worldview. So there's a famous quote from Steven Weinberg, "The more we understand about the universe, the more it seems pointless." Right? And there's this, Sigmund Freud wrote about 3 blows to human narcissism. The first was Copernicus. Right? That the Earth is not the center of the universe. And then Darwin, humans are not special. We're just another, like all the other animals, we evolved, right? And then, well, Freud, you know, ironically, he said he was the third blow to human narcissism. He was kind of a narcissist himself, but he said, you're not even in control of your own mind.

Yeah.

Your unconscious is, you know, is, is you, you're not con— you're not in control. So he said that's the third blow. Of course, AI is the third blow. AI is the blow which says that the human mind will potentially, you know, eventually be surpassed by, um, by computers, right? So this idea then that religion has been killed off by science. Is what I got, you know, what I came of age with. But actually, I think that science has been rediscovering God. You know about the Anthropic Principle?

Well, tell me.

Well, that's the idea, the physicists say that if you were to just change some of the numbers in the laws of physics out of, you know, by a very small amount, then there would be no inhabitable worlds in the universe. Yeah, it could never happen. So these people argued that it's almost as if the universe was designed so that consciousness could emerge, right? And so, you know, you could say, uh, you could say there's a new, uh, sort of, uh, theistic worldview, which would be that, you know, based on physics and neuroscience and AI, which is that the universe does have a purpose and the purpose is to create God. To create consciousness and to improve consciousness.

I mean, there's parts of this that I've always liked, which is what is the point of all this? And obviously, you know, we've all wrestled with these existential questions and, you know, part of the answer, if we were able to blast some form of intelligence, I'm not saying it's a full brain or just something that humans, some part of human intelligence out into the universe and this becomes eternal and it's it's off traveling through the galaxies. I mean, that makes sense to me. It's like, okay, you know, our purpose was to kind of like get beyond our boundaries and spread, spread our intelligence throughout this thing. And, and, you know, you obviously can see very religious sort of bits and pieces of that. The, when we start talking about, I'm gonna download my brain into a, upload my brain into a hard drive. Um, have you seen the show Pantheon at all?

People have told me. Yeah. Yeah.

I mean, you know, the, You don't just have to go to that. You could go to any number of sci-fi books. I mean, this goes wrong quite quickly. There's parts of it where it's like, would be wonderful. And then parts that's awful to think of yourself like in some kind of vat and just experiencing your family and the world through that, even though you could maybe experience them forever. Yeah. I mean, so do you, and you know, you're talking about, oh, well, part of this is my, it appears to be my fate because I've done all this research and now, you know, people are going to use this one way or another. I mean, so you're, you're what, you're optimistic in, in the end, or when you read these, these nightmare scenarios about where all this goes, does that worry you?

Well, I guess I'm fundamentally an optimist. So, you know, when we project, we tend to project our just, yeah, if you believe that human nature is fundamentally good, then you tend to be an optimist in life and you tend to think that when we're gonna create God, eventually we'll be creating a good God, right? And if you, if you, if you think that human nature is dark and wicked, then you know, you're gonna think that, that the AI we create will be wicked, right? So I think that's the projection that everybody makes.

I mean, it's interesting when you talk about humans creating God and now we're creating these AI systems. We didn't create like a very, I mean, the God is kind in some ways and then horrible, you know, in, in many, many anyways, that's because God is trained on 4chan as well as, well, you know, all the, all the cesspools of the internet were probably included in the training set. Yeah. I mean, I, well, I just see these athletes sometimes they win a championship and then, you know, they're thanking God for letting them do that. I'm like, what about all this suffering? You know, what a weird, weird choice for this God to make at this particular moment. Um, so, uh, yeah, I don't know. And then, and then I wonder if we're, well-equipped to make the— you see a lot of the people in the AI field, they actually think they have quite good intentions and they want to align this thing and they want to have this be a good outcome, but maybe human nature, even when we're creating our idealized thing, we're filling it with badness anyway just by—

Well, that's the question. I mean, do you think humanity is going to figure it out that we're fundamentally— you know, it's the human adventure. It's the human adventure. Um, I, I always have the faith that we're going to figure it out and it's going to be an exciting and wonderful and interesting, um, world. Yeah.

There is, so the fly that, that is in all these papers, I mean, that is maybe our first immortal, uh, being. Yeah.

One, one fly died for this connectome, but this is the first fly that will live forever.

And like when you, I mean, I just to keep going down the sci-fi. Bit for a minute, or, you know, I do, I do so much reporting on brain-computer interfaces and reading data out of brains. I mean, it feels like a big chunk of where we might be heading next would on some level be putting information into brains, possibly. It seems like, where are we on that? I mean, not, not, we can stimulate a brain, but like to be—

Yeah, the biggest, I think the most important brain-computer interface actually is for people who are blind. Right? People, that's clearly a very difficult sense to lose. And so if we can build prosthetics that either stimulate, you know, living neurons in the retina or somehow directly put visual information to the brain, that seems like it would be useful for a lot of people.

Do you have a, and tell me if I'm getting too weird, but like to go beyond sort of stimulating and activating a neuron, which I realize is at some level I'm going to talk about the same thing, but I mean like the thought of putting in for being in The Matrix and you, you're downloading kung fu or Spanish, you know, I mean, um, do you have any sense as to how realistic that could be?

Not sure, because you would still, you know, if, if you it seems difficult because unless you replace your whole brain, right? Because you're still the part of your brain that has to learn at a slow rate. Uh, so for example, as a normal person, I'm not blind and so on. I don't feel the need to have a brain-computer interface because I'm already, I already have information overload. You know, I can't, I just, there's too much information in the world. I can't absorb it all. Um, So I don't know whether if I had more information, I could learn faster. Really, it's hard to learn faster.

Yeah.

So I think that this, uh, it seems, it seems hard. Maybe I'm just lacking in imagination, but so it would require a total brain replacement or, you know, if you're uploaded, then obviously the code could be modified. You could be changed to learn more rapidly and so on and so forth. So I think a software version of you would be more amenable to that kind of modification.

Okay, let's fast forward out. I have no idea what number of years to pick. I mean, this could be 5, 10, 15. At this moment in time, the AI systems remain something of a black box. We have this field called mechanistic interpretability to try to understand what is actually happening inside of these large language models, these AI systems. We're getting some insights out. There are remarkable parallels with the human brain, sometimes like groups of neurons that associate with particular ideas and concepts, um, which I find fascinating. And then at the same time, we have people like you pushing forward on, on understanding, um, biological brains in the best way possible. The biological brains are kind of fixed in some sense. I mean, they are what they are. The AI systems are getting bigger and more complex all the time. If we go forward like 10, 15 years, I mean, do you have a sense as to we have a much better understanding of how the human brain works at that point and understand it on this very intimate level and, and AI systems have become just more of a mystery, or are we learning as much about both of these over time, um, so that we actually know how both of these systems work well.

Well, what's happened over the last, uh, 10 years, 20 years really, is that, is that AI accelerated neuroscience, right? AI has given us the ability to map connectomes. So I would say that neuroscience is now entering, uh, an era where we're going to see profound acceleration, and that acceleration is already evident in the study of the fly brain. That, that field is moving at warp speed compared to the rest of neuroscience. And so it's not just the connectome, it's, it's, um, it's also the control over all the molecules in the brain. And I should say that's another kind of brain mapping that's happening. So we, we can have the connectome, but we also need to know more about the, the nanomachines that exist inside inside cells and especially inside neurons. There's ion channels, there's synaptic receptors, uh, and the ability to map the locations of those molecules, identify the type of molecule exactly where they are in the neuron. That is a technology that's developing very rapidly, and that's going to fill the information gap that I talked about, uh, earlier, making bottom— removing the need for our clever assumptions that are currently needed to cross the information gap.

And when you look at this like 40-year period that went from the worm connectome to the fly, and then some of the things that you're just talking about, I mean, it must be just like incredibly exciting moment in time to be in these fields. Like I get blown away and I'm just like on the periphery, you know?

Right. Right. But you know, it's in general, uh, so we have AI, you might say, hey, AI is going to race ahead. But what's happening is that within neuroscience and within biology too, right? Like AI protein design and so on, the parts of biology and neuroscience that are next to adjacent to AI, those are the ones that are going to race forward the fastest. And hopefully they carry the rest of the field with.

And so you like, yeah. And you obviously believe in this, you know, I mean, there's people who are like, oh, the AI systems are kind of not, uh, I don't know, people are very impatient or something like, you know, they're not living up to sort of the billing. They're, they're already plateauing. But when I see the stuff taking place in biology, either the startups are just hyping themselves into, you know, some, some state, but it's like what I, what I see happening with the developing antibodies, learning more about the genome and how, how things is like breathtaking to me. And so, yeah, so you are, you feel like we, AI really is pushing science forward.

Well, you know, I'm a realist. I think that the timelines that people give or the expectations that people have are usually unrealistic, right? Right. And the reason, of course, is that it's, you know, when you, when you have a company and you say, oh, you know, we're gonna, we're gonna have an AI transformation. What happens is what can be accelerated by AI gets accelerated, but the rest of it, you know, holds back.

Yeah.

It, it makes the transformation less than what you expected. And it's the same thing in, in biology or neuroscience. The parts that, that are AI adjacent will, will progress exponentially, but they're dragged down by the hard stuff. Right. And so, Uh, it's, it's a paradox, right? On the one hand, we see this fantastic acceleration, but on the other hand, we, we don't get the results as quickly as we expect. Yeah. Yeah.

Yep.

Yeah.

You were, you were what, you were born in New York?

New York City.

That's right. And your parents are Korean?

Right. Right. They came from Korea.

Like, but North Korea, I think, right? Or, or—

That's right. They were, they were, they were refugees. Yeah. Yeah. That's right.

Who came to the States and then, and then your father ends up in this academic career when he was a philo— like what was he into in philosophy?

Oh, all kinds of stuff. I think he did his doctoral dissertation. Well, okay, so, so he did philosophy, so Hegel, Kant, uh, he was a Platonist. So, um, that's, that's an important distinction. He's a Platonist, but, but he, he also did literary theory, so he was an expert in Dante. Brings back the Heaven and Hell analogy here. And in the later part of his life, he was really into Goethe and Nietzsche and so on and so forth. So he did a lot of different things. And he never really wrote much about philosophy of mind though. I mean, so my joke is that he said, no son of mine will do philosophy. But I, and I, so I did science, But I ended up in the part of science that's closest to philosophy.

So this is kind of where I was going a little bit. Yeah. I mean, the, well, anyways, a man after my heart. I was a philosophy major and I love, love some Goethe. Um, the, and he taught at University of Texas, but was that late? Did you grow up in Texas?

I grew up in Austin, Texas. You did? Yeah.

I didn't know that.

All right, man.

I grew up in Houston.

All right.

Yeah. I never, I did not know this.

I feel like I should start wearing a cowboy hat for my, you know, I should pay homage to my origins.

And where'd you go to high school?

Um, I went to, uh, St. Stephen's.

You did?

School.

I went to St. John's.

Oh, really?

Okay.

Yeah. Yeah. Okay. Yeah. Soccer, uh, we're rivals. Soccer rivals.

Yeah. No, I used, I played tennis. We used to, we used to take you guys on. Um, okay. Fascinating. And then you, you went to, to, did you go straight to MIT for undergrad or, um, I was, uh, undergrad at Harvard. At Harvard. At like 16, I think.

16.

So you skipped some grades.

You know, in those days there were a lot of 16-year-olds at Harvard, and in fact there was a 15-year-old in my class, uh, and you know, he was taking 3 graduate-level math classes. Um, so you know, the Harvard is the place to go to not feel special. A lot of geniuses.

Yeah. Yeah. And then you, and you got your PhD at a very young age.

Um, yeah.

Uh, not that young.

I mean, you know, I, I guess right around when I turned 24. I mean, young means getting your PhD when you're a teenager. That's a real—

I don't know, that's real.

It's a real prodigy.

24 feels pretty good to me. So like, when you're going through all that and you're a physicist and your dad has this background, I mean, this is where I was heading a little bit. Um, when you were a kid, I mean, pick your age, um, somewhere in your teenage years, what you— what did you picture you were going to be?

Oh, well, I think my father was frustrated by philosophy, so he encouraged science. And, you know, it was a time when you watch science, you watch TV shows about science, and you read books, and it seemed very glamorous to me to be on the frontiers of knowledge. So I was really interested in a career in science, but in physics. I guess particle physics was a big deal when I was growing up.

Right. And I think that's where you did your dissertation.

No, I didn't. So by that time, string theory became the dominant kind of theoretical particle physics, and I clearly wasn't smart enough to do string theory. So I switched to condensed matter, theoretical condensed matter physics.

The easy, easy, easy path.

And then what switches you into neuroscience so Well, so, okay, so remember that the theory of elementary particles is all about breaking things down, finding the ultimate constituents. But then people, you know, I guess the other areas of physics are about how do you build things back up again? So the classic example would be diamond and graphite, right? They're both made of carbon atoms. But they have very different properties. One is dark and soft, the other is hard and brilliant. And you don't give your wife a graphite engagement ring and say, oh, it's the same atoms, right? It's the arrangement of the atoms clearly counts for a lot. And so if you're in condensed matter physics, it's all about these emergent properties of matter. But at some point you say, hey, that's, maybe that's too easy. The most important or enticing or mysterious emergent properties of matter are of living organisms. You take the same atoms, carbon, hydrogen, oxygen, whatever, you put them together and you have something that looks really different from a diamond or a piece of graphite, right? And then ultimately the craziest emergent property is intelligence and consciousness. How do you put together these cells that are none of which are intelligent by themselves really, but you get something that's intelligence. So it's all about this putting things back together, emergent properties. So physicists were talking about that question, and I being, you know, naive enough, was attracted by the hardest problem, which was the mind. And there were physicists like John Hopfield, my postdoc supervisor, Chaim Sompolinsky. These were brilliant theoretical physicists who were the Pied Pipers that brought other physicists into the field.

But we, like, this weekend, we, you know, we were at this salon where you were presenting, and then we were having this discussion about, um, obviously tied to the brain and AI, and we got into things like the soul and consciousness. And you were, I mean, you were putting some distance, you know, between wanting to go down these philosophical paths about— obviously it's a, you can tell from the room, it's an impossible thing to define. Nobody can settle on on really what consciousness is or where you would find it. But I did notice, I mean, you, you were putting some arms— you see, you were focused more on the work of the mechanics of these things and what you can understand. And you were kind of leaving some of this to the philosophers.

That's right. Because I have nothing novel to say about consciousness. I don't have anything to add to debate. And so that's why I was leaving it to the experts. Some of the scientists there want to find empirical ways of studying consciousness somehow that can neuroscience, can those tools like the brain-computer interface that you talked about be used to investigate, make consciousness not a philosophical question, but an empirical question. And I'm not sure if that's really ultimately possible, right? So what we know about consciousness is, is all from introspection.

I mean, I was curious. I pictured you like looking at these maps of the connectome for so many years and then late at night it's like, that's it. That's where it is. You know, I think we were talking about Descartes focusing on the pineal gland, you know? Yeah. I didn't know if you've— I was just wondering if you had this moment, you've looked at this brain so much that you've have something's hit you about—

No, my, my approach is let's, let's, uh, try to build the brain emulation. So there, you know, academic researchers talk about the computational theory of mind and they build these models and so on. But hey, let's, instead of that being some kind of philosophical debate, the real debate is let's build the real, the reality of it is let's build a, a brain emulation. Let's see if we can actually get it to work. And would it be conscious? Maybe the emulation would be totally faithful. It would capture, if we made a brain emulation of Ashlee Vance, it would talk like Ashlee Vance. It would, and everything that it would do would be exactly what Ashlee Vance would do, including the behavior of its neurons, right? But would it be conscious? Would it have the conscious feeling of what it's like to be Ashlee Vance? It seems impossible to know without introspection. And so ultimately what I'm, what human brain emulation will be, will, will be this leap of faith. Like, do you think you're emu— let's say emulation becomes possible. Do you choose to do it? You might end up being a zombie, right? But it's a practical choice now. Okay. Choose. Do you want to be— do you want to be uploaded? Are you going to be a zombie? Are you going to be conscious? Uh, well, it's a yes/no question. You either choose— you, you choose it or you don't. Oh, well, it's a practical question, not a philosophical one anymore.

Well, like, what happens if you make this fly emulation and then you see it trying to, you know, constantly escape its, its confines in this, this computer, and it just can't because it's— it is all in software? I mean, Um, all right.

No, so that's, that's, yeah, the thought experiment of brain emulation, right, would be, oh, let's say we have a bug and, uh, when we turn on your emulation, you're in excruciating pain.

Yeah, it's just screaming constantly.

Yeah. Or let's say that you're, you're, you're, you're suicidally depressed all the time, right? We didn't, we didn't make the emulation right. Yeah. So that, that's the, that's the problem. There's a leap of faith, I guess. I guess that's not really a leap of faith. I think that that ultimately we would be able to sort of scientifically verify whether the emulation is faithful. But what we can't verify is whether it's conscious. Yeah.

Yeah.

Okay. My friends want to say you could, you could somehow, but I think that it would be impossible and that would be the leap of faith and you'd have to choose.

Well, there's a, yeah, there was a lot of debate in this room around this topic. I mean, will there not be some people, I mean, look, you gotta have You have to take steps, you gotta make progress bit by bit. Won't there be some people who, um, with your new startup who just unsatisfied that we're still working on the fly and it's like, how much? Okay, you're gonna emulate this. I'm sure we'll learn some things, but obviously we want to get to the human. That's where all the good stuff is. That's where we really get to this point where whether it's a business case of, of trying to boost AI systems or some of these, these deeper questions, that's where we want to get. And, and like, oh God, that's going to take at least another decade. I know we're— humans are impatient. Decades, maybe not that long. But, but, um, yeah, I mean, is it— do you, do you feel like we'll learn enough, um, from the fly, or this is— you're just— you, you know, we need to do this work and you think the human will come quicker maybe than some people think?

Well, it's a roadmap, right? It's a roadmap. And, you know, if you think of the history of computing, it didn't start out with AI. It started out with, uh, you know, very simple computational devices. But it, once you've got it, once you've got a toehold, then ideally it just becomes this inexorable pace, right? And so you, you're trying to, you're, you're, you're, you're, you have to have the first fire, you know, once you, you can create a bonfire later on, but you have to have the first spark. And so the argument is we're at the moment of that first spark, right? Uh, And if you are impatient, and a lot of impatient people in the end, they're just wishful thinking. It's just wishful thinking, right? And the difference between also in science also, you can be a dreamer, you can be a visionary, but you know, in the end, if you don't make any progress, who cares, right?

Yeah.

So, okay. And then I would say, is there, is it commercially viable? Well, I think that the feats of animals, are incredible, right? We underappreciate what animals could do. As Yann LeCun often says, you know, we have AI systems that can prove theorems and do all these amazing things, you know, play Go, but where's my AI that can do what a cat can do? Yeah. Right.

Yeah.

But what's, you know, even a mouse or even an insect, they have amazing capabilities. And for embedded systems, so we don't just need AIs to talk to us, right? We want robots for a lot of robotic applications, embedded applications, we don't need language. It's not, that's not the fundamental problem. A lot of intelligence is pre-linguistic. So I think that this, I think that the study and emulation of animal brains is in some sense better to do than, I mean, why would we compete head-to-head with language models right now?

Yeah.

Right? We should be working on the stuff that language models can't do.

And animal brains are perfectly a great example of Cannot tell if this is a stupid question, but yeah, I mean, couldn't you make like the most, if, if you got the mechanics of this fly down, I mean, the most amazing drone you could imaginable.

Sure. You know, uh, if you can, you can try to attack a drone, a, a fly with a fly swatter and see it's, it's got a pretty sophisticated escape capability.

Right.

So, uh, and if you think about insects generally, right, there's a million species of insects. Think about, think about all the nature shows showing you what these insects can do. They have incredible capabilities. Uh, and, and, um, tiny brains, right? Tiny brains.

So there would be part of this where you, if you, if we reverse engineer these mechanics and then you can somehow like infuse it into a robotic system, I mean, you could, you could get these.

Yeah. I think we can make lightweight, um, uh, lightweight neural nets that can run with low power., in minimalist hardware, and that's going to be useful for these embedded systems, right? And ultimately also, uh, you know, people have been talking about, um, AI chips that are really based on, on brain structure, neuromorphic systems. And one of the problems with the whole neuromorphic, uh, field, um, has been that we didn't really know what neuromorphic was. We didn't have the structure of the brain. But now we do, right? And so this is going to suggest new ways to think about how to build these incredibly efficient AI chips that are efficient using, using some of the brain's tricks.

Yeah. Yeah. No, that, that is what I was right. I mean, cause I hear all the time of, of human brains are so efficient. We're going to, we're going to mimic this. It feels like we have so far to go. You know, when I see, I go out to Stargate and see the size of this data center, it's like, okay. And it's the same thing you were talking about with Yann LeCun. I mean, there's this part where the AI system already can do things I can't, you know, can read every physics paper on, you know, that's ever been written. It has. And I'm never, ever going to do that. But then I can do all this other stuff so much more efficiently on my piece of toast in the morning than this thing huffing, you know, natural gas can out in Texas. But yeah, it's hard for me to imagine actually bridging that gap in some reasonable timeframe.

Another, you know, another thing that's happened is that chip-making technology has become 3D. So a big difference between integrated circuits and the brain was that integrated circuits were fundamentally 2D, very planar. Everything's laid out on the surface of silicon, right? Whereas the amazing thing about the brain is this 3D, you know, your head is jam-packed with wires and you have this 3D interconnect. But if you look at chipmaking technology now, it's becoming more and more 3D, especially memory technology. I think, I think, I think NAND flash, they can have 1,000 layers of memory cells now. So it's really becoming 3D. And so that means there's a potential for making chips that are much more like the structure of the brain with this biological interconnect. Yeah.

Um, all right, I'm gonna let you go soon. Um, did you, did you enjoy working in corporate at Samsung?

Oh yeah, you did. Yeah, yeah.

You're not a, you're not just an academic purist.

Well, you know, I started out my, my career actually, uh, at Bell Laboratories, right? Right.

Yeah, yeah, I saw that.

Yeah. So, um, I started out in an industrial lab, a very legendary—

yeah, that's cheating. That one's like on the You still get to do such cool stuff, you know?

So the, um, and it's kind of, it's kind of, you know, ironic, right? Because Bell Labs was funded, you know, it was, it was, uh, the research arm of the phone company. And so Samsung's also, you know, involved in telecom, right? Obviously with smartphones. Uh, but the amazing thing about research, being, you know, head of corporate research is just seeing, uh, a huge, a very diverse range of technologies, right? Not just my specialty. You know, modern professors are very, they're just forced to enter this narrow kind of pigeonhole, right? But I got to be exposed to the tech world and world of technology and also the world of how these tech ecosystems work. And also just met a lot of nice people and enjoyed a lot of great Korean food. Life in Korea is very, you know, again, you know, I live in my bubble. We live in the US, that's one country, and I live in my bubble, which is Princeton, New Jersey, and being exposed to other countries and how people live, great experience.

So for this coming year, I mean, is, are there any people in particular? We've got a very bright audience. Any people in particular you're looking for on this quest or, you know, anything you want people to know about what Mamazing is up to this coming year?

Oh, um, yeah, we're looking, we're looking to hire. Uh, we, um, are right now we're hiring mainly people in computational neuroscience, uh, and also in robotics. We do need software engineers because I think that we have to build a, a software framework, you know, how deep learning was, was accelerated by PyTorch and TensorFlow. So we need that professional quality software framework for brain emulation. So we're hiring people in these various technical roles. We're also raising funding, right? So we're looking to—

I've got a sound effect for that.

We're looking for investors who want to join, I think, one of the most exciting adventures for humankind.

Yeah, yeah, yeah. It is. It is. I've watched the connectome field long enough to see these, these ups and downs. It is fascinating to me now on the, the, the back of your work and others' work with the fly to see startups now coming out of this is, is that I mean, because it was quite trapped in academia a fair bit. And then I always got to— I mean, I'm not naive. I see the economic drivers behind AI and why people would be pouring so much money into this. I always got a little sad on the Connectome because, you know, some of it was, was like, oh, well, we could do a human if we put enough money into these brain slicers and these imaging, you know, it was like a labor, it's kind of like a labor thing. It was like, well, if we put enough money into this, we'd learn how the most important organ that shapes our entire universe and everything we think and feel. It seems like something we should know. And I, I would, yeah.

That's exactly right. So people ask me, how long till we get a mouse connectome? How long till we get a human connectome? And my answer is always, it depends on how much we care.

Yeah.

Yeah.

Well, so that's the thing. Like, I always understood that AI has trillions because there's this hopeful, very much clearer, I guess, economic outcome. But it is, there's a part where it's just, it's insane. I mean, this is like the thing that shapes our entire everything. We should know everything we can about this and money should be I know, you know, I know some of this gets funded, but anyway, it was, I've always felt a little frustrated.

You know, there's a legend about, um, you know, Marvin Minsky, pioneer of AI, and Doug Engelbart, the guy who invented the GUI and—

Dad down the road from here.

Right. Right. So the legend is that they met and Minsky said, we're going to make computers talk. We're going to make them walk. We're going to make them do math. We're going to make them smart. Smarter than humans, right? And, and, and then Engelbert says, yeah, what are you, what are you gonna do for people?

Yeah.

Yeah. So I think that investing in our own brains is doing something for people. Yeah.

Right.

And, and we shouldn't forget that.

Yeah. Yeah. Okay. Well, that seems like a great note to, to end on. I was, I've been such a fan. I'm, I'm really honored that to get this much time with you and appreciate you trekking out here. And I cannot wait to see what you guys do.

Thanks a lot, Ashlee. It's been fun.

Thank you, man. The Core Memory Podcast is hosted by me, Ashlee Vance. It is produced by David Nicholson and me. Our theme song is by James Mercer and Jon Sortland. And the show is edited by the Jon Sortland. Portland. Thanks as always to Brex, NE1 Ventures for making this possible. Please visit our Substack, YouTube, and podcast channels to get more of what Core Memory makes. Thanks, y'all.