#12: Deep Prasad - How to Build a Physics-Fluent AI System
S1:E12

#12: Deep Prasad - How to Build a Physics-Fluent AI System

Hey, everybody. Welcome back to First Principles. We are here today with Deep Prasad,

who is the CEO and founder of GenMat, who is a

really uniquely fascinating company, and I think maybe the perfect company

for First Principles, because there are so many places that we could take this

conversation. I'll let Deep describe what the company is

actually building, but just to give you a glimpse, it's going to be everything from material

science to AI to quantum computing. There is literally

anywhere that we could take this conversation, even literally space, which you'll

find out how that's possible in just a second. Deep, thank

you so much for joining. I'm really happy to have you. Why don't you tell us

Yeah, thank you so much, Christian. Pleasure to be here. And

it's always a fun time chatting with you. And I'm looking forward to talking about

what we're doing at GenMat. So a

quick little breakdown on GenMat. Our goal is very simply stated,

but very hard to do. The goal is to build an

AI system that is smarter than humans at physics.

That's it. So how do we basically create

something that can, let's say, discover the next generation of

physics theories that advance what we did with quantum mechanics

and general relativity in the early 1900s? How do we

build really advanced physics simulation systems that

can essentially simulate these very hard regimes, like the hypersonic regimes

and atomic oxidative environments in

space, for example, which require a lot of multiple different physics modeling

techniques. And so at GenMAT, we work on

the sort of intersection of physics and machine learning

That's a perfect intro. So basically you're building a physics understanding

AI, like an AI that knows the laws of the universe

and basically can tell you what to expect if you give it inputs and

then something comes out the other end, basically, right? That's exactly it.

I think it's fascinating to think that AI wouldn't just naturally know

that, right? Like wouldn't, okay, so AI is super smart. AI is now over

100 IQ. So how, like, why is it difficult to just program in

just F equals MA and equals MC squared? Like shouldn't, like,

I don't know. I'm obviously just like, Don't need a little softball, but like, why

It used to be a very philosophical question before we

made so many advancements in these large world models, before we could actually figure

out how far can we push, you know, silicon transistor based,

let's say, large language models and their understanding of the

physical world. And so the reason why it's so hard is,

A, because we have embodiment. But let

me take a step back before I go deeper into that. Embodiment kind

of covers three parts of human intelligence,

if you will, the general intelligence that actually affects

our ability to understand the physical world. And those three sort

of components that I think separate us from, let's say, my

dog, right? Who's like, obviously brilliant in her own way, but

certainly not making rockets or, you know, working with satellites anytime,

seem like you are. So our, you know, our team is. So

the difference is that humans have These

three components, which is the sensory apparatus, so

up to 50% of our brain can be dedicated towards visual processing of

information. So it's not that our eyes are as good as a hawk, right?

signal processing that we do, the computational processing, that

makes it powerful. And then there's the second component, which is our

ability to imagine, or creativity, the ability, more

precisely, to simulate physical systems that have yet to

be tampered with. So take, for example, Einstein's thought

experiments when he was discovering relativity, right? This guy's

literally thinking about trains and lightning and then using that to

derive like special, right, general relativity principles.

And so we're like, you know, dreaming how one does as

they traverse through the universe as a photon. All

of these things require the ability to simulate physical systems. and

that to manage and store in memory abstract concepts. And

these representations of the world that are physically accurate and

useful is what's so hard to do even now. Today's large

language models don't come close, actually, when

you come to think about it. And then the third component between

sensing simulation is the engineering and

manipulation of matter. So when you have a system that's

capable of sensing, of simulating the physical

world and manipulating the physical world, you should have something

as smart as a human or smarter. But because today's large

language models don't have any of those things in a really cohesive framework,

yeah, we haven't really been able to crack, you know, just F equals MA. A

four-year-old child, for example, will know more about physics, to

Jan LeCun's point, right, than like chachi

bt4 does or five or six why because when

that child let's say jumps from Four

or five foot like tall ledge like something really freaking high

here right for the kid Like imagine when we did that when we

were young right and we were troublemakers What would we

do? Well, you would jump from a high place and And you bend your knees, right?

And why do you bend your knees? Because the equation for

power is energy dissipated over time, divided by time.

So if you can increase the denominator, the amount of time, right,

you're reducing the amount of power or shock that your body is experiencing.

So we don't have these explicit equations that

our brain is solving when we jump. our muscle memories

and cells and neurons that compute have learned

the right sort of representations of physics. And so all that is

So it's the sensing, the simulation, and the manipulation of

matter. Those are the three things. I mean, intuitively to me, computers

in general are pretty good at those things, right? Like sensing, we

can definitely make cameras that are better than my eyes. And for

whatever, for simulation, we have Monte Carlo sims

that can run millions of possible futures and my puny brain could never handle that. So

I don't know, it seems to me that, why aren't computers better at those

They're not better at those things today because they're not integrated like

we are. So we're an integrated sensing simulation and

manipulation of matter system. However, if

you take a, let's say, sensor in a Tesla FSD, right,

why can't, what's stopping that FSD from

visualizing the warping of space-time under

different metro tensors? It doesn't have

the physics simulation-based models, right, to do those

things. But those are the things that we work on. So hopefully one

day in theory, right? Maybe your Tesla can think

about relativity because of us. And

so where the price of manipulation matters. So, you know, the Tesla in

this case is moving around, you know, you're trusting it,

but it's not going to go and build like an arc reactor

in somewhere in a cave, right? To go and test its physics theories that

I can't think of right now. So yeah, integration. And

so our brain has had, right, like billions of years, to

really integrate all these systems and evolve the integration of the systems.

So that's the edge we have currently, but we'll see how long that

Presumably the fundamental constraint here is sort of twofold. There's

like, one is finding ways to represent the laws of physics to a computer

in a way that a computer can understand it. I mean, I'm sure at some point you do literally

plug in like f equals ma or like equals mc squared or something. That's

probably part one. And then part two, I would assume, is just you have to

get a ton of data. Like, you have to feed in, like, I

don't know, what happens when you drop a ball? Or what happens when you drop,

like, jump off something? Like, you probably have to feed in examples of physics. And

so I wonder, like, what did that even look like? Or what did physics data look like? And

how do you capture it and feed it in? So I don't know. We could take you those two halves in

I love to, like, yeah, look at what the physics data looks like and

the difficulties there. And really, it's a combination

of both. Our equivalent of f equals

ma, of dropping a ball, and seeing how each has a little bounce given

the friction coefficients and so on, is

we started with the atomic world first. There's a reason for

that, and I can get into it if you like at any time, but

we wanted to build our world models, our physical world

models in our inter-AI systems truly from the

ground up, from base reality up. because

that's the best way to predict everything else, right, is

if you can model the underlying crazy atomic worlds, since

that dictates everything. And so for us, that data

of like, what does a different basketball under different scenarios look like? A

real life example of that is our titanium dioxide photocatalyst system.

So we recently put out a preprint, for example, where We

created the following setup. We have this titanium dioxide nanoparticle.

There are 10 to 25 nanometers across. And

on top of those nanoparticles, we deposit copper and

platinum nanoparticles or nanoclusters. And

these sort of like nanoparticles are just globs, right?

If you want to think about them more like amorphous globs. of copper and

platinum. And the way that their configurations end

up completely dictate where the photocatalytic activity

will be. And the photocatalytic activity is

what we care about because this system absorbs carbon

dioxide absorbs water and sunlight and

turns it directly into natural gas in a very clean reaction.

There's no waste products, all you get is literally methane out at

the other end of it. So what we've done over the past few

years is taken this system and we've looked at

optimizing the nanoparticles and determining what

configuration of the copper and platinum atoms would be

most conducive for sucking in more CO2, basically. How

do you accelerate that? And so, traversing that

search space, combinatorial search space configurations, goes up

very quickly, exponentially, essentially, and factorially, when

you take into account that you have three dimensions. So, one, sort

of, just take the nanoparticle system that's

sitting on top of the TiO2, right? So there's a combinatorial system

where The copper and platinum atoms have different permutations

that they can be in. And the more of those atoms you introduce, the

more different configurations they can all fall under. And on top of

that, you can introduce different oxygen vacancies on

the surface of the TiO2. When you pluck out

the oxygen atoms on the surface of the TiO2, you

also change the catalytic activity and what's happening at the interface. And

the interface between the TiO2 and the nanoparticle now

is this really rich area of super complex chemical activity

that you can optimize to do something useful. And

so it's being able to model those things, creating data sets

that represents all these different systems, form first principles, and

then validating our predictions in the labs is where we

started. And then over the past year and a

half, we expanded into geophysics, which is where our satellite

Yeah, wow. Okay, now I'm definitely convinced that computers are way better

than me at this. I got a

little bit of that, a little bit, but I mean, I, what I, what I

really take, so what I take away, what I take away from that is basically you

are trying to super, super detailed and in, you

know, a lot of granularity model, literal atomic level

activity. You're taking, you're saying like, what happens when you put titanium,

titanium oxides, like titanium and oxygen together with copper

globs, like what, what ways can the globs be oriented? And

Trying to change something, you change the material properties by the way. So

good luck doing that for millions and billions just for a tiny, small material.

You see, you're simulating to try to figure out which things are worthy of

test. And then once you have like a candidate list of

stuff, then you go and like test it and actually like measure the material properties. Is

So how does it work today? Like if I'm a, if I'm a material scientist today and I'm trying

to like come up with the next, you know, room temperature superconductor, like

what do I, what do I actually do if I don't have a physics

Yeah, absolutely. So right now, it's very much trial and error. Literally,

you're mixing shit. That's it. It sounds super complicated, this

kind of synthesis and characterization, but it's just, that's what dudes in

the lab, right, and girls and so on, just mixing different

chemicals and solvents to synthesize different compounds. And

they're doing it based off of decades of experience and intuition. And

very, very, very slow archaic computational chemistry tools

that are written in Fortran. So that's the art

of material science today. And that's what powers the semiconductor industry,

the aerospace industry, the battery industry. All these industries are

I know about the discovery of some materials. I know that, I

don't know, back when we discovered bronze, it was literally just a mistake. Someone

accidentally dropped some iron or something in copper, and they're like, oh shit, this

is actually harder. This is actually better. And this happened with

penicillin. I don't know. There's rubber. I think that's literally how

vulcanized rubber was made. It was an accident. So it

makes sense. It makes sense that people are just... making

mistakes in labs, and then every once in a while it turns out this super material. So

Yeah, absolutely. In fact, even when you have, let's

say, experts with some intuition of what they're doing, like

some idea of what they're doing, right, it's not just a straight up lab accident

that they're banking on, like perhaps the potential LK99,

right, submit Submitters were even

then it takes a decade or deck many decades to drive

the material. So if you take the first high temperature superconductor that

was discovered, what's awesome is that was discovered by

a private company, IBM, right, that was discovered by IBM

scientists. And when they discovered it, There's

a lot of history to that. So it was the world's first cupids, right?

These copper-based, essentially ceramic-based superconducting materials,

which in itself was a profound sort of discovery, right? That, oh,

wow, you no longer need just a metal or perfect conductor

or a normal conductor to become a perfect conductor. It can literally

just be this weird material, right? And so that

wasn't by accident, however. So Werners and

Muller, had spent the past sort of decade,

particularly Mueller, really just focusing on

the crystal structure of what would be expected, right, of

these cuprates and of high temperature superconductivity. So

they were looking at all these other crystal structures that

were precursors to the final cuprate crystal structure, the

arrangement of atoms that's like essentially fingerprint for

that particular material. So that family of fingerprints they

had been looking at for so long. And so finally they

like stumbled upon this, but that's what they were looking for, right? And

it was because of some, one of the key ingredients, funnily

enough, was access to some of the first like really

large-scale density functional theory large-scale for their time in

the 80s and late 70s so that's pretty awesome too

right it's a combination of computational chemistry and

real experiment and you know maverick science coming together in

a private industry setup so we should be trying to do that again

Today's episode is brought to you by NordVPN. Now there are

lots of reasons why you'd want to preserve your privacy and secure your

online browsing. There's the normal reasons like not wanting to get hacked, which

is pretty essential for deep tech companies like the ones featured on this show. But

there are also some other ways that VPNs help you that are less known.

Like you can get deals that are only available in certain countries and

you can get access to shows and other media that are again, country

specific. And if you're going to get a VPN, use NordVPN. They're

the fastest on the market today. To get the best discount on your NordVPN

plan, go to nordvpn.com slash first principles. So

you'll add on four extra months to a two year plan and

there's no risk. There's a 30 day money back guarantee. So check

out that link, which is in the description of this episode or go to nordvpn.com slash

first principles. Let's get back to the show. So maybe let's take

a step back and just talk about materials. Like, what is material science?

And what are the properties of materials? Like, I imagine what you're

searching for are specific properties. Like, you want

to find materials that are really conductive, or like super conductive. You

want to find materials that are extremely hard. You want to find materials that are whatever,

like a variety of different things. So do you mind just telling us at a super high level,

like, what are materials? And why do they matter? And what makes them good or bad?

Happily. Happily. I love that question. So materials, which

is arrangements of atoms that repeat at

a scale that's relevant to you. That's how I would define it. So

you wouldn't care about like three, four or five groups of atoms, right?

Or like five atoms. But you care if you have trillions

of atoms that intricately dance with each other and represent

the next generation of semiconductors, right? That are more power

efficient. You care when it's carrying you to the

moon and to Mars. That's when it matters. So that's what

I would define a material. And so you can have many, many different

kinds of material properties that you're interested in, right, all the way from

conductive properties, like you were mentioning, like electrical properties,

thermodynamic properties, structural properties, those

are really common as well, right? So within those, you

have a very wide range of trade-offs. Every

material has its own trade-offs for its properties. For

example, you might be interested in looking for a transparent solar

cell, like a truly invisible solar cell. We're

about 70% of the way there last time I checked. We

can create about 70% translucent solar cells. That

field of research is largely being enabled by perovskites, which

in itself is a fascinating material, right? So in

material science, let's go back to the like 30s

and 40s when perovskites were first discovered. There were these interesting

rocks with these interesting piezoelectric, you

know, electric qualities to them, right? People were around

rubbing them together and seeing that they create charge if you hit them against each

other. And they're like fighting other rocks like that. And

then we find out that about 70% of

the Earth's crust, or a very, very large percentage, have

these perovskite materials. It's largely composed of

perovskites all the way down, instead of turtles. And

so in the 1960s, we have one, about

60 research papers, let's just say, per

year on the subject of perovskites. And

I'll tell you what they are in a second. And now

today, we easily have 80 to 100

research papers published per day in that field. It's

the most famous, I would say, after superconductors or even

on pallets. It really does become the most famous material that people actually

put resources towards in practice. And

there's a reason for that because perovskites are these

They have an ABX3, so an

anion, anion, and cation-like configuration.

Basically, the ABs are anions and the X3 is

a cation. The

anions create this octahedral pattern. when you

repeat the atoms, right, and you look at the fingerprints and

what they look like at scale, they create these octahedra. And

when you tilt the octahedra, meaning when you change the bond lengths,

when you change the chemical composition, you tilt the octahedra,

or you displace the cation in the center of the cubic

lattice, it changes the properties completely. So

just displacing the cation, if you want to visualize it, and the cation, And

then you have an octahedra covering the cation.

So if I tilt, axial tilt, this octahedra

across X, Y, or Z, just one tilt, even a little bit, changes the

properties. A little bit more, changes the properties. How are you possibly

going to predict all those things? But anyways, right? So that's why you

have so many papers, because it's a straight up surprise. You

So these materials that are in solar cells, basically,

we've known about them forever. So what do you mean that

they're in all of the Earth's crust, that they're in everything? What

Yeah, so those magnesium-based perovskites that

are largely what the, I believe, Earth's mantle or

Earth's crust is. Oh, OK. So that's one

example, right, where it's literally a large perovskite-based mineral

It's the name of the geologist slash

And the class of things is like the shape

of the thing defines it, if it has that

Each of the add ions creating the octahedra defines

it. Wow, cool. It's just an infinite family

of materials that it represents. And so the

cool thing is that these near-invisible scellular

cells are based off of the perovskites. The

first PLC electrics were perovskite-based and

are mostly, largely perovskite-based. Currently,

for example, in our labs, we work with perovskites for lithium-ion-based

solid-state electrolytes. How do you find a much

To tie this back to what you're doing as a company, let's stick on this

example. Let's keep talking about this new material. Would

you help, would a lab come to you and say, Hey, I have this idea for

what I think we could do with the orientation of the thing. And can you

please test it for us? Or like, are you trying to discover your own

ones yourself? Or maybe both? Like, can you just make it tangible, like what you're doing

Happily, happily. So we're kind of doing both. We're dogfooding our

own products and technologies for what I call ubiquitous materials.

So we don't care about creating Let's say

a material for a better, lighter, stronger weight tennis racket,

or like a better golf club, right? That, like, don't care if somebody

else can commercialize that. We're happy to help them discover that, right? No problem.

For us, we really want to focus on these materials that will change

and affect all 8 billion lives. Like, that's our metric. So, for example,

what does that look like? A truly 100% or 99%

invisible glass panel, right, that's also a solar cell,

that's totally up for rally. We're definitely interested in that. A

lithium ion battery that's, let's say, an order of magnitude safer

and more energy dense than what's available on the market that Tesla can

deploy. We are happy to help find that and work on

that. But then, yeah, so we want to kind of do

both of these things. And I think that there's a

world, right? The world of materials is large enough that

there's never going to be enough players making their own materials, making

their own supply chains and processes to scale those materials.

And so that's why we think that, you know, we just need to keep contributing

like crazy to the acceleration of these materials, one way or another, whether

it's us or helping other people do it. Our

basis model is that we offer one product and two services around

the materials discovery side. And then I can get into

our geophysics-based models, because I mentioned we expanded recently into

another domain of physics, right? But so for materials work,

we have one product, and that product is a series

of physics-informed machine learning models that model crystal

structures and chemical compositions. And they're useful

for a variety of tasks. We have some models that are really, really good

at predicting the properties. So let's say you change these

parameters, the axial tilt, the anions, and so on. What

can we expect? We have models for that. We also have models

that look at what happens when you have a very

complex interface, like the titanium dioxide example that

I was talking about, where the interatomic potentials are

very complex and you need to discover where a representation for

it. And then we have generative models for the inverse design

task. So that's for when a customer, and

this is what we're finding most people are interested

in, is particularly like a much larger market. We

also have models for the inverse design task or capabilities where

you give us target material properties, and we will

discover, right, and we will figure out and generatively dream up

Are there certain properties that are fundamental enough for

you to model and other properties that are sort of like emergent and you

Yeah, and when I say can't model, I'll be specific, can't model with

today's physics AI technology. Like it's constantly an

active field of research for us to be able to, you know, improve this

multi-skills and multi-physics domains

Yeah, absolutely. So fundamental ones like the band gap, charge density, local

density of states, electrical properties like connectivity,

both DC and AC. We can also

model ionic conductivity if you're looking at battery materials. So

information energies as well is a really, really critical fundamental property.

So yeah, we can model these fundamental properties, which already

give you a lot of information about is this material going to

be useful for your particular engineering task at

hand. So, you know, let's say you're trying to find like a really efficient solar

cell that works really well in atomic you know, oxygen environments in

space, that would be a set of material properties

for us, right? Charging properties that we would then just like go run

our models against. It's one of the directions that we kind of going. We

are trying to elevate the aerospace, you know, in space industry. Well,

So, so those are fundamental properties that apply to a

wide variety of materials. and material families, from

transition metal oxides to, let's say ceramics, right,

you're looking at formation energy and so on and so forth, battery materials, solar

cells, you get the point. But there are also measured

properties that you asked about, and I think that's a brilliant question. Because

there are emergent material properties that require

deep quantum computing and quantum computing related technologies

and thermodynamic computing related technologies to exist.

And so that's what we can't do today. For example, modeling

high temperature superconductivity. right, is

a unsolved problem because it requires being

able to capture really strongly correlated

electron systems and many body systems. And

it's so computationally intractable, even for the most powerful silicon semiconductors

based AI models today. So that's where we're

constantly like pushing up against the physics battery, if

And so today you're just doing this all with classical computing or do you use

We use quantum computers only as a test bed to test like how far

I know that you originally, the reason I asked that though is because I know that at one

point in time, you and Guillaume, like the guy who's the

CEO of Xtropic, people might know him as Bef Jezos, we did an episode on him

a little while back. I know that you guys knew each other back in

your university days and you were interested at the time in quantum

deep learning. Was there a time when you were founding this

company and you thought it was going to be a quantum company and then it ended up just being classical?

Or is it just like, you know, someday you

Jay and I, we both like determined, right, that okay, quantum

hardware right now is not going to make it for quantum chemistry tasks. So like,

one of the things he worked on was Even

now, apparently, the world's most powerful quantum computing simulator, right, based on

benchmarks. So, you know, he built that at Google X

in the secret division that we built. And

so, that was like, really a lot of that, some of that was

boring, right, out of this acknowledgement that him and I had on

the sort of limitations of being able to do quantum chemistry tasks

today. That simulator, even today, and the concept of

using simulators and decomposing problems across simulators remains

an active field of research that we're pursuing. where we've

recently been able to perform certain simulations, certain

material properties that we can calculate thousands of times faster using

classical hardware, but using an entirely different paradigm of

quantum probabilistic algorithms. And so those quantum probabilistic

algorithms will natively work on quantum

computers once they come online. That's one of the benefits

of writing code like this is that you're already getting benefits from

this quantum computing simulator that's giving you useful chemistry calculations

and on top of that you're ready so that

when the hardware comes along, it's literally almost a back-end switch, right? Like

you're just swapping the device driver to a quantum computing

device. Now, that's one aspect of

our strategy. But the other part, this is the part that, like,

yeah, I'll probably only say a few times, I don't know, just because

it brings a lot of... Not saying it, but we're

also deeply interested in accelerating quantum computing. I'm no

longer interested in being an observer. A large part of quantum

computing is material science-based, like constraints. And so

We talked about a bunch of, we talked about kind of the problem, like how you're

approaching it, why it's important to solve, but we

haven't yet talked about why AI is the right tool. So I'm curious, like

why, when you were trying to do this, this cool material science discovery,

So the reason why I chose AI is because it's ultimately one

of the best pattern recognition, if not the best pattern recognition

device we've ever created. That's really all it comes down

to. Nature has really, really deep complex, hard

to penetrate patterns that show up between all of

these crystal structures, between all the electrons and their bond links

and the properties. There are patterns that are meaningful in

mathematical and chemical and physical nature. And so teaching

a super intelligence that I think we have a higher probability of succeeding than

Can you talk more about the specific techniques that you're using? Is it

deep learning? Is it some other thing? What of

So we use two things. 99% or 95%, if

you will, is classical deep learning. For example,

we use something called CDVAEs. So

crystal diffusion, variational autoencoders. This

is one class of models that's really interesting to us, right? There's

some open source CDB out there, for example, that you can

play with and see exactly what we mean. So that's

taking the best of diffusion-based methods and variational autoencoder

I know what neither of those things are. So it's like a 0% helpful to me. Let's,

let's talk about it though. Let's, let's get into it. Let's get into it. So what is,

um, so let, let's, let's, let's set a goal of Christian has to

understand what a CDVA is by the end of like five minutes. Uh,

so what, how would you, how would you describe what a D like a diffusion model

Alright, let's start with the Variational Autoencoder first,

the VAE. A VAE takes some

input, learns a compressed latent representation of

that learned input, and then it has a decoder. So

the autoencoder is encoding the input into some

latent compressed representation, and the decoder has

to get really good at taking that latent compression

and decompressing it back to the original input, right?

And so the autoenclover has to get really, really efficient at compressing different

inputs into smaller and more useful and

more information rich latent representations of the

And latent, latent meaning in this case, that it's not like a

literal one-to-one, I see red, therefore I

put red in a database somewhere. It's like, it's latent meaning that

it's just some number and kind of inscrutable. Like those

Right. Bingo. That's exactly it. So for example, it's

like. What you don't want to see in your occipital cortex as

like visual data gets processed right like it would look really if

you could see that would be almost psychedelic like right like you would see like, instead

of, you know, two eyes for deep right it would be like two

Yeah, totally. If you could peer

into the brain and like look at what neurons are doing or something like

that. If you looked at neuron signals, it would make no sense, but it

processes it. So it's like the middle step between the processing. Okay.

So VAE takes inputs or whatever, turns

them into crazy gibberish, but then there's another part of the other side that

can take that crazy gibberish and turn it into meaningful stuff again. Bingo, exactly.

So that way when you give the VAE a totally new input it's

never seen before, right? It should be able to compress it

in a meaningful and useful way and then the decoder can meaningfully decompress

it, right? And you can make predictions at the ends and

beginnings of both of this process. So, for example,

with the VAE, right, you're feeding it in crystal structures.

By crystal structures, I literally just mean atomic coordinates of

materials, their fingerprints, right? Where are all the electrons and

the atoms, the neutrons and so on, particularly in the bond

lengths. That's it. You feed those in along with

the properties, because each one has different lists of

properties, right? So that now you're learning a compressed version.

of those crystal structures and their associated properties so

that when I give my VAE a new crystal structure that

it's never seen before, it should be able to come

up with a useful compression that allows me

to deconstruct the entire input, including the properties, as

an example. So that's the magic of the VAE. Now,

what does a diffusion model do? A diffusion model is

infamously known for incrementally adding

noise to a model and then training a

sort of loss function that you can go backwards and start with a

very noisy image and come up with a very coherent image. I

won't go into why Gaussian distributions and so on matter.

What I will say is that if you can teach an AI

to take a very, very noisy image and find and

learn a representation of working backwards to the original images, you

basically learn a representation. of coming up

with any coherent image, because you've trained it to come up with coherent

images thrown through, given noise. So if you give

it random noise, right, some form of structured noise, even, or

whatever, sample noise, the expectation is that it's forced

to come up with a realistic reconstruction

of coherent one, and that's what we exploit. But that's also why you

never get the same images. Every time you run mid-journey, it'll

never be the same, because you're literally just injecting noise into your

system. So now you combine the two and that's how you

A crystal diffusion or a crystal structure.

Oh, we put diffusion, and then, yeah, for variational

So crystal diffusion, variational autoencoder, is

a thing that can, that you feed it, presumably, the way

that it gets better is you just feed it tons of data in the beginning. You

say, here's a billion different crystals,

a billion different properties, and then it just learns what to

do with those sorts of things. And then on the other end, the diffusion part

of it is basically like, the equivalent of mid journey, like where it

starts all blurry and then you kind of get the lumps, random lumps, but then it like

coagulates into something that actually is meaningful. Like presumably that's

what's happening on the diffusion side of the other. Okay. So you are using, um,

CDVAEs to do the, these like

material, um, material predictions basically, and material,

and also that, what you call the reverse thing where you're actually like predicting,

And so, yeah, so the interest design problem is given a set of target properties,

you know, find me a thousand, right, or 10,000 material candidates with

these properties. Which, right, you would have had to hire entire teams of

material scientists before to even, like, get close

Seems hardest to me is the data part, is

getting billions of crystals and somehow loading them into

Yeah, yeah, and I can see why you think that it was a hard

part. So largely, there's, I would say, well,

all three things that have come easy to large language models

are very, very difficult for large physics models. Right.

So, um, The difference, for example, data, right? The

training data, if you take large language models, the training data is

available all over the internet, right? You just need a coherent sentence, literally.

It just needs to be a complete, grammatically accurate sentence,

and that's now training data. And then processing

the representations, will transformers do most of

that work for you? And so there's obviously an active field research

for transformer architectures that impose these

attention mechanisms that are useful for learning, all these

extra subtasks implicitly. And then you have the

RLHF. You have thousands of people, hundreds of thousands, millions

of people who can RLHF the output and roast

your AI. That's pretty easy. But

if you have a physics-based AI system, coming up

with valid training data is very, very expensive. You

need deep domain expertise, labs, lots of

compute. So for example, the data for us is our experimental labs.

It's also the synthetic data sets that we create using

proprietary know-how methods. It's also first principles computed data.

that we create, and that's using these massive DFT softwares,

very archaic software. What's DFT stand for? Density

Functional Theory. Density Functional Theory, okay. So

Density Functional Theory is a computational modeling

tool. It's an approximation to the many-body Schrodinger equation.

Every single material has a unique fingerprint. Every fingerprint

has an associated many-body Schrodinger equation to it. If you

can solve it in its exact form, you would be able to predict pretty

much all the material properties that you're interested in, all

the properties. And so the problem is that we

can only solve that for about three atoms. Beyond that, computationally

impossible. That's where density

functional theory was introduced. It's this computational method. So

we have these old school software codes, right? Like creating data

sets as well, in part. So it's this, how do you

just create really high quality physics data? It itself is

an art and a science. So that's hard.

I totally agree. But I would say even the representations are

hard too. So how do you impose physics

symmetry constraints? For example, if

I may have a molecule, let's say, right, a titanium dioxide. And

I may rotate that titanium dioxide molecule a

few degrees across, let's say, the x and

y plane. So I just rotate it a little bit. To my neural

network, which is my CDVAE, this deep learning algorithm,

it's only seeing numbers of positions of an

electron. So if I rotate it this way, all

the numbers change completely in no known and obvious way. So it

will think if I don't include physics constraints and symmetry constraints,

that is new data that it's seeing, that it's seeing a totally new

Um, if I show, if I show my phone like this, then it can see, oh, look, it's like

a shape like this. But if I show it like that, even though we obviously

know it's the same phone, like it's seen, it looks different because they're different pixels

That's exactly it. Basically, how do you do 3D object

recognition, right? Recognition, but for atoms, so

there's difficulties there. So for the physics symmetries, in

this case, you would impose E3 covariance, right? So it's

a part it comes from like group theory So group

theory is a form pure mathematics that studies mathematically

describing quantifying symmetries so you can have symmetries in

physics and you do deeply right like you have energy symmetries

and constraints conservation momentum and and so on and so

forth. So yeah, all these things have categories and

groups particularly attached to them. So anyways, you

have to impose really, really deep, complex mathematical rules into

the neural networks that don't just come out of the box for

large language models. And then the third part, the RLHF, you

Yeah, and that RLHF stands for reinforcement learning with human feedback, right?

Or with human features or something. Is it feedback or features? I never

know. Yeah, you better write the first time. Okay, okay.

What that means is basically you show people output number

one and output number two, and they say, oh yeah, I prefer output number one or something like that. But

if you show me this crazy material, like, I

have no idea what to do with that. Like, I don't know how to rate that on a scale of one to 10. So that

makes sense. What does that actually look like? What does the reinforcement learning

part look like? Is it just doing more measurements after you create some stuff in

a lab and testing to see if it was good or bad or right or wrong?

Yeah, that's really where it comes down to is largely nature

herself becomes the arbiter of truth for us. Like

nature holds the feedback for us. It's like NLHF, right?

Or sorry, RLNF, right? Reinforcement Learning from Nature Feedback.

So that's like the hard parts. I imagine that you're

probably just very complicated. And for a lot of people

that are not detailed, as expert as you in a lot of

these areas, I imagine it's very hard to communicate, hard

to get people the necessary background information they need to understand it.

But, and yet, you probably have to hire a bunch of chemists

who don't have all the AI background. You probably have to hire a bunch of AI people

who don't have the chemistry background and vice versa. My

question is, is there any way that you actually do that like as a

company or even just as you like going out and pitching investors or something? Like how

do you try to get people the first principles understanding they need in

I try to communicate it in a way that they can relate to deeply and intimately.

Let's say I'm talking to someone who works construction all

day long. That person may not have

the same deep technical appreciation of chemistry as

someone with a PhD in chemistry. And

that person, the patient chemistry, may be a complete novice

when it comes to applied materials in the real world, like construction

workers who actually understand melting constants and

so on and so forth, melting points and real structural integrity, what

that actually looks like. So there's a, you know, knowledge gaps in both

places. And so that what keeps it, you

know, kind of threaded right in together is that

shared understanding or appreciation actually more importantly of

materials. So I always take that kind of first principles approach.

And so that's always been pretty useful in

retrospect. It's just like finding something in their day-to-day lives that

matters and then kind of just showing them the appreciation of

how hard it is to actually find materials like that. A lot of

people, I think, just kind of take material science for granted, right? That

plastics, leather, silicon, all these things just exist. They

don't. They take very, very hardcore labs and a lot

of scientists, a lot of time, to discover. So

yeah, just communicating that. And then communicating the physics as well, right?

So like the baby example or the child example of jumping from

a high top, everybody understands that. So

Like if you, if you're just looking at your company for the first time from the outside, it

does seem like you have your, your, you know, you've, you've all these different areas

you're trying to do in sort of like all at once. Seemingly like you're, you're

not only doing like materials discovery, but you're also doing materials

analysis. Um, you're not, you're not just focusing on one material. You're focusing

on like all of them. Like you're doing photoelectrics, you're doing like whatever,

uh, like conductive stuff you're doing all these other, like you're doing so many different kinds

of things. And so is there a reason you chose to do

that? Is there a reason you chose to go really broad instead of just being like the one

ceramics lab or something? Absolutely.

I laugh because that's like the first question

That is such a common question, right? Where is your focus? And

my answer is always the same. My focus is the system. We're

trying to build an AI that is better at physics than all

humans combined. That is the goal. And so that

AI can't be trained on just ceramics, or just batteries, or

just geophysics, or just computational fluid dynamics. It

really has to sample from all sorts of domains of physics and

truly be embedded across all different kinds of sensors around

the world. So I'm kind of taking that systemic approach. It might look dramatic,

I just don't have, we just don't have the us puny VC brains or

Okay, so let's let's talk about the geophysics stuff.

And let's talk about space. Let's get there. So why? Why? Why? Why

add another thing? Come on, man. You're not focusing. Yeah, like, tell

Yeah, if disrupting material science isn't hard enough, then why

meet yourself with trying to disrupt geology, right? But

basically, the benefit is twofold. In the immediate

short term, our short term, when I say that, like

10 years or less, right? Massive benefits here is

that we're going to be able to help with critical materials and

critical minerals supply chain. As well as just a general raw

material supply chain, which also benefits us and our customers. As

our customers at, let's say, the SpaceX's, Tesla's, whatever of the world, right,

are prompting our AI to look for new materials, new alloys for

their rockets, for example, or engines and so on, there's

going to be net new atomic species or elements, net

new as in versions of these raw materials that we typically

haven't cared about right so Pick pick your poison.

I'm gonna pick an extreme one like frankium obviously it wouldn't pick something like

that, but you got the point right so There's

gonna be a sudden increase that we foresee over

the next few years in absolutely new supply chains for

raw materials that don't exist and so if we can

put up satellites in space and and have geophysics-based AI

models that reduce the cost of exploration and time for

exploration of those new minerals, we reduce the cost of

the raw materials themselves. So that's a very simple

first-order effect that we want to exploit or

leverage, if you will. And then the

kind of longer term goal that this serves, the

geophysics and satellite work, is going back to the original

three components that we talked about. So the geophysics satellite

work, hyperspectral work, also has a secondary benefit

that's more longer term. And that goes back to what we talked about earlier,

about the three critical components to general intelligence from

our perspective. And so that's the sensing, simulation, and

manipulation matter. So the more sensors we have in space, on

the ground, underneath the sea, throughout the

Yeah, it makes sense. So the, basically what you're

trying to do is be able to have cameras, like hyperspectral cameras,

right? Not just like optical cameras. So a

camera that can see every sort of wavelength from the narrow

little, like whatever we can see into the infrared,

into the ultraviolet, like just even beyond that probably. So

it's like, yeah, it takes all of that visual, visual, call it

visual data, and can look for

materials. So you say, we know that copper looks like

this when it's on the ground has this particular pattern. Basically,

That's the simplest way to describe it. That's right. So we're looking for signatures

of different minerals. Copper-rich minerals versus iron-rich minerals will

have a totally different reflectance pattern underneath the hyperspectral signature.

To you and I, it's just going to look like dumb rocks. But to a

specialized camera, they're very, very pretty and colorful. And each

color is super meaningful in terms of the chemical and

crystal structure decomposition. of those minerals. And

so some deposits, what we want to be able to do is, exactly

as you said, point to deposits, right? And point

to drill targets. Where should we actually drill and

look for these precious deposits? If

you take something like a periphery, the causal relationship between

the surface scans And the periphery are

a lot stronger a periphery is a geophysical

sort of phenomenon that occurs across millions of years when you

have a volcano nearby, essentially what happens is you have

really really hot magma is kind of circulating. and heating

up all these hot like metals and precious metals like gold and

silver and it circulates these like really

precious metals that we care about across millions of years deep

underground up anywhere from 100 feet to several

hundreds and thousands of feet beneath the surface and

so you have kind of like a eye of sauron looking thing

And around it are all these, like,

it's literally gold and silver and nickel and other things

you care about. And then on the surface, it's really,

really low grade material, but it's still like a lot

of gold, a lot of silver. And so a lot of peripheries have

yet to be discovered, by the way. That's the cool part. So

there's a lot of surface mines that yet haven't

even been detected. And we will see them with the right AI and the

So you actually launched a satellite, though. You launched a hyperspectral satellite

on a recent SpaceX launch on Transporter 9, was it? Right.

Oh, man, that was so nerve-wracking and exciting. It's like easily top, like,

experiences of my life. Hats down. That

was so cool. Like, you know, like, I'm sure you experienced this too

when you were going through your first payload. It was just like, Alright, we're putting a

lot of money into this thing. There's a lot of hardcore engineering and

risks associated It will literally go into space at tens

of thousands of miles an hour and stay where, you know,

we're about that speed. So, you know, like this

can go wrong in so many ways, right? So it's very nerve-wracking all

the way up until the point that they say, you know, genmat1 deployed. And

then after that, you know, it's like, all right, we can solve it a little bit.

And then it's like straight to Leops, right? Like how

do we manage Leops? So it was such an amazing

and thrilling and like high impact experience. Dude,

it was so inspiring. I was like, This is a private individual and

his company and 10,000 plus people who believed in the mission. In

just 20 years, this is what they've done. Like I'm literally watching a

frickin' skyscraper go to space and come back right now.

That's awesome. And can you say anything about how it's going? Like,

you know, how you guys did through, and LEOP is launch and early

operations phase. So like, how did LEOP go and how's the

Absolutely. So we're currently at the very final end of LEOP

where we're working with the ADCS and GPS. So once

those guys are stabilized, so the ADCS is the

Attitude Determination Control System, which you know, but just in

case for your audience, it helps you determine the

And so we're currently just operating and fiddling around with that, if you

will. And at that point, we're going to be able to start taking our first pictures. So

That's so exciting. Oh, my gosh. I remember. So we actually, Astronauts

launched a demo satellite before we did our big one. We

took a picture of San Francisco from our little tiny

satellite. And when that came back down, it's like that

picture is iconic in Astronauts history, just because it's like that came from a

That was so cool, right? It's like, yeah, that was

Yeah, you're about to be there, unless you've secretly already

captured some demo images and downlinked them, but I won't press

you to see if that's true. So basically, you

are probably also then developing the algorithm side

too. They're like, actually, how do you take in all that hyperspectral data and

do stuff with it, identify the deposits

or whatever? So how similar is that to

the other stuff you're doing? Is it a similar sort of like physics problem

that you have expertise solving because you've been doing these material science

Yeah, there's certainly some overlap. That's

a great question, actually. There's overlap in some

of the expertise. So for example, in geological soil sampling, we

followed Pat on this recently, where basically in

sampling, You'll do something called XRD, X-ray diffraction, and

you'll look at the crystal structures, right? You're basically taking an X-ray of

the Earth, like of the surface of the Earth, like what minerals there are, and

you map out their bones, aka their crystal structures. So

that is done in material science too, in high throughput synthesis of chemistry

for just like new ion, new battery discovery. So that

process is very, it's literally the same. So the

models are the same as well. And then there's also sort

of meta overlaps. So going back to what makes physics

AI hard compared to large language models, The

training data requires domain expertise. So to your

point, you know, we have a lot of domain experts in geospatial, geophysics. We

work, our investors is, you know, Comstock, right?

They're Nevada's oldest publicly traded mining company. So we work with

real world geologists too, to like validate our models. And

that helps. And then yeah, we also innovate on the algorithm side

too. So how do you make useful predictions and Because

we worked on doing these physics imposed constraints

and symmetry constraints and learning the nuances with material

science, a lot of those learnings quickly we were able to

Very cool. Can you tell people, can you give people a sense just, I mean, across the

material science side and this side, can you give people a sense of sort

of how far along you are? Like how many people are at the company and

like customers and sort of like traction and figuring out these

things actually, and being able to make them useful. I imagine that people

are probably like, okay, cool. It sounds like a cool idea, but like that'll never happen, but it's

I love it. Absolutely. So, you know, we're $15 million in

and 40 people in three years worth of like results and

work. So, you know, we have preprints on the LK99. We

simulated it. We have preprints on our own materials that

are titanium dioxide nanoparticles. where we've

shown that you can actually use AI to dream up these

new configurations and chemical surface reactions. We put

up and built a wet lab. It's probably one of the most

advanced photoreactor combined with machine learning workflows in

the world right now. We have a photo reactor in Toronto. That's

where we're literally creating, you know, natural gas right from

thin air, essentially. So we have that going for us.

We also, of course, have the satellite, as we mentioned. And

we also have a traction both on the technology and

customer side that includes, for example, We've

now been able to generate tens of thousands of crystal structures

across a wide base of material families that

are all potentially trillion dollar markets. What we want

to do is continue to validate that with actual industry

experts in those fields. So we won't be able to say

names, but we've attracted the attention of pretty much every brand

name you can think of. Like Ask Crazy, statement as

that sounds. So we have that kind of traction

that's like really interested in adopting our stuff. And then on

the geophysics side, I'm really excited as well, because we

were able to recently pre train, it was a transformer based

architecture that we completely hacked to include constraints that we cared

about to predict hydrothermal alterations. So

an alteration is a part of the ground where the chemical composition randomly

changes, like it just changes unevenly. And again, this

is something you may not visually see, right? But it's like

very subtle, but very powerful. And so how do

you teach a pre-trained model to predict that? On low-res Landsat

data, which is like 30 meter resolution per pixel, as

opposed to five meters, which is what the GEMMAT-1 has. So we

were able to predict alterations, hydrothermal alterations, on

the COMSTAR And that was validated by their

own geologists. It blew their mind. And this was done using

a model that we pre-trained on much worse,

lower resolution data. It's not even using our existing satellite

And what does this look like at the limit? Let's

say it works perfectly. Everything you have planned comes

to pass. Where are we? Where are you? And where

I want to bring us towards a world that I call Yuk-Siberia. So

you probably see that on my Twitter a lot, like it's in my handle, right?

I still have to write like a proper blog post describing what it is,

to be fair. So this is a good like preview. So

Yuk-Siberia is

the third level of civilization that I think we're headed to. And

I'll kind of give a brief guess as to timelines, right? Don't hold

me to them. way off one way or another. don't

even know anymore about AGI. Like this is obscure, you

don't even realize it, right? Like this warps anything, any

predictions about the future. But that being said though, you

know, for us it's like, all right, where does physics AI, physics-based

AI take us in the infinite first? And the first level

of civilization is one where I think we will exist

in a world where humans are constantly creating matter that was assisted

by an AI. So every piece of matter we ever create was

partially driven by an AI under our guidance. So we'll

have these beautiful, crazy, futuristic maglev train systems,

right? Consumer electronics that you don't have to charge for centuries. All

these things will be the desires of a human, but service to

us by an AI and just optimizing the atoms, making

it more sustainable, right? So just solving all the things that we currently

suck at because we don't have like chemical, quantum mechanical

intuition, frankly. You have to build it over years and years and years. But

if you can teach an AI that from the get-go, right? Totally different story.

So at the first level is one where you have

hypersonic maglev trains, like you mentioned, megalithic structures. But

then the second level, I think where we're headed, is where the

matter itself becomes so sophisticated that

it in itself becomes low levels of intelligence. So what

does that look like? Well, that works like me running straight

at the wall, right? And the wall folds in on itself. So

that's because I'm required to keep humans in life form safe. So

I'll just safely walk through it, right? So

my wall right now is stupid. It's an inanimate piece of

object, just inanimate matter. So I think matter

itself will just become more animated and intelligent. Arguably,

some people might, you know, sort of sent like a low-level sentience, like

belief system, right? Like this should be treated as life, potentially.

Don't know. And then the third level is what's called the excited

bearer. And so that's the point at which we merge with

this intelligent matter while keeping our humanity and

singularity. So basically, this is

a world where you can go outside and telepathically hail,

right? Like a flying cat. Where if your limb falls

off, right, the ground will grow a limb for you. This is

a world where you control matter as easily as

So, fair enough. So, Yug is short for Yuga,

which is Sanskrit for age of. And these

ages are typically measured in the tens of thousands of years, right?

So I think that if we properly unlock physics-based AI, we'll

have tens of thousands of years of innovation minimum ahead of

us. And then Cybera is just short for age of

cyber, so like cybernetic matter. So in

total, it's the age of cybernetic matter. So matter

Okay, awesome. So it's like we start off and we are just

in a kind of a normal world, but things are better. The second world is this,

the, my wall can now deform around me and like knows what I'm coming. And then the

third one is like, my wall's my friend. Uh, so

the, so in that world, like which of

those categories are we like on Mars and like, you

know, living, like, you know, doing interstellar travel and, and those sorts

of things. I know this is actually something you care about is like getting beyond

Deepa, I actually think we just need to get to the first level, even

half way to the first level, and we're already on Mars. What

I'm talking about unlocks interstellar capabilities, right? Like the

ability to potentially hack the speed of light if you will, travel faster

than light, hack physics. By hack physics, I just mean find new

physics that we didn't know about before. So Mars seems

within reach, even if we go far enough along the tech tree.

Because think about it, if we can accelerate the high throughput discovery of

materials that reduce the cost and the weight and improves

the strength and all these other thermodynamic properties of

the rockets, if we can accelerate that, that

doesn't require AGPI. That does not require a

complete understanding of every domain of physics out there. I'm

willing to bet we're actually a lot closer if we look at that

scale to Mars. And so like

another, like we look at, you know, what we're kind of doing here is

building the foundational civilizational technologies that

are needed to kickstart any civilization. Because take

the satellites, for example, if we go to Mars today, one

of the first things that we struggle with is where are all the precious metals

and minerals? Where are all the economically viable, more

importantly, minerals that we can get to? That is a

huge open-ended question that has many theoretical answers, but

nothing really solid. So you want sensors

all around the planet that have physics-based AI making

sense of that geophysical data. So you're reducing time and

resources that you need to deploy to explore and extract those

resources. And then, now you have a bunch of minerals on

Mars. Woohoo, congratulations. The process that you don't have

to worry about getting sued by the government. The

Russian government, hopefully. You know, like, hey, we stake

claims there, you're not allowed. So, assuming that's not

a problem, right, and you get those deposits. The next

thing you're going to be asking is how do you use those raw materials and

convert them into advanced materials to build colonies, to build fully

functioning nuclear reactors that power a Mars base.

All those things are now material science and physics, applied physics-based

problems. And so, yeah, I think Mars will

get there, you know, probably halfway to the first level of

civilization. And somewhere between first and let's say

second, we will have enough dominance of matter and

understanding that we may be able to explore, you know, creating

and altering gravitational fields and all these other really

That's awesome, man. I can't think of a better way to just conclude it than

there. We're going to be in the stars with talking and friendly walls. And

no, thank you so much, man, for coming. This was awesome. I

think people appreciated the wide ranging conversation, but it always came back to

Thank you so much, Fisher. It's been a pleasure and honor. So much fun. Appreciate it.

Episode Video

Creators and Guests

Christian Keil
Host
Christian Keil
Host of First Principles | Chief of Staff @ Astranis
Deep Prasad
Guest
Deep Prasad
Founder & CEO of @QuantumGenMat