20 Years of R&D. Matched in 3 Months by 2 Friends. | Advaith Sridhar, Discovered Materials • DD15

it was coming. I think it was very good at copying my style, so my emoji usage and the length of my messages and things like that.

It wasn't that great at content. it didn't have, it wasn't really responding to the message the way a person would, but I think that's because of two things. One is that it was all small models, 8 million parameters, which aren't that smart

Mm.

anyway. And also it didn't have access to my calendar or know what I was doing at the time. It didn't have access to many important context pieces. So my friend would text me saying, Hey, are you free for dinner today? And it would not know. So.

Ha ha.

It would

just hallucinate something back, and which is when they'd catch it.

That you literally used as much insights and data possible and then you cloned yourself You just let the algorithm do its work and then see how of how many of them are gonna react in this way.

Yeah, yeah. So I did tell them that for the next day, it's either me or the LLM talking to you. I don't know which one you figured it out, but they didn't know who they were talking to. Yeah.

А, okay, that's nice

And so then I asked them after each of the messages, who do you think it was that was talking to you? So that's how that was set up.

Actually on hacker news that you shared and somebody's response made me really laugh. He was like, I responded my messages three days later. Would that do the same thing if it

Yeah,

that's a true test that is actually an AI. Does it have to take as long? My texting is also super sloppy. So three days is good for me usually.

your co-founder is at the same time your best friend, Akash. Did Akash also get exposed to your experiment?

He did, I think. Yeah, yeah, I'm pretty sure he did. So he was one of the people who didn't fall for it at all. And his texting habits are more sloppy than mine. So it was not very successful because I'd send him one, he'd reply a week later. I'd reply a week later. So didn't really fit in the timelines of it. Yeah,

Okay, the conversation doesn't continue.

we are very much like talk in person kind of people than like talk on text.

Mmm.

But is there anything that connected you with Akash? Akash for the context your co-founder and your best friend from your college years, right?

Well, yeah, so we grew up together in the same college. We lived in the same dorm as well. So I guess we've been connected for a long, long time. Since I moved to the Bay area two years ago, I've also been visiting him much more in person.

I also went for his PhD defense, which is where I learned a little bit about material science for the first time. And I didn't understand much, but it was still a great experience.

Just like being in the company of cool people like her.

But this is the moment that was it a calculated thing or was it a happy accident on the way when you guys were looking into the field?

I feel like it was a happy accident. I'd say I went for his PhD talk at the same time I was seeing my work getting automated by Claude. And so I went from me writing

a lot of my code to Claude writing all of my code almost overnight, sometime in November last year. And that made me realize that if you have these verifiable domains, then AI can really go after them and can get better

Yeah.

and better at them.

Same time went for his PhD talk, didn't understand much, but understood that that's also, material science is also a verifiable domain because you can

Mm.

come up with a guess for a material and know if it's a good guess or a bad guess based on the properties of that material. So I

Mm-hmm.

got excited about it then. He didn't believe me. Scientists are usually skeptics. so.

Yeah.

Yeah. And then one day his workflow got completely automated where he went from writing his fancy PhD code to having Claude write his code for him. And I think that's when it like a daily hit him as well that, my God, this is going to be a big thing. And so that's when we decided to start this.

so it's actually did you try to help him out and then you kinda get into the revital and then you kinda realize my gosh, there's so much potential actually we can leverage in this.

Yeah, in fact, one

of the first things we were trying was he had some unpublished work. So he'd spent a year chasing this track of finding materials for some domain, but it didn't go anywhere. And so he never published it. However, we asked Claude to do some other tasks and some other unrelated tasks, but then it came to some of the same conclusions that he came to after a year of doing his work at Stanford. And so...

That's when he was like, yeah,

I was like

this thing just did in three weeks what I took a year to do. And so it was a little freaky at points over there.

actually it's interesting to come across that on your own sphere, because when DeepMind GNoME when they published the paper, it was November 2023, right? If I'm not mistaken.

Mm-hmm.

Mm-hmm.

d- and then most of the things that we figure

out iron or glass magnets that happened all

Mmm.

sedendity in in the bottom line of it if you take a look at how they came up with it even the Teflon that our moms dads

Yeah.

cooking is a pure happy accident they figured that somebody changed the measures

then it started

Mm-hmm.

to push the limits, and all of a sudden they realized that hold on a minute, this is a very special element that we came up with. Let's take a look at it. And then in your case, with the genome, it also pressed 800 years of work into it with two to two million combinations, right? If I am not mistaken.

Yeah, yeah, GNoME was really nice because it was one of these first. So.

GNoME was a lot of computational material science work. And that's generally been very, very slow to do, to try and predict new materials on a computer and then simulate their properties.

Mm.

GNoME was really nice because it sped up that process by orders of magnitude by training these ML models for it. And yeah, they were able to discover, think they added 50,000 new materials to the materials project, which was like...

30 % of all the materials that existed in it before. they overnight the number of materials that humanity knew just jumped up by 30%, which is insane. It is, I'd say with GNoME there have been critiques of the worker as well, which say that many of these materials, they're not really sure if they're useful or not, but it's definitely a great first step and it's needed.

If you want to find the next superconductors or next silicon and so on, we need to be able to expand the top of the funnel first and then also increase

the number of these that we're able to actually make and then test and then deploy it to chips.

I'm right now having so many this trees which one should I pick and start conversation from because the it's a it's a vast thing and I'm really curious about your approach to solving this. Perhaps we can start with a little bit of a big picture, right?

you're talking about coming up with new materials, and that is gonna

Mm-hmm.

be way more strong. In your case, I think now the new name is discovered material.

Materials. Is it sealed?

Yeah, yeah,

As of last week, we have now discovered materials at dot com.

I honestly when you shared that, I kinda thought at the beginning power gram suggested that we go with discovered materials. But I was like, is this sarcastic? Is this real? Or but then it's

It is.

It is very on the nose, I will say. It's the most obvious name you can have. But PG

has a great quote on this where he's like,

as a startup, the goal is to be as impressive as you can and to seem larger than you are. And just having the domain name for discoveredmaterials.com is quite cool.

And so I feel like it's a little bit of doing that as well.

Yeah, it's on the nose, I think it describes what we do well. So I think it's a good fit for us. We've been through this discussion so many times as well. naming things is hard.

It is definitely hard. And then there's thousands of

Yeah.

alternatives out there. Whatever you try to come up with, there's always something somebody is using somewhere.

then discovered materials. if you take a look at nature, let's picture periodic table. Over there we have one hundred and eighteen elements. And out of one hundred eighteen elements, some of them are super rare, some of

them are super expensive, some of them are toxic, then okay, let's cut them out, right? Because it's gonna

Mm-hmm.

be more difficult to manage them as a use as a component. And then we

Mm-hmm.

have a safe space of 60 elements.

And then

Mm-hmm.

out of sixty elements that we try to it sounds very low, but actually it's not how many components do we need in order to come up with new materials like the magnets, the magnetic things, the glass, let's say. And overall, historically,

Mm-hmm.

it's between five and ten. So and then if you say, okay, how many combinations can I get that is gonna have five components to make a new material?

out of the sixty materials just trying to depict how big it is is

five and a half million combinations like whose lifetime is gonna be enough to do those combinations even if you go for five and imagine if you need more than five is it a good way of demonstrating

Yeah, I feel like even that

how how big

Yeah, I feel like even that is an underestimate. So definitely there's maybe 60 elements which you can play with. In fact, that number is probably higher because, for example, we use hafnium oxide in our chips and hafnium is by no means earth abundant, but the performance is there. And so people are still willing to use them in silicon chips. let's

say we have around 60. Now for

Mm.

chemical compositions of materials, you can pick any combination of elements in these 60. And so you have this cometary explosion there itself.

for chemical compositions.

Yes.

But

then even within chemical compositions, the structure of the material plays a huge, huge role in its properties. For example,

Yes.

diamond and graphite are both carbon, but have very, very different structures and therefore different properties. Graphite is like soft electrical conductor, and you can basically powder it in your hand, whereas diamond is a harder substance we know, and is an electrical insulator.

Mm.

So structure determines properties so much.

There's also...

For example,

The crystal

part, right? The crystalliz the the atoms and how they

Yeah.

combine. There are two elements that literally the same. just the positioning of the atoms are different than one another, similar to your example. One is totally something different and the other is totally different that you can't even use in adjacent projects.

Correct, correct. And it's not even just that. So this is at the nanoscale, where you look at the structure of the atoms in the crystal. But even at the microscale, the size of the grains that these crystals form together, all of that also plays a huge role in its properties. And then on top of all this is even all sorts of unintuitive things. For example, with graphene, there's something called magic angle graphene, which basically people found out that if you take

Uh-huh.

becomes a superconductor.

normally at any other angle, it's not a superconductor. But if you take two layers and then slightly place it at the right angle like this, suddenly graphene is a superconductor. And so there's all of these

Mm.

weird, unintuitive properties that materials have. And that's what just explodes the.

It's too large to count, the number of combinations and ways in which you can change these materials.

Yeah, because in indeed this is like the step one. Okay, which which elements do we combine? And then okay, step two, at what form at what structure do we combine them? And then once you even have a clue about how is it gonna perform, and then comes the all real life execution part of it, which is as difficult as the discovering the piece. in the essence of it, I find similarities with biotech, especially with the

Mm-hmm.

drug discovery, but I

I think this is a bit more difficult in terms of space and in terms of the

How say reward rewarding part? Because with the when you come up with a let's say a drug that you know how to make a new thing. And then in the end of the day, your patent really also keeps you for a while. but

Mm-hmm.

once the patent phase is over, okay, you still establish yourself in the market. And also drugs are very crucial for health. And people are open to

Mm-hmm.

take the risk, people are open to run, and also it's very rewarding financially. How do you feel about the

material science also maybe you should mention us what you are exactly targeting with discovering materials because the good part about I really like you and Akash are handling the very specific part of it not like the entire discoveries around any sort of materials.

Yeah, yeah, there are many similarities to drug discovery. As you said, both processes take huge amounts of money today and a huge number of years, 10 years and more than a billion dollars before a new material makes it to a fab or a new drug makes it to market. There are differences as well. For example, as you said, in drugs, most of the money comes from the formula of the drug itself and the patent around that. In materials, it's generally

we found this material, but the money comes from the process around making that material and then depositing it. So for example,

Mm.

most of the big companies in this space today make the tools that are used to deposit materials, or they make money by being the only ones who sell the material at a very high purity. So either I have this amazing material that I know how to make it, and I'm the only one who does.

Mm.

And

people like Tokyo Electron, Japanese companies,

Because

Ha ha

even though it's just silicon, there's so much that goes into the right process and pulling the silicon out of a vat of molten silicon and the temperature at which you do it. And all of that affects the properties of the material that come out. So the magic over here is in the process to make it rather than in the final

Mm.

material itself. And that's how it differs from drug discovery.

Actually I took some notes.

Lane then literally what I wrote here, once we figured out let's say an ideal material, but then the rest of the chain, correct me if I'm wrong, is like the following. You need to have the product reliability and then you need to have the manufacturing consistency, and then comes the long term stability, because it's one thing it's today, and then long term, and then qualification cycles, and then the shelf life, and then comes the in customer integration, because most of the time

It's not like one person figures it out, pass it to the other, and then okay, that's your end user. No, it's always like a chain of people building together on top of each other's work. And at the same time, the people that you are now targeting, yeah, tell us a little bit about that too, Or even one step back, what are we focusing on right now? Are we focusing on coming up with the AI infrastructure to help?

People discover new materials, or are you focusing on actually I will do the discovery part myself and then I will let you take care of what you need to do with it that all the manufacturing, all the supply chain parts is your problem?

Yeah.

how how do you position yourself now as the car?

Yeah, the

goal is to be the R &D lab that comes up with all this amazing IP on new materials and how they can get integrated into chips. And so the goal is not to then go scale the manufacturing for it. Building a factory is very different from doing research. And we want our core competency to be the research part of this. so in some ways, material sciences

the game industry or the music industry, you have these... You're going after hits, right? One of these materials will be a big hit and will make all the money back, whereas maybe, say, another 50 materials that you made weren't very useful.

Mm.

In that sense, you want to be the person who knows how to reliably make hits rather than a one-hit wonder. The goal is to focus on the...

Yeah.

infrastructure that lets you reliably make hits, which is focusing

Yeah.

on the AI components of this and betting that AI will get better and better rather than saying, we're going after graphene or going after say XYZ other material and making that the core of the company.

Yeah, good day.

is gonna evolve ten years from now on, hopefully when everything works miraculously well. But you still need to have your core and in this case that your core is just the most recent

wave in the material science. People are trying to figure out the software let's say that's gonna help them run their labs

Mm-hmm. Yeah.

mm-hmm.

And in that way, it is similar to drug discovery again. for example, Genentech started off by finding the IP for a bunch of diabetes medication. And then Eli Lilly was doing the manufacturing for them. And then eventually they built that manufacturing capability and brought it in-house. So perhaps that's one way in which we could take this as well. But initially, we are going to be just focused on the research and the IP.

help us understand,

how is it gonna be? what are the milestones that you are trying to achieve in the short term? And then maybe perhaps a little bit of not too long term, but more like a midterm. Because things change

Mm-hmm.

so fast. Probably what we are talking is gonna be half of it gonna be changing in the next months when you start launching and getting more active with your customers who are co-building with you.

Yeah. So the way all of AI works is that you choose a set of tasks that are super important, and then you benchmark the model ability to do that task well. And if the models are in greater date, you go and generate a bunch of evaluation data, RL environments, and so on to get the models to be better at those tasks. for example, what we are working on right now, and I think this will come out sometime soon, is

judging a model's ability to do computational material science, which is, hey, I

Mm.

want high thermal conductivity as a target metric. And then from there, seeing does the, can the model come up with good materials that fit this metric? This has been like the Holy grail of the field for a long time. It's called inverse design in the field where

Mm-hmm.

you start off with the properties you care about or the application you care about and then say, okay, can I make a

target material that meets those properties. So we've been benchmarking, Opus, Fabel, GPT 5.6, all of their ability

Uh-huh.

to do this task. I think it's been quite exciting. So we will

Mm.

have a bunch of results on this going forward, as well as ways to improve the models on these tasks. And that's the first stage, where you start off in computation.

be able to make the things that you come up with. And then once you make them, you have to make them reliably, show that they can be deposited in a chip, show that they can be etched from a chip, show that they can be used in the specific target applications with high purity, with repeatability, and so on. So all of these are benchmarks or tasks

Mm-hmm.

that need to be flagged posts that we need to hit. And we will build evaluations to see how good the models are at each of them, and then go and knock them down one by one.

one topic I normally don't go into competitors because I'm a full believer of like if you wanna solve something and if you have a religion approach that is unique to your background, your team composition, then you should not worry about it. However, you should learn from history. it's a recent

Mm-hmm.

history, but then you should have some learnings from what they do, what they tried, and what happened. When I take a look at this current scene around especially

Startups, millions are being poured into it. And actually, anyone who happened to be in AI field always actually got excited around AI for RD because this is the most

Mm-hmm.

thrilling part and the most difficult, meaningful problem at the same time. That's gonna needless to say, brings our lives in a more advanced level, right? And then

Mm-hmm.

when you take a look at this space and I see

See the ones like

periodic labs, Lila in Boston, or in Europe. I think there's only one that I found really doing I think it's Rara Factory. Yes, that's the one in Europe.

Mm-hmm.

And then the list continues, and I have a pretty good list. But then big players also wanted to tackle. Meta wanted to tackle, Google wanted to tackle, DeepMind, no need to say that making waves,

Mm-hmm.

but and they have a tremendous amount of experience. I'm not saying

I'm not saying internal

Yeah

intelligence. Do you know what I mean? So, how do you see yourself?

Mm-hmm.

In your case, Akasha and you are amazing payers because it one is knowing the material science and the other is all about agent existence. And then okay, voila, you can really complement each other in a very strong way. How do you see your positioning like in comparison to those

Mm-hmm.

giants and periodic laughs, like one of the founders?

co-created chpt and the other genome. do you know what I mean?

yeah, I I wonder how you wanna articulate this.

Yeah,

well, so a couple of things there. First is, as you said, I don't think at this stage of company, it is competition that kills the company. It's more about their ability to execute. Secondly, it's

great that there's so many people who are excited about this because it shows that there's something here. And if a lot of smart people think that, hey, there's going to be something in this space, then that's going to be great for everyone. Thirdly,

There's a lot of materials out there. at the 30,000 feet view, that's all AI for material science. That's like saying all of AI for biology is the same. But as you zoom in,

I know.

you realize there's many, many hard problems in this field that nobody has figured out how to do. For example, training AI on RL environments for the physical world. how do you get, what even does a lab loop look like where?

AI is doing experiments and getting feedback from the real world. that is something that no one has figured out. Famously, AI is very sample inefficient today. So you need millions and millions of examples before you can get something done in a lab. And more infamously, lab experiments are very not repeatable. So one of the biggest problems that a semiconductor fab like DSMC has is that just from batch to batch in the same machine with the same tool with

all of the ingredients that they've been working with for years, all the same materials, it's very hard to get the same consistency from one wafer to the next wafer. And even in the same

Mm.

day, half an hour after each other, because things just keep changing. And so all of these are super hard problems.

Have you ever seen

Fav's insight?

I was at CMU last week where they were maybe did a tour of their NanoFab. And it is so cool. the thing about semiconductors is that everything happens at the nano scale or the micro scale. And at these small, small dimensions, just the smallest thing can cause big variations in the quality of the samples that you make. for example, everything operates in a clean room. It's like a hundred times

Yeah.

cleaner than like the operating room of a hospital.

And you need to be in full PPE gear and so on, all because a skin cell might fall from your hand. And that would completely contaminate all of the samples that they work with. So there's many hard problems in the field. And many of these problems people don't have solutions for.

it doesn't matter that it's great that there are so many people in the field. But the mountain to climb is much larger than what people estimate.

Yeah.

the common thing that I I kinda see between you and the ones who have the strong experience and also tackling the same space like periodic labs, you guys are kinda really saying, if we figure out semiconductors, we are gonna figure out so many others too. So let's let's focus on this one. At the same time, maybe it can help us understand better why do we tackle semiconductors, right?

There is also all the reasoning behind this, not in terms of everything could be more efficient, more high power, but there is a blocker there that you want to actually tackle. Maybe then we can now go one step further on what exactly you want to tackle.

Yeah, yeah. Particularly in chips, the main problem that everyone faces today is the problem of heat. So if you look at the latest generation of NVIDIA GPUs, the Blackwells run at around 1,000 watts of, they call it TDP or thermal design power So they're giving out, it's crazy. It's like a one inch cross one inch chip. And this gives out the same amount of heat as a house in the US does. And it consumes the same amount of power as a house in the US.

The Rubin's coming out, I think at the end of this year, they gave out 2000 Watts. So it's doubling. And so it's two houses of power given out by the small chip. The Rubin Ultra

Mm.

is coming out next year. I've to be 4,000 Watts. So it's quadripling. the amount of, yeah, yeah. if you look at

Is it quadrupling really? I didn't I didn't know that. gosh.

any of these graphs of the power consumed by GPUs, it's going up like an exponential. Yeah.

Kinda like this.

And that's causing huge heat problems for them. None of the materials were ever designed keeping in mind that, we're going to run this at 4,000 watts. And so they all start

Mm.

melting, or they block the heat from escaping the chip, which then causes damage to the chip. So.

maybe taking a step back. The reason people are pumping up this power is because it makes the chips way more efficient. So it's interestingly, it's good for the planet for these chips to give out more heat because they give out four times the heat, but they have more than four times the benefit in terms of their efficiency.

overall, the equation plays out. But anyway, yeah, they're all ramping up their heat. This heat is causing huge problems and we need new materials.

to fix these heat problems.

I read your white paper and I had to read it again. the couple couple

Okay

of parts because I'm not a material science expert. But then there are I'm not going into the numbers, but there there are big promises in there. But at the same time, what I couldn't understand, are you

What workflow are you attacking? Because if you take a look at Laila in Boston, that they kinda have the whole loop and they simulate things physically in the physical environments. that requires also a certain level of capital, certain level of establishment. at the same time

Even in order to understand whether your verifiers are really good, they are not gamed, they are reflecting the reality the moment we take them out.

I I like trying to understand this maybe first human in the loop workflow, what exactly how does it look like? And at the same time you start with the TMIs. you need to remind me the thermal interface materials. I'm very bad with aberrations. Thermal interface materials.

the thermal interface materials. Yeah.

These are the like making sure that things are running in between the components smoothly, right?

Yeah, yeah, they're basically the thermal paste.

But if I

exactly if I aim for them, how are you gonna bring me back to semiconductors? And also, what is the workflow that you are looking into? If you can just combine these two to help us understand how your solution is gonna look, maybe take me as a customer.

there's many different questions

over here. Let

me walk through the thermal interface material part first, and then after that, we can talk about the Lila sciences and all of that.

Hmm.

Yeah, so in a semiconductor chip, there's 300 different materials that go into a chip. So it's not just one thing. There's materials at different layers. There's the part that TSMC makes, which is in the fab. As well as once it comes out of the fab, there's a whole packaging stack that happens.

where they integrate the wafer or the chip with the different components on the PCB. This packaging stack is what's actually undergoing the most innovation right now because a lot of heat problems can be solved at the packaging layer. Particularly

Mm.

these TIMs, they're basically thermal based. So if you ever built your own PC, you see the TIM is used in that. It's between the GPU and the cold plate of the system.

Anyway, so it turns out these stems are responsible for 40 % of the total heat resistance for a chip. So basically as heat is making its way out of a chip into the cold plate, 40 %

Mm-hmm.

of the total resistance it faces is from this thermal paste. So it was a great target for us for YC because it's easy to get set up. It's a big problem today. And we can really go build out this whole loop within three months.

Yeah.

With just these things, you could go after a real industrial problem with it. So yeah, it was great in that sense. The white paper is also all about our approach over there. And interestingly,

Mm.

what we found out was that, so, oh, by the way, so on the customer part. yeah, it was cool

because in three months, we were able to discover these terminal interface materials that are.

on par with what Dow, Honeywell and Henkel are selling today. Dow, Honeywell, Henkel, these guys are like the three big players in the market. They've been making this for over 20 years. They've probably spent millions of dollars of R &D on it. And

Mm.

me and Akash were able to do this in three months out of a garage lab in Oakland, which I thought was super cool.

Kudos for that.

Yeah, all of their materials that they use are also proprietary. they don't like, there aren't even patents on these materials because they don't want anyone to know.

what goes into making them. So I think it was super cool that we were able to do that.

The

one thing we learned while doing this is that I think the approach that's the consensus today is possibly incorrect. The current consensus approach is, we need these self-driving labs, and we need to do everything autonomously, have these closed synthesis loops

Mm.

where people are making materials. However,

None of the gains that we saw came from a self-driving lab. We did buy a robotic arm, and we had these experiments running where the arm was moving things around and mixing things and getting results. That wasn't very useful. What was useful for us was actually modifying the equipment each time.

Actually, this is one of the most surprising parts of your white paper that taking a look again from scratch, first principles, okay, what are the equipment, what are the processes other than the things that we can in a logical level improve? That I was I

Mm-hmm.

was actually quite surprised because I thought you were going heavy on, okay, we are gonna do the combinations much faster, more efficient way. let you continue.

I think that helps in

computation, where you do the combinations really fast and so on. But in experiment, all of our gains came from not automating a process, but constantly changing the process.

Mm.

So we switched from magnetic stirrers to mixers, we switched from mixers to furnaces, we switched from furnaces to blade coating, and they're all completely different processes. And each of them is what gave us the step jump in performance. I think this is also true.

Have occurred are those simulations?

Because you're not it's not that you're going out and then trying this physically, but then in the end they are also all sim running in simulations, right?

These simulations tell you what material to go after, but they don't tell you the synthesis recipe for that particular material. And so the synthesis

Mm.

recipe is what you perfect in the lab and what you need all these experiments for. So the simulation doesn't help you much on coming up with recipes. However, I think agents can get really good at that task, and so that's what we are pursuing right now.

Okay.

when it comes to benchmarks I'm highly skeptical. It always tell me your incentives and I'll tell you your where you're gonna lead me, kinda. one step back, how do we tackle it today? Today we have the machine learning super cheap, machine learning models super cheap at the same time, giving us results in shorter time.

Mm-hmm.

North Star for everybody to understand. But even there we are talking about

Months, not talking about weeks of development. And then there is this another end of the spectrum which takes really long time. These are the literally trying things in real life and then figuring things out. So and then you position yourself between the DFT and MLs, right? in the very sweet spot. is it correct way of depicting the big picture?

picture, but how is it gonna then do the cover the weakness of ML but then get closer to the DFT level of results?

OK, yeah, so there's two different things happening here. So you

do things in computation and in experiment.

Mm.

The computational work helps you expand the top of the funnel, which is the number of candidates that are interesting that might be useful for you. Whereas the experiment is the throughput of, you make these materials in real life? In computation, what you want is to be as fast as possible. It doesn't matter for you if you

don't find, let's say there are hundred new superconductors out there, room temperature superconductors. It doesn't matter if you miss 50 of them, all you need to do is find one and then that one is enough for you. So

Mm-hmm.

in combination, it's super important just keep screening at high throughput and to find some winners, which is why you can't do DFT on everything. DFT takes a really long time to converge. Order of magnitude

Mm.

days for each candidate.

And so what you want is to do fast high throughput screening using ML models. Turns out you can keep fine tuning these ML models with the FT data. And that's what the graph that we showed where

Yeah.

we picked a bunch of ML models out there to start with. And they realized they were bad and then fine tuned them and increased accuracy.

Mm.

That's what that graph was showing. Basically, increasing the number of candidates to try by having good fast verifiers.

So that's happening in computation. In experiment is a totally different, after you go through the funnel of

Mm.

trying to make these and that's like a totally different path. So not to be confused,

But what sort of evidence is

the simulations don't help you with experiment. Yeah.

Yeah, but what sort of evidences now do you have? Because it always comes down to the same problem. It's one thing to say, yes, I'm hopeful, I'm highly confident this is gonna materialize in a new material that is gonna help with XYZ, versus okay, then let's see how it is gonna work in real world. Because your success hinges on that, yet your majority of your work is on the part where you try to figure out what the next thing is gonna be.

B. How do you bridge the gap in between?

So this is why we have the real world lab as well, the lab in Oakland where we were making things. As you said, it's super important to ensure that whatever we come up with in computation, we are able to actually make and that it can exist in the world.

over there what sort of evidences you produce so far to support your improvements claims, let's put them this way.

right, right. the improvement, the thermal interface materials, the one that we have, which are on par with Tao, Honeywell, and so on. all of that is physically tested and measured in a lab with standardized equipment. And it's not a computational claim. in fact, we built our own characterization machine as well, where we actually simulated what a data center setup would look like, where we had a GPU, a cold plate on top of it, a liquid cooling system for it.

And we were making these materials and testing them in that setup. We also tested them on

Mm-hmm.

a third party setup in Livermore, where they have more of these machines. So all of the claims that we make are tested using third party equipment, as well as our own equipment on this stuff.

how about the part that when you test things on samples that they're out of your training data? does it also

Mm-hmm.

give you a good indicator? It's a very biased question, but I just want to lead you to digging into that part.

Yeah, so when you say outside training data, there's two things that could happen. There's false positives and false negatives.

Mm.

We don't care about the false negatives. So basically, as I said before, it doesn't matter if we miss finding 100 great materials. All we need to do is find one winner. And

Mine.

so all of our evaluations is very skewed towards

Hey, can we reduce the number of false positives that we find and we don't care about the other metrics? it's all about

Mm.

precision and not recall. Hopefully I got that correct. So

Hehehehe

that's how we measure as well.

So I'm not sure what you mean by outside the training data because the final measurements we do are, can this work in an application? And those measurements are what we test in the lab and care about. And in computation,

Mm.

the metric that we care about is only false positives and not the false negatives.

Mm-hmm.

how do you cover the data part overall? Do you say that that we have the biggest database, but then your comparison is a public database? I'm certain that there are already privately held databases in the same field, do you still find this as one of your strengths? what do you mean when you say we have the biggest data points available?

Right, right. So that particularly was the material. So basically we are collecting this thermal conductivity data for various materials and our comparison was against the materials project, which has a thermal

Mm-hmm.

conductivity database as well. And what's bigger than that claim is valid. There are a bunch

Hehehehe

of other research papers out there, which are coming up with thermal conductivity data, but not at the DFT fidelity, which the materials project has.

Interestingly, after we put out the white paper, the materials project itself released another 10,000 calculations or something like that. So

Mm.

that we had no longer the biggest one, but it is cool to see.

That is actually the

things how things go so fast, right? This is also what

Yeah.

I'm trying to say. But probably by the time we speak and then even by the time I put it out or conversation, we will have a couple of things that change already. this is I think this is the really exciting part of the times that we are living and building in.

Agree, agree. It's also important on where you collect data. the materials project is a nonprofit research initiative. Their goal is just to keep collecting and gathering more of this. The important data is one that improves your models, particularly

the false positives that I talked about before. And so all of the data we collect is targeted at reducing these false positives. But yeah, the more data that's out there in public domain, the better it is for everyone.

Got it, got it. did you ever have a moment like you felt you hit the wall? Because those problems we are smiling and talking about them, but they are incredibly hard.

help us understand what are you leveraging? How do you approach this? What is your system looking like? there are so many key decisions to make. also were there any moments that you

is not correct.

this is really hard wall stop. reassess continue.

Yeah, yeah. feel like particularly in this field, the material you go after really matters because each of these has a really long sales cycle. There's all this qualification that needs to happen. if you go after a domain where it's possible for people to reverse engineer what your material is and the process wasn't that hard, then you could lose all the advantage that you've built in finding it.

So the starting points really matter. Particularly, we did all this work on thermal interface materials. But I don't think it's the right path for us to go after, because two reasons. One is.

the market for these stims is around a billion dollars today.

But that's not that much no matter what material we go after, the probability of success is around the same and it's low. And so you want to ensure you maximize the other term, which is the size of the market that you're going after. so,

Mm.

which is why we were like, hey, we are making good progress on these stims, but maybe we should reangle ourselves to materials that people will care more about than they do about the packaging layer.

So we're actually pausing all our work

Mm.

on these TIMS right now, but still on this track of thermal materials. We still believe that thermals is going to be what drives the future of chips. However, we don't think TIMS is necessarily the best way to get there. And so we are readjusting ourselves right now on that.

so then actually my initial my previous question kinda made sense because I was trying to understand how are you gonna go get from the TIMs to semiconductors. I think it served you well.

Because then you got to demonstrate your capability, understanding, and how fast you can execute. Because literally you have milestones during the six specs, first six specs, and those those are not easy. you immediately could f put yourself on a comparison with the big heavy players on a chart where everyone would like to be in that sweet spot.

Yeah,

yeah, 100%. As I said before, with many things in semiconductors, you need to have the fab or some version of expensive instruments to get started.

That's not possible in three months. And so the packaging layer was a really great starting point for us to go prove a lot of our thoughts in this space.

Yeah, so what is next now? then what are the key learnings that you had you had? and then

okay. Yeah, yeah.

what is gonna be next version?

Yeah, the key learning was that one is, hey, whatever material you choose, the qualification cycles are super long. We've been talking to a bunch of companies in this space. They've all been around for five years or more without seeing significant revenue yet because they are still in qualification with the fabs. This is also

Mm.

the key problem we are going after, which is reducing this lab to fab time. I feel like in this process, one thing we've learned is that

many of the improvements that we can make to this process are pretty material agnostic. And so we are reorienting our approach right now to be focused more on, we bet on improving the models in a bunch of different spaces rather than choose a particular material and throw all our bets behind that?

you kinda actually reverse it now. you don't say I'm going after this certain material characteristics. So you say no, this is actually material agnostic. So I just need to figure out how to come up with the meaningful way of finding the new materials that would work in real life. Is it? Or

It's still going to be thermal materials. So in that sense, it's much more focused than, hey, we're going after any material possible. It's very much thermal materials for chips. But within thermal materials for chips, I don't think there is much of a benefit to saying, hey, let's go particularly after, aluminum nitride versus let's

Mm.

see how we can improve processes like deposition and etching and reduce the cost of barrier to entry for these processes.

as well as the speed of improving them. I think that is a much better play than going after say an aluminum nitride as a winner and then seeing where that goes. It's about balancing how much of a materials company you want to be versus an AI company. And so

Mm.

that has real consequences on where your value add comes from. our value add comes from using AI to speed up these processes.

more sense.

Okay, got it.

When you view something new, even a simple soft layer, that you always need a feedback loop. In the way that you performed during your batch, it was a very, condensed, intensive very fast-paced environment. But

Mm-hmm.

feedback loops did you create for yourself so that it helped you the most? because in situation of what you're building, it's not just you can throw

It over the fence to someone and tell me your feedback. maybe you can help us understand how the conversations in your batch helped you with your partner, when you talk to them. How did they give you the right pointers so that you could first see?

Yeah, yeah, with respect to feedback loops, it's actually...

easier than I expected. We have been in chat with many, many customers at say an Intel, say at a bunch of these chip manufacturers, say at the IBM which you spoke about before. So.

Mm-hmm.

people at these companies are all super willing to give their time and they're all excited about AI and the ways in which it can help them because as you said before, this process takes millions years as well as so much money. And so they're always excited to see how they can speed it up. I think our backgrounds also really helps over here where just having Akash with his PhD and from Stanford and so on, really helped open some of these doors. So I think we've, one of the things we've done well in the last three months is I'd say is.

listen to industry and where the industry is moving and then accordingly make our bets. And so I think that's been great. With respect to the our YC batch mates, I used to think that having people with a lot of semiconductor related knowledge is going to be the most helpful thing for us. Though

Ha ha.

I think my YC experience has changed that I feel like many of the challenges that startup founders go through are

fairly similar, as long as you're in hardware, for example, many

Mm.

of the challenges you all, you guys will face will all be around the same, especially when it comes to things like fundraising and so on. And in that sense, our partners have been invaluable in giving us feedback on this stuff.

when it comes to fundraising, what was the most difficult part for you? And this actually comes from someone who has spectacular angels backing up. That is actually very meaningful because you had a really, really good line of angel investors backing you.

Yeah, yeah, I was quite grateful that happened. I think the hardest part was just the number of meetings you have to do. YC makes it fundraising a lot easier because they cram all of fundraising into two weeks around demo day. But in those two weeks, it can be fairly exhausting.

Ha ha ha.

I've done my pitch easily more than 100 times now. And pitching 100 times in two weeks is not easy.

So how how many calls did you get per

It was a rough two weeks, but it got done.

day? were you very

It was.

strict with your timing?

I tried to be, but then a lot of investors are in India or in Europe or somewhere else in London. And so the time zones also gets to you. So it ends up being from 6 a.m. in the morning to 10 p.m. in the night. So it was a lot of calls over those two weeks.

But what what was the most surprising call you had? Did you ever have a moment like, is this a joke? do they really understand me or they are just having the call with me?

Hmm. No, I'd say like,

I take it from the horror stories and then we can talk about

a very good investor example.

Yeah. Yeah.

feel like, um, around that time, horror stories about VCs was also trending on Twitter. So I was also excited to see some, I was like, I have a hundred calls. someone's got to be weird in this, but actually I felt like it was all they were all actually super attentive. It's like every call felt like I was teaching something to somebody, but the student was actually being attentive and asking questions and so on. And so I feel like overall the process was.

Pretty good. Yeah.

I do know some weird stories from people in my badge. someone who took a call from their bathroom while they were, because they had to go to pee during the call. So.

no, and I

forgot to mute themselves.

Forgot to mute themselves

and the camera was accidentally on. Nevermind.

Yeah.

There are horror stories from the batch, but I personally, thankfully did not have any.

gosh, this is really different level. gosh.

Yeah.

I must say then congratulations in a way because you managed to attract the right minds to you to talk to.

How many people did you talk that you didn't need to educate on the call? I think that would be an interesting call. When they had done their homework so well, so you didn't need to go through this typical, this is the page, this is the this rather you go straight to the,

it may be two or three investors at Max had actually seen the deck before, and then gone through it and then asked interesting questions from it. So it definitely is rare. I feel like especially for semiconductors, even people who know semiconductors may not know thermal interface materials, for example, that is definitely a niche.

Mm.

I got for every call that was at some point where I was educating people on what there's definitely something new that they saw, which they don't see in their usual pitch decks.

But in in the end of the day, I think they don't necessarily have to be expert either, but in the end of the day they do invest in you and Akash. And did you also

Mm-hmm.

feel it with the what percent of the investors did you feel like yes, they try to see me before any commercialization prospects in the future?

Mm.

And the only slide she saw on the deck was like, how much are we raising? that was the only thing she wanted to see. I think that's actually

makes a ton of sense because so many things change in the early stages of a company, including our name. So

Mm.

I feel like it really, you're looking for the meta scale of can this person be good when things change rather than whatever they're showing you right now. Though that does matter a little bit. However, I'd say that was the minority of people. A lot of people had just a bunch of questions around the particular approach we had, the market itself.

A lot of things that I knew would change three weeks after I spoke to them about it. But

Hehehehehe

I still had to be like, yeah, yeah, this is definitely, this is written in stone and because otherwise they would be like, huh? And I feel like that.

You kinda

need to show some confidence, right, in your approach, but at the same time

Mm-hmm.

things are gonna change. I think anyone who had been in your shoes know that very very, very clearly.

those conversations, I feel a difference. If I don't see anything about the person when I go into the room, From the questions that they pose, I can tell whether that person has technical background or whether that person has ever built something in their life, or it's not everyone is Michael Moritz.

he was surrounded

Mm-hmm.

with the best of the best and he was journalist, he was questioning so so well and seeing the minds rather than this you build

Mm-hmm.

this, show me this. No, no, no, show me first how you think and then we can talk about the stuff we can build with the thinking. Or

Yeah, yeah, not true.

Yeah.

to explain it to them itself would take half an hour. And so it was, there were definitely all kinds of people.

I don't know how it feels for you, but the part that your vision

on how this industry is gonna go is as important as where you started like what you picked because nothing is in set in stone you just said it but at the same time you need to act as if this is today set in stone i need to figure out if this is gonna work or not and then move to the next one

Yeah, yeah, you always need

to have a plan, but the plans can change and you should be open to that.

If you didn't have a cash, would you do you think you would pick material science?

Hmm.

Good question. feel like six months back I would have said, no, no way. there's nothing that I know about this space. Now I am a little more comfortable with it, but it's still going to be incredibly hard without him. But maybe, yeah, I would say I'm before it would have been a hundred percent no. Now it's like an 80 % no. Perhaps at

Hm

some point it will flip.

to the other side, but definitely glad to have him. I would not want to do this without him.

What what would what would you do if you were not supposed to work on what you're working on? what would be the next thing that comes to your mind?

Hmm.

I feel like AI for science has always been exciting. And that's the reason why I'm in this space. It would probably still be in AI for science.

Mm.

More details of that, honestly, I haven't had too much time to think about things other than this. but

Yeah, yeah, yeah, yeah.

yeah, I would say still AI for science, just maybe a different aspect of it.

Yeah, yeah. Well I'm I'm gonna ask a very specific question though.

what sort of ecosystem are you creating for yourself so that you still stay sane? Because a lot is happening and a lot is sometimes going against the plan, most of the times

And what would you tell to someone who is listening they also want to pursue a problem that they set their eyes on? where should they start? and what sort of people should they put around themselves? Because, in my opinion, this is the most important thing. nothing, nothing else really matters other than the people around you.

Mm-hmm.

I'd say for me, two things have really, really worked. One is, Akash is awesome. And so definitely start things with people who you've known for a long time and are amazing.

Mm.

the second, I feel just being in the Bay area or in San Francisco has also helped quite a bit. there's just so many serendipitous interactions that happen with people working on the frontier. And so, it really does make a difference that.

Yeah.

a relief from the stress of running a company. And so I'd say underrated,

Ha ha ha.

but very important is to have someone who you can just go to regardless of how things are going.

Mm, like somebody who understands before you make a sentence, right? Like looks in your face

Yeah, yeah.

and then says, Okay, today

You

I I cover ya, I got you, don't worry.

Yeah, yeah,

super underrated, but very important.

one question actually, people in our community posed: your tech stack. Because I

Mm-hmm.

heard this in a

couple of times people happen to say but you never ask for your tech stack and I was always thinking but tech stack is something so personal how you like doing things what you value and so on but at the same time it gives you an understanding into what you find most valuable. Do you want a little bit of open up your tech stack for us.

Yeah, sure. I feel like it's one of these things that changes so fast these days because as cloud gets better at software engineering, the tech stack changes

Mm.

every two weeks. My most interesting amusing realization is that I don't open VS code that much anymore. All of my coding just happens directly from the cloud code desktop app, which is shocking

Mm-hmm.

for me.

That d when did this

really start happening?

it used to be like the Cloud Code on the terminal in VS Code, so I could also see the code. But then over the last maybe month, it's just, it's been only the Cloud Code app and nothing else. Like any, I have actively pushed away from doing anything locally to ensuring that Cloud can do everything it wants to on the Cloud. So

Mm-hmm.

it's mostly either my mobile phone or

Yeah.

me texting it on the desktop app.

That's how I do all of my coding right now. SuperBase has also been great. We use it for all of our database object storage, all of that. I feel like they've really capitalized

Mm-hmm.

well on the whole agentic coding thing. And there's a bunch of different libraries that we use for the material science computational side of it. Most of them written in Python, one written in Fortran, because the community, the best version of a

Okay.

lot of this code is still Fortran for some reason.

We have an active project now where Claude is rewriting the Fortran code base into Python.

I

but these are the moments you're so grateful you have claude.

Yeah,

yeah, definitely. but yeah, there's nothing too, special within the tech stack. I'd say it's, it's a lot of pick what the agents are good at and then just let them cook on the site.

didn't you take up on Sam Atman's offer?

Oh, we have been going back and forth on it. Now it's been, I think the deadline has passed. We have not taken it up. But I still don't think it's a great deal. And maybe they'll receive an email from us sometimes saying, hey, is that office still open? We're still undecided.

I but I what is f what is your hestation?

The main hesitation is around whether we will end up using 1.5 million of open AI credits specifically. In fact, our biggest

Mm.

compute costs right now is renting GPURs and not one of these model providers. So

Uh-huh.

if it was $1.5 million of GPURs, I would be all over it, $1.5 million of tokens, I'm not so sure.

regarding the companies that you're gonna be collaborating together, what sort of collaboration did you guys aim to do? And are you allowed to give us names?

Mm-hmm.

we know the process for depositing them. And you should use them in your chips because they solve a burning problem for you. We have an advisor as well who has successfully done this model in the past with materials that have made it to the fab processes of.

a HBM memory and chips and so on. So he guides us a lot on this, on the way this collaboration would work. The key here is that there's many problems that these chip manufacturers are facing, which affect their yield, which affect their performance, which they really need help with. And so if

Mm.

you can solve that for them, then the business models will come out of this.

I kinda enjoy conversation with you, but I'm also realizing the time is passing too. I would like you to reflect back on your batch and before

Mm-hmm.

you applied by YC what is the one thing from outside you thought you knew, but you totally changed your mind about.

The more you start getting into this and then you change your mind and then no, this is not how you your mental model should not be like this, no. Or any specific

Hmm.

thing.

I'd say so there's this code that first time founders worry about product and second time founders worry about distribution. I think

that has been more and more true for me through the batch. I've realized that.

we worry about the technical aspects of it and those are important, but what's way more important is ensuring that you're able to bring these out to the market and sell these to people, especially in the chip industry where there's only giant players and there's no real small players. And

so I think it's more important maybe to focus on the distribution than it is to focus on what our technical approach is going to be here. And that's definitely a realization

And what sort of distributions

for me.

then what was your remedy for that? What sort of distributions you guys are now trying to double down on?

Well, I feel like a lot of quote unquote sales is just like sales is simple, but not easy. have to put in a lot of work into reaching out to people and talking to them and building connections and relationships with these companies. But none of it is rocket science. You just have to make sure you spend the time doing it. and so basically the way the most impact immediate impact it's had is on our calendar where the amount of time we spend on.

reaching out and talking to people has gone up dramatically over these three months.

But I bet this is one of the things you didn't automate, am I correct?

Yeah, yeah, Claude is horrible at doing, especially

Ha ha.

it's the moment it writes something, it's written by Claude. So yeah,

as well as the kind of people it reaches out to. We did try, but no, this is best done. Bye.

Got it. is there anything that I haven't asked you but you would have loved to talk about?

Well, no, I feel like this was super comprehensive. So I honestly cannot think

N

of anything else.

this is this this is really really super insightful for me at

One thing that I'm seeing, you guys are very frugal. Do you think the

Mm-hmm.

next fundraising rounds target is

Ride them out, in comparison to the other fundraising rounds that people went for in the same space with you.

Yeah, there's definitely a giant amount of money being poured into the space and a lot of it will get burnt as well.

So there's two things here. The important question to ask, you need to scale something. And I think that is known. The important question is what needs to be scaled. And I don't think that question has been answered yet. And before you answer that question, it's not smart to go spend a bunch of money scaling things. So I think the answer really depends on what needs to be scaled and that is to be seen. Perhaps we will go

Mm.

raise a giant amount of money as well.

But I think there are ways to get around this without having to spend that much. And those are some of the key questions that we will be answering going forward.

Okay. Yeah, that makes sense actually. I always bring a question from the previous guests

the question I will ask you is from Elias.

Zettascale the founder of Z

From Zettascale? Yeah, yeah.

Exactly.

So what do you think the future looks like in 10 years? What do you want the future to look like in 10 years? Or are those two the same?

as in what I wanted to look like versus what it will look like in 10 years? OK.

Yes.

Well, I think what the future will look like in 10 years is,

so the reason we are focusing on chips is that there's two reasons. One is there's this whole recursive aspect to it where the models get better. Then they improve the chips that they were used to make them. And then because you have better compute, that makes the models better. And so there's this whole recursive.

cycle to material science for chips.

Mm.

And that gets me super excited.

today, the data center build out is so unsustainable. There's all these massive power requirements that generate so much heat and there's so much of water that gets used to operate these things. And if you can have super intelligence, but also in a more sustainable way, that would be so cool. So that's what really excites me.

And I hope it happens. The second thing is that

Improving compute is probably the most high leverage technology thing you can do because compute is used in so many different applications. it can be used

Mm.

to improve biology. It can be used to improve robotics. It can be used to improve self-driving cars. It can be used to improve AI, which goes and improves a bunch of other things. So in that way, making chips better is probably the best thing to work on. And so I hope that we make these chips better and that results in one, the recursive super intelligence as well as.

helps result in a bunch of technological change in other fields.

When you mentioned robotics, I just realized I didn't ask you about your opinion on world models. Are you guys also leveraging world models?

no, not in any... I feel like world models is early. It is super promising. But not in any way right now.

Mm-hmm. But also isn't it

one of the but isn't it also one of the really greatest plays to leverage?

I feel like people use the world model definition very loosely. there's a bunch of different things that people, all