Discovering New Materials With AI

Jonathan Godwin  is the cofounder and CEO of Orbital Materials, where they're using Generative AI to develop a pipeline of new materials for carbon removal and energy transition. He was previously a Senior Research Engineer at Google DeepMind.

Jonathan's favorite book: The Making of the Atomic Bomb (Author: Richard Rhodes)

(00:07) The Process of Discovering New Materials
(04:20) Building the Foundation Model
(06:42) The Impact of Google's GNoME Project
(08:49) Adding Materials to the Pipeline
(11:08) Computational Screening and AI
(13:43) Material Structures and Properties
(18:41) Material Formation and Degradation
(20:47) Materials for Carbon Removal and Energy Transition
(23:44) Ensuring Material Safety
(27:13) Exciting New Materials
(28:07) Breakthroughs in Material Science
(30:09) The Future of AI in Materials Discovery
(34:07) Rapid Fire Round

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.511)
Jonathan, thank you so much for joining me today.

Jonathan (00:05.518)
Thanks so much for having me.

Prateek Joshi (00:07.863)
Let's start with the process of discovering new materials. Can you explain how scientists do it today? What does the journey typically look like without the help of AI?

Jonathan (00:23.758)
Yeah. Well, so I don't think the process has changed really for well over a hundred years, we kind of use this Edisonian approach where, you know, a chemist or a material scientist will have an idea, you know, derived from chemical intuition, the sorts of science that we've known about for a long time. And then we'll try and make that thing in the lab and you'll get a result back that will inform your chemical intuition to go do another experiment.

And that's that kind of slow trial and error process that really relies upon either just being able to do lots and lots of experiments, or the genius of a few chemists. So it's totally different to the way that you might think about other engineering sciences. It's really extraordinary how much that hasn't changed.

Prateek Joshi (01:16.415)
That's interesting. So let's say a scientist is looking at it and there's a certain use case like, hey, we want a material to do X. And then they figure out, okay, there's a goal. And you know, the chemistry, there's trial and error, there's experimentation. So how can AI help? Or specifically what parts of this process can be used?

sped up with AI.

Jonathan (01:48.718)
Yeah. Well, so, you know, the reason we don't have computational tools commonly used in the discovery of new materials, sometimes they're kind of used to help explain in an explanatory way, what's going on, but in that discovery process, not used because the physics of what's going on is just so complicated. Um, it's too large. It's too complex to do a kind of bottom up physics simulation to predict the outcomes of an experiment.

Um, and so, you know, that the approaches that you would, would think of that have worked quite well in other engineering sciences, like designing an air foil or, you know, designing a bridge where you've got, you know, Newtonian physics, we can simulate that pretty well. We can simulate that high fidelity is just not true at the atomic scale. Cause you've got these quantum effects going on that make that really, really hard. But what AI is incredible at doing.

is being able to take into account and learn that complexity. You know, we've seen that in things like Alpha Fold, where this bottom up physics approach never worked. And then AI was able to calculate and compute that complexity that now labels us to design proteins. And the same is true in material science. So AI is taking all of that work that we've done and accounting for this additional complexity that allows us to design new advanced materials.

Prateek Joshi (03:16.595)
That's amazing. And I think it's a good stopping point to talk about orbital materials. You're the co-founder and CEO of the company. Can you explain what the company does?

Jonathan (03:30.99)
Yeah, so we develop a foundation model for material science, something that allows us to design new materials. And we go and develop our own pipeline of advanced materials for important applications, broadly speaking, within the energy transition. So you can think of these things like areas of batteries or carbon capture materials or materials that power green industrial processes.

Prateek Joshi (03:56.167)
you're building a foundation model, right? And you call it Linus. And as you said, it's tailored for material science. So can you explain how you built it? Meaning what data went into it and how did you train the model? And also what was the goal in mind as it regards to the end user?

Jonathan (04:20.269)
Yeah. So, you know, I use this phrase foundation model, um, in the way that you can find it from, uh, uh, the, the people who came up with that, that, that phrase. And that really means a large scale model that's trained on lots and lots of data. Um, that is then used for a bunch of downstream purposes. So, you know, large language model is a great example of a foundation model trained on all of the internet and be used for any sort of language.

And then you add in different modalities that might use slightly different architectures, things like diffusion models for image or video generation. So we really take inspiration from those models and we apply them really to material science. Now, where you might have images or text in material science, you have atomic configurations. So if you look at a material under a very, very powerful microscope.

The thing that gives a material its functional properties most of the time is the configuration of those atoms. And so we have a very, very large data set of those sorts of atomic configurations and for inorganic materials as well as organic molecules, because the really important thing often in material science is the interaction between organic molecules and inorganic. And we train a kind of structure property map.

we can generate for a given property we want to target a material that will have that property. That's kind of similar in some ways to the way you think about training an image generative model. You have a label that describes what the image is and you have the image and you train a machine learning model to take that description and generate the image. We're just doing something very similar and we use a lot of the same techniques but really for material science.

Prateek Joshi (06:11.519)
Right. And Google made a big announcement with their genome project back in November 2023. They said they discovered 2.2 million new crystals, including 380,000 stable materials that could power many of the technologies. So can you talk about what that means for this subsector? And also specifically, what does that mean for you?

as a startup doing this.

Jonathan (06:43.598)
Hmm. So, you know, I thought that was an extraordinary paper and I know that team worked with that team and they're incredible people. So I was really, really happy to see that out there. And I think that it goes to show a lot of the things when we were starting orbital materials, a lot of people would say, well, scale doesn't work in this context. You know, it works in language, it works in images. I don't think it's going to work in material science. And that paper was really, really great.

at showing people that scale is really important. So I was so happy that came out, and I think it's a great tailwind and a great example of how we can take the same ideas and same approaches that have worked elsewhere and apply them in material science. Now, I think that we're doing two slightly different things though, to what the Genome Project is doing. That project was really focused on finding stable crystals, stable materials. That's great, but it didn't say anything about the functional properties.

So you've still got to figure out whether that's going to be a good material for what you want to use it for. And that's incredibly important for us to developing a pipeline of materials for specific important applications. You know, that, that project really didn't, didn't say anything about that. And that's, I'm sure not a criticism. They've done an extraordinary work, but, but for us as a startup.

we're really focused at making materials that matter. And so we've got to add in different types of data. We've got to do this generative part where we start with the functional property and then generate material. And the genome project did not kind of look at those sorts of machine learning problems that we're kind of more familiar with from generative models and foundation models.

Prateek Joshi (08:21.855)
Right, that's an important distinction. And you've mentioned your pipeline of new materials, and I wanna dig into that a little bit. So how does a material get added to your pipeline? What needs to be true for that to happen? Is it driven by customer demand? Is it driven by just science? Or is it a combination of both?

Jonathan (08:49.358)
Yeah. So I think, you know, you can think of what we try and do is similar to other companies and kind of hard tech areas that develop a pipeline of, of, of hard tech products. You know, we have an extraordinary competency in AI. I think our focus has been right at the beginning, materials applicable for the energy transition. And that actually ends up sort of forming a subset of the materials, you know, and that's where we have built up our own wet lab in New Jersey.

where we've got really some of the world's leading scientists in these areas making these materials. And so there is a kind of economic focus about where we think the large economic markets are for new advanced materials, where there are growing markets and big demand and where new materials can unlock, you know, really new incredible products.

And then on the AI side, I think there are kind of areas where AI is better right now and areas where AI is going to be better in five years time. And so we've deliberately tried to target ourselves in areas where we're really confident that AI is production ready right now, and we can make really significant progress and accelerate that development cycle. So it's at this intersection. And I think that's going to be the case. You know, if you look at enzyme design companies, enzyme design companies working in a very similar sort of area to us, you know,

I imagine they go through a similar sort of techno economic analysis process and then try and find partners that can help them develop that new pipeline.

Prateek Joshi (10:23.119)
Yeah, that's a fair point. And also, there's a lot of parallels to drug discovery companies. It's a combination of what you can do well, what's doable, and also what does the customer want. So it's a good combination, it's a healthy combination, because as a business, obviously, we're building great products, but they're to stay viable, like financially viable, which means that you have to build and deliver products that the market...

needs. That's a fair point. So when it comes to materials discovery, computational screening plays a key role. So can you explain what that is? And also, how can AI help here?

Jonathan (11:08.078)
Yeah. So computational screening, I think this is used a lot in when you think about AI and drug discovery as well. The kind of idea here is that you've got lots and lots of different candidate molecules or materials. So in the case of a molecule, any new configuration of atom creates a new molecule, any new kind of crystal structure creates a new material.

And so you kind of get this, people call it this combinatorial explosion where the search space of materials is so large that there's no possible way that you can go and make all of them and see which one is better. So the idea that people have had the kind of theory of change in a lot of AI companies has been that we can evaluate these materials super quickly using AI. We can have a structure property map and I can find the structure with the optimal property.

Um, for my, uh, my application or in the case of kind of drugs, a molecule that kind of has the best binding energy to a protein, that's kind of how screening works. Um, I think that, you know, screening is probably the first generation of how. You know, atom atomistic design works with AI and design and materials works. I think generative models are the next step.

Because in this screening approach, you look through millions of materials and you just pick the needle in the haystack. And that's very different from a generative approach where you say, well, I want this functional property. And then the AI designs a candidate that it thinks has that property. Now it's kind of similar in analogy to thinking about an image, like the screening approach is saying, well, I'm going to randomly select pixels in an image until I have something that looks like a cat. It's obviously going to take a really long time. And that's really, really hard.

The generative approach is to say, generate me an image of a cat and a cat comes out. And that's what we've seen with images. And I would, I would expect that when we're talking about AI and in design of materials and a couple of years time, everyone is going to be primarily around generative design and you know, that's something that we care very deeply about. And there's been a hypothesis from the very beginning.

Prateek Joshi (13:15.063)
Right. That's actually a great piece of the puzzle here is doing that. Basically, hey, I need this and you just generate it. Like, don't go through the whole entire space. Just generate this thing. And you mentioned material structures and properties a couple of times. And maybe we can go a step deeper.

When you look at a new material, what properties are even relevant in the first case? Meaning if you are an AI product or a scientist, like what properties do you look for and what's good and what's bad here?

Jonathan (14:02.094)
Yeah, that's a really good question. So obviously the properties are going to depend upon the application that you care about, but let's say that you're trying to do something, which I think is extraordinary, which is to go and produce aviation fuel that is sustainable aviation fuel, so a catalytic chemistry process. And you've got a really advanced material in there that speeds up and makes economic economically possible that process. I think, you know, the history of human development over the past 100 years.

has almost been the story of discovering these materials. The way that we feed the world is by discovering a way to produce ammonia and at the core of that is catalysis. So I'm gonna use that as an example, but you know, there are gonna be other examples when you think about energy generation, so solar panels and semiconductors, yeah, other examples if you're thinking about batteries and energy storage. If we just take that example of a catalyst,

What you're really looking for is a material that's going to be really highly stable, because it's probably going to be under really intense conditions, so you really need to have a sense of the stability of that material. It needs to have this kind of stickiness factor. It needs to be sticky for the right molecules with the right level of stickiness. So it just holds a molecule perfectly in place for another molecule to come and react with it.

And that's a kind of incredibly difficult thing to go and calculate. But that kind of Goldilocks zone of perfect stickiness is the thing that you want for that catalyst. And then you want it to be kind of easy to make. So you want it to be made out of elements that are abundant in earth. Because if this is gonna be a really scalable chemical process, you don't want your catalyst to be the most expensive and most rare.

element on the Earth's crust. But that actually, you know, sometimes kind of happens. That's one of the big problems that we have with hydrogen at the moment. So that's an example of the sorts of things you look for with catalysts. And then it needs to have a large surface area. So it needs to have a lot of these areas where...

Prateek Joshi (15:52.999)
Right, right, right.

Jonathan (16:08.494)
molecules can go and sit and wait for other molecules to react to them. So surface area is another, another important thing. So all of these kinds of multi-dimensional optimization problems, which make the whole area so fascinating, so difficult, you know, has a lot of analogies, as you say, in drug discovery.

Prateek Joshi (16:25.831)
Right. And another key element in material science here is material formation and degradation. And you need to understand that when you build and develop new materials before it gets delivered and like in production, you need to understand that. So how do you understand and predict the

Prateek Joshi (16:54.847)
Can AI help? If so, what can it do to make it faster, better, cheaper?

Jonathan (17:00.654)
Yeah, so, you know, I'm optimistic that many of these things are going to be solved by AI in the near future. Now our company is starting with a few things and is growing that capacity. So I think one thing that we're really interested in are materials where you can kind of design in a, design that synthesis and formation route. So certain materials, you know, are designed and synthesized with the use of these things called template molecules.

Um, and template molecules are kind of like little molds that offer a kind of structure for a specific geometric material to form around. And so that, you know, molecule design, which is kind of very similar to, to, to sort of designing small molecules that bind to proteins.

That molecule design is something that's kind of tractable with AI. So we're kind of interested in materials classes, I think, where those sorts of approaches are attackable because that gives us then a really strong kind of synthesis pathway to go in making the designs that we generate.

Prateek Joshi (18:09.663)
Right. Another interesting topic here is smart materials. And for people who may not know, these materials are designed to have properties that can be changed in a controlled fashion, like temperature, stress, moisture, and many more metrics. Can you explain how these materials are manufactured? And also, where are they used in the real world?

Jonathan (18:41.294)
Yeah, I think that's a really good question. So I think, you know, one of the things that you kind of figure out when you go into material science is that there are so many ways to manufacture different types of materials. You know, so if you were to take a kind of thermoelectric material, so that is a material that generates energy.

when it encounters heat. So that's, I mean, that's a really exciting area of material science, because you can imagine cladding a building with thermoelectric materials and then the excess heat from heating in winter provides additional energy. So really, really exciting categories of materials. Making those materials is gonna be very, very different from creating like a smart material that might go into your body as a drug delivery process that is gonna change its properties depending upon the pH level.

or even depending upon light being shined on it. That's a really exciting, I guess you might call it drug area, but it's also a materials area of photosensitive materials. You get drugs accumulating inactively in a site, and then you shine light on it, and then they start becoming active and attack that cancer very, very locally. The synthesis methods for both of those types of materials, you can imagine, are very, very different. And so.

Uh, you know, I think that's one of the challenges you have when you're thinking about AI for material science and why it's so important to really embed, um, experimental wet lab work into the development of your pipeline. Um, happy to take, yeah, I don't know whether that answers your question, whether you had sort of follow up thoughts on that.

Prateek Joshi (20:17.831)
No, I think that's great. And at Orbital Materials, you have mentioned that you want to discover new materials for carbon removal and energy transition. Now, I have a two-part question here. One, what type of materials can help with that? And two, who is going to consume that product? Is it clothes manufacturer? Is it...

Jonathan (20:19.822)
Yeah.

Prateek Joshi (20:47.023)
Is it actually every industry? So basically two part question of what type of materials and also who is it for.

Jonathan (20:54.126)
Yeah. You know, so when you think about the energy transition, I think that you have, I think broadly speaking, sort of three areas that are really, really, really exciting, um, and you think about areas of kind of what we call gas separations of areas like carbon capture, methane capture, and environmental remediation. So that would also kind of count things like removing harmful chemicals from our water, water desalination.

extracting renewable natural gas from landfill sites. Those sorts of applications are really exciting. And I think the final area within separations that gets me really excited is in mining and resource extraction. So a lot of the rare earth metals that we really, really need and the rare elements that go into batteries or go into nuclear fishing plants like uranium,

They're very hard to mine and more efficient ways of extracting some of those elements from places like seawater or brines are really, really important ways for us to improve the resource efficiency and reduce costs of the energy transition. So I think areas of like those sort of separation applications are really, really exciting. The other area that I can kind of spoke about is catalysis. So new chemical processes. You know, I think that the, you know,

Almost all of our chemical processes at the moment are extremely polluting and some of them are way too expensive. We need a way to produce way more ammonia for fertilizer at a cheaper price in a sustainable way. So that really requires us to have a new chemical process and new types of chemical companies. It turns out that those sorts of materials that you think about

are very kind of similar. Like the sorts of materials you have for kind of separations and catalytic processes, you know, they're very, very similar. And even when you get into areas like energy storage, some of those materials actually have kind of similar properties. You might think of them as kind of porous materials. Those sorts of materials are the area that I think we're really, really excited about because they have such a breadth of application in those energy transition.

Jonathan (23:18.062)
in new economies and new market demands. So Carol.

Prateek Joshi (23:25.948)
Yeah, that's actually great. And when you develop a new material, how do you ensure the safety, or rather how do you stress test it to make sure that when it's deployed in the real world, I mean, in extreme case, it's not gonna blow up. So how do you ensure that?

Jonathan (23:44.942)
Yeah. So I think that one of the things that I think we do is we have AI and then we have a bunch of, you know, more traditional parts of our company that do some of that testing. So some of the toxicity things you can predict with AI, but a good chemist is going to know that they're toxic anyway. And then we have kind of rigorous testing procedures that allow us to check that that is safe.

Also, the applications that we're really interested in are not things like putting, I think maybe something that comes to mind is PFAS chemicals, these are known as forever chemicals. We thought for decades that they were totally inert and totally harmless. I think people even suggested that they could be used as diet pills because they weren't going to react with anything in your stomach.

but we're going to fill you up. It turns out that actually these things, surprise, surprise, are extremely harmful for human health. And now we have them coated on our saucepans and they're in our wastewater, they're in all sorts of challenging areas. We wouldn't expect to be, you know, our target is not to create materials that have that sort of usage. We wouldn't create, you know, coatings or areas or things like that. We're really kind of focused, as I said, in

in those kind of application areas. And it's easier to control and check the safety in those applications.

Prateek Joshi (25:17.299)
Right. And quickly for people who may not know, can you explain what PFAS is?

Jonathan (25:24.814)
Yeah, so PFAS are a group of chemicals, of molecules, that are used in a huge number of industrial processes. The thing that makes PFAS really applicable for industrial processes is that they're really inert. So they actually are on Teflon. They're part of the coatings that go on your waterproof coat, that's the reason that it's waterproof. They go on...

non-stick coatings on your on your saucepans or at least they're used to. And they're also used in areas like semiconductor manufacturing really extensively. So they've got these kind of remarkable properties. It turns out though that so they're kind of in everything and they've been used for decades. We have realized that they're extremely extremely harmful for

human health for a bunch of animal health as well. But they are now everywhere. They're in the water that you drink. They've been found in Antarctica. They're at the bottom of the oceans because they have just been so prevalent in so many aspects of our daily lives. And so it's a really important thing that governments around the world are investing billions of dollars, companies are investing billions of dollars in removing PFAS.

from their industrial processes and then cleaning it up from our environment.

Prateek Joshi (26:56.223)
Right. And there's so many new materials keep getting discovered across industries and use cases. Very exciting. So what has caught your attention? Like what new material has caught your attention and why?

Jonathan (27:13.582)
Yeah, so I mean, I think we think, of course, as a scientist, the idea of, you know, outside the areas of the energy transition, which I think are incredibly exciting, there's always this promise of room temperature superconductors. And so I got so excited last year, you know, there have been a number of really incredible papers, as well as the buzz last year with that sort of archive dump.

Prateek Joshi (27:32.521)
Right.

Jonathan (27:42.254)
But every single time that comes out, it always strikes some excitement because the belief is there. I think this is the thing that I realized last year. There's still a strong belief that we can find a room temperature superconductor that might be manufacturable and might enable all sorts of incredible new devices. Yeah.

Prateek Joshi (27:59.135)
That's great. Same here. Every time I see that, it's really exciting and it has certain allure to it, certain charm. It can't seem to get away from it. Right, right. Okay. Moving on to the science part of it, what breakthroughs in material science are you most excited about?

Jonathan (28:07.31)
Yeah. Yeah. It's a heresy grail. Yeah.

Jonathan (28:23.182)
Yeah, so I mean, I think within material science, the thing that I care huge about is the energy transition. I think that is an extraordinary thing that I get very excited about. And so we have a bunch of important discoveries to make in that area. Grid scale energy storage.

because we're producing an extraordinary amount of renewable energy, we need a way to go and store that. So grid-stellet energy storage materials surrounding that. New chemical processes, new ways of capturing harmful greenhouse gases. All of these things are extraordinarily important. We need better breakthroughs for all of those applications. And then outside, I think, of those sort of areas,

I'm really excited about materials for quantum computers. My background is primarily in computer science and machine learning. So I've always felt maybe I should be working on quantum instead of AI. I caught the AI buzz and I'm not regretting it. But this idea of really incredibly powerful quantum computers is so exciting. And at the core of that are...

really important material science problems. You know, materials unlocked the computers we use today. We would not have, we wouldn't be having this podcast if it weren't for huge breakthroughs in semiconductor manufacturing and discovery of extraordinary semi-conducting materials. And I think the same thing is really true when you think about quantum. So I'm watching that really closely. I think it's incredibly exciting.

Prateek Joshi (30:09.651)
Right. And you made a very interesting point about grid scale energy. Basically, a lot of renewable energy is being produced, but if we can find a way to store it without being glossy and hopefully ship it to long distances, I think that's going to unlock a whole bunch of new things. So in an extreme, an ideal case, will it be so that we can tell the AI product that,

Prateek Joshi (30:40.211)
for a long time, energy for a long time, can it be that way? Or how specific do you have to be when you specify a goal to an AI product?

Jonathan (30:50.126)
Yeah. So I think that the right way to think about it is that you're going to have to be pretty specific anyway, right? Because the sorts of requirements that come through on the grid are going to provide a lot of constraints on the sorts of things that you could use for grid scale energy storage. So I think there's going to be that level of specification that doesn't really come from

having to fine tune an AI, but just comes from the understanding of the requirements of that application. And so always these things come down to how well do you really understand the goal that you're trying to achieve. But I would believe, and I have a strong belief, and I'm hopeful that we get to talk some more about our pipeline later this year, that we can really start to do this sort of multi-dimensional optimization.

of materials in natural language to give really, really strong candidates that really follow through into something you can go and make, including things like manufacturability, including things like stability, like lifetime, that's incredibly important, like properties that change with temperature, all of these things.

I think are really on the horizon for materials. And I've seen extraordinary progress over the past few years. The reason I started Orbital Materials is because of that rate of progress. And I have not seen it die down at all. If anything, it's accelerating far faster than I could have predicted. And so I don't think that is an absurd goal for us to be reaching towards.

Prateek Joshi (32:32.999)
Amazing. I have one last question before we go to the rapid fire round. And it's about the immediate future of this of the subsector. So when it comes to using AI for materials discovery, what's coming next in this field, both from within orbital materials, but also from other companies who may be working on this on this problem.

Jonathan (32:59.598)
Yeah, so I think, I think, you know, material science, new synthetic routes, that go and move into the, um, uh, the wet lab are incredibly important. So I would expect not just on this design part, but also in the synthesis methods, the synthesis design, um, the testing rigs and the process design.

that surrounds everything that you think about material science, those things are going to start being attacked and start being attacked really, really successfully. So, you know, we're really starting with this generative model of materials. I think that, you know, you're going to see that stretch into all parts of the lab and the process of actually going synthesizing them to the point where you end up with something fully automated. So I'd imagine that materials discovery process to be to have a robotic lab.

and an AI brain that is controlling that robotic lab and doing automated discovery of incredibly important and commercially valuable advanced materials.

Prateek Joshi (34:07.047)
Amazing. With that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready?

Jonathan (34:17.23)
Go for it.

Prateek Joshi (34:18.635)
Question number one, what's your favorite book?

Jonathan (34:23.022)
So I really like, it's called The Making of the Atomic Bomb. It was written before I was born by Richard Rhodes. It has an extraordinarily comprehensive description about how the atomic bomb was made and introduced me to some of the, I think, extraordinary contributions that science has made in the 20th century.

Prateek Joshi (34:46.795)
Amazing. What has been an important but overlooked AI trend in the last 12 months?

Jonathan (34:54.958)
So, of course, I'd say apart from AI for material science, I think AI for the hard sciences has been underplayed. We've got some extraordinary stuff in language, in videos, in images, but long context windows and all those advances are gonna have huge impact in areas like genomics. And we're starting to see the fruits of that. I think that should be a bigger story.

over the next 12 months and underappreciate this story from the previous 12.

Prateek Joshi (35:26.211)
Yeah, I agree with that one. What's the one thing about material science that most people don't get?

Jonathan (35:35.15)
I think people don't get the scale. You know, everything we use, almost everything that we derive in our daily lives is from material science. So, you know, the way that we feed the world is materials, the, uh, you know, the chips on our computer and materials, the screens, these are all inorganic materials, Silicon Valley was built on semiconductor companies, um,

And so the scale and the economic opportunity and the demand for new advanced materials is something that is kind of passed by in the consciousnesses of most people. Despite the fact that we rely on it every day and it is perhaps the most important sector. I think we're understanding that more when we look at the chip wars going on. And that just how extraordinarily difficult it is to go and make Semiconductors and build semi fabs.

Prateek Joshi (36:27.347)
Right. What separates great AI products from the merely good ones?

Jonathan (36:34.126)
So I was thinking about this question and I don't think that there is a distinction, right? Like I think, sorry, I don't think there is a distinction between AI products and normal software products. The differentiators between those two ultimately come down to how well they serve their customer. And I think that the good and the great AI products ultimately are derived from customer benefits. So I don't see that there is something unique about AI in that respect.

I think the teams look different, the pipe you hire, you're different, the kind of motes that you have are different, but at the end of the day, I think these are all, all come down to the same, same sort of thing.

Prateek Joshi (37:14.313)
Right. What have you changed your mind on recently?

Jonathan (37:20.846)
You know, I think the...

I was not anticipating how quickly the context length of large language models can increase. This is a kind of slightly nerdy point. And that's changed my mind upon, I think how quickly we're going to get to systems that can really reason over large amounts of complex data. And that opens up a huge amount of applications that I thought were perhaps, you know, nine to 18 months away because RAP pipelines weren't so good. And it's brought this forward.

So I think that's super, I mean, that's another thing that I think is super underappreciated about that sort of breakthrough that we see with, you know, ring attention and this million context length from Gemini.

Prateek Joshi (38:07.335)
Right, what's your wildest AI prediction for the next 12 months?

Jonathan (38:13.358)
So I think on that note, like I can imagine that it's really plausible that we have a full length feature animation generated from AI from a single prompt within the next 12 months. I think that it may not be something that's easily served as part of a commercially available LLM, but I can imagine one of the big AI labs having a really, really long context link they're using research and showing that off.

Prateek Joshi (38:43.103)
Final question, what's your number one advice to founders who are starting out today?

Jonathan (38:51.214)
I would start company, start a company where I think this is something that kind of Sam Altman said and it's something that, you know, was part of our decision-making and going and building a wet lab, which is to do something where if AI goes as fast as it has over the past two years, over the next five years, that you think is still going to reap those tailwinds but provide you with defensibility. I think some of the

some of the things that are going to be short-term, incredibly impactful, are not going to last a long time. So be careful about that would be my advice.

Prateek Joshi (39:28.691)
Jonathan, this has been such a wonderful discussion. It's amazing how spectacularly huge the world of materials is. The entire world runs on it. And I'm glad that AI is finally entering and not just entering, but like entering with force and it's bringing it to life. So really exciting what you're doing and thank you so much for coming onto the show and sharing your insights.

Jonathan (39:53.838)
It's a pleasure, thank you so much.

Prateek Joshi (39:56.141)
Good one.