E567 How Frontier AI Is Accelerating Materials Discovery with 83 Sciences' Ian Naccarella

0:00The oldest podcast is brought to you by all this international, supplying your expert AI and digital transformation staffing needs across the US and Europe. Today, you are listening to our AI in Action series. We're leading minds in AI from across the world share their story, success, and advice. AI in Action cuts through the hype and explores the true impact of artificial intelligence in our world today. You're listening to the oldest podcast, AI in Action, I'm your host, JP Valentine. Our guest today is Ian Nakarella. Ian is the co-founder and CEO at 83 Sciences. Ian, welcome to the show. Thanks so much for having me. I'm excited to be here. We're excited to have you. So Ian, let's start with yourself, please, as we do all of our guests.

1:01Could you give us a little bit of an overview of your background in technology from where you got started, some of the roles you've held along the way? Most importantly, what's led you to where you are today is co-founder and CEO at 83 Sciences? Yeah, absolutely. I think come about it from the angle of engineering mixed with the business side of things. I think my career has really straddled both of those, probably more along the line jumping back and forth than most other folks. Global engineering did both my bachelors and masters at Stanford. Thought maybe I'd want to go get a PhD. At one point during my honors research, I realized that I was maybe not as competent in the lab as though my other friends. So, after graduating, knew I wanted a number of different experiences, went to the Boston Consulting Group in San Francisco, was really mostly in the climate and sustainability practice area during that time. Had the chance to work on a lot of really cool projects in the climate space. I ended up going to a battery startup, a pretty late stage called SEALA Nanotechnology is out in the Bay Area, where my role was really working closely with the R&D teams to

2:03figure out how to bring the second product to market. So SEALA Series F startup when I was there, but they wanted to be more of a technology platform than a single product company. So I spent a lot of time with the R&D teams figuring out how do you bring that second product to market? Do that I wanted to go a little bit more early stage, didn't really have a sense of how to do that, ended up going to business school at Harvard, which is what brought me into the Boston ecosystem. Spent a lot of time while I was there doing this fellowship at MIT, those really geared around helping professors commercialize their new technologies in the climate tech space. As I was doing that fellowship, ended up meeting Eric R.C.T.O who had just finished up his PhD at MIT. He done a lot of really interesting stuff in the AI for materials characterization, AI for materials discovery space. We got to talking, worked on a bunch of different projects, first from room temperature superconducting to clothes that cool you as you're being active, working with a bunch of different professors,

3:05landed on this AI for materials discovery piece which brought us full circle back to essentially Eric's PhD. We started working on that pretty quickly looped in Yang Kang who was one of my former colleagues from BCG. He was essentially the node that was running their AI implementation practice and really has a very good sense of how do you get folks actually use these AI tools. That's the core of the team and then applied to Y Combinator and A16Z speedrun, ended up electing to go with Y Combinator and work two thirds of the way through that batch now. That's exciting. So, the next round then helps paint the picture of who's involved on the foundation level which is always critical and congrats on the Y Combinator. Let's set the stage. Tell us all about 83 sciences who you are, what you do, mission of the business. The mission of the business is can we bring these frontier AI models into the real world. So what actually inspired us is Google, Microsoft, Meta, they've been putting out these models

4:06for a couple of years at this point that are can we do materials discovery? Think of it like a search engine for a new material. You say I want a material with these properties and they run around and they spin up something and they do a bunch of modeling and they say, hey, here are some crystal structures that have the properties that you're looking for. Big deal got a nature paper out of it, ton of traction there. However, a bunch of labs then went out and tried to synthesize some of these crystals and realized that there wasn't a single material that was proposed, that was both unique and synthesizable. And so if you follow the academic literature, there's tons of back and forth here, there's some retractions. It's actually pretty entertaining to read. The core pain point there though is that none of these models are trained off of real world experimental data that contains the failures. And when you're training models only on successful outcomes, you end up with this incredibly optimistic view of whether or not a material can actually be synthesized. So we're working with academia to get the full set of experimental data so that you can

5:11learn those synthesis pathways and really refine these models to figure out what is synthesizable and what is useful to the real world. I want to talk about the aoy in a moment, but you just touched on what's useful throughout the real world. What help us visualize 83 science is successfully making this more readily available to the real world. What are some of the business use case applications that we could expect to see? 100%. So if you think about, I'm going to go back to the battery space because that's part of my background. If you think about a new battery company that maybe has a solid state electrolyte or a new electrode material or something like that, a lot of times they've got all this R&D spun up around their core differentiated IP, but there's all of these other products. They're binders, there's additives, there are things that are generally called co-products that enable the performance of their innovation that they have to spend R&D time on, but they

6:13don't really want to spend R&D time on, right? If they could just buy it off the shelf, they would love to do that. And so we plug into a lot of these processes and we say, hey, you've got your core IP, that's great. Maybe there's something upstream, maybe there's something downstream, maybe there's an additive that has the properties you need that works very well with your material. Let's go run at that. We'll propose some candidate materials that they don't have precious metals in them. They've got a supply chain. They're not on these restricted lists. And let's incorporate that into your innovation. And that's really our bread and butter right now. We're going to grow from that space, but we've seen a lot of interest in solving that type of problem for these companies. Makes absolute business sense because if you zoom out at an executive level, if there are off the shelf solutions, which will allow for R&D with better results in a quicker, faster turnaround, everyone's going to want a piece of that because it will reduce burn rate insurance, gets them to product voyability or solution quicker, faster, makes sense.

7:15Okay. So, I'm going to give you a little fun stuff, very complex technological problems that we're trying to solve here. Obviously, aoy is at the forefront of this. It gives us incredible tools to do so. Speaking to an audience of aoy and data professionals, help us look behind the scenes. What's the how? I appreciate you may not be able to give away everything, but help us visualize how you guys are going to execute what is the steps involved and that the people involved to because that's a lot of what it comes down to. There's a bunch of stuff going into this. Let me tackle it at a high level. So, the frontier models, they've got these diffusion models that propose new candidate materials. We're very much shifting from the old world, which was let me train a classical ML system on, hey, there are these inputs, center it at this temperature for this amount of time. These are the elemental inputs and let's start predicting the properties, the yields,

8:18the outcomes of this experiment. There's a bunch of companies that do classical machine learning for materials discovery for process optimization. There's not that many companies yet that are really leveraging the frontier models to the extent of their capabilities. It's a new field. It's a super exciting field. You're starting to see folks implement this in their work streams, but that's really where we want to live. We don't want to be the ones that are developing, hey, here's an ML algorithm that helps you improve the yield of this very specific process in the chemical industry. We want to be the ones that say, hey, you need a new material. We've got this general frontier model that we then filter with real world experimental data that can tell you, hey, it seems like this is the boundary condition for avoiding side phase production in your crystal lattice. Let me give you a super tangible example here. One of the things that our CTO Eric is world class that is crystal structure determination.

9:18If you work in really anything in an organic chemistry, you do a lot with crystal structures. Fickering out the crystal structure of the material you've made is a non-trivial process, particularly if you've made something unexpected. Right now you do a guess and check. You've got this bank of these are what we know crystal structures look like when you run an X-ray diffraction on them. Let's see, does this look like what we expected it to? Okay, check the box there. If you're producing impurities, if you're producing side phases, what we do is we help you build that data analysis pipeline to really understand what it is that you've made and we relate it to the synthesis processes that made it so that you now get a really good understanding of how does synthesis relate to crystal structure, how does crystal structure relate to properties. And based off of that, we get a really good understanding of this is the synthesis process that takes you to this material that has these properties that you're looking for.

10:21You are listening to the All This Podcast. When you're looking to scale your team or if you are interested in showcasing your company in a future episode, reach out today. Or if you're in the market for a new role, visit our website to view open positions. www.allthis.com. Thinking about the steps to build a business like 83 Sciences, you really have to break it down into the core foundational level and you've obviously assembled a stellar leadership to him and yourself and Eric. And I know there's a few more pieces of the puzzle to slot in. So, focusing on the engineering component. A lot of our audience will listen in to hear about new and exciting companies. We're doing innovative things within a oil and we're doing growing. I want to look at the near term role map the next 12 to 18 months. Twofold. One, what do you exote about when you think about what's in store, what you're going to build? Part two of that is part of pieces of the puzzle that we still need to find in order to

11:25have the roster complete. To your point, we're growing the team. We're in YC right now. Demo Day will happen on September 10th. He wanted to interest in tracking that type of stuff. We'll raise our seed round pretty probably before Demo Day, maybe shortly after Demo Day, the check will actually land. And we're really looking to grow the technical side of the team. So I think we've got access to one of the most exciting databases of experimental results out there. We're working with a bunch of labs where technical folks in the chemistry space, you probably know some of the professors we're working with. We've got access to all this raw data. And right now, it's a relatively hands-on process, right? We go through a lot of the data. We've got systems in place for investing it, but there's still a lot of more manual work and understanding that goes into it. We're having folks right now build those pipelines, that auto and just the data that understand the context and really start driving this process at scale for materials discovery, which, if you think about the end goal of AI for science, that's the whole thing, right?

12:27Can you feed in the data? Can you automate all of that stuff I talked to before, the relation of synthesis, to characterisation to properties and then spit that out, in commercially relevant products? Yeah. That sounds fun. We could probably spend another 20 minutes talking about possible use case applications and various industries, but let's keep it to the here and now. One of the amazing opportunities that those exist currently at 83 Sciences is engineering team, founding engineers. So I'm sure once demo day comes, you're going to have a lot of interest, but as it stands right now, that key role is one that we're talking about. Speaking to an audience of potential future employees, the caliber of yourself and Eric and the backing of a voice is of the highest order. You want to assemble a team of eight players. So you're trying to attract the elite level talent, which no doubt will be very interested

13:28in what they've just heard, but there's also a lot of competition for that same caliber of individuals. So if you were out for a coffee right now with somebody who you felt could be a good addition to the team on the engineering side, what would you tell them about the mission, the opportunity, the work that would get them excited and interested to join 83 Sciences over somebody at a great company is also out there. We've had some of these conversations, I think like usually if we're talking to a candidate there, they've got an offer from Anthropic Open AI Deep Mind Meta Fair. The way that I would sell it is I think this is one of the few opportunities you'll have to join at the early stage of a company that's doing some of the most exciting work that you could do in AI for science, right? Sure, you could go join a periodic or a lila or a cusp, but you're going to join as an entry level engineering talent. And I'm sure you're working on interesting problems there. You're not going to have the ability to set up the data infrastructure to run this the way that you are most interested in.

14:29I would argue that you'll probably also be focused on these specific, these specific pain points that we know that those teams have and you'll probably be working on a very narrow scope. This is an opportunity to build a company in your vision, recruit a team and build this with us. Ian, thank you so much for coming on today and to all of us. It's been an absolute pleasure. I appreciate you sharing your backstory, the story behind the formation of 83 sciences, the mission where you're at in your journey and the great insight into what you're building and why it could be a great place to work for the right type of engineer. So we wish you, your team and everyone at 83 sciences the very best of luck. Not just on demo day, but everything else to come and look forward to watching you build an amazing company and look forward to having you back on the show again in the near future. Awesome. Thank you, likewise. It was great being here and hopefully we've got some exciting updates in a couple of years for you guys.

15:31Thanks for listening to this episode of the oldest podcast. If you enjoyed today's episode, don't forget to subscribe, rate and review. We are available on Apple podcasts, Spotify and any Android podcast of choice. You can also head over to our website, www.allthis.com to listen to more podcasts, view our open roles and stay up to date with industry news. Thanks for listening and stay tuned for more great episodes coming very soon.