Periodic Labs emerges from stealth to build AI scientists for materials discovery

by the TV. Congratulations on the overnight success. We'll talk to you soon. Success. Have a good one. Thanks, guys. Cheers, Victor. Our next guests are already in the reream waiting room. We'll bring in William and Dogus from Periodic Labs. I hope I'm pronouncing that correctly. Welcome to the show. How are you doing?

Hey, good. Thanks for having us. Appreciate it. Fantastic. We have been so excited for this. This feels like the antidote to uh the timeline today. Yes. So uh yeah, there's a lot of negativity about certain AI products. Uh why a lot of positivity about periodic what you're building. Yeah.

I mean I mean first of all that that's very important technology too. So like don't want to like pile on the hate over there. Totally. Yeah. Yeah. I mean we we've talked about we've talked about the good and the bad. The good is that it's it's it's a very it's very entertaining outputs.

There's logical business reasons why you would do it. Clearly, there's demand economic flywheels and yes, world models for training robotics, training, you know, who knows, maybe a robotic surgeon in a decade will be trained on uh AI video in some way. But anyways, let's talk about you guys.

Quick intros on yourselves and then the company. That'd be great. Hey guys, my name is D. Um I have a background in using machine learning to study physics and after my PhD I joined Google to do deep learning research and that's where Liam and I met.

So that was u eight years ago and at Google I did some deep learning research but I continued to do my PhD passion to kind of move all the advances from deep learning into solid state physics.

And then in 2020, I actually started the team to just focus on material science because I felt like LLMs were getting really good and we could probably really benefit from porting these advances uh to the physical R&D.

And then I've been uh doing that until you know six months ago when Liam and I left our jobs and started this company. Amazing. Yeah. And uh I go by Liam. Uh I have a physics background academically. Spent many years at Google Brain.

worked on a mix of generative models, reinforcement learning and created some of the first trillion parameter neural nets. These are sparse models um really fun collaborations with Jeff Dean, Nome Shazir and others.

Then in late 2022 uh went to OpenAI and a group of us took precursor chat products and we turned that into chat GPT. Um like I've told many many people it was legitimately a low-key research preview. Uh I mean there's like the prior up to that point was there had been no successful chatbot. Yeah.

So there was a poll um as to how many people would use this thing and people were guessing numbers like 10,000 50,000 someone's like oh maybe a million people will use chatbt which we exceeded in like a week.

Um, and I mean reflecting back on that, I think obviously there's good luck, good timing, but also boring things done really well. Like we're really focused on uh training details, data, evals. Um, there was no novelty to it.

Like ultimately people have been building chat bots for many decades, but just crossed a critical boundary. Um, yeah, and so left six months ago to start periodic. What do you think of that gu idea of you just need to like generate random ideas and look at connections between them to find scientific insight?

Do you think there's anything there? Is that directionally correct? Is that a path that people will go down? Are you going down that path at all? I mean, I think it's incredibly promising. Like I think linking these desperate ideas is really promising.

However, unless you actually have experiment in the loop, you're just sort of thinking. Sure.

So you can imagine like hey here's a a really promising thing to try but as every scientist know it's like until you actually take that idea into like to try it to act you're really no further along to like building up a scientific understanding.

So for periodic we're it's that's our opinion which is we need to have experiment in the loop. That's super key. There's so many different experiments. Uh are you selecting for a few different environments? you're going to get a large hydron collider or a wet lab like the science is very broad.

So have you narrowed it down at all? So one way of thinking about it is you know the llams today are really good at math and logic. Yeah. So the next obvious frontier is theoretical physics. Sure.

And there are different energy skills of theoretical physics but we wanted to take the part that is very relevant to human life which is quantum mechanics you know shinger's equation. So this affects all the inorganic materials, organic molecules, uh drugs.

Um so we're starting at the level where we're just designing how atoms come together and how their properties can be utilized in devices. Just start with the simple stuff. Just uh yeah. So so yeah, just drilling in deeper.

What does what does success look like for you guys as a team over the next even even you know I can imagine the five year what success looks like 5 years 10 years right these sort of massive ambition uh how do you how do you sort of make make uh you know immediate progress now that you guys are heavily funded and and have the resources to really move on on these opportunities I think you can kind of describe it along technical and commercial goals from technical goals The highest level way to describe it is have a system that can intentionally design the world around us.

Um so right now we've built systems that understand the internet. They have been trained on math and code. We want to take these same type of techniques and build system that can uh produce things of like properties of interest.

So you're trying to optimize for a new um material for a semiconductor industry or space company or defense company's looking for something for like you know a new heat shield for a missile or you know anything. It's like how can we intentionally design the world around us and and that's sort of um a goal.

Um some of our higher risk goals are around novel discoveries that Doge can talk more about as well. Yeah.

And even even be before you get into that, like how you know you guys are coming from organizations that were totally okay with years and years and years of just deep research and not worrying about immediate commercial applications.

From you guys with the new company, I'm assuming you're you're open to having much longer kind of commercial timelines and just, you know, basically experimenting, right? Some of the stuff is unpredictable. You talked about chat GPT being a research preview and it turned into a hit product.

So yeah, there has to be some you know takeaway from that process. I mean one comment on that is we can basically produce through this technology through experimental data a foundational understanding of physics of chemistry of atoms just the world around us. And you can sit at different levels of abstraction.

Some some levels might be more kind of like scientific like exploration and others might be more towards advanced engineering, advanced manufacturing, engineering. And we're going to have both pieces as part of periodic.

So the labs allow us to produce deeper understanding of different like physics, chemistry, material science systems. And our guidance there is like are we able to produce new discoveries? Are we able to advance science?

But however that data also gives us a better foundation uh when we're actually working with customers. So we want to make sure that our labs our prioritization the tools we're teaching our agents do have some grounding in sort of the uh commercial commercial. Exactly. Yeah.

Like technology and capital are very intertwined. It's not like technology doesn't develop in a bubble and we can accelerate science maximally when it's like a a very commercially successful enterprise. Yeah, I went on a serious emotional roller coaster with the LK99 saga.

Uh give I and then I uh there was a you know proposed breakthrough in in superconductors, room temperature superconductors. It was very excited on the internet for a few days. Uh then it was found not to be such a breakthrough. Um, and then I kind of fell out of that news cycle and stopped tracking it.

Uh, get me up to speed on where we are on superconductors. What's at stake if we can have a breakthrough there? Is that something you're interested in working on? Do you think that this is actually even a tractable problem that you can make any sort of prediction about where we are in terms of progress there?

I'm super interested in that. Yeah, you know, we're very excited about superconductors and being able to discover a novel exciting superconductor requires us to do well on many dimensions of physical sciences which we're very excited about.

So, one of these is being able to synthesize materials reliably, being able to characterize materials reliably, being able to predict their properties, design their um kind of defects and for us you know it's also a very uh important scientific goal.

So currently the highest TC superconductor at ambient pressure is about 135 Kelvin and under pressure you can make it a lot higher temperature but then it's not very practical because you can't really apply that kind of pressure. Yeah.

And it's very interesting to wonder you know can we bring that up to 200 Kelvin for example and if we can even before it impacts products I think it teaches us a lot about the universe because it's a very large macroscale quantum property.

Um so you know we have real good characterization tools in the lab and that you know immediately prevents us from an issue like alken 999 because we can actually measure its properties at different temperatures and for us the LLM being able to use the characterization tools and infer the relevant physical insights is crucial.

So we're teaching our LM agents you know the ability to synthesize the abble to characterize the abble to hypothesize the next step.

Um so it'll be really exciting I think and yeah lab as part of this is is so key right it's the lab is providing our grading function yeah so again like that that keeps us grounded that makes sure and you know it's sort of a very interesting objective to put a lot of optimization pressure against yeah what uh what uh have you stackranked the impact of super room temperature superconductors or even higher temperature superconductors in terms of the impact I remember people talking about quantum computing, but then I also saw a demo of just some like hoverboards and like skateboards that were using magnets and superconductors to or or super cooled uh magnets to kind of like float.

Uh where do you think the impact would be if we get a breakthrough there in the next couple years? I mean the impact is huge like whenever you think about a futuristic technology like fusion. Yeah.

um you know low um low uh loss energy transmission, quantum computing as you said and even you know when you talk to chip companies like think about the chip design of 10 years from today. Yeah.

Superconductors always come up because it's such a important quantum mechanical property but at the same time it's just very easy to understand it lowers your resistance to a point where losses are very low. Yeah. Um, so for us it's probably one of the most impactful things we can do on the solid state physics side.

Can you guys give us a white pill like a pump up speech? I feel like so much of you know that there there's been uh common chatter about you know all this capex all this compute that that we're developing as as a you know a country and a and a and a human race.

A lot of it's obviously going to go to image generation and and memes and things like that and that's fine. But you guys are doing the thing that that has been promised by you know the the AI community broadly for a long time. How how excited should people be about the potential?

I think you guys are both very modest, but like talk about talk about the impact in in your words because you know I I I take it it it it means a lot coming from you given that this is your scientific discovery is like you know the entire focus right and and commercializing it versus you know labs uh historically that are maybe working on some of the same stuff but they're also working on a million other things right they're also working on they could be working on codegen they could be working on uh you know new versions of search, etc.

But this is your guys' whole thing. Uh, so I'd like to to hear um yeah, just just that that uh optimistic vision of what's possible. I mean, ultimately, science is unbounded. That's one of the biggest drivers of progress for humanity.

It's not like we optimize and we can now match the performance of a human on some task. It's you're just creating new technology, new abundance by being able to kind of control the physical world around you.

And we think that the technology is at that stage where we've seen incredible things on models from mathematical reasoning, code reasoning, also the ability of some of these agents to start doing like real valuable work, but the world around us still kind of like largely looks the same.

It's like not really moving too quickly. And I think as the cost of intelligence decreases, Doge and I are thinking the bottleneck increasingly becomes contact with the real world.

Bring data, bring atoms like into this end-to-end system that's going to lead to this new ability to just accelerate scientific progress, accelerate the technology around us.

And I mean it's like the applications of you know scientific developments are are endless right like you know becoming more of a you know uh space traveling society like you know new computation um it's you know it's it's unbounded I think that's what excites us.

Yeah like there will be no end to this right like it's not like we can finish science actually the more jobs never finished exactly so it's a really uh fun place to be because of that. I love that. Well, congratulations. We're gonna ring the gong for you. 300 million.

Last last question I have before you guys jump off and uh you you keep it to 30 seconds. What's your philosophy around building and sharing your guys' work? We've seen different approaches from the labs. Thinking machines has been experimenting, sharing some of their progress. SSI been very silent.

Some of the other labs are also, you know, you know, constantly sharing um you know, just just small products, etc. But how do you guys think you'll approach it given your focus? Yeah. So, we're building tools in the physics space.

We're building tools in LLM space and we will definitely share whenever it seems like it will enable the community. We also have an academic grant program where we want to support academic groups that are also passionate about this space. Uh because we feel like, you know, science is endless.

So, there's so much to do and we'll, you know, do it together. Amazing. Well, thank you. All guys, thank you so much for coming