Arena Physica launches Heaviside, the first foundation model for electromagnetism, trained on a self-built data factory

you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange? Just do it. And let me also tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more. While Railway automatically takes care of scaling, monitoring, and security. And without further ado, we have Pratt from Arena Physica. who's the CEO and co-founder. How you doing?

Hey, Jordy and John. Great to see you again.

Good to see you again.

Great to have you.

Welcome back. Um,

yeah, it's been a bit.

It has. Please, uh, reintroduce yourself. Give us a little backstory and then we'll go into the company.

Yeah, perfect. So, yeah, Pat Rad, CEO and co-founder of Arena Physica. Uh, you know, we were on your show announcing some fundraising about, I'd say just about a year, a year and change ago, maybe a year and a quarter. And, uh, companies evolved a bit since then. we were you know started with bringing AI to hardware engineering. So AI hardware engineers that's evolved actually quite a bit um as we've seen the whole AI space evolve which is what is you know if we think about modern hardware it's software defined right like that's that's happening everywhere

but software definined hardware is really like electromagnetically governed like the key problems are it's the electronics interfacing with the firmware interfacing with the mechanics and it needs to obey physics and behave well and turns out humans um can really use one of those forces of physics really well electromagnetism um and it's a rate limiter for progress. And so what we released yesterday is the first foundation model for electromagnetism. Think about it as a a large similar to a large language model in architecture. But the training tokens, they aren't words, they're materials, geometries, and electromagnetic fields. So learning us helping us speak another language basically the kind of language of the universe.

Oh, so what's the data set for that? Like is there you go the website for sure. So the April Fools actually app

April Fools. It's a new LLM.

It's a new chat box. No.

Yeah. No, for sure. The data set doesn't exist, which is why this has been hard.

Yeah. So, what did you do? Is this simulated or did you have to go collect this? How does that work?

Yeah, that's a component. So, so the reality is like what we've what we've done is basically build a data factory. We spent about a year and a half building that. and um you know how do we how do we get all of the data that a model of this sort of architecture a transformer type model needs. So it starts by creating random patterns. So that's layer one of the cake if you would is large amount of random generation and that's different geometries with different material stackups uh and then running them through simulation to see how they behave. So that's sort of think about it as layer one of the cake. But you know the number of possibilities here explode. You add more layers. You think about modern design it's happening on silicon. how does you know how how do they actually behave in the real world? Um you can't explore that whole universe. You can't randomly explore it and this has been one of the big bottlenecks is there's very few of these experts in the world. Uh you know I think Starlink is a great demonstration of what you can build when you've actually mastered such capabilities but few people uh you know have achieved that outside of SpaceX and it's bottlenecked by experts and simulation. So the second layer of the cake is our experts have been creating designs that work and then the final step is we actually fabricate those designs and then we pipe that back into the system. And so I think you guys might have something over there.

We do. We do. We received something in the mail but we we have weird news because it was intercepted. Somebody opened it and stole maybe half the package. I don't know if there's two things. I only have one, but I do have a screwdriver. And this

it was an amazing presented in quite an amazing way. I think somebody basically was like, "Okay, this this looks valuable."

I think that's what happened. So So uh I will I will open this. You can tell me what this is and then uh if the other piece of the puzzle is missing, you can break that down and we can go uh find the thieves

uh at some point.

I will talk through the other piece. That is a booklet that describes the background of what you're opening.

They stole the booklet. stole the book. I know. Weird.

That's extremely weird.

For the metal box,

the metal box is probably a nation state.

I have no idea. That's just so weird. We're not making this up. Like we opened it. We we we received it and it had been ripped open and the booklet had been stolen. But I mean, I hope that's not like intellectual property or something. Uh but inside it says, "Caution, electromagnetic super intelligence. Hand handle only if you are ready for the future." Love it. There we go.

And so let's open this up.

All right. All right. So, break it down for

Yeah, break it down. What What is this?

Okay, great. So, um I'll fly through a bit some of the booklet for background and then I'll tell you what that is. But if you look at what the booklet is focused on, so what can we do with electromagnetism? What can we do with structures like the one you're holding? What you're holding is um an RF circuit.

Okay?

So, a circuit used for radio frequency communication. And if you think about what's happening right now, communication and sensing is kind of the big bottleneck for robotics, for satellite communication, for a lot of the bottlenecks in data centers. And that's gated by this um small group of experts and these really slow simulators from about the 80s, right? And so um

what we've what we've done here is actually u create a design from scratch from nothing but a prompt. So the way it works and it's actually live on our website on our research page arenaphysica.com/ressearch. You can read the technical blog and you can play with the product. You can literally type in a prompt and that goes to an LLM and then the LLM basically passes it to our foundation model. So the LLM interprets the prompt. It's like oh John's trying to build a 10 GHz band pass filter for his new satellite company. Um, and it pass you guys like hydrate it. You guys sort of like hydrate the the prompts uh to make sure it's structured and has like the necessary information or or what what does that pass off look like?

Yeah, it's a it's a great question. So the prompt that's where we actually will rely on the LM. So let's say you give an incomplete description. You're like, "Hey, I'm trying to build this space antenna, but you didn't specify that it's a filter." The LLM you will make some good assumptions and say, "Are you looking for this?" Um, and then that prepares sort of the inputs that we need. Um, and then then it calls our model. So, in a way, what I'm really excited about is is I see a future where you have multiple AI foundation models working together. And, you know, we've got one that speaks English really well, which is your interpreter, which is your front door.

Um, but now it needs to go and create a structure which is physically valid. So, that's where it calls Heavyside, our EM foundation model. And so what you're holding is a 10 GHz band pass filter uh which is one of those key components inside a modern phase array system which is used for radar.

Okay,

for satellite communication um uh and so you you think about there's a lot of conversation about data centers in space. We're not going to fiber optically cable them together. Uh you know you're you're going to need to transmit that data again between them wirelessly. And so that that industry basically started off in about the the sort of 60s with uh with phased array radar. Actually I think the first one was back in World War II. Um, and so, uh, what what you know what you've got in your hand is like, if you look at it, it's an alien geometry.

It's a strange structure. You know, it looks it doesn't look like something that came from a human brain, right? It's like it doesn't look like a nice geometric pattern. It's not a line. It's not a coil. Um, the crazy thing is that it works.

And and Jordy, just to be clear, you're not supposed to touch that with your bare hands.

Yeah, it's over.

It's highly radioactive. I'm kidding. No, I'm not I'm not convinced this isn't alien technology,

but uh but cool story.

No, it's very very cool.

Uh so I mean I I I appreciate the example of like the uh like connectivity and space data centers. Obviously that's like frontier. Um, but walk me through the actual like market map industrial scale of like I imagine there are like certain power law buyers of uh of RF equipment like it's every it's in everything from like that Bluetooth connected toaster to your phone to your car. It's everywhere. But what is the actual shape of the industry? where what when I'm thinking about like subdividing the industry and getting down like the satellite industry is here, the smartphone industry is here like walk me through the map of the like the potential customer set.

Totally. Yeah. If you think about that there's a few different buckets. So if you can harness electromagnetism in RF, which is let's say application one, um you've got satellite communication. So phase rays for a satcom. So that's your user terminal, your satellite. So you think about a Starling satellite, I believe um it needs to track 64 different locations on the ground. So make 54 64 different beams. Um you have to do that with a phased array. Um so anyone launching anyone in in the space industry, which I think is it's exploding right now, um needs communications between Earth and space. So phased arrays are key there. Radar, radar is another really big one. So if you think about that, you know, we've obviously got Rathon as the the the incumbent here. But if you think about advanced radar, they're incredibly expensive today and we have few of them on ships. You know, we should have backpack based radar for soldiers, for counter drone. We should have a whole variety. This this we want to debottleneck this field and basically democratize it almost like AWSify that for companies building for space for uh radar radar makers also for companies um you know if you think about um data centers chipto-chip communication. So these channels that let chips talk to each other are incredibly difficult to get right. In some way you're trying to build a very bad antenna there cuz you don't want it to pick up random interference. You want it to send the signal you want. So just those alone are are massive markets. Actually we looked at the forecast for um the phase ray and RF components and it looks like this line. And I feel like it's something I would have done um right out of school as oh great let me take the historic numbers and propagate them forward. But it doesn't seem to account for the fact that we're expecting a boom in space. We're expecting growth in in in AI data centers. And if you think about um imaging, you know, I heard something from from our chief scientist here that blew my mind. I always thought about radar as detection. You're trying to detect something, right? And he said radar is an imaging platform. So as you get to higher frequencies um you know like LAR um you've got this high frequency beam that's echolocating things like a bat and it's useful for all weather imaging. So when light's getting blocked, you can actually use a radio wave to go kind of create your LAR map. So if you think about that being used for robotics, which is another application. So each of these pieces we feel haven't been sized into the forecast. And so that's I think where where a lot of a lot of the opportunity lies. And when we talk to customers, so you know, we work with um AMD and Andrew as examples. Um, so our our mission is partner with companies at the frontier that are trying to challenge existing cost models and push frontier technology harder. Those are sort of the places where we're seeing early adoption.

And do you want to be like selling intellectual property like ARM or software for chip design or the actual finished chips? How vertically integrated do you want to be? Yeah, I would say like realistically as of where we are today, our model is closer to the model of like a Palunteer uh or an ARM. Obviously, my own heritage last company was bought by Palanteer. Um and so partnering with companies using our whole tool chain. So this is the latest product, Heavyside, the EM Foundation model, but Atlas, our current product is an agentic product that lets you use LLM agents to debug parts of the hardware process. So that plus our FTEEs that are not just software engineers but also electrical and RF engineers uh for deploy as partners to drive major outcomes. Uh what we are seeing and what you're holding as a piece of this is you know the the existing IP industry is let me license you IP and this is where it's an open question. We haven't figured it out but um

you now have an IP generator as a machine

and so we're really excited about uh the ARM kind of model as I think probably the most approximate to where we are today. and our way of having the most impact because we can empower more teams. Like our mission is to empower more teams to build more advanced stuff. If we think about a lot of the AI doom and gloom, we're like where do the jobs go? You know, the conclusion we come to is companies just need to be more ambitious. Like that's how we create jobs. Let's assume we're working with AI. How do we enable you to do that? How do we enable others and more companies to do this? The the the last thing I want to say on that is if you look at what happened with language models, it was all possible because of scaling laws and like you know before language models there was a lot of machine learning where we were trying to go and solve problems at that use case level right we were solving like language translation summarization um spam detection separately but once you you pointed a huge model at language as a substrate and open published that seminal paper in 2020 on scaling laws you could scale it up and it started to generalize now these things can do remarkable things we believe we will see and we're seeing early hints of scaling laws for electromagnetics and this might be true for physics in general electromagnetism being one of the four fundamental forces here but that's sort of what we what we think is possible so each of the industries you talked about doesn't need a specific solution we think in the future as you scale this up this central brain could power any industry that's bottlenecked by electromagnetism you know phase array radar satcom and uh data center interconnects being a starting point Yeah.

Wild. What's the reaction been from the experts that you've mentioned? Are they experiencing a chat GPT moment with uh the new model yet or you still need to iterate? How are people responding?

Yeah, it's a great question. I think there the reaction's definitely been really positive in some pockets and and and deeply skeptical in in other pockets. Uh as you can imagine, getting this stuff right is is really difficult. Right. We're um I would say we're at like a GPT1 moment here. We're not at a chat GPT moment. We have about, you know, tens of millions of data points in the training set. As I mentioned, this data doesn't exist. We had to make it all in the data factory. So that's sort of one of the keys. And so we we're finding as the models get larger and as the data set gets bigger, they're capable of more. But today, what you're holding, a human could design an equivalent component like what you're seeing with a normal structure and it would work. We're not yet at a superhuman point with the model, but what we are is transforming the cost structure. um this is 800,000 times faster than a commercial solver uh for instance and the expertise is not locked up in an expert. You could go and just uh you know get that kind of on demand from from the AI. So that's what we're seeing today and I think it will take some time to move up that that stack. But that's why I'm super excited about the scaling laws and I think from you know as I mentioned AMD and Android Androl I think we're fortunate enough to get uh partners and customers that I think are really on the frontier that are leaning in and believing in this and so we're super excited about that.

Can you talk about the FTE model in this case? I mean I feel like a lot of people don't even understand the the typical engagement of having Palunteer FTEEs inside of an organization. They might be more familiar with like McKenzie coming to the organization doing a delayering or or some management consulting engagement like how long is a product life cycle development life cycle how long do you imagine that a typical engagement with FDES might be is it going to be the same as Palanteer or different like what how is this unique

yeah so right now you know we we definitely see the engagement models being you know you're we're partnering on an array of problems so we start with one for example let's say that new product uh innovation cycle is you know for something like a new chip or a new data center that might be in the number of like a year it might be a couple of years but what we're seeing is then those architectures are used across multiple product lines by a lot of these companies and so uh could we become a partner uh long term across all of these electromagnetically limited products that's sort of where we see sort of that that outcome driving you know that driving really big outcomes across the stack

yeah that makes a lot of sense well congratul ratulations, Jordy. Anything else?

Amazing progress. We will put this to use.

We will

as many of these as you send.

Entire show will run through that. We'll figure it out.

We'll do it.

It might take a lot of queries for us. Incredible. FTE. You have to send one over next time.

And uh let's let's uh have you back before another year passes.

Yeah, we'll talk to you soon. Have a good one.

Awesome. Thanks, guys. Great to see you. Great to see you.

Goodbye. Cheers.

Let me tell you about Console. Console builds AI agents that automates 70% of IT, HR, and finance support, giving employees instant resolution for access requests and password resets. And clan,

let me also tell you about Lambda. Lambda is the super intelligence cloud building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. And I think we're in the Lambda Lightning round now. Or are we in something serious? Let's Let's fire up the Lambda Lightning round. We love the Lambda Lightning round. The cloud

is working overtime.

Kick. Yeah. What's going on?

Acquired FaZe Clan.

Oh, that's announced a strategic acquisition of FaZe Media.

Yeah.

A globally recognized creating private.

Okay.

The tax is real. It's not April Fools.

Oh,

I'm just throwing that. I don't know.

Okay. Yeah. Kick. I don't know. Can we dig into this? Is this an April Fool's joke? Somebody already asked rock. Yes, 100%. It's an April Fool's PL prank. The press release is dated 412026. There's zero real news of confirmation of any kick phase deal and it fits the classic nostalgic streamer or acquisition troll vibe. Nice catch.

I'm so I didn't realize this was

Is this supposed to be funny?

It's on their Wikipedia though, so I don't know. Actually,

I I think this might be real. I don't know. I mean it seems like it seems like if FaZe was taken private and is looking for you know a new home integration into kick like this doesn't seem like that crazy. It's not like if it is a if it is an April Fool's joke. It's not like hilarious. I'm not like oh wow I never would have imagined that Kick and FaZe teamed up like that's there are people that are making that joke. There was a some company high frequency trading firm that was joking about being acquired by Anthropic or something. It was like a crypto wallet company.

Crypto wallet company like that that feels a little bit more nonsequiter, right?

Well, Adam Faze has a good story. He says Faze Media owns the trademark for his born last name Faze and sent him a cease and desist in 2023 when he originally named his production company FaZeworld. Says, "I want my last name back."

I'm rooting for you. I'm rooting for Adam Faze to be able to use the brand.

You should have bid on the IP, buddy.

You should have You should have bid. Yeah. I mean, also, what were your parents doing? Not trademarking your name in every possible category as a child. This is deep alpha.

I need to lock up cou later.

Tonnie on a roll today says it's never been a better time to be in the fake wood panel industry. If you sell fake wood slats that can be quickly installed in a room somewhere, business is booming with no signs of slowing down.

I love the slat walls. We got to get one in here so we can just like step back and go and hang out and do a little slat wall podcast or whatever.

Wait, that would have been that would have been the April Fools. What were we thinking?

You know, we struggled with April Fools this year. We we think it's a little bit overplayed. It's not that it's it's not that funny. Uh and kind of every single day is April Fools for us and it's a good day for us to just get serious, right? Get in the white suit,

focus on the market. The market's up. We're in white suits. Dylan would be serious. But, you know, the the actual play would have been to rebuild like the typical podcast set with like the bookshelves, the slat walls, but then make it miniature, right? Oh, yeah.

So, we're sitting there. We're sitting there. We look like, you know, giants.

Giants.

Giants of the industry. Uh, do you know why the slat wall is so popular?

Isn't it uh Hubman?

Uh, yes. But do you know why he selected it?

I do not. So when you are recording a podcast, you want good audio quality. You don't want reverberation. You don't want a lot of flat walls that will bounce the sound back to you and create echo and distortion and hollowess in the sound. You want a nice bassy response. You need a nice microphone, but you also need a sound treated area. In the Ultradome, we have some sound treatment over there. We have some sound treatment over there. We do have a hard floor. Maybe that would make the audio quality better. Who knows? But we uh uh for a long time people would put up that sort of it's a egg crate style. I don't know if we have a camera. Maybe the the reverse PTZ would would do it. But uh the egg crate style uh soundproofing works incredibly well. And if you're ever in a professional recording booth as doing voice over for a film or recording an a song, uh you will be in a booth that has a lot of spike. Yeah, it has a lot of spiky. Yeah, you can see it right here. So, this sort of egg crate, it creates a lot of little holes for the for the for the sound to get caught in effectively. And uh very fun, Jordy. And uh and and the problem is that that makes it look like you are in a sound booth. It's not very aesthetic. It doesn't look like a natural environment. And so there were companies that said, "Let's get the best of both worlds. Something that's aesthetic but still sound treating." And so they launched these wood slat walls. Of course, the space in between the wood captures the sound a little bit, acts as a little bit of deadening. And then of course you can also hide um you can also hide soundproofing material behind the wood slat wall. And so it was a very logical way to have an elevated aesthetic while still getting the benefits at least some of the benefits of sound treatment. Then Andrew Huberman did it went mega viral and everyone sort of copied it over and did the same thing. And uh it's become a huge trend. He wound up painting his black, I think.

And who knows, now they got sick of being

now, you know, if you're in the black paint industry, that's where the money is is because everyone with the fake wood panels is going to be painting it black. Um, we have yet to do this. I've looked at this in the past, but I've never I've never actually done it. Anyway,