0:00 All right. Thanks, Duann. Can everyone hear me? So, yeah, I got I got super grateful to Van for setting this up and having us here. Of course, putting electrical engineering and electrons right after lunch. Let’s see. But, yeah, so just a quick introduction. I’m Pratap. I’m the CEO and co-founder of Arena. So, yeah, what we’re going to cover today a little bit, you’ve heard a lot of really awesome things about architecture, structural design.

0:27 We’re going to talk about electromagnetic fields. They surround us. First we’re going to talk about why they matter. And we’ll just go through some demos of what we’re actually doing and then we’ll talk about some of the problems and how we’re how we’re how we’re trying to tackle those and basically building you know what we hope to become basically a version of artificial intuition. So with that let’s begin.

0:48 So first of all, you know, just simplifying it, everything is becoming a computer or in some some industries, everything is already a computer. So, you know, in some some dinner discussions here, I think some of you have more facts from proprietary libraries than I could find on the internet. But let’s say these are at least directionally right on the F-22. So, if you compare kind of a mainline fighter jet from the Vietnam War era, the F4 Phantom to the F22, you comparison becomes quite striking, right?

1:16 You you’ve gone from something with you know 10% of the cost and let’s ignore the fact that the cost also was 10x less was avionics to 50% of the cost. So a 5xing in just the density but then the cost went up 10x. So you basically have a 50x increase over the last 50 years and how much compute basically was you you you assume compute including like wiring firmware is really running the show.

1:39 I think the stats from the F-35 are that the this is the first time in the jet age that the avionics costs more than the jet engine. Just to put put that in in in in kind of context for everybody. I think the wiring length is also quite an interesting one from from you just 14 miles of wiring to 850 mi of wiring. It’s really really coiled in there.

1:58 So complexity’s exploded. Really what you’ve got is this electronic nervous system on any piece of modern hardware especially when you go autonomous or semi-autonomous. And then as you know a lot of your smart friends probably spent the last 10 20 years going into application layer software that has become increasingly removed from the metal. So you’ve actually had a huge talent talent loss in the electrical engineering and the firmware side.

2:23 So you have this huge explosion in complexity at the same time as you have this talent cost. So you’ve got an acute problem basically on on on the sort of nervous system side. And so what you’re seeing, and I know again from some of you in this room who are at some of these folks is vertical integration. So a lot of companies are trying to say, I want to do hardware better.

2:42 I want to do hardware faster. And it happens to correlate with which are the most valuable hardware companies in the world are choosing to vertically integrate. So if you contrast this with an old automotive or airspace OEM, they basically, you know, in some ways they’re almost procurement companies. They’re buying parts from tier 1, tier 2, tier three suppliers. But actually as you move down the stack, there’s actually very little getting built inhouse.

3:04 And so I mean Apple is a great example now where you know they’re making their own silicon. If you look at Tesla also making their own silicon. And then you look at SpaceX and companies like Andreal actually building their own like their own flight computers, their own sensors, their own boards. So moving down the embedded layer. So you’re seeing this, we’re probably going to see a a lot more of this in the future.

3:23 But this is sort of how companies are trying to tackle this problem of how intertwined the kind of electrons have gotten into the guts of the into the guts of the machine. It’s the only way you can really really move fast. So what does this mean? It means that you have engineering teams that for the first time could actually think about and are thinking about the sort of full path of the electron, right?

3:44 And if you think about the purpose of electrons on a practical level, they’re doing two things. And they’re doing a lot more, but like they’re really doing two things that we care about, which is they’re carrying power and they’re carrying data. And so really, you can trace that all the way from the chip. So you’ve got a little silicon chip here through the embedded system. So again, in aerospace, you folks would be calling that an LRU.

4:07 In in automotive, that would be an ECU, but basically now you’ve got this hardened box with something that’s operating like an onboard computer, an onboard sensor, an actuator, and then that’s plugged into all those kilometers or miles of wiring. Basically that’s sort of inside the vehicle, right? So for the first time an engineering team is not doesn’t have to be siloed. You could actually ask the question of what’s happening.

4:26 You could actually debug something all the way from what went wrong? I’m seeing an error code in my first drive test to what happened and and and this is amazing because if you’re doing iterative development, you own the whole cycle. You know, the root cause could be anywhere along this line, right? So that’s that’s been pretty profound. And also it means that you’re you’re the the sort of breaking down of the barriers between design, the test, manufacturing and repair.

4:48 I mean these are sort of like artificial constructs that we created based on how we built our orgs. The electron does not care that your team has a different incentive structure is located somewhere else than someone else’s. It actually doesn’t care. So so we have to understand this problem. So you know I’m going to start just a little bit with showing showing what the what what our product does today.

5:08 We started with tests. We started with this first place where you know if I quote Mike Tyson he says everybody everybody has a plan till you get punched in the face and you know test engineers get punched in the face a lot so the designer has done something amazing really optimized then you have a test engineer so usually the design is late so they’re compressed on schedule but they’re still launching on time so they’re under time pressure and they basically are bringing up a new system so we do a lot of work with some of the large AI superclusters that are being set up this is a just an example from public data we just bought bought a bunch of GPUs and CPUs and built a little server.

5:42 So, a small approximation of what you see, but we really are running a bringup test here. So, you’re like, great, I’ve got all these parts. I’ve plugged them in together. I’ve loaded my firmware on. I’ve wired the fans in. Let’s set it up. Everything is going to be perfect, right? And of course not. So you actually have this hard problem where like your design should work. You ran simulation on the parts you could simulate.

6:02 And it never does work. And now you’ve got this race to debug, find the root cause. So that’s sort of where where where our product sort of started, right? Which is using agents to talk to tools to make this entire process a ton easier. And really like starting to model the entire system effectively as a graph. So test naturally leads to debug. So once you’re helping people run testing and saying, “Hey, what’s going on?” They’re like, “Great, like where’s the problem?” And now if you’ve got an integrated stack, your problem may be in your board, right?

6:32 So actually what you’re doing now is you can drill down from oh great I’m seeing this thing fail to my schematic. So I’m now looking at the schematic and so you can see here an example and you know this quote this this guy’s not actually called Jim I’m sure sure in some team it is but you know you know this is a simplified again public data example from I think a TI board but if you look at this you might have let’s say this is a component for a radar you have someone who specialized in like the the the matching network or the amplifier you don’t know who he is maybe doesn’t work at the company anymore and you’re like this it’s probably this part and so here literally just dra drag and select and start asking questions.

7:08 Now, for those of you who are like, “This is just an LLM. Please try this with an LLM and then let’s talk.” It’s not LMS are useful for part of this, but really are pretty bad when you get to anything on the inside. So that’s sort of that was the next step, right? Which is great. Let’s like understand what’s happening in test. Let’s make that a little easier for amazing engineers who are on the frontier but are in a lot of pain.

7:30 And then those engineers, let’s let’s help them try and root cause. And we want a root cause because we want to try and understand what’s going on. And then the third module of the product came to life and I wish I could take credit for the idea. It’s not an idea from us. The idea came from our customers and they said, “Hey, can you actually make changes like could you autonomously make changes?” So, what you see on the left is actually a real kind of a device farm.

7:55 So, you have a lot of boards here. Again, if we stay with the data center example, you see the fans and what you’re doing is you’re seeing software workloads being run on the screen. So, those are real software workloads. We’re running them in test. Obviously, the boards aren’t hooked up to everything cuz they’re they’re you know, the rest of the machine isn’t ready. And you can literally say, “Okay, great.

8:12 I can interpret test data, but can you now suggest what action you might take next? Could you then load that onto the device, make a change to the settings, and then could you could you adjust the results?” And so what was really surprising to us here u was actually at the beginning it was supposed this whole thing was supposed to be a timesaver just a timesaver and I think what we saw is for the first time especially in one of the big problems which is performance depending on whatever your metric is but performance per watt is in many applications a really really big problem and you know of course everyone’s got a lot of AI in there so assume there’s a human assume there’s an AI but we were just blown away at how much better you could actually do and so that was that was I think the thing that surprised me most wasn’t even the numbers it was the way the engineers talked about it they were like oh that was beautiful that was counterintuitive that was creative which is not what we typically associate with a machine and it reminded me of you know something that inspired us when we started the company in 2019 which is Deep Minds AlphaGo and you know if you think about there’s the lots of games but there was game two move 37 and this is sort of the sequence if you read the expert commentary as someone who doesn’t play go but really enjoyed reading the commentary is this is what people said they’re like oh I thought that was a mistake that’s stupid and then they’re like that’s a very strange move and then eventually I think they converge on this is really beautiful this is really elegant so we’ve seen systems like this kind of work before we’ve seen them sort of cross over from like a brute force loop to something that starts to look like emergence and it’s super exciting and we want to get carried away and be like great we’ve discovered stuff we can do stuff but now I started out as a physicist so like a lot of you will be like well really is that how much does it really know right and so you know this is really powerful it’s super useful all those systems are deployed they’re in production those aren’t pilots they’re scaling they’re working this is becoming default in how engineers work but can you come up with a truly new idea and all of the things you saw I was giving it a goal I was telling it what to do I was saying hey I want you to go and now help me debug this thing this flight failed this this bring up bring up is failing.

10:24 I want you to beat this target, but I was the one coming up with with the goals. The engineers coming up with the goals. How do you come up with a truly new idea? Like like that’s what if if if it really understood what we’re trying to understand here, you would be able to come up with something more creative. And so so what is it that we really want?

10:42 It forced us to sort of think about what is the question that we’re asking. And if you think about a lot of our tools today, they’re all I mean since the dawn of like computers, they’ve all been about better deduction, right? Like deductive reasoning, solve, compute, calculate, predict, right? Here’s a set of things given this input, tell me what will happen. What we want is something that can like postulate.

11:07 If I’m aware of how these equations behave, if if I’m aware of sort of the the conceptual understanding of how physics works, I should be able to suggest really creative things. And so, you know, I want to share with you something very personal to me. This is this is an experiment called the quantum mirage. And so, this is what drew me into physics in the first place.

So, my adviser at Stanford actually was was was the the first person to run this. It was on the cover of Nature way back 25 years ago in 2000. But let me just explain what’s happening and then why this is a good example of intuition. So on the left you see 80 cobalt atoms. They’re they’re arranged in in an ellipse on basically like a copper metal sheet, a metal surface.

And why don’t you see the atoms on the metal sheet? Well, quantum mechanics turns out atoms aren’t exactly real. So that will be like an electron C. So you’re not actually going to obviously you’re not going to see atoms, right? So that’s basically what you see over there. But you know the the analogy that inspired him if you think about like what a person can do is we know the shortinger equation.

12:14 We know what we know Maxwell’s equations. And what what Harry did is he said well you know these wailing walls you you guys have you’ve been to those wailing walls? You know you go stand in some like really old place in Europe and you whisper in a corner and then you’re in the other corner and then you hear the sound of the person whispering. And that’s a property of sound waves, right?

12:35 And if you think about, you know, quantum mechanics, we’re like, okay, great. Like matter waves also interfere. And so if you think about cobalt, the the reason cobalt, and we don’t need to nerd out about cobalt here, but is like it has a different magnetic signature. And so what you’re saying is you’re saying, great, could I see would I see another magnetic sign? Would I see basically like the echo at the other location?

12:56 And on the right, so the left is just topology. It’s just showing you geometry. The right is showing you the magnetic information. And so you see this thing which is you see the very bright spot at the second focus and you really see basically like a ghost atom. What’s crazy about this is actually if you account for light delay this confused a lot of physicists and there’s better answers out there today.

13:16 That should take there should be a small amount of time that elapses before you put that there and the other one shows up. In reality there is zero time that elapses. So this has led to a whole field quantum communication and so a ton of that is like you know first we’re like oh great this violates causality how the hell does this happen it does happen but it has led to quantum communication and so this is a great example of something which like we’ve known the equations for a long time but it took a human mind a human genius to go and postulate and come up with that right and in that case that mind was Harry’s mind but how amazing would it be if we had machines that could actually really understand these equations and help us like we have to wait, you know, a few decades for a genius to come along and push a field forward.

14:02 What could we as humanity benefit from if those geniuses came or had a little help, right? Like how much closer could we pull the future, right? So that’s that’s really what we’re after. That’s the goal. It’s hard, but you know, why are we around? We might as well try. So how do we get there? Obviously, we all know the answer, right? It’s generative design for sure. We’re going to do generative design.

14:24 And of course, we’re going to use LLMs because language is the primitive intelligence. So, we will take LLMs and we’ll do generative design with LLMs. And then we’re very data hungry. So, we’ll get a huge amount of data. We’ll buy a bunch of H100s and everything will work beautifully. H100s are amazing. But, so I don’t know if everyone here I mean I think one of one of the reasons I’m very excited about being here is I think this is probably a rare room where these are actually probably not necessarily mainstream viewpoints.

14:52 In most rooms we’re in, this is the mainstream viewpoint. So the first thing and I know I have a few minutes left so I’ll try and hurry is the problem is actually not generation that’s relatively solved the problem is verification. So quickly I’ll quote Jason we who’s now at meta but was at open AAI one of the creators of 01 and deep research. So I think this this this verifiers rule which is fairly standard now in AI is I think a fairly good conjecture.

15:16 Basically all tasks that are possible to solve and easy to verify will be solved by AI. And you know, we had a bunch of great presentations over the last two days about materials problems and fluid problems where the famous and notorious Navier Stokes equation, you know, bears its fangs. You we don’t deal with mad Navia Stokes, but we we deal with, you know, his evil twin brother, Maxwell.

15:37 And so if you think about that, you also have a bad set of PTE. So that’s that’s part one, right? Salt. It’s hard. You need simulators. But I mean, we know the answer to this. This is just a question of solving hard but there’s a ton of good stuff for this and there’s a ton of open source stuff for this and then you have the bigger problem which is simulation deviates from reality.

15:56 So your real materials actually break the cycle. So one is we need to figure out verification. Two is especially in our world, you know, for software engineers, language is kind of a primitive because code is code is text, but we think it’s just one langu, you know, animals did not learn to move by reading about it. Movement is probably a primitive. Our friends at physical intelligence are taking that belief to the moon.

16:22 We think fields are a primitive. You saw Harry’s work on the quantum mirage. That happened without anyone speaking about it. That was happening already, right? So what if your tokens aren’t words? What if your tokens are fields? So that’s the second thing. What would you get? You’d get large field models. There are four fundamental forces of physics. Our goal is to just do humbly try to do one of them, make a dent in one of them.

16:45 And so that’s basically the second bet, which is it’s not an LLM. And the third part is subtle and data is not the same thing as information, which is really what fuels intelligence. And so I’m going to steal from chapter one of basically the Bible of machine learning, Chris Bishop’s book, Pattern Recognition and Machine Learning. And basically the definition of information is actually, you know, it’s funny that in the textbook, surprise is the word they use, but information is the degree of surprise associated with an event.

17:15 These two sentences have the same number of characters, the same number of bits. One tells you a lot more information than the other. And so that’s really what we’re after. I will steal another scene from one of my favorite video games. Our main conference room, the office is named civilization for this reason. But you know when you start, if any of you have played, like you start, you have a settler, you have a little portion on a map and they’re there and you know everything around it and you can either go and like take all the minerals and build all your cities or you could go up or you could go left or you could go down and you might find minerals and gold or you might, you know, find barbarians and get killed and you you have to take a risk and you have to do this and but but actually the the the real world is the place where you can mine the most amount of information.

17:54 So you can explore in the real world. You saw some of those examples when that optimizer went crazy and discovered new Optima. You saw Harry’s work with with the quantum. We’ve seen this work before. And so to fuel this thing, we’re building basically an information factory. And there are three parts to this. So the first part is synthetically generated designs. I’ll explain the charts on the right at the end if we have time.

18:18 Otherwise after that gets you partway there. Then you have expert-seated design. So think about like procedural generation in video games. You have artists basically creating these sort of beautiful new designs. And they especially, as I said, with the talent gap, there’s like literally 10 people in the world, especially for some sub fields of of this, who are good at it. There’s a lot of people, but there are like 10 people who are legitimately good.

18:38 And they’re expensive, but but those people then can see designs you can then procedurally amplify. So, you now like enrich the data set. So, the second is much higher information. The first is much higher data. You do need both. And then the third is we actually then fab them. So, we fabricate the designs. And this is not human assisted. These are completely autonomous generated designs. We fabricate them.

18:59 They don’t all work, but more of them are working than there were. And then you you pipe that data back into the same system. And so this is a sneak preview. We’re going to launch this in the coming months. But what you’re basically doing is, you know, think about this as your frequency resonance plot for the mechanical folks. It’s the equivalent for EM. You’re basically giving in scattering parameters like how much energy am I putting in, how much is being transmitted, reflected, bounced out.

19:23 You’re seeing actually just one layer on one grid as an example here of like the process running and a design coming out from noise. You can visualize it and as you saw in the beginning with the product, we have a lot of LM agents that we chain with our physics models so they can talk to each other. So you can interrogate design choices. You can drag and drop like I showed you on the circuit and say why did you do this?

19:41 Why did you do this really weird checkerboard QR code thing in the corner? And you can actually get an answer that makes sense. So so that’s basically where it’s going. You’re seeing some really weird alien designs. So we’re really excited to share this with all of you. You’re going to be able to interact with it online. Follow us if you want to get early access coming out before the year is over. So, that’s all I got. Thank you. To see the full recording of this and previous presentations, as well as information about future CDFAM events, visit CDFAM.com.