An AI Engineer Is Here, and It Might Redesign the Physical World - EP 20 Paul Eremenko

Is it Lee, like L-E-E?

It's L-I.

L-I. Where does—

Long Island.

Long Island. Because you're—

We got him, we adopted him in Long Island.

Oh, okay.

That was the only reason. No other affiliation to Long Island.

And Lee goes with you everywhere?

Lee goes with me everywhere. He was a shelter dog and he had separation anxiety.

There he is.

And we got started working on his separation anxiety. We adopted him like fall of 2020. And started working on his separation anxiety, and then the pandemic started.

Okay.

And then we gave up working on his separation anxiety, and by the end of the pandemic, I had separation anxiety.

So, well, yeah, I was gonna say the pandemic was probably good for his separation.

Well, in the sense that we were never separate, right? And so, so it was indulgent of his separation anxiety, but as a result, it never, it never cured, and I developed it as well. So now we have mutual separation anxiety. When I go without him, I have like phantom dog syndrome.

Well, so if people see Lee cruising around, hanging out with us, you— oh, we are like on your— well, okay, I got two questions. I mean, one is, and I am going to do an intro here in a second, but the— you fly, you flew in your own plane, right?

I did not.

No, no, no.

I flew JSX, which is almost as good.

I thought when we were emailing, it sounded like you were— I didn't know if you were just hopping in. On your own plane.

Okay. No, but I do fly. I do fly. I enjoy it very much. And it's been a lifelong, lifelong thing.

Okay. We'll get into that. All right. I'm going to kick us off with a proper, proper intro. We try to.

Let's do it.

All right, Paul Eremenko. Thank you so much for being here. I guess in some ways this is vaguely home turf for people who don't know. We record this at the Playground Global offices. They were kind enough to gimme some space to build my podcast studio. I know they've, they've invested in your one, at least one of your companies in the past. That's right. Absolutely. Yeah. So you've spent time in here.

Some of my favorite people in Silicon Valley here are here.

Okay. And real quick housekeeping item before I do the proper, proper intro. This is all made possible by, uh, E1 Ventures, kind enough to, uh, to lend some support to the podcast. So here we go. Um, this one's exciting. I've known you for a while now. You have this tremendous background, which I'm gonna give to people so they know a little context of where you're coming from. I mean, it's hard to find— there's a lot of impressive stuff on here. You're born in Ukraine, moved to the United States when you were about 11.

That's right.

You have this very impressive family. I learned this week going back to some Russian generals who were—

Marshal of the Soviet Union.

Of the Soviet Union. Your father, I think, is a quite famous mathematician. That's right. You got astronautics degrees from MIT and Caltech, and then also a law degree from Georgetown.

I think. Yep, that is true.

Okay, so I'm doing all right so far. And then you were the CTO at Airbus. I think you became CTO when you were about 35. You're one of the youngest.

That might be true. I'm not sure. Yes, I have never claimed that. But if, if somebody did the math, then it's true.

The internet. The internet says this. You were also the CEO of Airbus's venture capital, like sort of innovation arm.

Yeah, sort of the Silicon Valley outpost. Yeah.

Called ACube.

ACube. That's how I got into Airbus.

You were also the CTO of United Technologies Corporation. You worked at DARPA, you worked at Google on a lot of this stuff is aerospace, but Google was really a lot of projects, but phones and trying to rethink hardware more broadly, I think. Is that fair to say?

That's right. Yeah, that was my only sort of major detour out of aerospace up until the current startup.

And then we met while you were running Universal Hydrogen the last time we were together was at your hangar that you had in Hawthorne. I think your neighbors were Harrison Ford and SpaceX, Elon.

That's right.

Right. So that's quite the pairing to be surrounded by. And then now you've got a new startup. We'll get to the startup in a bit. Why don't you give us like the elevator pitch on what you guys are doing?

Sure. Yeah, we're, we're trying to build an AI engineer for physical, for physical systems. So for, for those of your listeners who are familiar with maybe Devin in the software I wrote the first story on Devin.

Yes.

We are doing Devin for physical product or physical system engineering.

And the name of the company, it's P1 AI.

P1 AI.

P1 AI. And the agent is called Archie, I think. Yep. And you guys launched basically this year.

So we founded the company at the end of last summer. We raised a small pre-seed kind of last August, started recruiting the team in September and really got going in October of '24. And just last month, or I guess now in April, 2 months ago, we raised a fairly sizable seed round and are kind of off to the races and de-stealthed and significantly growing the team and actually launching it into the wild pretty soon here.

Okay. We will get into that because as you know, I'm fascinated by engineering and I'm covering AI all the time. And I love the idea of an AI for the physical world. But I want to start— I mean, you've lived such an interesting life when you go through your background. One part that I was— that just grabbed me, moving from Ukraine to the US at 11. 11 is an interesting age to do that. You know, if you come when you're much— I moved from South Africa when I was 3. I mean, I feel more or less American, I guess, even though my family's from a little all over the place. But Um, yeah, I don't know. I mean, that must have been— did you end up in New York, or where did you end up?

No, actually, we ended, uh, we ended up in Lexington, Kentucky.

Okay.

Of all places, which probably is why I have no accent, right? Is there was really no Russian-Ukrainian diaspora.

I was gonna ask about the accent. I mean, because you're— you would— you would usually— there's some hints. You would just never, never know.

Yep, yep. No, I, I plunged, uh, plunged straight into American culture, assimilated as quickly as possible, and, uh Yeah, we initially landed— my dad is a mathematician, as you mentioned, and so initially he had a visiting appointment at University of Kentucky at Lexington for about a year. And during that year we left— it was the Soviet Union— and during that year in 1990, the Soviet Union disintegrated. And so we had passports from a non-existent country, which led to some interesting sort of legal machinations that ensued over, over the next years. But ultimately, um, my dad landed at Purdue, where he still is. And, and I grew up, like, did high school and then flight training and things like that in, in Indiana.

And you— I mean, I ticked off a couple of things from your family. I don't know. I guess this is a hard question to answer, but it seems like a very prestigious family with a long line of exceptional people.

So I think both on my dad's and on my mom's side, it's a, it's a, it's an intelligentsia family, I guess I would put it that way. A lot of, a lot of military folks, again, on both sides. Some of whom did some fairly impressive, prominent things during World War II. And then doctors, a lot of doctors. And then my father, of course, is a mathematician, and I describe myself as a failed mathematician. I think he had high hopes that we would be one of those father-son pairs, and I disappointed sorely my freshman year at MIT in that regard.

But he's like next level. I forget which award he won about—

Humboldt Prize.

Right. I mean, Was he recruited? The Soviet Union was falling apart. I mean, maybe he just wanted to get out for a better life for his family, or was he, was he pulled over here by the U.S.?

I think it was a, I think it was a combination of the opportunity being there, and, and like I said, we left fully intending to return after a year. It was a visiting gig for a year hosted by University of Kentucky. And then during that year, things became so complicated, uh, back home, back in the motherland, yeah, uh, that it just made sense to stay, right? Like, it wasn't clear whether there was going Civil War, or what, you know, it was a— if you were— if you recall that period, right, it was, uh, it was, it was pretty exciting.

Yes, yes, I lived through that. Wild times. Um, okay, so you— your dad wanted you to be a mathematician. I assume based on your academic record you did pretty well in high school. Were you a sciency, mathy kind of kid?

Uh, I was a sciency, mathy kind of kid. I was also a little bit of, uh, sort of a, I guess, a nonconformist.

Okay.

So I didn't, I didn't really play by the rules very much, and So I ended up graduating from a very good high school, but not— I think I wasn't— I didn't make the top quarter of the class.

Okay.

But nonetheless, you know, MIT and a couple other schools looked at the holistic, I suppose, at the holistic application and let me in.

Yeah. I mean, what stood out? You were flying, I think, from a pretty early age, weren't you?

Yeah. I started flying, I think, when I was 14. You could solo when you were 16. And so I soloed on my 16th birthday and you can get your pilot's license at 17. And I got it on my 17th birthday. I probably shouldn't have. And in fact, I screwed up a landing during my checkride and the examiner was like, well, try again. And second one was a little bit better, not super great. And we taxied in and I was like, ah, crap, I failed for sure. And he turned to me and said, you know, I know your instructor, go do another, another, another couple hours of landings practice with her. But happy birthday.

Okay.

And past me.

I hate to make generalizations like this, but I feel like this one holds true. I think most people I know who got a pilot's license in their teenage years, they do tend to be pretty extraordinary people that accomplish quite a lot in their lives. I guess this is some early indicator of drive and/or just ability to get things done and a fascination with, I don't know, I guess like adventure and science and it makes sense, I suppose.

Yeah, I mean, I think in all honesty, I was and probably still am a space cadet. But if you're a 14-year-old kid in Indiana, there's not that much you can do with space, right? So, you know, I was fascinated with telescopes and stuff like that, but that's not nearly as cool. And so I think flying was about as close to doing spacey stuff, uh, as you could do.

Yeah.

So it was an early manifestation of that.

And so how does it— how does this— clearly you're getting these aeronautics degrees, um, this is sticking with you as you're going through school, which you, you do get a law degree as well. As you're going through all the schooling, what are you thinking you're going to end up doing in life?

Well, I, I did also have a very early fascination with— and I promise this is not revisionist history with AI. And when I came to MIT— MIT has a pretty, pretty active undergraduate research opportunity program called UROP. And, uh, and I, uh, my first stop was at the AI lab, um, and I was like, I want to do something at the AI lab. And what year is this? This was 1997.

So, okay, probably in some kind of winter perhaps, or, or no, maybe that's like It was—

yes, it was certainly winter. Uh, AI Lab was doing a lot of robotics. Rodney Brooks, uh, by the way, uh, who's also part of the Playground family here, uh, was I think the head of the AI Lab at the time, and he was very much focused on mobile robotics, as I guess he still is. Um, and, uh, but the, the folks that I approached at the AI Lab were not terribly welcoming. Um, and so my second stop was at the Aerospace Department, and I was very interested in aerial robotics. And joined the MIT Aerial Robotics team. And then by sophomore year, I think I was leading the MIT Aerial Robotics team.

Okay.

Which were autonomous helicopters to perform various, various kinds of, kinds of missions, the kinds of things that you— we have drones doing today. This was the very, very early days of that in, in kind of the late '90s.

Okay. Okay. And okay, so you're into all this different stuff. You— and, and the AI point is well taken. I actually— it was a little after the time that you were describing, but I found some like PBS interview with you, and I want to say it was like 2013, and someone is proposing this idea of, uh, quite sophisticated AIs, and they're like, you know, wouldn't that be, um, that'll never happen. I don't know, you, you were already, you know, um, promoting this idea back then.

That was, yeah, that was already my DARPA days. I think, I think the episode, if I recall, was called Rise of the Drones. Yeah, right. Um, and, uh, yeah, I mean, I've always been a, a proponent of, uh, of autonomy and, and very interested in autonomy technology.

So A doomer.

Uh, well, let's talk about— how deep do you want to go down that rabbit hole? I mean, look, I think, uh, I think, I think we are on the cusp of what Vernor Vinge called the singularity, right? And I, I am super grateful to be alive at this time, uh, because I think it's probably one of the most interesting times in, in human history. But I do think that Vinge's sort of description of the singularity as a point at which technological change progresses so quickly that we can't really see past it, so kind of like an event horizon of an astrophysics singularity, I think that point is very valid. And so I don't think we know what is on the other side of the singularity, but I think it is nigh.

So we don't know, and you're excited. But there's— you're— so you're— I mean, it sounds like you're not necessarily a doomer, but you're not discounting the fact that this could go awry. I guess that's what you're saying.

I think it absolutely could, but I think we have the opportunity to shape it. And that's one of the reasons that I really wanted— I felt the urgency to dive into the AI world in the last couple of years, sort of like headfirst, is because I do think that there is an opportunity. You know, it's not out of our control.

Yeah.

It's not fate. It's our future to shape and build. And I want to be here for that.

So, but you do have this kind of 15, 20-year period where you're very heavily, I mean, you couldn't be more in the aerospace world in many ways. Was that what you expected to do coming out of college?

Yes, I think it was. I don't think I could have forecast the specific sort of sequence of events, but I remember even in high school being asked like, what do you want to be? What's your dream job when you grow up? And my answer was, I want to be the chief engineer at Boeing. Um, and so I missed the target a little bit, right? I ended up, I ended up at Airbus, um, but kind of, kind of checked that off the bucket list. Um, uh, so I, I was pretty, pretty goal-oriented, mission-oriented, right? Uh, now the fact that I ultimately landed in that particular role at that particular company was probably a series of fortuitous coincidences.

What I mean, to get there at 35 as ostensibly an American, how did that happen?

I don't think anybody has ever called me ostensibly an American.

What did they see? What had you done? What had you been doing right before that that, you know, made you attractive to Airbus and to take on such a big role?

So initially I was recruited to Airbus to kind of found their Silicon Valley presence, which is this innovation center called aCubed. And I also helped a little bit with Airbus Ventures. And, and that went well. I came out of Google to do that, but I had, I had an aerospace background. I was at DARPA mixed with kind of Silicon Valley exposure through my times at Motorola and at Google. And and it went well, right? We acubed, we launched a number of very exciting projects. I think we launched one of the very first eVTOLs that actually flew, the Vahana.

Vahana, right? Yep.

We launched Uberization of helicopters as a service, right? So you could do sort of one-click dispatch of a helicopter and maybe a ride to the helipad and a ride from the helipad, right? So kind of make it part of a one-click point-to-point transit. We launched a bunch of autonomy work, which ultimately led to the first fully autonomous flight of an A350 gate-to-gate, which was a program that we continued when I became CTO and came to fruition a couple of years later. So we did a lot of exciting stuff. And I think Tom Enders, who was the CEO of Airbus at the time, saw that and said, this is what Airbus engineering and the Airbus technology organization should look like. And invited me to Europe. And I think that my American-ness was a positive in that particular context, but it was a very narrow window in which it was a positive. In general, it probably is not a positive for success at Airbus. Just culturally, or culturally, it is a very Franco-German company and there's a lot of tug of war between the French and the Germans. And because Tom's sort of mandate for his new CTO was, I really want a drastic change and rethinking of the way we do technology. Uh, he wanted somebody who wasn't French or German because if there was going to be restructuring, if there were going to be, you know, shifts of where we do things and, and job, job losses, job creation, etc., he didn't want somebody who would be accused of one-sidedism, right?

But the system rejected some of his desires, or it, it went relatively okay?

I think it went— I, I, I'm pleased to say that a lot of the changes that we made, and we did make some pretty pretty drastic rethinking of how Airbus does technology development. Most of the things have persisted to this date.

Okay. This is one thing I wanted to get into. I don't know where you're at in terms of being able to publicly speak about Airbus and Boeing or your— I know we've had chats, sort of private chats. You have opinions on these things. So anyway, I'll go there and We'll see where we end up. Like, what did you find with the Airbus culture when you, when you, when you got there? How would you describe it to people?

So it is very mission-focused. I think the people that come to Airbus really love airplanes. They really love flying. They really love Airbus. It is the national pride of France, at least. I think in Germany it ranks maybe a little lower on the, on the list of most desirable places for an engineer to work, right? Because there's just a lot of automotive companies and things like that. But at Airbus, it is the— for engineering grads from top schools, it is the number one employer of choice, right? So it is really getting the crème de la crème of, uh, of engineering talent. And both the French and German education systems are very, very good, right? So the quality of engineering talent was extraordinary at at Airbus, and people were really passionate and dedicated to, to the mission, I would say. A bit more on the, on the negative side, it is also a hyper-political organization. It was originally created as a, as a, as a government, uh, as a government-formed entity with government ownership on across 4 governments: French, German, British, and Spanish. Spanish share is relatively small. The British exited their share Um, so, but there is still very, very strong, uh, undercurrent of sort of national identity politics, uh, that makes things difficult.

Yeah. And what I look— I mean, the thing that I know better is ArianeSpace, um, the European sort of similar, vaguely similar structure in, in a lot of ways. I mean, it's quite—

and is Airbus is a, is a partial subsidiary, right?

Exactly, part of Airbus. And, and, and kind of French-led, different countries chipping in. But, you know, I mean, Arianespace, oh man, you know, just a rough, um, sort of miscalculated on the SpaceX of it all, quite, quite dramatic.

As did ULA, right? To be fair, I probably like—

as did everybody. Yeah, the universe. And, and, you know, to SpaceX's credit, sort of pulled off things people weren't expecting. But I like, when I look now at the European space program in particular, it just, it feels, ah, there's part of me that thinks it's like a bit of a lost cause. You know, you've missed reusable rockets. It's like still barely even talking about it. They have a couple small rocket startups. Um, yeah, I mean, this part is, I was just in Europe last month and people are like asking me how this is gonna go and you don't want to just be some, it's not like I'm not even like pro, uh, I'm not, I'm not a super nationalistic person. You know, this isn't like a rah-rah US thing. It's just kind of like, man, I mean, maybe you guys want to focus on something else. I don't know if this is going to pan out for you very well. That when you compare Airbus, you said you wanted to work at Boeing as a kid. Can you— where do you have a sense as to which one you feel is more innovative between Airbus and Boeing and has its act together at this moment in time?

Well, I think— so look, I think that both companies have big company syndrome, right? That's just a function of being a very large organization that's existed for decades. Uh, Boeing is much older, obviously, right? So it probably has more ossification. Um, but, uh, but on the other hand, Boeing's, uh, internal politics are probably somewhat simpler. I did spend a little bit of time in Boeing actually in my career. I did a summer internship between undergrad and grad school.

Okay.

At, uh, working for a guy named Bob Liebeck, who is the father of the blended wing body.

Oh, okay, cool.

Down here in, uh, here in California in Long Beach. Um, uh, and so I got a little— and I, of course, I know a lot of, a lot of Boeing, Boeing execs. Um, so I think, uh, you know, I, I think both are phenomenal companies, right, that, that on the commercial aviation side have produced remarkable products, right, that have changed the course of of human history, right? The fact that we have like the modern commercial air transportation system, which was largely built by, by those two companies, is, is a marvel, right? The level of safety and, and cost efficiency in that, uh, the, and the, the resulting sort of globalization and interchange of cultural interchange, commerce, etc., that has, that it's precipitated probably rivals the internet, right, in terms of impact on 20th century human history.

Sure. Yeah.

Um, so, so I have deep, deep respect for both companies. I think both have had challenges throughout their history. I think Boeing is going through a challenging period. Kelly Ortberg, by the way, was a colleague of mine at United Technologies. He was the head of Collins Aerospace, which was one of the, one of the four businesses at UTC, and is— I have huge, huge respect for Kelly. I think he is a, he's a very strong engineer, but he's also just a phenomenal leader and a nice guy. Um, and I, I think he'll do great things at Boeing if, if given, if given the opportunity, if the board, if the board supports him.

Yeah, like, total tangent, when I was looking you up and I was reading about United Technologies, it traces its roots all the way back to Litton Industries, these, the Litton brothers that lived up here in Silicon Valley. This is wild to me, the California sometimes will, will surprise you, you know.

Well, you know, so, so the origin of UTC is actually United Aircraft and Technology Corporation, which was early 20th century, 1920s, 1930s.

This is sort of the era I'm talking about. Yeah, that's right.

Yeah. And it used to own— it was a, it was a, it was a full-stack aerospace company, right? So it owned Boeing and it owned United Airlines. In the early days.

Yeah.

And it was one of the first applications of the, of the, uh, antitrust, uh, the, the big Antitrust Act and movement. And, um, uh, and, and it was, it was forced to divest, uh, several of these businesses. And the decision was made along the Mississippi, right? It was not, not more sophisticated than that. So UATC, as it was called at the time, was allowed to keep everything east of the Mississippi and had to divest Boeing and United Airlines.

Okay. I did not know that. If you— okay, if you have to— when you look at the future of commercial aviation, I don't know, what do you see happening between sort of Airbus, Boeing, China? Like, what does the next 10, 20 years look like in your head?

Well, so I could give you a pre-AI and a post-AI answer, right? And it is entirely possible that like everything about the engineered world is going to change in the next several years. And, and, you know, frankly, the reason that I founded the startup that, that I did is to help, is to help that change, help make that change and help steer that change. Um, and so it is possible that, that we're talking about a completely different engineered world 10 to 20 years from now, right? And so, so, so my predictions predictions that I would have made a few years ago, or certainly when I was at Airbus, about what is the future of commercial air transport, as you suggested, are— would be like decadal-type predictions, because it takes, you know, like a new, a new generation of airplanes materializes roughly every decade. And so I could have probably told you what the next generation of airplane looks like and maybe kind of guessed a little bit at the one after that. Um, but I'm not sure that in a post-AI world all of that's going to hold anymore.

Okay. Well, I definitely want to get into that. All right. One, just a couple of real quick questions before we did. Now I'm extra intrigued. Just on this new generation of aircraft point, I mean, when I was writing about Universal Hydrogen, you guys were trying to change the fueling infrastructure for aircraft. And some of this would hinge on what the next sort of workhorses of commercial aviation would look like, the next generation planes from companies like Airbus and Boeing. I hope I'm not like outing you somehow. I mean, you were— you, if I remember right, you weren't super bullish on these companies' abilities to, to, um, kind of create these next-gen planes and to be innovative in how they were pushing things forward. Is that fair?

Yeah, maybe. So let me pull on a couple threads in, in your question, right? So, so first of all, the next, the next airplane, right, is probably coming in the next 5, 6, 7 years. I think right now Boeing doesn't have the capital and Airbus doesn't have the incentive to launch a new airplane program because Airbus is selling the A321 family like hotcakes and it's doing very well, very well for the company. Why would you cannibalize that? Yeah, right. And Boeing has a lot of sort of structural cultural issues that it needs to work through over the next few years. But probably early 2030s, there will be a new airplane. 3, 4, 5 years ago, if you had asked me, I would have said that there's a reasonable possibility that that airplane could be a hydrogen airplane. I no longer think that, not as a function of Airbus or Boeing, but actually as a function of hydrogen cost. And we can talk about why hydrogen cost didn't follow the curve that we thought it would in 2019 and 2020. But I don't think that hydrogen is going to be a competitive competitive aviation fuel anymore.

Okay. Which was— so the goal of Universal Hydrogen, it's been a while since I wrote about, so I might get some of this wrong, but I remember you guys were— you wanted to prove it out as a viable fuel source for planes and you were going to create an infrastructure where this could be delivered to airports to— you were trying to make it easy on the industry to switch over, right? So you were doing a lot. You were, you were kind of retrofitting some planes to show how they could fly with with hydrogen as a fuel source. And then you were also working on this infrastructure piece of actually delivering the hydrogen safely, effectively to airports and making it easy for the planes to refuel and swap out some of this infrastructure. Is that—

Yeah, that's exactly right.

In the ballpark?

Yep, exactly.

And so you guys were going at it. As I mentioned, it was pretty cool. You were in Hawthorne. You had stuff going on in France and in California. And you were in Hawthorne right next door to Harrison Ford, which is probably, apparently is like a possibly dangerous hangar to be next to if the internet is to be believed. So what— it didn't work out. Yeah. I mean, let's get into it. What, what, what happened?

You—

because I know you were so excited. I remember when I was reporting on you guys, you were so excited. I felt like you explained it really well. When I sat back, I was like, man, this is going to be a journey because there are a lot of things that have to fall into place to make this work. So why don't you tell us like how far you got and then what ended up happening?

Yeah. So, so as you correctly summarized, the goal was first to develop a hydrogen delivery infrastructure that would be easier than sort of rebuilding the current jet fuel delivery infrastructure. Right. So the idea was we modularize hydrogen, put it in these like Nespresso-style pods, much bigger, but, but modular pods and use intermodal freight, the existing intermodal freight network to deliver it. Right. So any airport that can take intermodal cargo, which is pretty much every one of them could take, uh, could take hydrogen and kind of plug it into the airplane as a fuel source. And hydrogen is a phenomenal energy storage mechanism, right? It's not really a true fuel in the sense of like you don't extract it from the environment. Well, there are— there actually are some, some pushes to try to get hydrogen out of the environment, but typically you create it, right? So it's, it's, it's a, uh, it's created— it takes a lot of energy to create hydrogen. But it is a very, very efficient storage mechanism, energy storage mechanism, which is why it's great, great for aerospace. And so we were creating these modular, modular pods and then, and then to sort of prime demand for our modular pods, we converted a regional airliner to fly on hydrogen and flew the sort of world's biggest airplane on fuel cells.

I remember that. Yeah, awesome proof of not just concept, but yeah, getting the work done.

Yeah. So product-wise and technology-wise, the company was a huge success, right? Like we delivered on every, every product milestone. We had a pretty clear certification path with the FAA for all of this. We flew in record time of, I think, 2 years from, from getting first, first, first financed and, and kind of everything was, was humming on schedule, on budget. We had an incredibly talented team, which I'm still very, very proud of. What all of this was predicated on was that the cost of hydrogen production would drop very, very rapidly, right? So when we were sitting in 2019, 2020 conceiving the company, hydrogen was pretty expensive and the expectation was that it would, by 2024 or '25, it would reach parity with jet fuel. And you might ask why.

I was going to ask why, right?

'Cause it's not obvious. It's not obvious that it should, right?

I made a fan with Lee over here. Excellent.

He'll be a regular. And so, so, so how do you make hydrogen? You make hydrogen by electrolyzing water. At least that's how you make green hydrogen. So hydrogen, you can also crack hydrocarbons to extract the hydrogen, but then you have the carbon left over. And so it's not a particularly clean process. The clean way is you take renewable electricity and you electrolyze water. So you break it into hydrogen and oxygen. Oxygen you can release, it's useful. Or you can, you can sell it for other applications.

I hear it's good.

And then the— it is good. And then hydrogen, hydrogen you can use to power airplanes or other things. And so the cost of hydrogen is principally the cost of renewable electricity. And the thesis, again, 5, 6 years ago was that there's this huge growth in renewable production and a lot of renewables you size for peak grid consumption, right? So you will build a wind farm to power, you know, a nearby city or region and you size it for peak consumption, you know, evening hours, right? Or midday heat or whatnot. And then at night, wind still blows, not as much, but still blows. You're still producing electricity and typically you would just dump it 'cause there's no massive grid scale storage, right? And so you could then take that off-peak electricity and produce hydrogen very, very cheaply. So that was the thesis. And most of the hydrogen cost projections were based on very affordable off-peak renewables being scavenged at near zero electrical cost, right? Because otherwise it was getting dumped, uh, for, for hydrogen. And then data centers happened, right?

So, okay, AI. Yeah, yeah.

And so, so that thesis did not pan out, and the cost of hydrogen production, of green hydrogen production today, is nowhere near that of, of jet fuel, just because we went on some mad energy consumption dash. Yes. And, and it is also— I mean, there's more nuance to it than that, right? Also the mix of, of renewables. So you, you have solar, for instance, right? And solar then actually doesn't get produced at night. Okay. So you don't necessarily get as much off-peak. So there's also a balance of what kinds of renewables go into it. But the bottom line is electricity prices for hydrogen production did not follow the curve.

Okay. And so, okay, so we're not gonna have a hydrogen plane industry, some massive hydrogen-fueled aviation industry. Anytime soon.

Uh, I don't think so.

Yeah. And you were not— if I remember, you weren't like super bullish on the SAFs either, these, these alternative fuels.

Yeah, SAFs are a little bit of a, a little bit of a shell, a carbon shell game, right?

Right. Uh, yeah. So what does that mean? I mean, that means we're— that means for what, the next 10, 20? Okay, well, we'll get into it.

Yeah, we can come back to the question of what happens to the aviation industry if things were holding normal ways.

You, you, it sounds like the future you're talking about would probably be just a, you know, quite a more still quite traditional commercial aviation industry for the next 10, 20 years.

I think there will be innovation. I think there'll be innovation in other aspects of it, right? I do. But I, yes, I do think that it'll be hydrocarbon powered. I think some of those hydrocarbons will be SAFs. I think a lot of SAFs are kind of a collective delusion. In terms of their cleanliness, right? There's no— it's an offset scheme, right?

Yeah. I don't want to go nuts on this point, but I mean, they get— they sort of get like a free pass as far as I can tell in the press. And you were kind of like, the math does not check out at all.

Certainly does not check out for SAFs. But it makes, you know, to the extent that it makes people feel better, great.

Right. Is it great? This is what I'm wondering. I feel like this is the, you know, any time the aviation industry is like, oh, well, we're going to be more climate friendly. It's, it's, it all, if we're not doing hydrogen, it all revolves around SAFs. And then you're not the only person I talk to who's like, this is mostly bullshit. So I don't, I don't know where that, that leaves. I feel like that leaves us at a pretty, you know, pretty much the same. Maybe the planes get more efficient and things like that.

But otherwise, I mean, I think the energy, the energy discourse has changed, like the global energy discourse has changed drastically over the last couple of years. Again, with the advent of AI and data center energy demand and, and things like that. Um, so I think the, you know, you don't hear people talk about the Paris Agreement as much anymore and about actually trying to hit net zero or true zero by, by 2030 because it clearly isn't going to happen because we're in an arms race. Uh, it clearly isn't going to happen. And I think the conversation around climate change has veered much more towards mitigation, right? Resiliency and mitigation.. And so that's probably the reality we live in.

Okay. Okay. All right. So Universal Hydrogen is winding down. You've been thinking about AI for many, many years. You're a hardware head. You're an engineering head. I mean, I guess I could see in my own head, you know, how this idea began to formulate. But why don't you tell me the origin story of P1?

Yeah, so the origin story goes way, way back. I mean, I grew up on hard sci-fi and reading the hard sci-fi. I was promised Dyson spheres and starships, right? And so when the AI revolution started to really take off in the 2020, early 2020s, I was like, I'm working on hydrogen airplanes, great. You know, I'm doing my thing, but I'm excited that clearly something is really rapidly, you know, there's been a step change, uh, it's going to rapidly accelerate. And, uh, and so clearly the biggest impact of AI is going to be in designing starships and Dyson spheres, or at least the built world, right, as a stepping stone to that. And I want to meet the people who are doing that. And, and I tried, except there aren't any people, and there weren't any people, and there still aren't very many who are doing that.

Even though you, you, you'd come from this world of DARPA and Google X and, and all these places that would be populated with like-minded type folks, I would assume.

That's right. And, and so the— but the AI world was very much working on sort of chatbots and, and, you know, cat— better cat video generators. I'm pleased to say that over the last couple of years, it's evolved much more towards, you know, medical diagnostics and drug discovery and those, those kinds of things, software engineering. But still very seldom do any of the frontier labs talk about physical, physical product engineering. And so I started poking around and being like, why?

I mean, robotics is kind of the closest they seem to be getting on the Frontier Lab side.

Sure. Yes. Robotics has been there and remains and maybe is the immediate next big focus area. But robotics doesn't help you build the built world. It helps you interact with the existing built world, right? But it is a different thing in kind. And so I asked the question of, okay, well, like, why isn't anybody working on this? Like, clearly I'm not the only one that thinks that that is potentially the biggest sort of impact that AI is going to have on, on, on our— on the human experience is through the built world. And the answer was training data, right? Like, if you want an AI, uh, you know, the, the, the enabler of this revolution in AI is large, very large training data sets. Yeah, and the reason that we were getting chatbots and cat videos is because there's a lot of language, uh, written, written language, and a lot of, uh, a lot of cat videos out there, right? And, and if you want an AI that, that can help you design an airplane, it needs to see a million airplane designs, and there just haven't been a million airplane designs since the Wright brothers.

Yeah, yeah, no, I can— this was— these were going to be my questions. I mean, so if you The two things, as you mentioned, that have taken off first, obviously we kind of mathematically solve language. We spent the last 40 years vomiting our every thought onto the internet, you know, along with just these corpuses of knowledge that existed. And then code, similar thing. We've got these huge repositories to lean on. These have been the things that have worked so far. I was poking around your website also. Okay, I mean, there's not a lot on the website yet. It's awesome. It's like—

Thank you.

I'm glad you like it. Interface.

I wrote it, I hand-coded it lovingly over a weekend.

Yeah. So at the moment it's kind of bare bones. It's just a little information about the company, but you do point to a couple of videos where you can see some of these early Archie demos. And yeah, I mean, so that as I was watching, going through them and I heard you talking about this a bit, I'm still not clear on where you would get the training data. So if you, if I'm gonna— let me back this up for one second. So, and I will out myself as knowing some things, but not enough. I mean, today if I want to go design an airplane, I am firing up something like ANSYS or something like that, right? I mean, I'm trying to build like a computer model of this thing. I want to get the materials that I'm going to use. And we do have software that gives us some of the properties of those materials. And then obviously we get to computer simulations and we can get kind of far. Right? And so what I was trying to understand is, is if you AI all this, what the next step looks like. Is that one right way to think about this?

Uh, so I think, I think you have to conceptualize physical product development, whether it's an airplane or some other, some other piece of industrial hardware, um, as a more of a sociotechnical system, right? So there are humans and there are tools, and there are things that humans are good at, and there are things that humans rely on tools for. What are humans good at? Humans are very good at taking a mass of requirements and figuring out what are actually the key design drivers among those requirements, postulating some solutions or some approaches to meeting those key design drivers, and then doing a first order, like, here's what the answer is going to look like. Then knowing what tools are available, what is the range of validity of those tools, and what are the inputs to those tools. Tools, like how do you drive the tools, right?

Yeah.

And so we are focused squarely on that cognitive automation. So what I just described, on doing what the human does, because there's no point in competing with the tools. The tools are there. The human uses the tools. Archie should use the tools, right? So we don't try to replace any tool. We don't try to encompass, supplant, compete with, right, make it— make any tool better. The tools are what they are. Um, there are a lot of them. So at Airbus, I think we had 700 to 800 active engineering tools on a new program, and the whole ecosystem was several thousand tools, engineering tools. So it is a very complex ecosystem. Some of those tools are Fortran that goes back decades. Some of them are very complex, modern sort of design software. And so the best thing that the agent can do the AI agent has learned, has learned what are the tools, when to apply them, and how to apply them.

Okay.

And then focus on that upstream reasoning piece that we still require thousands and tens of thousands of humans, uh, uh, in a, in a com— in a large industrial OEM, uh, to do.

So, okay, so if I'm, I'm, I want to be a— and I'm picking something totally random, so, you know, this may or may not be a good place to start— I, I am a rocket startup, huge debate inside the company. Are we going to build this thing out of carbon fiber or are we going to build it out of aluminum, stainless steel? At some point, yeah, you've got all these tradeoffs in these materials. I mean, this is like an active debate that goes on today and it is something that is, that is not solved. And people put their thesis forward as to why one course is better than the other. Where— so where would you— is that a play? Where would you plug in to this part of the process?

So initially we probably wouldn't, right? That's probably an architectural decision that a chief engineer at the company would ultimately make, right? And some of it may end up being data-driven, some of it may end up being a judgment call.

Yeah.

Or it is human bias, right? That isn't the first— that isn't the low-hanging fruit for Archie. Eventually, of course, we do expect and want to develop, sort of move up the the engineering expertise scale, but the low-hanging fruit is doing what junior engineers do.

Okay.

And junior engineers do a pretty rote, uh, sort of application of the process that I just described, right? Which is requirements, pick out key design drivers, do first-order sizing. And that sizing may not be of like the whole airplane or the whole rocket engine. You may be sizing a left bracket kind of thing, right? Or you may be sizing a pump, right? Or, or something else. Right? So, so it's kind of a fractal— everybody does those things, right? But, but it's kind of fractal, so you may be doing it at a different, different level of abstraction in within the system, the system design tree. Um, and, uh, and then, and then you, you, you have postulate a couple of solutions, right? You do a first-order sizing and then you go do— you refine the, the, the design.

So I'm picking a pump for a certain job and Archie's helping me Archie's helping me through this problem to, to, to size this thing right and pick the right, the right product.

That's right. And can go do a bunch of design tradeoffs, right? And say, okay, like, actually, can you build me a matrix, right? And, and, and try varying this versus that, right? And oh, by the way, we have a supplier that we are in a contract dispute with. So can we not use that supplier? Make sure that doesn't factor in, right? And give me like the minimum cost version and give me the minimum schedule version. Right, those, those kinds of design explorations that you would typically go assign to a, to a relatively junior engineer, that's what we're targeting as, as initial, initial capability. And there's a huge market for that.

Okay, yeah, no, I saw the example on the website. I mean, you have an HVAC system, it has 6, 7 core components. It's like, we're gonna, we're gonna change something here, what does that do down here? Or we want a certain amount of performance out of it, how do we get that without you know, what are the changes we can make to get that performance without, without changing the motor or something like that? And it seemed to be advising on these, these various tradeoffs and the cost benefits. It was pretty cool. I mean, you could even, like you said, I mean, you could get the actual parts list, see who supplied it, how much it cost, and then your options of what else you could buy.

That's right. And that was the toy sort of the toy problem that we set for ourselves in the pre-seed. It wasn't really tied to any particular customer use case. Case, we just picked a domain that was pretty multi-physics, right? So you have refrigerant flows and airflows and thermal interactions and electrical, right? It kind of has all the physics that a much more complex system like an airplane would have, but at the, you know, 10-15 component level. And, um, uh, and the physics are pretty linearizable, right? So you're talking about, like, if you're talking about an airflow, it's a one-dimensional flow through a duct rather than a three-dimensional flow over a wing. So much simpler to analyze. So that was our toy problem. The first application domain for us is data center cooling systems.

Oh, okay.

Which there's a certain sort of poetic elegance, right, to AI helping design its own data center. Right.

Okay. All right. Wait a minute. So you've gone from fighting the good fight for climate change and now, well, okay, so I guess you'll make us more efficient data centers. Is that—

I think you can, for sure, you can make more efficient data centers, right? I do think, like, again, look, I think AI changes everything, right? Like, it is really a technological discontinuity. And so I think our approach to the whole climate conversation needs to change, and appropriately, I think, is changing.

So, okay, so if data center design is kind of the first step, how do you actually get training data to make something like that effective?

Yeah, so this is sort of our, the novelty that we as P1 AI bring to the table table is the creation of these large synthetic training datasets for physical product domains that are physics-based and supply chain informed. And this is the idea behind this goes back 15 years to my DARPA days, actually. And so I joined DARPA right at the, at the beginning of the Obama 1 administration. And this was a couple of years, uh, just this basically just before the Joint Strike Fighter was, was fielded. And we were going through kind of all of the, the teething pains and the, the, the long tail of the cost and schedule overruns, uh, that the JSF experienced. And the, the, the big issue at the Pentagon was breaking the cost curve, right? Because if you plot the cost of like the F-16, the F-22, the F-35, it's a very steep exponential. And the question was, how do you afford the next one? Like, what does this mean? JSF was the most expensive thing that the human species has ever built, right? And was massive overrun both in terms of schedule and budget. And so I came to DARPA with a very specific proposal, which was we have like clearly the growth in cost is largely a function of product complexity, right? We're just putting a lot more and compactifying it in a very small space. In terms of electronics, engines that perform at higher temperature points, right? There's just a lot more. We expect more performance and that drives a significant growth in complexity. And as a result of that complexity, historically, the way we do physical systems is we build an engineering organization that mirrors the product. And if the product is more complex, the engineering organization is bigger. There's a lot more interactions in the product. There's a lot more interactions in the engineering organization and it doesn't scale well.

Right?

And but where, where, where has it scaled? What's, what's our existence proof that it can scale better than that? And the existence proof for me was semiconductors.

Okay.

Is the advent of VLSI design and EDA, electronic design automation, in the '80s led that whole industry. And this was kind of from the 8088 to the 386, you know, going by Intel, Intel processor numbers.

Yeah, this is what, right, like Carver Mead and Lynn Conway. Exactly.

Yeah, exactly. Yep. Conway and Mead VLSI design. And the, the idea was that you can raise the level of abstraction on the design process, right? You can, you can model the, the fab, right, with a physics library of like what, what could go— what can the fab make, what could go into your system. And you have these rich multi-physics models of the constituent components, and then you can assemble your system from those, from that physics library, and you would have a computable abstraction of your physical, of your physical hardware so that an individual designer could deal with that abstraction rather than dealing at the gate level.

I see.

Okay. With, with the hardware.

So this is like Cadence for, for—

That's right.

For all of the hardware.

That's right.

Okay.

And in fact, the founder of Cadence and Synopsys, uh, who was a Berkeley, Berkeley professor, Alberto San Giovanni Vincentelli, was, was an advisor to the DARPA program that I started. But I came to DARPA with this idea of like, hey, let's take VLSI, and try to apply it to a fighter jet, right? Like, so make it very multi-physics, much larger systems, much more nonlinear physics, much tighter coupling to the environment than you get in a semiconductor, right?

Did we do that?

So we did.

Okay.

I would say that the program was maybe 60, 70% successful.

Okay.

If a DARPA program is 100% successful, you didn't try hard enough, right? Like you weren't shooting high enough. Yeah. So, and a lot of the stuff that we developed permeated the industry. A lot of it was open source. A lot of it permeated the industry. One of the more bizarre sort of surreal moments in my Airbus tenure was, uh, from the Airbus, uh, business intelligence organization, which collects like open-source intelligence on what's happening, competitors showing me a slide that Boeing briefed at a conference. Um, and within it was like verbatim copies of my, my DARPA work.

Yeah.

And I was like, huh, this is very meta, right? Like, uh, but it's— but it really permeated, I think permeated the industry. And probably, frankly, you know, like, I did a lot of cool stuff at Airbus, but probably the longest lasting legacy was transitioning Airbus engineering into a model-based enterprise for the next airplane. Okay, right. And using this kind of like compositional model-based synthesis to design, to design the next airplane.

So, okay, this helps me understand a lot. Um, how— okay, well, but then how on earth do you use synthetic data to get to somewhere accurate on something like this.

Right. So, so back to— yeah, so, so sorry, I told you the dark story, but I didn't—

that really helped me wrap my head around it.

Yeah. So that was 15 years ago, right? Like I said, those techniques permeated— have, have permeated the industry to a significant extent, particularly sort of at the bleeding edge of complexity. So more aerospace, less industrial chillers. And, and so when the question arose, like, hey, so I need a million airplane designs, there haven't been a million airplane designs ever made to train my AI model. I was like, well, I kind of know how to make a million airplane designs that are going to be physics-based and, and supply chain informed. Because if the physics library that you're composing your system out of is tied to what's available in the supply chain, then every design that you compose will sort of by definition be realizable. Okay, it may never be realized, it may never have been realized, like historically, but it should be theoretically, if you wanted to go build it, you could build it. Now, some of those designs to train an AI model, you actually want to sample densely around the area where most of the work is going to be happening. And most of the work that engineers do are incremental changes to existing designs. So you want to sample very densely around those kind of dominant designs, but you do also want to sample sparsely in the corners and edges of the design space because you learn something, right? It's sort of like, uh, you know, we train college students when, uh, when they study engineering, we really train them on synthetic— mostly on synthetic data. Like, you may go build a one airplane, right, model airplane of some sort, uh, as part of your undergraduate education, but you do a bunch of designs, you do a bunch of practice designs, and that's synthetic data. And some of those designs might be wacky, and intentionally wacky because that teaches you about a corner of the design space that may be a terrible idea, right? But it teaches you why it's a terrible idea. And so you get sort of like the gradients of why it's a terrible idea. And so sampling like which million airplane designs you choose to make using these compositional model-based synthesis tools is actually pretty tricky, pretty tricky questions. And that's a lot of what we're working on at P1AI.

So you have smart are humans guiding that process?

Sure, the synthetic data generation process has a fair degree of human involvement. We're working on trying to automate some of that human involvement, but that's not a bad place. Like, we're not shy to have human involvement in that process. Think of it as a professor, as a professor guiding a student.

I didn't mean it as a knock or— I, I, uh, yeah, no, I, you know, I'm trying to understand how you get good synthetic data. Yeah, because it feels like you would want an expert in the loop on—

Absolutely. Yeah. And so we have a number of sort of domain subject matter experts who are helping us figure out how do you sample the design space? Are we sampling it the right way? And sanity checking a lot of the designs that we produce. They can't sanity check like a million designs or 10 million designs, but they can sanity check some of the seed designs. In the space.

And then there's like a reinforcement learning aspect to the synthetic data on all those designs. Is that—

so it's a, it's a, it's a good question. So the synthetic data itself, right, is kind of AI agnostic, right? So it's just a data set of a million designs and a performance vector for each design, right? And by the way, the performance vector— so we simulate each design to make sure that it's feasible, it is in fact a physically realizable design, um, and we, we have a performance vector so you can train on a design and the performance that design, right? Because the design process is a mapping from the design space to the performance space, or from the performance space to the design space. And the good news, by the way, is that you can use fairly low-fidelity simulation for generating this, because again, we're trying to replicate cognitive— we're trying to do cognitive automation, replicate the human cognition, and that is low fidelity. So a human does not solve the Navier-Stokes equations to model the flow over a wing, right? The airflow over a wing. Humans have a fairly low-order surrogate model for how to estimate lift on a wing or drag on a wing, right? And so we only need to be at that level because we know that there will then be a Navier-Stokes solver to do the detailed design. Okay. Right. So, so we produce this synthetic dataset. And so to your question about reinforcement learning, then, then the question is like, what do you do with it? If you just throw a million airplane designs at an LLM, it does not magically become a good engineer. Yeah, right.

Yes.

And so the AI architecture that you, that you train is, uh, is today a federated set of models that are good at individual parts of what comprises engineering reasoning, right? Because there is spatial reasoning, there is performance estimation, right? There is the multi-physics aspect of it. And so we have dedicated models for all of that, and then we do use a big LLM as a reasoner, kind of as a front end that knows about all of those other neural models as well as all of the non-neural tools that a human would use.

Okay. And you're, you're basing this off of a, uh, like an open-source foundational model for that part?

That's right. We do use OpenWeights models because our customers don't generally want their proprietary— because eventually— so, so we built synthetic datasets, uh, to, to to initially train the model. But of course it gets better as it encounters real data.

Okay.

From, from a customer and the customer data, we want the customer to have complete control over. And so for that reason, we do use open weight models that we self-host.

And I, and I'm sure— so you're starting kind of with this low-level engineer type jobs. I'm sure there's, you know, a giant path where this could go on that.

The—

if I was going to go in a data center design is part of this. It sounds to me like part of this is, is It's helping with tasks much in the same way it sounds like with coding. You're kind of having the AI take on some tasks that a human normally does, and then that's useful for all sorts of reasons, and the humans can focus on some other stuff. Is it also like an iteration thing? I mean, you know, if I'm thinking about building the new Stargate, and I just want this in the most innovative way that this AI can because you've trained it on all these models, that it can just run through a lot of design choices a human doesn't even have time or the inclination to think of? Or are we not— is that further down the line?

No, for sure, right? And, and it is faster than a human. Um, we— what we haven't quantified is the error rate, right? And that's next for us, is, is we're going to put this out in the wild. Uh, and by in the wild, I mean in an actual engineering workforce with one of the industrial OEMs that data center cooling systems. So these are industrial chillers basically, and let them use it. And we want to quantify the error rate.

Okay. Because that's my big— it's still my massive question around AI. And I don't want this to sound as cynical as it does because I sort of feel like we're in such early innings with this technology. But it's like, you know, the AI can write. I mean, sometimes Claude writes really nice things. I've seen Sam Altman demoing a new version of OpenAI's writing tools that seems much more sophisticated. But, you know, when I ask it to write something for me, like, on one hand, it's miraculous. It's like, oh my God, you can do— asking you to combine Shakespeare and Hunter S. Thompson and describe this bar of chocolate and you can do it, right? So hats off. When I ask it to write something for me, I'm, you know, I wouldn't publish it, right? I'm still like, no, that's not— I feel like I write Right. Better than that. It sort of lets you down. Robotics, it's like, does awesome stuff, kind of lets you down a little bit. Coding does awesome stuff. But if you talk to the world's best coders, they're like, no, it's not replicating what I do. And I know you said where you're starting, but feels like we're early days with this technology and there's just a lot of belief that, that we just keep throwing GPUs, we keep improving this technology. We do get to some new level where, um, it isn't making these mistakes and feeling still a bit rudimentary?

Yeah, I think coding is the best analog, right? And, uh, and sure, you know, I don't, I don't think anybody claims that, that the AI copilots or agent or the more autonomous agents like Devin fully replace every, every kind of software developer, but it is, it is seems incontrovertible, right, that they are certainly replacing the low end the low end of that profession. And so if, and if you think about it, you know, so, so developing a semiconductor is like coding in VHDL or Verilog, right? Those are languages where you represent, represent the hardware and the physics of, of what happens in the microprocessor. And in our case, we have similarly sort of a language programmatic representation of a physical system with, with all of the different physics domains. Again, air flows, fluid flows, electrical, thermal, et cetera. And, and so it doesn't seem all that crazy that it should be able to write programmatic representations of physical systems as well as writing, you know, programmatic representations of algorithms.

Okay.

Right. Doesn't seem like such a great leap.

Okay. Okay. Let's get weird. I mean, you were saying, you know, when I asked you about the airplane question of what these look like, I mean, it sounded like where you were going to go was that you think if these tools get good enough, they start to think of things we really haven't thought of and start pushing the field. I mean, expand that thought. I know you were talking about Starship. You've been— this has been a quest since your childhood. I mean, you wanted a craft where we would be exploring through the solar system and the universe, right? I mean, this is like at your core.

Yeah. I mean, like I said, I grew up on hard sci-fi.

Yeah.

And I'm just excited that I think that starships are within the realm of attainability during my lifetime.

Yeah, I mean, so let's do the thought exercise. I mean, it sounded— tell me when I was asking that question and you're like, I could give you two answers. One is the non-AI answer and one is the AI answer of where someone like a Boeing and Airbus is going. What does that AI answer look like to you?

Yeah, by the way, I think I undersold the non-AI answer, right? I do think that there will be cool things in the next airplane. Like, I think it'll be single-pilot capable, for instance, right? Like, I think there'll be incremental degrees of autonomy, uh, better, better cabin experiences, uh, more efficient engines, etc., right? Just to be—

that's just, just for the record, right?

Um, but yeah, I think there could be, there could be a complete discontinuity. Uh, I think it's difficult to see past that discontinuity, um, but, uh, but it, it does seem plausible, right, that you get, uh, you get a superintelligence that can help create things that we cannot create today, right? Design engineering systems that we cannot design today. Um, and that's the promise, right? That's, that's the exciting part. You want me to tell you what they look like?

How confident, how confident are you on that? Okay, so let's take rockets, right? I mean, SpaceX has done amazing things, but not so much really because it rethought how the rocket it is designed in mass, I mean, reusability, yes. But otherwise, you know, the physics of when the Falcon 9 first came out is like a lot of rockets that came before it because you have this tiny margin to play with. We know what happens when you put liquid oxygen and kerosene together and there's only so much you can do, right? The tube has to be, the tube has to do certain things. The fuel can only do so much. And so, And AI comes along, I mean, is it that much smarter that then it's got like a rocket design that we just never even thought of before? Or aren't we kind of constrained by physics on a lot of this? I don't know. What do you think?

I think that there are other ways of getting to orbit than chemical reactive propulsion.

Okay, so I shouldn't fixate on the rocket.

I should go— For instance, right? I mean, I'm giving you examples. Obviously, I am not super intelligent, right? I am just moderately intelligent. So I can only conjecture, right? But one direction, and I know you and a previous guest talked a little bit about non-traditional propulsion technologies, right? The kinetic stuff to me doesn't compile. We looked at a lot of this at DARPA in my time at DARPA. The kinetic stuff doesn't quite compile for me, right? Because you basically have to be at orbital velocity, but like, at sea level. And so you get, first of all, you have very high Gs, right? To even get to orbital velocity in a very short period. And then, but then you're moving at Mach 20 in like sea level atmosphere. So that's not really great thermally. And then you have to circularize, right? You have to have an injection burn, also unclear how you do that. So you have a little motor, but then your rocket motor has to survive some unfathomable Gs, right? There's a lot of issue, unless you're throwing slugs into space, right? Right? There are a lot of issues with that. But beamed propulsion, I think, does have a lot of promise and hasn't had a huge amount of investment into it. So we're talking about a laser beam or a microwave beam where you can accelerate— you can have a gradual acceleration profile, right? There are questions about focusing the beam and atmospherics and birds and all sorts of other things, right? So there are a lot of engineering challenges, which is why it hasn't been been, uh, scaled up. But conceptually, it makes, it makes a lot of sense, right? And any, any propulsion mechanism where you can leave the big energy source on the ground, right, for sure is appealing.

Yeah.

Um, uh, and so that could, that could be a direction, right? Um, certainly you could imagine fission and fusion as alternative energy sources for, for propulsion, right? So I think there are many directions that have been limited by just our our capacity for engineering as a species.

Right. Okay, so I was kind of approaching this from the wrong direction. It's not do the same thing better, it's do new, new things. Um, okay, give me your, like, things go as well as they possibly could, you know, both for, for you and your vision with this company and also AI. I mean, this is total— this is going back to your sci-fi roots. I mean, seriously, like, what do you think 20 years from now? I don't know, pick number, 20, 40 years, I don't care.

I think it's much less than that.

Much less than that. And what? And what? The material world is drastically changing?

I think it's drastically different. Yeah, I think, and look, I think the key to it ultimately is going to be recursive self-improvement of these models. And there are some early and promising results, and there are big questions about how do you steer that recursive self-improvement so that the so the models sort of don't get away from you and learn, you know, Asimov's laws of robotics, which I think are important.

Yeah.

Or well thought out in their time. But I think we will, like, there's no, I see no theoretical barrier to an AI that can self-improve. And once it can do that, you will hit a very steep, like a much steeper exponential than the one we are currently on, right? In terms of the capabilities of those models.

And you think it's much sooner?

It's hard to predict, but I, I think at the pace, at the current sort of human-driven pace of model self-improvement, maybe it's 5, 10, 15 years away, right? But if you hit it, once you hit that self-improvement, uh, moment, I think the slope of your exponential changes drastically.

So our previous guest was Zach Abrams, and he comes from the fintech world, and he was talking about, I don't know, we could be on the verge of the wealthiest being on planet Earth being an AI that's started a company. Yeah, I mean, unless I'm wrong, I mean, I'm kind of hearing it sounds like you think, um, I mean, maybe the next great airplane company is, is started and run by an AI almost. I mean, it could be, right?

I, I think the, uh, you know, I, look, I, again, I think that there will be this, this moment that, that Vingy called the singularity, beyond which it's, it's, it's almost impossible because it is going to be so sensitive to sort of initial conditions, right? It's almost impossible to predict what is on the other side. Um, but, but I think we're going to live through it, and I think that there is an opportunity for us for sure to try and— I think it's inevitable, right? Um, the question is when, when does it happen, and, and can we make fine course adjustments as we go through to ensure that what emerges on the other side is in fact a, you know, a cornucopia utopia versus a, uh, we all get turned into paperclips.

So I don't ask everyone this, but you seem like the kind of person that has thought about these things. What, what is the ultimate goal of all this? Is it, uh, when you're thinking about Starship as a kid, I suspect you were thinking probably more about human exploration of solar system and things like that. Is, is that the goal of all this, or is it just to push some intelligence that we helped create out into the universe? I don't know, what would you—

it's a deep, deep philosophical question, right? Um, well, first of all, I, I am, I'm a big fan of von Neumann probes also, right? So yes, I love, I love the, the vision of generation starships, right?

What is, what's a von Neumann probe?

A von Neumann probe is a, is a self-replicating a spacecraft, right, that can sort of perpetuate itself, uh, travel, right, then reproduce and build, build newer versions of itself.

And I assume von Neumann actually came up with that, with the idea. He came up with a lot of things. Okay, okay, okay, okay, wait, so sorry I cut you off. So you're into that? So that would be the sort of spreading intelligence through the—

Sure. And is that human intelligence, right, or is it, is it intelligence that humans help build? Yeah. It's an interesting question. Yeah.

Well, I— this is a philosophical— I mean, we could sit here with beers for many hours talking about this stuff. I feel like, you know, the last— I don't know. I feel like most of the people who maybe grew up watching the Apollo program and it all got into space and NASA and all these, you know, when I was growing up, it tended to be human exploration. Of space, and that was the spirit of all this. It was that part of us that wants to explore, and I'm sure some of that still exists, but I can feel it in my bones right now that we're kind of seeding, you know, mostly the like, can we send some intelligence that we at least had some part in off into the universe? That seems like the central quest to me that we're gravitating towards.

Sure, and I think there's also a question of what does it mean to be human?, right? Uh, and perhaps an intelligence that's trained fundamentally on, on the, the human experience, interacting with humans, and on the human, uh, body of, uh, of creative work, right?

Maybe it is.

Yeah, maybe, right? 4chan, um, right? Maybe, maybe that is, maybe that is a human intelligence, right? It's just non-biological.

Yeah.

Or maybe with technologies like Neuralink, right, maybe there is a hybrid of, uh, of biological and non-biological human intelligence.

You're pro or pro-con?

I'm pro almost all technology.

Yeah, yeah, yeah. So you would, if it got good enough, you would consumer elective surgery at Neuralink?

100%. Yeah, 100%.

What would you want to do with it?

Well, I think I, like, I, I think, I think the idea of, of an uploaded intelligence is, is very interesting, right? I think the idea of, of potentially merging multiple Yeah, uh, intelligence is also, also very interesting. Um, and I don't know, I would be, I would be excited to explore.

Yeah. What do you— like, there's that Ilya quote where he's like, you know, if you valued intelligence as, as like the main, as one of the main properties of humans that got you excited, those days are done. Um, yeah, I find this one— do you get the spirit what I'm saying?

Well, I, I think so. But again, I, I think that, uh, that the, that the artificial intelligences that we're creating are not substantially less human, right? Because they're trained on the same, on the same things that we're trained on.

No, no, no. Yes. And if you're— if we're talking about some sort of merge, I mean, this is a little to the side of what I was—

but even if there's a bifurcation, right, you just have a biological and a non-biological version, right? But it's trained fundamentally on the same, same thing, same data set.

I do find— I do struggle with this. I've got two teenage boys and it's like, oh, you want them to do well in the world. And, you know, you get fooled into these. It's like, oh, he did well on the SAT. Good for you. And then I just— I do totally feel like we're entering something where it's like, good luck. I mean, it's like intelligence seems like the last thing, one of the last things that you can sort of be openly It's like one of those human— most of our human traits you're really not supposed to like hold against each other these days, you know. And, and but intelligence, it's like okay to be smarter than— so it's like, oh, I hired this person because they're smarter than the other one. You can, you can, uh, hold that against people. I just, I feel like all these little judgment calls are about to, to go away.

What are you, what are you urging your kids to study?

I'm so curious. I battle with this a lot, you know, especially because I'm in Oddly, I'm in the fortunate position of getting to interview all these people like you and all the people at the Frontier Labs. And we live in Silicon Valley. And so hopefully, God, if anyone has some insight into where this is heading, but I'm at a loss. I mean, he's— my oldest gravitated towards computer science and I'm not going to stop him because it's great on so many levels. He's thinking through things, he's learning this stuff. But then I'm like, man, where does that— I don't know where ends up for you. I feel like— I don't think I've ever heard anyone give satisfying advice on this because it usually chalks up to you have to be adaptable, you've just got to go with the flow, adjust.

Human creativity, right?

Yeah.

Which AI will never replicate, which I don't buy.

And it's like, some of this might be true and it's great to say, but it's like, how do you— I don't know, how do you actually impart that? I mean, you know, and then for a while I was like, geez, is you should just, let's go farm, you know? And then, and then now everybody's, now I'm seeing Optimus and some of these androids that I'm like, ah, that's gonna get pretty good pretty quick, I think. I don't know. I thought our hands would, would hold out a little bit longer than I think they're going to. What would, I mean, sorry, it's such a stupid question. Like, yeah, for what would you tell a kid to do?

I, I don't think I have any good advice, right? I mean, I probably would say, frankly, I, I probably would still at this juncture say computer science so that you understand how these systems are built. Yeah, right.

Yeah, yeah, yeah. God, it's wild. Well, Paul, thank you. I want to be mindful of your time. You've been so generous. And so just to cap things off where you are with P1, it sounds like so this year the goal is to get out with a customer, start testing the software, and you're going to go data center chillers first and then, and then just keep pushing on.

Yeah. Move on to other verticals. Right. And eventually we want to end up with aerospace and defense, and then starships and Dyson spheres.

Okay, okay, I'm excited. Um, I get the first ride on any— this is part of the contract you signed by coming on here. I get to get off this place. Um, thank you, Paul, so much for your time. I appreciate it. Thank you, Lee. I made a friend.

You sure did.

Is this— is this like normal, or did I do really well?

He likes smart and kind people.

Ah, well, there That was the right answer. Thank you, Paul, so much for coming by. Appreciate it.

My pleasure.

Thanks, Ashlee. Core Memory is hosted by me, Ashlee Vance. The show is produced by David Nicholson and me. Our theme music is by James Mercer and John Sortland, and the show is edited by the John Sortland. Thank you all. All for listening. As mentioned, the show is sponsored by Yuan Ventures. Thanks.