Data Centers in Space, AI Excavators & Fixing AI Slop | Philip Johnston, Boris Sofman, Spiros Xanthos

19:51Spiros Xanthos:democratic part, somewhere between ICE, the Immigration and Customs Enforcement Agency, and Iran is AI, artificial intelligence. Notion is the AI-powered, connected workspace for teams. It brings all your notes, docs, and projects into one space that just works. And with AI built right in, you spend less time switching between tools and apps and more time creating great work. And now with Notion's custom agents, busy work that used to take hours or never got done at all, runs itself. Custom agents automate all of your team's repetitive workflows and they live inside Notion already. Maybe you want to keep track of what everybody's working on.

20:32Spiros Xanthos:Maybe you want to see which pages are getting edited. Maybe you want an analysis of what your team is working on. I'm constantly, constantly getting disturbed by pings and pings from Slack, by team members with all these questions. Now, Notion's Q &A agent can research the answers from anywhere on the platform and get back to those people directly in Slack. You can design custom agents on your own, but Notion has a bunch of pre-trained ones ready to go. Try custom agents now at notion.com slash twist. That's all lowercase letters, notion.com slash TWIST. And when you use our link, you're supporting our show and keeping it free and vibrant.

21:12Spiros Xanthos:Notion.com slash twist. survey came out as well, saying that companies, and I think these are ones with 500 million plus in revenue, only have a 9 % plan to cut jobs due to AI. I think this is the key part of this survey, while 55 % expect to increase hiring. And the reasons they give in this KPMG study, reasons people don't trust AI include fear of misinformation, privacy, data concerns, lack of transparency, job displacement, lack of human oversight, and a general trust of corporations. So let's talk a little bit about the trust issue. Spiros, we'll start with you. Why are Americans so concerned here?

21:58Spiros Xanthos:And then how do you think about job displacement, which obviously seems to be a big part of the uneasiness, as well as what should our industry do to communicate, hey, there are benefits here for society, and are we not doing that enough?

22:10Philip Johnston:First of all, I think I completely understand why this is a sentiment that people have around AI. It is very disruptive. It has actually a huge potential impact on jobs, and I don't think the world is prepared for what to follow, right? I am a huge believer and a huge optimist in AI, right? And I think that the net benefit of the world is going to be significant, right? One of the biggest maybe advancements we've seen in technology, definitely in our generation. So, but I think it comes with a lot of downsides and it comes with a lot of downsides for people who maybe are not involved in it, right?

22:42Philip Johnston:And I don't think we do enough to probably explain or think through what the world is going to look like in the next three years. Like models advance at a pace faster than maybe even those of us who are in expected, right? This means that a lot of the, I guess, you know, we have a physical world and virtual world here, right? I think it means that a lot of, you know, white-collar jobs are going to be highly automated. And it means also a lot of the physical world jobs are going to be highly automated, right? And of course, that's desirable, right? And probably it fills a gap, like Boris was saying.

23:15Philip Johnston:But, you know, we have to think through the implications of all of that, right? And, you know, I don't think we as an industry do much to think that through or explain to the rest of the world, to be honest. I was just going to ask Spiros what he thinks we should do to explain it better. I mean, yeah, it's quite a difficult one to convey to the general public. Is there anything you had in mind specifically to get the word out? We know we focus too much in what's happening inside the industry and not maybe enough in explaining what are the benefits of the world, right? I think the story what Boris was describing to me is a real one, right?

23:51Philip Johnston:Like maybe Jason was saying, for a generation, people stopped doing essentially this physical world, right? And now we're limited by that, right? Our economy is limited by that. And I would say in the world of software that I am in, I believe that AI is actually going to result in a lot more technology, a lot more software. It's going to make a lot of things easier and simpler and cheaper. Things that are not accessible to people today are going to be much more accessible, right? You know, like happened with every generation, right? You know, my phone or Jason's phone is a lot more powerful than the phone that anybody can have anywhere in the world, right?

24:22Philip Johnston:That wasn't the case maybe 30 years ago. And I think AI is going to become, like, intelligence is going to become actually abundant and everybody is going to have access to it, right? And the rest of the world is going to benefit as a result. But, you know, that's not going to happen overnight. The benefits are going to be seen over a few years, right? But I think we should be focusing a lot more on the benefits.

24:39Boris Sofman:If you look back at some of the revolutions that have happened, go back like 150 years. You have everything from internet, mobile, computer, industrial revolution, every single one of those, there was an immediate fear that the disruption is going to be absolutely massive. And it's always a lot easier to identify the places that end up getting immediately shaken up and displaced. But it's a lot harder to see the secondary side effects. And so mobile phones created this astronomical economy with Uber and Instacart and all of the other applications and services that became possible. Same thing with the Internet and computer.

25:16Boris Sofman:You go back to the Industrial Revolution, it was massively disruptive where almost every single type of traditional factory job or even farming job got completely kind of shaken up and everybody thought that there's going to be the end of labor. What ended up happening is by the end of it, it was disruptive locally, but when everything settled, the number of jobs actually increased, the average salary doubled, the productivity skyrocketed way more than that, and you actually had a big net positive for overall society. And so one of the elements here might actually be, like you said, Spears, like highlighting the wins where we need the drug discovery and the cancer treatments that end up being giant winds that were discovered for the first time by AI and not by scientists doing massive experimentation.

26:01Spiros Xanthos:How do you frame it at your company, Boris? And I'm curious, the folks you're selling into, I'm sure, are like, yeah, we need this because we're behind on these three projects. But then, and I don't want to stir the pot here, there might be in some countries people who have very strong union protections and say, hey, you can automate this, you know, excavator all you want, but we need to have a supervisor and we need to have somebody in that cabin. So automate it if you want, but we still want our two guys in there making 75 bucks an hour. What's the vibe like when you go bring this to market?

26:35Boris Sofman:Most of our partners are actually employee owned where they literally have like stock options in their own company. And so they succeed when the company succeeds. And the physics of these companies right now are that they're turning down lots of jobs because they physically can't stop the work. And so they see this as massively expansionary. And so instead of pushing their people to work 60 plus hours a week, where you start to get fatigue and safety issues, and then turning down the rest of the jobs, they're actually able to actually take on more work, utilize their equipment and their teams better.

27:06Boris Sofman:And so it becomes genuinely expansionary because the demand just exceeds supply.

27:09Spiros Xanthos:Which is very different than, say, Uber drivers and Lyft drivers who, you know, I was talking to Dara, a great Iranian American.

27:18Boris Sofman:Yeah, he's amazing.

27:19Spiros Xanthos:Amazing. And he was saying, yeah, you know, we're still growing in San Francisco, LA, wherever Waymo is, but we don't need as many drivers there to match that. And we're going to have our robotaxis. So we're just not doing driver recruitment in those markets because we don't want to make it harder on those folks. So then people see the writing on the wall. in Austin, Texas. I think we have five people testing right now, maybe six in this tiny little million person, you know, enclave, uh, midsize city, or even on the smaller side of cities. So it's super interesting, by the way, back to our other discussion.

27:52Spiros Xanthos:I, I just asked Claude to tell me what are the, what's the, the industry term, uh, for our previous discussion about automation. It's automation dependency and skill degradation, automation dependency, right? Create skill, degradation in aviation. That's what they talk about in aviation is this automation bias, the tendency to overtrust and defer to automated systems when a manual intervention would have been better. And so we see that in code, we'll see that in robots, and we see it in CAD, right? That's going to be the skill in some ways, Philip, isn't it? Knowing when to stop trusting this, knowing when to blindly trust the system.

28:31Spiros Xanthos:Hey, it's going to get it right. It's like a spreadsheet. two plus two is four. We haven't seen it make that mistake in a long time. But hey, when we're making this more complicated thing, then we should have human in the loop. And then, hey, for this, if we're making a specific bolt that goes onto this, you know, satellite, it's got to be human designed and then maybe AI tested. I'm just giving an example here. Yeah, 100%.

28:56Philip Johnston:I think that That will be where the sort of niche labor is going to be that skill set of knowing when you can use the models and when you need to intervene.

29:06Spiros Xanthos:Yeah. You know, having had this discussion countless times with David Sachs and, you know, he's in the administration doing AI. And I'm like on the streets watching startups and I see the startups just saying, this is unbelievable. We don't have to hire any more people. We can get to a million dollars with just the three co-founders. We're not going to hire. It's like too much time to hire, too much culture, you know, friction. We'll just build it ourselves. We can do it all with AI. And then on the other side, you're seeing the big companies say, yeah, you know, maybe we'll lay out 5 % of people.

29:36Spiros Xanthos:But we think AI is not ready for prime time. The startups are the most curious and they have the least amount of resources. So, of course, they will figure out how to apply this technology first, along with criminals, gamblers, and, you know, the people at the bar on Mos Eisley. I was explaining this to my wife last night, Spiros. I was telling her how much I'm into this one specific piece of technology. And I won't say what it is right now. She said, why are you so into that? I was like, well, I always look for two things. When these new technologies come out, what are the hackers and the tinkers, the gadget people, the GitHub people, the Reddit, Hacker News people?

30:19Spiros Xanthos:What are they like to play with? And number two, what are the gamblers I know? and the criminal element how are they adopting this technology because those two groups of people they're the earliest of adopters i was actually talking about stable coins actually i can say it and i was like you know this is why i know stable coins are like legit now the hackers are all triggering out how to save money and the gamblers i know selling poker disputes for us they're like just yeah send me send me tether i don't need anybody to know about this traffic send me zcash and I'm like what's Zcash? and they're like oh that's really anonymous if you really want to sell your poker debts use Zcash in this wallet nobody will even know it ever happened pretty hilarious

31:03Philip Johnston:by the way I don't even remember Jason last time I saw you were playing poker

31:06Spiros Xanthos:where were we? we were at the KOTU conference?

31:09Philip Johnston:yeah the KOTU conference

31:10Spiros Xanthos:yeah that was fun we had a good time that was a profitable event for me but I've been thinking about this in my debates with Saks because you know steel sharpened steel kind of situation. We battled over this over and over and over again. And I think I know how to solve the PR problem. I think I figured it out, Boris. The three things that the technology industry has had the least amount of impact on, because it's the most regulated here in the United States, we all know, healthcare, education, construction, Boris, we have you here today, so you're in one of those. Those three are where consumers just spend so much money.

31:50Spiros Xanthos:And if we could communicate as an industry, hey, your house is going to be cheaper to build and faster to build, so we're going to build more houses and they're going to be cheaper. Hey, your child's education, you don't need to hire a tutor. This adaptive learning, your AI tutor is going to really level the playing field with some rich parents who have tutors there all weekend long to do better on the SATs, yada, yada. And then on a healthcare basis, where are the startups saying, hey, come to, you know, this pod and this micro healthcare center, and we will do all of your health stuff, you know, very easily, very cheaply.

32:29Spiros Xanthos:And that's happening overseas. People go to Turkey, they go to Mexico, and they have these health retreats. Now, I don't know if you heard about it, but my wife, second reference here, went to one recently where she got some stem cells, she did this, she did that, all of that available, like, for the same equipment, but for lower price in a, you know, cash kind of way. So what do you think, Boris, if our industry could handle those big three, and show declining prices and more availability, this might turn things around?

32:58Boris Sofman:I think so. You're hitting on like some of the most foundational things of life, right? Like, what do people care about, you know, their health, education, shelter, food's the last one, right? And then you have a lot of the foundations of life covered. And particularly when the cost of living is going up, this is actually pretty meaningful. But these would be big wins. And ideally, it starts to impact everybody where we're feeling the impact as drivers of the businesses that go and develop products, but that doesn't you know, broaden out to the rest of society immediately, it has a delayed effect.

33:37Boris Sofman:These would be astronomically valuable, particularly around education, when you have such diversity of availability, even within a metropolitan area. Yeah, it's, I do worry because there's a backlash that could become bipartisan against technology over the next five, 10 years if this isn't actually thoughtfully managed in the right way, just because it's such a lumpy impact in terms of parts of society. And so it's important to actually celebrate these wins and actually try to accelerate the ones that impact the overall foundation of the country.

34:15Spiros Xanthos:Philip, we had the administration, which is obviously super engaged in AI and the build-out. They said, hey, we're going to get this agreement done. And I saw Sachs announce it, I think, last Monday. Electricity is not going to be impacted. Your electricity price will not be impacted by data centers. So that's kind of a reassurance of a fear. It's not aspirational. But at least I think the industry is starting to get attuned to, let's lower the fear about some things. And I think the next thing we all agree here is maybe amplify the benefit. But what's your take here on our PR crisis in AI? And is it going to get worse, better?

34:54Spiros Xanthos:And what should we do?

34:55Philip Johnston:I think that definitely goes a long way to helping it. Like if people started seeing their energy prices going through the roof, that's really would have been a disaster for the PR thing. I think one thing we need to look at as well is sort of Andrew Yang style universal basic income, you know, and in some senses, we already do it. I mean, I don't know if, what was this thing? Brad Gerstner has this, you know. Oh, the Trump accounts, Invest America,

35:26Spiros Xanthos:which they brilliantly rebranded as Trump accounts and magically they sailed right through. Very interesting, very interesting. By the way, StarCloud is now Trump Cloud or StarTrump. Yes, you've just got all your approval. you've just been approved to put a hundred thousand satellites in space

35:46Philip Johnston:i will very happily rename it to trump cloud in that case i mean it's just an idea djt cloud you never know things could go easier for you with the trump accounts i mean that's essentially a form of ubi um and i think there'll be more of that over time and they'll be financed by taxing essentially tech companies i mean that's basically how they're financing these things now

36:11Spiros Xanthos:So controversial and Philip that people are talking about outside the United States because there's been now Paradoxically or randomly Wuhan in China has had protests over self-driving and there's been signaling from the CCP That they're going to give out a certain number of licenses and they're going to tax them in a different way for self-driving cars Then New York State said hey, you can't get professional advice from your large language model and I was like okay, that doesn't make any sense. Like legal advice or health advice, why can't a large language model present some information with a disclaimer that it's up to you to check it?

36:50Spiros Xanthos:Just like a podcast or a book. But we're starting to see the regulations. Boston also, I think, was a little anti-self-driving. So it's bubbling up. But Spiros, what do you think is some way, the proper way for the government to be involved, if at all? Or are you just, hey, free market, let it rip, we'll figure it out afterwards with unemployment, et cetera. Because we have done that in the past. When we had COVID, we extended unemployment, I think, indefinitely for a period. You know, we can react in the review mirror, but any thoughts on taxation and other systems for thinking about job displacement?

37:27Philip Johnston:Yeah. So first of all, I think that, you know, starting maybe with my perspective on this, like I think AI is the way out of our trouble as a country, in my opinion. And I think like, Like, you know, all other discussions aside, you know, if we didn't have AI, we would be in much deeper trouble than we are today, you know, the government included. So in that sense, I think we would need to lean in and we need to, you know, the government should, for the most part, get out of the way and, you know, let the industry work because we're playing at a global scale as well, right? It's not just what happens inside the US, it's if the US falls behind and somebody else essentially wins.

38:02Philip Johnston:And that's not the future we want to some extent. So now that said, obviously, you know, we need to be very thoughtful of what happens with job displacement. And, you know, I don't think like taxation and slowdown is the right answer, right? The right answer is probably just letting the technology evolve as fast as possible. And then, you know, figuring out ways to deal with what happens to people who are essentially on the wrong side of this, right? And we should actually doing everything we can for those, right? Like the simple example we gave with electricity is obviously a very simple maybe thing to do.

38:33Philip Johnston:And it makes a lot of sense, right? And I think there are probably other similar things we have to think through, you know, especially with real world automation, right? And, you know, robot access and all of that, right? What happens to everybody who was making a living out of that? But I don't think the answer is stop robot access, right? The answer is obviously let robot access happen and then deal with, you know, or help all the people that maybe lost their job as a result, right? In a very short period of time. You know, that's my perspective. Now, I mean, I don't know if I have the answer what needs to happen in terms of like the actual implementation of something like this.

39:02Philip Johnston:But I do know that we cannot slow down technology, right? We're going to pay the price if we do that.

39:05Spiros Xanthos:All right, let's talk a little bit about real world models. We have seen massive disruption in a number of industries because of, you know, guess the next word. Models, co-pilots, legal. My Lord, what you can get done with these things is amazing. And agents being able to use them to navigate the real world or to navigate the online world has been phenomenal. But Jan LeCun, I believe is how it's pronounced. Lukun. He's raised a billion dollars to bet on world models. His new startup, Advanced Machine Intelligence, raised a billion at a$3.5 billion valuation. Largest seed round ever in Europe. And he wants to build real world models.

39:51Spiros Xanthos:Obviously, Borussia operating in the real world and FSD is operating in the world. And then there was a second story, Figure, which is a$40 billion dollar startup that maybe has one or two customers. BMW is a little controversy around that. And we can play the video here. The founder, I believe it's Brent. Is it Brent or Brett? He shared this video. Elon responded and said, is that remote operated or not? And it created a little back and forth. But here you see this robot just tidying up, spraying some Windex on a table and extremely slowly and lethargically wiping it in a circle, wax on, wax off, and then throwing it over his shoulder.

40:37Spiros Xanthos:Or their shoulder, its shoulder. I don't want to misgender this robot, but they, them, seem to be doing, they're the slowest housekeeper ever in the world, but this will be the worst it's ever been. Boris, when you see that, what did you think? And then just generally world models. Obviously, NVIDIA and Jensen are building some open source ones and trying to just help everybody understand the real world so they can build apps. What are your thoughts here when you see this video?

41:04Boris Sofman:It's impressive to see the sort of progress that, you know, human beings in general are making. But it's always very hard to know in a video like this how much has been optimized for this particular setup and how versatile is it when you get into, you know, general households. because that's where it actually gets hard when you're thrown into the infinite permutations of what the world actually offers. And you see incredible demonstrations like this from a number of companies that then have trouble on the last 5 % or 10 % that makes it still pretty far from a commercializable product.

41:35Philip Johnston:What did Brett say? Just to jump in, what did Brett say when Elon asked? Did he say it is teleoperated or he said it wasn't? Probably.

41:41Boris Sofman:He couldn't have possibly said that.

41:42Philip Johnston:He said it wasn't. Yeah, he said it wasn't. If it isn't, it's very impressive. Yeah, but there is the question of is it teleoperated? Has it been optimized just for this particular video also, right?

41:51Boris Sofman:Because a lot of these companies have almost like shops of people like teleoperating robots in order to build the data, and then it can be hyper-fed to a particular application. Oh, I see. And then you go and you do it autonomously, but you're learning from a lot of demonstration. In a lot of ways, it's interesting because an LOM has become so generalized when it's applications, but it's partially because there's this infinite amount of data with relationships between words that have existed in a fairly similar kind of context and feature space across all the use cases on the internet. And so when you then tune it for law or for medicine, there's a lot of carryover.

42:29Boris Sofman:When you think about the physical world, that data doesn't exist. And you also have the complexity of unique hardware, unique sensors, unique kind of applications. And so you go and you kind of teleoperate, you can learn any specific thing, but it's almost like hiring a team of writers to train open AI versus using the internet, right? And so you're, you know, putting drops in the ocean.

42:48Spiros Xanthos:Which, by the way, people are doing, right? That's scale AI. We're investors in Micro One. We seeded the company. They're literally hiring really talented people to find that last 5 % and iron it out and say, yeah, these legal questions, it's hard. And they're just like, hey, let's iron out this last little edge over here. But yeah, when I saw this, the first thing i thought was you know robo taxis and waymo operating in a very narrow constrained geography is this is the same thing right spiros just let's just train it on this seven by seven square mile area and uh get it perfect before we keep extrapolating and then even in that case with waymo and robo taxi they all have remote monitors some people say it's two to one three to one four to one i think baidu said it's like 50 to one or something they're

43:37Boris Sofman:already at the women keys is a bit different because even if like just take san francisco the amount of infinite chaos that can happen in a city like san francisco is almost infinite like the the horrors that we've seen is just you know you couldn't even begin to explain like we literally had like enough people jumping on the front hoods of our cars that we had rates behind it in every city there would be like a civility rating on a on a city right and oh really at way It's wild, yeah. Singapore, nobody's jumping on the cars. No, it's like they're very disciplined in Singapore. Nobody's jaywalking.

44:11Philip Johnston:Where is San Francisco, Boris? Where did San Francisco rank?

44:13Boris Sofman:Way higher than Phoenix and then Austin, LA. Like literally, it would just happen.

44:18Spiros Xanthos:It was right above Calcutta and Baghdad. Yeah, exactly. Right there.

44:24Boris Sofman:But you can't escape the long tail safety challenges. And so anything that can happen in driving will happen in San Francisco. And so it is like a very broad solution. Just as a robot taxi, you have physical infrastructure, which is why you stamp it out by cities. That's a much broader scope in the driving domain than like the household kind of cleaning situation. But you have a very verticalized solution, right? So you have a deep focus on driving in this case. And so you have these like vertical solutions that require a giant amount of data and a giant amount of focus. This is where world models are actually pretty interesting.

44:59Boris Sofman:It's not immediately applicable. But when you think ahead, like, you know, some number of years, the big breakthrough they can normalize physically, I development might be a giant breakthrough in simulation where you actually have a realistic representation of the world where now data is not as, you know, you can break through this bottleneck of data in a more robust way. But there's a long way to go to get there. But it's such an incredible and tantalizing unlock if you can get there.

45:25Philip Johnston:Well, yeah, I think, you know, maybe Jason, what we've seen happening, let's say, with like RLMs, right, is that, you know, as we improved, like I said, the reasoning, then we applied RL to all sorts of problems, right? So they took a general purpose model that is very good at reasoning, gave it tools, and then how did it try a task as many times as it needed, right? Math being the simplest, I guess, right? Because you have an answer always, but we took other problems that maybe the answer wasn't that obvious and let the models play with it. And then, you know, we made them very, very good at one particular task, right?

45:51Philip Johnston:One particular task at a time. And now we've allowed them to write code on our behalf, right? Or do math or, you know, all sorts of other things. I suppose that's applicable to the real world as well, right? Especially with developed models that, you know, virtual, let's say, simulation models where essentially we can try things, right? So I think the future is inevitable in the real world as well, right? But I'm not an expert in that to know the timeline.

46:12Boris Sofman:Yeah. And the thing is, you just can't experiment that way in safety critical. like you wouldn't be able to do RL on a public road, you know, driving, but you can in picking things up. And, you know, there's been attempts at it. When you get into simulation, that's where you have some like incredible breakthroughs with some of the best applications of walking technologies and quadcopters and incredible balancing. That's actually a simulation train because you were able to actually simulate physics very accurately. And then you can do tricks like blur it where you may not know the exact physics of the world, but you're learning a superset of the world's physics.

46:45Boris Sofman:and then when you get onto the real world you're superhuman in your ability to walk and balance so

46:49Spiros Xanthos:to give an example of that if you're a quadcopter company you're archer or joby you have the simulation you understand the physics now you can do edge cases like what if we had a crosswind followed by a lightning strike followed by a bird strike and let's see if sully can still fly you know the the plane that's right my understanding my understanding of how you know

47:11Philip Johnston:self-driving cars essentially got trained is this right you had a model that essentially you know simulate the road and you know what others did around you right and then the the car had to

47:20Boris Sofman:react basically right isn't that the case yeah and it's it's a lot more heavily on imitation learning uh where uh you're capturing this super complicated relationships between all the things on the road pedestrians vehicles cyclists but um it's incredibly delicate and subtle because your safety is a function of your interactions with everybody else and so you have to model all these complicated interactions. But with enough data, you're able to capture these patterns in a way that's way better than traditional kind of engineered solutions with heuristics and search.

47:48Spiros Xanthos:Philip, you must be doing simulations like this or planning to do them for when you have, I don't know, how many thousands or tens of thousands of these satellites do you plan on having out there in space? And you have to now account for, I don't know, are there 7 ,000 or 8 ,000 Starlink satellites out there now? How many satellites are going to be out there and how do you account for all the collisions and possibilities. Oh my gosh.

48:14Philip Johnston:Yeah, we've just filed for a constellation of 88 ,000 with the FCC. Elon's just filed for a constellation.

48:20Spiros Xanthos:You said 88 ,000.

48:21Philip Johnston:88 ,000. So you're a little suspicious. You didn't go 8, 8, 8, 8, 8. Yeah, no, that would have been, yeah, we should have. 80 ,000 just wasn't enough. Yeah, the Chinese would be happy if we'd called it 88. um but uh no no elon's just filed for a constellation of a million um and it's all gonna be in this yeah yeah for a million of this for his ai satellite constellation um and so with that you can deploy we can deploy at least 20 gigawatts of uh capacity with that um i think elon's targeting uh 60 or 70 gigawatts of capacity with that so maybe 100

48:58Spiros Xanthos:the physics and i guess when that happens because you have so much i don't know how you i guess you would have to be low earth orbit or at least be mid-earth orbit how do you think about that because you you don't need to transmit all of them don't need to transmit down to earth and you i guess could hit your eye with starlink or you know bezos's and amazon's infrastructure there but is it low earth orbit mid-earth orbit what's the best place for these

49:25Philip Johnston:to exist it's low earth yeah the reason you want low earth is so that you can serve all inference workloads for example if you want to do you know voice agents for ai or hgbt or video generation actually yes most most workloads would be fine in medium earth orbit but um you know for video generation would be totally fine because it's going to take 10 seconds anyway to generate the video but things like voice agents you'd definitely want to have explain to us the

49:52Spiros Xanthos:how large that band is. Is it 100 miles wide low earth? Is it 500 miles? Is it 10 ,000 miles? I've never actually asked that question because that would determine, hey, there could be, if it's 100 miles, you could just give each person a mile and you could have a million in each of those 100 miles and you'd have 100 million satellites. How does it work?

50:14Philip Johnston:Yeah, exactly. So it's from about 400 kilometers. Usually people say low airfield, but it's up to about 2000 kilometers. anything above that to about 20 000 is medium medium orbit but it's a lot you can you can probably put at least 10 terawatts of um of capacity there that's 20 times the entire uf power grid um and then once you've once you've filled that up you can go to medium earth orbit you can go to cislunar orbit you can go to the lagrange points so there's an enormous amount

50:39Boris Sofman:of capacity you can deploy so can you like um i'm so curious about this like can you explain the unit economics of this where it's obviously more expensive to kind of get this out into space But then you presumably save a lot on energy, maybe cooling, although it's not super. What is the long-term convergence of this in terms of the advantages over Earth data centers?

51:02Philip Johnston:Yeah, so the main one is on energy and infrastructure. So the best comparison is with solar on Earth, because solar is the cheapest form of energy we have on Earth. So the biggest three costs there is, number one, the cost of permitted land. That's the largest cost, especially in North America. Number two is the cost of batteries and backup power. And number three is the cost of the solar cells themselves. So in space, number one, we don't need permitted land. So your biggest cost is gone. Don't need batteries and backup power because you're 24-7 in the sun. So your second biggest cost is gone.

51:27Philip Johnston:I need eight times less solar cells because one square meter of solar panel in space produces eight times the energy of one on Earth. The main additional cost we have is the launch cost. Yeah. And so there's a break-even point where the launch cost gets below the cost of permitted land batteries and solar. We see that's around$500 a kilo break-even. And Starship is targeting marginal launch costs for SpaceX of$10 to$20 a kilo. So yeah, it's well within range of what's coming.

51:53Spiros Xanthos:Just to recap that, getting to space is the cost. The energy cost is, it sounds like, is it 80%, 90 % less?

52:03Philip Johnston:Yeah, 90 % less. 90 % less on the, and then does that include the heat dissipation cost in the energy there?

52:11Spiros Xanthos:or is that like another level of energy that gets taken out?

52:15Philip Johnston:Just the infrastructure cost of what we're doing is around$5 million per megawatt. Terrestrially, if you build new infrastructure, you're talking about$15 to$20 million just for the infrastructure. So that's cooling towers, chillers, batteries, backup power, none of which we need. And so that's about 4X cheaper. But then you add in the fact that our infrastructure costs includes all of our energy over the next five, six years. So that's all of the solar panels and radiators included. so in total we're talking about 10x less in terms of both infrastructure and energy

52:44Boris Sofman:yeah so we designed the chips to be radiation hardened as well um or no we're using both

52:50Philip Johnston:standards so we launched an h100 uh in november last year and it's working uh remarkably well so far we just trained the first model in space and did a bunch of other things like that and those

53:01Spiros Xanthos:prototypes cost low tens of millions i think to make and put up we did this one for like two and

53:07Philip Johnston:a half million dollars it's this one that you the one you see on the screen here is two and a half million dollars so that was uh about 300 grand for launch for spacex and thank god for elon and spacex because we would never exist without them um about a million dollars for the bus and then another few hundred grand for the computer hardware inside and what will these cost at scale if you

53:25Spiros Xanthos:put a thousand of these up and then you put your ten thousand up what are they going to cost at

53:30Philip Johnston:scale do you think in five years we'll be launching these 200 kilowatt um three tons blades that fit on Starship. So we can launch 50 of them per Starship. So it's about 10 megawatts of compute per Starship. Just the infrastructure on that is about$50 million. If you were to do that on the ground, it would be about$200 million. And then obviously we don't pay for energy over the life of them. The chips are by far the most expensive part of what we're doing. So the chips on that would be another few hundred million dollars. But yeah, it's weird. Now for the crazy question.

54:01Spiros Xanthos:is it possible to fabricate some of this on the moon eventually 100 100 elon is absolutely a factory to make h100s on the moon i mean now it's we're kind of really stretching

54:18Philip Johnston:not h100s you probably would not do that you do you do the h100s probably on the ground for quite a long time because they're they're they're pretty light anyway but the the things which are you have a lot of aluminum and silicon and you need both of those for the solar panels and the radiators that's the biggest mass and also the satellite bus all of that is like 80 or 90 percent of the mass of the satellite chips is quite a small proportion so you definitely manufacture the solar panels and radiators and satellite bus on the moon and then shoot them back to earth do the

54:44Spiros Xanthos:elements to do that exist on the moon sorry for yeah yeah not going to graduate school but

54:51Philip Johnston:literally that exists on the moon yeah literally silicon is about 30 of the lunar regolith and which is necessary for the solar panels. And then aluminum is the rest.

55:01Spiros Xanthos:So this isn't, Spiros, as crazy as it sounds. I think you're taking a part of the, I loved your investigative Columbo analysis of Sam's quote. Because Sam doesn't have access to space. And if he wanted access to space, he would have to bend the knee to Elon to get access. And you have to genuflect. So that's untenable at this point, since they're still in a lawsuit. but he didn't say a decade he said the this decade so that was i think a very cool parsing

55:34Philip Johnston:on your part hey you know until so just say until a few months ago sam was actually quite bullish on space data centers he he for example when he released 03 the first task he got it to do was to design a radiator for a one gigawatt space data center that was uh that was the video interesting

55:52Spiros Xanthos:You know, I want to hit on this Andrew Kuparthi weekend project. He's got a side hustle Spiros. And maybe you could explain his post, his GitHub repo, and what he incepted in the world. Because it has inspired everybody from Toby Lutke, from Shopify to start playing with this. and all the open claw people are like, oh, I set up an agent myself. I might as well set up my own language model and start doing experiments. Again, it's like the distance between science fiction and our reality just seems to compress every day or every week.

56:35Philip Johnston:Yeah, I think in simple terms, what Prathit did is to essentially build an agent that tries experiments on his behalf, right? And these are simple experiments, right? Intuning a model, basically. But I think, you know, obviously this is a prototype, right, to show what the future looks like maybe, right? So essentially in this case, I think the agent, the research agent, right, that he developed can modify essentially the code to essentially try an experiment. Then he can run it, evaluate the output, run for five minutes, I think, see if it improves or not based on like some test he has. But essentially this can run on its own, you know, overnight.

57:12Philip Johnston:It tries simple things, right? It's not going to lead to any breakthroughs yet. But I think it's mostly like a vision for the future, right? Where maybe like even these harder problems that are all human intuition based, let's say, and long years of experience can be delegated to agents, right? I think what we've seen happening with OpenClaw or even maybe like coding agents, right? We can let them run for a while. You know, a single individual can probably consume in tokens a lot more than their own salary, right? Yes. I've seen this. I think it's happening, by the way. I think we're going to see it a lot more this year.

57:46Philip Johnston:Maybe if, let's say, coding agency is the canary here, we see that engineers probably consume more in tokens than their salary, but then they produce 10 times what they could on their own, right? So that equation works out. And I think what we see here maybe with Carpathia is that can be applied to many other things, right? Eventually, it's going to be applied to all types of maybe virtual work. So in a way, I completely understand the excitement around this, right? Because it proves that maybe even like very hard problems and research can be delegated to agents, right? That, you know, maybe they're not as smart as we are, but they can work, you know, nonstop.

58:21Spiros Xanthos:And the next thing Carpathia did, Boris, was to say, what would happen if we had a thousand of these as like essentially graduate students running experiments and then had them talking to each other? And this is the idea that I think is getting particularly interesting with open source and this hacker movement in AI, Boris, is we're seeing, you know, from Shanghai to Brooklyn to the Bay Area to Austin meetups for OpenClaw. and everybody from knowledge workers to moms, you know, homeschooling. We had a homeschooling mom on who's using OpenClaw to, you know, do, you know, tasks to help her with her homeschool, with her kids.

59:04Spiros Xanthos:I mean, it has infected everything. And now we have this, you know, model building going out to the edges. And here, Andrew Karpathy says, I tried a few setups, eight independent solo researchers, one chief scientist giving work to eight junior researchers, et cetera. Each research program is a Git bench. Each scientist forks into a feature branch. Git work trees for isolation, et cetera, yada, yada, yada. This is like, Boris, we're down the rabbit hole now. We're going to be firing off entire graduate school teams. There's only 3 ,000 PhDs, I think, of note in AI in the United States, and you all are battling for them to be on your teams.

59:49Spiros Xanthos:this could go to 30 ,000 or it could be 300 million if this continues. Yeah.

59:54Boris Sofman:It's like you start thinking about like modularity of your agent systems versus your software and architectures at your point. What's fascinating about these is it almost more abstractly just points to the fact that the scientific method is being automated and you have a fairly well-structured problem in terms of like optimizing a ML model where you have dimensions you can push on. You have a very clear way to evaluate it and kind of iterate on it. It makes me excited for, you know, the fields we talked about, like medicine, where it's exactly, you know, what is, you know, done at large scale in universities and research labs.

1:00:28Boris Sofman:And, you know, at the end of the day, like these models have an ability to interpret giant permutations of dimensions of, you know, the data coming back and potentially be a lot better on the next experiment than independent, you know, kind of like isolated students or researchers might be. And so to me, it feels like a next step of automating, like just a scientific method of running experiments, accumulating those results and taking next steps. But now you can paralyze it in a way that's shocking. And that's what's probably most exciting about it.

1:00:58Spiros Xanthos:All right. Let's end on the Claude, or I should say Anthropic versus the military and the government. Obviously, you gentlemen have been probably watching this back and forth. I'm curious, Philip, if you have a take on what's occurring here, you know, you have some conscientious objectors or concerns. Hey, don't use our software to build murder bots. Don't build our software to build a police state. We had a meal, Michael, from the Department of War on All In last week. I'm unsure if you guys saw it or not, but he sort of explained, like, we need to use these tools how we want to use them. And we already have a system of law.

1:01:37Spiros Xanthos:We don't need Dario to be this interpreter of the law. We have the law and just sell us the bullets and we'll point the gun in the right direction. What was your take on this whole back and forth? I'm curious.

1:01:52Philip Johnston:Yeah, I think it's a really tough spot for Anthropik and for Dario in particular. I think the military definitely, if they're paying for technology, has the right to use it how they want. I do think, though, that private companies are allowed to say that they don't want to work with the military and not have the military threatened to cut off all of their other customers by saying they're now a supply chain risk. I think if they go down that route, it's a dangerous path for the military because what it means is new companies coming up will be much more hesitant to sign any contracts with the military.

1:02:27Philip Johnston:If they know that, oh, if this goes bad, we're going to lose all of our, you know, everybody now are going to be called a supply chain risk. Whereas if we just don't engage in the first place, you know, then there's no downside for us. So, yeah, I think the military should be careful that we know threatening to call people supply chain risks. But, yeah, I definitely think they have the right if they pay for technology to use it how they want. Yeah. And to just give you guys the definition of this statute, USC 3252, the statute defines supply chain risk as the risk that, quote, an adversary may sabotage, may loosely introduce unwanted function or otherwise subvert a covered system.

1:03:08Spiros Xanthos:Pretty aggressive, Spiros. how much of this do you think is performative on the two parties? Because, you know, this administration can be a bit effervescent in their responses and performative. And let's face it, Dario and Anthropic, you could make the same critique of them. So what are your thoughts trying to get through these two?

1:03:27Philip Johnston:I think so. I think there are, you know, big egos. And, you know, first of all, Anthropic was a supplier. Like they gave the models to the government. Sure, they opted in. Especially when they were a smaller company, right? And I do think like there was a bit of, I mean, I understand maybe the dynamics inside the company, right? Or the Valley, right? And the balance they're trying to keep. And, you know, I respect that to some extent, right? But I do think, you know, it is performative to some extent, right? And I do think also the government maybe is going too far, right? Anthropic has amazing models.

1:03:54Philip Johnston:They should be able to use them. I think like, you know, punishing Anthropic with the supply chain risk maybe is extreme, right? It doesn't help either party, right? Like Anthropic obviously loses as a result. And they are to blame to some extent for getting to that point. but also the government would rather have access to their models, right? I don't think that makes sense either on the other side. You know, there are great alternatives, right? And, you know, open AI, especially the new models seem to be catching up. But, you know, there's no reason to not have access to aerobic models. I think, you know, just maybe like being a little bit cooler on both sides maybe would have helped them in this case.

1:04:24Boris Sofman:I agree with Philip's take. It's like, you know, they're free to do business with the government if they want to. But if they do, you know, the government has, you know, can use the product however they need to. It feels a little bit surprising how broad and heavy handed the response is, because the unfortunate part of this is that those models are actually some of the best for certain applications. There was a lot of invested time by a lot of other agencies to be able to leverage them. And now having to replace them, even if it's temporary, is actually a giant waste of effort and distraction.

1:04:55Boris Sofman:And it's not just the government itself. There's a lot of companies that now indirectly do business with the government that are in an area of ambiguity on whether they're allowed to use these models if they go and do business with the government. So there's a lot of secondary impacts here that in the end will probably get unwound at some point and will end up with a lot of wasted effort that's unfortunate. The sooner the better, by the way.

1:05:16Spiros Xanthos:I don't know if you guys saw this story. I called cap on this one. Anthropics Clause from Robert Wright from Non-Zero News, which I've never heard of. And he said, Anthropics Clause helped select hundreds of targets for the opening wave of Iran strikes. There's a good chance that one of them was the elementary school where more than 100 girls died. My latest non-zero news piece. This can't possibly be true because the military would never use an LLM to pick targets and do that blindly. Anybody have thoughts on this? I'm trying to have my team, you know, track this down. But it got almost a million views on X, so I'm not sure what the state of this is.

1:06:04but any thoughts on would you even consider using this

1:06:08Spiros Xanthos:to pick targets at this moment in time, Philip? It feels like not wise.

1:06:14Philip Johnston:I would be very surprised if they were using Claude to pick targets in Iran. I mean, for one thing, Claude is not trained on data to do with Iran, particularly I would imagine. And even if it was, I mean, the logic gates for the LLM in particular, maybe other models have logic gates which would work for that but yeah i would find that very surprising to me to me this is like blaming uh microsoft for excel when it's used to to short targets right like it's the same thing here right maybe somebody put some data into into cloud to to try to short them out right or you know figure out type yeah they could have done that right they could have they could have they could have said rank these cities by population size like and claude could do that for sure um but that's not that's not no way human using this one yeah I mean, maybe they have more proprietary data, right, that they have to go through, let's say, right, to make decisions.

1:07:06Yeah.

1:07:07Philip Johnston:But again, that's akin to Excel and bringing Microsoft then for the use of Excel in military.

1:07:12Spiros Xanthos:Also, I use this Zebra pen to circle on the map where to drop the bomb. So I think Zebra G750 is responsible for me marking on the map where I wanted the bomb to go. I mean, just absolutely crazy. And I think, yeah, this makes absolutely no sense. All right, listen, gentlemen, a great job. uh philip boris and spiros you guys were awesome thank you for deep diving into all these topics uh star cloud bedrock robotics and resolve ai i uh wonder are you gentlemen hiring and if so for what positions i always like to give you the ability to uh make the pitch to come work at your companies uh who are you hiring for philip and we are like to work there we're very much

1:07:55Philip Johnston:hiring we are hiring across engineering power electronics mechanical engineering thermal engineering, as well as some government relations, mission operations people. If people want to come and build Dyson Spheres and Matrioshka brains, they should come and work at StarCloud.

1:08:09Spiros Xanthos:Got it. Okay. Boris, your best pitch? Yeah. Dig ditches with AI.

1:08:17Boris Sofman:Oh, and a lot more than generalized to a lot of other tasks and machines. So hiring all over, it's a giant autonomy problem. So a lot of machine learning, simulation, infrastructure, hardware, also operations, a general counsel. So there's a lot of interesting positions, but it's a fascinating problem with a lot of dimensions to it.

1:08:38Spiros Xanthos:And a good time to get on the rocket ship because, hey, the company could go places. Your stock options might be worth something. That's my editorializing. Spiros, what are you hiring for? And what's the culture like over there? And what's it like to work there yet?

1:08:51Philip Johnston:We're in person in San Francisco, and we're trying to change the way software works. We just recently announced our series A at a billion dollars. We're hiring, essentially, Resolve is building agents, and we're trying to also collect the right data and train models for, let's say, running production software. We're hiring in infrastructure, let's say, agentic workflows and post-training, but also go-to-market. Resolve is now working with some of the largest technology and financial services companies in the world, companies like DoorDash, Salesforce, et cetera. So we are looking for folks who want to maybe come and help us, you know, expand in everywhere, especially in the US.

1:09:28Spiros Xanthos:All right. There you have it. And we will see you all next time on This Week in AI. Go to thisweekinai.ai and you'll find the YouTube channel. You'll find your Spotify links. You sign up for the daily email to get inside information on. We're going to start writing profiles in our daily email of the next wave of AI companies here on the program. we've got like you know the killers who have all you know have established companies but in the newsletter we're going to start covering the very early stage companies basically these companies five years ago this week in ai.ai see you next time everybody bye bye