Title: Astrid Wilde: A Bright AI & Robotics Future Show: Value Hive Podcast (audio podcast; host Brandon Beylo, Macro Ops) Guest: Astrid Wilde - founder, Inheritance HQ / Inheritance AI (robot-training data for physical AI); public-markets investor who posts holdings on X Date: 2026-09-11 (Spotify episode date) URL: https://open.spotify.com/episode/04qniefmyhr2p4VTOUoWeU Length: 1 hr 14 min (~74:45) Note: Spotify auto-generated transcript ("accuracy may vary"). This is an AUDIO podcast with no YouTube upload found, so the (m:ss) cues below are Spotify's own transcript cues - they are NOT YouTube timestamps; the analysis page's "At" cells open the Spotify episode at that second. CAPTURE METHOD: same as cnbc/archive/2026-sep-11 - the Transcript tab holds the whole episode in the DOM but virtualizes it with CSS content-visibility:auto; forcing content-visibility:visible + contain:none on all 892 transcript children made the full text readable. Captured in two get_page_text pulls: pull 1 from the cold open to cue 52:12 (the tool's 50,000-char cap), pull 2 after hiding the first 618 panel children so it began exactly at cue 52:12 and ran to the sign-off ("Yes, been a pleasure."). The join is at a cue boundary with no overlap or gap; nothing invented. Episode title verified on the page before capture. Speaker diarization was numeric; mapped here as Speaker 1 = Brandon Beylo (host; Astrid addresses him as "Brandon" at 30:55), Speaker 2 = Astrid Wilde. Spotify's auto-chapter titles interrupt the text (and sometimes swallow the speaker label) - kept as [Chapter: ...] markers, speaker inferred from content. ">>" marks a speaker change inside a cue. Auto-caption garbles left verbatim and corrected only in the analysis: "Astrude Wilde" = Astrid Wilde; "Asher" = Astrid; "Entropic"/"anthropic"/"anthropics" = Anthropic; "Fennec" = Fanuc; "Sara Lee need" = necessarily need; "SKYH" = Sky Harbour Group ticker as read by the host; "EVE bit" = EV/EBITDA; "Isaac SIM"/"Mujoko" = Isaac Sim / MuJoCo; "Left 4 Dead Space" = left-for-dead space; "commodities bowl" = commodities bull; "PUD 8" unresolved (a bitcoin-miner issuer, not mapped to a ticker); "LLMS" = LLMs. Light remove-only cleanup (pure stutters like "I, I" / "the, the"); wording otherwise verbatim.
[Chapter: Astrid Wilde's Journey and Inheritance AI's Mission]
(0:00) [Brandon] Astrude Wilde, This podcast has been a long time in the making. I appreciate you being flexible with me and having my schedule with two young kids. And I'm in my daughter's nursery right now because my office was in the basement and now we're trying to like finish the basement.
(0:17) So I had to move upstairs. I don't have my mic set up, but I wanted to get this conversation in the books because there's so much happening in AI and robotics and the release of Astra the latest GPT model.
(0:32) I just like, I just had this feeling like maybe I need to like really talk to somebody that is at a much deeper level working on this basically 24/7 and kind of looking at things and like kind of has an inside baseball look at it.
(0:48) And anytime I see your tweets, and like I was saying earlier, I was like, anytime I see your tweets, I just feel like I'm not bullish enough on AI. And so I think to kind of frame this conversation, let's start with who you are, what you do for a living, how you got to where you are.
(1:03) And I think your company right now is Instinct HQ or in Inheritance HQ or something like that. So yeah, we can even touch on that and then dive into everywhere we want to go next.
(1:16) >> [Astrid] Sure. Yeah. Well, I guess most of my time, my day-to-day life is people coordinator, slash owner operator. At Inheritance AI, we'd primarily the most material aspect of our business is robot data.
(1:36) And so taking what we call the, forgive my jargon, but it's human priors, which is like all of the embedded information and knowledge that a human has about actions or doing stuff in the world. How can we take that innate or learned behaviors from humans and move that into an embodied intelligence or which we now think robots or physical AI would be the bigger category.
(2:04) But I think that robots is like two different, entirely different design architectures. One kind of robot would be literally your dishwasher, which doesn't Sara Lee need to know much of anything, but it does a job and it does it really well.
(2:21) And I think that that's the perfect robot. But there's so many, so many, so many cases where the task or the action is more complicated than just doing one thing with a deterministic set of rules.
(2:37) And we need more quote unquote intelligence baked in to be able to handle edge cases. And that world is the world in which we deal with. And so we're trying to cap, take the things that humans do and package that up into a nice form such that it's easy to train on for these large AI companies that I'm sure you and the listeners are all familiar with.
(3:05) >> [Brandon] Yeah. And I think I've seen, I mean, you've seen examples of that with companies that basically are paying people to wear cameras or to wear, you know, hand, you know, gloves with sensors so that they can drain robots. And then you've got, you know, the NVIDIA simulators, right, where they can run things.
(3:25) And you mentioned it, you mentioned the kind of a interesting point where there's things that are very easy for robots to do. And you can kind of think about it like auto manufacturing, like where you're seeing a lot of the initial robots being deployed. It's like, hey, you're an assembly line, your only job is to do this thing.
(3:42) And then when it comes to something like folding laundry or cleaning the house, like that's a multivariate super complex problem. So how do you even begin to like solve that and then create the environment and create the data to then train the model?
(3:57) >> [Astrid] For solving the problem, the answer is we don't solve the problem. The bitter lesson, so to speak, is that we don't actually have to solve our problems anymore. We just have to provide the model with enough data of high quality and varied enough and the magic of the attention mechanism or this miracle of deep learning kind of takes care of the rest.
(4:25) So when we're capturing data, sometimes the data that we're capturing, we would refer to as primitives where let's say in tennis, the for an individual player who's playing tennis, there's a lot going on at any particular point to keep track of.
(4:44) You have to know about the score and that might determine your strategy moment to moment. You have to know where, where's the other person relative to this ball that I'm going to hit. But all of the things that make up the play space that is tennis are very simple and they involve standing on 2 feet that involves swinging racket, forehand, backhand, above below account, understanding the spin of the ball because that's important for both when you make contact of where, how, where do you need to make contact with the ball and where's the return targets on the other side of the court.
(5:28) But those are the kind of learned behaviors, so to speak. The primitives are just moving around, tracking a ball and swinging. And there are so many, so many tasks that it's economically inefficient for us to capture the actual full task end to end, But where it's very easy for us to capture the primitives.
(5:51) And so sometimes we're just dealing with primitives, sometimes we're dealing with full tasks. And you've seen there's a whole ecosystem around both. And so in the capturing full tasks like you said the example of maybe you give somebody a camera on their hat or tape to their forehead or some other hacky solution and they go about just doing their job.
(6:17) And the problem with that approach is people are very messy. So you might have somebody that's on an assembly line and they just turn on the camera and then there's an 8 hour unedited segment of video that includes bathroom breaks, that includes people pulling up something obscene and watching it while they work.
(6:41) That includes them stopping to talk to their Co worker for 5 minutes. And these are not the behaviors that we care about that are generating economic value. And so there's a whole pipeline of people that comes after the people that capture that data to split this up into meaningful segments of where's the important information happening, from when to when, what is that important information?
(7:07) And ultimately that's the real grunt work of this that takes the really, really messy real world people data and turns that into something that we can then train on. [Chapter: The Future of Non-Humanoid Physical AI]
(7:22) [Brandon] So within the physical AI space, like where are you most excited about the applications looking out 18 to 24 months? I think a lot of people are kind of hooked on the humanoid aspect where you've got like this kind of Tesla optimist robot friend, you know, kind of in your home.
(7:40) But my base case is that the really cool stuff is going to be the stuff that doesn't look like humanoids at all, but it's just robots that allow people to do things either a lot more quickly, more efficiently and, you know, can create things where you don't need as many people to get stuff done.
(7:59) >> [Astrid] Yeah, I think a great example of this is a company Matic. They're a robot, the original robot vacuum company. If you remember Amazon a while ago tried to purchase a company that was doing a robot vacuum called Roomba.
(8:16) That is, I think since gone bankrupt as a result of not having a bailout by Amazon. But this is a perfect example of this. This thing is handling edge cases. It has some kind of innate sense of space and time and it has cameras so it can sense when things are going, but also can work operate in the dark.
(8:39) It's in, unlike the Roomba, which is really loud when it turned on at 3:00 AM and would wake you up, or the cat, like this thing's almost perfectly silent. But it does one job and it does it really well and it handles it in a variety of circumstances. That's the future that I am really excited to live in is where all of these menial labor tasks where the things that involve humans doing something repetitive and mindless, where we're not actually solving a problem.
(9:07) It's just busy work getting progressively automated away 1 by 1 by 1. And it's going to be little chips like that over and over and over again. And there's a lot of stuff both in the home that I think is going to return a lot of time for people to spend with their children or for whatever they else they want to do with their leisure time.
(9:26) And in the factory, that right now is just, we need a person doing this because the person's the cheapest person to do this. Yeah. That I just want to. I'm so excited to live in a world where people aren't doing mind numbing, repetitive labor for years of their lives.
(9:45) >> [Brandon] Yeah. And so within that you've got the home, you know, home tasks, you've also seen stuff in construction and you've seen stuff in like mining. Have you you know within inheritance, do you guys focus on a specific domain or is it just kind of general use case and then you see where the opportunities are for you guys to kind of have an impact?
(10:07) >> [Astrid] Wherever there's customer on the other side with the data problem that they're willing to pay to solve. So the broadly speaking, the narrow to narrow it a little bit. Our focus is anything that is a task that a human would perform because of all of our models are tuned to recognize human action specifically.
(10:28) And so if you took a bunch of videos of somebody operating a forklift, for example, we're almost certainly not going to be helpful in this case. I do know the right people to introduce you to. But more broadly speaking, this like menial human labor tasks of like assembly and sorting and anything that involves your proprioception and ability to manipulate objects.
(10:56) But it's like fundamentally a human is the one doing the action rather than a human augmented device. [Chapter: Manufacturing Capacity and AI Industry Concentration]
(11:02) [Brandon] Yeah. And so one of the things and I feel like this, you know, the word bottleneck's obviously been just beaten to death during this whole AI kind of cycle. But with physical AI I think the bottleneck argument is pretty interesting. Like it was kind of memory then it was, you know, then now it's kind of power and energy with the crypto miner data center play from your seat again, looking out 18 to 24 months.
(11:31) Like, are there any bottlenecks within the physical AI robotic space that you think people aren't necessarily discussing that you think are probably more important than the mainstream ones that we see today?
(11:43) >> [Astrid] The biggest bottleneck is manufacturing capacity right now for a lot of things. And that's like fundamentally the at the root of like the power bottleneck and the memory bottleneck and also the device slash sensing bottleneck.
(12:01) So right now this hasn't made its way yet to public markets, but a big shift that's taking place over the last 18 months is if you were trying to purchase a sensing device, let's say it's a camera or some kind of sensor that you're to put in gloves, for example, for some kind of a data capture. 18 months ago, if you put in a purchase order or emailed somebody at one of these companies to even get like a single unit for testing, you would get 3 people that would respond to your e-mail and get on
(12:32) a call with you. And they'd be happy to talk with you for 1/2 an hour and like understand exactly what you're doing and get you the device within a week. And today those same people won't even talk to you unless we're talking about more than 100 or 1000 units, depending on what the SKU is.
(12:50) And so, and things that are affecting our business as well is there's a huge backlog that's developing for everything at a very, very low to low level. Everything, every form of sensor is starting to sell out.
(13:06) And we're getting estimates of, oh, we'll have more than three months or we'll have more than nine months or we'll have more next year. And that's not a problem that we had 18 months ago.
(13:19) >> [Brandon] Yeah. Do you think that almost like accelerates the consolidation within physical AI and compute and then these models? Because the whole like one of the arguments, I guess with AI now is like, you just kind of have a race between it seems like anthropic and open AI.
(13:35) And it's just these two giant powers. And there isn't room for anyone else. Because now it becomes like a scale problem of, oh, like who can just get the most compute? And then whoever can get the most compute is usually the person that has the most money and who can raise the most money. Oh, like the two biggest companies. And so there's like the chicken and the egg problem.
(13:52) And so it kind of squeezes out a lot of these smaller potential models, smaller companies, researchers. And like, how do you think we solve that? Because I don't know, like, I'm interested to hear your thoughts. Like, should we be OK with a world where it's, you know, AI and is just dominated by two players or, you know, kind of what's your take there? [Chapter: Why Individuals Gain Leverage with AI Tools]
(14:13) [Astrid] I think the good news is that neither of these two players and broadly speaking, they're very few exceptions to this among the like top ten players that they're basically not interested in actually solving customer problems.
(14:30) They're interested in solving intelligence, broadly speaking, because they think that that's the point of highest leverage. But Entropic's not going to invent a vacuum that can go clean your whole house. They're not going to invent a new novel dishwasher or stove or they're not going to solve lawn care or invent a drone that can detect a disease by flying over trees.
(14:58) Like this is not the level that they're thinking at. They're trying to train something that's very broadly applicable and can be applied to a large range of problems so that other people have a much lower cost of doing just that, of solving actual real world problems.
(15:17) And I'm much less concerned probably than the median person about the concentration of this large general intelligences because of that posture. And because it's either we're going to end up in a world where they feel like they've done their job and now they just start trying to spin up these factories of people that are solving all of the really narrow specific problems.
(15:46) Or they just continue down the trajectory that they are for the next 5 to 10 years. And there's a whole bunch of really, really low hanging fruit for everybody to pick where companies that used to require years of work, now a couple of people can spin up in a couple of months with no starting capital.
(16:04) And that's like the work that we're doing even it would not be possible without these large models that are have effectively reduced our cost of everything software very, very near to zero. I mean systems that we built last year and spent probably $200,000 to create are you could replicate now if you had some domain specific knowledge in like 3 weeks and $20,000, which is pretty phenomenal difference.
(16:34) And then we're still like able to solve the same problems. And it's this kind of this big recursive loop that you're right, it does like increase the concentration of power, quite literally, like they literally are buying the power assets.
(16:51) But also the you get what I'm pointing at, but I'm super not worried about it or thinking about it. The thing that I'm thinking about is how much leverage an individual person with very little money has to affect change in the world is just going to continue to increase even faster than this larger concentration of.
(17:09) >> [Brandon] Power and I think that's what I pick up on when I read your posts where you know your quote posting and stuff like interesting things that are happening like this morning, someone with using Astra created like, you know, some kind of similar thing where like with their hand, they could scroll up and down where the camera is kind of looking at what they're doing and you know even using kind of AI to paint a picture of things.
(17:32) And so I get the sense of like you're very optimistic. I don't know if you know, maybe I don't want to call you like an AI maximalist or something like that, but you're definitely more optimistic, at least in my feed too. Like when I scroll, you know, like I said earlier, it's like whenever I read kind of some of your tweets about AI, I'm like, man, like maybe the future is gonna just be way more interesting.
(17:57) And that means like the demand for kind of the whole AI stack today is going to be much higher in the next, you know, 24 months or so.
(18:05) >> [Astrid] Yep, I couldn't have said it better myself. That's I'm. It's funny that you call me an AI maximalist or even use that term jokingly, because among friends I've been known to be like a very, very skeptical that this trend was going to continue. [Chapter: Astrid's Journey to Investing and The Power of Patience]
(18:21) [Astrid] For a long time. I was super skeptical that synthetic data was going to allow us to breakthrough the bottleneck that was the Internet. We had this nice pre training corpus and then we had this like lull effectively and I was super skeptical that that was going to continue.
(18:38) I was dead wrong. So wrong in fact, that I started a company.
(18:41) >> [Brandon] I was about to say like you like you receptacle, but then you, you know started inheritance, which is like betting that this AI thing that you're skeptical on like actually accelerates. Yeah, OK. So then outside of kind of what you're doing privately with inheritance, you comment and you know, on early September you posted about kind of your public stock portfolio.
(19:07) And I think it's interesting like looking at your current holdings that are greater than 1%, I mean, you hold a ton of cash. But I was surprised to see like looking at kind of what you do on the private side and what you post about. I was surprised to see your portfolio at the same time because you've got, and this is as of September 4th, right?
(19:25) So obviously you're subject to change whatever, but you've got really NVIDIA, which makes sense. QXO you've got Nintendo, SKYH, Amazon, couple undisclosed positions but like to contrast with what you tweet about. I would assume that your portfolio would look like situational awareness.
(19:44) >> [Astrid] And.
(19:44) >> [Brandon] It's and it's not like it's actually kind of a little bit esoteric. Like you've got H, you know, you've got hero in there, which is a name that I've followed. You've got Nintendo, which is one that's interesting. So like walk me through how you trade public markets and how you view that and different investments there.
(20:00) >> [Astrid] Yeah. There's really only fundamentally 2 things that I do. And both of them have a couple of different ways in which they manifest. But broadly speaking, the thing that I've done for the last six years is front run earnings. And that when I say front run earnings probably means something very different to a quant that's listening to this or somebody at a pod shop, it's like, oh, I know that this quarter is going to be a beat $0.10 higher per share than the current St. estimate.
(20:29) This is broadly speaking, not the thing that I do. There's a couple of times that I have done this, but these are very, very, very rare exceptions. Usually the thing I'm looking for is a very large change.
(20:44) It's material for the business that's going to last for a long time for a specific company or sector. So this happened with during COVID for shipping for everything energy for the Russian Ukraine war, the what the ongoing stuff that's in the Strait of Hormuz, like the broad theme that I'm looking for is what are things that are going to materially change in a business before consensus has caught up to it.
(21:18) And so most of that for the last like 3 or 4 years has just been front running that the quote unquote AI trade of buying all of the crypto miners before they converted to high-powered compute, buying all of the power assets before everybody figured out that that was the bottleneck buying last year, buying all of the memory companies before everyone figured out that we were sold out for the next 4 years like that.
(21:44) But that like mindset of looking for something that's out of material inflection and a large trend that's going to last for a long time is thing one that I do. Thing 2 is just sit around and wait for thing one to happen.
(21:59) >> [Brandon] Well, that's super important because that's like my biggest hang up and where I'm trying to improve is just the waiting. Yeah. And like, 'cause like, you know, you come to like the market's open five days a week. There's this tendency to feel like you should always be doing something.
(22:15) There's always a trade. But like most of the time you're just waiting, learning and just kind of again, like you said, waiting for that thing, one, that inflection to happen. So how do you develop that? Like if it's hard for someone like myself to have that patience day-to-day, how would you advise someone to improve that skill? [Chapter: Why Less Market Research Can Lead to Better Outcomes]
(22:38) [Astrid] Find other things to do.
(22:40) >> [Brandon] Yeah, I.
(22:41) >> [Astrid] I mean that sincerely. I so literally, yeah, when I started this process of publicly documenting what I was holding public equities wise back in 2019-2020, the that was what I was doing like 6 hours a day, 10 hours a day was essentially I was a private equities analyst.
(23:08) And The funny thing that I found over the last few years is the number of hours that I spend reading earnings releases and the news, it said, has no material impact on my performance.
(23:23) >> [Brandon] Isn't that amazing?
(23:24) >> [Astrid] It's incredible. So now I spend like maybe 15 minutes a day optimistically.
(23:30) >> [Brandon] Wow.
(23:31) >> [Astrid] And it's still the same kind of historical CAGR for the last two years as it was before, before that. And I think literally the amount of time that I spend worrying about it, about what I own and what I might be missing out on is also directly proportional to the amount of time that I spend on this So sincerely.
(23:54) I mean this without a smidgen of insincerity. Find a new hobby.
(24:01) >> [Brandon] Yeah, well, investing is one of those weird professions where just pure increase in time spent and effort does not translate to better outcomes. Where like in most other domains, that equation pencils like in tennis for, for example, like I, when I was seriously playing and playing, I was playing six to seven hours a day, five days a week when I was in my late teens.
(24:34) And the amount of improvement I had during, you know, those two summers or so, it was insane. But when you go to investing like you like that, it's just, and it's just such a harder equation because I think investing attracts people like myself who like, really love to dive deep and really love to like work and grind at things.
(24:59) But then you get that frustrating reality where it's like, it doesn't matter sometimes if you do more and more work. Like your outcomes aren't necessarily predicated on how much more work you do, which is very uncomfortable.
(25:10) >> [Astrid] Yeah, yeah. I don't have anything more to add. You said it perfectly. [Chapter: Trust, Maintenance, and Niche AI Companies]
(25:17) [Brandon] So then how do you spend your day? I know you, you know, you run inheritance. So maybe part of the answer is like people should just go start companies.
(25:26) >> [Astrid] You think this is the best time ever in history to start a company? You don't even have to raise venture for most things anymore because especially if it's like a software product spends $1000 in credits and you can create the thing that you needed to.
(25:45) Most of the hard problems are not technical problems, they're people problems and money moves at the speed of trust. So everything that you can do to increase trust and deliver value in excess of what you're charging somebody is that's a business.
(26:01) >> [Brandon] So you don't think software is dead?
(26:04) >> [Astrid] No, no, far from it.
(26:05) >> [Brandon] OK, so explain that if I was someone that said look like eventually AI is just gonna solve all the software problems. We won't need Intuit, we won't need Salesforce, everyone to create their own little software for their own edge use cases. Yeah, why? Why would I be wrong about that thesis?
(26:21) >> [Astrid] You're absolutely right that people are going to create their own software. And every day my feed, because I follow so many creative people is like, oh look, I've recreated Adobe Light Room in a weekend. Oh look, I've created Adobe After Effects in a weekend. Oh look, I've made my own CRM and like this is going to happen.
(26:40) This is going to continue happening and this is going to accelerate. It's also going to be the case that the most successful businesses, the things that matter or a material for a business are not saving $50.00 or $300 per month on a SAS subscription.
(26:58) Like the place, the things that really matter for big successful businesses are spending time on the things that matter, not immaterial things that only affect the bottom line. And that also increase your liability and your ongoing maintenance cost.
(27:17) Where a big portion of what you're paying for when you have like Salesforce, for example, is other people are going to maintain this thing for you and other people are going to the accept some degree of liability if something goes wrong.
(27:33) And that's not something that you can just spin up trivially, although you could replicate, you can build your own CRM, do it. It's great. I've done it many times to meet like weird edge use cases.
(27:50) But as soon as you're doing something that's large where the input cost of the software is no longer material to the business outcome, then these customers are not going anywhere.
(28:02) >> [Brandon] So it's almost as like software itself isn't dead because you know what, whatever company it is, it's just kind of this like it's just this layer of skin that is over top of just like trust and like you just like a counterparty just needs to have someone they can trust.
(28:22) And it's easier to trust humans at this point than it is to trust, you know, 10,000 agents or something like, you know, like, oh, like our software is backed by 10,000 Grokbot agents who will always have your best interests at heart. It's probably a different value prop than it's trusted by humans, and maybe that philosophy doesn't always stand.
(28:39) >> [Astrid] Maybe that changes yeah, but I even as it changes I would be shocked if that is the thing that Anthropic for example or open AI because you named them or decide to that's going to be our business is spinning up all for every individual use case.
(29:00) I think that the increasingly very large companies that need a niche needs and take on this like liability and trust aspect are going to continue to do exceptionally well, whether that's cognition or whether that's Harvey or things that like, no, we have, yes, we're an AI company, but we really only do this one thing and we do it really well.
(29:27) And you can think of them like a software company. [Chapter: Cultivating a Network for Investment Ideas and Niche AI]
(29:31) [Brandon] And I kind of thinking about kind of that niche idea like we do one thing really well, like kind of going back to physical AI, that's kind of what really excites me about physical AI is the ability for low like tons of niche physical AI robotics companies.
(29:48) Because there's on one hand, there's the argument that this whole physical AI race is really just a race between two companies, Tesla. And then, you know, there's kind of figure AI and there's also some Chinese models.
(30:04) But we can get into kind of like the bifurcation between Western and Chinese robots. And like that may not happen, but you can kind of think about it as like, there's this race between just two giant companies. And then on the other hand, it's like, well, with these tools, especially like with Astra, and then even with like 3D printing and everything like this, maybe the really exciting stuff is like the niche construction or niche mining where it's like, hey, we've built the best robot at drilling holes for mining companies.
(30:40) >> [Astrid] Yep, Yep.
(30:41) >> [Brandon] And like that to me is what's very exciting. And like, that's kind of the future I hope to live in versus like everything is just a battle between Tesla and then figuring.
(30:55) >> [Astrid] I think that you will be very pleasantly surprised about the very near future then, Brandon.
(31:02) >> [Brandon] So going back to before the discussion about software and I know I'm kind of going a little bit of place, but like I just have a ton of questions.
(31:09) >> [Astrid] That I'm definitely like.
(31:10) >> [Brandon] How do you spend your time? Like, like what's your ideal day? I know you run inheritance, but so like, even looking at public market stuff like that, like how do you spend your time? What do you consume and how has that evolved over time?
(31:24) >> [Astrid] I have been terminally online person my entire life. There's through if you knew the right websites to look for, There's literally a public record of my life since like the age of 12 to today, an unbroken record all the way through.
(31:40) And so the I was on Twitter before Twitter had a real text interface. It was or a user interface. It was just a group texting platform. I was one of the early users of Tumblr, of Live Journal before that, if you remember, and Blogger and like I'm out there.
(32:02) And one of the wonderful things about being a publicly online person for so long is that it has been this very long standing filter that has brought interesting, increasingly so people into my life. And the majority of my ideas now for investing don't actually come from me having to seek anything out.
(32:25) People just bring stuff to me that they find interesting. And it's such a wonderful joy to be in that position where I'm just constantly getting already pre screened, pre filtered ideas thrown at me all the time.
(32:40) But most of them I just passed because I either don't know anything about it, don't have enough context to know, hey, is this a material inflection for a business going to last or I don't have like a preconceived idea of something and I just don't have time to spend a lot of time diving into it because I'm focused on building my own business.
(33:03) But the things that I do have a sense of what's going on, like what's going on in the straight, for example, or what's going on with energy demand in Europe, like, or anything having to do with AI and data centers.
(33:18) These are things that I'd like very much in my wheelhouse and have a preconceived idea. So when somebody brings me something that's on theme and an interesting special situation with a company, for example, like Nebius is a great example. Yeah, I.
(33:33) >> [Brandon] Just want to ask you for an example.
(33:34) >> [Astrid] Oh my gosh, the most will never. I probably will never see anything like this again in my lifetime. But you have a combination of, OK, this is clearly going to be one of the biggest providers of cloud compute in the United States. But the way that people are thinking about this company is they're not everyone just showed who owned any part of a Russian ETF or like who wasn't able to trade, suddenly just had this thing show up in their brokerage account one day and it's just clicking the sell button.
(34:10) And that was one of the most fantastical returns for allowing you to bet directly on the theme in a way where it's clear who the other side of the trade is. And even at the time when I was telling other friends about this, the people didn't get it or like, they didn't understand.
(34:31) They were just like, wait, I get some of my money back that I lost on this Russian thing. Great. I'm happy to do that. And whenever you're in a situation in public markets where you know who your counterparty is, you're probably going to make a lot of money.
(34:49) And if you know who your counterparty is and you have a materially different view, which in the case of data centers being a long standing trend, you can make phenomenal life changing money. [Chapter: Case Studies: Nebius, GameStop, and Peloton Opportunities]
(35:03) [Brandon] So how did Nebius come to you? Was it through your network or did you find it on your?
(35:07) >> [Astrid] Own wow. The way that this company came to me was because some of the people that are pre filtered into my network were betting that basically the Russian market would not get close to US investors. They were all dead wrong.
(35:24) For the record, I almost got, I was so close to getting burned on this too, because man, some of those prices were crazy attractive. If you had any kind of a view that this was not about to close on your face. And so then those people were saying, hey, we caught word that one of these, a portion of one of these companies is listening in the US and we're all going to own shares in it now.
(35:49) And I was like, whoa, OK, I'm just curious about that, what's going on there? And you look into it a little bit and it's like, OK, this is a real company. Wait a second. This is a real company in a sector I know and understand and. And I have friends in common with these people. OK, now I need to do some real work.
(36:08) Yeah.
(36:09) >> [Brandon] And it's yeah, it seems like it's just about checking again, like checking those filters. Like is it one thing I understand? Is it pre screened by people I know and trust that kind of follow the same thing I do? And then do I understand my counterparty? And then do I know kind of the risk reward set up here?
(36:26) Do I have? And then it goes back to that variant perception, right? Do I have a perception about this company or about the Steam industry that's totally different from the market? And do you see that? Like I know you own some Nintendo. So like is that an example of like a similar situation there? Because I know Nintendo's kind of been this weird variant perception story for a few years now.
(36:45) Like is it the hardware company? Is it this kind of software business like? I think that thesis would be kind of interesting as a current example maybe.
(36:53) >> [Astrid] Yeah, I don't think that this is a current example. Nintendo is something that I think of as a cash position that's also hedging my memory position. Like they're super exposed in a way that most other hardware companies are to memory prices and that they did not go out and contract.
(37:16) They've just been buying things on the spot market for forever. And that's not a very, very bad position to be on the other side of. And so the, but historically Nintendo's, I mean, for the last like 10 years, you can find value investors pitching this company.
(37:36) I remember like the first pitch on Nintendo, I heard involved like, oh, we're going to spin off the Mariners and like this is going to unlock all of this value. And it just hasn't worked out. But I'm not worried about this company going anywhere, I'm happy to just own shares for a while and mentally this helps me justify a larger position in SK Hynix.
(38:01) >> [Brandon] So where do you think that kind of variant perception is today, right? Like you've seen, I mean, I guess you've seen a lot of the memory stocks, you've seen a lot of the AI trade play out things like the Bitcoin miners. They, you know, a lot of them have pulled back 30 to 50% from their highs starting to turn around.
(38:17) But do you think are there any kind of fresh themes that maybe your network is pitching you today that have you excited or maybe have you interested in doing more work?
(38:27) >> [Astrid] The two that come to mind that I don't yet have a position in, well, I guess I do have one. Small positions in are builders or which I consider both a direct bet on US interest rates and on home building.
(38:45) And then the second theme would be specific consumer companies. And so maybe I'll start with the second one because it's easier to make more concrete. Peloton is a great example. I don't have a position in this company to be clear.
(39:04) GameStop is another one. Wow, I can't believe I'm saying this. If you pull up the financials of GameStop for the last quarter and you analyze things, which I actually think is fair in this case, you'll come up with a number that seems unbelievable, which is that right now GameStop is trading for an EVE bit of less than three.
(39:26) And that seems like that should not be correct. I don't have a view on this business, but it's clear that the folks running this are in a position where because of weird historical timings and a past meme bubble, they just have a lot of cash and a lot of their existing asset base of these stores that nobody's going to, it's going to be very cheap to close them down.
(39:53) And so if you have any kind of a view of what GameStop is becoming, which is kind of this collectibles store, I don't know how familiar you are with like Pokémon cards or other kinds of these large.
(40:08) >> [Brandon] I mean, I know like I feel like Pokémon specifically has had like a crazy resurgence and I know sport like sporting cards, NFLNBA, like that's a huge market that's growing really quickly. But I didn't know kind of GameStop almost like as this thesis of like the Center for Collectible Exchange.
(40:26) >> [Astrid] Increasingly they're becoming a player in that market. And so if you have a view on that market being durable and you think that the management is not entirely incompetent, then that's totally interesting as a like one off Peloton is another where this is in the same like mental bucket.
(40:45) I think for people of like, we had this crazy COVID boom and now it's, I don't know what, I haven't even pulled up the chart recently, but it's very silly.
(40:53) >> [Brandon] It's still at 5 bucks. It was, you know, at its high as $167, down 96% from its COVID highs.
(41:02) >> [Astrid] Yeah, if you forget everything that's happened historically and just look at what's the next two years of this business look like, I think that it's totally plausible that you're buying this for a single digit earnings multiple today. Again, you have to have a variant view that this is like a durable business and growing.
(41:22) If you don't think that the business is growing and it's in terminal decline and people are just going to continue sending or junking their existing sunk costs into these devices, then it's not interesting at all.
(41:38) But like this is totally unrelated to the AI stuff. And like these are both kind of single company consumer themes. But this is there's a lot of stuff that is sold at equally like 30 to 60 to 90% that's in this bucket of consumer category.
(41:58) That's really interesting. [Chapter: The Future of 3D Printing and Lifelong Learning]
(42:00) [Brandon] What is your take on 3D printing as an industry and as a space here over the next two years?
(42:05) >> [Astrid] It's still super early there. There's stuff that works. Anything that requires injection molding is still, we really just can't 3D print this stuff at scale effectively.
(42:21) But for consumers, it's great. It's great. I mean, my dad is a super into 3D printing. He has several and there's an entire room of his house that you just open every cupboard and it's just spools of material.
(42:36) And so he's always printing stuff and showing it off. But I don't know, I'm much more excited about things that are involved metals and that's a much more difficult problem to solve.
(42:52) There's a friend who has a novel kind of 3D printing that's useful for things like printing false teeth for people. That's interesting. I just, I don't see, it's feels like we're way, way too early still to see some kind of an explosive demand growth.
(43:11) >> [Brandon] Yeah, yeah, no, 'cause that's kind of a Left 4 Dead Space that I've always kind of kept on the back burner, 'cause it kind of it's kind of like that Peloton thesis where like the charts are very similar. Like during COVID, you know, these 3D printer stocks just went nuts.
(43:27) Now they're all down 90 plus percent.
(43:30) >> [Astrid] If you have any kind of any view, then this is very attractive space to go together, yeah.
(43:36) >> [Brandon] Yeah, this is interesting, 'cause I feel like, you know, you as an investor. It's, you know, it's much more complex than just, you know, this kind of AI person like you're, you know, you're investing in Nintendo, you've got these DCS and Peloton, you know, kind of how do you develop this wide range of interests?
(43:56) I mean, it sounds like you're a generalist, you know, you've got passions for robotics and physical AI, but you know, what's your, you know, what's your kind of reading schedule like? Like how do you stay, you know, updated on a bunch of different things.
(44:09) >> [Astrid] I just read all the time. I can't help it. I'm an info vore. This is my this is. I think part of the being an online person is you post and you read and I don't read any traditional news source.
(44:27) I know a lot of very interesting people working on niche problems that I would consider nerds who tell me lots of things. And so I come with a lot of prepared context into a lot of things and I'm just a very innately curious person. And so things that might catch my attention that another person wouldn't think anything about like hey, this weird Russian companies shares just showed up in my brokerage account flags as interesting to me that other people would be like what?
(44:55) Who cares? OK. And I don't think I have a more specific answer for you than that, unfortunately. I've just cultivated a cluster of wonderful interesting people who continue to bring interesting things to me and I have an innate curiosity that I just follow and even though it leads nowhere most of the time.
(45:19) >> [Brandon] No, I think it's great because again, it's like going back to the discussion like doing more work in public markets doesn't necessarily equate to better returns. Yeah.
(45:29) >> [Astrid] Just.
(45:30) >> [Brandon] Any really smart person I talk to, they always just give me the answer. Like just read read, read. That's all they do.
(45:36) >> [Astrid] Yeah, there's, it's really not that. There's nothing more complicated like maybe.
(45:43) >> [Brandon] Maybe that's if you approach.
(45:44) >> [Astrid] The reading and you think of it as work or like research that is towards some specific ends like increasing your. This is not gonna work out for you. But if you're just following your curiosity, then the world is your oyster. [Chapter: Why Robotics Opportunities Lie in Private Markets]
(45:59) [Brandon] Going back to physical AI kind of as we, you know, we're coming up on an hour the robotics trade is something I'm super interested in. A lot of the robotics names I've just got, you know, I've got a watch list that I watch and a lot of the robotic names are down a good amount from their highs.
(46:17) And I'm trying to get a sense of timing like, is this the right time? Are we too early kind of like, you know, 3D printing, like are we just so early where it doesn't matter what my view is? And then understanding with where within the robotics, the physical AI, right?
(46:34) Because everyone's talking about, oh, like actuators and the hands. And you mentioned kind of a Lidar sensors where, you know, like where's this bottleneck within physical AI? If you had to construct like a physical AI portfolio in public markets, what would it look like?
(46:51) Like where would you allocate your resources?
(46:55) >> [Astrid] Today in public markets there is very, very few options and most of the options are a derivative of the actual bet that you're interested in making. So the there are a few like robot manufacturers, Fennec, ABB come to mind is like, OK, these people are making real robots.
(47:17) There's, they have customers on the other side and basically everything else you're betting that the demand from physical AI use cases is going to materially lead to inflection in their business.
(47:32) But you're not betting on the theme directly. And I don't like doing this. I found this almost never works out whether it's like, oh, I have a thesis around solar and so I'm going to buy silver. That one happened to work out most, Most of the other ones really don't.
(47:50) And if what you're interested in betting on is lumber prices, you can bet on lumber prices or you can bet on companies that sell lumber having a material inflection in their business, But you're always taking on the operational risk and management risk for that specific company.
(48:07) And the further away you get from the bet you're interested in making them, the worse your risk return prospects kind of look. So if you have a view on open AI today, your options for doing that investing in that public markets are SoftBank, Oracle or you can make these more derivative bets on like power and compute and that's kind of your only options.
(48:34) But or you can just sit around and wait for the IPO. And I'm not particularly interested in the options available to me on that one. I would rather sit around and wait for the IPO, but other people who have different risk constitutions might have a different view.
(48:50) So unfortunately I have to decline the question and that if I'm trying to construct a robots portfolio today, the answer is all of the most interesting stuff happening in robots is still in private markets. And if you are the kind of person who gets is really into having a more hands on experience, then there's a whole ecosystem around funding this stuff.
(49:16) It's unfortunately mostly hype and very little real. But if you can find the real stuff, then there's a lot of interesting stuff going on. But it hasn't yet made its way to public markets directly.
(49:27) >> [Brandon] Man, I thought you were going to say something like, yeah, it would just buy Agility robotics and that would be it. But you wouldn't even look at agility. You just think it's too early.
(49:38) >> [Astrid] It's hard for me to say much more specifically because a lot of these companies that are in the process of IP owing or about or just IP owed our customers or soon to be customers of mine. So I have a direct conflict of interest, but I'm gonna decline the question. [Chapter: Underestimated Demand for AI Training Data in Private Markets]
(49:57) [Brandon] That's awesome. But within the private space, then where do you think is kind of the most interesting companies? I mean, obviously besides your own, but you know, you have an interesting view of the supply chain. So on the private side, like what's really exciting that you're seeing being developed?
(50:15) >> [Astrid] I think that demand for data continues to be underestimated. We saw a similar kind of underestimation over the last couple of years of demand for data from the language labs and the form of data that's people are broadly familiar with, maybe not at all.
(50:34) I is reinforcement learning and there's a whole bunch of companies in the ecosystem that spun up around creating these environments for teaching models to do a task. So when you saw the stuff that I've been sharing on Astra, for example, making stuff in Blender or in simulations or controlling directly controlling a robot to help it paint.
(50:55) The reason that the models have those capabilities is because teams of people created environments which converted what is like an abstract problem into a computation problem that the model could then train on.
(51:11) And I don't want to get super into the weeds of how that's done or possible because I think it will bore your audience, but maybe I'm happy if you're super curious to go.
(51:20) >> [Brandon] Into the no, I'm super curious. Like what's like, OK.
(51:23) >> [Astrid] Depends on the work. So if let's say your thing that you want to train a model, a language model to do is to be able to draw control the robot to Draw Something, the way that I would create an environment around that is I ultimately your the problem to solve for is how do I turn this into something that's a text as a like objective?
(51:50) And this feels like a completely unrelated to text, right intuitively, but it can be broken down South. One way that we represent images today through text is with the what's what's called the name of the file format is escaping me vectors.
(52:12) So you could represent images in this like compressed format which is jpegs or an uncompressed like image format which just has the pixels and their values of PNG. Or you can represent images in this vector format and image in this vector format are just text.
(52:31) It's just a text file. It's tiny. And so instead of having like a huge thing that encodes each direct pixel value, you can train this model to draw an image, quote unquote, by just saying here's the vector.
(52:49) And that's one way you could do it. Another way you could do it would be, hey, let's train a vision model to recognize values of these direct pixels. And so, OK, either way is totally feasible and people have done both.
(53:09) But the way I would create an environment would be, here's my training objective. Let's say I give the model a picture of something that I want the robot to draw. Then I would create an A training environment. And in this environment, I'm probably using an off the shelf robot simulator like an Isaac SIM or a Mujoko or there's a whole bunch of these.
(53:34) And then the actual objective that I'm creating around this environment, part of the packaging is it, I give it a score on how it performed at the task. And ultimately that's the creating the environment and having an output score that's deterministic is what goes into these training environments.
(53:57) And the way that you would compute the score for this specific task is OK, here was my input. And then on my output I have something in this environment which the virtual robot is using to draw and something that is a virtual canvas. And I could either directly record the pixel values that are pushed onto this canvas in this virtual environment, or I could re encode this as this vector image.
(54:21) And either way, I can then overlay this or compare this to my original objective. And this gives me a very concrete score. And so this kind of thing, even though this feels really unintuitive and like how could a language model learn this kind of thing?
(54:36) Because you can represent images as language and because you can have this really easy, easy to score training objective. This is just the kind of thing that language models are easy to train on if you have enough of these environments.
(54:52) And that's the operative question is what's enough? And the answer is we usually don't know until we get there. And people really have underestimated how much is required to get there. But the actual process of creating these environments and then training on them is not complicated.
(55:10) There's no magic sauce or secrets under the hood. It's all very, very simple stuff. It's all just these primitives that I was talking about, but it's just so, so, so many of them and so much compute thrown at the problem that eventually all of the connective tissue kind of works. [Chapter: AI's Impact on Compute, Capital, and Resource Demand]
(55:30) [Brandon] Yeah, I mean, it sounds like after hearing that, it's like maybe I should just buy more compute stocks because like isn't, you know, maybe the logical endpoint is just if you throw enough compute at anything, it'll figure it out. And then so then what's the AAI guess? It kind of goes back to like, what's the price people are willing to pay for intelligence, You know, maybe like intelligence with a capital I and I don't know if we know the upper bound of that price yet.
(55:57) And so then the function of that is like, OK, then that's just a compute problem. And so then, what's the upper limit of people's willingness to pay for computer?
(56:06) >> [Astrid] Yeah, I don't know the answer, but I know that human demand is unbounded and people will never be satisfied. They always want more. And we are living in this weird magical time where what once was a human labor problem is becoming increasingly a capital problem.
(56:22) And there's continues to be capital that's willing to put up the balance sheet risk for the answer to that question of.
(56:33) >> [Brandon] See two, it's capital and it's a resource problem too because like going back to 3D printing, you know, kind of like my initial interest in it was after watching that Astra video, my logic was OK 3D printing to me always seemed like this thing that engineers did in their free time because it was super fun and like engineers knew how to build stuff.
(56:56) Like my buddy was a aerospace engineer back in the Syracuse. I'm like, I just remember him 3D printing like this cup holder because the cup holders in his car weren't like ergonomically perfect. And I was like, OK, like that's something only an like a like an aerospace engineer would do in their free time.
(57:13) But now with lol, it's like anybody can kind of engineer products, consumer products. And so my thought was like, OK, so you've kind of taken the lid off of a constraint where now it's like the constraints for 3D printing or something like that are like the individual's creativity and then the raw materials necessary to kind of build whatever you want to express in the physical world.
(57:38) And you know, that kind of leads me down. Like I'm super bullish on physical commodities. And I just had my buddy Gavin on the podcast and his whole thing is like the physical AI and you know, kind of these LLMS as we enter the physical world, it's going to kick start a massive amount of demand for commodities.
(58:00) And I'm interested to see kind of like what your thoughts are if you even play in the kind of commodities metal space. [Chapter: AI's Role in Resource Extraction and Commodity Prices]
(58:05) [Brandon] But that's kind of another, you know, like you've got capital, but then you have resources, right, 'cause this isn't software.
(58:11) >> [Astrid] Yeah, the answer is yes, but it's so very rare. The usually the only reason I'm getting involved in a commodity is because there's not the direct bet that I want to make available in public markets. But the when it comes to increase in resource demand, yeah, I think that the lid has been blown off and that like the trends that we've seen kind of since the 1980s of resource demand specifically in the US and Europe has been kind of steady to down is completely reverse course.
(58:50) And I don't, I think that that's a material regime change that continues to not be priced into all number of things, whether that's companies that are out servicing the oil sector to companies that are involved in shipping coal and natural gas, like all across the board.
(59:10) But I'm super, super hesitant to have direct exposure to commodities for any large amount of time because the trend historically over time is any increase in demand that increases the price is going to result in people profit seeking.
(59:29) And man, are people in commodities crazy degenerates like nobody's ever seen before. And just like, oh wait, there's a bigger opportunity over there. Great. I'm going to arbitrage this as long as it exists. And that price goes right back down and people invent new stuff. And I continue to be surprised and delighted by the innovations that take place in the Permian, for example, that yeah, it seems that we just continue to unlock resources and our ability to process them cheaper that we already had that people thought was not feasible to extract and or not or refine.
(1:00:06) >> [Brandon] So just to, you know, again, I'm not suggesting you to be a commodities expert at all, but in terms of robotics and physical AI, is there a world where the robots, let's say, like for specialized things, like let's say underground mining, for instance, drilling, blasting, hauling, all that stuff.
(1:00:25) Let's say automation becomes very good and these robots become very good at what they do, mining these materials. Do you see a world where the efficiencies gained from automation and robotics are so high that it drastically lowers the cost to extract, which will then lower all most commodity prices?
(1:00:47) So like you have this massive demand influx, but if automation becomes so good, then the ease at which we can extract those materials to meet that demand, like you said, brings up brings the cost down. Like, do you like, is that a possibility in the next, you know, 5 to 10 years?
(1:01:03) >> [Astrid] Yeah, yeah, totally. Yeah. I think that, well, first the price has to go up such that people are motivated to solve the problem. But as soon as people are motivated to solve the problem, then the problem's going to go away. I'd never, never, ever find yourself betting against human ingenuity.
(1:01:20) >> [Brandon] Because it's an interesting concept because like on one hand, all the easiest material has been mined, but then on the other hand, we have so much more technology today than we did back when all the easy material has been mined. So unlike if you equate for technology, have we really mined all the easy stuff?
(1:01:43) It's something I wrestle with as you know, a commodities bowl as a natural resource investor. But you know, and again, like there's not many ways to play that on the public markets, right? Like you can invest I guess in like SanDisk or Sandvik and you know, Caterpillar or even stuff like that where they do a lot of automation things, but it's not a direct bet on, you know, mining automation technology.
(1:02:03) But maybe that's a space like in private markets that would be very exciting.
(1:02:07) >> [Astrid] That I will tell you that specifically mining, there's a lot of start-ups that have been funded in the last two years that are going after this opportunity to automate everything from resource discovery to extraction to transportation. Yeah.
(1:02:23) >> [Brandon] OK, well, I'll communicate with you after this then, because I would love to get intro to some of those companies just to kind of pick their brains. [Chapter: Debating AI's Marginal Value and Future Demand]
(1:02:30) [Brandon] But Asher, I know we've been all over the place, man, but like, this is exactly the podcast. I thought that would happen, just be all over the place asking you a bunch of questions. Do you think there's anything that I didn't discuss or any questions I didn't ask that you think are worth asking?
(1:02:51) >> [Astrid] It feels like the big error of omission is what's the marginal demand for higher intelligence? And to be clear, I don't have the answer to this question, but it does feel worth mentioning that as right now Open AI and Anthropic have models that are way better than anything that we've seen.
(1:03:14) And well, I'm sure we'll see them next year or sooner. And the question is what's the marginal value that they're going to be able to extract or are we already at a place where the models that are publicly available open source, we can rent, compute and host them privately or even on run on our own computers.
(1:03:38) Is that enough? At what point does the extraction from these large two companies or the rents that they're collecting on intelligence actually matter And do we reach a point where all of the future investment in compute and in training and in training broadly speaking, data collection like is all of that is all everything that we've done up until now completely rational and everything that we do into the future complete malinvestment?
(1:04:12) I don't know the answer. My guess is probably not, but I'm much this is a super super open end question and I'm way way bullish open source use continuing to explode as these models get more compute efficient and people are way underestimating what living in a world where creating like real time content for example, is effectively free.
(1:04:40) That's what it what that means for the advertising market. I don't know if you spent any time on YouTube or TikTok, for example, but like the median ad at this point is probably completely AI generated. And that trend is going to continue.
(1:04:57) And not only are the ads going to be AI generated, but they're going to be AI generated specifically for you. And that's the complete upending of that entire sector. That's ongoing. And it's only going to get accelerated as the cost to run models, it's drops down.
(1:05:13) But this open question is what? What is the marginal value of more intelligence to an end consumer? And who is the end consumer of that greater intelligence? I don't know.
(1:05:27) >> [Brandon] Yeah, I feel like for the average human, like the models we have today are basically AGI, like, you know, the goal post keeps getting pushed back, like the models we have today, like that's AGI.
(1:05:43) But then I think the marginal demand, and I always kind of look at this from like a national security standpoint, right? Like US versus China. Like I don't think that there's an upper limit to the marginal demand because it's like, if this person gets it, then you will have a structural disadvantage forever.
(1:06:00) And so then it's like, OK, like what are you willing to do to not put yourself in that situation? And like for that, like on a nation state level, I don't think that there's like an upper ceiling to the demand, which is probably why like the Frontier Labs will always see that demand. But then on an open source level, like the models we have today are like great.
(1:06:20) And it's probably like, it's almost like a Pareto distribution, right? It's like you can do everything you need with what you have today. And like if you spent a bunch more money to get marginally better, it's probably not going to make that much of an impact for you as the individual. [Chapter: Geopolitical Stakes and AI's Impact on Nation-States]
(1:06:34) [Brandon] But.
(1:06:34) [Brandon] Like if you're fighting a war, right? Like what's the cost of winning the war versus losing a war? Like there is no cost. You know, like there's.
(1:06:42) >> [Astrid] For what it's worth, I think the leadership at every hyperscaler thinks about this exactly the same way that there's no cap to the demand here and we just have to continue throwing more resources and more resets at this thing forever. And as long as literally every hyperscaler C-Suite is thinking about this in those terms, I think you can comfortably buy NVIDIA.
(1:07:05) We'll sleep well at night.
(1:07:07) >> [Brandon] Because like if you think about it in that context, like maybe that's what separates this AI and this kind of AI bubble from prior bubbles like telecom and railroads and things like that. Because I don't know.
(1:07:24) And again, like this is me being naive, right? I need to go back and do some historical work. But I don't know if those inventions and those technologies had the same national security implications as AI has today. And I think that's in two ways, like 1, like the world is so more interconnected now than it was back then.
(1:07:43) But then two, like the relations, like geopolitically, like it's very unstable now. And the two kind of competing nations like are basically at a Cold War technologically that didn't exist like back during the telecom boom.
(1:08:02) And then even you think about the railroad boom, it's like, OK, like, if the US built the most railroads, like, did they have a nationally strategic competitive advantage against, you know, whoever else they were competing with? And it's like, I don't know if that was true. And so then it's like, maybe you do have this just structural difference where what's good for the US in terms of national security kind of Trump's everything else.
(1:08:25) And that creates a different dynamic for this bubble than prior ones.
(1:08:30) >> [Astrid] My instinct is in a completely different place. I think that there absolutely was a national security element to both of the bubbles that you mentioned and further that this has nothing to do with capital allocation decisions at Google and Meta and Entropic Open AI.
(1:08:49) They're just looking for what's the marginal return on my investment. And the lowest number that I've seen any of those companies model is 20% like all in costs as they can get a almost contracted return.
(1:09:05) And as long as we're living in that world, resources are just going to continue to be thrown at this thing forever. And the price of lending for other use cases is going to continue to go up because they're competing with building a data center and getting a 2520% return.
(1:09:23) Why would you buy bonds?
(1:09:26) >> [Brandon] Yeah, just by just by data center debt, right? Just by like Bitcoin miner debt to Tier 1 hyperscalers like PUD 8 issued like 7% yield to worst debt. Like that's not bad when you know treasuries are getting 4 or 5% and you can get a Tier 1 tenant.
(1:09:44) So that's interesting. Like the whole telecom railroad had national security implications as well. I guess all technology does, right? Maybe that's the counterpoint. All technology has national implications. And like that's right. The consumer use case is probably like the last iteration of the technology that's been built in the shadows for the military complex.
(1:10:03) >> [Astrid] I think that the industrial complex attached to the war machine is way further behind on this, on the curve on this than you would give them credit for. Yeah.
(1:10:16) >> [Brandon] Wow, well, I feel like we could talk for like 7 hours about this. I know you've got a hard stop at 3:00, so I want to respect that. [Chapter: The Power of Scientific Advertising and Final Thoughts]
(1:10:23) [Brandon] But I had AI, had a blast. I learned a ton. I'm really glad we had to. We got the time to do this. Where can people go to find out more about you? You're on Twitter.
(1:10:33) >> [Astrid] Yeah, that's probably the place. It's just my name on Twitter, Astrid Wilds, Wilde and I post multiple times a day about all number of things, including most of the stuff that we just talked about. All right.
(1:10:48) >> [Brandon] And then the last question I have for you, Astrid, if you could have dinner with one person from the past or the present, who would it be and why?
(1:10:55) >> [Astrid] Give me Claude Hopkins. He's probably most well known as the author of a book called Scientific Advertising. But the man has reshaped so much of society, both directly and indirectly, that I can't help but think of the guy every time I'm brushing my teeth or drinking orange juice.
(1:11:14) Or like, there's so many things that are part of everyday life that didn't used to be that are only part of everyday life because of insane advertising campaigns and the shenanigans behind them was crazy. Like the fact that we drink orange juice and think of that as a drink even at all is literally the results of the efforts of one man trying to solve a problem where we had too many orange Groves in California and he was trying to prevent them all from getting chopped down.
(1:11:44) And like this is there's so much of bacon is another great example. Like there's so many silly.
(1:11:52) >> [Brandon] Hold on, what's the story with bacon? We still have like 10 minutes.
(1:11:56) >> [Astrid] It's the same results of an advertising campaign. In this case, it was in England where it like advertising that this is part of a normal breakfast and like through targeted advertising over a period of literally only a couple of years, this thing that used to be trash and thrown away was it became a sought after dish.
(1:12:20) I mean, there's so much like can so many consumer goods are just the results of better advertising and scientific advertising is a brilliant read. You just have to read it.
(1:12:36) >> [Brandon] Do you think these LLMS are just advertising, scientific advertising meant to distract those?
(1:12:45) >> [Astrid] Interesting question. The answer is no. I think the companies that are behind the LLMS are kind of weird religious organizations that are very bad at advertising actually.
(1:12:59) >> [Brandon] Well, that was like the whole Navier's or Navier Stokes. I was chatting with a buddy of mine who was like open AI just completely, you know, fumble that whole process. Like he said they should have reached out to the scientists and been like, hey, like we saw that you got basically to the goal line.
(1:13:17) Why don't you publish this and say, Hey, powered by or assisted by open AI And like, we'll give you all the credit. And instead like this whole Pandora's box of like open AI just taking it and then running with it and then publishing it like it's crazy.
(1:13:30) >> [Astrid] It the degree to which these people are people that are influenced by rumors and is wildly underestimated. I mean, the fact that they even went on this journey to try and solve this equation and threw probably 6 to $10 million at the problem is only because they thought that somebody else already had arrived at the solution, which they hadn't.
(1:13:54) And it's a very, very funny. But ultimately, in this grand scope of history and none of this probably matters, but it is a very there's many, many PR bungles and opportunities and anthropics entire like media strategy or thing that they tell employees to like talk about the fact that we think that we're all going to die in a few years like this feels and crazy and insane to me, and I'm trying.
(1:14:21) I can't imagine a future, but I'm soon get going to be living in one where these companies are publicly traded and we'll hear the most insane things you've ever heard on an earnings call.
(1:14:32) >> [Brandon] Yeah, no, it's going to be wild. It's going to be wild. So Astrid, thanks so much man. I had a ton of fun. Can't wait to release this and I look forward to chatting with you soon.
(1:14:43) >> [Astrid] Yes, been a pleasure.