The AI Revolution Call

Maybe. Here we go. Hello and welcome to the AI revolution. Call. I am Cody Willard. It's Bryce Smith, Craig Delon, John d', Oliveira, Cody Bryson. Smith. I'm sorry. Cody Bryson and Craig work at 10000 days. We run hedge fund stuff, invest in technology. John Dio is a legitimate, he actually contributes to society. He actually works in the field of AI and generates applications and services that benefit mankind and generate economic activity. Unlike us who push buttons and try to allocate. I guess we allocate capital and if we do it effectively the world benefits. But it's a derivative, unlike truly working. I have to leave at 3:25 today, take my eldest daughter to the doctor, she's not feeling well, no big deal, but we're going to go get her checked up. What's that?

Mountain time.

Mountain time? Yeah. 325 mountain time, 25 minutes. The look, we're talking AI. Before I got on I was just talking a little bit about how the bitcoin miners of the world, they're, I don't know, probably two dozen publicly traded bitcoin mining companies and they were grifty at best, fraudulent at worst and probably going bankrupt as of two years ago. And those stocks were down 99% from their highs. They rallied over the last year, year and a half, two years and some of them are up 10x20x and I fought it for a while because they were grifty, somewhat fraudulent companies. But the fact of the matter is the AI revolution is so huge that and the demand for energy is so huge those guys business model was trying to secure the best, cheapest energy to mine bitcoin mostly at a loss still. But now they can turn their bitcoin mining facilities and energy contracts into AI. And the hyperscalers are happy to buy basically an empty shell of an AI factory and then they'll go throw in their servers and then the, the energy connection into that data center has pre existing contracts at good prices for AI, not good prices for mining bitcoin though. So anyway, it is interesting. One of the things I challenged the guys this morning, what I'm trying to get here to, among other things is that we should be having deep thoughts about how these new applications and services and businesses that are going to be built on the AI platforms are going to change the world and create new economic activities, new business models. They'll disrupt some things like higher education, they'll change some things forever, make them more profitable, they'll create, I was just thinking, you know, Amazon web services in the year 1998 to the year 2005, Amazon was an Internet company. And people thought of it as an Internet company, but it was a retail Internet company. And nobody thought through the fact, not even Jeff Bezos. It was one of his is president of Internet Stuff Management. I t guy who came up with the realization that they had all this excess compute and Internet capacity for when it's not Christmas time or some big event. And they obviously, all Internet companies overbuilt so that they could try to handle as much traffic as possible, which meant that 99 of the time, 99 of the capacity was sitting idle. And he figured out he could turn that on and provide web services to other companies. Thus the name Amazon Web services. And now 90%, 80% of the value of Amazon, the publicly traded company, is in Amazon Web Services. And that's the type of thing that when you just model out and you're like, hey, how many? There's a billion people using AI and we can generate $20 a month in subscriptions and that's $20 billion. And hey, at 325 we'll leave. Yes, that'd be great. Thank you. All right, I don't have to leave at 325. The where was I?

AWS Cloud.

It's not linear. It's not. It is exponential. Even the number of businesses and business models and revenue streams that get generated out of the AI revolution, just like out of the Internet revolution, are going to be bigger than we picture and they're going to be different than we picture right now. And you've got to try at least as technology trend investors and AI people to think about what we're not picturing yet and try to extrapolate out different trends and different revenue streams and different modes of using the technology that is going to get here. And we're still the very archaic stage of AI and its use and how what we do with it and what we can do with it. But there's infinite number of business models to be built in the next 10 years on AI platforms.

Lately I don't feel like the growth is going to be smooth, though. I mean, to talk about disruptions, I feel like there's going to be more loss before gain, not necessarily in business value and things like that, but in jobs, you know, that I feel like there's, you know, from the programming side that, you know, Dario Amade, head of Anthropic, was talking about, you know, the. All the software engineering jobs going away, they corrected that and they said they really mean the programming Part of it. And that people will still have plenty to do with all the other aspects of software engineering.

That's where I'm always at, John. I feel like it's, it's just almost like a cliche at this point to worry about the impact on jobs from AI. But I mean, clearly it should, it is such an important thing in society, in the economy and, you know, people's personal lives and livelihoods that of course it, we should talk about it. I just. The amount of like, there's so much more work to be done. Every doctor that can save himself 30 minutes for record keeping a day, that's over a lifetime of service. That's 30 minutes or more patients a day times thousands, tens of thousands of days. And there are literally billions of people that don't have access to health care. There are billions of people that don't have access to teeth cleaning. There are billions of people that don't have access to blank. And all of that is, is. Can at least, should at least be addressed over time with AI and robotics.

Yeah, I think. Oh, sorry, Bryce, go. Well, I was just saying, you know, to your point, like to the healthcare point, I think that's one of the biggest things. Like. Yeah, I mean, there is going to be, you know, some job loss in the, in the, you know, interim. But you know, on the back end of all this, you know, I, I used to work in health care and one of the biggest parts of rural care and Medicare is getting people adequate health care and access to different types of health care, like pulmonologists. Like, how much can you pay a pulmonologist to move to little of nowhere, north Louisiana? You know, so the idea of maybe having like you know, next. Next to data centers or something, the next big thing they'll be building is not necessarily hospitals, but almost like big mega clinics. Every single, every single type of specialty that you have is right there done with robots or whatever. You know, the future for medical.

We're done with human doctors, but the person is standing in a holodeck like from Star Trek, and the doctor is literally. Yeah, they have the tools literally figuratively projected into this room with them and is basically for all intents and purposes there. And the robots can do what they can do. And if you have to go to a major facility, you know, you end up getting them there if they, if the, if the economic system will bear it.

Yeah.

And I guess that's it. Craig.

Yeah, my point with that is like the abundance that comes, you know, after the pain. I just, I I feel like it'll be worth it. You know, it's easier for me to say, but you know, I just, I just feel like.

Well, and I guess the point that I'm getting at is that whether it's construction workers or doctors or nurses or programmers, there is more work to be done. If you're a programmer who loses their job, you're not. You're telling me you don't have a single business idea that you couldn't build with AI apps and your talents now, like, go be creative, like beyond.

That's a big ask. There's, I mean, some people are creative, you know, and some.

Right. And I'm not saying everybody needs to be, but there are millions of people who are creative who will now be unleashed because they can create their own businesses with the lower barrier of entry and less capital because they can use AI tools.

One of the things I think is going to change a lot though is that AI and robotics is, is gonna, is so deflationary that it's going to make price things affordable for people that previously couldn't afford them. And so like you're going to have services and goods that were previously reserved to the wealthy and the cost starts to go down and I, it's not gonna. Demand is not gonna be a linear expansion with the cost decline. Right. So if you lower the cost of air travel by 50%, you're gonna get three times as much air travel or something. Right. As an example. See what I'm saying? Yeah. And I think as AI starts to work down the cost curve, the demand just explodes and therefore you're gonna have a lot of employment created in meeting the additional volume at lower prices.

Do we see any of that yet or is it.

I think we've, I mean, I think the history of the industrial revolution is sort of.

No, but he's saying here in the AI revolution, do you have any examples presently? I think if, if it's the dial up Internet phase, you know, there were web developers in high demand in the year 2000, 1999 and at this moment, if you have AI skills at all, you're in high demand. But eventually there were Amazon warehouses, eventually there were trucking jobs delivering goods as they exponentially expanded delivery services and you know, like it's it. I don't mean to make it sound like it's going to be easy and that there won't be political pressures to all of this and there won't be partisan nonsense arguing that it's the other party's fault the whole time and People will be mad at the other party if they're in that party and all that nonsense like oh yes, of course. And there will be disruptions and there will be job losses like professors. Again, I just talking about higher education like it's. There could be a secular change, a generational change going on in higher education that's being started right now, but it's not even having any impact. But somewhere over the next three to five years maybe, and somewhere over the next 10 to 20 years. Absolutely. But hopefully people won't be going into professorship jobs right as that secular train is taken off from the station. Here's a thought. We've spent the last, I've spent the last 25 years being a revolution investor trying to invest in the most impactful technological revolutions that change society and economies forever, create trillion dollar economies. We right now are focused on three revolutions. Elon Musk of SpaceX and. And Tesla, fame of political infamy. Was in an interview last weekend and he said other than Tesla as an investment, he doesn't have a portfolio. But if he did, it's a Tesla, Nvidia and Google, those happen to be our three biggest positions for, for many years, frankly. Google, I guess for the last eight months we made it one of our biggest positions. But it's been a big position and a core position. I've owned it personally since the day it came public. The other point he said were there are three revolutions you need to be invested in right now. Now there are three revolutions that we have been literally pounding the table on. Bryce, he said we should invest in the AI revolution. Craig, he said we should invest in the robotics revolution. And John with us, my other revolution. I always talk about space by the.

First two sounded like the same one but, but space, yeah.

So anyway, I just wanted to. That's my victory lap on that topic. I am the richest man, most successful business person in the history of the planet. Says that the approach that we're taking is probably the best approach for investors. I agree with him.

He could have just said it with the shorthand of, you know, just invest the way Cody does.

I mean that's what he was really.

Saying that about Tesla going to have a 30 trillion dollar valuation someday. Because Bryce's model literally put that together. We published a book about it and then six months later Kathy Wood starts talking about it and then a month later Elon starts talking.

I mean the difference between you and Elon is he's saying it and you said it 10 years ago.

That's the only difference Between Elon. No, but I. I did just think it was funny. And, you know, we're on here doing the AI Revolution call weekly, and we obviously always end up talking about robotics in space, too, because, you know, John, you were making a joke that AI and robotics are almost the same revolution. But one of the things we talk about a lot when we're out giving our investor presentations is that the. The Internet, the electricity revolution, beget the Internet revolution, beget the smartphone revolution, beget the AI revolution, which begets the robotics revolution, which will beget the space revolution. AI enables robotics, and robotics actually will go to the moon and Mars before people do and help build, you know, habitat for humans.

That's what I was talking about a long time ago. Like, everybody kept talking about going to Mars, and I was like, they're not listening to me. They need to send the robots to Mars first. And they. They knew that, I guess, but. Or else they finally heard me. So.

Yeah, no, you know, do you guys remember when they rolled out the Optimus and it was a human being dressed in a robot outfit and it was dancing around? Truly a PR disaster. It was an investor relations disaster. Like, the stock was down the next day and I was already a. You know, it was probably. It was my biggest position. It's been my biggest position in the hedge fund for seven years. And I was like, oh, man, I don't even want to own it tomorrow.

It's like throwing at the ball at the window.

That one I actually got a kick out of. And I've always thought he did it on purpose as a PR stunt because we would talk about it forever.

No, I don't. I don't know if that was on purpose.

Yeah, maybe not. But, John, you know, we haven't done the call in a few weeks. Like, what. What are your. What do you think the most important AI developments over the last couple three weeks have been?

Well, it's interesting you mentioned Google, because they kind of had a couple very good weeks and. Or even in one week, they had a couple very good things happen. They had, you know, gemini was it 3.13030. Oh, three zero. Gemini three zero. But, you know, I think even bigger is the. I hate the name, but the Nano Banana thing, the Nano Banana Pro, which they come with these names because they do, is they have the. They go on that website where things compete before they're revealed about what they are, and they're given, like, a code name. So it's like one of those tech names. But the. The breakthrough here in generating graphics, a couple things. One, is it doing it in combination with Gemini to where it's not just kind of pushing out whatever it can and then. Well, it didn't. It's not quite what I meant and so forth, but it's kind of in combination with the language model which is controlling the objects mentioned. So you can do stuff like say, here's a dozen different individual things, this dress and this picture and this whatever and this whatever, and make one image that contains all of these things. And it's put together as the stuff is not just sitting separately in the image. They make up a single picture. Whether the stuff brought together amazing and just showing the control that things have now. Another things that is that it can now be like perfect with spelling and numbers and words and things so that you can basically ask it to do anything, especially things like this, a big infographic you can like with one sentence, describe, you know, tell me about Warren Buffett's such and such that happened over time and blah, blah, and make it a treasure Mac map or a timeline or whatever, and it'll produce something like this. This is unbelievable.

Yeah.

Knowing how bad it was six months ago when you would ask it to do something and it even just put the word yes at the top and it would look like it was Russian crylock or something. And you never, you were always like, why can't it spell? And to now see this document produced at that quality with those full sentences. I, I, you know, unless you've been doing it, maybe it's not as mind blowing as it should be, but I'm telling you people, that's mind blowing.

Well, I, which visual, you know, that you can, you want to convey an idea and you want to make a visual around it, or you do things like, you know, I remember, like, when Sora first came out, I talked to my brother who's a Hollywood producer, and I said, you know, Hollywood's gone in 10 years. And now I see what is it two years later, maybe a little less that it definitely were on path to that. You know, you make a beginning image and ending image with the same character and, and say fill in, in between and storyboard the whole thing and start to really make things, you know, movie, like, out of it. It's. It's crazy.

And this is. Go ahead, Craig.

Oh, sorry. I just wanted to say I made these on. On exactly what you're talking about on nano. And it was like, I did, I asked it to first Gemini, create a whole outline of the investment philosophy, the wins, the Losses. The, the, the world views of some of the greatest, you know, investment minds, you know, like Stanley Drunkenmiller and, and Warren Buffett. And it really had these long outlines of like, everything, their, their influences, everything. And I really wanted to see how complex it could get. And it blew me away with not only how complex it got, but I mean, that looks like Stanley, like the other one looks like Warren. And it, it's, visually, the fidelity is incredible. And I was, I was, I, I.

It's a, like, those are quality products that you could literally print those out and give them out at a brokerage firm to your clients. And they would be like, oh, this is, you know, like I'd sit on a subway looking at that for hours, hopefully not too long on a subway.

But, you know, it was interesting. I did some, was working on Waymo recently. John and their CEO gave a talk three months ago and he was talking about how they. It's using the Gemini model that just made that, but in reverse because you have all these cameras taking images on a car. You put Gemini into the car and it can help the language model or the vision language model can help interpret the scenes that are happening in real time. You know, instead of just making images, it can also understand the image, you know, and I never thought about that. But I mean, you can take any image and put it in Gemini and like, yeah, I can see a stop sign. You can say, hey, there's a person, whatever, put that in the car. I'm going to help it decide and reason and inference in real time.

That's actually, it's fascinating because, you know, you think large language models, and we've talked on these episodes before about the idea of how the large language model is like triangulating the next most likely word in order to make sentences, in order to make paragraphs, in order to make reports. But it's written words and it's text based. And this is where like the fact that it can also then take image and use processing of large language thought. And this is where even like the human mind is magical because we do that, right? Like, we think in words, but we don't think in words. And you dream, but not in words, but sometimes maybe. And words exist, but reality would exist even without words. And visualizations would of things that you don't even have names for are still there. It's. And the mind does it all and sort of muddles through and, you know, helps you from not tripping on the crack in the sidewalk and have a conversation as you're Doing it?

Yeah, just the idea that you can get. I mean, I've done this with ChatGPT before. I think everybody has. Like, you know, one time my sister needed her car worked on, so I got under there. I was in a really, just really uncomfortable spot trying to get my phone to take a picture of this part and then he could walk me into how to take it off and do it. But, you know, Gemini just its ability to really understand and, you know, take the, all of the information that it can process about what it's looking at, the model of a vehicle, or even like just how to change something on your toilet. Look at it, walk you through it 100%. And there's, there's YouTube videos, YouTube channels just built on that idea of just taking something and just helping you walk through it. And it's just going to get easier for people as they get smarter over time.

Not to, not to take away from Google, because I thought they, they had an amazing week and did great and everything, but just when Gemini 3 came out, everyone was like, this thing just leapfrogged everything. And then two days later, Anthropic Opus was ahead of Gemini on the programming ratings rankings. So Gemini still great, you know, across the board for other things, but Anthropic retained its top position as the programming model.

Do you think it would have been. I mean, I feel like it would have been detrimental for them if they, if they really couldn't beat in the coding department. Because that's their whole gig, right? Like, that's their.

Yeah, well, there's more than just kind of the code itself, there's the whole kind of ecosystem around it, you know, with Claude code being the accepted, you know, thing being used by developers and everything. But yeah, they need to stay near the top for sure. It is where they aim. But it's just that Gemini did take that number one position and it did show, you know, we're ahead of open source China, which, you know, was, was catching up. And it looked like, oh, can we even stay ahead? And, well, we're ahead again. And from the, like, interview with Ilya this week, which was really good if you saw him being interviewed by Dwarkesh, he kind of conceded that the large language models and the transformer kind of architecture is not by itself going to bring about AGI. It doesn't matter. It's not. It is going to keep advancing, it is going to keep getting more powerful and making more use cases happen and so forth. And maybe AGI is less, you know, relevant, but for it to Come about we're going to have to see some other, some other kinds of advances as well as the. Is what he was agreeing.

I think.

Do you not think that we'll get them, those advances?

I do think we'll get them. Yeah, yeah, yeah.

I mean that's again it's like we, we can look at what large language models are now and it would that I'm sure there's got to be parallels to the Internet again where it was like you, you looked at a stagnant website that had some, mostly text and some bad jpegs on it and you thought that that would be the Internet. And eventually people had, you know, created HTML5 which enables you to play video games in 4K in real time with low latency on a website, in a.

Browser and react all the different things that came along to. For that front end stuff. Yeah, I think we will and I think we're not in a. I mean the, the large language models and so forth. They suck a lot of the oxygen out because a lot of time gets spent on those different hyperscalers, you know, competing each other with each other over that specific thing. But there's enough leftover of people around the edges trying all new kinds of things. And I think very interesting and some of the new things maybe what you know, it's sitting in a white paper or a research paper right now. The thing that's going to, you know, you know, break through.

And I always, whenever I see ratings of the various LLMs and large. The large language models and the you know, chat bots that agents etc. That you can use right now. I always just. It literally harkens back to 1998 and reading reviews of search engines and you're like AltaVista is a little bit better if you're looking for music, but if you're trying to find research you should probably use Excite, you know, like whatever the examples are. And it was all. None of it really mattered five years later when Google came along and trounced it all with its better algorithms and approach to the web itself. And it that those analogies are out there constantly. And I'm always. I think it's important for us to think them through and just how transcendent the like AI can become, the business models can become. In AI the models are. Are. I was watching Star Trek the Star wars the other day. Empire Strikes Back it flip connecting of Bluetooth speakers wanted to hear science fiction movie. That was my, that's probably my go to Star wars science fiction movie. So I go to Empire Strikes Back and I turn it on and I'm watching and I'm thinking about is I like, you know, they're like.

The thought.

I had was all of the different lightsabers tools, space travel, hyperspace. I never thought as a kid about the implications, the economic implications of all of it and like rules about technology and lock in costs and how the Empire would get locked into like their own version of like even military industrial complex contractors and. And then like the rebels can use new technologies and they could adapt faster and, and, but also just even like space travel and economics or in Star Trek when you are watching and you're thinking about like these guys are basically like peace military guys on Enterprise out there exploring new worlds and it's being financed by like prosperity somewhere. And that prosperity is being created by guys who are making the starship and like designing new phasers and better ways of teleporting and new. And it is interesting how the point I guess I'm getting at is you can almost guys when I talk about like we need to think about the new applications and services that AI enables a lot of it's actually already on TV and the movies and future dystopian things. But you take the good thing out of the dystopia and the. But what they never address is the economics behind it all. Like that's never in the movies. That doesn't make good movies. You need drama and yeah for good movies you don't need like you guys do realize that since everybody uses Land Cruisers, there is no tire business. There's no, there is no good year. And Tatooine.

Well the, the job is used big land things. So they have. I can't remember the name of it but there's probably like a rubber plant somewhere or something.

I don't know.

I don't know.

Unless it's metal.

It might be metal entirely. But the point being is fascinating that maybe that's part of what I'm getting at is when you're watching a science fiction movie to actually picture the society and economy in which Star Trek 22 year 2250 exists. And what that actually meant for them to get to the point where it was economic to have a crew of hundreds of people exploring stuff with tax dollars.

So there's two parts I was thinking of there. One is that even if you. We did some reason whatever stop where we are with AI development, you got what you're going to get. It's not going to go any further. Whatever we've talked about before that there's enough there for us to kind of develop the economy based on what we've just been given there for the next 40, 50 years. It's like, it's a ton of potential in what's already there. So the fact that we keep going up, kind of pushing along, let's get.

Some low hanging fruit in that example. Like literally what's already there is the ability for your computer to shop for you based on your past buying habits, Christmas time. The ability for your computer to generate an eight second clip of, of science fiction idea that you might have. Like what? Like the ability for your computer to see your browser and tell you what you're looking at and research more about that topic for you. What else? John?

Well, I mean I, I put on the side of technology inertia everything that's about to happen with robotics because, you know, we kind of already have the technology we need. They're just, it takes time to bring together the hardware and the software and to kind of train it up and so forth, but it's not, you know, there'll be new inventions and, and developments and things, but even without them, the robotics can, can happen. So I put that on that side of it and say we've got the, we've got the technology inertia for that already. And there's a lot of things then, then you can put with, you know, the technology inertia that, that affect the, the, our physical world kind of being developed and processed faster and better than it ever was. All of the medical, the ability for medical discoveries, you know, it's not about whether we made this or that discovery, it's that do we have already the power to continue to make those discoveries now? You know, the material science, you know, we now already have the ability to isolate one of 500,000 different materials that, with exact properties, you know, that can be used for something.

John It's a really good point to frame just how amazing what we already are experiencing is and how nascent it still is, because it's not being used in corporate America extensively. The tools themselves are still just being developed and getting refined and being tested out with a few employees and not an entire company with 20000 people working there. And but the tools that exist already are amazing and will get better, and people will develop ever more ways of utilizing what's already existing, and they're only going to get better.

Bryce, you're smiling.

Yeah. Bryce has got something on his mind.

Like what is it, Bryce?

No, it was nothing. It was A random tweet that popped up.

So before that, when I started talking about this thing, about the inertia of it, you were. I think you seemed like you were in agreement.

Yeah, I think, like, I. Like, it's like being on the exponential curve. Right. That we can't ever comprehend. It's like we. We assume, like, we. I guess people forget, like, that what's already happened is sort of a miracle. And then, like, the miracle that, like the fee. Like, what. The way things are right now, just the way things are going to be going forward. And the. The actual state of things is that we are, like, in a state of change currently, and people seem to forget that, like, really quickly, you know? And, like, even when you're looking from investment perspective and you're trying to build a model of, like, oh, how does this all work? And you, like, Cody was talking this morning about, like, trying to model the economic impact of vehicles using, like, how horses worked. Just doesn't. Doesn't work. Right. It doesn't compute. And I feel like that's what people are trying to do right now with AI in a lot of ways, because a lot of it doesn't fit within sort of the paradigm we've been in.

Yeah, it's sort of like right now people are like, this takes it to an extreme, but it's like people are like, looking at how people use Sears catalog, Sears, Roebuck catalogs, and they're like, this is what that means for chat, GPT and shopping. Yeah, I'm not really sure that it's relevant. I mean, there's. Yeah, it was amazing that you could send a catalog of items that you could. That people could buy around the country, and then they would order it from you and you would ship it to them. That was logistically revolutionary for its time. And you know what? I. As we were talking today, I just made a note and it had an epiphany about it. And Bryce, I wasn't going to bring it up, but your point underscores the importance of it, and it is. You guys know what Metcalfe's Law is? I'm sure most people watching this don't, but Metcalfe's Law is that the value of a network goes up by the square of the number of users on it. It's actually square minus. It's N minus 1, but you get the point. So if you have a network with two people, you can have two people talking, but if you have three people now, you have this connection. You have this connection, you have this connection. And you have, you get an infinite number of connections. By the time you have 10 connections, you've got a bunch of connections going on. When you get to 1,000 connections, you add one more connection, that's 1,000 more connections. I mean, I'm messing up the math. It's actually exponential. But the point being it gets ever more valuable exponentially and you just have to add one more thing to the network for it to go up yet another huge number every time. That's exactly what AI enables with applications and services and business models. Instant delivery of your app, business service, whatever it is to billions of people who can use it very easily and don't have to like install some software and new app on their phone and maybe a connector to get it to work. Metcalf. Thank you, Bryce.

So I thought the diagram was helpful to see how it compounds exponentially.

So there's a other angle on this which is sort of an inside out view. You, you run a business and what makes up, I always think in terms of small businesses, but you know, you, you what makes up what you're doing to get from idea to market and all of the workflows that happen, all of the value chains and all of the value add that's going on in your business to turn that idea into something to get the supply chains, to get the things built, to get plans happening and so forth. And with AI, more and more and more of those kind of use cases within it can be done by like agents, you know, can be done by AI. Whether it's like we see the example here with this new Google new version of their graphics product where I don't need a graphic designer, unfortunately to the graphic designer, but I mean literally I can get from the idea in my head to a beautiful finished product, whether it is an advertisement, whether it is packaging on something, whether it's the getting a view of what a product or clothing line will look like and things like this, it's like all of it can happen. Almost anything visual, it's okay, got that end of it covered. We already had the kind of end of reaching into spreadsheets and doing planning with those and so forth. So more and more you can see why there's this argument that we're going to have our first, you know, billion dollar kind of company, our first unicorn that has a single employee is going to happen sometime in the next, what, I don't know, 10 years or whatever, but it can happen.

Yeah.

John, I've got a question for you that I've been thinking about. A little bit lately with robotics and stuff. And one of the things is like, you're not. Context awareness. Dang it. The term just escaped me. Object permanence, that concept that humans like sort of innately do. I. I don't even. It's a miracle that our minds can like remember. Like I have a vague recollection of white, like what the Oklahoma City skyline looks like, even though I haven't been Oklahoma City in a decade or something. You know what I mean? Like it would, if something was radically different, I would know that, that like that's not how it used to look. Right. And, and I can tell you like in my parents house, what's in the top drawer of my mom's desk on the left hand side, you know what I mean? Like where she keeps the pins and things. For a robot, it can't possibly store all, every video, all it's ever is seen in memory and then fetch from memory like all of that, can it? I mean that's, that seems like an impossible amount of video.

Break it down. What's, how much does, you know, 24 hours of video take when it's compressed, you know, and now I don't need like 30 frames per second. I need you know, one frame per second or something. It's like, it's not that much.

So you don't, you don't think that's an issue?

No, I don't think the bigger issue is how you, how you pull this stuff, how you retrieve it.

Yeah, it has access to the cloud and it'll have super fast band, low latency bandwidth to the cloud anyway. So does it even need it internally? It'll know everything it could have, whether.

It'S internal or not. I think it's more about all of.

That video to the cloud.

I don't know. I think it's more about retrieving, you know, and what's the, you know, what's.

The.

Most compressed form something can take? And this is something we worked on all the way back when I was at sitecore and it's now a possibility and that is having that infographic that Craig did be described in a single sentence. Because it doesn't matter that it was slightly fading off on this edge like this with a point of this and the light coming this way. No, you know, or if you said, you know, a purple unicorn running through a meadow with woods in the back and snow on the mountain behind it or whatever, it's like you can, in a sentence, that's your compression for something much, much more Detailed, you peel out what matters and then you store it that way so you don't even need.

How does the AI know what matters?

It already has examples of, of doing that.

So, so like let me ask a more specific question. Like Google when they demoed their Google Glass last summer or whenever that was. And it's not really, it was just a highlight video, basically. Like if you have the glasses on and they're like, hey, where did I leave my keys? And if I, as long as I've had the glasses on all days, it's like, oh, you left the keys, they're on your desk straight, you know, right behind you. Whatever.

Yeah.

Like you think, like, how does the AI know that my keys matter? And like the location of the, of something else that I happen to look at during the day does. Doesn't matter or does that need to.

Be your keys move around so there has to be information about them being here, there, the other, that book on your shelf over there. You know, knowing about it once is good enough until it ends up, you know, moving and then so storing the fact that it's there once and having kind of layered information where I go all the way down to the layer of where's the last time it moved? And that's the only thing I actually have to, you know, know about it is it, you know, is, is down there. But yeah.

So I guess just to sum it up, you don't think the object permanence challenge is that big of a challenge?

I think that the part of the challenge has to do with retrieval and you know, you can store, you know, everything, but getting at the fact that, you know, like, it is sort of the thing you said about like what's novel that you're gonna go get at. It's. I do worry about memory related things, but I'm getting kind of constantly surprised at how good things can be. Like the agent system I'm working on, every agent now stores in its own little mini file system on virtual disk all the files that it's working with. So you know, six months ago that was in memory and taking up tokens and you know, was it going to be a piece of the context and so forth? And now it's off this separate thing and it's like this own little operating system where it goes and pulls in what it needs and has two separate files, one that has all the stuff like where the, you know, the glasses are and everything. And the other that's more like a table of contents onto that. Like here's the part that matters. Go grab out this part kind of thing. So there's. There's a lot going on in that area.

John, how big a chain difference order of magnitude or several orders of magnitude fewer tokens that it requires to do what you're describing with your agent. Since the memory is now local, that.

Plus another memory kind of advancement, make it that it's 10x. Because the kind of thing that Claude is doing, if I have a context window, that's whether it's 200k tokens or a million tokens, I kind of spin off another agent to go do something. It gets its whole new set for it. So it starts over. So every time you go and have multiple agents involved, you've got a clean slate. And then there's also this ability for it to go up. I'm running out. Let me compress what I've done so far by summarizing what work has been occurring and keep just what matters and get from a million down to 50,000 and go like it never looks like anything happened to the user and it just keeps going and going. So instead of it being like, sorry you ran out, it says don't, you know, you don't even think about it anymore. You just keep going.

The point of the story is that like the exponential acceleration of the technology is happening parallel in its efficiency and its effect ability to have exponential more memory also. And the number of tokens that you can run per unit of energy is getting 10x every year with better chips from Nvidia and Google. And the whole thing is feeding upon itself and you're getting like, you know, orders of magnitude of better token economics, which enables ever more application services and business models.

Agree. And one more thing to throw on there is what Google just did with their file system. I don't know if it's Google Drive or what, but you know, you throw stuff in and it's sort of like automatic rag, automatic retrieval, augmented generation. So instead of worrying about vectors and embedding spaces and things like this, it's just happening automatically. Go grow your file system, whatever, it doesn't matter. It's going to go find what it needs and pull the right stuff in. So that's, you know, another element. It's all the same thing you're saying it's sort of multiplying on itself in different dimensions at the same time.

You know, in this context, it makes the deep seek moment from a year ago or whenever it was seem silly in retrospect. And this is where our reaction at the time, even on These AI calls and in the first funds was that it was like a non event because you wanted 10x more efficiency per token for better AI applications and services. And even to this moment, if, and it's clearly happening on several layers of the stack in the AI platform itself, that you want all of these like constant exponential improvements in the unit economics so that ever better applications can run economically.

And as we said as many before, it seems to be the case that as many tokens as open AI as Google, as anthropic, you know, can. As many as they can come up with, they can sell. It's it, they're. They're not limited on that.

You know, it's. There's got to be a fun way of, of mocking the old communist adage of.

To each according to his needs.

Yeah, but it's the old. But the way it actually ends up working is how it's like to each according to the amount of time they have is how much work they get. Like it's always the one who works the hardest gets the most work. And that's. I don't know, there's got to be some.

I don't know, right.

To be a way to make fun of communism in all of this. That's all I would say.

Yeah.

So I guess. And the summary of all of this, of this kind of multiplying and the productivity in these different directions and everything is things are looking up.

You know.

Yeah, it's, it's amazing all of the new, the features that come out. It's like, I don't know how many times we've said this on this podcast, but the AI like every week seems like there's something that blows our mind. And in many other revolutions that's not the case or supposed revolutions, that's definitely not the case. Crypto.

And we would be remiss without addressing the Michael Burry, Jim Chanos Capex of the hyperscalers and the NEO scalers who are building all of these Nvidia clusters and now the depreciation schedule are making it look like earnings are better than they are and that the use of these. Let's not call use because they don't argue that they're still being used six years after they were deployed. They're arguing that they're profitable at that point. Price, would you just frame it and address this topic? He wrote a fantastic article for trading.

I mean I think they did, they, they did argue at one time that those things were worthless after a couple years. And then Nvidia and Google and everybody was like, no, we use them all the time. And they're like, well maybe they're getting used for six years, but they're not economic. And so it's like, okay, so Google is intentionally wasting. What is that?

What that doesn't.

Google's going to use something that look, that actually doesn't.

No, no, there are some costs, you know, so anyways.

But what's interesting I think is that there is the debate is this, that these very expensive chips, pieces of hardware, are not going to produce income for very long. They will rapidly depreciate as new generations of hardware come out, which Nvidia is going to release new generations every single year. That's the plan. And you know that Blackwell is like 15 times more power efficient on a token, you know, tokens per wattage.

Nvidia's latest chip in the marketplace. Blackwell.

Yeah, it's like somewhere like 15 times more power efficient than the hopper was, which is the last generation. And what was interesting is when I was dug in on this topic is that the way I see it is that the compute itself is sort of fungible and there is difference in efficiencies. But you can, because of the CUDA software system and the upgrades that come with that, you can run as an engineer, as a software person building AI, your AI could be running across an A100 H1 hundreds and you know, different architectures of Nvidia compute because the CUDA is a layer of abstraction that sort of makes it all fungible to you as a developer. Now, as from a power and efficiency perspective, those things are less power efficient. But as long as the demand, you know, it's a sunk cost. So as long as the price, the marginal revenue coming in from making those tokens on old hardware exceeds the marginal cost, which is fairly nominal. It's just the electricity cooling cost and the revenue still multiples of that, then they're going to keep running those old, those old chips. And the real question is only if, if the demand were to flatline or, you know, that eventually enough supply came on, then you wouldn't use the less efficient method of producing, you know, meeting that demand. But that's not the case. It seems like the demand for tokens is actually growing probably exponentially.

What's really interesting about that is that's actually an inherent side effect. I guess it may be even an inherent feature of it being parallel processing and not vertical process. The other is that it vertical processing, well, what do you call the C. No, linear. Whatever it is, the CPU runs one at a time. And those GPUs run all in parallel. And since they run in all in parallel, if you create a layer of software on top of this, of the stack, then it's like you called it fungible. That's what made me picture. It literally is like in a sponge and it's just moving from one side or the other.

Scaling up versus scaling out.

Yeah.

So why is it that if demand goes up that the price doesn't go up? To take advantage of the fact that there's all these people wanting like that these new models come out and they keep dropping their, you know, their per million, whatever prices, per million tokens sent, received. You know, it's like 15 this way and a hundred dollars that way and.

Right.

But it's, but it's down from what it was every time. It's down from what it was or it's the same. But there's so much greater processing on the thing that is, that's effectively down.

Because the cost per million tokens is going down from the producer side. I mean that's like the, the demand. So like the amount of tokens that a Blackwell can produce is multiples of, of the hopper. I've got a, you know, tokens per second. And so they can crank out. It's because the hardware is getting fundamentally more efficient. They can sell the tokens at a cheaper price.

And then also though, from a business strategy standpoint, this is the gold rush claim part of the cycle. And so they're willing to run AI at a loss as long as they can fund it with either cash flows or venture capital money. In the case of like anthropic or perplexity or something.

Land grab for customers.

It's a land grab. Get people locked in on your AI now and they'll be there in 10 years and you'll be charging, you know.

Yeah.

If your ARPU right now is 20 bucks a month for some customers and 0 for other, in 10 years you're planning that it'll be, you know, generating like Bryce was talking this morning on our call, $30 a month in ad revenue on your retail customers that you lock in today. And you're generating hundred thousands of dollars a month on your enterprise customers that you're locking in.

And there is, there is some lock in. I mean, there's. Because there are switching costs. It's not like, oh, every day I'm going to go, you know, switch this around. It's like there's work to be done if I want to go switch out A model. I can do it. I can stick another model behind.

Right, you can, but there's work to be done. There's potential glitches. There's just, it's, it's not costless. It is a cost. And so that's it. I mean, there's, that is so much. As long as you can spend them, as long as you can pro. As long as you can spend the money right now to lock in customers, you're going to do it.

I had a picture of, you know, this idea. Every time you come out with a new model and it's a new fast or whatever, it's kind of like Spider man grabbing the next or Tarzan grabbing the next Vine. It's like, now you swing on that one until you get on to the next one, and then you swing on that one. You know, it's like, as long as you keep that going because you got to get that next one there before people are like, you know, it's taking you too long. I'm ready to switch. It's like.

What's interesting, John, I think on that too is like, I had this thought the other day where I think I was like, k tokens are sort of a commodity, whether you like your anthropic tokens or your Google tokens or Xai tokens, whatever. But, and, but then I was like, okay, so you want more token. Like there's, there's like a lot of things are scaling at once, which is driving token demand. Number one, there's more user adoption, right? That's going up all the time. Number two, there's more use cases, right? You can use tokens in a chat bot. You can use it to make images, you can use it to make an app. App. So that's keeps expanding. The third thing is that the unlike most commodity, any other commodity, the quality of a token gets better over time. The tokens we produce today are not the same tokens that we produced last year. Right. Oil, on the other hand, is the same as it was when we produced it 100 years ago. And so as in that's the, the part of scaling that like, like all the quality of the intelligence itself gets better.

Okay, there. I think that's better. Because what I. I'm not sure, maybe it is true, but it seemed confusing to me that you were saying the token itself is better. Is it better or what it can do is better. It. There's more.

I mean, it's predict the predicted next word. The next predicted next token is objectively better with each model. Okay. All right.

Yeah. Okay, so the tokens themselves, what it's a semantics at this point?

Well, I mean, not per token.

Like they, they. They do more every, Every day.

Yeah. Think of it as like a hundred tokens or five hundred tokens or whatever. And say when that chunk of tokens comes back and it used to be stupid and now it's brilliant, those tokens are doing. Are worth more, you know.

Exactly. The bucket of tokens.

Yeah.

So I mean, that's the. I think there's. Bryce. We need to write this article up about just sort of how tokens are the pure commodity. But I think, correct me if I'm wrong, here's another point we should make.

That.

The more tokens in this cluster, the more valuable each token is. They actually get more valuable exponentially themselves because what they're creating feeds off of the other tokens abilities to do other stuff too.

I'm not that. Because there's like two kinds of. I mean, we talk about tokens as in the stuff I send in and out as the context for what I'm asking for something. Or am I talking about it more like, you know, the size of the model and what it, you know, handled in building the thing.

No, in this case, I'm talking about the actual usage of the tokens, the inference usage of the token compute itself. The tokens themselves are more valuable exponentially because there are other tokens doing brilliant stuff in the system with them.

Well, if I said that, I would go say that these advancements in rag or what Google did with that, or in memory and so forth, make it so that a company can build up these assets over time and that bigger chunk of assets, that every next thing that comes along isn't just happening in a vacuum, it's happening against that whole set. And so that's why there's so much value in it kind of growing as a. A thing that. That's. That's being referenced every time, you know, so in that sense, it's like. It's like the Metcalf thing. Yeah, yeah.

I just think there's something. There's a. Some. I. Maybe I'm gonna have to think on it and try to verbalize it better. But it seems to me there's some economic difference in tokens versus oil in this analogy. Because the tokens, the oil just does one thing and it doesn't matter how much of it you have. It's a quantifiable amount of energy it creates. And tokens, like, the more tokens, the more the technology creates Somehow I'm gonna have to work on my.

Well, yeah. If you think about back when there, when contest was only like these things first came out and it was. The context was like 1,000 tokens. That's all it could handle. Being even involved in the question and the answer and kind of thing. And that thousand goes up to a million. And now it can, you know, go and interpret a video and tell you about something that happened at some point within it. So it's like the, you know, more tokens being involved matters. Yeah.

Another way.

Sorry, Price. Maybe the tokens are like the city being built on top of the energy grid. Right. Like they're the actual build. Like the build. I don't. I'll have to figure this out. What were you say about Bryce?

I'm not sure this is what you're saying, but one of the ways I think about it too is like I have. I've analog. Anna analogized this. And to steal before, like with steel, you make. Steel is a commodity you can build a building out of, but you can also build a tractor out of steel. And then the tractor enables you to build bigger buildings and better buildings and building, you know, building cities at scale and which are also made out of steel. And tokens are both. Tokens are both the tool and they're also like the end product. Right. The app, and then the app itself, AI features within it which produces tokens, which you want.

No, that's good.

Cody, last word.

We're going to do a whole show about this next week, but here's the last word. That over the last 25 years, half of all the wealth that has ever existed was created. In the last 25 years, we doubled the amount of wealth that was created in the first 6,000 years of civilization. And we're about to do that again off this new base.

We did the same with information. Coincidence or no.

ME thinks not. Could be a booger, but it's rice. Next question, Final thoughts, final words.

I think I like to said some of this this morning, but I just think it's so early even for AI Revolution. Like I've told people even this weekend, like Thanksgiving and stuff, I was like, yeah, it's probably second inning or something of AI And I. I take that back. Like, the more I've thought about it, I'm like, man, it's still first inning of the A.I. there's like we're just barely scratching the surface and. And especially in terms of monetization. Yeah, monetizing AI really barely scratching the surface of that you know, and so I told Cody and then this Morning is like 95 of chat users or ChatGPT users are free and everyone's worried about OpenAI's economics. I'm like, they haven't even monetized 95% of their 800 million users since then.

It's like when the Facebook had all these users and didn't know what it was yet and didn't monetize it. You just had a edu and everybody, it was spreading like wildfire. But that was definitely the first inning of Facebook if I could ever imagine, you know.

So is that your last word, Craig?

No, no, this cool thing I read today, I just thought it was super interesting. So Aristotle's apparently an AI system by Harmonic and it solved this 30 year old math problem by just brainstorming for six hours and then proving it in a minute. And I don't know, that really kind of got to me because the stuff, you know, mathematical proofing and things that you can do in scientific research, you know, as computing gets more powerful and, and more quick, the, the amount of experiments, the amount of things that you can run iterations is just going to be exponentially faster for like the space revolution, you know, will this work? You know, especially when you have an integral intergalactic or space economy. Even if, you know, just in our lifetime having things up, you know, in this in orbit where you just, we start to get clutter, we get more data centers up there, we get, you know, just being able to process all this information, medical, you know, all of the different things that we can do with, you know, gene therapy and stuff, how, how many times we can iterate faster. You know, that was just really interesting to me that we're kind of at that breakthrough. We've seen that breakthrough this early. So that's something to kind of look forward to as we, you know, start to hopefully get into the second inning sometime very soon.

Amen. I think we should just cut it off there. And John doesn't get last words. I'm just kidding, Johnny boy. What do you got? What's the last words?

I think, I think there's going to be a tremendous need for breaking through the friction of how to put all this AI to use. There's a small number of people kind of with awareness of what it can do. There are people, you know, just on chat, asking questions, whatever. But for all of the things that can be happening in business and research, everything, there's got to be a lot.

More.

Done to help kind of bridge that gap and and like be, be the, the, the grease that, that, you know, has greased the skids to get this stuff really put to use. Not sure how yet, but. And it's not Jen.

I think the promise of AI is that it gets to a point where it's just so easy to use. We don't even think about it.

I don't think it's going to do it all by itself. I think that it's possible to go talk to it and so forth, but it's lacking at this point. Kind of the spark to go, here, let me help you get there and whatever.

And hey, hey, Siri, can you help me with anything ever? Sorry, I don't understand. Freaking knew it, man. I knew it. That's a perfect way to end. Thanks, everybody. Good night.

All right, see you later, guys. Thanks, John.

Peace.