YT YouTube2Text

YouTube2Text → Transcripts → What's Still Valuable When Intelligence Is Free

What's Still Valuable When Intelligence Is Free — Transcript

by Invest Like The Best · 15,071 words · 2,370 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I can see all these incredible ways to
0:02use AI and I use a good jillion tokens
0:05and I can see how I could use a good
0:06jillion more. Everyone uses many tokens
0:08as me. We are so short. Like it's going
0:10to [music] be this huge thing that's
0:11going on forever and everyone and and
0:12everyone who's in this build out is
0:13going to make a fortune, right? On the
0:15other hand, the cost per token is going
0:16to collapse and so maybe actually
0:19like we're totally over built, right?
0:21And like I can tell I can tell both
0:22stories like with [music] equal
0:24certainty.
0:39>> You called me saying it's [music] just
0:40intelligence was sort of the theme of
0:41our conversation.
0:42Describe that like simple idea.
0:44>> There's this idea that if intelligence
0:45starts at like an amoeba and it gets
0:47better and better and and then it's like
0:48a monkey and then it's it's like a
0:49person and then it's like Einstein and
0:50then it like goes zoom and like every
0:52and it becomes God, right? I think that
0:54that's something of an intellectual
0:56mistake. What people have wrong there is
0:58they're missing what are sort of the
0:59real limits of what intelligence can do.
1:02And and maybe another way to put it is
1:03like if you think thinking of
1:04intelligence as like an unbounded scalar
1:06like the height of a tree isn't correct.
1:08What it you should really think about
1:09intelligence as is more like a ratio
1:10like the like the roundness of a ball.
1:13And what the ratio is calculating is
1:17intelligence is is a gadget and we use
1:19it to make predictions about the
1:20universe. And so we start with this thin
1:22scattering of facts about the universe
1:24where we have
1:25relatively few of them in the world is
1:26in the universe is complex.
1:28And we use intelligence to interpolate
1:30between them. So as your intelligence
1:32goes up, your ability to sort of color
1:34in the spaces gets better. And some
1:37domains are really amenable to that,
1:39right? Mathematics, um you can, you
1:41know, solve it like, you know, those
1:43sorts of things. And in those areas you
1:44kind of get to perfect.
1:46Um but for most of the interesting stuff
1:49that is most of the questions in in the
1:52world. The mathematical term is they're
1:53they're complex dynamic systems. So like
1:56um what that basically means is it
1:57they're kind of like weather in the
1:59sense that you can't just guess your
2:01way. You can't make a closed form
2:03answer.
2:04You can try to simulate it. Um but if
2:06you simulate it, invariably your
2:08simulation will have some degree of
2:09error in your initial conditions or in
2:12each step going forward because you
2:14know, at a minimum your steps for
2:15example are going to be truncated. And
2:17so what happens is your error will build
2:19and invariably after the simulation runs
2:21for some period of time, you'll be off
2:23by some amount. And so the point is that
2:25you can't actually like for mixing cream
2:27into your coffee or predicting the
2:28weather or thinking about how groups of
2:31people, you know, who are all very
2:32complicated are going to interact with
2:33each other, intelligence just isn't that
2:36powerful.
2:37And and so my sense is that what what
2:39you get is
2:40you get this sort of increasing
2:42intelligence
2:44to the point where you kind of are
2:45making
2:46all the predictions that something that
2:48super smart could make which it turns
2:49out is a relatively small set of all the
2:51interesting things you'd want to know
2:52about the world. To get past that you
2:54then need to go out and get more
2:55information about the world and that
2:56just takes time. You have to like do
2:58experiments. You have to touch reality.
3:00But what one interesting corollary to
3:02that is that the zoom actually kind of
3:04becomes like one of the things you can
3:05turn your intelligence to is reducing
3:07the cost of intelligence which we're
3:08seeing. So what you actually end up
3:10getting is like everyone gets unlimited
3:13really great like very round ball so to
3:16speak intelligence for essentially free.
3:19And and obviously that's a sort of a a
3:21limit statement. It won't actually be
3:22free but the point is that you know,
3:24that you that that what we're seeing is
3:25is that like really good enough
3:27intelligence is rapidly falling in price
3:29and it turns out that really good enough
3:30intelligence will solve
3:32the vast majority of questions where
3:33people kind of need to use intelligence.
3:35>> Yeah, can you say more about why this
3:36notion of like a complex adaptive or
3:38dynamic system
3:40means that actually even perfect
3:42intelligence or whatever intelligence we
3:44have in a in a bottle 10 years from now
3:46or pick your horizon
3:47will still be limited in its
3:49applications?
3:50>> Yeah, so like non-linear dynamic system.
3:51So the the famous story here and how
3:53these were discovered and this this came
3:55into
3:56the branch of mathematics became chaos
3:57theory is um
3:59a
4:00back in the day there was a
4:01um
4:02there was a guy who who was building a
4:04weather forecasting model, you know,
4:06back on big computers that took up rooms
4:07and that sort of thing.
4:09And he runs the model and it makes
4:11predictions of pressure and temperature
4:12over some span of some forecasting
4:14period. And he's like, "Cool." And so
4:17then, just to make sure it's still
4:18working, they reset it and they run it
4:20again. And it starts out basically the
4:22same results, but then it's the results
4:24start to diverge and by sort of the
4:25third, fourth, fifth day of the
4:26forecast, um the results were entirely
4:29different. And the guy thought, you
4:31know, some version of like, "Darn, I I
4:32must have a bug in my system." Um and
4:35begins, you know, the process of trying
4:36to figure out why the computer isn't
4:38producing the same results in repeated
4:41runs.
4:42And it would turned out that the
4:44technician who had loaded the data into
4:47the computer
4:48in the um second run had rounded off
4:52some at some decimal, like one decimal
4:54less, all of the initial conditions for
4:56pressure and temperature and such.
4:58And you would have thought that
5:02okay, maybe that would have produced a
5:04small but consistent error over the
5:05whole forecast period. But what had
5:08actually happened was that small error
5:10through all of the, you know, uh
5:12repeated steps, you take your current
5:13conditions, you calculate your next set
5:15of conditions, they compound and you had
5:17an entirely different forecast by some
5:20small distance into the into the
5:21forecasting period and the
5:22unrecognizable weather. One one one
5:24model said it was 70° and sunny and and
5:26the other model said it was, you know,
5:2720° and pouring. People called it like
5:28the butterfly effect cuz at least in
5:29principle, you know, a butterfly could
5:31flap its wings in Africa and like that
5:32could lead to a hurricane in in North in
5:34in North America or whatnot. And it just
5:36turns out that like many systems that
5:40we're interested in like have this
5:41characteristic. You can know everything
5:43about pressure and temperature, but you
5:44can't model you can't necessarily like
5:46write down an equation that you can
5:47solve easily that would tell you that
5:50there's going to be a hurricane. You
5:51have to like model it. And the model in
5:53turn is is is going to be grained at
5:55some level unless you actually go to the
5:57step of actually having like the whole
5:59universe. And so ultimately like in the
6:00limit, if you want to simulate the whole
6:02universe, what you actually need is the
6:03whole universe, right? Now, you can't
6:05actually you don't actually do that.
6:06What you actually do is, you know, there
6:08are many, many circumstances where
6:09simulation is super useful for
6:11approximations.
6:12Uh and in those circumstances, you know,
6:13you can use intelligence to solve them,
6:15but there are many more
6:16um where it doesn't work. Um and so,
6:19intelligence is is limited in that
6:22instead of the the sort of fundamental
6:25unknowableness, I guess, about the
6:26universe. So, back to this notion of
6:27it's just intelligence. What does that
6:29mean about where you think we're going
6:31from a more practical standpoint?
6:32>> We're all going to have,
6:34you know, a team of, you know, 200 IQ
6:38people working full-time in our pocket,
6:41and that team will have, you know, it
6:42itself will have a team of that it can
6:43deploy of many more.
6:45And they're going to be trying to figure
6:46out all the little optimizations that
6:48we're too lazy to do.
6:50And every company, you know, will will
6:52have will make similar teams available
6:54to their employees in order to,
6:56you know, to to do their jobs. There's
6:57just a lot of businesses that
7:00where that's not good for them. I think
7:02there's um
7:03you know, variety of ways. So, in a lot
7:06of companies, I think that what happens
7:07is
7:08like
7:09consumer satisfies, right? They they
7:11Consumers are boundedly rational at
7:13present, right? They they only have so
7:14much cognitive space.
7:17And so, you sort of go to a consumer and
7:19you say, "Look, we'll take your money.
7:20We won't steal it. We'll pay you a
7:21decent interest rate. Like, you don't
7:22have to think about it. Like, and you
7:24can go about your life." And the
7:26consumer doesn't say, "Look, if I think
7:28about this every day, I could save, you
7:29know, like, $400 a year. So, therefore,
7:32I'm going to have to think about this
7:33every day." The consumer says something
7:34like, "Eh, sounds good enough." Right?
7:36I'm going to move on. But with an agent,
7:38like, you would think the agent will
7:39will will continuously optimize for
7:41them. And there's also companies where
7:43what they produce fundamentally is like
7:45some sort of artifact that is just like
7:46some sort of crystallized intelligence.
7:47So, I'm by far not the first person to
7:49realize that AI could be bad for certain
7:50software companies, but like, you know,
7:52just this idea that like what what an AI
7:53model is is is is it's kind of like this
7:55general-purpose intelligence tool. And
7:57what many software sort of tools are are
7:59kind of like crystallized and hardened
8:01like
8:02pieces, like, projections of
8:03intelligence.
8:05And so, you know, as I've been thinking
8:06about companies, you know, and and
8:08investments, I found it helpful to sort
8:10of you know, assume everyone has really
8:12great intelligence and ask what's left.
8:14And so, the line I've been using when
8:15people talk about some version of the
8:17pushback is often some version of but
8:19you don't understand how hard this is,
8:20you don't understand how complicated
8:22this is, and you don't understand, you
8:23know, how much how many smart people
8:25have had to think for so long about
8:26this. And I sort of my my my sort of
8:28hand-wavy and and sort of approximately
8:30right response is, you know, okay, sure,
8:32but that's all just intelligence, right?
8:34Um and to follow this logic all the way
8:35through to the end is if intelligence
8:37becomes cheap and ubiquitous,
8:39what you want to own are the complements
8:41to intelligence. And so, the complements
8:43to intelligence are um the they're
8:45they're data, attention, like love,
8:48affection, you know, whatever sort of
8:50emotional attachment you want. Um it's
8:52it's physical proximity, it's scale, you
8:55know, it's a little cynical, but it's
8:56it's it's regulatory capture, it you
8:58know, it's these sorts of
9:00things where for some businesses, what's
9:02what happens is
9:04their finite ability to deploy
9:05intelligence, be it economics, you know,
9:08just the cost of it, they're only a
9:09certain size company and they they don't
9:10have access to 10,000 of Google's best
9:12engineers for 10 years, or be it just
9:15the organization is behind or whatever.
9:16Like the ability to deploy way more
9:19intelligence to your business, and your
9:20competitors to deploy way more
9:21intelligence to their business, and your
9:22customers to use lots of intelligence,
9:24it can work out to be, you know, better
9:26for you, or you know, if intelligence is
9:28a barrier to you serving your customers,
9:30like the lack of it, or it can be um bad
9:33for you if intelligence if your
9:35customer's lethargy is kind of a
9:37fundamental element of your of your
9:39economic um value.
9:41>> Maybe pick like Meta or something as
9:42like the perfect example of a company
9:45that I think you believe owns perfect
9:47complements to
9:48>> Sure.
9:48>> ubiquitous cheap intelligence, just to
9:50like bring the example to life.
9:51>> I'm actually going to set aside for a
9:53moment like muse and agents and that
9:55whole thing and focus on business.
9:57Uh and and the core business at Meta,
9:58right, is is is personalized TV.
10:01>> Personalized TV.
10:01>> Yeah, it used to be social media, but
10:03now it's like it's personalized TV. And
10:05it turns out that, you know, people
10:07really like
10:08uh watching Instagram or whatnot. It
10:10just discovered that I love like dark
10:11British humor and it's great. If you
10:13take the a really high
10:16like like you people talk about like the
10:1710,000-ft or whatever. Let's talk about
10:19like Meta from space, right? Like the
10:20the highest view.
10:22They have 250 billion in revenue.
10:25And they spend 250 billion dollars a
10:26year. Some of it gets tax accounted as
10:28for its costs, some of it is accounted
10:29for its capex. It it doesn't really
10:30matter. Just they spend it all.
10:33The question you might ask is like
10:35what do they spend it on?
10:38And a really reductionist view is that
10:41they spend it all on intelligence.
10:43And what do I mean by that? Well,
10:45they're buying data centers. What is a
10:46data center? Well, it's kind of an
10:47intelligent producing tool. What else
10:49are they doing? Well, they're they're
10:49actually literally buying inference,
10:51right? That's just intelligence. What
10:53else are they doing? Well, they're
10:53paying really smart people to write
10:55algorithms. That's intelligence. They're
10:57writing code, also intelligence, right?
10:59Like they're surveying videos. That's
11:01intelligence-adjacent, sure. It's
11:02intelligence.
11:03And so in some highly reductionist sense
11:07and you shouldn't you shouldn't take
11:08these things literally, but I'm just
11:09kind of trying to highlight the really
11:10big directions here. All they buy is
11:12intelligence.
11:13And if you think forward with the fact
11:16that intelligence the price of
11:16intelligence is falling by, you know,
11:18like a factor of a hundred or something
11:19over the next five years.
11:22First principles,
11:23their costs kind of go away. And
11:26you know, that's just thinking about the
11:28sort of the cost side of things. Um, on
11:30the revenue side of things, there are
11:32all of the creators. You are a creator.
11:34How much more
11:37quality um, creative can every creator
11:41that makes content for Meta make if they
11:44have access to intelligence that's
11:461/100 the cost and 10,000 times more
11:50effective or whatever, right? The point
11:51is
11:52probably better.
11:54And then Meta in turn, maybe they don't
11:56take their costs to zero. Maybe they
11:59spend more or more than zero. And they
12:03do a far better job of picking the
12:06videos to show people next, which should
12:09make the content more, um, you know,
12:11better targeted and more enjoyable. Um,
12:14and then of course on the ad side, um,
12:17you know, advertisers could make far
12:19better designed, far better targeted,
12:22personalized ads that then Meta in turn
12:23can target much better. And advertisers
12:25can also build kind of much more
12:27slippery funnels or whatever to convert.
12:28The whole point of all this is that
12:30maybe instead of taking their cost to
12:31zero,
12:32if
12:33with much better intelligence and much
12:35lower prices, maybe they can like sell a
12:36few more ads. All of this is downstream
12:38of of what they possess,
12:40which is not, um,
12:43destroyed by ubiquitous intelligence,
12:45which is data and attention. They they
12:47just have this enormous amounts of data
12:49about, you know, all all these people
12:51and how they view they view videos and
12:52whatnot. And they have an enormous
12:54amount of attention and these things
12:55compound each other. And and so I don't
12:57think that the availability of cheap and
12:59ubiquitous intelligence is going to
13:00cause people not to be interested in
13:01watching funny their 20-second videos.
13:03Um, and so if anything, you know, they
13:05can win some time on the margin, you
13:07know, um, from other things people do
13:09with their time. I don't actually own
13:10it, but like, you know,
13:12a company that benefits
13:14in theory from having a lot more
13:15intelligence is is Grindr, right? Um,
13:19and you know, basically right, Grindr is
13:21the gay dating app, but
13:23it has been constrained over its life by
13:25the fact that it's just not that big a
13:26company.
13:27And it turns out they've got lots of
13:29attention and they've got lots of data,
13:31um, but their ability to build all the
13:34things that you could imagine building
13:36on what ultimately amounts to sort of a
13:38finite size, you know, team and budget
13:41and ability to like create things. Um,
13:44you know, it has has meant in some ways
13:46that Grindr can't be all the things you
13:47could imagine it could be. And again, I
13:49don't I don't actually own any Grindr or
13:50whatever, but we study I study lots of
13:52companies and let's just say it occurred
13:53to me that like this is another example
13:54of something where, um, you know, here
13:56the here the sort of mental model is
13:57like what what if you had 10,000 of of
13:59Google's best engineers to go work on
14:01your problem for 5 years, like what
14:02could you do? And Grindr is one where
14:04you, you know, maybe they could do more.
14:05>> Ramp is the only platform built to make
14:07your finance team leaner, faster, and
14:10better, saving businesses 5% annually on
14:12average so So can stay focused on
14:14growth. Ramp just opened for business in
14:16the United Kingdom. So if you're running
14:18a business in the UK, you can now use
14:20Ramp's AI-powered finance platform to
14:22manage cards, expenses, bills,
14:24approvals, and accounting all in one
14:26place. I run my business on Ramp and so
14:28should you. Learn more at
14:30ramp.com/invest.
14:33OpenAI, Cursor, Anthropic, Perplexity,
14:36and Vercel all have something in common.
14:38They all use WorkOS. And here's why.
14:41To achieve enterprise adoption at scale,
14:44you have to deliver on core capabilities
14:46like SSO, SCIM, RBAC, and audit logs.
14:50That's where WorkOS comes in. Instead of
14:52spending months building these
14:53mission-critical capabilities yourself,
14:55you can just use WorkOS APIs to gain all
14:58of them on day zero. That's why so many
15:00of the top AI teams you hear about
15:02already run on WorkOS.
15:04WorkOS is the fastest way [music] to
15:05become enterprise-ready and stay focused
15:07on what matters most, your product.
15:09Visit workos.com [music] to get started.
15:12Rogo is the AI platform purpose-built
15:15for financial institutions serving tens
15:17of thousands of bankers and investors at
15:19hundreds of leading firms worldwide.
15:21Rogo's AI agents autonomously execute
15:23large chunk of your firm's [music]
15:24workflows. They can screen deals, draft
15:27SIMs, run buyer outreach, and due
15:29diligence on data rooms. All with full
15:31security and compliance handled.
15:33Every action feeds Rogo intelligence,
15:35the firm's context layer, turning years
15:37of deal experience into a secure,
15:39governed institutional memory [music]
15:41that makes each new transaction stronger
15:42than the last. To learn more, visit
15:44rogo.com/invest.
15:47>> I want to talk about the labs.
15:49So
15:50back to the original point, I think lots
15:52of great investors would say
15:54um intelligence, specifically frontier
15:56intelligence, is not a commodity.
15:58Historically, the conversation has been
16:00about the gap in intelligence between
16:02closed-source and open-source.
16:03Increasingly, the conversation is about
16:05the gap between what's publicly
16:06available from the labs and what's
16:08privately used by them in inside their
16:10own businesses.
16:11They seem to have figured out how to
16:12stop or slow distillation. Lots of
16:14people would say that there's this
16:16escape velocity thing happening and that
16:19the frontier token is going to be way
16:21more valuable and useful and that the
16:23labs themselves can then distill their
16:24own models and provide the best version
16:26of the cheap intelligence, too. I'm I'm
16:28curious for you to riff on the labs as
16:30companies through this sort of mind of
16:33an investor.
16:33>> If they have
16:35models that are better and the ability
16:36to produce sort of
16:38adequate intelligence at way lower cost
16:39than everybody else and nobody can copy
16:42it, then it's going to be a great It's
16:43going to be a great business.
16:44>> I guess I have a couple of thoughts. You
16:45know, one is as I've watched models get
16:48deployed,
16:49it feels like one way to think about it
16:50is like you start out with these very
16:51inadequate models that are basically
16:53able to do a very small set of jobs.
16:55Then what happens is the models get
16:56better and they work their way kind of
16:57up the stack.
16:59And it with each successive step up, you
17:02know, the frontier model has unlocked a
17:03larger, you know, call it like
17:05addressable market. But behind it,
17:08what's happening is that once the models
17:10get saturate once the sort of whatever
17:11get saturated, you get this
17:13like you know, you would never use a
17:15frontier model for like a chatbot in a
17:17in a customer service application. You'd
17:18be an idiot. And so what's happening is
17:21um
17:22those
17:24like areas, they they clearly
17:25commoditize. And so there's like this
17:26wave
17:28of hard tasks where you really need a
17:29frontier model and then behind that you
17:31you don't.
17:32And so, you know, the friends I have who
17:34who are big, you know, frontier labs
17:37companies bulls or whatever and
17:38obviously these companies' financials
17:39aren't public or whatever and they might
17:40have an information advantage over me,
17:42but would say something like, "Don't
17:43worry, we can also along the whole
17:45efficient frontier produce cheaper
17:47tokens." Maybe they can right now, maybe
17:49they can't right now, maybe they will in
17:51a few years, maybe they won't. It's very
17:52I I find this this to be very hard to to
17:54think about. Certainly the trend has
17:56been rapid commoditization of of these
18:00things. If you would look at the rise of
18:01Chinese models and the sort of the the
18:03step every step. Uh and at a minimum,
18:06the cost has been collapsing. So, um
18:08even if they do sort of have a position
18:10where the tokens necessary to do tasks
18:12that are sort of well behind the
18:13frontier, it might not be a very big
18:15market um because it just doesn't cost
18:17that much in the scheme of things to
18:18deliver and it is competitive even if
18:22it's um even if they are a cost leader
18:24in a commodity in a shrinking commodity
18:27market.
18:28And then the other thing is that I think
18:29that you know programming was this huge
18:32use of models.
18:33As you push forward the question becomes
18:35like
18:36for the next next next like layers of
18:38intelligence
18:39do you still have bigger and bigger
18:41places to go after and the bull story
18:43would be some version of like yes like
18:44people will pay anything for a model
18:47that can like figure out how to cure
18:48cancer. Maybe that's true but I think
18:50figuring out to go to our prior
18:51conversation I think figuring out how to
18:52cure cancer requires touching reality
18:54and I'm not sure that more IQ gets you
18:56that. If you don't believe that there's
18:58some super incredible use cases where
19:00you can just charge an enormous amount
19:01for the next for like more intelligence
19:03then you start getting to this point
19:05where like the next step up in terms of
19:07like
19:08like you got coding and then maybe it's
19:09like I don't know something else and
19:10then but it gets into smaller niches.
19:12It's like this diamond shaped demand
19:15frontier where like initially you've got
19:16wider and wider uses and then actually
19:18it starts to get more and more rarified
19:19where you actually like how often do you
19:21really need a 300 IQ one might start to
19:23ask and you know look these things are
19:25all going to like we're going to find
19:26out and I don't want to be too dogmatic
19:28but it it strikes me as like totally
19:30reasonable that what we end up with is
19:32some version of like there are very few
19:33use cases for like super elite
19:35intelligence. There are lots of use
19:37cases for like very good intelligence.
19:39That very good intelligence is highly
19:42competitive and commoditized
19:44and if these frontier labs can somehow
19:46become like cost leaders in doing that
19:49there's a margin to make in being like a
19:50cost leader
19:52but then there's also just rampant
19:53deflation
19:55and then this question becomes of like
19:56like you know how big is the market if
19:59you know good enough intelligence is
20:01very very cheap then how much good
20:03enough intelligence are people going to
20:04buy?
20:04>> I'll just keep playing the devil's
20:05advocate.
20:05>> Yeah.
20:06>> So the two things people might say in
20:07this case are Jevons paradox right that
20:09like as it gets cheaper we're just going
20:11to want to use more of it so I'm curious
20:12what you think about that. The second
20:14would be
20:15you know historically
20:17the new uses of intelligence have
20:18typically been discovered by highly
20:21intelligent people. My my good friend
20:24says intelligence creates its own
20:25demand. What about that? Like what about
20:27the fact that maybe it's diamond-shaped
20:29for humans if humans are driving the new
20:32questions that the intelligence is set
20:34after to answer. But what if the thing
20:36itself is creating the new questions and
20:39we train it to do that really
20:40effectively and just sort of like
20:41explore in an open-ended way in the way
20:44that like all of humanity would be doing
20:47led by the people with, you know, 150
20:49plus IQ or something historically.
20:50>> Jevons paradox, right? It's just a
20:52sort of narrow example of of price
20:54elasticity of demand. And in in general,
20:56all markets are U-shape, right? Like so
20:58initially when the product is terrible,
21:00um there's no demand.
21:02And then the product gets better
21:04and people will pay a lot for but but
21:06but it's uh but they won't buy very
21:07much.
21:08And then as the price goes down,
21:11the the size of the market will will
21:13will rise because, you know, you know,
21:15it becomes um
21:17uh you you Jevons paradox essentially.
21:19But then you get to a point where
21:21further price decreases don't actually
21:24drive enough demand and the market
21:26shrinks and eventually, of course, if
21:28you take the price to zero, um then
21:30demand goes to the the size of the
21:32market goes to zero. You know,
21:35one example that I've I've thought about
21:37is um the cost of fiber optic
21:38communications, right? So
21:40or just communications. Like the cost of
21:42moving information used to be super high
21:44and people used moved very little of it.
21:47And and then, you know, through a whole
21:48series of innovations, the cost of um
21:52uh of moving information uh went down a
21:54lot and the amount of information moved
21:57uh went up a ton and the market grew a
21:58lot for a long time and and and then uh
22:01the market stopped growing. And the cost
22:03of moving information kept falling and
22:06the market kept shrinking and nowadays,
22:09you know, um
22:10Cogent Communications I think routes 30
22:13or 40% of all IP traffic across the
22:15internet and it's through one of its
22:17divisions and Cogent, you know, and
22:19Cogent which which is like not even its
22:20biggest division Cogent Communications
22:22is like a few billion dollar market cap
22:24company. This is like just not a very
22:26big industry. Now, I'm not saying that
22:27we're going to get quite that small in
22:29terms of like intelligence, but there
22:30you know, Jevons paradox is um is a
22:32statement about a particular spot on a
22:34curve, which doesn't necessarily have to
22:35be true. And you know,
22:38you know, it remains as your other
22:39question about like can and can more
22:40intelligence create more
22:42sort of desire for intelligence? I mean,
22:44I have a couple of thoughts on that.
22:45Like what may like certainly maybe and I
22:48um
22:49but I started out by saying like you
22:51need to touch reality to like answer
22:53questions, right? And so, if you wanted
22:56to like if you had a model and it was
22:58going to think more about physics, it
23:00could certainly think lots of
23:00interesting thoughts about physics, but
23:01eventually it's going to need to do
23:02experiments.
23:04And that's not to say we couldn't also
23:05do experiments, but now we're bound by
23:08like our ability to
23:09>> real world stuff.
23:09>> physically do experiments. And so, I
23:11think there's going to be this
23:12incredible time where we sort of there's
23:13going to be so many cool discoveries,
23:15which you're already starting to read
23:16about where like this paper was written
23:18and this paper was written and this
23:19paper was written and if somebody had
23:21read all three papers and thought about
23:22it, they would have realized this, but
23:24nobody had. But it turns out the models
23:26read all the papers and they'll they'll
23:27they'll like figure that out and there's
23:28going to be we're going to there's going
23:29to be all kinds of great stuff that
23:30falls out of that. But at some point,
23:32we're kind of kind of finished that and
23:34and then it's going to be about like,
23:36"Hey, you know, there's some way to
23:38think about things that are outside of
23:39what we know
23:41and we have a hypothesis, but a model,
23:43even a super smart model, won't be
23:45better at guessing that necessarily.
23:47We'll we'll be better at guessing that
23:48than the best smartest person, but the
23:50smartest person's odds are super low
23:51anyway. So, it'll be able to come up
23:53with great ideas for experiments to do
23:55and they'll be worth doing,
23:57but that'll limit our ability to expand.
24:00The other thing I just think about is
24:01like my own use of like AI.
24:04Um a lot on a lot of levels. Like one,
24:06just to go back to the commoditization
24:07thing. I've totally commoditized the my
24:09use of models. For what it's worth, I
24:10just it's cuz I'm lazy about how I do a
24:12lot of my like building for stuff. I'll
24:14I'll tell like a GrokBot to build things
24:17and cuz I don't want to use up all my
24:18GrokBot wheat tokens, I then have it run
24:21all of them I subscribe to all the
24:22things and I have it use all this and
24:24and I'll have it build using those
24:25subscriptions on like a server.
24:29And so I actually don't know
24:31what model I'm using anymore. Like I
24:33used to know all the which this model is
24:35better, this one is worse. I no longer
24:36know. All of this is being optimized on
24:38the back end for me and all I know is
24:39that stuff is getting built and like
24:41it's not using up my subscriptions. And
24:42so like we're good. There's this batch
24:44of companies that I'm curious what you
24:45think about which is effectively real
24:47world experimentation companies that
24:48create a loop that's automatic. There's
24:52this happening in robotics, there's this
24:53happening in several places in bio in in
24:56bio labs and you know automated wet
24:57labs, automated testing labs. Um any any
25:00commentary on like that?
25:01>> Super great idea, right? Like brand of
25:03company. Yeah, like super great idea.
25:05Closing the loop in the physical world
25:06is going to be great and it's going to
25:08accelerate the process of learning and
25:10and and this is a great idea. But again,
25:12like there is there are speed limits. I
25:14I guess I'm my my sort of the the thing
25:15I keep coming back to is some version of
25:17like this is amazing and I'm super in a
25:20sense like AI pill in the sense of like
25:22what I think these are things are going
25:23to do and they're useful for. But I'm
25:25but I'm also sort of cognizant of the
25:26idea that like things bump up against
25:29>> The way I like to tell this to people is
25:30I say look, you're really smart so you
25:31must understand. Being smart has limits.
25:33>> Being smart has limits and and so the
25:34ability to just think your way to
25:37solutions is is limited. And so if
25:38you're going to build an
25:40an automatic biology testing laboratory
25:43and you're going to connect it to a
25:44computer and the computer is going to
25:45come up with an idea and and then it's
25:47going to run an experiment and it's
25:48going to iterate. That's awesome. But if
25:50that experiment requires breeding mice
25:52and mice need a year to grow up and die,
25:54that experiment going to take 10. And it
25:56might be a great experiment, you might
25:58learn a lot and that's great. Still
25:59going to take time.
26:01And you know, the lab's only so big and
26:03so you might have 5,000 great ideas,
26:05you're going to pick 20.
26:06And you you the model may be better. You
26:09you your odds your odds of that
26:10experiment working might even be 1%.
26:13Turns out that might be 10 times better
26:15than the average researcher and that's
26:16amazing. It's still 1%. None of this is
26:19to be pessimistic. This is like super
26:21amazing. Like discovering all the things
26:23that we should know that cuz it's
26:24already in our corpus that we forgot is
26:26awesome. Discovering all the things um
26:28like and like better directing our
26:30research is awesome. Automating that
26:32research to move faster is awesome.
26:35We will learn. It will push the
26:36frontier. But the But like this is not
26:39um
26:40This is not like suddenly the model like
26:41looks looks around and says, "You guys
26:43missed it. Like here you know, here's
26:45like you know, some
26:47new world that like you guys can't
26:49comprehend or
26:50whatnot." Like it I don't I don't think
26:52that's where we're we're headed. It
26:54seems like the market has been
26:55especially in private markets and sort
26:57of the leading edge of investors
26:58thinking about AI very focused on
27:00investing in supply related stuff. Like
27:02how are we going to build and serve all
27:03the intelligence versus like what I'll
27:06call demand related stuff.
27:08Someone watching their personalized TV
27:10on their on their meta app or whatever.
27:12What do you think about that first batch
27:14of companies as investments?
27:16>> Super hard. Right? Like I mean, I've
27:18sort of been I've joked I've been
27:20willfully ignorant about them and
27:21probably I'm sort of worse off for it.
27:22But like
27:23you know, there's this big buildout.
27:25And then you're trying to forecast like
27:27inference demand and I've I can't
27:29forecast supply or demand within several
27:31orders of magnitude, right? Like this
27:32idea of like how much more inference
27:34will people use
27:36as intelligence gets better and cheaper
27:38is like a really really hard thing. And
27:41equally hard is how much better and
27:43cheaper inference is going to get. And
27:46then even if you could somehow guess at
27:47those things, then there's the whole
27:48value chain and you sort of need to be
27:51at the bottleneck in order to capture
27:53the rents. And then of course finally
27:55you get to like your actual competition
27:56with your competitors in that particular
27:58piece of value chain.
27:59And like it's just a super hard thing. I
28:01think about my use of AI as a consumer,
28:04you know, increasingly I'm doing stuff
28:06like instead of um like having the AI
28:09chatbot like talk to my daughter like
28:11you know, or you know, text with my
28:12daughter about how to
28:14you know, do math. Now I have it make
28:17like a three blue one brown style video
28:20explainer so that you know, she can see
28:23like a nice five-minute or 10-minute
28:24video on like you know, powers or area
28:26or you know,
28:28circumference or something. It's super
28:29cool. I don't think I would do that if
28:31it was going to be $5, right? And so
28:34there's this thing where like um
28:36there's real price elasticity. The point
28:39I'm trying to make is like if a lot of
28:40demand comes from the idea that like you
28:42know, it's one thing to
28:43like I'll film myself working out and
28:46with my Muse app and it'll like show me
28:48pointers on my form and I'll do that for
28:50very little money, but I won't do that
28:52if the price point is real and I'm
28:53probably much less sensitive to the
28:54price than like most people. So I think
28:56there's like real like
28:58like AI demand is going to be really
28:59sensitive to the to the price. Which
29:01intrinsically I think does limit in some
29:03sense like the the amount in total
29:05consumers are going to be willing to
29:06spend on this.
29:07>> Maybe this is a time to talk about
29:08capital cycles or something.
29:10It sure does seem like the smartest
29:12people in the world and many or most of
29:14the best entrepreneurs in the world
29:16are heavily personally invested in this
29:20supply build-out thing. You know, if you
29:21watch SpaceX's
29:23uh
29:24investor day yesterday, like this is
29:25basically the theme of it.
29:27Arguably maybe the greatest entrepreneur
29:28of all time is dedicated to effectively
29:31doing this, building out the data center
29:32supply, etc.
29:35This is just like a historical question
29:36like um if that's all true and there's
29:39so much uncertainty about supply and
29:40demand and there are other ways to go
29:42build businesses, products, make money
29:45why do you think so many of like the
29:47best entrepreneurs, the smartest people
29:49are so completely dedicating their lives
29:51to the build-out of the supply?
29:54>> This is a huge infrastructure build-out
29:57of like a huge and important technology.
30:00And it would be impossible for there to
30:02be a huge infrastructure build-out of a
30:04huge and important technology
30:06if there weren't a strong price signal
30:09and a ton of talent going at it.
30:11So you may have your causality
30:13backwards. You know, obviously these
30:15thing things work both ways, but like
30:17could you imagine a telling of the story
30:19where nobody interesting who'd made no
30:21money was building out the AI, right?
30:23That's probably probably the best
30:24answer. Um
30:26And and to be clear like
30:28this is a huge and important project.
30:30It's only happening because super smart
30:32and talented people are focusing their
30:33lives on it. It's great that they are.
30:36They will have and deserve to make lots
30:38of money.
30:39It's just very, very hard and from where
30:42I sit and there are probably people who
30:44are smarter and figured it out, but like
30:46for where I from where I sit, it is
30:47very, very hard to figure out where one
30:50would invest along this and like how
30:53like whether or not you'll you'll make
30:54any money doing it. It's not to say that
30:56there isn't like plenty of opportunity
30:58potential there. I've just sort of
30:59decided to focus my attention more on
31:01the longer-term equilibrium results that
31:03come from what the world goes to.
31:07You know, instead of betting on like
31:08which railroad tycoon is going to make
31:10money and how many railroads we need,
31:12I'm trying to think about, okay, once
31:13you have railroads, like what, you know,
31:15that like I I I'd love to have the
31:17foresight to figure out like that like
31:19CPG brands come next, right? Like that's
31:21kind of the um the way I've been trying
31:23to think about it.
31:24>> Can you riff a little bit about what
31:24you've learned about capital cycles?
31:26Like it it
31:27and innovation cycles like this. I'd
31:29love your perspective on having studied
31:30so much market history. If we're in the
31:33middle of one of these, maybe the
31:34biggest one ever,
31:36what does history teach us about these
31:38cycles and how they work
31:40that's interesting to you?
31:41>> You could think about American society
31:43right now as like a household that's
31:44like saving money for a house, right?
31:46Like we're building a house and we're
31:47all kind of feeling pinched, right?
31:49Because all these resources are
31:50dedicated to like building out like this
31:52big AI infrastructure. All these
31:53electricians and, you know, plumbers
31:56and, you know, everything else are going
31:57to do these tasks and as consequence
31:59there's higher costs elsewhere in the
32:00economy.
32:01Um
32:02and undoubtedly we're going to get a lot
32:04of value for it.
32:06And we're going to keep pushing it
32:07because no one really knows. We're going
32:09to keep pushing it until eventually like
32:11we overdo it.
32:13And along the way
32:15you're going to make people have to make
32:16tons of money because otherwise people
32:18won't keep pushing it.
32:20And because things have lead times and
32:23because it's sort of a winner's curse,
32:26you know, invariably
32:28the last person to stop
32:30will regret it.
32:33And it's just so hard, right? It's just
32:35so hard. Um and then you You that with
32:37the fact that there's just so much
32:38dynamism in the technology here, which I
32:40think was less true for sort of other
32:41historical like infrastructure
32:42build-outs, right? But like um
32:44you know, there's just so much change in
32:46in terms of how you do these things.
32:49You know, I mean other people have made
32:50the point which is I don't think is is
32:52necessarily wrong that like even going,
32:54you know, we talked about the models,
32:55but even at like the chips level or
32:56whatever that, you know, the potential
32:58for um
33:00AI to commoditize even things like
33:02Nvidia or sort of recursive
33:04self-improvement everywhere like at the
33:06>> Yeah, like everywhere. And a lot of it
33:07ends up being like just intelligence.
33:08Like Nvidia literally quite literally
33:10like makes
33:12>> flat files that they send to Taiwan,
33:13right? So like
33:14Now,
33:15their flat files are very hard to make.
33:16Um but like
33:18what they have is a lot of people who
33:19think really hard about it, right? And
33:21so in theory you could make them. And
33:22also then the question becomes like at
33:24what point are you
33:26Like what if you step back a little from
33:27the frontier? Like then the cost of
33:29making something comes way down.
33:31>> In that world, would your model be that
33:34physical physical and scale are sort of
33:37the things that you'd want in a business
33:39that you were to back on the supply side
33:40of this equation? TSMC?
33:42>> I don't want to carry myself around as
33:44some sort of TSMC expert. I mean
33:45certainly what they do has better
33:48attributes and I think in that regard.
33:49They've They've also got this sort of
33:50interesting problem where they did they
33:52did this super hard thing that requires
33:54like all this process power built over
33:55many years and is super hard to
33:58replicate. Um
33:59and much less amenable to like an AI
34:01because just because it's it is so
34:03physical and because it it it's all
34:04these little things that are many which
34:06aren't like even legible to an AI,
34:08right? It's like it's like some guy in
34:10who kind of is in the system kind of
34:11knows that you have to do this, right?
34:13Because there's so much demand and
34:14because they haven't been able to meet
34:15it, they've basically caused everyone in
34:17the world to seek to copy their moat.
34:20It's like a not a great fact. Um but I
34:22you know, I think more generally sort of
34:25stepping outside the AI build-out like
34:26if you think about um
34:28you know, businesses We talked about the
34:29meta example. Um
34:32just the idea of businesses that have
34:33like physical proximity or physical, you
34:36know, um scale might might benefit if
34:38they can use intelligence to serve their
34:40customers better or have like lower
34:41costs as as a consequence. I want to go
34:43back through the natural complements to
34:46intelligence and just have you explain
34:47why why it's a complement to
34:49intelligence so that people can go away
34:50and think about okay, what companies
34:51have this feature that might be
34:52interesting. Good example to go talk
34:54about physical proximity. How would you
34:56think about DoorDash or Uber or these
34:58businesses that have local density
35:00effects that matter a lot.
35:03Does intelligence being ubiquitous sort
35:06of naturally magnify
35:08all of the scale winners of this type
35:10where physical proximity is some
35:12is some sort of key advantage?
35:14>> Well, so if you're the low-cost producer
35:16of something, I'll give you a a sort of
35:18stylized example. So like at one
35:20If everyone has an AI agent, then what
35:22happens is many decisions get offloaded
35:25to the AI agent cuz the AI agent helps
35:27you like optimize, helps you pick up all
35:28the like pennies and dollars and nickels
35:29and stuff that you like were leaving on
35:30the ground.
35:31And so let's take the example of So So I
35:33think about it consumer choices in like
35:35a spectrum, right? Like on the one hand
35:37is like maximizing my bank rate, being
35:39deposits. Like
35:41This is something where I think I can
35:42almost entirely like defer to the AI's
35:44judgment. Like as long as it's FDIC
35:46insured,
35:47I just want the highest rate.
35:49On the other hand, spectrum end of the
35:50spectrum might be like picking my
35:52favorite sweater. This is something
35:53where it's just all of my judgment. The
35:54AI can source me things to look at, but
35:56in the end, you know, I need to decide.
35:58It's going to I can't really delegate
36:00that judgment to it. And there's a
36:01spectrum. And so let's just pretend for
36:03a moment that garbage collection is
36:05something where I can defer the
36:06judgment. You might say, "No, I I really
36:07care about the guy and I trust the
36:09service and they're driving to my house
36:11or whatever." But let's just set all
36:12that aside for a second. So
36:14right now, if I'm in the waste
36:16collection business,
36:17um
36:18I might be great at it, but it's still
36:21really hard to get people to bother to
36:22switch. They sort of have like I doubt
36:24you know what you pay for garbage
36:26collection like off the top of your
36:27head. Um many people do, but like I
36:29doubt you do. And so what happens is you
36:32have to do all this work to get people
36:33like pay attention to bother to like
36:35switch garbage collection companies even
36:37if you're offering a better service at a
36:38lower price.
36:40You could imagine that like my agent,
36:42their agent could talk to my agent and
36:43be like, "Hey, I'll save you a 50 cents,
36:45you know, a month." And the agent will
36:46say like, "Will the garbage go away?"
36:48And I'm like, "Yes, the garbage will go
36:49away." And like, no problem, and I'll
36:51change. So, the first-order effect is
36:52that you move to the low-cost provider.
36:55But the general market effects are
36:56actually sort of interesting, too,
36:57because now let's say that there's a
36:58street. And before, like, it would be
37:00more efficient for society if all of the
37:02people on my street had the same waste
37:04garbage collector, cuz the guy could
37:05drive up the street and pick up the
37:06trash, and he could offer all of us a
37:07lower price make and make and make more
37:09more money. But it's really hard for us
37:11to coordinate that. Because what happens
37:12is like all of us are just lazy about
37:14switching garbage collectors, and so we
37:15all kind of have one, and um you know,
37:19the guy with with two guys on the street
37:20can't call everyone else up and say,
37:22"Hey, I've got two guys on the street.
37:23Everyone else has just one. I'm I'm the
37:24low-cost producer. Let me
37:26roll this up." So, a second-order effect
37:29might be that, you know, the guy get who
37:31has two two people on my street can go
37:33to everyone else on the street and say,
37:34"Hey, I'll offer you 50 cents, you know,
37:35a a lower price on the garbage." And
37:37everyone will switch. And so now you'd
37:38get this like like local, you know, like
37:41everyone moves to the low-cost producer,
37:42who in turn gets to be like a lower-cost
37:44producer. So, um
37:47the that actually creates in some sense
37:48like a better like business. Um where
37:50the where the low-cost guy wins a bunch
37:52more volume at like much lower margins.
37:54Um
37:55so like, you know, a lot of companies
37:57have like, you know, have like mixed As
38:00you study them, like a lot of companies
38:01have like mixed effects that point in
38:02different directions. Um
38:05uh
38:06you know, because because of these
38:08things.
38:08>> What do you think the impact of agents
38:10is going to be on your world? What will
38:13be your role? How will it be different
38:14in 5 years? Like, will the best agent be
38:16able to build a portfolio of Carvana and
38:18Hilton Grand Vacations and Meta and in
38:21the same way that you would, cuz you
38:22trained it how you think?
38:23>> better. In terms of what I do, right?
38:25Like, I think that in the long term
38:28we'll probably get competed out.
38:30>> Pricing pricing risk as the thing you
38:31do?
38:32>> Yeah. When computer-aided chess came
38:34along, there was this long span of time
38:36where freestyle chess was better, where
38:38where the human working with the
38:39computer program was able to outcompete
38:42either the human or the computer by
38:44itself.
38:45Um and this is just because humans
38:46thought about things differently and
38:47computers thought about things
38:48differently and they were synergistic.
38:50And I think in investing for for a while
38:52>> We're in that era now.
38:53>> We're in that era. And that's because,
38:55you know, there are things that I
38:57understand in the world in ways that the
38:59model just doesn't.
39:01Um, by virtue of being a person and
39:03having lived a life. And there's
39:05information I have about the world that
39:07just isn't in um, models cuz as much as
39:10as much information is on the internet,
39:12not everything is actually on the
39:13internet. While that's true, which is
39:15going to I think be a while,
39:17um, you know, the ability of me to use
39:20models to um,
39:22assist me and then make judgments will
39:24make better judgments working together.
39:26Um, you could you could imagine a a
39:28circumstance though um, in the future
39:32where that gap diminishes to the point
39:34where it doesn't matter anymore. And
39:36then, you know, models, you know, I I
39:38don't think models will be able to
39:39predict the future as we talked about
39:40like the future is in some ways very
39:42intrinsically unpredictable.
39:44Um, but the model's distribution of
39:45outcomes may be um,
39:48the best that you can do and that adding
39:51a human to the forecast would be kind of
39:53like adding my judgment to whether or
39:54not it's going to rain in 7 days. You
39:55know, Cliff,
39:56I don't really care what you think. I'm
39:58I'm using the forecast. Like this is
39:59this is this is going to be better.
40:01Whatever However you fudge it, it's just
40:02going to make it worse. That day may
40:04come. It's a very interesting like
40:05moment there. Like
40:07there's a lot to investing, right? You
40:08think about like judging management or
40:10something, right? But like maybe maybe
40:12the models will be able to judge
40:13management. They'll be able to interview
40:15them. They'll be able to figure it out.
40:16And then and then it gets really
40:17interesting like how do you raise
40:18capital? You'll be like, you know, I I'd
40:20like to run an amusement park and the
40:21model will say like, "Sure. Here are
40:23like six locations that have a 51%
40:25chance of an amusement park work park
40:26working." And, you know, as long as
40:28you're like really committed to it. And
40:29I've assessed you and you have like a
40:3151.2% chance of like succeeding cuz
40:33you're actually slightly below average,
40:34right? And you'll be like, "Okay, well,
40:36what can I do better than average?" And
40:37I'll say, "Well, you know what? Have you
40:38thought about like running a you know, a
40:40go-karting park?" But um,
40:42you know, unfortunately it I don't think
40:44it's going to affect uh, my career that
40:45much cuz I think it'll take a while.
40:47>> What about replacing price as the
40:49organizing principle and signal that
40:52coordinates economic activity? Price is
40:54a miracle, right? Cuz it's a signal
40:56wrapped in an incentive.
40:57>> What does that mean?
40:57>> So, price is an is an incentive. It
41:01causes you to want to buy more or less
41:02of something because it goes up or down.
41:04Um and it's also a signal cuz it says
41:06something about the relative value of
41:07something and the scarcity of something
41:09and the cost to produce something. It's
41:10like a miracle of a mechanism to
41:11coordinate things. Um
41:14and uh I think you'd be really
41:15hard-pressed to [music] find a better
41:16system even if even with like unlimited
41:19uh
41:19unlimited intelligence. Um it'd be
41:21pretty hard.
41:23>> Vanta automates compliance so your team
41:25can spend less [music] time on security
41:26reviews and more time getting customers.
41:29Vanta cuts audit prep by 82% and gives
41:32you instant up-to-date proof of trust
41:34without all the manual work so you can
41:36close deals faster and with less
41:37friction.
41:38Customers report a 526%
41:41[music] return on investment and more
41:43than 16,000 companies use Vanta
41:45including Ramp, Harvey, [music] and
41:47Snowflake. Get started with $1,000 off
41:49at vanta.com/invest.
41:51[music]
41:52Ridgeline is the only AI-native
41:54investment management platform that
41:55brings the front, middle, and back
41:57office together in one system. [music]
41:58It gives firms one place for their data
42:00and uses AI agents to automate
42:02repetitive work like onboarding,
42:03reconciliation, post-trade compliance,
42:06and reporting [music] so teams can spend
42:07more time on investing, serving clients,
42:09and growing the business. Ridgeline is
42:11built to give you time back that [music]
42:12can change the trajectory of your firm
42:14and be the inflection point that takes
42:16your growth from incremental
42:18to accelerated. Rideline also wants to
42:19help asset and wealth managers
42:20understand and improve their firm's
42:21growth [music] trajectory so they built
42:23an interactive dashboard that lets you
42:25see how your firm compares to 16,000
42:27peers. If you file a Form ADV, your firm
42:30is already in it. Visit
42:31ridgeline.ai/trajectory
42:33to see where your firm stands. [music]
42:35Then talk to Rideline about how you can
42:36do more with the team you already have.
42:39>> Maybe another way of asking the question
42:40[music] is kind of like investing
42:42if ubiquitous intelligence gets so much
42:44better and so much cheaper
42:46and in some way um
42:49it can replicate what you do and you're
42:50really good at pricing risk. A friend of
42:52mine,
42:53Dan, was saying there's sort of three
42:54levels to this. There's I put in a
42:56prompt and I get something back. That's
42:58level one. Level two is I set up a
43:01watchtower. So, I set up a you know, my
43:03instinct every night runs like a set of
43:04tasks and it happens in the background
43:06on a schedule. And level three, which
43:08produces the most use of intelligence,
43:10is it's just always on proactive,
43:12figuring out its own way to use its own
43:14intelligence, and that just like goes on
43:16forever and it's sort of unlimited in
43:18how much it can do.
43:20If I think about that level three
43:21applying to like everything, like what
43:23things should be priced, efficiencies,
43:25optimization, all this kind of stuff.
43:27>> Does that add something interesting to
43:29the price?
43:30>> Sharpens markets, right? There's a lot
43:31of like stickiness to prices, right?
43:33It's a well-known macroeconomic
43:34phenomenon. It's like price stickiness.
43:36There's
43:37a lot of like lazily priced stuff. Think
43:39about deposits and such. I mean, it's a
43:41really competitive world out there, but
43:42it's actually not as competitive as in
43:44theory like it could be. So, I think
43:47you know,
43:48if everyone has agents who are working
43:51on their behalf and constantly sort of
43:52optimizing, I I think you know, the
43:55price of garbage collection will move
43:56around a lot more. As all the input
43:58costs of garbage collection evolve, and
44:00every you know, a garbage collecting
44:02company is constantly like re-running
44:04its financials and trying to figure out
44:05its true marginal cost and bid every
44:08customer where it could like serve for
44:09like you know, more than marginal cost
44:12or you know, and every agent is
44:13switching to the lowest cost provider.
44:16That's just a much like the price of
44:17garbage you won't just pay $60 a month
44:19for 12 months and then have the guy come
44:21back to you and say it's a 5% increase,
44:22it's going to be 63 next year, and you
44:23don't even notice and you just like you
44:25know, like it's they're just picking up
44:26your garbage. Your your garbage
44:28collection will vary by day and and you
44:30know, and
44:31one week it'll be whatever time it
44:33there'll be all this interesting much
44:35more accurate, much more rapidly moving
44:38prices, I would imagine.
44:39>> Have you seen anything interesting in
44:40capital markets innovation? There there
44:42was this awesome economist story you're
44:43talking about the behemoth that is
44:45American capital markets and how good it
44:48is at adapting to new stuff and creating
44:50new new instruments, new ways of
44:52financing things. And that that's been
44:55>> [clears throat]
44:55>> as much a part of the American story as
44:57the technology innovations that
44:59typically happen to cause, you know,
45:01like that there's this amazing
45:02interrelationship between Wall Street
45:04and Silicon Valley would be one way of
45:06thinking about it. Anything interesting
45:08percolating for you in
45:10in markets, in capital markets, in ways
45:13of financing things, things like this?
45:15>> Well, I mean, I just it's fun to take a
45:16moment to pause and maybe congratulate
45:18ourselves as as a as capital markets for
45:20what we're doing with AI. Like this
45:23I mean, how you know, people were
45:24worried about this, the decline in
45:26American dynamism. I think we just like
45:27redeployed
45:29what is it? 3% of GDP, right? Like right
45:31into this like highly speculative
45:32venture basically on the on the turn of
45:34a dime. Um
45:35it's just this remarkable thing. Of
45:37course, who knows? Maybe it should be
45:39two, maybe it should be six, right? But
45:40it is amazing that we can redeploy
45:43resources at this scale um this quickly.
45:47And there's no one else in the world
45:49that's been able to do it. And um
45:52you know, certainly the the inability to
45:56move investment and productive resources
46:00um can be a binding constraint
46:04on productivity growth even with
46:07especially with rapid technological
46:09progress.
46:10And and so um
46:13you know, at least with regards to
46:14capital and obviously
46:16you know, we talk about capital, but the
46:18capital investments are tied to
46:20people doing different things. Like
46:23Americans are doing different things
46:25today. Like every day, like people are
46:27waking up and going to some place in
46:29I don't know, Pennsylvania and like
46:30building a data center, right? And like
46:31that wouldn't be happening if it wasn't
46:33for this whole infrastructure we have
46:36that's somehow been able to move this
46:37enormous amount of capital. It's
46:38incredible. As like deep innovation, I
46:40don't I think the fundamental idea that
46:42like people are taking risk on
46:44speculative ventures and sort of slicing
46:45up the pie in different ways is sort of
46:47irreducible. I think people will come up
46:48with lots of different names for what
46:49are essentially the same things over
46:51time. I don't think that's like deeply
46:52different. Um but it is incredible. I
46:54should also note that like I'm sure
46:55we'll discover all kinds of
46:57like fraud and bad behavior and stuff
46:59eventually like when when like the water
47:01goes out, but I think given the
47:03magnitude of this, I suspect like it's
47:06so clean. Like it's amazing. Like if you
47:07read about like the railroad build-out
47:09or something, like the amount of
47:11you know, like
47:12waste and and fraud and like whatever.
47:15Invariably there is some, but but like
47:17the like
47:18this is like there's a lot of people
47:20trying to basically do the right thing
47:21here and like the governance parts that
47:23seem to kind of we just take for
47:25granted. Like it's incredible. Like
47:26America's capital markets and the
47:28capital markets in general I think are
47:29just like incredibly under appreciated.
47:31Like the idea that you know, you know,
47:33the any person can sort of buy $100,000
47:35of stock and
47:37get, you know, a portfolio of
47:39well-run, you know, well-motivated
47:41teams, you know, sort of building the
47:43companies that are they're sort of
47:44basically doing mostly the right things
47:46and do it basically instantly and at
47:48basically no cost and then sell it as
47:50soon as they're done. Like
47:51you know, is amazing and then the fact
47:53that that that all those companies and
47:54and other ways people invest like can
47:57then allocate society's resources like
47:59pretty well like in the scheme of things
48:01is just I mean,
48:02you know, just incredible. Like the
48:04whole system is just very underrated.
48:05>> Why has America so dominated this, do
48:08you think?
48:09>> My pet theory and everyone has their own
48:10theories, um,
48:12it is some sort of
48:15interplay between genetics, cultural
48:18evolution, and, um,
48:21institutional evolution.
48:22So basically like the people who came to
48:25the United States
48:27were the subset of people
48:29who were most interested in taking risk
48:31because taking going to the United
48:32States was very risky.
48:33And I think that um,
48:36risk taking has a meaningful like
48:37genetic component.
48:39And and so I think that
48:42that
48:43tendency is just more common amongst,
48:46you know, Americans. When you're in a
48:48place where
48:50um, risk taking is more the norm,
48:51there's also a whole bunch of like other
48:53things that come from that. Like um, it
48:55it's it's more the norm and then that
48:57leads to build-ups of sort of cultural
48:59and institutional things that reinforce
49:02that. It's understood that people can
49:04take risk and fail. That leads to
49:05bankruptcy laws that make sense. That
49:07leads to labor laws that support that.
49:08That leads to
49:10you know, venture capital that supports
49:12that. And all of those things reinforce
49:14that. Um and then that leads to success
49:16stories which we So you get kind of this
49:18whole soup. Um but that's my pet theory.
49:21>> So if there's a moon base at some point
49:23you're going to be bullish on that
49:24community's ability to
49:26do really well economically.
49:27>> Yeah, actually that's a great that's a
49:29great way to put it. The United States
49:30has
49:31has been persistently like
49:34innovative in in like a very amazing
49:36way. And so if there's a moon base I
49:38wouldn't be shocked if it's ours. Even
49:39though the Chinese are brilliant and and
49:41numerous um and the Europeans exist.
49:45>> [laughter]
49:50>> When you see rates go up like the the
49:53you see that great long-term chart of
49:55rates let's say the 10-year rate in the
49:56US. And it just looks like this for
49:58forever basically our whole lives and I
49:59was going uh comment on that chart.
50:02>> Well, I mean there's so much and I'm
50:04certainly no macroeconomics expert nor
50:06do I spend much too much time thinking
50:07about it but I'll make a couple of
50:08observations. Like you know, one
50:09obviously is just you know, in general
50:12when society has like you're you're
50:15society and you're you're building it
50:17you're making a giant investment in the
50:18future.
50:20Means you're going to have to consume
50:21less today. Like fundamentally all of
50:22the resources going into this AI
50:23buildout
50:24>> are about the future.
50:25>> Are about the future and they're and
50:26fundamentally like the the plumber
50:28building that today is not making
50:30consumption today. Like he could be
50:32making today's present consumption.
50:33Instead he's
50:34making a data center which is future
50:36consumption.
50:37So one piece of a very complex puzzle
50:39that I have thought a bit about is this
50:40idea that rates are higher cuz we we
50:42we're inducing people to save for the
50:43future.
50:44Um and the way you do that is you pay
50:46them.
50:47>> Simple.
50:48>> [laughter]
50:49>> People would kill me if I didn't ask
50:51about your portfolio a little bit which
50:52is uh
50:53so interesting to me. It's like 80%
50:55Carvana, vacations, [clears throat] 5%
50:57Meta. It's basically your portfolio. But
50:59like rough justice it's like five
51:01things. Sure. And one thing is 80%. And
51:05so if you think about your portfolio um
51:08and those companies which obviously you
51:10know extremely well
51:11and you were advising them on how to
51:13think about using
51:15>> Yeah.
51:15>> AI to transform their businesses, what's
51:17that speech look like?
51:19>> Well, it starts by making sure that
51:21everyone is vibe coding. Starting at the
51:23board and the CEO and like everybody
51:26needs to start building some, you know,
51:28they need to have um
51:30hackathons. And of course I I brought
51:32this to to Carvana and they said, "Yeah,
51:34we call them pizza parties." And I don't
51:35mean like the developers, I mean like
51:36everybody needs to have like hackathons.
51:38Um and it's not so much that you're
51:39going to build hugely useful software
51:41necessarily that comes out of these
51:42hackathons, but it's that you're going
51:43to teach everybody what can be done. And
51:45that's going to change their
51:47expectations, I think, for what can be
51:48done. Um and then the other thing, the
51:50second thing which I I I tell them is
51:52the CE- CEO needs to basically turn to
51:55the head of like IT and say something to
51:57the effect of "So, I want um my Grok
52:02bot, MetaMuse, agent, whatever to be
52:04able to access every system in the
52:06company on a read-only basis. Everything
52:09I could look up like if I like asked
52:10somebody I want to be able to get that
52:13in a prompt." And the guy's going to
52:14say, "That's really hard. We have like
52:15umpteen trillion problems with that."
52:17And he's going to say, "That That's my
52:18point. Like I want that."
52:20And you know, that's a big process and
52:22it's system by system and all the rest,
52:24but if you can get to that um then that
52:28gets you
52:29like read-only access.
52:30>> starting line.
52:31>> Yeah, to the starting line. That That
52:32gets you read-only access. Um and then
52:34the next logical thing is the CEO or
52:36whatever needs to like use this
52:37everywhere and everyone once is going to
52:38get jealous and you're going to say,
52:39"Oh, you can have it, too, but we need
52:40to work out your um permissions."
52:43>> [laughter]
52:44>> And and then that in theory percolates
52:46down and now you've got permission to
52:47read-only access and then invariably
52:49someone is going to say something like,
52:50"I'd really like to put data into the
52:52system." And you're going to say,
52:53"That's fantastic. We got to work out
52:56write permissions. Um and then someone's
52:58going to say, 'I'd love to change the
52:59system cuz we just this vibe coding
53:00thing and I'm confident we could do
53:01it.'" And you're going to say, "That's
53:03also fantastic. We need to work out, you
53:05know, a process by which we're going to
53:06manage changes to the system.
53:08And, you know,
53:10what I just described
53:12like
53:13like sure, it's like four bullet points,
53:14but it's like
53:16that's like a
53:16>> Years of work.
53:17>> It's like impossibly [laughter] hard.
53:18So, you know, the question is can are
53:20you in a business where
53:22you have time [clears throat] for that
53:23to happen before someone who starts for
53:25the jump ball, someone else that starts
53:26from go beats you. Exactly.
53:29And and so it's it's good to have a like
53:31a
53:31um
53:33you know, a really good business.
53:34>> It it seems like one of the conclusions
53:36from that is it's the best time ever to
53:38start companies both because of capital
53:40market conditions that people will fund
53:43at fairly low cost of capital highly
53:45experimental things, but also just
53:47because with these tools if you start
53:49without the baggage, you can go so far
53:51so fast. You think that's roughly right?
53:53>> I mean, yeah, it's the best time to
53:55start a company until tomorrow, right?
53:56So, I agree. It's amazing. I'd also just
53:58say that like just thinking from my
53:59perspective, right? Like
54:01you know, people are such a pain in the
54:03butt, right? Like as a if you're running
54:05a company like they're such a pain in
54:06the butt. And the idea that you have all
54:08these agents and stuff it it makes it so
54:10you can do so much more
54:11like with so many fewer people. Of
54:14course, you know, one of the challenges
54:16always is like if you can sort of easily
54:18do something, then like so can everybody
54:21else.
54:22So, there has to be some
54:25wedge by which thing you build
54:28um produces like enduring value. Does it
54:30have a path where you end up with some
54:31sort of like
54:33block in or proprietary data in this
54:36world where all the agents are happily
54:38you know, changing providers you know,
54:40every day like
54:41what gives you any enduring
54:44um
54:44economics. But then of course then
54:47there's the further thing, right? Is it
54:48you know, obviously if you if you do the
54:50thing where it seems like there's an
54:51obvious path to enduring economics, then
54:53you that's a bad thing to do because
54:55everyone is trying to do that. So, it's
54:57very competitive and the really great
54:59outcomes are the ones that make no
55:01sense. You shouldn't do it at all, but
55:02then it turns out there is a path or the
55:04world turns. do
55:05>> are the most underrated sources of
55:07future competitive advantage then? Like
55:09if if ultimately good businesses need
55:12sustainable edge sustainable competitive
55:15advantage to have great profits.
55:18Anything underrated? Underrated. We
55:20talked about attention and data and
55:22physical proximity. These are examples
55:24of things that might be super or more
55:26valuable in the future than they are
55:27today because of cheap intelligence.
55:30Is there any
55:31type of it maybe it's regulatory
55:32capture. Um are there any types of edge
55:35that you think are underrated?
55:37>> I get the sense that a lot of these are
55:38appropriately weighted. Um
55:42you know, I guess one interesting
55:44question I've had I'll I'll two of these
55:45I was like one is does this make margins
55:47like if you think of corporate margins
55:48in aggregate or do they go up or down?
55:50>> to say like this.
55:50>> Right, do they go up or down?
55:52And you know, I
55:55I think everything we're talking about
55:56would sort of argue down, right? Uh
55:58there's many businesses where you know,
56:00you know, they're making margin because
56:01>> More consumer surplus.
56:02>> More consumer surplus.
56:03>> Everywhere.
56:03>> What's interesting is if you go back and
56:06read a lot of the research reports from
56:07the '90s when the internet was coming
56:09along. They said they said the same
56:10thing.
56:11Um so the base rate here might not be
56:12great. Like maybe we're wrong, but I
56:14think the main reason they were often
56:15wrong was like was this aggregation
56:17economics of Van Thompson aggregation
56:19economics, right? So maybe absent that
56:21maybe it wouldn't have been so wrong.
56:23And I do think that like a lot of
56:24aggregation theory breaks down when you
56:27know, cuz that basically relies on the
56:28idea that consumers only look at the
56:29first thing on Google, right? But it
56:30turns out if your agent looks at
56:32everything on on search then being
56:34ranked first doesn't matter and
56:35um you know, all the sort of core things
56:37like quality and and price and
56:39everything else um service become more
56:41salient.
56:43Um but you know, the other thing is that
56:44maybe like
56:46another sort of downstream effect of
56:48this is that it'll
56:49you'll be able to build more complicated
56:50businesses, right? So so like right now
56:52a lot of industries
56:54are more horizontal in structure because
56:57the complexity of running your slice is
56:58very high
57:00um and and that you know,
57:02makes it hard to to do what's above and
57:03below you and not have undue complexity.
57:05With um AI agents and and basically
57:08cheap and ubiquitous code and
57:09intelligence like you could build the
57:11whole vertical stack. People trying to
57:12do this I think in
57:14other countries but like you can have
57:16a full stack pet care business that does
57:18everything from make pet food
57:20all the way up through pet insurance and
57:22pet hospitals and pet vets and pet and
57:24all these things that historically were
57:26multiple layers but now because you can
57:28solve the coordination and reduce the
57:30principal agent problems
57:31by having um
57:33by having AI do it like maybe you those
57:36sorts of businesses are possible and
57:37then maybe those businesses are are
57:39higher margin in in the end and more
57:41instead of more efficient and more
57:42consumer surplus. You know, Carvana is
57:44sort of an example of that built before
57:47this was you know easy to do but maybe
57:49people can do it in other verticals in
57:51[snorts] which case maybe margins go up
57:53and there's consumer surplus. Um
57:55The other one is the fun one of things
57:56here is just how this affects like
57:58wealth and income distribution, right? I
58:00can tell tell two stories you know on
58:02the one hand
58:04I think a lot of people have talked
58:05about this idea that it concentrates
58:06income distribution, right? You've got
58:07all these people who made money on like
58:09various labs and and like whatever data
58:11Nvidia and um
58:13this idea that there are like increasing
58:14returns to small groups of people. Um
58:17but the counter argument is if we have
58:19cheap and ubiquitous intelligence
58:22um a couple of things happen. One is
58:23there's all these rent takers in the
58:24economy. Think you're the software
58:26company that charges bowling alleys like
58:29you know whatever, right? And those guys
58:31all get smoked which sucks for them but
58:33it does lower the cost of bowling. And
58:35so then you know that increases living
58:37standards for everyone who bowls, right?
58:39Um and that's flattening because the my
58:41my spending on bowling is small
58:43proportion to my income but like people
58:44who make less money spend more
58:45proportion to their income on on
58:46bowling. And so the it's it's a
58:48flattening effect on on income. If like
58:50all these like little rent pieces like
58:53go away
58:54the decline in corporate profits like
58:56all the money that the average American
58:57is giving to Bank of America goes from
58:59away from Bank of America the
59:00capitalists lose you know like the the
59:03average consumer wins, right? So that's
59:06leveling.
59:07The other piece is like there's a whole
59:08professional class, right? Who basically
59:11you could define their role as
59:14>> Intelligence.
59:15>> Intelligence, right? They take big
59:16corpuses of information and apply them
59:18to certain circumstances in kind of a
59:19fairly
59:20like, you know, straightforward way. Um,
59:23and and if those all get commoditized,
59:27one, everyone gets the best ubiquitous,
59:30which is one good for everybody, but
59:32two, um, particularly good for people
59:33who can't afford the best right now or
59:35any.
59:36Um, but then three, those individuals,
59:39they earned these extra rents because
59:40they were smarter. They then have to
59:43drop to their next best alternative.
59:46Uh, and in theory then that might
59:47flatten
59:49the curve. Now, the counterargument to
59:51that might be something like, well, yes,
59:53but returns to
59:55initiative and ambition and intelligence
59:57could go up because now I can do so much
1:00:00more because these models are like a um,
1:00:04you know, like a bicycle for the mind on
1:00:07steroids, right? Like they they give
1:00:09people who are smart and and have
1:00:11initiative like the ability to
1:00:13accomplish, so much more. Maybe that
1:00:16actually causes like increasing returns
1:00:18to that subset of people and it steepens
1:00:22um, the curve. Truman asked for a
1:00:24one-handed economist. I'm, you know, I
1:00:26see I seem to very distinctly have two
1:00:28hands on everything.
1:00:30Um, but uh,
1:00:31>> [laughter]
1:00:32>> you know, I I I think that's a big
1:00:33interesting question.
1:00:34>> They're super interesting questions and
1:00:35and like I I've found that a lot of
1:00:36situations I look at like the there's a
1:00:38lot of cross currents. Like on the one
1:00:39hand everything we started the
1:00:40conversation with the AI buildout. On
1:00:41the one hand, I can see all these
1:00:43incredible ways to use AI and I use a
1:00:46good gajillion tokens and I can see how
1:00:48I could use a gajillion more and
1:00:49obviously I am in the far right end of
1:00:51the distribution. If everyone uses as
1:00:53many tokens as me, we are so short. Like
1:00:55it's going to be this huge thing that's
1:00:56going on forever and everyone and and
1:00:57everyone who's in the buildout is going
1:00:58to make a fortune, right? On the other
1:01:00hand, right? The the amount I get per
1:01:03token is like going to the moon and it
1:01:05turns out I'm a pretty unusual guy in a
1:01:08lot of senses and people are going to
1:01:09adapt more slowly and the cost per token
1:01:10is going to collapse. And so maybe
1:01:12actually, uh, like we're totally
1:01:15overbuilt, right? And like I can tell I
1:01:16can tell both stories like with equal
1:01:19certainty.
1:01:19>> I have
1:01:20a good friend who has been running this
1:01:23very cool experiment where if you break
1:01:25investing down into picking what to own
1:01:27and then how much to own of each thing
1:01:29that you've picked
1:01:31overly simplified but kind of
1:01:32interesting interesting way of thinking
1:01:33about it that he's been running this
1:01:35experiment where he continues to pick
1:01:38the things to own
1:01:40but then runs his own portfolio
1:01:42construction as he would have in the
1:01:43past in in one lane and in the second
1:01:45lane is just completely handing that off
1:01:47to an AI harness thing that he's built
1:01:50and says, "You do all the portfolio
1:01:51construction but stick with the stuff
1:01:53that that I've picked."
1:01:54And what he's found so far is that that
1:01:57is drastically outperforming his version
1:02:00of portfolio construction which itself
1:02:01is quite sophisticated. And the simple
1:02:03argument that he would make is well, AIs
1:02:05are just really good at optimization and
1:02:07they've read every paper on portfolio
1:02:09construction and and all this kind of
1:02:10stuff and bet sizing and Kelly criterion
1:02:14and like they they know it all and that
1:02:16in the future perhaps we make that
1:02:18transition where the investor's job is
1:02:21just picking what to own and not how
1:02:23much of
1:02:24of each to own. And I ask you this
1:02:26question what you think about this idea
1:02:28and the future because you have such an
1:02:29unusual portfolio where again 80% of it
1:02:32is in one specific asset and I guess is
1:02:35if you were in lane two your portfolio
1:02:37would
1:02:38the AI would weight your portfolio very
1:02:40differently than you have chosen to do
1:02:42it.
1:02:42>> I'm underweight.
1:02:44>> [laughter]
1:02:48>> So I'm just I'm dying to know what you
1:02:50think about
1:02:52about this idea.
1:02:53>> You know,
1:02:54>> underweight Carvana.
1:02:55Should be 90.
1:02:56>> Time time scales probably are
1:02:58interesting on that to think about too.
1:03:00Um,
1:03:01you know, as I think about how AI is
1:03:02going to encroach into this business I
1:03:04think it's going to start at shorter
1:03:05time scales
1:03:06and work its way up to longer time
1:03:08scales.
1:03:09And so be be you know, cuz one of the
1:03:11challenges is that like the
1:03:14there's no verification loop.
1:03:16Like your your friend's ability to say
1:03:17that
1:03:19is based on results over I mean,
1:03:20certainly not 20-year results.
1:03:22I strongly suspect that like
1:03:25the march of AI into markets and I mean,
1:03:28they're started, right? Like, I don't
1:03:29know how Renaissance Technologies did
1:03:30it, but like I assume
1:03:32>> Seems like they might have discovered a
1:03:33transformer before everybody [laughter]
1:03:34else. It's like I'm not kidding.
1:03:37>> It wouldn't shock me if people who
1:03:39invest over shorter time scales starting
1:03:41with with seconds and working their way
1:03:43out to minutes, hours, weeks, months
1:03:45find themselves handing more and more
1:03:48delegating more and more to models.
1:03:50Makes perfect sense in that you could
1:03:51you could definitely imagine bifurcating
1:03:52that between
1:03:54trying to enrich
1:03:55you know, the model. I think other
1:03:57people do things like this already and
1:03:58this is just a more sophisticated
1:03:59version, but you can
1:04:01you can have a a fully sort of
1:04:02quantitative strategy and then you can
1:04:04enrich it with what you think is is
1:04:07qualitative kind of human-designed
1:04:08analysis and then you can have a
1:04:11quantitative strategy to to incorporate
1:04:13the qualitative. I think many of these
1:04:15multi-strat strats like do these sorts
1:04:16of things. And what he's doing is kind
1:04:18of a just a slightly version of
1:04:19different flavor of that. It makes makes
1:04:20perfect sense. I'm not There are so many
1:04:22ways, you know, to be successful and and
1:04:24models have just added more dimensions
1:04:26of freedom.
1:04:27>> But what about your own portfolio? Say a
1:04:28word more about how you feel comfortable
1:04:31having a position. Forget the company
1:04:32that it is
1:04:33>> Sure.
1:04:33>> that's that big in the face of like so
1:04:36much
1:04:38research that would suggest that's a
1:04:40crazy way to
1:04:41build a portfolio.
1:04:42>> A lot of that research I think is is
1:04:45more looking at
1:04:48price changes in terms of like
1:04:49securities prices
1:04:50>> Mhm.
1:04:51>> and less focused on
1:04:54like thinking about it the way someone
1:04:56who
1:04:58you know, owns a company might think
1:04:59about it. I mean, we're talking about
1:05:00Carvana. Were it not for my sense that
1:05:02like Carvana's
1:05:04infrastructure and brand and everything
1:05:06else is is just so differentiated that
1:05:08there are
1:05:09few scenarios where it works out really
1:05:10badly.
1:05:12Um
1:05:14you know, then I wouldn't feel so
1:05:15comfortable owning so much of it. It
1:05:16might just be as simple as that. Like
1:05:18like over thinking over
1:05:20you know, that that portfolio approach
1:05:23would have worked absolutely, but it's
1:05:24certainly if you discovered Walmart or
1:05:26Home Depot or Fastenal or something back
1:05:28in the day, you wouldn't have
1:05:30taken full advantage of it.
1:05:32Um
1:05:33and that's kind of my sense, you know,
1:05:35with with Carvana. If you read history,
1:05:37of course, all the people who did really
1:05:38great
1:05:39tended to be very concentrated. But of
1:05:41course, the history is lit that you
1:05:42don't read the history of
1:05:44Yeah, you don't read the history of you
1:05:45didn't. I also think that a lot of that
1:05:47literature is is sort of what most
1:05:48people who are looking for average or
1:05:50slightly better than average average
1:05:51results should do and it's totally the
1:05:53right thing to do.
1:05:54Um but that's not what I'm looking for.
1:05:57I like Carvana a lot. It wouldn't upset
1:06:00me any if
1:06:01the stock went up a bunch and
1:06:03I felt compelled to diversify. That
1:06:05would also be okay. It's just more of a
1:06:08a trade-off um between expected returns
1:06:11and
1:06:12sort of knowingly taking a
1:06:14um
1:06:15an unusual risk. What do you think about
1:06:17education? You and I have kids of
1:06:18roughly similar ages. Comment a little
1:06:20bit on how you're thinking about their
1:06:22educations, the institutions that have
1:06:24driven education in this country that
1:06:25have been really important in that whole
1:06:27path-dependent story you've described.
1:06:29What do you think the present future of
1:06:31education looks like or should look
1:06:32like?
1:06:33I've been
1:06:34like talking to the head of my kids'
1:06:36school about this and it more and more
1:06:38it seems to be that his thinking and you
1:06:41know, and and my thinking are
1:06:44heading in this direction of like the
1:06:46way you're educating kids is getting
1:06:47more bimodal in terms of what you're
1:06:49looking for. Right. Let me step way
1:06:51back. What is the benefit of
1:06:53intelligence?
1:06:54>> [snorts]
1:06:54>> Personal intelligence, like your own
1:06:55intelligence, in a world where there's
1:06:57infinite intelligence from models.
1:06:59And
1:07:01you might say, "Oh, it's useless. Like
1:07:02doesn't matter. Like someone with a 50
1:07:03IQ can navigate the world just as well
1:07:05as someone with a 150 IQ." And I don't
1:07:06think that's quite right.
1:07:08And intuitively, it doesn't feel right.
1:07:10But I think
1:07:11just to nail down where it is is that
1:07:13the more intelligence you have and and
1:07:15more crystallized intelligence you have,
1:07:16which would be sort of education,
1:07:19the better you are at communicating with
1:07:22a model and telling a model what you
1:07:24want. So, like I
1:07:27would be have a very tough time like
1:07:28having a deep conversation with the
1:07:30model to like create leading like
1:07:32quantum physics experiments because I
1:07:34don't know anything about quantum
1:07:35physics.
1:07:36And the model could try very hard to
1:07:38explain things to me, but in the end it
1:07:39would be explaining things in crayon and
1:07:40Nerf. And like it just like I we just
1:07:44like it would just be very hard to get
1:07:45to leading edge and like have a and get
1:07:47like the best out of what the model has
1:07:48to offer.
1:07:50Um whereas in an area where I do have
1:07:51some expertise, um the model and I can
1:07:54communicate kind of effectively and I
1:07:56can direct the model to my to my will
1:07:58like much better.
1:08:00And so, there's like a super important
1:08:03like you to direct a model you you to do
1:08:06your will, which I think is going to be
1:08:08an important skill in the future, you do
1:08:10have to understand something um about
1:08:12the world. You have to be you know and
1:08:13and smarts help for that. And education
1:08:16helps for that. Like you can't ask it to
1:08:19you know, assume this is a stochastic
1:08:21process unless you know what a
1:08:22stochastic process is. Now, you you
1:08:24certainly it could teach you what one
1:08:26is, but that's a lot of friction instead
1:08:28of this figuring this all out. You don't
1:08:30have to work out the math, but you have
1:08:31to know what it is, right? And so, when
1:08:34I've been thinking about education,
1:08:34there's like some this idea that like
1:08:36you're teaching your kids a bunch of
1:08:39foundational skills
1:08:41that build on one another
1:08:44where I think in the end the sum total
1:08:45of those they have to be able to do them
1:08:47kind of without models. Not because
1:08:48they're going to do them without models
1:08:49in the future, but because if they don't
1:08:51know how to do multiplication, it's
1:08:53going to be really hard to understand
1:08:54matrix multiplication. And if they don't
1:08:55know what matrix multiplication is at a
1:08:57at a high enough level, they don't have
1:08:58to be able to do it really well, but
1:08:59they have to be able to understand what
1:09:01the model is doing when it's doing it.
1:09:02And that's so they can direct the model.
1:09:04And so, increasingly I'm of the view
1:09:05that like we should spend a bunch of
1:09:06time with kids basically separated from
1:09:08these models teaching them basic skills
1:09:10that we think are important and what
1:09:12those are is a really hard thing, but
1:09:13the idea being that yes, model can do
1:09:15it, but you need to know it so that you
1:09:16can tell the model what to do.
1:09:19And then on the flip side, some English
1:09:21like I don't like models because like
1:09:23the the just cheat on their Shakespeare
1:09:24essays. I'm like you're not being
1:09:26ambitious enough. Right? If the kids are
1:09:28cheating on their Shakespeare essay,
1:09:29>> Give them something harder to do.
1:09:30>> Give them something 100 times harder and
1:09:32more ambitious and have them use models
1:09:33to their fullest extent. Right? Have
1:09:36them bring the work product back and
1:09:37then make sure that they didn't just
1:09:38give a prompt to the model do the
1:09:39homework. Make sure they own it. Ask
1:09:40them why it's like this, right? Force
1:09:42them to like justify it, right? And
1:09:45I think that's another really important
1:09:47thing cuz kids need to be good at using
1:09:48them and they also have frankly just
1:09:49need to be taught they can do like
1:09:52anything.
1:09:52>> You and I get to talk every so often on
1:09:53the phone every time I wish it was
1:09:55recorded so that other people could hear
1:09:57it cuz I learn so much and
1:09:59I think in this particular conversation
1:10:00today, which is sort of a version of a
1:10:02call we had several months ago, it's
1:10:04just such a valuable set of food for
1:10:06thought about the future of intelligence
1:10:08and what that means for markets and
1:10:10companies and uh is a little bit not in
1:10:13opposition to but a different
1:10:15perspective than the most common one
1:10:17that you're hearing, which is sort of
1:10:19unlimited, you know, unlimited build out
1:10:21forever, short everything forever, um no
1:10:24cap on the demand for intelligence
1:10:26>> [music]
1:10:27>> and uh I've loved talking to you about
1:10:28it both on the phone and today, so thank
1:10:30you so much for your time.
1:10:31>> Thank you.
1:10:35You know how small advantages compound
1:10:37over time? [music] That's true in
1:10:38investing and just as true in how you
1:10:40run your company. Your spending system
1:10:42is your capital allocation strategy.
1:10:44Ramp makes [music] it smarter by
1:10:45default. Better data, better decisions,
1:10:47better economics over time. See how at
1:10:49ramp.com/invest.
1:10:50As your business grows, Vanta scales
1:10:52with you, automating compliance and
1:10:54giving you a single source [music] of
1:10:55truth for security and risk. Learn more
1:10:57at vanta.com/invest.
1:10:59The best AI and software companies from
1:11:01OpenAI to Cursor to Perplexity [music]
1:11:03use WorkOS to become enterprise ready
1:11:05overnight, not in months. Visit
1:11:07workos.com [music] to skip the
1:11:08unglamorous infrastructure work and
1:11:10focus on your product. Ridgeline is
1:11:12redefining asset management technology
1:11:14as a true partner, not [music] just a
1:11:15software vendor. They've helped firms 5x
1:11:18in scale, enabling faster growth,
1:11:19smarter operations, [music] and a
1:11:20competitive edge. Visit
1:11:22ridgelineapps.com
1:11:23to see what they can unlock for your
1:11:25firm.

About this transcript

This page contains the full transcript of What's Still Valuable When Intelligence Is Free by Invest Like The Best, generated from the public captions YouTube serves with the video. The transcript has 15,071 words across 2,370 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.