The State of AI with Marc & Ben

0:00:00 Most of the content created on the Internet is created by average people, and so kind
0:00:03 of the content on average, you know, as a whole on average is average.
0:00:09 The test for whether your idea is good is how much can you charge for it?
0:00:12 Can you charge the value?
0:00:14 Or are you just charging the amount of work it’s going to take the customer to put their
0:00:18 own wrapper on top of OpenAI?
0:00:22 The paradox here would be the cost of developing any given piece of software falls, but the
0:00:27 reaction to that is a massive surge of demand for software capabilities.
0:00:31 And I think this is one of the things that’s always been underestimated about humans is
0:00:37 our ability to come up with new things we need.
0:00:41 There’s no large marketplace for data.
0:00:43 In fact, what there are is there are very small markets for data.
0:00:47 In this wave of AI, big tech has a big compute and data advantage.
0:00:52 But is that advantage big enough to drown out all the other startups trying to rise up?
0:00:57 Well, in this episode, A16Z co-founders Mark Andreessen and Ben Horowitz, who both, by the
0:01:02 way, had a front row seat to several prior tech waves, tackled the state of AI.
0:01:07 So what are the characteristics that’ll define successful AI companies?
0:01:12 And is proprietary data the new oil, or how much is it really worth?
0:01:17 How good are these models realistically going to get?
0:01:19 And what would it take to get 100 times better?
0:01:23 Mark and Ben discuss all this and more, including whether the venture capital model needs a
0:01:27 refresh to match the rate of change happening all around it.
0:01:31 And of course, if you want to hear more from Ben and Mark, make sure to subscribe to the
0:01:35 Ben and Mark podcast.
0:01:37 All right, let’s get started.
0:01:41 It is kind of the darkest side of capitalism when a company is so greedy, they’re willing
0:01:46 to destroy the country and maybe the world to, like, just get a little extra profit.
0:01:50 When they do it, like, the really kind of nasty thing is they claim, oh, it’s for safety.
0:01:55 You know, we’ve created an alien that we can’t control, but we’re not going to stop
0:01:59 working on it.
0:02:00 We’re going to keep building it as fast as we can, and we’re going to buy every freaking
0:02:03 GPU on the planet.
0:02:04 But we need the government to come in and stop it from being open.
0:02:08 This is literally the current position of Google and Microsoft right now.
0:02:13 It’s crazy.
0:02:16 The content here is for informational purposes only, should not be taken as legal, business,
0:02:20 tax or investment advice or be used to evaluate any investment or security and is not directed
0:02:26 at any investor or potential investors in any A16Z fund.
0:02:30 Please note that A16Z and its affiliates may maintain investments in the companies discussed
0:02:35 in this podcast.
0:02:36 For more details, including a link to our investments, please see A16Z.com/disclosures.
0:02:41 Hey, folks, welcome back.
0:02:44 We have an exciting show today.
0:02:45 We are going to be discussing the very hot topic of AI.
0:02:49 We are going to focus on the state of AI as it exists right now in April of 2024, and
0:02:53 we are focusing specifically on the intersection of AI and company building.
0:02:57 Hopefully, this will be relevant to anybody working on a startup or anybody at a larger
0:03:01 company.
0:03:02 We have as usual solicited questions on X, formerly known as Twitter, and the questions
0:03:05 have been fantastic.
0:03:06 We have a full lineup of listener questions, and we will dive right in.
0:03:11 First question, so three questions on the same topic.
0:03:14 Michael asks, “In anticipation of upcoming AI capabilities, what should founders be
0:03:18 focusing on building right now?”
0:03:20 Gwen asks, “How can small AI startups compete with established players with massive compute
0:03:25 and data scale advantages?”
0:03:27 Alistair McLea asks, “For startups building on top of open AI, etc., what are the key
0:03:32 characteristics of those companies that will benefit from future exponential improvements
0:03:36 in the base models versus those that will get killed by them?”
0:03:39 Let me start with one point, Ben, and then we’ll jump right to you.
0:03:42 Sam Maltman recently gave an interview, I think maybe Lex Friedman or one of the podcasts,
0:03:45 and he actually said something I thought was actually quite helpful.
0:03:48 Let’s see, Ben, if you agree with that.
0:03:49 He said something along the lines of, “You want to assume that the big foundation models
0:03:54 coming out of the big AI companies are going to get a lot better, so you want to assume
0:03:57 they’re going to get like a hundred times better.
0:04:00 As a startup founder, you want to then think, “Okay, if the current foundation models get
0:04:03 a hundred times better, is my reaction, oh, that’s great for me and for my startup, because
0:04:08 I’m much better off as a result, or is your reaction the opposite as it, oh, shit, I’m
0:04:12 in real trouble.”
0:04:13 Let me just stop right there, Ben, and see what you think of that as general advice.
0:04:16 Well, I think generally that’s right, but there’s some nuances to it, right?
0:04:22 So I think that from Sam’s perspective, he was probably discouraging people from building
0:04:28 foundation models, which I don’t know that I would entirely agree with that, and that
0:04:34 a lot of the startups building foundation models are doing very well, and there’s many
0:04:38 reasons for that.
0:04:39 One is there are architectural differences, which lead to how smart is the model, there’s
0:04:43 how fast is the model, there’s how good is the model in a domain.
0:04:47 When that goes for not just text models, but image models as well, there are different
0:04:53 domains, different kinds of images that responds to prompts differently.
0:04:58 If you ask Mid Journey and Ideagram the same question, they react very differently depending
0:05:03 on the use cases that they’re tuned for.
0:05:07 And then there’s this whole field of distillation where Sam can go build the biggest, smartest
0:05:13 model in the world, and then you can walk up as a startup and kind of do a distilled
0:05:18 version of it and get a model very, very smart at a lot less cost.
0:05:22 So there are things that, yes, the big company models are going to get way better, kind of
0:05:29 way better at what they are.
0:05:31 So you need to deal with that.
0:05:34 So if you’re trying to go head to head full frontal assault, you probably have a real
0:05:37 problem just because they have so much money.
0:05:41 But if you’re doing something that’s different enough or a different domain and so forth,
0:05:49 for example, at Databricks, they’ve got a foundation model, but they’re using it in
0:05:54 a very specific way in conjunction with their kind of leading data platform.
0:06:00 So, okay, now if you’re an enterprise and you need a model that knows all the nuances
0:06:07 of how your enterprise data model works and what things mean and needs access control
0:06:14 and what needs to use your specific data and domain knowledge and so forth, then it doesn’t
0:06:19 really hurt them if Sam’s model gets way better.
0:06:21 Similarly, 11 Labs with their voice model has kind of embedded into everybody.
0:06:28 Everybody uses it as part of kind of the AI stack.
0:06:31 And so it’s got kind of a developer hook into it.
0:06:35 And then they’re going very, very fast at what they do and really being very focused
0:06:40 in their area.
0:06:41 So there are things that I would say like extremely promising that are kind of ostensibly, but
0:06:47 not really competing with OpenAI or Google or Microsoft.
0:06:52 So I think it sounds a little more coarse green than I would interpret it if I was building
0:06:57 a startup.
0:06:58 Right.
0:06:59 Let’s dig into this a little bit more.
0:07:00 So let’s start with the question of do we think the big models, the god models are going
0:07:03 to get 100 times better?
0:07:04 I kind of think so and then I’m not sure.
0:07:07 So if you think about the language models, let’s do those because those are probably
0:07:11 the ones that people are most familiar with.
0:07:13 I think if you look at the very top models, you know, Claude and OpenAI and Mistral and
0:07:19 Llama, the only people who I feel like really can tell the difference as users amongst those
0:07:26 models are the people who study them, you know, like they’re getting pretty close.
0:07:31 So you would expect if we were talking 100x better that one of them might be separating
0:07:36 from each other a lot more, but the improvement, so 100% better in what way?
0:07:42 Like for the normal person using it in a normal way, like asking it questions and finding
0:07:47 out stuff.
0:07:48 Well, let’s say some combination of just like breadth of knowledge and capability.
0:07:52 Yeah.
0:07:53 Like I think in some of them may are yeah.
0:07:55 Right.
0:07:56 But then also just combined with like sophistication of the answers, you know, sophistication of
0:07:59 the output, the quality of the output, sophistication of the output, you know, lack of hallucination,
0:08:03 factual grounding.
0:08:04 Well, that I think is for sure going to get 100x better.
0:08:08 Like that.
0:08:09 Yeah.
0:08:10 I mean, they’re on a path for that.
0:08:11 The things that are so against that, right, the alignment problem where, okay, yeah, they’re
0:08:18 getting smarter, but they’re not allowed to say what they know.
0:08:21 And then that alignment also kind of makes them dumber in other ways.
0:08:25 And so you do have that thing.
0:08:27 The other kind of question that’s come up lately, which is kind of do we need a breakthrough
0:08:32 to go from what we have now, which I would categorize as artificial human intelligence
0:08:40 as opposed to artificial general intelligence, meaning it’s kind of the artificial version
0:08:45 of us.
0:08:46 We’ve structured the world in a certain way using our language and our ideas and our
0:08:50 stuff.
0:08:52 And it’s learned that very well, amazing.
0:08:55 And it can do kind of a lot of the stuff that we can do, but are we then the asymptote?
0:09:02 Or you need a breakthrough to get to some kind of higher intelligence, more general intelligence.
0:09:08 And I think if we’re the asymptote, then in some ways it won’t get 100x better because
0:09:15 it’s already like pretty good relative to us.
0:09:18 But yeah, like it’ll know more things, it’ll hallucinate less on all those dimensions,
0:09:22 it’ll be 100x better.
0:09:24 There’s this graph floating around.
0:09:26 I forget exactly what the axes are, but it basically shows the improvement across the
0:09:29 different models.
0:09:30 To your point, it shows an asymptote against the current tests that people are using that’s
0:09:33 sort of like at or slightly above human levels, which is what you would think if you’re being
0:09:37 trained on an entirely human data.
0:09:39 Now, the counter argument on that is are the tests just too simple, right?
0:09:42 It’s a little bit like the question people ever run the SAT, which is if you have a lot
0:09:45 of people giving 800s on both math and verbal on the SAT, is the scale too constrained, do
0:09:49 you need a test that can actually test for Einstein?
0:09:52 Right.
0:09:53 It’s memorized the tests that we have and it’s great.
0:09:57 You can imagine SAT that really can detect gradations of people who have ultra-high IQs,
0:10:02 who are ultra-good at math or something.
0:10:03 You can imagine tests for AI, you can imagine tests that test for reasoning above human
0:10:07 levels, one assumes.
0:10:08 Yeah, well, maybe the AI needs to write the test.
0:10:11 Yeah, and then there’s a related question that comes up a lot, it’s an argument we’ve
0:10:15 been having internally, which is also I’ll start to take some sort of more provocative
0:10:18 and probably more bullish or as you would put it, sort of science fictionary predictions
0:10:21 on some of this stuff.
0:10:22 There’s this question that comes up, which is, okay, you take an LLM, you train it on
0:10:25 the internet.
0:10:26 What is the internet data?
0:10:27 What is the internet data corpus?
0:10:28 It’s an average of everything, right?
0:10:29 It’s a representation of sort of human activity.
0:10:32 Representation of human activity is going to kind of, because of the sort of distribution
0:10:34 of intelligence in the population, most of it somewhere in the middle.
0:10:37 And so the data set on average sort of represents the average human.
0:10:40 You’re teaching it to be very average, yeah.
0:10:42 Yeah, you’re teaching it to be very average.
0:10:43 It’s just because most of the content created on the internet is created by average people.
0:10:46 And so kind of the content on average as a whole on average is average.
0:10:51 And so therefore, the answer is our average, right?
0:10:53 You’re going to get back an answer that sort of represents the kind of thing that an average
0:10:56 100 IQ, you know, kind of by definition, the average human is 100 IQ, it’s IQ is indexed
0:10:59 to 100.
0:11:00 That’s the center of the bell curve.
0:11:01 And so by definition, you’re kind of getting back the average.
0:11:03 I actually argue like that may be the case for the default prompt today.
0:11:06 Like you just asked the thing, does the earth revolve around the sun or something?
0:11:09 You get like the average answer to that and maybe that’s fine.
0:11:12 This gets to the point as well.
0:11:13 Okay, the average data might be of an average person, but the data set also contains all
0:11:17 of the things written and thought by all the really smart people.
0:11:20 All that stuff is in there, right?
0:11:21 And all the current people who are like that, their stuff is in there.
0:11:24 And so then it’s sort of like a prompting question, which is like, how do you prompt
0:11:26 it in order to get basically, in order to basically navigate to a different part of
0:11:29 what they call the latent space, to navigate to a different part of the data set that basically
0:11:33 is like the super genius part.
0:11:35 And you know, the way these things work is if you craft the prompt in a different way,
0:11:37 it actually leads it down a different path inside the data set, gives you a different
0:11:40 kind of answer.
0:11:41 And here’s another example of this.
0:11:42 If you ask it write code to do X, write code to sort of list, you know, whatever, render
0:11:46 an image, it will give you average code to do that.
0:11:48 If you say write me secure code to do that, it will actually write better code with fewer
0:11:53 security holes, which is very interesting, right?
0:11:55 Because it’s accessing a different purpose of training data, which is secure code.
0:11:58 Right.
0:11:59 And if you ask, you know, write this image generation thing the way John Carmack would
0:12:01 write it, you get a much better result because it’s tapping into the part of the latent space
0:12:04 represented by John Carmack’s code, who’s the best graphics programmer in the world.
0:12:08 And so you can imagine prompting crafts in many different domains such that you’re kind
0:12:11 of unlocking the latent super genius, even if that’s not the default answer.
0:12:15 Yeah.
0:12:16 I think that’s correct.
0:12:18 I think there’s still a potential limit to its smartness in that.
0:12:24 So we had this conversation in the firm the other day where you have, there’s the world,
0:12:28 which is very complex.
0:12:30 And intelligence kind of is, you know, how well can you understand, describe, represent
0:12:35 the world?
0:12:36 But our current iteration of artificial intelligence consists of human structuring the world and
0:12:45 then feeding that structure that we’ve come up with into the AI.
0:12:50 And so the AI kind of is good at predicting how humans have structured the world as opposed
0:12:56 to how the world actually is, which is, you know, something more probably complicated,
0:13:01 maybe the irreducible or what have you.
0:13:05 So do we just get to a limit where like, it can be really smart, but its limit is going
0:13:10 to be the smartest humans as opposed to smarter than the smartest humans.
0:13:14 And then kind of related, is it going to be able to figure out brand new things, you know,
0:13:21 new laws of physics and so forth?
0:13:22 Now, of course, there are like one in three billion humans that can do that or whatever.
0:13:28 That’s a very rare kind of intelligence.
0:13:30 So it still makes the AI is extremely useful, but they play a different role if they’re
0:13:37 kind of artificial humans than if they’re like artificial, you know, super-duper mega-humans.
0:13:45 So let me make the sort of extreme bull case for the 100, because okay, so the cynic would
0:13:50 say that Sam Altman would be saying they’re going to get 100 times better precisely if
0:13:53 they’re not going to.
0:13:55 Yeah, yeah, yeah, yeah, yeah.
0:13:57 Right?
0:13:58 Because he’d be saying that basically in order to scare people into not competing.
0:14:01 Well, I think that whether or not they are going to get 100 times better, Sam would
0:14:06 be very likely to say that like Sam.
0:14:08 For those of you who don’t know him, he’s a very smart guy, but for sure, he’s a competitive
0:14:13 genius.
0:14:14 There’s no question about that.
0:14:15 So you have to take that into account.
0:14:17 Right.
0:14:18 So if they weren’t going to get a lot better, he would say that.
0:14:20 But of course, if they were going to get a lot better to your point, he would also say
0:14:22 that.
0:14:23 Yes.
0:14:24 Why not, right?
0:14:25 And so let me make the bull case that they are going to get 100 times better or maybe
0:14:28 even, you know, on an upper curve for a long time.
0:14:31 And there’s like enormous controversy, I think, on every one of the things I’m about to say,
0:14:34 that you can find very smart people in the space who believe basically everything I’m
0:14:38 about to say.
0:14:39 So one is there is generalized learning happening inside the neural networks.
0:14:42 And we know that because we now have introspection techniques where you can actually go inside
0:14:46 and look inside the neural networks to look at the neural circuitry that is being evolved
0:14:49 as part of the training process.
0:14:50 And you know, these things are evolving, you know, general computation functions.
0:14:54 There was a case recently where somebody trained one of these on a chess database and, you
0:14:57 know, just by training lots of chess games, it actually imputed a world model of a chess
0:15:00 board, you know, inside the neural network and, you know, that was able to do original
0:15:04 moves.
0:15:05 And so the neural network training process does seem to work.
0:15:06 And then specifically not only that, but, you know, Meta and others recently have been
0:15:10 talking about how so-called overtraining actually works, which is basically continuing to train
0:15:15 the same model against the same data for longer, you know, putting more and more compute cycles
0:15:18 against it.
0:15:19 You know, I’ve talked to some very smart people in the fields, including there, who basically
0:15:22 think that actually that works quite well.
0:15:24 The diminishing returns people were worried about about more training.
0:15:26 And they proved it in the new Lamar release, right?
0:15:29 That’s a primary technique they use.
0:15:31 Yeah, exactly.
0:15:32 Like one guy in the space basically told me, basically, he’s like, yeah, we don’t necessarily
0:15:35 need more data at this point to make these things better.
0:15:37 We maybe just need more compute cycles.
0:15:39 We just train it a hundred times more and it may just get actually a lot better.
0:15:41 So one day the labeling, it turns out that supervised learning ends up being a huge boost
0:15:47 to these things.
0:15:48 Yeah.
0:15:49 So we’ve got that.
0:15:50 We’ve got all of the kind of, you know, let’s say rumors and reports of various kinds of self-improvement
0:15:54 loops, you know, that kind of underway.
0:15:56 And most of the sort of super advanced practitioners in the field think that there’s now some form
0:15:59 of self-improvement loop that works, which basically is, you basically get an AI to do
0:16:03 what’s called chain of thoughts.
0:16:04 You get it to basically go step by step to solve a problem.
0:16:06 You get it to the point where it knows how to do that.
0:16:08 And then you basically retrain the AI on the answers.
0:16:10 And so you’re kind of basically doing a sort of a forklift upgrade across cycles of the
0:16:14 reasoning capability.
0:16:15 And so a lot of the experts think that sort of thing is starting to work now.
0:16:18 And then there’s still a raging debate about synthetic data, but there’s quite a few people
0:16:21 who are actually quite bullish on that.
0:16:23 Yeah.
0:16:24 And then there’s even this trade-off.
0:16:25 There’s this kind of dynamic where like LLMs might be okay at writing code, but they might
0:16:29 be really good at validating code.
0:16:30 You know, they might actually be better at validating code than they are at writing it.
0:16:33 That would be big help.
0:16:34 Yeah.
0:16:35 Well, but that also means like AI is maybe able to self-improve.
0:16:36 They can validate their own code.
0:16:37 Yeah.
0:16:38 Yeah.
0:16:39 They can validate their own code.
0:16:40 And it’s, and we have this anthropomorphic bias is very deceptive with these things because
0:16:43 you think of the model as an it, and so it’s like, how could you have an it that’s better
0:16:46 at validating the code, the writing code, but it’s not an it.
0:16:48 What it is is it’s this giant latent space, it’s this giant neural network.
0:16:51 And the theory would be there are totally different parts of the neural network for
0:16:54 writing code and validating code.
0:16:56 And there’s no consistency requirement whatsoever that the network be equally good at both of
0:16:59 those things.
0:17:00 And so if it’s better at one of those things, right, so then the thing that it’s good at
0:17:04 might be able to make the thing that it’s bad at better and better.
0:17:06 Right.
0:17:07 Right.
0:17:08 Right.
0:17:09 Right.
0:17:10 Right.
0:17:11 Sure.
0:17:12 Sure.
0:17:13 Right.
0:17:14 Sort of a self-improvement thing.
0:17:15 And so then on top of that, there’s all the other things coming, right?
0:17:16 Which is it’s everything is all these practical things, which is there’s an enormous chip constraint
0:17:17 right now.
0:17:18 And the AI that anybody uses today is its capabilities are basically being gated by the availability
0:17:22 of chips, but like that will resolve over time.
0:17:24 You know, there’s also your point on like data labeling, there is a lot of data in these
0:17:27 things now, but there is a lot more data out in the world.
0:17:30 And there’s, you know, at least in theory, some of the leading AI companies are actually
0:17:32 paying to generate new data.
0:17:34 And by the way, even like the open source data sets are getting much better.
0:17:36 And so there’s a lot of like data improvements that are coming.
0:17:39 And then, you know, there’s just the amount of money pouring into the space to be able
0:17:41 to underwrite all this.
0:17:42 And then by the way, there’s also just the systems engineering work that’s happening,
0:17:45 right?
0:17:46 Which is a lot of the current systems.
0:17:47 You know, we’re basically, we’re built by scientists.
0:17:48 And now they’re really world-class engineers are showing up and tuning them up and getting
0:17:51 them to work better.
0:17:52 And you know, maybe that’s not a, maybe that’s not a.
0:17:55 Which makes training, by the way, way more efficient as well, not just inference, but
0:18:00 also training.
0:18:01 Yeah.
0:18:02 Exactly.
0:18:03 And then even, you know, another improvement area is basically Microsoft released their
0:18:05 five small language model yesterday.
0:18:06 And apparently it’s competitive.
0:18:08 It’s a very small model, competitive with much larger models.
0:18:10 And the big thing they say that they did was they basically optimized the training set.
0:18:14 So they basically de-duplicated the training set.
0:18:16 They took out all the copies and they really optimized on a small amount of training data,
0:18:19 on a small amount of high quality training data, as opposed to the larger amounts of
0:18:22 low quality data that most people train on.
0:18:23 You add all these up and you’ve got eight or 10 different combination of sort of practical
0:18:27 and theoretical improvement vectors that are all in play.
0:18:31 And it’s hard for me to imagine that some combination of those doesn’t lead to like
0:18:33 really dramatic improvement from here.
0:18:35 I definitely agree.
0:18:36 I think that’s for sure going to happen, right?
0:18:38 Like if you were still back to Sam’s proposition, I think if you were a startup and you were
0:18:43 like, okay, in two years I can get as good as GPT-4, you shouldn’t do that.
0:18:48 Right.
0:18:49 That would be a bad mistake.
0:18:51 Right.
0:18:52 Right.
0:18:53 Well, this also goes to, you know, a lot of entrepreneurs are afraid of, well, I’ll give
0:18:55 you an example.
0:18:56 So a lot of entrepreneurs, here’s this thing they’re trying to figure out, which is, okay,
0:18:58 I really think, I know how to build a SaaS app that harnesses an LLM to do really good
0:19:02 marketing collateral.
0:19:03 Let’s just make it very similar.
0:19:04 A very similar thing.
0:19:05 Yeah.
0:19:06 And so I build a whole system for that.
0:19:07 Will it just turn out to be that the big models in six months will be even better in
0:19:11 making marketing collateral just from a simple prompt, such that my apparently sophisticated
0:19:16 system is just irrelevant because the big model just does it?
0:19:18 Yeah.
0:19:19 Yeah.
0:19:20 Let’s talk about that.
0:19:21 Like apps, you know, another way you can think about it is that the criticism of a lot
0:19:24 of current AI app companies is their quote unquote, you know, GPT wrappers, they’re sort
0:19:28 of thin layers of wrapper around the core model, which means the core model could commoditize
0:19:31 them or displace them.
0:19:32 But the counterargument, of course, is it’s a little bit like calling all, you know, old
0:19:36 software apps, you know, database wrappers, you know, wrappers around a database.
0:19:40 It turns out like actually wrappers around a database is like most modern software and
0:19:43 a lot of that actually turned out to be really valuable in there.
0:19:45 It turns out there’s a lot of things to build around the core engine.
0:19:47 So yeah.
0:19:48 So Ben, how do we think about that when we run into companies thinking about building
0:19:50 apps?
0:19:51 Yeah.
0:19:52 You know, it’s a very tricky question because there’s also this correctness gap, right?
0:19:56 So you know, why do we have co-pilots?
0:20:00 Where are the pilots?
0:20:01 Right?
0:20:02 Where are the AI?
0:20:03 There’s no AI pilots.
0:20:04 They’re only AI co-pilots.
0:20:05 There’s a human in the loop on absolutely everything.
0:20:09 And that really kind of comes down to this, you know, you can’t trust the AI to be correct
0:20:16 in drawing a picture or writing a program or, you know, even like writing a court brief
0:20:24 without making up citations, you know, all these things kind of require a human and
0:20:30 kind of turns out to be like fairly dangerous to not.
0:20:33 And then I think that so what’s happening a lot with the application layer is people
0:20:37 saying, well, to make it really useful, I need to turn this co-pilot into a pilot.
0:20:43 And can I do that?
0:20:44 And so that’s an interesting and hard problem.
0:20:48 And then there’s a question of, is that better done at the model level or at some layer on
0:20:53 top that, you know, kind of teases the correct answer out of the model, you know, by doing
0:20:59 things like using code validation or what have you?
0:21:01 Or is that just something that the models will be able to do?
0:21:04 I think that’s one open question.
0:21:06 And then, you know, as you get into kind of domains and, you know, potentially rapprocent
0:21:11 things, I think there’s a different dimension than what the models are good at, which is
0:21:16 what is the process flow, which is kind of in database for all to so on the database kind
0:21:23 of analogy, there is like the part of the task in a law firm that’s writing the brief,
0:21:31 but there’s 50 other tasks and things that have to be integrated into the way a company
0:21:38 works, like the process flow, the orchestration of it.
0:21:42 And maybe there are, you know, a lot of these things, like if you’re doing video production,
0:21:46 there’s many tools or music, even, right, like, okay, who’s going to write the lyrics,
0:21:51 which AI, I’ll write the lyrics and which AI, I’ll figure out the music, and then like,
0:21:56 how does that all come together and how do we integrate it and so forth.
0:22:00 And those things tend to, you know, just require a real understanding of the end customer and
0:22:08 so forth in a way, and that’s typically been how like applications have been different
0:22:13 than platforms in the past is like, there’s real knowledge about how the customer using
0:22:19 it wants to function that doesn’t have anything to do with the kind of intelligent or is just
0:22:26 different than what the platform is designed to do.
0:22:30 And to get that out of the platform for a kind of company or a person turns out to be
0:22:35 really, really hard.
0:22:36 And so those things, I think, are likely to work, you know, especially if the process
0:22:41 is very complex.
0:22:42 And it’s something that’s funny as a firm, you know, we’re a little more hardcore technology
0:22:47 oriented, and we’ve always struggled with those, you know, in terms of, oh, this is
0:22:52 like a some process application for like plumbers to figure out this, and we’re like, well,
0:22:59 where’s the technology.
0:23:01 But you know, a lot of it is how do you encode, you know, some level of domain expertise and
0:23:07 kind of how things work in the actual world back into the software.
0:23:13 I often think I’ve been told founders that you can think about this in terms of price,
0:23:16 you can kind of work backwards from pricing a little bit, which is to say sort of business
0:23:19 value and what you can charge for, which is, you know, the natural thing for any technologists
0:23:23 to do is to kind of say, I have this new technological capability, and I’m going to sell it to people
0:23:26 and like, what am I going to charge for it is going to be somewhere between, you know,
0:23:29 my cost of providing it and then, you know, whatever markup I think I can justify, you
0:23:33 know, and if I have a monopoly providing it, maybe the markup is infinite.
0:23:36 But, you know, it’s kind of this technology forward, you know, kind of supplier supply
0:23:40 forward, you know, pricing model, there’s a completely different pricing model for kind
0:23:44 of business value backwards, and sort of, you know, so-called value pricing, value-based
0:23:49 pricing.
0:23:50 And that’s, you know, to your point, that’s basically a pricing model that says, okay,
0:23:53 what’s the business value to the customer of the thing?
0:23:56 And if the business value is, you know, a million dollars, then can I charge 10% of
0:24:01 that and get $100,000, right, or whatever?
0:24:04 And then, you know, why is it cost $100,000 as compared to $5,000 is because, well, because
0:24:09 to the customer, it’s worth a million dollars, and so they’ll pay 10% for it.
0:24:12 Yeah, actually, so a great example of that, like, we’ve got a company in our portfolio,
0:24:19 Crest AI that does things like debt collection, okay, so if I can collect way more debt with
0:24:28 way fewer people with my, you know, it’s a co-pilot type solution, then what’s that
0:24:36 worth?
0:24:37 Well, it’s worth a heck of a lot more than just buying an open AI license because an
0:24:43 open AI license is not going to easily collect debts, or kind of enable your debt collectors
0:24:50 to be massively more efficient, or that kind of thing, so it’s bridging that gap between
0:24:56 the value.
0:24:57 And I think you had a really important point, the test for whether your idea is good is how
0:25:00 much can you charge for it?
0:25:02 Can you charge the value?
0:25:04 Or are you just charging the amount of work it’s going to take the customer to put their
0:25:10 own wrapper on top of open AI, like, that’s the real testimony of, like, how deep and
0:25:17 how important is what you’ve done?
0:25:19 Yeah, and so to your point on, like, the kinds of businesses that technology investors have
0:25:24 had a hard time with, you know, kind of thinking about, you know, maybe accurately is sort of,
0:25:28 it’s the company that is, it’s a vendor that has built something where it is a specific
0:25:32 solution to a business problem, where it turns out the business problem is very valuable
0:25:36 to the customer.
0:25:37 And so therefore they will pay a percentage of the value provided back in the terms for
0:25:43 price for the software, and that actually turns out you can have businesses that are
0:25:47 not very technologically differentiated that are actually extremely lucrative.
0:25:52 And then because that business is so lucrative, they can actually afford to go think very
0:25:56 deeply about how technology integrates into the business, what else they can do.
0:26:00 You know, this is like the story of a Salesforce.com, for example, right?
0:26:03 And by the way, there’s kind of a, a chance, a theory that the models are all getting really
0:26:09 good.
0:26:10 There are open source models, there are, like, that are awesome, you know, Lama, Mistral,
0:26:16 like these are great models.
0:26:18 And so the actual layer where the value is going to crew is going to be, like, tools,
0:26:24 orchestration, that kind of thing, because you can just plug in whatever the best model
0:26:28 is at the time, whereas the models are going to be competing, you know, in a death battle
0:26:33 with each other and, you know, be commoditized down to the, you know, the cheapest one wins
0:26:39 and that kind of thing.
0:26:40 So, you know, you could argue that the, the best thing to do is, is to kind of connect
0:26:47 the power to the people.
0:26:49 Right.
0:26:50 Right.
0:26:51 So that actually takes us to the next question, and this is a two-in-one question.
0:26:54 So Michael asks, and these are, and I’ll say these are diametrically opposed, which
0:26:57 is why I paired them.
0:26:58 So Michael asks, why are VCs making huge investments in generative AI startups when it’s clear
0:27:03 these startups won’t be profitable anytime soon, which was a loaded, loaded question,
0:27:07 but we’ll take it.
0:27:08 And then Kaiser asks, if AI deflates the cost of building a startup, how will the structure
0:27:12 of tech investment change?
0:27:14 And of course, Ben, this goes to exactly what you just said.
0:27:16 So it’s basically the questions are diametrically opposed, because if you squint out of your
0:27:20 left eye, right, what you see is basically the amount of money being invested in the
0:27:23 foundation model companies kind of going up to the right at a furious pace, you know,
0:27:26 these companies are raising hundreds of millions, billions, tens of billions of dollars.
0:27:29 And it’s just like, oh my God, look at these sort of capital, you know, sort of, I don’t
0:27:33 know, infernos, you know, that hopefully will result in value at the end of the process.
0:27:37 But my God, look at how much money is being invested in these things.
0:27:39 If you squint through your right eye, you know, you think, wow, that now all of a sudden it’s
0:27:43 like much easier to build software.
0:27:45 It’s much easier to have a software company.
0:27:46 It’s much easier to like have a small number of programmers writing complex software because
0:27:49 they’ve got all these AI co-pilots and all these automated software development capabilities
0:27:53 that are coming online.
0:27:54 Yeah.
0:27:55 So on the other side, the cost of building an AI like application startup might, you
0:27:59 know, crash.
0:28:00 And it might just be that like the, you know, the Salesforce, the AI Salesforce.com might
0:28:04 cost, you know, a tenth or a hundredth or a thousandth of the amount of money that it
0:28:07 took to build the, you know, the old database driven Salesforce.com.
0:28:10 And so yeah, so what do we think of the dichotomy, which is you can actually look, you can actually
0:28:14 look out of either eye and see cost, either cost to the moon as like for startup funding
0:28:19 or cost actually going to zero.
0:28:21 Yeah.
0:28:22 Well, like so it is interesting.
0:28:24 I mean, we actually have companies in both camps, right?
0:28:27 Like I think probably the companies that have gotten to profitability the fastest, maybe
0:28:33 in the history of the firm have been AI companies or been, you know, AI companies in the portfolio
0:28:37 where the revenue grows so fast that it actually kind of runs out ahead of the cost.
0:28:44 And then there are like, you know, people who are in the foundation model race who are
0:28:49 raising hundreds of millions, you know, even billions of dollars to kind of keep pace and
0:28:54 so forth.
0:28:55 They also are kind of generating revenue at a fast rate.
0:29:00 The headcount in all of them is small.
0:29:02 So I would say, you know, where AI money goes and even, you know, like if you look at open
0:29:09 AI, which is the big spender in startup world, which, you know, we are also investors and
0:29:16 is you know, headcount wise, they’re pretty small against their revenue.
0:29:20 Like it is not a big company headcount.
0:29:22 Like if you look at the revenue level and how fast they’ve gotten there, it’s pretty
0:29:27 small.
0:29:28 Now, the total expenses are ginormous, but they’re going into the model creation.
0:29:33 So it’s an interesting thing.
0:29:35 I mean, I’m not entirely sure how to think about it, but I think like if you’re not building
0:29:40 a foundation model, it will make you more efficient and probably gets profitability
0:29:45 quicker.
0:29:46 Right.
0:29:47 So the counter the counter and this is a very bullish counter argument, but the counter
0:29:51 argument to that would be basically that falling costs for like building new software
0:29:55 companies are a mirage.
0:29:57 And the reason for that is this thing in economics called the Jevons paradox, which I’m going
0:30:01 to read from Wikipedia.
0:30:02 So the Jevons paradox occurs when technological progress increases the efficiency with which
0:30:07 a resource is used, right, reducing the amount of that resource necessary for any one use.
0:30:12 But the falling cost induces increases in demand, right, elasticity, enough that the
0:30:17 resource use overall is increased rather than reduced.
0:30:20 Yeah.
0:30:21 That’s certainly possible.
0:30:23 Right.
0:30:24 And so this is you see, you see versions of this, for example, you build in your freeway
0:30:27 and it actually makes traffic jams worse, right?
0:30:30 Because basically what happens is, oh, it’s great.
0:30:31 Now there’s more roads.
0:30:32 Now we can have more people live here.
0:30:34 We can have more people that, you know, we can make these companies bigger and now there’s
0:30:36 more traffic than ever.
0:30:37 And now the traffic is even worse.
0:30:39 Or you saw the classic examples during the Industrial Revolution, coal consumption, as
0:30:43 the price of coal drops, people use so much more coal that actually the overall consumption
0:30:48 actually increased.
0:30:49 And people are getting a lot more power, but the result was the use of a lot more coal
0:30:53 in the paradox.
0:30:54 And so the paradox here would be, yes, the cost of developing any given piece of software
0:30:59 falls, but the reaction to that is a massive surge of demand for software capabilities.
0:31:05 And so the result of that actually is, although it even, it looks like starting software companies,
0:31:08 the price is going to fall.
0:31:09 Actually, what’s going to happen is it’s going to rise for the high quality reason that you’re
0:31:12 going to be able to do so much more, right, with software.
0:31:16 The products are going to be so much better and the roadmap is going to be so amazing
0:31:19 of the things you can do.
0:31:20 And the customers are going to be so happy with it that they’re going to want more and
0:31:22 more and more.
0:31:24 So the result of it, and by the way, another example of Jevons Faradok’s playing out in
0:31:27 another related industries in Hollywood, you know, CGI in theory should have reduced the
0:31:31 price of making movies.
0:31:33 In reality, it’s increased it because audience expectations went up.
0:31:36 And now you go to a Hollywood movie and it’s wall-to-wall CGI.
0:31:39 And so, you know, movies are more expensive to make than ever.
0:31:41 And so the result of it, you know, so, but the result in Hollywood is at least much more,
0:31:45 let’s say visually elaborate, you know, movies, whether they’re better or not is another question,
0:31:48 but like much more visually elaborate, compelling, kind of visually stunning movies through CGI.
0:31:52 The version here would be much better software, like radically better software to the end user,
0:31:57 which causes end users to want a lot more software, which causes actually the price of development
0:32:01 to rise.
0:32:02 You know, if you just think about like a simple case like travel, like, okay, booking a trip
0:32:07 through Expedia is like complicated, you’re likely to get it wrong, you’re clicking on
0:32:12 menus and this and that and the other and like, you know, in AI version of that would
0:32:17 be like, you know, send me to Paris, put me in a hotel I love at the best price, you know,
0:32:22 send me on the best possible kind of airline, an airline ticket and then, you know, like
0:32:29 make it like really special for me and like maybe you need a human to go, okay, like we’re
0:32:35 going to, you know, or maybe the AI gets far complicated and says, okay, well, we know
0:32:40 the person loves chocolate and we’re going to like, you know, FedEx in the best chocolate
0:32:45 in the world from Switzerland into this hotel in Paris and this and that and the other.
0:32:50 And so like the quality, you can, the quality could get to levels that we can’t even imagine
0:32:56 today just because, you know, the software tools aren’t, aren’t what they’re going to
0:33:00 be.
0:33:01 So that’s right.
0:33:02 Yeah, I kind of buy that actually.
0:33:04 I think I brought the argument you’re both is how about, yeah, or how about I’m going
0:33:10 to land in whatever Boston at six o’clock, I want to have dinner at seven with a table
0:33:13 full of like super interesting people.
0:33:15 Yeah, right, right, right, right, right, right, right, right, like, yeah, yeah, yeah, yeah,
0:33:21 no, no travel agent would do that for you today, nor would you want them to.
0:33:24 No.
0:33:25 No.
0:33:26 Right.
0:33:27 Well, and then you think about it, it’s got to be integrated into my personal AI and
0:33:33 like, and this is, you know, there’s just like unlimited kind of ideas that you can
0:33:37 do.
0:33:38 And I think this is one of the kind of things that’s always been underestimated about humans
0:33:43 is like our ability to come up with new things we need.
0:33:48 Like that has been unlimited.
0:33:50 And there’s a very kind of famous case where John Maynard Keynes, the kind of prominent
0:33:56 economist in the kind of first half of last century, had this thing that he predicted,
0:34:01 which is like, nobody, because of automation, nobody would ever work a 40 hour work week,
0:34:08 you know, like good, because once their needs were met, needs being like shelter and food.
0:34:15 And you know, I don’t even know if transportation was in there.
0:34:17 Like that was it.
0:34:18 It was over.
0:34:19 You would never work past the need for shelter and food.
0:34:23 Like why would you?
0:34:24 Like there’s no reason to, but of course needs expanded.
0:34:27 So then everybody needed a refrigerator, everybody needed not just one car, but a car for everybody
0:34:32 in the family.
0:34:33 Everybody needed a television set, everybody needed like glorious vacations, everybody,
0:34:38 you know.
0:34:39 So what are we going to need next?
0:34:41 I’m quite sure that I can’t imagine it, but like somebody’s going to imagine it and it’s
0:34:46 quickly going to become a need.
0:34:48 Yeah, that’s right.
0:34:49 By the way, as Keynes famously said, his essay I think was economic prospects for our grandchildren,
0:34:55 which was basically that.
0:34:56 Yeah.
0:34:57 You just articulate it.
0:34:58 So Karl Marx said another version of that, I just pulled up the quote.
0:35:00 So that society, when you know, when the Marxist utopia socialism is achieved, society regulates
0:35:06 the general production.
0:35:07 That makes it possible for me to do blah, blah, blah, to hunt in the morning, fish in
0:35:12 the afternoon, rear cattle in the evening, criticize after dinner.
0:35:18 What a glorious life.
0:35:20 What a glorious life.
0:35:21 Like if I could just list four things that I do not want to do, it’s hunt, fish, rear
0:35:26 cattle and criticize.
0:35:27 Yeah.
0:35:28 Yeah.
0:35:29 Right.
0:35:30 And by the way, it says a lot about Marx that those were his four things.
0:35:32 Well, the criticizing being his favorite thing, I think it’s basically communism in
0:35:37 a nutshell.
0:35:38 Yeah.
0:35:39 Exactly.
0:35:40 I don’t want to get too political, but yes, yes, 100%.
0:35:43 And so yeah, so it’s this, this, yeah, do you, what they, what they have with Keynes and
0:35:46 Marx in common is just this incredibly constricted, it’s incredibly constricted view of what people
0:35:50 want to do.
0:35:51 And then, and then correspondingly, you know, the other thing is just like, you know, people,
0:35:53 people who want, people who want to have a mission.
0:35:55 I mean, probably some people just want to fish and hunt, but you know, a lot of, a lot
0:35:58 of people want to have a mission.
0:35:59 They want to have a cause.
0:36:00 They want to have a purpose.
0:36:01 They want to be useful.
0:36:02 They want to be productive.
0:36:03 It’s actually a good thing in life.
0:36:04 It turns out.
0:36:05 It turns out.
0:36:06 Yeah.
0:36:07 In the startling turn of events.
0:36:09 Okay.
0:36:10 So yeah.
0:36:11 So yeah, I think that I’ve long felt, you know, a little bit of the software eats the
0:36:13 world thing a decade ago.
0:36:15 I’ve always thought that, I’ve always thought that basically demand for software is sort
0:36:18 of perfectly elastic, possibly to infinity.
0:36:20 And the theory there basically is if you just continuously bring down the cost of software,
0:36:24 you know, which has been happening over time, then basically demand, you know, basically
0:36:26 is like basically perfectly correlates upward.
0:36:29 And the reason is because, you know, kind of as we’ve been discussing, but it’s kind
0:36:32 of there’s, there’s always something else to do in software.
0:36:35 There’s always something else to automate.
0:36:36 There’s always something else to optimize.
0:36:37 There’s always something else to improve.
0:36:40 There’s always something to make better.
0:36:41 And, you know, in the moment with the constraints that you have today, you may not, you know,
0:36:44 think of what that is, but the minute you don’t have those constraints, you’ll imagine
0:36:47 what it is.
0:36:48 I’ll just give you an example.
0:36:49 I mean, I’ll give you an example of playing out with AI right now, right?
0:36:51 So there have been, and we have, you know, we have companies that do this.
0:36:54 You know, there have been, you know, there have been companies that have made AI, you
0:36:56 know, that have made software systems for doing security cameras forever, right?
0:37:00 And it’s like, for a long time, it was like a big deal to have software that would do
0:37:03 like, you know, have different security camera feeds and store them on a DVR and be able
0:37:06 to replay them and have an interface that lets you do that.
0:37:09 Well, it’s like, you know, AI security cameras, all of a sudden can have like, actual like,
0:37:13 semantic knowledge of what’s happening in the environment.
0:37:14 And so they can say, you know, hey, that’s Ben, and then they can say, oh, hey, you know,
0:37:18 that’s Ben, but he’s carrying a gun.
0:37:19 Yeah.
0:37:20 Right.
0:37:21 Right.
0:37:22 And by the way, that’s Ben and he’s carrying a gun, but that’s because like he hunts on,
0:37:24 you know, on Thursdays and Fridays, as compared to that’s Mary and she never carries a gun
0:37:28 and like, you know, like something is wrong and she’s really mad, right?
0:37:32 She’s got a, yeah, really steamed expression on her face and we should probably be worried
0:37:35 about it, right?
0:37:36 So there’s like an entirely new set of capabilities you can do just as one example for security
0:37:40 systems that were never possible pre AI and the security system that actually has a semantic
0:37:44 understanding of the world is obviously much more sophisticated than the one that doesn’t
0:37:48 and might actually be more expensive to make, right?
0:37:50 Right.
0:37:51 Well, and just imagine healthcare, right?
0:37:53 Like you could wake up every morning and have a complete diagnostic, you know, like
0:38:01 how am I doing today?
0:38:02 Like what are all my levels of everything?
0:38:04 And, you know, how should I interpret them, you know, better than, you know, this is one
0:38:09 thing where AI is really good is, you know, medical diagnosis because it’s a super high
0:38:14 dimensional problem.
0:38:16 But if you can get access to, you know, your continuous glucose reading, you know, maybe
0:38:21 sequester blood now and again, this and that and the other, yeah, you’ve got an incredible
0:38:26 kind of view of things and who doesn’t want to be healthier, you know, like now we have
0:38:32 a scale.
0:38:33 That’s basically what we do, you know, maybe check your heart rate or something, but like
0:38:39 pretty primitive stuff compared to where we could go.
0:38:41 Yeah, that’s right.
0:38:42 Okay, good.
0:38:43 All right.
0:38:44 So let’s go to the next topic.
0:38:45 So on the topic of data, so a major Tom asks, as these AI models allow for us to copy existing
0:38:50 app functionality at minimal cost, proprietary data seems to be the most important moat.
0:38:55 How do you think that will affect proprietary data value?
0:38:58 What other moats do you think companies can focus on building in this new environment?
0:39:01 And then Jeff Weishaupt asks, how should companies protect sensitive data, trade secrets, proprietary
0:39:06 data, individual privacy, and the brave new world of AI?
0:39:09 So let me start with a provocative statement, Betsy, if you agree with it, which is, you
0:39:15 know, you sort of hear a lot, this sort of statement or cliche is like data is the new
0:39:18 oil.
0:39:19 And so it’s like, okay, data is the key input to training AI, making all this stuff work.
0:39:23 And so, you know, therefore, you know, data is basically the new resource.
0:39:26 It’s the limiting resource.
0:39:27 It’s the super valuable thing.
0:39:29 And so, you know, whoever has the best data is going to win, and you see that directly
0:39:32 in how you train AI’s.
0:39:33 And then, you know, you also have like a lot of companies, of course, that are now trying
0:39:36 to figure out what to do with AI.
0:39:38 And a very common thing you’ll hear from companies is, well, we have proprietary data, right?
0:39:42 So I’m, you know, I’m a hospital chain or I’m, you know, whatever, any kind of business,
0:39:46 insurance company or whatever.
0:39:47 And I’ve got all this proprietary data that I can apply, you know, that I’ll be able to,
0:39:50 you know, build things with my proprietary data with AI that won’t just, you know, be
0:39:54 something that anybody will be able to have.
0:39:56 Let me argue that basically, let’s see, let me argue in like almost every case like that,
0:40:01 it’s not true.
0:40:02 It’s basically what the Internet kids would call cope.
0:40:04 It’s simply not true.
0:40:05 And the reason it’s just not true is because the amount of data available on the Internet
0:40:10 and just generally in the environment is just a million times greater.
0:40:16 And so, while it may not, you know, while it may not be true that I have your specific
0:40:19 medical information, I have so much medical information off the Internet for so many people
0:40:24 in so many different scenarios that it just swamps the value of quote, your data, you
0:40:30 know, just, it’s just, it’s just like overwhelming.
0:40:31 And so your, your, your proprietary data as, you know, company acts will be a little bit
0:40:35 useful on the margin, but it’s not actually going to move the needle.
0:40:37 And it’s not really going to be a barrier to entry in most cases.
0:40:40 And then let me cite as proof for the, for my, my belief that this is mostly cope is
0:40:45 there has never been nor is there now any sort of basically any level of sort of rich
0:40:49 or sophisticated marketplace for data, market for data, there’s no, there’s no, there’s no
0:40:54 large marketplace for data.
0:40:56 And in fact, in fact, what there are is there are very small markets for data.
0:40:59 So there are these businesses called data brokers that will sell you, you know, large
0:41:01 numbers of like, you know, information about users on the Internet or something.
0:41:05 And they’re just small businesses, like they’re just not large, it just turns out like information
0:41:09 on lots of people is just not very valuable.
0:41:11 And so if the data actually had value, you know, it would have a market price and you
0:41:15 would see it transacting and you actually very specifically don’t see that, which is
0:41:19 sort of a, you know, yeah, sort of quantitative proof that the data actually is not nearly
0:41:23 as valuable as people think it is.
0:41:25 Where I agree, so I agree that the data, like just as here’s a bunch of data and I can sell
0:41:34 it without doing anything to the data is like massively overrated, like I definitely agree
0:41:42 with that.
0:41:43 And like maybe I can imagine some exceptions, like some, you know, special population genomic
0:41:49 databases or something that are, that were very hard to acquire, that are useful in some
0:41:53 way that’s, you know, that’s not just like living on the Internet or something like that.
0:41:57 I could imagine where that’s super highly structured, very general purpose and not widely available.
0:42:04 But for most data in companies is not like that.
0:42:07 And that it tends to not, it’s either widely available or not general purpose.
0:42:12 It’s kind of specific.
0:42:14 Having said that, right, like companies have made great use of data, for example, a company
0:42:20 that you’re familiar with, Meta, uses its data to kind of great ends itself, feeding
0:42:26 it into its own AI systems, optimizing its products in incredible ways.
0:42:31 And I think that, you know, us, Andreessen Horowitz, actually, you know, so we just raised
0:42:35 $7.2 billion and it’s not a huge deal.
0:42:40 But we took our data and we put it into an AI system and our LPs were able, there’s a
0:42:47 million questions investors have about everything we’ve done, our track record, every company
0:42:53 we’ve invested and so forth.
0:42:55 And for any of those questions, they could just ask the AI, they could be wake up at
0:42:58 three o’clock in the morning, go, “Do I really want to trust these guys?”
0:43:02 And go in and ask the AI a question and boom, they’d get an answer back instantly.
0:43:05 They’d have to wait for us and so forth.
0:43:07 So we really kind of improved our investor relations product tremendously through use
0:43:12 of our data.
0:43:14 And I think that almost every company can improve its competitiveness through use of its own
0:43:21 data.
0:43:22 But the idea that it’s collected some data that it can go like sell or that is oil or
0:43:30 what have you, that’s, yeah, that’s probably not true.
0:43:36 I would say, and you know, it’s kind of interesting because a lot of the data that you would think
0:43:41 would be the most valuable would be like your own code base, right?
0:43:46 Your software that you’ve written, so much of that lives in GitHub.
0:43:49 Nobody is actually, I don’t know of any company, we work with, you know, whatever a thousand
0:43:55 software companies and do we know any that’s like building their own programming model
0:44:00 on their own code?
0:44:02 Like, or, and would that be a good idea?
0:44:05 Probably not just because there’s so much code out there that the systems have been
0:44:09 trained on.
0:44:10 So like that’s not so much of an advantage.
0:44:14 So I think it’s a very specific kind of data that would have value.
0:44:17 Well, let’s, let’s make it actionable then.
0:44:19 If I’m, if I’m running a big company, like if I’m running an insurance company or a bank
0:44:23 or a hospital chain or something like that, like how, or, you know, a consumer packaged
0:44:27 goods company, Pepsi or something, like what, how should I validate like, how should I validate
0:44:32 that I actually have a valuable proprietary data asset that I should really be focusing
0:44:36 on using versus maybe versus in the alternate, by the way, maybe there’s other things that
0:44:40 maybe I should be taking all the effort I would spend on trying to optimize use of that
0:44:43 data and maybe I should use it entirely trying to build things using internet data instead.
0:44:47 Yeah, so, so I think, I mean, look, if you’re right, if you’re in the insurance business,
0:44:53 then like all your actuarial data is both interesting and then I, I don’t know that anybody publishes
0:45:00 their actual, or actuarial data.
0:45:03 And so like, I’m not sure how you would train the model on stuff off of the internet.
0:45:08 Yes.
0:45:09 That’s good.
0:45:10 Let me, let me, can I challenge that one?
0:45:11 So that, that would be good.
0:45:12 That’d be a good thing.
0:45:13 That’d be a good test case.
0:45:14 So I’m an insurance company.
0:45:15 I’ve got records on 10 million people and, you know, the actuarial tables and when they,
0:45:17 when they get sick and when they die.
0:45:18 Okay.
0:45:19 That’s great.
0:45:20 But like there’s lots and lots of actuarial, general actuarial data on the internet for
0:45:24 large scale populations, you know, because governments collect the data and they process
0:45:28 it and they publish reports.
0:45:29 And there’s lots of, there’s lots of academic studies.
0:45:32 And so like, is your, is, is your large data set giving you any additional actuarial information
0:45:38 that the much larger data set on the internet isn’t already providing you?
0:45:41 Like are your, are your insurance clients actually actuarially any different than just
0:45:46 everybody?
0:45:47 I think so.
0:45:48 Cause on intake on the, you know, when you get insurance, they give you like a blood test.
0:45:55 They’ve got all these things we know if you’re a smoker and so forth.
0:45:58 And in the, I think in the general data set, like, yeah, you know who dies, but you don’t
0:46:02 know what the fuck they did coming in.
0:46:05 And so what you really are looking for is like, okay, for this profile of person with
0:46:09 this kind, with these kinds of lab results, how long do they live?
0:46:13 And that’s, that’s where the value is.
0:46:15 And I think that, you know, interesting, like, you know, I was thinking about like a company
0:46:20 like Coinbase where, right, they have incredibly valuable assets in the terms of money.
0:46:27 They have to stop people from breaking in.
0:46:29 They’ve done a massive amount of work on that.
0:46:32 They’ve seen all kinds of break-in types.
0:46:34 I’m sure they have tons of data on that.
0:46:36 It’s probably like weirdly specific to people trying to break into crypto exchanges.
0:46:42 And so, you know, like, I think it could be very useful for them.
0:46:45 I don’t think they could sell it to anybody, but, you know, I think every company’s got
0:46:51 data that if, you know, fed into an intelligent system would help their business.
0:46:57 And I think almost nobody has data that they could just go sell.
0:47:02 And then there’s this kind of in-between question, which is, what data would you want to let
0:47:08 Microsoft or Google or OpenAI or anybody get their grubby little fingers on?
0:47:13 And that I’m not sure.
0:47:19 That’s a — that I think is the question that enterprises are wrestling with more than — it’s
0:47:24 not so much should we go like sell our data, but should we train our own model just so
0:47:29 we can maximize the value?
0:47:32 Or should we feed it into the big model?
0:47:35 And if we feed it into the big model, do all of our competitors now have the thing that
0:47:39 we just did?
0:47:40 And, you know, or could we trust the big company to not do that to us, which I kind of think
0:47:47 the answer on trusting the big company not to F with your data is probably I wouldn’t
0:47:52 do that.
0:47:55 If your competitiveness depends on that, you probably shouldn’t do that.
0:47:58 Well, there are at least reports that certain big companies are using all kinds of data
0:48:02 that they should be using to train their models already.
0:48:04 So.
0:48:05 Yep.
0:48:06 I think like I think those reports are very likely true.
0:48:10 Right.
0:48:11 Or they have open data, right?
0:48:12 Like this is, you know, we’ve talked about this before, but, you know, the same companies
0:48:17 that are saying they’re not stealing all the data from people or taking it in an unauthorized
0:48:22 way, refuse to say open their data.
0:48:26 Like why not tell us where your data came from?
0:48:28 And then in fact, they’re trying to shut down all openness, no open source, no open weights,
0:48:32 no open data, no open nothing, and go to the government and try and get to do that.
0:48:36 You know, if you’re not a thief, then why are you doing that?
0:48:39 Right.
0:48:40 Right.
0:48:41 Right.
0:48:42 What are you hiding?
0:48:43 By the way, there’s other twists and turns here.
0:48:44 The insurance example, I kind of deliberately loaded it because you may know it’s actually
0:48:48 illegal to use genetic data for insurance purposes, right?
0:48:51 So there’s this thing called the GenoLaw Genetic Information Non-Discrimination Act of 2008.
0:48:58 And basically, it basically bans health insurers in the U.S. room actually using genetic data
0:49:02 for the purpose of doing, you know, health assessment, actual assessment of, which by
0:49:06 the way, because now the genomics are getting really good.
0:49:08 Like that data probably actually is, you know, among the most accurate data you could have
0:49:12 if you were actually trying to predict, like, when people are going to get sick and die.
0:49:15 And they’re literally not allowed to use it.
0:49:18 Yeah, it is.
0:49:20 I think that this is an interesting, like weird misapplication of good intentions in
0:49:28 a policy way that’s probably going to kill more people than ever get saved by every kind
0:49:37 of health FDA, et cetera, policy that we have, which is, you know, in a world of AI, having
0:49:46 access to data on all humans, why they get sick, what their genetics were, et cetera,
0:49:51 et cetera, et cetera, is the most, that is, you know, you’re talking about data being
0:49:55 the new oil, like that is the new oil, that’s the healthcare oil is, you know, if you could
0:49:59 match those up, then we’d never not know why we’re sick, you know, you could make everybody
0:50:05 much healthier, all these kinds of things.
0:50:08 But you know, to kind of stop the insurance company from kind of overcharging people who
0:50:15 are more likely to die, we’ve kind of locked up all this data, a kind of better idea would
0:50:23 be to just go, okay, for the people who are likely to, like we subsidize healthcare, like
0:50:29 massively for individuals anyway, just like differentially subsidize, and, you know, and
0:50:38 then like you solve the problem and you don’t lock up all the data.
0:50:41 But yeah, it’s typical of politics and policy, I mean, most of them are like that, I think.
0:50:47 Yeah.
0:50:48 Well, there’s these interesting questions like insurance, like basically, they’re one
0:50:50 of the questions people have asked about insurance is like, if you had perfectly predictive
0:50:53 information on like individual outcomes, does the whole concept of insurance actually still
0:50:57 work, right, because the whole theory of insurance is risk pooling, right, it’s precisely the
0:51:03 fact that you don’t know what’s going to happen in the specific case that means you build
0:51:06 these statistical models, and then you risk pool, and then you have variable payouts depending
0:51:10 on exactly what happens.
0:51:11 But if you literally knew what was going to happen in every case, because for example,
0:51:15 you have all this predictive genomic data, then all of a sudden it wouldn’t make sense
0:51:18 to risk pool, because you just say, well, no, this person’s going to cost X, that person’s
0:51:22 going to cost Y, there’s no.
0:51:24 Self-insurance already doesn’t make sense in that way, right, like insurance, the idea
0:51:29 of insurance is kind of like the, it started with crop insurance where like, okay, you
0:51:35 know, my crop fails, and so we all put money in a pool in case like my crop fails so that,
0:51:41 you know, we can cover it.
0:51:43 It’s kind of designed for it to risk pool for a catastrophic unlikely incident.
0:51:49 Like everybody’s got to go to the doctor all the fucking time.
0:51:53 And some people get sicker than others in that kind of thing.
0:51:56 But like, the way our health insurance works is like all medical gets, you know, paid for
0:52:02 through this insurance system, which is this layer of loss and bureaucracy and giant companies
0:52:08 and all this stuff when like, if we’re going to pay for people’s healthcare, just pay for
0:52:14 people’s healthcare.
0:52:15 Like, what are we doing, right, like, and if you want to disincent people from like
0:52:20 going for nonsense reasons and just up the copay, like, it’s like, what are we doing?
0:52:27 Just, well, then from a justice standpoint, from a fairness standpoint, like, would it
0:52:31 make sense for me, you know, would it make sense for me to pay more for your healthcare?
0:52:35 If I knew that you were going to be more expensive than me, like, you know, I’m directly, you
0:52:38 know, if you, if everybody knows what future healthcare costs is per person, there has a
0:52:43 very good predictive model for it, you know, societal willingness to all pool in the way
0:52:46 that we do today might really diminish.
0:52:47 Yeah, yeah.
0:52:48 Well, and then like, you could also, if you knew, like, there’s things that you do genetically
0:52:54 and maybe we give everybody a pass on that, it’s like, you can’t control your genetics,
0:52:58 but then like, there’s things you do behaviorally that like, dramatically increases your chance
0:53:02 of getting sick.
0:53:03 And so maybe, you know, we incentivize people to stay healthy instead of just like paying
0:53:09 for them not to die.
0:53:12 There’s a lot of systemic fixes we could do to the healthcare system.
0:53:17 It couldn’t be designed in a more ridiculous way, I think.
0:53:20 Well, it couldn’t be designed in a more ridiculous way.
0:53:22 It’s actually more ridiculous in some other countries, but it’s pretty crazy here.
0:53:27 Nathan, Nathan Odie asks, what are the strongest common themes between the current state of
0:53:31 AI and web 1.0?
0:53:33 And so let me start there.
0:53:34 Let me give you a theory, Ben, and see what you think.
0:53:36 So I guess it’s questionable, you know, because of my role in, you know, Ben, you, you with
0:53:39 me at Netscape, you know, we get this question a lot because of our role early on with the,
0:53:43 with the internet.
0:53:44 You know, the internet boom was like a major, major event in technology, and it’s still
0:53:47 within a lot of, you know, people’s memories.
0:53:49 And so, you know, the sort of, you know, people like to reason from analogy.
0:53:53 So it’s like, okay, the AI boom must be like the internet boom.
0:53:55 Starting an AI company must be like starting an internet company.
0:53:58 And so, you know, what, what is this like?
0:54:00 And we actually got a bunch of questions like that, you know, that are kind of an analogy
0:54:03 questions like that.
0:54:04 I actually think, you know, and then Ben, you know, you and I were there for the internet
0:54:06 boom.
0:54:07 So we, you know, we lived through that and the bust and the boom and the bust.
0:54:10 So I actually think that the analogy doesn’t really work for the most, it works in certain
0:54:14 ways, but it doesn’t really work for the most part.
0:54:16 And the reason is because the internet, the internet was a network, whereas AI is a computer.
0:54:24 Yep.
0:54:25 Okay.
0:54:26 Yeah.
0:54:27 So, so some people understand what we’re saying.
0:54:29 So, you know, like the PC boom or even, I would say the microprocessor, like my best
0:54:35 analogy is to the microprocessor or even to like the original computers, like back to
0:54:39 the mainframe era.
0:54:40 And the reason is because, yeah, look, what the internet did was the internet, you know,
0:54:43 obviously was a network, but the network connected together many existing computers.
0:54:47 And then of course, people built many other new kinds of computers to connect to the internet.
0:54:50 But fundamentally, the internet was a network and then, and that’s important because most
0:54:54 of, most of the sort of industry dynamics, competitive dynamics, startup dynamics around
0:54:59 the internet had to do with basically building either building networks or building applications
0:55:03 that run on top of networks.
0:55:04 And this, you know, the internet generation of startups was very consumed by network
0:55:07 effects and, you know, all these positive feedback loops that you get when you connect
0:55:12 a lot of people together.
0:55:13 And, you know, things like met, you know, so-called Metcast Law, which is sort of the value of
0:55:17 a network, you know, expands, you know, kind of the way it expands is you have more people
0:55:20 to it.
0:55:21 And then, you know, there were all these fights, you know, these fights, you know, all the
0:55:23 social networks or whatever fighting to try to get network effects and try to steal each
0:55:26 other’s users because of the network effects.
0:55:29 And so it was kind of, you know, it was dominated by network effects, which is what you expect
0:55:32 from a network business.
0:55:34 AI, like, there are some networks effects in AI that we can talk about, but it’s more
0:55:39 like a microprocessor.
0:55:40 It’s more like a chip.
0:55:41 It’s more like a computer in that it’s a system that basically, right, data comes in, data
0:55:47 gets processed, data comes out, things happen.
0:55:50 That’s a computer.
0:55:51 It’s an information processing system.
0:55:52 It’s a computer.
0:55:53 It’s a new kind of computer.
0:55:54 It’s a, you know, we like to say the sort of computers up until now have been what are
0:55:58 called von Neumann machines, which is to say they’re deterministic computers, which is
0:56:01 they’re like, you know, hyper literal and they do exactly the same thing every time.
0:56:05 And if they make a mistake, it’s, it’s the programmer’s fault, but they’re very limited
0:56:08 in their ability to interact with people and understand the world.
0:56:11 You know, we think of AI and large language models as a new kind of computer, a probabilistic
0:56:15 computer, a neural network based computer that, you know, by the way, is not very accurate
0:56:19 and is, you know, doesn’t give you the same result every time.
0:56:22 And in fact, might actually argue with you and tell you that it doesn’t want to answer
0:56:25 your question.
0:56:26 Yeah, which is very different in nature than the old computers.
0:56:31 And it makes you get kind of compulsability, you know, the ability to build things, big
0:56:37 things out of little things more complex.
0:56:40 Right.
0:56:41 But, but the capabilities are new and different and valuable and important because they can
0:56:45 understand language and images and, you know, that all these, all these things that you,
0:56:48 you see when you use.
0:56:49 All of that means we can never solve with deterministic computers.
0:56:53 We can now go after, right?
0:56:55 Right.
0:56:56 Yeah, exactly.
0:56:57 And so I think, I think, Ben, I think the analogy and I think the lessons learned are
0:57:00 much more likely to be drawn from the early days of the computer industry or from the
0:57:02 early days of the microprocessor than the early days of the internet.
0:57:05 Does that, does that sound right?
0:57:06 I think so.
0:57:07 Yeah.
0:57:08 I definitely think so.
0:57:09 And that doesn’t mean there’s no like boom and bust and all that because that’s just
0:57:12 the nature of technology, you know, people get too excited and then they get too depressed.
0:57:18 So there’ll be some of that, I’m sure.
0:57:19 There’ll be overbuild outs, you know, potentially eventually of chips and power and that kind
0:57:24 of thing.
0:57:25 You know, we start with the shortage, but, but I agree.
0:57:27 Like I think networks are fundamentally different in the nature of how they evolved in computers
0:57:34 and the kind of just the adoption curve and all those kinds of things will be different.
0:57:39 Yeah.
0:57:40 So then, and this kind of goes to where, how I think the industry is going to unfold.
0:57:43 And so this is kind of my best theory for kind of what happens from here of this kind
0:57:46 of this, you know, this, this giant question of like, you know, is, is the industry going
0:57:49 to be a few God models or, you know, a very large number of models of different sizes
0:57:52 and so forth.
0:57:54 So the computer, like famously, you know, the, the original computers, like the original
0:57:58 IBM mainframes, you know, the big computers, you know, they were very, very large and expensive
0:58:03 and there were only a few of them.
0:58:05 And the prevailing view actually for a long time was that’s all there would ever be.
0:58:09 And there was this famous statement by Thomas Watson senior, who was the creator of IBM,
0:58:13 you know, which was the dominant company for the first like, you know, 50 years of the
0:58:16 computer industry.
0:58:17 And he said, he said, he said, I believe this actually true, but he said, I don’t, I don’t
0:58:21 know, I don’t know that the world will ever need more than five computers.
0:58:24 And I think the reason for that, it was literally, it was like the government’s going to have
0:58:27 two, and then there’s like three big insurance companies.
0:58:30 And then that’s it.
0:58:31 Yeah.
0:58:32 Who else would need to do all that math?
0:58:34 Exactly.
0:58:35 Yeah.
0:58:36 Who else would need to, who else needs to keep track of huge amounts of numbers?
0:58:38 Who else needs that level of, you know, calculation capability?
0:58:41 It’s just not a relevant, you know, it’s just not, not, not a relevant concept.
0:58:44 And by the way, they were like big and expensive.
0:58:46 And so who else can afford them, right?
0:58:48 And who else can afford all the headcount required to manage them and maintain them?
0:58:51 I mean, this is in the days, I mean, these things were big, these things were so big
0:58:53 that you would have an entire building that got built around a computer, right?
0:58:57 And they’d have like, they’d famously have all these guys in white lab coats, literally
0:59:00 like taking care of the computer, because everything had to be kept super clean or the
0:59:03 computer would stop working.
0:59:05 And so, you know, it was this thing where, you know, today we have the idea of an AI God
0:59:08 model, which is like a big foundation model that, you know, then we had the idea of like
0:59:11 a God mainframe, like there, there would just be a few, a few of these things.
0:59:14 And by the way, if you watch old science fiction, it almost always has this sort of conceit.
0:59:19 It’s like, okay, there’s a big supercomputer and it either is like doing the right thing
0:59:22 or doing the wrong thing.
0:59:23 And if it’s doing the wrong thing, you know, that’s, that’s often the plot of the science
0:59:26 fiction movies is you have to go in and try to figure out how to fix it or defeat it.
0:59:30 And so it’s sort of this, this idea of like a single top down thing.
0:59:33 Of course, and that helped for a long time.
0:59:35 Like that held for, you know, the first few decades.
0:59:37 And then, you know, even when computers, computers started to get smaller.
0:59:40 So then you had so called mini computers was the next phase.
0:59:42 And so that was a computer that, you know, didn’t cost $50 million.
0:59:45 Instead, it costs, you know, $500,000, but even still $500,000 is a lot of money.
0:59:50 People aren’t putting mini computers in their homes.
0:59:52 And so it’s like midsize companies can, can buy mini computers, but certainly people can’t.
0:59:56 And then of course, with the PC, they shrunk down to like $2,500.
0:59:59 And then with the smartphone, they shrunk down to $500.
1:00:01 And then, you know, sitting here today, obviously you have computers of every shape, size, description,
1:00:06 all the way down to, you know, computers that cost a penny.
1:00:08 You know, you’ve got a computer in your thermostat that, you know, basically controls the temperature
1:00:12 in the room and it, you know, probably cost a penny.
1:00:13 And it’s probably some embedded arm chip with firmware on it.
1:00:16 And there’s, you know, many billions of those all around the world.
1:00:18 You buy a new car today.
1:00:19 It has something new cars today have something on the order of 200 computers in them.
1:00:23 And maybe, maybe more at this point.
1:00:25 And so you just basically assume with the chip today, sitting here today,
1:00:28 you just kind of assume that everything has a chip in it.
1:00:30 You assume that everything, by the way, draws electricity or has a battery
1:00:33 because it needs to power the chip.
1:00:35 And then increasingly, you assume that everything’s on the internet
1:00:37 because basically all computers are assumed to be on the internet or they will be.
1:00:41 And so as a consequence, what you have is the computer industry today is this massive pyramid.
1:00:46 And you still have a small number of like these supercomputer clusters
1:00:49 or these giant mainframes that are like the God model, you know, the God mainframes.
1:00:53 And then you’ve got, you know, a larger number of mini computers.
1:00:56 You’ve got a larger number of PCs.
1:00:57 You’ve got a much larger number of smartphones.
1:00:58 And then you’ve got a giant number of embedded systems.
1:01:01 And it turns out like the computer industry is all of those things.
1:01:03 And, you know, what is it, you know, what size of computer do you want is based on
1:01:08 what exactly are you trying to do and who are you and what do you need?
1:01:11 And so if that analogy holds, it basically means actually we are going to have AI models
1:01:16 of every conceivable shape, size, description, capability, right?
1:01:20 Based on trained on lots of different kinds of data, running at very different kinds of scale,
1:01:24 very different privacy, different policies, different, you know, security policies.
1:01:28 You know, you’re just going to have like enormous variability and variety.
1:01:32 And it’s going to be an entire ecosystem and not just a couple of companies.
1:01:35 Yeah, let me see what you think of that.
1:01:37 Well, I think that’s right.
1:01:38 And I also think that the other thing that’s interesting about this era of computing,
1:01:42 if you look at prayers of computing from the mainframe to the smartphone,
1:01:47 a huge source of lock in was basically the difficulty of using them.
1:01:53 So, you know, nobody ever got fired for buying IBM because like, you know,
1:01:58 you had people trained on them, you know, people knew how to use the operating system.
1:02:03 Like it was, you know, it was just kind of like a safe choice.
1:02:07 Due to the massive complexity of like dealing with a computer.
1:02:12 And then even with a smartphone, like the read, you know, why is the Apple computer
1:02:19 smartphone so dominant, you know, what makes it so powerful?
1:02:23 It’s well, because like switching off of it is so expensive and complicated and so forth.
1:02:27 It’s an interesting question with AI because AI is the easiest computer to use by far.
1:02:32 It speaks English.
1:02:33 It’s like talking to a person.
1:02:35 And so like, what is the lock in there?
1:02:38 And so are you completely free to use the size, price, choice, speed that you need
1:02:45 for your particular task, or are you locked into the God model?
1:02:49 And, you know, I think it’s still a bit of an open question, but it’s pretty interesting.
1:02:56 And that that that thing could be very different than prior generations.
1:03:01 Yeah, yeah, that makes sense.
1:03:03 And then just to complete the question, what would we say?
1:03:05 So, Ben, what would you say are lessons learned from the internet era that we lived through
1:03:08 that would apply that people should think about?
1:03:11 I think a big one is probably just the the boom bus nature of it.
1:03:19 That like, you know, the demand, the interest in the internet, the recognition of what it
1:03:25 could be was so high that money just kind of poured in and buckets and, you know, and
1:03:32 then the underlying thing was in internet age was the telecom infrastructure and fiber
1:03:37 and so forth got just unlimited funding and unlimited fiber was built out and then eventually
1:03:42 we had a fiber glut and all the telecom companies went bankrupt and and that was great fun.
1:03:49 But you know, like we entered in a good place and I think that that’s something like that’s
1:03:53 probably pretty likely to happen in AI where like, you know, every company is going to
1:03:58 get funded.
1:03:59 We don’t need that many AI companies.
1:04:01 So a lot of them are going to bust.
1:04:02 There’s going to be a huge, you know, huge investor losses.
1:04:06 There will be an overbuild out of chips for sure at some point.
1:04:11 And then, you know, we’re going to have too many chips and yeah, some chip companies will
1:04:15 go bankrupt for sure.
1:04:17 And then, you know, and I think probably the same thing with data centers and so forth,
1:04:21 like, well, be behind behind behind and then we’ll over build at some point.
1:04:26 So that that all be very interesting.
1:04:29 I think that and that’s kind of the that’s every new technology.
1:04:34 So Carlotta Perez has a great kind of has done, you know, amazing work on this where
1:04:39 I like that is just the nature of a new technology is that you overbuild, you underbuild it, then
1:04:43 you overbuild and, you know, and there’s a hype cycle that funds the build out and a
1:04:49 lot of money is lost.
1:04:50 But we get the infrastructure and that’s awesome because that’s when it really gets adopted
1:04:54 and changes the world.
1:04:55 I want to say, you know, with the internet, the other the other kind of big kind of thing
1:05:01 is the internet went through a couple of phases, right?
1:05:04 Like it went through a very open phase, which was unbelievably great.
1:05:08 It was probably one of the greatest booms to the economy.
1:05:11 It, you know, it certainly created tremendous growth and power in America, both, you know,
1:05:16 kind of economic power and soft cultural power and these kinds of things.
1:05:21 And then, you know, it became closed with the next generation architecture with, you
1:05:26 know, kind of discovery on the internet being owned entirely by Google and, you know, kind
1:05:31 of other things, you know, being owned by other companies and, you know, AI, I think
1:05:36 could go either way.
1:05:37 So it could be very open or like, you know, with kind of misguided regulation, you know,
1:05:42 we could actually force our way from something that, you know, is open source, open weights,
1:05:48 anybody can build it.
1:05:49 We’ll have a plethora of this technology will be like, use all of American innovation
1:05:55 to compete or we’ll, you know, we’ll cut it all off, we’ll force it into the hands
1:06:02 of the companies that kind of own the internet today and, you know, and we’ll put ourselves
1:06:08 at a huge disadvantage, I think, competitively against China in particular, but, but everybody
1:06:14 in the world.
1:06:15 And so, so I think that’s, that’s something that definitely, you know, that we’re involved
1:06:20 with trying to make sure it doesn’t happen, but it’s a real possibility right now.
1:06:24 Yeah.
1:06:25 There’s sort of an irony is that networks used to be all proprietary and then they opened
1:06:29 up.
1:06:30 Yeah, yeah, yeah.
1:06:31 Landman, Apple Talk, Net Buoy, Net Bios.
1:06:34 Yeah, exactly.
1:06:35 And so these are all the early proprietary networks from all individual specific vendors
1:06:38 and then the internet appeared and kind of TCP/IP and everything opened up.
1:06:41 The AI is trying to go the other, I mean, the big company is trying to take AI the other
1:06:44 way.
1:06:45 It started out as like open, just like basically just like the resource.
1:06:48 Everything was open source and AI, yeah.
1:06:50 Right.
1:06:51 Right.
1:06:52 And now they’re trying to, they’re trying to lock it down.
1:06:53 So it’s a, it’s a, it’s a fairly nefarious turn of events.
1:06:56 Yeah.
1:06:57 Yeah.
1:06:58 Very nefarious.
1:06:59 You know, I can, it’s remarkable to me.
1:07:01 I mean, it is kind of the darkest side of capitalism when a company is so greedy, they’re
1:07:08 willing to destroy the country and maybe the world to like just get a little extra profit.
1:07:12 You know, and they do it like the, the, the really kind of nasty thing is they claim, oh,
1:07:17 it’s for safety.
1:07:18 You know, we’ve created an alien that we can’t control, but we’re not going to stop
1:07:22 working on it.
1:07:23 We’re going to keep building it as fast as we can and we’re going to buy every freaking
1:07:26 GPU on the planet, but we need the government to come in and stop it from being open.
1:07:32 This is literally the current position of Google and Microsoft right now.
1:07:37 It’s crazy.
1:07:38 And we’re not going to secure it.
1:07:40 So we’re going to make sure that like Chinese buys can just like steal our chip plans, take
1:07:44 them out of the country.
1:07:45 We won’t even realize for six months.
1:07:46 Yeah.
1:07:47 It has nothing to do with security.
1:07:48 It only has to do with monopoly.
1:07:49 Yes.
1:07:50 The other, you know, just been going back on your point of speculation, so there’s this
1:07:54 critique that we hear a lot, right?
1:07:56 Which is like, okay, you idiots, basically it’s like you idiots, you idiots, entrepreneurs,
1:07:59 investors, you idiots.
1:08:00 It’s like there’s a speculative bubble with every new technology, like basically like
1:08:04 when are, when are you people going to learn to not do that?
1:08:06 Yeah.
1:08:07 There’s an old joke, there’s an old joke that relates to this, which is the foremost
1:08:10 dangerous words in investing are, this time is different.
1:08:13 The 12 most dangerous words in investing are the foremost dangerous words in investing
1:08:17 are this time is different, right?
1:08:18 Like, so like, does history repeat?
1:08:21 Does it not repeat the, my sense of it, and you referenced Carlotta Perez’s book, which
1:08:25 I agree is good.
1:08:26 Although I think, I don’t think it works as well anymore.
1:08:28 We can talk about some time, but, but you know, is a good, at least background piece
1:08:31 on this.
1:08:32 You know, it’s just like, it’s just incontrovertibly true, basically every significant technology
1:08:35 advance in history was greeted by some kind of financial bubble, basically since financial
1:08:39 markets have existed.
1:08:40 And this, you know, by the way, this includes like everything from, you know, radio and
1:08:43 television, the railroads, you know, lots and lots of prior, by the way, there was a,
1:08:47 there was actually a, a so-called, there was an electronics boom bust in the 60s called
1:08:51 the, it was called the Tronix, every, every company had the name Tronix.
1:08:55 And so, you know, there, there was that, so there, you know, there was like a laser boom
1:08:58 bust cycle.
1:08:59 There, there were all these like boom bust cycles.
1:09:00 And so basically it’s like any new tech, any new technology, that’s what economists
1:09:04 call a general purpose technology, which is to say something that can be used in lots
1:09:07 of different ways.
1:09:08 Like it inspires sort of a speculative mania.
1:09:10 And you know, and look, the critique is like, okay, why do you need to have a spec speculative
1:09:14 mania?
1:09:15 Why do you need to have a cycle?
1:09:16 Because like, you know, people, you know, if people, some people invest in the things
1:09:18 they lose a lot of money, and then there’s this bust cycle that, you know, causes everybody
1:09:21 to get depressed.
1:09:22 Maybe it delays the rollout.
1:09:23 And it’s like two things.
1:09:25 Number one is like, well, you just don’t know, like if it’s a general purpose technology,
1:09:28 like AI is, and it’s potentially useful in many ways, like nobody actually knows upfront
1:09:32 like what the successful use cases are going to be, or what successful companies are going
1:09:36 to be like you actually have to, you have to learn by doing.
1:09:38 You’re going to have to miss this.
1:09:39 That’s venture capital.
1:09:40 Yeah.
1:09:41 We, we…
1:09:42 Yeah, exactly.
1:09:43 Yeah, exactly.
1:09:44 So yeah, the true venture capital model kind of wires this in, right?
1:09:46 Yeah.
1:09:47 We, we basically, in core venture capital, the kind that we do, we sort of assume that half
1:09:50 the companies fail, half the projects fail.
1:09:52 And you know, if, if, if any of us, if we are any of our…
1:09:55 Tell a completely, like lose money, lose money.
1:09:57 Exactly.
1:09:58 Yeah.
1:09:59 And so like, and of course, if we are any of our competitors, you know, could figure
1:10:02 out how to do the 50% that work without doing the 50% that don’t work.
1:10:05 We would do that.
1:10:06 But you know, here we sit 60 years into the field and like nobody’s figured that out.
1:10:10 So there is, there is that, that unpredictability to it.
1:10:13 And then the other, the other kind of interesting way to think about this is like, okay, what
1:10:16 would it mean to have a society in which a new technology did not inspire speculation?
1:10:20 And it would mean having a society that basically is just like inherently like super pessimistic
1:10:25 about both the prospects of the new technology, but also the prospects of entrepreneurship
1:10:29 and you know, people inventing new things and doing new things.
1:10:31 And of course there are many societies like that on planet earth, you know, they’re just
1:10:35 like fundamentally like don’t have the spirit of invention and adventure that, you know,
1:10:40 that a place like Silicon Valley does and, you know, are they better off or worse off?
1:10:44 And, you know, generally speaking, they’re worse off.
1:10:46 They’re just, you know, less future oriented, less, less, less, less focused on, on building
1:10:50 things, less focused on, on figuring out how to get growth.
1:10:53 And so I think there’s a, at least my sense, there’s a comes with the territory thing.
1:10:57 Like we, we, we, we would all prefer to avoid the downside of a speculative boom bus cycle,
1:11:01 but like it seems to come with the territory every single time.
1:11:03 And I, at least I have not, nobody I’m aware, no society I’m aware of has ever figured out
1:11:07 how to capture the good without also having the bet.
1:11:09 Yeah.
1:11:10 And like, why would you?
1:11:11 I mean, it’s kind of like, you know, the, the, the whole Western United States was built
1:11:16 off the gold rush and like every kind of treatment in the like popular culture of the gold rush
1:11:23 kind of focuses on the people who didn’t make it anymore, but there were people who made
1:11:27 a lot of money, you know, and found gold and, you know, in the internet bubble, which, you
1:11:33 know, was completely ridiculed by, you know, kind of every, every movie.
1:11:38 If you go back and watch any movie between like 2001 and 2004, they’re all like how only
1:11:46 morons did.com and this and that and the other, and they’re all these funny documentaries
1:11:51 and so forth.
1:11:52 But like, that’s when Amazon got started, you know, that’s when eBay got started, that’s
1:11:58 when Google got started, you know, these, these companies, you know, with that work
1:12:03 started in the bubble in the kind of time of this great speculation, there was gold in
1:12:08 those companies.
1:12:10 And if you get any one of those, like you funded, you know, probably the next set of companies,
1:12:15 you know, which included things like, you know, Facebook and accent, you know, snap
1:12:19 and all these things.
1:12:20 And so, yeah, I mean, like, that’s just the nature of it.
1:12:24 I mean, like, that’s what makes it exciting.
1:12:26 And you know, it’s just a, it’s an amazing kind of thing that, you know, look, the transfer
1:12:33 of money from people who have excess money to people who are trying to do new things
1:12:39 and make the world a better place is the greatest thing in the world.
1:12:43 Like, and if we, some of the people with excess money lose some of that excess money and trying
1:12:49 to make the world a better place, like, why are you mad about that?
1:12:53 Like that, that’s the thing that I can ever, like, why would you be mad at, you know, young
1:12:59 ambitious people trying to improve the world, getting funded, and some of that being misguided?
1:13:06 Like why is that bad?
1:13:07 Right.
1:13:08 Right.
1:13:09 As compared to, yeah, as compared to, especially as compared to everything else in the world
1:13:12 and all the people who are not trying to do that.
1:13:14 So you’d rather, like, we just buy, like, you know, lots of mansions and boats and jets.
1:13:19 Right.
1:13:20 Like, what are you talking about?
1:13:21 Right.
1:13:22 Right.
1:13:23 Exactly.
1:13:24 We’re donating money to ruin us.
1:13:25 Yeah, ruin us causes.
1:13:26 Right.
1:13:27 Such as, ones that are on the news right now.
1:13:31 Okay.
1:13:32 So, all right.
1:13:33 We’re at a minute 20.
1:13:34 We made it all the way through four questions.
1:13:35 We’re doing good.
1:13:36 We’re doing great.
1:13:37 So let’s call it here.
1:13:38 Thank you, everybody, for joining us.
1:13:39 And I believe we should do a part two of this, if not parts three through six, because we
1:13:42 have a lot more questions to go.
1:13:43 But thanks, everybody, for joining us today.
1:13:45 All right.
1:13:46 Thank you.
1:13:46 Bye.
1:13:47 Bye.
1:13:48 Bye.
1:13:49 Bye.
1:13:49 (upbeat music)
1:13:51 (upbeat music)