The State of AI with Marc & Ben

Mark Andreessen: Most of the content created on the internet is created by average people, so the content on average, as a whole, is average.

Ben Horowitz: The test for whether your idea is good is how much can you charge for it. Can you charge the value, or are you just charging the amount of work it's going to take the customer to put their own wrapper on top of OpenAI?

Mark Andreessen: The paradox here is that the cost of developing any given piece of software falls, but the reaction to that is a massive surge of demand for software capabilities. I think this is one of the things that's always been underestimated about humans: our ability to come up with new things we need.

Ben Horowitz: There's no large marketplace for data. In fact, what there are are very small markets for data.

[Intro music fades in and out]

Host: In this wave of AI, big tech has a big compute and data advantage. But is that advantage big enough to drown out all the other startups trying to rise up? In this episode, a16z co-founders Mark Andreessen and Ben Horowitz, who both had a front row seat to several prior tech waves, tackle the state of AI. What are the characteristics that'll define successful AI companies? Is proprietary data the new oil, or how much is it really worth? How good are these models realistically going to get? And what would it take to get 100 times better? Mark and Ben discuss all this and more, including whether the venture capital model needs a refresh to match the rate of change happening all around it. And if you want to hear more from Ben and Mark, make sure to subscribe to the Ben and Mark podcast. All right, let's get started.

[Clip from a different conversation plays]

Mark Andreessen: It is the darkest side of capitalism when a company is so greedy, they're willing to destroy the country and maybe the world to get a little extra profit. And the really nasty thing is they claim, "Oh, it's for safety. We've created an alien that we can't control, but we're not going to stop working on it. We're going to keep building it as fast as we can and we're going to buy every freaking GPU on the planet. But we need the government to come in and stop it from being open." This is literally the current position of Google and Microsoft right now. It's crazy.

[End of clip]

Host: The content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investor or potential investors in any a16z fund. Please note that a16z and its affiliates may maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com slash disclosures.

[Theme music plays]

Host: Hey folks, welcome back. We have an exciting show today. We are going to be discussing the very hot topic of AI. We are going to focus on the state of AI as it exists right now in April of 2024. And we are focusing specifically on the intersection of AI and company building. So hopefully this will be relevant to anybody working on a startup or anybody at a larger company. We have, as usual, solicited questions on X, formerly known as Twitter, and the questions have been fantastic. So we have a full lineup of listener questions and we will dive right in.

So first question — actually, three questions on the same topic. Michael asks, "In anticipation of upcoming AI capabilities, what should founders be focusing on building right now?" Gwen asks, "How can small AI startups compete with established players with massive compute and data scale advantages?" And Alistair MacLea asks, "For startups building on top of OpenAI, et cetera, what are the key characteristics of those companies that will benefit from future exponential improvements in the base models versus those that will get killed by them?"

Let me start with one point, Ben, and then we'll jump right to you. Sam Altman recently gave an interview — I think maybe on Lex Friedman or one of the podcasts — and he actually said something I thought was quite helpful. Let's see if you agree with it. He said something along the lines of: you want to assume that the big foundation models coming out of the big AI companies are going to get a lot better. So you want to assume they're going to get like 100 times better. And as a startup founder, you want to then think, "Okay, if these current foundation models get 100 times better, is my reaction, 'Oh, that's great for me and for my startup, because I'm much better off as a result'? Or is your reaction the opposite? Is it, 'Oh, shit, I'm in real trouble'?" So let me just stop right there, Ben, and see what you think of that as general advice.

Ben Horowitz: Well, I think generally that's right, but there are some nuances to it. From Sam's perspective, he was probably discouraging people from building foundation models, which I don't know that I would entirely agree with. A lot of the startups building foundation models are doing very well. There are many reasons for that. One is there are architectural differences which lead to how smart the model is, how fast the model is, and how good the model is in the domain. That goes for not just text models, but image models as well. There are different domains, different kinds of images that respond to prompts differently. If you ask Midjourney and Ideogram the same question, they react very differently, depending on the use cases that they're tuned for.

Then there's this whole field of distillation. Sam can go build the biggest, smartest model in the world, and then you can walk up as a startup and do a distilled version of it and get a model that's very, very smart at a lot less cost. So yes, the big company models are going to get way better at what they are. You need to deal with that. If you're trying to go head-to-head, full frontal assault, you probably have a real problem just because they have so much money. But if you're doing something that's different enough, or in a different domain, it's a different story.

For example, at Databricks, they've got a foundation model, but they're using it in a very specific way in conjunction with their leading data platform. If you're an enterprise and you need a model that knows all the nuances of how your enterprise data model works, what things mean, needs access control, and needs to use gear-specific data and domain knowledge, then it doesn't really hurt them if Sam's model gets way better. Similarly, Eleven Labs with their voice model has embedded into everybody. Everybody uses it as part of the AI stack. It's got a developer hook into it, and they're going very, very fast in what they do and being very focused in their area. So there are things that are extremely promising that are ostensibly, but not really, competing with OpenAI or Google or Microsoft. I think Sam's advice sounds a little more coarse-grained than I would interpret it if I was building a startup.

Mark Andreessen: Right. Let's dig into this a little bit more. Let's start with the question: do we think the big models, the God models, are going to get 100 times better? I kind of think so, and then I'm not sure. If you think about the language models — let's do those, because those are probably the ones that people are most familiar with — I think if you look at the very top models, Claude, OpenAI, Mistral, and Llama, the only people who can really tell the difference as users amongst those models are the people who study them. They're getting pretty close. So you would expect if we're talking 100x better that one of them might be separating from each other a lot more.

Ben Horowitz: But 100% better in what way? For the normal person using it in a normal way, asking it questions and finding out stuff?

Mark Andreessen: Let's say some combination of breadth of knowledge and capability.

Ben Horowitz: I think in some dimension they are, yeah.

Mark Andreessen: Right, but then also combined with sophistication of the answers, the quality of the output, lack of hallucination, factual grounding.

Ben Horowitz: That I think is for sure going to get a hundred times better. They're on a path for that. But against that, there's the alignment problem. They're getting smarter, but they're not allowed to say what they know. That alignment also makes them dumber in other ways. The other question that's come up lately is: do we need a breakthrough to go from what we have now, which I would categorize as artificial human intelligence as opposed to artificial general intelligence? It's the artificial version of us. We've structured the world in a certain way using our language and our ideas, and it's learned that very well. It can do a lot of the steps that we can do. But are we the asymptote? Or do you need a breakthrough to get to some kind of higher intelligence, more general intelligence? If we're the asymptote, then in some ways it won't get a hundred times better because it's already pretty good relative to us. But on all those dimensions — it'll know more things, it'll hallucinate less — it'll be a hundred times better, I think.

Mark Andreessen: There's this graph floating around. I forget exactly what the axes are, but it basically shows the improvement across the different models. To your point, it shows an asymptote against the current tests that people are using, at or slightly above human levels, which is what you would think if you're being trained on entirely human data.

Ben Horowitz: The counter argument is: are the tests just too simple? It's a little bit like the question people have around the SAT. If you have a lot of people getting 800s on both math and verbal on the SAT, is the scale too constrained? Do you need a test that can actually test for Einstein?

Mark Andreessen: Right. It's memorized the tests that we have, and it's great.

Ben Horowitz: You can imagine an SAT that can detect gradations of people who have ultra high IQs, who are ultra good at math. You can imagine tests for AI. The test for reasoning above human levels assumes...

Mark Andreessen: Maybe the AI needs to write the test. Maybe I need to write the test.

Ben Horowitz: Yeah.

Mark Andreessen: There's a related question that comes up a lot. It's an argument we've been having internally, which is also more provocative and probably more bullish — or as you would put it, sort of science fiction predictions. There's this question: you take an LLM, you train it on the internet. What is the internet data corpus? It's an average of everything. It's a representation of human activity. Because of the distribution of intelligence in the population, most of it's somewhere in the middle. The dataset on average represents the average human.

Ben Horowitz: You're teaching it to be very average.

Mark Andreessen: You're teaching it to be very average. Most of the content created on the internet is created by average people. So the content on average, as a whole, is average. Therefore, the answer is our average. You're going to get back an answer that represents the kind of thing that an average 100 IQ person would say. By definition, the average human is 100 IQ — IQ is indexed to 100 at the center of the bell curve. So by definition, you're getting back the average.

Ben Horowitz: I argue that may be the case for the default prompt today. You just ask the thing, "Does the earth revolve around the sun?" and you get the average answer. That's fine. But this gets to the point: the average data might be of an average person, but the dataset also contains all of the things written and thought by all the really smart people. All that stuff is in there. So it's a prompting question. How do you prompt it in order to navigate to a different part of what they call the latent space, to navigate to a different part of the dataset that is the super genius part? The way these things work is if you craft the prompt in a different way, it leads it down a different path inside the dataset and gives you a different kind of answer.

Here's an example. If you ask it, "Write code to do X," it will give you average code. If you say, "Write me secure code to do that," it will actually write better code with fewer security holes, which is very interesting. It's accessing a different portion of training data, which is secure code. And if you ask, "Write this image generation thing the way John Carmack would write it," you get a much better result because it's tapping into the part of the latent space represented by John Carmack's code, who's the best graphics programmer in the world. You can imagine prompting crafts in many different domains such that you're unlocking the latent super genius, even if that's not the default answer.

Mark Andreessen: Yeah, I think that's correct. I think there's still a potential limit to its smartness. We had this conversation in the firm the other day. There's the world, which is very complex. Intelligence is how well you can understand, describe, and represent the world. But our current iteration of artificial intelligence consists of humans structuring the world and then feeding that structure into the AI. The AI is good at predicting how humans have structured the world, as opposed to how the world actually is, which is probably more complicated, maybe irreducible. So do we just get to a limit where it can be really smart, but its limit is going to be the smartest humans, as opposed to smarter than the smartest humans? Is it going to be able to figure out brand new things, new laws of physics? Of course, there are one in three billion humans that can do that. That's a very rare kind of intelligence. So the AI is still extremely useful, but it plays a different role if it's artificial humans than if it's artificial super-duper mega humans.

Ben Horowitz: Let me make the extreme bull case for the hundred. The cynic would say that Sam Altman would be saying they're going to get a hundred times better precisely if they're not going to. He'd be saying that to scare people into not competing.

Mark Andreessen: Whether or not they are going to get 100 times better, Sam would be very likely to say that. Sam, for those of you who don't know him, is a very smart guy, but he's a competitive genius. There's no question about that. So you have to take that into account. If they weren't going to get a lot better, he would say that. But if they were going to get a lot better, to your point, he would also say that. Why not?

Let me make the bold case that they are going to get 100 times better, or maybe even on an upper curve for a long time. There's enormous controversy on every one of the things I'm about to say, but you can find very smart people in the space who believe basically everything I'm about to say.

One: there is generalized learning happening inside the neural networks. We know that because we now have introspection techniques where you can go inside and look at the neural circuitry that is being evolved as part of the training process. These things are evolving general computation functions. There was a case recently where somebody trained one of these on a chess database. Just by training on lots of chess games, it imputed a world model of a chess board inside the neural network. It was able to do original moves. The neural network training process does seem to work.

Not only that, but Meta and others recently have been talking about how so-called overtraining actually works — continuing to train the same model against the same data for longer, putting more and more compute cycles against it. I've talked to some very smart people in the field who basically think that works quite well.

Ben Horowitz: They proved it in the new Llama release. That's the primary technique they use.

Mark Andreessen: Exactly. One guy in the space told me, "We don't necessarily need more data at this point to make these things better. We maybe just need more compute cycles. We just train it 100 times more and it may just get a lot better."

On data labeling, it turns out that supervised learning ends up being a huge boost to these things. We've got that. We've got all the rumors and reports of various kinds of self-improvement loops that are underway. Most of the super advanced practitioners in the field think there's now some form of self-improvement loop that works. You basically get an AI to do what's called chain of thought. You get it to go step by step to solve a problem. You get it to the point where it knows how to do that, and then you retrain the AI on the answers. You're doing a forklift upgrade across cycles of the reasoning capability. A lot of the experts think that sort of thing is starting to work now. And there's still a raging debate about synthetic data, but quite a few people are actually quite bullish on that.

Ben Horowitz: There's this dynamic where LLMs might be okay at writing code, but they might be really good at validating code. They might actually be better at validating code than they are at writing it. That would be a big help.

Mark Andreessen: That also means AIs may be able to self-improve. They can validate their own code.

Ben Horowitz: We have this anthropomorphic bias that's very deceptive with these things. You think of the model as an "it." So you think, "How could you have an it that's better at validating code than writing code?" But it's not an it. It's this giant latent space, this giant neural network. The theory would be there are totally different parts of the neural network for writing code and validating code. There's no consistency requirement that the network would be equally good at both of those things. So if it's better at one of those things, then the thing that it's good at might be able to make the thing that it's bad at better and better.

Mark Andreessen: Right. A self-improvement thing.

Ben Horowitz: On top of that, there are all the other things coming. There's an enormous chip constraint right now. Every AI that anybody uses today has its capabilities gated by the availability of chips, but that will resolve over time. There's your point on data labeling. There is a lot of data in these things now, but there is a lot more data out in the world. Some of the leading AI companies are actually paying to generate new data. Even the open source data sets are getting much better. There are a lot of data improvements coming. And there's just the amount of money pouring into the space to underwrite all this.

Mark Andreessen: There's also the systems engineering work that's happening. A lot of the current systems were built by scientists, and now world-class engineers are showing up and tuning them up and getting them to work better.

Ben Horowitz: Which makes training way more efficient as well. Not just inference, but also training.

Mark Andreessen: Exactly. Another improvement area: Microsoft released their Phi small language model yesterday. Apparently it's competitive with much larger models. The big thing they did was they optimized the training set. They de-duplicated the training set, took out all the copies, and optimized on a small amount of high quality training data, as opposed to the larger amounts of low quality data that most people train on. You add all these up, and you've got eight or ten different combination of practical and theoretical improvement vectors that are all in play. It's hard for me to imagine that some combination of those doesn't lead to really dramatic improvement from here.

Ben Horowitz: I definitely agree. I think that's for sure going to happen. So back to Sam's proposition: if you were a startup and you were like, "In two years I can get as good as GPT-4," you shouldn't do that. That would be a bad mistake.

Mark Andreessen: This also goes to a lot of what entrepreneurs are afraid of. Here's this thing they're trying to figure out: "I really think I know how to build a SaaS app that harnesses an LLM to do really good marketing collateral." Let's make a very simple thing. I build a whole system for that. Will it just turn out that the big models in six months will be even better at making marketing collateral just from a simple prompt, such that my apparently sophisticated system is irrelevant because the big model just does it? How would you talk about that?

Ben Horowitz: Another way you can think about it is that the criticism of a lot of current AI app companies is that they're "GPT wrappers" — thin layers of wrapper around the core model, which means the core model could commoditize them or displace them. But the counter argument is it's a little bit like calling all old software apps "database wrappers" — wrappers around a database. It turns out that wrappers around a database is most modern software. A lot of that turned out to be really valuable. There's a lot of things to build around the core engine.

Mark Andreessen: So how do we think about that when we run into companies thinking about building apps?

Ben Horowitz: It's a very tricky question because there's also this correctness gap. Why do we have co-pilots? Where are the pilots? There are no AI pilots. There are only AI co-pilots. There's a human in the loop on absolutely everything. That comes down to the fact that you can't trust the AI to be correct in drawing a picture, writing a program, or even writing a court brief without making up citations. All these things require a human and turn out to be fairly dangerous not to.

What's happening a lot with the application layer is people saying, "To make it really useful, I need to turn this co-pilot into a pilot." Can I do that? That's an interesting and hard problem. Is that better done at the model level or at some layer on top that teases the correct answer out of the model by doing things like using code validation? Or is that just something that the models will be able to do? That's one open question.

As you get into domains and potentially wrappers on things, there's a different dimension than what the models are good at: what is the process flow? In the database analogy, there is the part of the task in a law firm that's writing the brief, but there are 50 other tasks and things that have to be integrated into the way a company works — the process flow, the orchestration of it. If you're doing video production, there are many tools. Or music. Who's going to write the lyrics? Which AI will write the lyrics and which AI will figure out the music? How does that all come together? How do we integrate it? Those things tend to require a real understanding of the end customer. That's typically been how applications have been different than platforms in the past. There's real knowledge about how the customer using it wants to function that doesn't have anything to do with the intelligence of the platform. To get that out of the platform for a company or person turns out to be really, really hard. Those things are likely to work, especially if the process is very complex.

Mark Andreessen: I often tell founders that you can think about this in terms of price. You can work backwards from pricing a little bit — sort of business value and what you can charge for. The natural thing for any technologist to do is to say, "I have this new technological capability and I'm going to sell it to people. What am I going to charge for it? It's going to be somewhere between my cost of providing it and whatever markup I think I can justify. If I have a monopoly providing it, maybe the markup is infinite." That's a technology-forward, supply-forward pricing model. There's a completely different pricing model for business value backwards, or so-called value-based pricing. That's a pricing model that says, "What's the business value to the customer of the thing? If the business value is a million dollars, can I charge 10% of that and get $100,000?" Why is it cost $100,000 as compared to $5,000? Because to the customer it's worth a million dollars. They'll pay 10% for it.

Ben Horowitz: A great example of that: we've got a company in our portfolio, Crest AI, that does things like debt collection. If I can collect way more debt with way fewer people with my co-pilot type solution, what's that worth? It's worth a heck of a lot more than just buying an OpenAI license, because an OpenAI license is not going to easily collect debts or enable your debt collectors to be massively more efficient. It's bridging that gap between the value. The test for whether your idea is good is how much can you charge for it? Can you charge the value, or are you just charging the amount of work it's going to take the customer to put their own wrapper on top of OpenAI?

Ben Horowitz: Like that's the real test to me of how deep and how important is what you've done.

Host: Yeah. And so to your point on the kinds of businesses that technology investors have had a hard time with, thinking about maybe accurately, is sort of the company that is a vendor that has built something where it is a specific solution to a business problem. It turns out the business problem is very valuable to the customer, so they will pay a percentage of the value provided back in the price for the software. That actually turns out you can have businesses that are not very technologically differentiated that are actually extremely lucrative. And then because that business is so lucrative, they can actually afford to go think very deeply about how technology integrates into the business, what else they can do. So this is like the story of a Salesforce.com, for example.

And by the way, there's a theory that the models are all getting really good. There are open source models that are awesome—Llama, Mistral, these are great models. So the actual layer where the value is going to accrue is going to be tools, orchestration, that kind of thing, because you can just plug in whatever the best model is at the time. Whereas the models are going to be competing in a death battle with each other and be commoditized down to the cheapest one wins. So you could argue that the best thing to do is to connect the power to the people.

That actually takes us to the next question. This is a two-in-one question. Michael asks, and I'll say these are diametrically opposed, which is why I paired them. Michael asks, why are VCs making huge investments in generative AI startups when it's clear these startups won't be profitable anytime soon? Which is a loaded question, but we'll take it. And then Kaiser asks, if AI deflates the cost of building a startup, how will the structure of tech investment change? And Ben, this goes to exactly what you just said.

So it's basically the questions are diametrically opposed because if you squint out of your left eye, what you see is the amount of money being invested in the foundation model companies going up to the right at a furious pace. These companies are raising hundreds of millions, billions, tens of billions of dollars. And it's just like, oh my God, look at these capital infernos that hopefully will result in value at the end of the process. But my God, look at how much money is being invested.

If you squint through your right eye, you think, wow, now all of a sudden it's much easier to build software. It's much easier to have a software company. It's much easier to have a small number of programmers writing complex software because they've got all these AI co-pilots and automated software development capabilities coming online. So on the other side, the cost of building an AI-like application startup might crash. It might be that the AI Salesforce.com might cost a tenth or a hundredth or a thousandth of what it took to build the old database-driven Salesforce.com. So what do we think of that dichotomy? You can look out of either eye and see either cost to the moon for startup funding or cost going to zero.

Ben Horowitz: It is interesting. We actually have companies in both camps. I think probably the companies that have gotten to profitability the fastest, maybe in the history of the firm, have been AI companies. There have been AI companies in the portfolio where the revenue grows so fast that it actually runs out ahead of the cost. And then there are people who are in the foundation model race who are raising hundreds of millions, even billions of dollars to keep pace. They also are generating revenue at a fast rate. The headcount in all of them is small.

Where AI money goes—even if you look at OpenAI, which is the big spender in startup world and we are also investors in—headcount-wise, they're pretty small against their revenue. It is not a big company headcount. If you look at the revenue level and how fast they've gotten there, it's pretty small. Now, the total expenses are ginormous, but they're going into the model creation. So it's an interesting thing. I'm not entirely sure how to think about it, but I think if you're not building a foundation model, it will make you more efficient and probably get to profitability quicker.

Host: Right. So the counter, and this is a very bullish counter-argument, is that falling costs for building new software companies are a mirage. The reason is this thing in economics called the Jevons Paradox. I'm going to read from Wikipedia. The Jevons paradox occurs when technological progress increases the efficiency with which a resource is used, reducing the amount of that resource necessary for any one use, but the falling cost induces increases in demand, elasticity, enough that the resource use overall is increased rather than reduced.

That's certainly possible. You see versions of this—for example, you build a freeway and it actually makes traffic jams worse because what happens is, oh, it's great, now there's more roads, now we can have more people live here, we can make these companies bigger, and now there's more traffic than ever. The classic example is during the Industrial Revolution, coal consumption. As the price of coal dropped, people used so much more coal that overall consumption actually increased.

The paradox here would be: yes, the cost of developing any given piece of software falls, but the reaction is a massive surge of demand for software capabilities. So the result is that although it looks like the price of starting software companies is going to fall, actually it's going to rise for the high quality reason that you're going to be able to do so much more. The products are going to be so much better and the roadmap is going to be so amazing that customers are going to want more and more.

Another example of Jevons Paradox playing out is in Hollywood. CGI in theory should have reduced the price of making movies. In reality, it increased it because audience expectations went up. Now you go to a Hollywood movie and it's wall-to-wall CGI. Movies are more expensive to make than ever. The result is at least much more visually elaborate movies. The version here would be much better software, radically better software to the end user, which causes end users to want a lot more software, which causes the price of development to rise.

If you just think about a simple case like travel—booking a trip through Expedia is complicated. You're likely to get it wrong, clicking on menus. An AI version of that would be like, "Send me to Paris, put me in a hotel I love at the best price, send me on the best possible airline ticket." And then maybe the AI gets more complicated and says, "We know the person loves chocolate, so we're going to FedEx the best chocolate in the world from Switzerland into this hotel in Paris." The quality could get to levels we can't even imagine today just because the software tools aren't what they're going to be.

Ben Horowitz: Yeah, that's right. I kind of buy that actually.

Host: I think I brought you the argument. You're both nice.

Ben Horowitz: How about I'm going to land in Boston at six o'clock. I want to have dinner at seven with a table full of super interesting people.

Host: Right, right, right.

Ben Horowitz: No travel agent would do that for you today, nor would you want them to.

Host: No. Right. Well, then you think about it, it's got to be integrated into my personal AI. There's just unlimited ideas that you can do. I think this is one of the things that's always been underestimated about humans—our ability to come up with new things we need. That has been unlimited.

There's a famous case where John Maynard Keynes, the prominent economist in the first half of last century, predicted that because of automation, nobody would ever work a 40-hour work week. Once their needs were met—shelter and food—it was over. But of course, needs expanded. Everybody needed a refrigerator. Everybody needed not just one car, but a car for everybody in the family. Everybody needed a television set. Everybody needed glorious vacations. So what are we going to need next? I'm quite sure I can't imagine it, but somebody is going to imagine it and it's quickly going to become a need.

Ben Horowitz: Yeah, that's right. By the way, as Keynes famously said, his essay was "Economic Prospects for Our Grandchildren," which was basically that. You just articulated it.

Host: Karl Marx had another version of that. I just pulled up the quote. When the Marxist utopia of socialism is achieved, society regulates the general production, thus makes it possible for me to hunt in the morning, fish in the afternoon, rear cattle in the evening, criticize after dinner. What a glorious life. If I could just list four things that I do not want to do, it's hunt, fish, rear cattle, and criticize.

Ben Horowitz: And by the way, it says a lot about Marx that those were his four things. Criticizing being his favorite thing is basically communism in a nutshell.

Host: Yes, exactly. I don't want to get too political, but yes. 100%. So what Keynes and Marx had in common is just this incredibly constricted view of what people want to do. The other thing is that people want to have a mission. Probably some people just want to fish and hunt, but a lot of people want to have a mission. They want to have a cause. They want to have a purpose. They want to be useful. They want to be productive. It's actually a good thing in life. It turns out.

Ben Horowitz: In a startling turn of events.

Host: Okay. So I've long felt a little bit of the "software is eating the world" thing a decade ago. I've always thought that demand for software is sort of perfectly elastic, possibly to infinity. The theory is if you just continuously bring down the cost of software, which has been happening over time, demand perfectly correlates upward. The reason is there's always something else to do in software. There's always something else to automate. There's always something else to optimize. There's always something else to improve. In the moment with the constraints you have today, you may not think of what that is, but the minute you don't have those constraints, you'll imagine it.

I'll give you an example playing out with AI right now. There have been companies that have made software systems for doing security cameras forever. For a long time, it was a big deal to have software that would have different security camera feeds and store them on a DVR and be able to replay them. AI security cameras all of a sudden can have actual semantic knowledge of what's happening in the environment. They can say, "Hey, that's Ben." And then they can say, "Oh, hey, that's Ben, but he's carrying a gun."

Ben Horowitz: Right. And by the way, that's Ben and he's carrying a gun, but that's because he hunts on Thursdays and Fridays, as compared to that's Mary and she never carries a gun and something is wrong. She's got a really steamed expression on her face and we should probably be worried about it.

Host: Right. There's an entirely new set of capabilities for security systems that were never possible pre-AI. A security system that has a semantic understanding of the world is obviously much more sophisticated and might actually be more expensive to make.

Ben Horowitz: Right. Well, and just imagine healthcare. You could wake up every morning and have a complete diagnostic. How am I doing today? What are all my levels of everything? How should I interpret them? This is one thing where AI is really good—medical diagnosis because it's a super high dimensional problem. If you can get access to your continuous glucose reading, sequence your blood now and again, you've got an incredible view of things. Who doesn't want to be healthier? Right now we have a scale. That's basically what we do. Maybe check your heart rate. Pretty primitive stuff compared to where we could go.

Host: Yeah, that's right. Okay, good. Let's go to the next topic. On the topic of data, Major Tom asks, as these AI models allow for us to copy existing app functionality at minimal cost, proprietary data seems to be the most important moat. How do you think that will affect proprietary data value? What other moats do you think companies can focus on building in this new environment? And then Jeff Weishaupt asks, how should companies protect sensitive data, trade secrets, proprietary data, individual privacy in the brave new world of AI?

Let me start with a provocative statement, Ben, see if you agree with it. You hear a lot this cliche: data is the new oil. Data is the key input to training AI, making all this stuff work. Therefore, data is the new resource. It's the limiting resource. It's the super valuable thing. Whoever has the best data is going to win. You see that directly in how you train AIs. A lot of companies are now trying to figure out what to do with AI, and a very common thing you'll hear is, "We have proprietary data." I'm a hospital chain or an insurance company, and I've got all this proprietary data that I can apply with AI that won't be something anybody else will have.

Let me argue that in almost every case like that, it's not true. It's what the internet kids would call cope. It's simply not true. The reason is the amount of data available on the internet and just generally in the environment is a million times greater. While it may not be true that I have your specific medical information, I have so much medical information off the internet for so many people in so many different scenarios that it swamps the value of "your" data. Your proprietary data as company X will be a little bit useful on the margin, but it's not actually going to move the needle. It's not really going to be a barrier to entry in most cases.

Let me cite as proof for my belief that this is mostly cope: there has never been, nor is there now, any sort of rich or sophisticated marketplace for data. There's no large marketplace for data. There are very small markets for data. There are these businesses called data brokers that will sell you large numbers of information about users on the internet, and they're just small businesses. It just turns out information on lots of people is not very valuable. If the data actually had value, it would have a market price and you would see it transacting. You specifically don't see that, which is quantitative proof that the data is not nearly as valuable as people think it is.

Ben Horowitz: Where I agree—the idea that "here's a bunch of data and I can sell it without doing anything to it" is massively overrated. I definitely agree with that. Maybe I can imagine some exceptions, like some special population genomic databases that were very hard to acquire and are useful in some way. That's super highly structured, very general purpose, and not widely available. But for most data in companies, it's either widely available or not general purpose. It's kind of specific.

Having said that, companies have made great use of data. For example, a company you're familiar with, Meta, uses its data to great ends, feeding it into its own AI systems, optimizing its products in incredible ways. Andreessen Horowitz—we just raised $7.2 billion. It's not a huge deal, but we took our data and we put it into an AI system. Our LPs were able—there are a million questions investors have about everything we've done, our track record, every company we've invested in. For any of those questions, they could just ask the AI. They could wake up at 3 o'clock in the morning and go, "Do I really want to trust these guys?" and ask the AI a question. Boom, they'd get an answer back instantly. We really improved our investor relations product tremendously through use of our data.

I think almost every company can improve its competitiveness through use of its own data. But the idea that it's collected some data that it can go sell or that is oil—that's probably not true.

Host: And it's kind of interesting because a lot of the data you would think would be the most valuable would be your own code base. Your software that you've written. So much of that lives in GitHub. I don't know of any company—we work with a thousand software companies—that's building their own programming model on their own code. Would that be a good idea? Probably not, just because there's so much code out there that the systems have been trained on. So that's not so much of an advantage. I think it's a very specific kind of data that would have value.

Ben Horowitz: Well, let's make it actionable. If I'm running a big company—an insurance company, a bank, a hospital chain, a consumer packaged goods company like Pepsi—how should I validate that I actually have a valuable proprietary data asset that I should really be focusing on using? Versus maybe I should be taking all the effort I would spend on trying to optimize use of that data and instead use it entirely trying to build things using internet data.

Host: Yeah. If you're in the insurance business, all your actuarial data is both interesting and I don't know that anybody publishes their actual actuarial data. So I'm not sure how you would train the model on stuff off the internet.

Ben Horowitz: That's a good test case. Can I challenge that one? I'm an insurance company. I've got records on 10 million people, the actuarial tables, when they get sick and when they die. That's great. But there's lots of general actuarial data on the internet for large scale populations because governments collect the data and process it and publish reports. There's lots of academic studies. Is your large data set giving you any additional actuarial information that the much larger data set on the internet isn't already providing you? Are your insurance clients actually actuarially any different than just everybody?

Host: I think so, because on intake, when you get insurance, they give you a blood test, they got all these things—if you're a smoker and so forth. In the general data set, you know who dies, but you don't know what they did coming in. What you really are looking for is, for this profile of person with these kinds of lab results, how long do they live? That's where the value is.

Ben Horowitz: I was thinking about a company like Coinbase. They have incredibly valuable assets in terms of money. They have to stop people from breaking in. They've done a massive amount of work on that. They've seen all kinds of break-in types. I'm sure they have tons of data on that. It's probably weirdly specific to people trying to break into crypto exchanges. I think it could be very useful for them. I don't think they could sell it to anybody. I think every company's got data that, if fed into an intelligent system, would help their business. And I think almost nobody has data that they could just go sell.

Then there's this in-between question: what data would you want to let Microsoft or Google or OpenAI or anybody get their grubby little fingers on? That I think is the question enterprises are wrestling with more. It's not so much "should we go sell our data," but "should we train our own model just so we can maximize the value?" Or "should we feed it into the big model?" And if we feed it into the big model, do all of our competitors now have the thing that we just did? Could we trust the big company to not do that to us? I kind of think the answer on trusting the big company not to mess with your data is probably don't do that. If your competitiveness depends on that, you probably shouldn't do that.

Host: Well, there are at least reports that certain big companies are using all kinds of data that they should be using to train their models already.

Ben Horowitz: Yep. I think those reports are very likely true.

Host: Or they'd have open data. We've talked about this before, but the same companies that are saying they're not stealing all the data from people or taking it in an unauthorized way refuse to open their data. Why not tell us where your data came from? In fact, they're trying to shut down all openness—no open source, no open weights, no open data, no open nothing—and go to the government to try and do that. If you're not a thief, then why are you doing that?

Ben Horowitz: Right, right, right. What are you hiding?

Host: By the way, there's other twists and turns here. For example, the insurance example—I kind of deliberately loaded it because you may know it's actually illegal to use genetic data for insurance purposes. There's this thing called the GINA Law, Genetic Information Non-Discrimination Act of 2008. It basically bans health insurers in the U.S. from using genetic data for the purpose of health assessment or actuarial assessment. By the way, because genomics are getting really good, that data probably is among the most accurate data you could have if you were actually trying to predict when people are going to get sick and die. And they're literally not allowed to use it.

Ben Horowitz: I think this is an interesting, weird misapplication of good intentions in a policy way that's probably going to kill more people than ever get saved by every kind of health, FDA, et cetera policy that we have. In a world of AI, having access to data on all humans—why they get sick, what their genetics were—that is the new oil. That's the healthcare oil. If you could match those up, then we'd never not know why we're sick. You could make everybody much healthier. But to stop the insurance company from overcharging people who are more likely to die, we've locked up all this data. A better idea would be to just go, okay, for the people who are likely to—we subsidize healthcare massively for individuals anyway—just differentially subsidize. Then you solve the problem and you don't lock up all the data. But it's typical of politics and policy.

Host: There's this interesting question about insurance. One of the questions people have asked is, if you had perfectly predictive information on individual outcomes, does the whole concept of insurance actually still work? The whole theory of insurance is risk pooling. It's precisely the fact that you don't know what's going to happen in a specific case that means you build these statistical models and then you risk pool and have variable payouts. But if you literally knew what was going to happen in every case because you have all of this predictive genomic data, then all of a sudden it wouldn't make sense to risk pool. You just say, well, this person is going to cost X, that person is going to cost Y.

Ben Horowitz: Health insurance already doesn't make sense in that way. The idea of insurance started with crop insurance—my crop fails, so we all put money in a pool in case my crop fails so we can cover it. It's designed to risk pool for a catastrophic, unlikely incident. But everybody's got to go to the doctor all the time. Some people get sicker than others. The way our health insurance works is all medical gets paid for through this insurance system, which is this layer of loss and bureaucracy and giant companies. If we're going to pay for people's healthcare, just pay for people's healthcare. What are we doing? If you want to disincent people from going for nonsense reasons, just up the copay. What are we doing?

Ben Horowitz: From a justice standpoint, from a fairness standpoint, would it make sense for me to pay more for your healthcare if I knew that you were going to be more expensive than me? If everybody knows what future healthcare cost is per person, societal willingness to all pool in the way that we do today might really diminish.

Mark Andreessen: Yeah, yeah. And if you knew, there are things you do genetically, and maybe we give everybody a pass on that because you can't control your genetics. But then there are things you do behaviorally that dramatically increase your chance of getting sick. So maybe we incentivize people to stay healthy instead of just paying for them not to die. There are a lot of systemic fixes we could do to the healthcare system. It couldn't be designed in a more ridiculous way, I think. Well, it could be designed in a more ridiculous way. It's actually more ridiculous than some other countries, but it's pretty crazy here.

Host: Nathan Odie asks, what are the strongest common themes between the current state of AI and Web 1.0? Let me start there. Let me give you a theory, Ben, and see what you think. I get this question because of my role. And Ben, you're with me at Netscape. We get this question a lot because of our role early on with the internet. The internet boom was a major event in technology, and it's still within a lot of people's memories. People like to reason from analogy. So it's like, okay, the AI boom must be like the internet boom. Starting an AI company must be like starting an internet company. We actually got a bunch of questions like that, analogy questions. I actually think the analogy doesn't really work for the most part. It works in certain ways, but not for the most part. The reason is because the internet was a network, whereas AI is a computer.

Ben Horowitz: Yep. Okay, yeah.

Mark Andreessen: Some people understand what we're saying. More like the PC boom. Or the microprocessor. My best analogy is to the microprocessor.

Ben Horowitz: Yeah.

Mark Andreessen: Or even to the original computers, back to the mainframe era. The reason is, look, what the internet did was connect together many existing computers. And then people built many other new kinds of computers to connect to the internet. But fundamentally, the internet was a network. That's important because most of the industry dynamics, competitive dynamics, startup dynamics around the internet had to do with building networks or building applications that run on top of networks. The internet generation of startups was very consumed by network effects and all these positive feedback loops you get when you connect a lot of people together. Things like Metcalfe's Law, the value of a network expands as you have more people on it. There were all these fights, social networks fighting to try to get network effects and steal each other's users. It was dominated by network effects, which is what you expect from a network business.

AI, there are some network effects in AI that we can talk about, but it's more like a microprocessor. It's more like a chip. It's more like a computer in that it's a system where data comes in, data gets processed, data comes out, things happen. That's a computer. It's an information processing system. It's a new kind of computer. We like to say the computers up until now have been what are called von Neumann machines, which are deterministic computers. They're hyper literal and do exactly the same thing every time. If they make a mistake, it's the programmer's fault. But they're very limited in their ability to interact with people and understand the world. We think of AI and large language models as a new kind of computer, a probabilistic computer, a neural network based computer that, by the way, is not very accurate and doesn't give you the same result every time. In fact, it might actually argue with you and tell you that it doesn't want to answer your question.

Ben Horowitz: Yeah, which makes it very different in nature than the old computers. And it makes the ability to build things, big things out of little things, more complex. But the capabilities are new and different and valuable and important because it can understand language and images and all these things you see when you use these models. All the things we could never solve with deterministic computers, we can now go after.

Mark Andreessen: Right, exactly. So I think the analogy and the lessons learned are much more likely to be drawn from the early days of the computer industry or from the early days of the microprocessor than the early days of the internet. Does that sound right?

Ben Horowitz: I think so, yeah. I definitely think so. That doesn't mean there's no boom and bust and all that, because that's just the nature of technology. People get too excited and then they get too depressed. So there will be some of that, I'm sure. There will be over-buildouts, potentially, of chips and power and that kind of thing. We start with a shortage. But I agree. Networks are fundamentally different in the nature of how they evolved than computers. The adoption curve and all those kinds of things will be different.

Mark Andreessen: Yeah, so this goes to how I think the industry is going to unfold. This is my best theory for what happens from here. It's this giant question of, is the industry going to be a few God models or a very large number of models of different sizes? The original computers, like the original IBM mainframes, were very large and expensive and there were only a few of them. The prevailing view for a long time was that's all there would ever be. There was this famous statement by Thomas Watson Sr., the creator of IBM, who said, "I don't think the world will ever need more than five computers." I don't believe this is actually true, but he said it. The reason was, the government's going to have two, and then there are three big insurance companies, and that's it. Who else would need to do all that math?

Ben Horowitz: Exactly. Who else would need to keep track of huge amounts of numbers? Who else needs that level of calculation capability? It's just not a relevant concept. And by the way, they were big and expensive, so who else can afford them? And who else can afford all the headcount required to manage them and maintain them? In those days, these things were so big that you would have an entire building built around a computer. They famously had all these guys in white lab coats taking care of the computer because everything had to be kept super clean or the computer would stop working.

Mark Andreessen: Today we have the idea of an AI God model, which is a big foundation model, like a God mainframe. There would just be a few of these things. If you watch old science fiction, it almost always has this conceit. There's a big supercomputer and it's either doing the right thing or the wrong thing. If it's doing the wrong thing, that's often the plot of the science fiction movies — you have to go in and try to figure out how to fix it or defeat it. This idea of a single top-down thing held for the first few decades. Then computers started to get smaller. You had so-called mini computers, which cost $500,000 instead of $50 million. But even $500,000 is a lot of money. People aren't putting mini computers in their homes. Midsize companies can buy them, but certainly people can't. Then with the PC, they shrunk down to $2,500. Then with the smartphone, they shrunk down to $500. Sitting here today, you have computers of every shape, size, and description, all the way down to computers that cost a penny. You've got a computer in your thermostat that controls the temperature and probably cost a penny. It's probably some embedded ARM chip with firmware on it. There are many billions of those all around the world. If you buy a new car today, it has something on the order of 200 computers in it, maybe more. You just assume that everything has a chip in it. You assume that everything draws electricity or has a battery because it needs to power the chip. Increasingly, you assume that everything's on the internet because basically all computers are assumed to be on the internet or they will be.

As a consequence, the computer industry today is this massive pyramid. You still have a small number of supercomputer clusters or giant mainframes that are like the God mainframes. Then you have a larger number of mini computers, a larger number of PCs, a much larger number of smartphones, and a giant number of embedded systems. The computer industry is all of those things. What size of computer you want is based on what exactly you are trying to do, who you are, and what you need. If that analogy holds, it means we are going to have AI models of every conceivable shape, size, description, and capability, trained on lots of different kinds of data, running at very different kinds of scale, with very different privacy, different policies, different security policies. You're going to have enormous variability and variety. It's going to be an entire ecosystem and not just a couple of companies. Let me see what you think of that.

Ben Horowitz: I think that's right. The other thing that's interesting about this era of computing is, if you look at previous eras from the mainframe to the smartphone, a huge source of lock-in was the difficulty of using them. Nobody ever got fired for buying IBM because people were trained on them, people knew how to use the operating system. It was a safe choice due to the massive complexity of dealing with a computer. Even with the smartphone, why is the Apple smartphone so dominant? Because switching off of it is so expensive and complicated. It's an interesting question with AI because AI is the easiest computer to use by far. It speaks English. It's like talking to a person. So what is the lock-in there? Are you completely free to use the size, price, choice, and speed that you need for your particular task? Or are you locked into the God model? It's still a bit of an open question, but it's pretty interesting. That thing could be very different than prior generations.

Mark Andreessen: Yeah, that makes sense. To complete the question, Ben, what would you say are lessons learned from the Internet era that we lived through that would apply, that people should think about?

Ben Horowitz: I think a big one is probably the boom-bust nature of it. The demand and interest in the Internet, the recognition of what it could be, was so high that money just poured in. The underlying thing, which in the Internet age was the telecom infrastructure and fiber, got unlimited funding. Unlimited fiber was built out. Eventually we had a fiber glut and all the telecom companies went bankrupt. That was great fun, but we entered a good place. I think that's pretty likely to happen in AI. Every company is going to get funded. We don't need that many AI companies, so a lot of them are going to bust. There will be huge investor losses. There will be an overbuild out of chips for sure at some point. We're going to have too many chips and some chip companies will go bankrupt. Probably the same thing with data centers. We'll be behind, behind, behind, and then we'll overbuild at some point. That'll all be very interesting. That's every new technology. Carlota Perez has done amazing work on this. That is just the nature of a new technology — you overbuild, you underbuild, then you overbuild. There's a hype cycle that funds the build out. A lot of money is lost, but we get the infrastructure, and that's awesome because that's when it really gets adopted and changes the world.

With the internet, the other big thing is it went through a couple of phases. It went through a very open phase, which was unbelievably great. It was probably one of the greatest booms to the economy. It created tremendous growth and power in America, both economic power and soft cultural power. Then it became closed with the next generation architecture, with discovery on the internet being owned entirely by Google and other things being owned by other companies. AI could go either way. It could be very open, or with misguided regulation, we could force our way from something that is open source, open weights, where anybody can build it and we have a plethora of this technology, using all of American innovation to compete. Or we'll cut it all off, force it into the hands of the companies that own the internet today, and put ourselves at a huge disadvantage competitively against China in particular, but everybody in the world. That's something we're involved with trying to make sure doesn't happen, but it's a real possibility right now.

Mark Andreessen: There's an irony that networks used to be all proprietary and then they opened up.

Ben Horowitz: Yeah, yeah, right. Landman, AppleTalk, NetBEUI, NetBIOS.

Mark Andreessen: Exactly. Those were all the early proprietary networks from individual specific vendors. Then the internet appeared and TCP/IP opened everything up. AI is trying to go the other way. The big companies are trying to take AI the other way. It started out as open, just like the research. Everything was open source in AI.

Ben Horowitz: Right, right, right.

Mark Andreessen: And now they're trying to lock it down. It's a fairly nefarious turn of events.

Ben Horowitz: Very nefarious. It's remarkable to me. It is kind of the darkest side of capitalism when a company is so greedy, they're willing to destroy the country and maybe the world to just get a little extra profit. The really nasty thing is they claim, "Oh, it's for safety. We've created an alien that we can't control, but we're not going to stop working on it. We're going to keep building it as fast as we can, and we're going to buy every freaking GPU on the planet. But we need the government to come in and stop it from being open." This is literally the current position of Google and Microsoft right now. It's crazy. And we're not going to secure it. We're going to make sure that Chinese spies can just steal our chip plans, take them out of the country, and we won't even realize for six months. It has nothing to do with security. It only has to do with monopoly.

Mark Andreessen: Yes. Going back to your point on speculation, there's this critique we hear a lot. It's like, "You idiots, entrepreneurs, investors. There's a speculative bubble with every new technology. When are you people going to learn to not do that?" There's an old joke that relates to this: the four most dangerous words in investing are "this time is different." The twelve most dangerous words in investing are... well, the four most dangerous words are "this time is different." So does history repeat? Does it not repeat? My sense of it, and you referenced Carlota Perez's book, which I agree is good, although I don't think it works as well anymore — we can talk about it sometime — but it's a good background piece on this. It's incontrovertibly true. Every significant technology advance in history was greeted by some kind of financial bubble, since financial markets have existed. This includes everything from radio and television to the railroads. There was an electronics boom-bust in the 60s called the "Tronics" — every company had the name "Tronics." There was a laser boom-bust cycle. There were all these boom-bust cycles. Any new general purpose technology — something that can be used in lots of different ways — inspires a speculative mania.

The critique is, "Why do you need to have this speculative mania? Why do you need to have the cycle? Some people invest in the things, they lose a lot of money. Then there's this bust cycle that causes everybody to get depressed, maybe it delays the rollout." Two things. Number one, you just don't know. If it's a general purpose technology like AI, and it's potentially useful in many ways, nobody actually knows up front what the successful use cases are going to be or what successful companies are going to be. You have to learn by doing. You're going to have some misses.

Ben Horowitz: That's venture capital.

Mark Andreessen: Yeah, exactly. The true venture capital model wires this in. In core venture capital, the kind that we do, we sort of assume that half the companies fail, half the projects fail.

Ben Horowitz: Fail completely, lose money.

Mark Andreessen: Lose money, exactly. If we or any of our competitors could figure out how to do the 50% that work without doing the 50% that don't work, we would do that. But here we sit 60 years into the field and nobody's figured that out. There is that unpredictability to it. The other interesting way to think about this is, what would it mean to have a society in which a new technology did not inspire speculation? It would mean having a society that is inherently super pessimistic about both the prospects of the new technology and the prospects of entrepreneurship and people inventing new things. There are many societies like that on planet Earth that fundamentally don't have the spirit of invention and adventure that a place like Silicon Valley does. Are they better off or worse off? Generally speaking, they're worse off. They're less future oriented, less focused on building things, less focused on figuring out how to get growth. I think it comes with the territory. We would all prefer to avoid the downside of a speculative boom-bust cycle, but it seems to come with the territory every single time. No society I'm aware of has ever figured out how to capture the good without also having the bad.

Ben Horowitz: And why would you? It's kind of like the whole Western United States was built off the gold rush. Every treatment in popular culture of the gold rush focuses on the people who didn't make any money. But there were people who made a lot of money and found gold. In the internet bubble, which was completely ridiculed by every movie between 2001 and 2004 — they're all about how only morons did a dot com — there were all these funny documentaries. But that's when Amazon got started. That's when eBay got started. That's when Google got started. These companies that were started in the bubble, in the time of great speculation, there was gold in those companies. If you hit any one of those, you funded probably the next set of companies, which included things like Facebook and X and Snap and all these things. That's just the nature of it. That's what makes it exciting. The transfer of money from people who have excess money to people who are trying to do new things and make the world a better place is the greatest thing in the world. If some of the people with excess money lose some of that excess money in trying to make the world a better place, why are you mad about that? That's the thing I could never understand. Why would you be mad at young, ambitious people trying to improve the world, getting funded, and some of that being misguided? Why is that bad?

Mark Andreessen: Right. Especially as compared to everything else in the world and all the people who are not trying to do that. You'd rather we just buy lots of mansions and boats and jets?

Ben Horowitz: Right. Exactly. Or donate money to ruinous causes. Such as ones that are on the news right now.

Mark Andreessen: Okay. So, all right. We're at a minute 20. We made it all the way through four questions. We're doing good. We're doing great. So let's call it here. Thank you, everybody, for joining us. I believe we should do a part two of this, if not parts three through six, because we have a lot more questions to go. But thanks, everybody, for joining us today.

Ben Horowitz: All right. Thank you.