Reid Hoffman on Foundation Models: Who Wins & How Do Incumbents Respond

Some number of you like, it’s that distant before pandemic prehistory time. Like, I was young then. Yeah. Yeah. Well, young now

answer is yes. By the way, what I have most wanted to aspire to be is a public intellectual, which is how do you speak the truth about who we are and who we should be as individuals in a society? So I aspire to that in every aspect of my life. Now that being said, I do of course have commitments to the board of Microsoft, so the things that I can’t talk about. Sure. But as opposed to like what you can say objective is like, I just say I can’t talk about it versus giving a false answer or a misleading answer.

And then the second is, obviously, I learn a bunch from the world from that. Like, there’s a perspective that I will learn from the intensity of Microsoft that may give me a blue colored lens as I’m looking at something, but I myself aspire to be a public intellectual, to speak truth so that we collectively become better. Why do you think you aspire to be

but No. No. No. Well, that that was that was the goal. I get great joy from discussion where we are discovering truths together. Teaching is part of that, right? Some teaching is we’re in a seminar and we’re talking, we go, ah, we now understand some important aspect of the world better, whether it’s about ourselves, whether it’s about the world, and that’s why I think, to some degree, one of the fundamental parts of the meaning of life is, which is how do we become wiser, more intelligent, more compassionate, etcetera, but that’s because of the discovery of truths about these things that we share on this this journey that we call life together.

make

Both. On one sense, it’s a great companion catalyst in these things. It’s just like, for example, don’t proxy your judgment of truth to what you happen to have found in a search engine. It can inform your perception of truth. Google told me red wine was fantastic for me. Yes, exactly. And it’s similar, like it could be bad in that regard, because you might say, Well, the AI said x. And you’re like,

I would, to some degree, hope so. I mean, one of the things that I think that the kind of the libertarian kickoff of the Internet has underplayed is that we need to have shared media be how we learn together. And so some of that is you need to say, it isn’t just propositions in the wild that were us, a thing is said, that you can, like for example, get better from COVID by taking hydrochloroquine. It’s like, no, no, no. Just because that proposition is said and said by the then president of The United States doesn’t mean that it has any scientific validity and isn’t actually in fact destructive to your health.

We need to have better sources of how we’re collectively learning. Like, example, when we as human beings try to figure out how to get to truth, we use panels of human beings to do it. It’s like a scientific panel observing a study or deciding on a paper. It’s a blue ribbon commission and trying to figure out what the facts are in a public circumstance. It’s a jury in a criminal trial. We bring a group of people together, and that brings a set of balanced perspectives. We need to have that kind of thing, and we need to have the brands and institutions that reflect that kind of collective approach to discernment of truth so that we learn together.

we can in various ways. Like, we can amplify that way. So for example, you know, questions around what’s someone’s professional CV? Well, LinkedIn is working its way towards being a higher and higher source of, yes, if it’s there, it’s got a very high likelihood of being accurate.

It is, classically, for Gustav, it’s a good lens into the future, because call it five years from now, there won’t be a professional doing professional activity who, if they’re not using AI to help them with it, is being fully competent at the profession they’re doing. The speed and intensity, which would allows you to process information, find information, analyze information, make decisions, is only going up year by year.

language is interesting language. I do think that it won’t be that it’s kind of like a grain of wheat or a pound of copper. Foundation models will be different. They’ll actually have somewhat different strengths and weaknesses. Like, some of the most interesting work that’s being done right now is kind of comparing and contrasting Gemini with ChatGPT, with Copilot, and Bing Chat, and it’s kinda saying what because even between the OpenAI, are the differences? How do they operate? What

That’s not necessarily a bad thing, but part of the thing is to be choiceful about what your analyses are relative to what’s important that you solve, and understand that everything is changing at a fast pace. So if you say, for example, like what I’ve heard from some people who’ve done Gemini is better at producing fiction, OpenAI is better at the kind of Wikipedia like reports, three months from now it may be different. Me as a user, that might say, I’m trying to solve a certain kind of thing, like if I’m trying to write a science fiction novel, maybe I’m going to go to Gemini and work on that mostly.

But by the way, of course, one of the things for products, for individual users, using multiple is good. Because that’s where I was going with is it commoditized? In fact, think there will still be differences, but I think our behavior, both in creating products and as individuals, will be almost more like kind of a conductor in an orchestra. It’ll be like, well, you know, a little bit louder on the bass, a little bit softer on the cello as a way of navigating. It won’t be that they’re all just vanilla each other.

Now I do think, because we have multiple ones that are competing with each other, there will be an intensity of providing it at low price, fast response, etcetera, because if provider one isn’t working that well, I’ll just go to provider two.

I do think that every cloud provider and frankly, every scale software company is gonna need to have a serious provision within like, they’ll want to have some internal teams and some internal talent doing things. It doesn’t necessarily need to be that everyone’s doing frontier models, but because I think it’s a blend of models that will be creating the quality of the highest tune cognitive services, the elevation of cognitive capability, I think that everyone’s going to need to participate. Whether it’s acquisitions, whether it’s, you know, developing and refining open source models, whether it’s building your own, you’ll see all of

What is certainly the case is that part of what’s happening with AI is it isn’t like we’ve discovered the algorithm. It’s that we’ve realized that applying scale compute, turning these things into learning machines off a lot of data with scale teams is the thing that has generated all the magic. And so compute is obviously a very, very central part of that. And as yet, all j curves turn into s curves, but at the moment, we see each new level of scale of compute bringing new serious capabilities to the table.

And so therefore, really matters, and compute will matter in training. Compute will matter in inference and providing. We’re going to be on a journey, maybe forever, as human beings of massively increasing the amount of compute available to us in all

Moore’s law was specifically transistors. Right? And so that de acceleration of Moore’s Law still happened, but what people didn’t realize is, well, I can just build much, much bigger data centers, and I can tune a new set of chips to a much different set of functions, where those functions, mathematical, graphic processing units, can actually in fact do things that are extremely important to us. And so I think we’re already in the equivalent of a resumption of the acceleration of Moore’s Law, and of course, you know, if we want to be a little bit like advancing the ball in the future, you know, what happens when quantum computing becomes very real and apply this.

There may be other areas where it’s even much more accelerated.

So I think that there is literally no doubt in my mind that there’s viable businesses that will play out. Purely in the foundation model? Purely in the foundation model. And also the others, but the foundation model as well. In part because, look, at some point, it is the classic venture thing of you start investing, and you may even have bad operating margins. Like for example, if you said the first x thousand stays at Airbnb, the operating margins were terrible, right? It was kind of money losing.

You have to get to a certain scale, you have to bring a certain amount into it to then make it into the amazing business that it is. It’s like the food

Of these things have various elements to say the very first of it looks like very bad. Now why conviction on even the frontier models? And the answer is because, look, software ultimately, that almost always has good operating margins. Software is cheap to execute. It’s expensive to AI to develop it. Now, if you just say, well, we’re not getting as much value out of it five years from now, the development curve costs will get more slower and more amortized as we’re turning into the

Yes. Exactly. So if what we were was seven big tech companies heading to three, then I would be concerned, right, because of kind of the monopoly effect. But I actually think we’re seven big tech companies heading to like 15. And so, yes, the big tech companies are becoming bigger. All of them are compounding. It’s a function of globalization. It’s a function of we’re all in one big global network. It’s a function of the more and more importance of tech and AI. All of this stuff leads to kind of massive size.

But size is always relative. Right? So if you kind of said, well, there is only one Uber company, and it’s called Microsoft or it’s called Google, well, that would be concerning. But the fact is Microsoft, Google, Amazon, Apple, they’re all competing with each other ferociously. And part of the thing I think we’ll see, because people frequently say, wow, but that’s it. And it’s like, well, but actually, look, Salesforce is coming along. Like, there’s a number of companies that are seeking to get into that that set.

What’s more, if you don’t take only a US perspective, you go, well, what about ByteDance? ByteDance is a huge company. Right? You know, Alibaba, Tencent, etcetera. So we’re heading towards globally more and more very large tech companies, and that actually is a feature, not a bug. Part of the benefit you have is take this AI revolution we’re in. It’s because you have these large companies that can invest in multiyear risky projects, that can create things that are massively valuable for society. When they started, there was no clarity that there was any

So, look, VCs always are looking for the areas that not off the mainstream kind of major company already established. Like for example, if you came to me and said, I’ve got an idea for a new smartphone, and I want you to fund my new smartphone to compete with the iPhone. You’d be like, okay, good luck. Or I have a new kind of mobile or desktop search company that I want you to invest in. Okay.

And so, sure, there are things in these large tech companies, like, and with AI, that you shouldn’t be doing startups to try to compete with them on, right, because that’s where their dominant position, and there’s network effects and modes and enterprise integration and list of customers and all the rest of the stuff. You go, no, no. You don’t attack a fortress at the front door. That’s what disruption is. It’s like, what’s the new territory or the diversion of the river or something else as kind of a way of doing that?

And I think there’s a lot of that available with InVenture. It’s just not, oh, I’m gonna go build a frontier model to take on Google search.

I think it’s almost 0% the way a frontier model created in the way the frontier models exist today. Right. I think that wave is gone. But on the other hand, to think that the only pattern for these scale learning systems happens to be these attention transformers together with a set of techniques like, you know, mixture of experts and other kinds of things, and say that’s the only one true path. It’s like, well, there might be other paths, and so therefore, that may also get you alternative paths to building interesting frontier models, and those may still happen.

think that part of the thing is to make sure because a lot of people try to build these models and build them poorly. So it isn’t just put in data, put in compute, press button, go away for six months, come back and go, I’ve got GPT-four. There’s a whole bunch of art and science, and what I mean by art is like, well, you train it this way first, then you train it that way second, as part of the kind of the evolution of this. And there’s a bunch of stuff that isn’t written down in the published papers and so forth about how to do this kind of thing.

So quality of model really does matter, and it’s a small number of teams and a small number of folks who are doing it that well so far. It’ll broaden. There’ll be more. By the way, size of compute, as per the earlier question, is really gonna matter. People frequently mistake what quality of data means, because they think, oh, you have to train on the New York Times. And actually, in fact, most of it is quantity of human data, not because you’ve had any particular journalist who’s who’s got a particularly nice way of writing with words that makes anything particularly useful.

networks. Yeah. In terms of professional data, yes. But, you know, there’s lots and lots of very interesting piles of data. For example, all of the academic publications. I mean, there’s tons of things that are very There’s GitHub with

think NVIDIA is a great company and Jensen’s a great CEO. I don’t actually myself trade in the public market that much. I’m a private market investor, so I don’t actually really make those operative Yeah. Right? So it might be. I think their chips are extremely valuable in this scale AI transformation that is happening. Without all of the NVIDIA inventions and chips, we wouldn’t be nearly as far along as we are. Now, the real question is buy at this price.

So Satya and Mustafa had a conversation, and it was kinda funny because both of them called me afterwards. So I had this really fascinating conversation about what is the kind of work we could do together. Do you think we should possibly do think about this? And so I had separate conversations with both of them because obviously I’m on both boards, and then they said, well, should the next conversation be the three of us? And I’m like, no. I think I have to keep my hat separate because I have to talk to Mustafa from a Greylock hat, from a Inflection board hat.

I have to talk to Satya from a Microsoft public board member hat, and I have to make sure that I’m playing each role effectively. If I’m in the room, it’s like, you you can almost like bring two baseball caps to go, okay. It doesn’t quite work that way, so it’s better to do that. And so they, in a set of conversations, kind of got to the position of saying, well, Mustafa had already been worried that the notion of the massive increase of scale of frontier models was going to be beyond startup companies’ ability to monetize with an agent infrastructure.

It’s going to take a long time for agent infrastructure to monetize. It’s like, look, we have this really great product, we have this really great agent, but if we’re going to keep pace with that, we’re going to be massively in the red for a long time, and yet that’s what we really want. We want to make sure this great agent that really helps everyone is going to continue and grow and thrive, and that was part of the, okay, well maybe we could do that at Microsoft.

While at Inflection, part of it was we’d been thinking of like, well, maybe the real business is around an AI studio. You have a number of different models, some of which are custom trained, like Inflection, other which are open source models, and you’re kind of being a B2B AI studio model. Because that’s what we had been thinking about doing with Inflection anyway, given the economics of the business. So I was like, well, that business could then be funded by this transaction. The people who wanna build agents can go do it at Microsoft, and the people who wanna do b to b studio can

Look. There’s a decay rate to the model, but the model is still valuable. Right? Because part of the idea with Inflection, and still Inflection 2.5 is, you know, a unique GPT-four class model that is the best one that prioritizes emotional intelligence along with analytic intelligence. So its EQ is as good as its IQ, and its IQ is pretty good. And so yes, that, and by the way, that learning and so forth, but part of it was of course that one of the things that Microsoft was looking for is year by year, how is Copilot iterated into the kind of agent that is this new agentic universe, this new world that we all see coming.

And Mustafa and Karin and some of the key team at Inflection, that’s what they wanted to build. So I was like, okay, let’s build that. And we’re starting their whole strategy is building on the OpenAI models that, you know, Microsoft and OpenAI are already working on together. Do you think that Inflection could have been a standalone business? Well, is a standalone business as a B2B AI studio, but as an agent, it was a very risky proposition. What does it mean like a B2B AI studio? So large enterprises can come in and So someone comes in and says, I have a customer base.

I would like to provide AI to them in the following way. Can you help provide the set of models through the APIs that I could then use to service my particular business need, whether it’s a community, a customer service function, an application, etcetera.

The answer is both differently. So and I’m not trying to weasel out of a question. Look. If you’re saying what really matters is, like the driving function here will be the billion dollar compute run to make this model, it’ll be very difficult for startups to do that. Even in this more modern global venture capital where startups can raise billions of dollars, if it’s a billion dollars per compute run before you’ve got your revenue really working in a way that that you can predict where that becomes very valuable, that becomes very difficult to finance.

So that’s all large companies. That’s, you know, Google, Microsoft, or, you know, kind of their partners, OpenAI, other other folks. But that doesn’t mean that there isn’t a lot of room for startups. Startups can do all kinds of things. They might build smaller models. They might build models in different ways, they also might use the frontier models through APIs, but they can do things, can take on markets, because any company can only take on like a couple of markets really in a serious way. Sure. So there’s a lot of markets that, you know, Google, Microsoft, Amazon, etcetera, don’t get to.

And so it’s like, okay, startups are opportunity there. And what if you were creating something, you took a risk in getting there, whereas people thought you were crazy? Like, for example, think about early Airbnb where people said, that’s that’s nuts. No one’s gonna ever rent a room from some other human being. And you say, no. No. We take a risk when we get there, and then when you’re when you’re right, you’ve created a very valuable business.

Well, that’s interesting. Blitzscaling is still applicable. Chris and I have written about this a little bit in Blitz Capital, but the Blitzscaling creation of the future, it’s like you’re doing capital raises that literally the general received wisdom is that looks totally insane. And you have to have that in your tool set some. So that was, for example, what Sam Altman was doing with OpenAI, where you go, wait, you’re gonna put billions of dollars into this five zero one c three effort that doesn’t really have a direct commercial goal in order to build this stuff because this scale application of compute with this small focused team is going to create something that’s going to be an Arcimedian Leverage to Change the World?

Okay. Sounds good. Yeah, sounds good. I mean, like for example, the very first OpenAI Microsoft deal was you know, Satya and Kevin Scott brought to the Microsoft board. We want to put a billion dollars into the investment in this vehicle from a five zero one(three) that doesn’t really have any particular plans at making its equity worth anything, and we can’t demonstrate any valuable product right now that’s useful, but we have conviction that it should, and this is what we should be doing. That must be a

You need to be able to fundraise as a founder, so you have to have that in your skill set. There are some fundraisers who are not particularly good founders, especially when they’re, you know, quasi fraudulent.

No. It’s a little bit like what you were saying in blitzscaling. This is the thing that standard studies of business don’t really track, which is when you think you might be successful at scale and you still have tons of risk, you go, generally speaking, within the software world, and maybe even the hardware and a bunch of others, raise a ton of capital and start spending it on the presumption you will be successful even though your probability might be 20% or 30%. And to do that, you must be a great fundraiser.

is essential. I think very few companies pull their selves into a coherent shape and deliver in a focused way without some scope of competition. There’s exceptions. But, like, for example, Google Brain invents the kind of transformer and publishes it, and it’s kind of bouncing around as kind of a research project. And it takes OpenAI developing it, scaling it, and releasing it as ChatGPT for Google to go on Red

The successful entrepreneurial things, generally speaking, is you outpace your competition. So when you have a lot of ferocious competition, that is challenging. But you’ve actually put your thumb on something that most entrepreneurs don’t really fully recognize, which is to some degree, it’s the most valuable market you can against bad competition. Yes.

Well, you know, one of the things that was very interesting, and I think Sam doesn’t get enough credit for this, Sam was very, very much wanted to have a board that was separate governance from himself. A lot of founders tend to set up as, no, the board is my rubber stamp vehicle. I’m in control. I’m doing this stuff. And part of that, for example, was when Elon had announced to the company that he was leaving and not financing it anymore, and Sam called me and I said, Look, I’ll cover the salaries, and he asked me if I join the and I said, yes, I’ll join the board.

And so then he said, well, will you come to a fireside chat? Because the company does you know, only a few people here know you. Will you do that? I was like, sure. Of course. You know, they they should get to know me and we should have mutual trust. And literally, one of the questions that Sam asked me that he hadn’t prepared me for is we were sitting in this entire company meeting, and he says, so what happens if I’m not doing my job well? And I’m like, well, I would work with you to try to improve it.

He said, no, but what happens if I keep not doing my job well? I’m like, okay. Well, then I would fire you. And he’s like, okay. Great. I’m like, you just asked me in front of the whole company, will I fire you? No. It’s like it’s like, well, that’s a little strange. It’s the only time I’ve ever been asked this question in any of my boards in front of a company thing. But it’s reflecting what Sam has always been trying to do with the OpenAI board, which is it’s a genuine independent governance board.

People don’t realize, with all of what I thought was the set of mistakes around the November debacle, that one of the positive attributes of that, that was because Sam had made an effort to say this is a separate governance board. Now I think he was making too much of a AI safety, who are the most AI safety people, make sure they’re on the board. Some of those people should of course be on the board. But you also need to have good board knowledge. Good knowledge about what are the kinds of things that a scaling company and a scaling organization, because by the way, it’s a chaotic mess, that’s part of what split scaling is about.

And what things do you kinda go, okay. That’s fine. And what things you learn from and which things are important?

Well, look, that in a week is a lot. But by the way, you know, having done this at PayPal, having done this at LinkedIn, having done this at Airbnb, having done like, all of this stuff, it’s always a chaotic hot mess. Is it? Yes. Sometimes a lot of it happens in one week, and sometimes it’s different. The real question is how do you learn and adapt? Like, do you go, okay, that was a screw up. Now let’s fix it, and let’s be better.

Well, all of those are very good examples of ones that have done it very well, because look at the look at where the companies have become and what they’ve done. I would say I don’t know if I could say best. There were different problems. I mean, we had at PayPal, we had eBay trying to shut us down, and that was where all our volume was. We had at Airbnb, they had this very early trust and safety thing where a person had stolen a credit card and literally totally trashed a place, right?

Like had a kind of like a drug party binge leaving the place in not quite the equivalent of flaming ruins, but kind of physically totally trashed. And so what do you do? Because you need trust in a marketplace. So each of these things had a different emergency that you were like, all hands on deck, we have to solve this, and we’re dead if we don’t. What do you do when shit hits the fan? You’re a very calm, collected individual. Well, so for me, at PayPal, that was my primary job.

Yes. I was like the go solve this problem. There was a lot of go solve this problem. So, like, go make sure eBay can’t kick us off their platform. Okay.

Tuesday morning. Should be done by Tuesday morning. Help us launch Japan was another PayPal one. The lawyers were coming back week by week with a longer list of here’s all the criminal penalties and all the things that if we launch PayPal in Japan, you know, no executives could land there because they’d get arrested and so and you’re like so literally, Peter came to me and said, the lawyers are screwing this up. Figure out if we can launch Japan and kill it if we can’t, but otherwise get us launched.

And I was like, okay. We figured out, we launched four weeks later. What was Peter’s

Oh, look, super smart. I’d say his biggest strength is a willingness to be contrarian based on first principles and an absolute maximization of speed. Can we have this done by the end of the day? But, you know, that kind of thing. And I think his biggest weakness is probably thinking that that’s the only way you play the game. Often the way.

Yes. Well, one of them is, I think even shorter than that, every single person who has a smartphone will have one or more personal agents, personal AI agents, that is their agent, but it is, like, there for them, and on a whole range of things, like not just where should I go for a restaurant meal, you know, tonight, or hey. I had this odd conversation with this person at work. Can you help me talk about it? To hey, I’m writing an essay on this, can you critique this and get me more information?

But just a whole range of things. Like, for example, I have this skin legion, you know, should I be concerned about this? Every person today will have an agent that they interact with and consult with every day multiple times.

And by the way, this is part of the whole realm of data, is that now when you tell people, oh, we’re gonna have to integrate all this data, go, oh my god, and they feel like it’s a lack of control a loss of control. And that’s frequently what they mean by loss of privacy is a loss of control. And actually, in fact, as you begin to realize of actually, in fact, like if I had a biosigns monitor that would tell me and tell a doctor if I had somebody like for example, I would love to know if it’s like, uh-oh, you might be fifteen minutes away from a heart attack.

That would be very valuable knowledge to know, and I would like it to escalate to the ambulance shows up. Like, it’s almost like you think it was like a support version of minority report. Like, the ambulance shows up, you’re why are here? Oh, we’re here for you, but we’re about ten minutes early. We can wait.

Right? So people think, oh, it’s all good. Some people say, it’s all going be different this year, all going be different next year. It’s like, next year’s going look a lot like this year. But as you compound year by year, and as you get certain phase shift changes, then all of a sudden the world looks entirely different. I mean, for example, consider how just driving over here through London, I was reflecting on two thirds of the people I saw on the street were looking at their phone doing something with it.

Think about what like twenty years ago, if you had portrayed that, they said, Oh, that’s the Jetsons. That’s the city of the future. It’s like, No, it’s right here.

No. And part of that reason is because if you look at what the AI revolution is, it’s a shift from we program the computer, and obviously we’re still programming but as opposed to programming knowledge into it, we’re giving it a learning algorithm. The learning algorithm then goes through a whole set of training sequences to become a cognitively sophisticated device. That particular quote presumes that the algorithm is something like that, as it were, is coherent, that you kind of look at the algorithm, and actually, in fact, what we want is a really robust learning system that has learned a lot of really great things, has incorporated a lot of human feedback into that, and is making good actions and judgments based on that.

And by the way, that’s what we hope for from our leaders, whether they’re leaders in politics, leaders in, you know, investing, leaders in entrepreneurship, leaders in academia, leaders in journalism. That kind of we are learning machines that have learned good judgment in these questions. Do you think we dramatically overestimate the concerns

Broadly, yes for the following reason. The problem that I have with these critics is that the real issue is AI is a human amplifier, right? So it’s part of the thesis that I’ve been going out and trying to get people to understand is actually in fact it’s a human amplifier for a large number of years. I don’t know if a large number of years is ten or five hundred or whatever, but a human amplifier. And amplifying bad people too is the question. So whether it’s a terrorist, a criminal, a rogue state.

And so we go, oh, the robots are coming. It’s like, well, I’m a lot less worried that the robots are coming than Putin is coming with his AI enablement. Right? So let’s focus on the on the bad humans and what’s happening there, and then pay attention some to the science fiction side.

So I’m glad to say that every single person so far is wrong. I think part of the reason they’re saying that is the American electorate right now hasn’t really fully remembered the corruption and incompetencies of Trump. They’re like, oh, the world is kind of a fragile place. I’m unhappy with how there’s all this global conflict. I’m unhappy with the tragedy that’s going on in Gaza, a stack of things. Because of the huge stimulus thing from COVID, prices have gone up, I’m unhappy that there’s been an inflation of prices.

I have some agitation, and completely legitimately agitation. So they’re like, well, I’m agitated. Biden’s president. I’m agitated. When people start remembering that Donald Trump agitated for an insurrection that killed police officers, he was convicted by a jury that said, not only did you do sexual assault, but you slandered about it in a jury thing. And they start like, it’s a jury that when can they begin to remember that, they will begin to get much more negative. So it isn’t that it’s Are

I hope not. I think there’s real work ahead of us to remind people about what a Chernobyl Trump is. But I think if we successfully remind them and then also, by the way, you know, Biden has passed more bipartisan legislation than any president in decades. He has done stuff for the climate. He’s done stuff on inflation reduction. He has helped assemble the world on Ukraine.

But there’s normally a leader, and Biden is Well, but Biden selects his staff. We have had a whole history of presidents where the staff is what does everything. Mhmm. And he is the person who’s like, look, I’ve had a two hour lunch with Biden. Like, it’s very popular to say, well, he’s not cognitively with it. In a two hour conversation, he was asking me questions about AI, he was explaining stuff that was going on in nuanced detail in Israel. He was kind of saying, look, this is what I care about your average worker, and this is what we need to be doing, and how can you technology industry be helping the average worker?

Look, it was a two hour robust conversation.

There’s almost a little bit of implicature in the question that income equality is the most important feature. I think that AI will raise incomes both for working class people, and for, you know, middle class people, and for wealthier people. I think it’ll raise incomes across all of them. And so therefore, I’m actually in fact quite positive on that, and I can make arguments for how AI provides productivity increases and is positive on each of these levels. Now there are some jobs, you know, when we’ll just want the AI only to be doing it versus the human, so for example, you know, driving.

We actually want AIs driving because they don’t get drunk, they don’t get tired, they have a lot more sensors to pay attention to the world around them, like they can have LIDAR and infrared, so they can see the kid who is running out from behind the parked car, or as a human can’t. There’s all kinds of reasons why we would want that. In each of these cases, like for example, if you look at, like one of Greylock’s portfolio companies, Cresta, did a study of which people in the customer service were most benefited by early AI, and it was the entry level people.

It was the people coming in and learning the job much more quickly, and that’s that’s a pattern that you can see across this. So I think it’s beneficial across the entire thing. Now, the inequality question, I think a much more fraught question for the following reason, which is I don’t know of any human society organization where we don’t run on the basis of inequality. Right? Like better investor, worst investor. Better entrepreneur, worst entrepreneur. Better business, worst you know, getting into college, etcetera, etcetera. We run our entire human society runs on inequality.

Even places ostensibly communist, who’s in charge? Who gives give other people orders, etcetera? Like, human beings are tribal creatures, so we have inequality inherently in the system, and we try to orient it so that people can, through hard work and talents and everything else, rise up as much. And I think that equality of opportunity and equality of ability to make amazing things of themselves, like you with 20 BC, is a really, really good thing. But the end result is inequality. Then you say, well, but how much inequality is too much?

Well, obviously there’s an answer to how much inequality is too much when it’s like, well, I now am the autocratic ruler, like say I’m Putin, and I say half of the country belongs I own and belongs to me, and everyone has to do what I say or I kill them. Okay. That would be a problem. And so there’s problems, but it’s like, for example, people say, well, but a CEO shouldn’t make more than 20 times what the entry level and it’s like, where do you get your magic number?

I want equality of opportunity. I want massive economic opportunity. I want ability to do your best work and express your talent, and I think AI can do all of that. Does it benefit more wealthy people or more working class people? I think we want it to benefit both massively.

that

risk? Mean, is one of the reasons why the vast majority of stuff that I talk around technology and AI is we will have much better tools to make much better things for human beings and society in the future. Let’s make sure we get there to building those tools. So for example, your average press person, your average government official will say, What’s the most important thing in AI? We make sure that those big tech companies don’t dominate with the AI. That’s the most important thing. I was like, well, that’s interesting.

I think your most important job is I have a line of sight to a medical tutor, a medical assistant, probably better than your average general practitioner, on every smartphone. Don’t you think your actual job is how do you get that on everyone’s smartphone, that every single person in your country, maybe in the world, has potential access to good medical advice? It may not be perfect medical advice. Wouldn’t you think that’s your top job? Or how about a tutor for any age of per of human being in any subject?

Wouldn’t having access to all of that for every child that you have in your

Yes. That’s the orientation, and that’s the blind spot from the press and the blind spot from the dialogue. But

When you talk to the right smart people, the answer is yes. And so for example, in Europe, recently, have you seen some of it? Macron. He’s been doing a good job here. He has. Yeah. And so the White House executive order, very good. The UK AI safety institute, leading in the world, like, really, really good, helping The US, etcetera.

I would say medium sized countries, the basic answer is don’t try to just replicate everything. Don’t try to try to say, we are going to have our own GPT-four model or whatever the current time slice of it, and that will be our own Bangladeshi, you know, kind of version of this. Because and just like, for example, say, well, you’re to have your own OS for your phone, or you’re going to have your own OS for the computer. It’s like, no, that doesn’t make sense. What it does make sense is to say, okay, given that these technology trends are happening, and that there’s an important you know, the way that I describe artificial intelligence in an economic sense is it’s a steam engine of the mind.

It has, just as the steam engine gave us physical superpowers, the AI gives us mental superpowers, and we’ll have a cognitive industrial revolution across this entire world, and it’ll be much faster. This is one of the things that’s like, oh, a little vertigo, much faster than the industrial revolution because of the speed that it can spread throughout the world because we have the Internet and mobile in a globally connected world. So when you’re a Indonesia or something else, you say, okay, well, we know that the tech companies are building these things.

We know that these things are possible with open source models. How do I most help my citizens? How do I most help my industry and my people participate in this? And then be making those moves. And so, for example, you you don’t say, well, I’m going to build my own supercomputer. It’s like, well, how do I, you know, kind of get the right access from the hyperscalers to good compute that helps my industry and helps my citizens? How can I kind of nudge them to say, I want you to provide systems and services that fit my cultural values in here and make that possibly work?

So I think it’s just like a startup developer. You say, well, I’m not necessarily going to build my own frontier model, but I’m going to build upon these platforms that are being constructed. And that was the reason I was using, you know, mobile platform and a I’m PC gonna build upon the platforms being constructed for things that help my society.

I don’t think that it is an unusual threat to US democracy. Although I do think the key question is is, you know, China bans Western, you know, kind of social media and other companies with it. So I think it’s not an unfair thing to say, hey, look, if you’re banning ours, we can ban yours. I think it’s I think that’s the the baseline of the kind of the fairness side. Now that being said, part of the Chinese approach to these things is they view that all of their companies essentially work for their governments.

Now they think that’s true of the West too. They don’t realize, no, actually, in fact, we we operate differently. Like, Google does not work for The US Government. Microsoft does not work for the US government. There’s ways they’re accountable by law, but they have to go through a legal process for that kind of thing. And so given that legal process doesn’t exist within China, one worries about what security vulnerabilities and other things that happens because the companies don’t have the kind of autonomy that our companies have, that they have where, for example, they can sue and say, no.

We’re not gonna respond to this this subpoena or this request from government because we think that’s an illegal order, which you can’t really do in China. So that kind of thing does prevent present a possible concern and risk. I don’t have any data that suggests that they’re doing anything right now on it, but it’s possible. And so you wanna you wanna close down that possibility. Now, you know, what the US government proposed was to say, get to a form of governance that is not the Chinese form of governance that is within the Western system, divesting as a version of that.

Like if they became a if US TikTok ByteDance became a public company launched on, you know, the NASDAQ or something else, that would give it the kind of governance that would say, okay, great. We understand that the governance is by, you know, kind of public rule of law, not potentially corruptible by a government that doesn’t have that public rule of and that’s a good place to be.

Well, think likely, and I have no inside information here, I think ByteDance waits for the election because they presume that Trump could get bought off by investors since, you know, he’s a coin operated person in his entire life, not just as in his presidency. He is not.

But he is completely coin operated. Yeah. Yeah. Right. So self coin operated. Like, you know, to some degree it’s like that Churchill line. It’s not insulting that he’s he’s for sale, it’s that he’s for sale for so cheap. Have you met him? No. And, you know, it’d be difficult given that the harm that I think he’s doing to the country and the world is, I think, is amongst the greatest ills that has made a huge negative damage in the world.

There’s two different questions there. So one, Zuckerberg is one of the best meta strategists. I deliberately pun on Meta, but meta strategists, like, kind of like, what is the game we’re playing of any technology leader on the planet? I learned from him on this, he’s super good. Matter of fact, one of the things I have to do is I have to send him the speech that I just did in Perugia, because I was referring to Carthago de Lendesk, and I hadn’t realized that was actually one of the things he put on his T shirt and all the rest until I saw a picture recently.

But he’s really great meta strategist, and I think is seriously underappreciated for that. Now in the question of public markets, I think part of what you have to do is you have to build communication and credibility over time with public markets where they go, okay. We you haven’t surprised us. You may have made a new announcement, but you’ve been kind of building towards it. And so they tend to react when you say, well, look, we’re doing really well, but our profitability is gonna go down because we’re gonna invest in all this stuff.

They go, uh-uh. We Like, don’t understand how to process it, and so the surprise response is a negative response. I think obviously the public market should have responded differently, because I think he is a great meta strategist, and his turn into AI is going to matter fundamentally here is a 100% correct.

Why don’t we watch first just because then I’ll both put my answer and my commentary on the read AI answer together.

Absolutely.

I’ve changed my mind on how fast generative AI is advancing. I initially thought significant breakthroughs in creative fields would take longer, but AI tools have rapidly evolved in writing, art, and music.

So on this one, it’s funny. It’s almost like you get accused Read AI of being like the wise person with the elephant. Because it’s an AI, it’s like, what I most changed in my honor is AI and how it’s accelerating. And it’s like, nope. Actually, I haven’t changed my mind on AI stuff in the last twelve months, one year. I would say that my answer to this question would be that it’s almost much more of the response to AI, and I think I’m confident that this will still happen, but, like, trying to get people to see the positive sides like, it’s like, look, most people are like, oh my god.

I why are we willing to say AI stuff? It’s dangerous. Like, risk discourse. Risk discourse. Like, look. We have amazing things in front of us, and that’s why it’s worth navigating the risk. And the slowness by which persuading journalists, academics, other people to say, hey, look, by the way, your discourse around AI is like the discourse when the printing press was launched. People would say, this spreads misinformation. You’re like, okay. By the way, the printing press was a good thing. You might have noticed.

So let’s think about what could possibly go right, and the slowness by which they’re getting to what could possibly go right.

The biggest misconception is that AI will completely replace human jobs. In reality, AI will augment human capabilities, transforming jobs and creating new opportunities we can’t even imagine yet.

So I broadly agree with this one, but I do think that what people kind of presume is that with a massive increase in cognitive capability that that will just kind of create, you know, since we’re in Britain, a bunch of people on the dole. And actually, in fact, think what it’s doing is transforming a lot of jobs. And just as we discussed, learning how to do prompts and learning how to do it, that will be an essential tool. Like, for example, if you ask people to say, most jobs require some facility to do with reading.

And so, yes, you’ll have to to be able to use these tools in some useful way, but I think there will be lots of new and interesting human jobs.

Well, I think we’ll have a whole swath of jobs that are in the kind of prompt engineering, AI, research assistant directors.

On me. Yeah. Okay. My prediction will be we’ll get this one wrong, but we’ll see.

Let’s see how this goes. Because because by the way, as as I told you, I hadn’t listened predict this? Yeah. I I don’t know actually what it will say. I literally have zero idea, and it’s one of the reasons why I’m I’m Surprise. Yes. And so part of what we decided in doing, this fun thing, is I literally have not listened to any of these answers before right here, right now.

I’d say Satya Nadella. His leadership at Microsoft focusing on empathy, innovation, and a growth mindset has profoundly influenced my views on leadership and culture, especially as I’ve watched him successfully navigate rapid technological changes while fostering an inclusive and innovative environment.

It’s not bad, actually. Yeah. I’ll give it pretty good points. It was it was much better than I was expecting. It’s better

Yes. Incredible to read AI. So points to that. Look, probably the thing I’ve most learned from Satya has been how do you do cultural transformation of massive scale organizations? And he has demonstrated that that is doable and possible in ways that I don’t know if anyone ever has before. How does he do? Part of the benefit is he comes to Microsoft as an an outsider insider. His whole career was through Microsoft, but he was in charge of Bing. He was in charge of Azure. He was the one who was launching the new strange projects versus the person sitting on top of Office.

That gave him both this kind of outsider insider perspective, which was, okay, I understand what is great about Microsoft, and I also understand how Microsoft needs to grow and evolve. And those two things together have made him amazing in his transformation, and he’s willing to take Reid’s really bold risks. Like for example, it’s like, I’m going to go put a billion dollars into a five zero one(three) on a partnership thing with OpenAI. He does these bets that are super smart, even though I think the vast majority of public company CEOs would never make that bet.

What’s the boldest bet

Oh, what’s immediately coming to mind is some investments that didn’t work out. I think I’m not gonna call them out because it seems ungentlemanly as a partner and as a VC, but there are a couple of investments that I’ve made that I thought, oh, this is gonna be great and huge, and was completely fumbled. What did you get wrong? They’re different. In one case, it was the people. In another case, it was the difficulty of the market. Part of the thing that you learn as you get more experience in VCs is that sometimes markets are diff are just wrong, like difficult.

Now here’s one that I got wrong initially that will be fun to share. We were talking about bad competition being the central thing. So I think I was one of the first couple people that Elon Musk pitched SpaceX to, and I literally kind of laughed at him thinking because by the way, to my defense, his pitch was, I’m gonna be the first person to send life to Mars because I’m gonna send a turtle to Mars. And you’re like, that’s not a business. That was the pitch.

Right? Good. Good. Do you have a slide two? Exactly. Is there another slide? And there was a fool in the room, but the fool wasn’t Elon, the fool was me. The thing that I learned from watching Elon’s rocket success, pun intended, was that if an industry has had decades of technological innovation that are not reflected in the market, because the structure of the market locks in this anti innovation kind of oligarchy, it’s hard cause there’s something that’s locking it in. But if you can break it, it’s huge.

Elon is one of those people who takes the whole concrete wall down with him when he’s going through it. So if someone can do that, it can be really good. And that was an instance of a mistake, because I should have said, yes, I will invest in this. And now, of course, I’ve updated to when I think it’s somebody who could take on the structural change of a market because the technology has already been built, then I will consider investing in things.

Look, they think they’re both very smart. There’s a deep value in allocating capital to the kinds of products and services and companies that could be transformations of the world. Maybe there’s other things that would better, like being entrepreneurs or being CEOs or other things, but I think it’s a real contribution.

him

Keith, super smart. He’s worked for me twice, once at PayPal, once at LinkedIn. I provided him as instrumental reference for working at Square and, you know, a sack of other things. I called Keith because he’d he had worked with the founders longer and earlier than I. I just literally met them a couple days before, and was, you know, kind of sorting through the, you know, questions of what are the questions that a smart person might have once they’ve sat with this project for more than a couple of days, and have you got answers to those questions?

How do you think about it? Which is the what I do on every investment that I do. And you know, Keith was quite helpful to my making my decision, because he was like, Look, I met them, I’ve talked to them, they’re smart, they’re first principle thinkers, they’re not hiding any skeletons in the closet. Helpful. Yes, that is, as you know as an investor, extremely helpful, because you’re like, I met them three days ago. It seems very exciting, right? And I want to offer them a term sheet right now.

And so all of that was quite good, which is, you know, the reason I invested and joined the board. And I was trying to persuade Keith to get a little bit more closely involved too, and I think he was a happy investor but wasn’t Keith said he said one line

Maybe it’s something along the lines of that it actually has that the marketplace demonstrates the network effects that we we expect to have.

Oh,

First time founders often don’t iterate quickly enough based on customer feedback. Rapid iteration and adaptability are key to finding product market fit and achieving sustainable growth.

That’s interesting. Not bad. Not bad. It is definitely one of the major failure points. Yeah. Right? Because the over belief that your vision is the one perfect vision versus getting data. Now part of the thing where that’s a little limited is part of what I suggest to founders is that they should always be getting, trying to get disproving data to their investment thesis on their business and on their product, but even by going around and asking smart people. Like saying, what do you think of my product service?

What do you think will go wrong with my business? And aggregating. That doesn’t mean that everyone who gives you something is right. Maybe most of them are irrelevant, trivial, wrong. But when you begin to hear the pattern of it, that’s the way you go, oh, I need to adjust. And you might even learn that you need to adjust

That would be another classic one is distribution, especially in mobile Internet software, is more important than quality of product.

Yeah. So, look, that’s always the challenge in startups, and it’s one of the reasons why the Internet has been much better for entrepreneurs than mobile ecosystems, because the mobile mobile ecosystems are controlled by solitary providers in ways that the Internet is not. And that’s why we’ve seen a lot more generativity on the Internet than mobile ecosystems.

AI to these 20 VC interviews. Probably just get him to do it. Yeah. It’s crazy.

Ten years, I’ll be leading an AI driven public health initiative to predict and combat pandemics and pioneering AI education systems to reduce global inequality. I’ll be shaping international AI ethics regulations and involved in space exploration projects using AI to support humanity’s expansion to other planets.

So Reid AI will certainly be doing that, because that’s I’ll be sending Reid AI to do all those things. Dude, we’re gonna be very big behind it. Yes. Exactly. That sounds

Are you happy now? Yes. Part of the earlier questions of, like, I’m a student of decision making, so I’m always looking back on decisions I have made, and it isn’t that I would necessarily remake them, but how would I learn to make decisions better in the future in everything I’m doing? And and that’s the reason I wouldn’t like, I just don’t know if I’d say, well, would I make that decision differently? Would I end up in a different life path that would be better than the one I’m in?

No. I I concur. Look, I I think he is driven. And by the way, anyone who achieves that much amazing success is driven. And by the way, that level of drivenness is very stressful. You have to go with that stress. And Jensen said something similar, you know, would you advise people to be entrepreneurs? And he was like, oh my god, it’s so hard.

that. And he also said we should lower our expectations because it will make us more resilient. No, no, no. Keep your expectations. And by the way, think in retrospect, he said, no. No. I’d still do it again. Like, think after he thought about it more, he was like, okay. No. No. No. So where do you actually want to be in ten years? Well, think the part of the Reid AI answer that’s absolutely correct is how do I help not just industry, but society and global ecosystem realize the human amplification benefits of AI?

And some of that’s public policy, but how can we have a new golden era of humanity? Just as the last chapter in my book impromptu, my speeches in Bologna and Prussia about homo techne, we evolve ourselves through technology. How do we have that that grand step forward as humanity? That is what I would like to be continuing to contribute. And in ten years, maybe I’m doing that more as a thinker, as an adviser, than as a board member, investor, inventor. Although,

are the things that I love and have learned about the Silicon Valley approach to building the next world, the the our future world, and what are the things that I think Silicon Valley still needs to learn? If you were to answer that question, how would you answer it? So on the first, I have learned so much by just by the fortuity of having gone to Stanford and under like, learned about software entrepreneurship. Like, I I never would have gone and sought out Silicon Valley. It was literally, oops.

I’m a student at Stanford, and a bunch of my friends have have have started doing this, and I’ve started paying attention and said, hey. That’s cool. Let me go do that. There’s a ton. One, every scale problem that you think of, 30 to 80% of the solution is technology. So criminal justice, economic justice, medical, climate change, that having an orientation of how do we deploy technology to help that scale problem, because technology changes the possibility curve, changes the cost curve, etcetera. So how do we do that?

And Silicon Valley is one of the most intense places in the world of how do we organize corporations and organizations and startups to tackle new technological problems? How do we take risk, allocate initial early capital in the seed? How do we evaluate it where we grow it? How do we build networks and teams around it? How do we have an entire entrepreneurial network within Silicon Valley where the coopetition of it is very important, right? Because both competing and cooperating like you know, the classic kind of theory is, no, no, you should compete and you shouldn’t talk to each other.

It’s like, No, no, you should talk to each other all the time and you should compete intensely, right, is kind of the angle on this. And that whole approach to how do we solve problems, and you know the classic aphorism is to a hammer everything looks like a nail, but it is amazing hammer, and it works on a lot of things. Not everything, but a lot of things. So that’s what I’d say is the things that I’ve learned from Silicon Valley. And obviously there’s whole bunch of entrepreneurial lessons like your distribution strategy for a software thing is more important than your product.

And then on the things that Silicon Valley needs to learn, by nature, you know, the whole like, hey, we’re pirates, we’re disruptors, is like, hey, we’re building this new thing, just trust us. It’s like, look, now technology is really centrally part of the fabric of human life. You can’t just have your engagement with the world be, hey, I’m gonna go build something that’s gonna totally change your life, and just trust us. You’ll like it when we give it to you. You need to be in conversation more.

You need to be in conversation more, whether it’s through the press or through direct, whether it’s with government, and saying here’s our theory about how the world should be. Here’s how we’re operating. Here’s our intent, here’s the things that we’re doing. Doesn’t have to be revealing your secret product plans, but an explicit kind of sense of here is how we’re trying to be in the mix of society and these are the things that we care about so that people can be in dialogue when they say, well, actually, fact, this part of it.

Like for example, whether you’re a phone provider or a social network, we care about what you’re doing with kids, and we don’t think you’re being smart yet about what you’re doing with kids, and we’d like you to tell us more about how you’re navigating, like, mental health issues or peer pressure, the fact that kids bullying at school can follow them home in these things, what are you doing to help with that? Right? Because that really matters to us. As an example, I’m being in dialogue about that.

We’re no longer just the pirates and disruptors. We are part of the genetics of what happens in society, and we need to take with power comes responsibility. We need to take responsibility of that, and the minimum beginning in responsibility is dialogue.

Well, it was a genius invention that came about of necessity because the Chinese government didn’t want to have any social networks with more than 500 followers. So how do you drop out the whole following mechanism and still get a high scale people are producing content, individuals, and individuals are consuming content, and how do you make that happen? Because the way that social networks like LinkedIn and others made this happen is through I have a graph. I have a graph of LinkedIn connections. I have a graph of followers, etcetera, and they couldn’t do it that way.

They couldn’t do followers all on YouTube. So what they did is they paid for so much content to be created, but then that created a scale effect in the entire thing because TikTok watches you the moment that you launch the app. How many seconds did you spend on this? Did you swipe, you know, on this, etcetera? And it’s learning about you to the next thing because it’s matching you to the flow of current content in this stuff. Now it’s a social network, you know, as much as there’s individuals who are creating stuff that’s being consumed by other individuals, but it’s not a social network in that I’m connected to my

It has been an absolute pleasure,