Reid riffs on AI adoption, sensitive data, and digital twins

0:28I'm Reid Hoffman. response to some listener questions. Hope you enjoy.

0:38Yes, Faison, I'm traveling from Chicago. So the question I had was, especially in this evolving world of AI, what are the biggest misconceptions founders have about scaling companies quickly? Oh, it's interesting. The usual general misconception is that scaling is just a straightforward game that's just kind of like the next step. Like once you establish product market fit, scaling is a relatively, you know, kind of like just, oh, hire more people, have a larger organization, reorganize. And there's a lot of things. It's one of the reasons why I actually think that the phrase product market fit is interesting and useful, but scale product market fit is what really matters when you're projecting a scale.

1:28And what does that look like? But then the question is, another kind of misconception frequently is like you've proven everything and then you just, again, add rocket fuel, go scale. Most of the time you go, I've got a high enough probability of scale product market fit that I'm just going to go. right and we see lots of companies like that doing that in various challenging ways so you know classics from the book are like Uber and you know Airbnb and others and so like for example very early days you know no one few people here will remember this but like in early days it was oh yeah yeah Facebook's scaling a whole bunch of people are using it but is it going to be a good business right was actually in fact and they were going no no we'll figure out the business as we go.

2:17It'll happen. Let's go. And so actually, in fact, you will take some scaling risk. You will figure out some questions as you're going. Now, in an age of AI, maybe there's a few other ways to kind of add to that. One is, you know, because, of course, the general discourse is around size of model and the hyperscalers, there's an over, like, okay, can I get, like, is size a model the thing that really mattered to me? And I think in a lot of businesses it won't matter. Now, there's this interesting question about how long and what shape will open source models be provided. The reason why there's a lot of open source models right now is because there's a lot of challengers who, it's the way they get into the game.

3:03But once, you know, open source providers start realizing that they're taking their compute resources, which by the way gets more expensive as it gets larger, and they're providing them to everybody, including competitors and all the rest, I think there will be a, well, maybe we shouldn't do that. Or maybe we'll try to experiment with different kind of license modeling and other kinds of things. So you can use open source models. But like, by the way, I think there will be tons of competitors, which means that it's kind of like the multi-model thing as a way of doing startup is a good thing because one of the reasons why things we're seeing with the various hyperscaler models now is like, oh, this one's better at this.

3:45Then four months later, no, no, this one's now better at this. That's a pattern that's going to continue. You have to adjust that variability. That will be a little bit of this question of how scale plays into what you're building because you have to have the dynamism and being able to shift as you need. Now, obviously, we have network effects to defend your business. You can be slower at adoption. Maybe a really key one is, you know, Apple and artificial intelligence. You know, it's kind of, you know, if anyone's got a good result from Apple intelligence, I'd be curious. I haven't yet heard one.

4:20But, you know, they have such a network effect lock-in to the business, they have time to play it and obviously amazing devices and all the rest. So anyway, so those are some reflections. Awesome. Thank you. And we're going to dig deeper into how different industries approach AI. Is Priya Morali here? Hi, that's me. I'm Priya. Oh, stand up. Yes, great. Reid Arger, thanks so much for being here. From Cobalt ID, my question is for founders building in cautious, data-sensitive industries, how can founders go about earning the trust they need to access buyer-sensitive data? and what are some of the missteps you see them take that commonly break that trust?

5:01Great question. And obviously one of the other challenges is because we have a general atmosphere of mistrust of tech companies overall and what people are doing with data and how that's being described. That's an additional difficulty on this particular game. But I did actually have an episode on this in Masters of Scale. And maybe it was the program, it was the Daniel Ek and Spotify one, might have been the build trust quickly. And to some degree, solving this kind of problem can be a competitive edge in differentiation for your business. To give you an example, we were in kind of a similar way.

5:34One of the questions early days in PayPal was, how do you get people to trust using PayPal to do transactions? And I won't name the executive that came up with the idiot idea. But the idiot idea was the buyer protection program, which was, we guarantee your purchase up to$2 ,000 and basically with no conditions. And it was kind of classic software person who had no understanding classic industry. And so you'd have people on eBay where like Ari and I would be colluding and she'd put up a plasma TV and I would buy it and she wouldn't ship it to me. And then I would get my money back paid by PayPal and the losses were hemorrhaging.

6:16And so this problem landed on my desk as a, oh my God, we have to shut this down. Maybe everyone's going to lose confidence in PayPal. Maybe it's going to go away. I was like, okay, look, how many days can you give me to solve this problem? And Peter Thiel was like, five. And I was like, fine, I will go try to solve this problem. And what we came up with was, what we did is we reconstituted the PayPal buyers protection solution. Instead of guaranteeing you$2 ,000 in every situation come hell or high water, we should switch to doubling the eBay insurance, which meant we still had a buyer protection program.

6:59And, of course, it was on eBay, so it was fair. So if you use PayPal, you could double the insurance that you got with eBay. But, of course, what you had to then do is show that the eBay insurance, where you, like, offloaded a whole cost structure on eBay because, like, you had to prove that eBay thought that it was worth doing. And then you show that the thing was actually$370 versus$250 and we'd pay the$120. And those are instances of kind of how to think about how to build trust. Part of the reason Spotify could come out of Europe and not the U.S. is because it guaranteed Denmark and Scandinavia as an area.

7:38And buying it in, so like, for example, getting insurance, Lords of London, other kinds of things as ways of doing that. Basically, it's fundamentally to building trust quickly. it's like, again, ideation, creative, but it's what's the kind of thing that will give the relevant constituencies confidence that you're holding yourself accountable to the failure points or worries they might have, you know, buying something on, you know, PayPal or anything else. And that's the kind of thing you're looking at. And a lot of people also just don't understand. It's like one of the funniest conversations I had a number of years ago is, this was the Silicon Valley person.

8:19What has Google given me for my data? And it was like, well, free search. So there's a lot of misconceptions. Not that I don't love everyone in the room. This is my favorite AI person in New York. Allie Miller, are you here? Hey. And Allie and I did a podcast. She has great ideas. It was fun. Oh, good. Okay. Well, then let me put you on the spot and see if you have good ideas for me. No, I never do. I would like to know what signals we should be looking for and perhaps on what timeline where AI is good enough at embedding itself. This is part of the software writing, but AI embedded itself inside of business and technical processes.

9:04So much so that the painfully slow human behavior of adoption and adaptation no longer matters. great question not a surprise given that you do this uh ai leadership quite well over a number of things and i will essay to try to be not dumb in my answer um what you're essentially looking for is where are there coherent loops that can move much faster that don't have organizational change organizational hiring as part of them part of the reason why you know i invested in sierra with Britt Taylor is because I think the front end to customer service is more easily modulares of all from how it interfaces with the rest of the company.

9:49Obviously, there is a period of connection. And there's a way that you can very quickly move it to being the front end to how you're interfacing with the company and that kind of thing. And so it's one of the reasons why I think the customer service thing is one of the places where it will really kind of speed up what that changes. One of the ones that I think will be interesting to see how it plays out, and one can play out multiple hypotheses, good range for multiple startups is like sales. And part of it is like, just like customer service, you could say, well, that's kind of inbound. What about outbound with sales?

10:22Of course, people might get really pissed off about being called by AI agents, might get legislation and other things, a bunch of different risks and how that plays out, you know, thing. But that's another one where a modular function where you can get that loop going. Like one of the classic things is like, you know, the Bezos two pizza teams and other kinds of things that what the coherence of a smaller group can accomplish if the group's working pretty intensely will now have a higher throw weight. And there's always, that's part of why reason why you have Dunbar number, you know, 150 is like how much do people hold in mind one interesting i think prediction actually this would be a fun thing for everybody in the room to do think about what year you will never have a professional meeting where you're not having an ai agent listening to you and playing a role um and that's like the literally every single meeting that you're doing that has any kind of professional thing by the way you You might also do it when you are having a talk to your kids, and I'm asked that a different question.

11:30I don't think that year is too far off. If that's the case, this is one of the things I wrote for the MIT Tech Review like a decade ago as an anticipation of where AI is coming. If that's the case, then the scalable coordination between teams might get a lot easier because it's literally like this team is having this meeting, and this team is having this meeting, and the AI agents go, oh, wait a minute, and immediately create notification, or even if it was like two hours later, that kind of transition and how the speed and how that operates will be interesting. That being said, the way the accelerations will happen will be where the loops close in tight ways and where people feel competitive need.

12:19So part of the reason why AI adoption has been so slow right now is that most people don't feel competitive need. But like once, for example, you know, I'm sitting here, you know, doing my coding and I see ARIA 10X-ing my speed and delivery. What now, Reed? Yeah, with, yes, exactly. And I'm like, how am I working? Oh, she's being much smarter about how she's using AI. I'm going to then start using AI, too. And that will be another thing that will drive it. But obviously, it's a very good question in a complex space. Well, I mean, a lot of Ali's work is working with Fortune 500 companies to have them adopt AI more quickly.

12:55And I could imagine that perhaps 90 % of this room is already there, can't imagine a work day without AI listening constantly. And yet, I think other people are going to be sort of dragged, A, kicking and screaming, B, only when they absolutely have to because of market need. That being said, as a parent, I mean, last week, I was having a conversation with my nine-year-old, we were talking about monarchy because of the No Kings marches. And we were explaining what monarchy was. And he goes, what are some other archies? And I was like, oh, I don't know. And then I was like, oh, patriarchy. And I explained what it was.

13:28And he goes, oh, it's so good that that doesn't exist anymore. And I was like, ooh, I wish an AI was listening to that conversation and could have just piped in like, Aria, here's how to explain patriarchy to your nine-year-old. So I think as a parenting agent, like there's hope.

13:51I think we have time for about two more questions. Esther in the front row. So this is not a political question. So this is not a political answer. Exactly. But it is an economic one. There's a lot of value created, whether it's Google or who else, by basically what's a collective asset of content on the Internet. What do you think of the idea of some kind of like mining rights or water rights or taxes that would, LLMs, not people who use their own data sets for training, but the large language models and I don't know, maybe some other things where they would fundamentally pay some kind of taxes on the revenues they generate using the collective asset.

14:38And then that would go to teachers and child care workers and people who are now desperately underpaid. I think that trying to figure out how to tax on all the data stuff, it's such a fast-moving thing. Like one of the things that currently seems like a pattern is the larger scale you get on the mixture of expert models, the less general data it actually needs. And so then which general data and all the rest is one instance. Like what the patterns are of how data fits into this is challenging. And I also believe in simplicity. I think it's probably fine to say, hey, look, because of technological leverage, some companies are going to have enormously profitable business models.

15:23Let's have a little bit more tax there and direct it towards public goods. I think it's just simpler to do it something like that, kind of wherever it is. One of the underlying tensions is because labor and capital, and then when you get services labor, services labor gets a lot less kind of scale leverage and yet services labor really matters in terms of how we care for each other and so forth. So we go, okay, as that balance changes, because one of the things that AI does is it does change some of the labor calculus towards capital. It's like, okay, should we try to figure out how to get some more adjustment to services labor?

16:02One of the mistakes I think most policymakers do is they try to be overly coding specific. I think it's almost like the mistaken software of hard variable coding. Just kind of let's try to do a general program across that. So I'm positive on that. I don't think it needs a justification of, well, these things can't work unless they use what is this currently common good. I think they live in society, right? And it's important that we have a society that brings people along. Yes, right there. Hi, Reid. Thank you so much for coming today and for inspiring many of us to start companies. The way you inspired me is by posting the video of you talking to your digital twin last year.

16:45And I thought that this is so cool and this is the future of sort of AGI interface. So my question is, could you share a little bit about your experience building this digital avatar? What were your learnings? What are some of the things that could go right about this technology? and what advice you could give in terms of blitzscaling for someone building in this space and trying to compete with incumbents like Cajun and Synthesia? Thank you. Well, blitzscaling will be more specific and hard, so I'm not going to give much of a good answer there. But, you know, it's a good question to be asking.

17:17It's like, when is your probability of the set of things of scale product, market fit, business model, etc., coming together, can you convince people to raise enough capital and go? And what blitzscaling is set as a clock is what your competitors look like. It's one of the problems that you have is if you're doing a capital spend competition with hyperscalers, that's a very difficult, risky business. You better get it right. Now, on the Read AI thing, it was funny because the idea basically came along because I have as a general principle that one of the things that people happen to do is kind of technology good, technology bad versus how do you shape technology to be good and how do you shape technology to be less bad?

17:59and one of the things that I find very frustrating about most tech critics is they go bad and you're like well maybe in some ways but the question is is always a question of sequencing dynamism which things do you fix first which things do you improve first it's like wow you just said medical agent and sometimes that medical agent's gonna say something that's wrong to this person and you're like well yeah but you have to look at this kind of thing of well how often that person have got nothing, how often will that person get wrong some other way and all the rest. So you have to kind of look at it on kind of a systematic basis.

18:31The real question is how do you shift it? And when I was thinking through it, I was like, oh yeah, and there's this technology called deep fakes, which is almost like labeled bad, bad. And the only positive case that most people can imagine for it now is like making a younger Tom Hanks, you know, playing in a movie or doing CG or something. And I was like, ah. And all the rest of it's terrible. Evil, evil, evil. And I was like, well, is that true? And I went, well, okay. Well, let's start exploring with it. Because I understand that this one's harder. Because obviously, you know, you can abound by the tremendous number of different harms and deep fakes.

19:09Everything from fiscal to misinformation, fake news, to, you know, revenge porn, all the rest of the stuff. It can be really, really bad. And so I was like, I was thinking about it. I said, well, let's just start experimenting. I mean, literally the way we kicked it off was, let's just do something and then see. And, you know, so like, for example, you watch the first one. It ends with me saying, well, I thought I'd hate this more than I do. And then as we started iterating, this is one of the things about always be thinking about a positive outcome. So last year I gave a speech at Perugia that we then had read AI given nine languages.

19:42And it was kind of because a human connection. and it was kind of amusing to watch myself speak Hindi, Chinese, Japanese, Italian, you know, languages which, like, I don't think I know a single word of Hindi. And so that was kind of an instance of positive use cases. And then it made me start thinking about, like, all right, you know, the likely thing is, like, we won't have voicemail anymore. You know, I'll have read AI that essentially answers the phone and does the initial talking and has actually a better interface for that. And I was like, oh, right. That's the kind of thing that will have.

20:17And those will be some of the positive use cases of this. And look, obviously you can argue with like, well, wait, there's so many negatives in this one versus the positives and what the issues and the negatives definitely need to be managed. But that's how I got into it. And part of it is a general principle of part of the thing you have to do as an entrepreneur is when most people think it's idiot bill and you think you have a good idea, that's potentially a great idea. That's part of the contrarian and right. The trick, of course, it's easy to be contrarian sometimes, harder to be right, but that's the kind of thing to play out.

20:45One of the really fun things about Read.ai was that we also just used commercially available technologies. We called 11 labs and hour one, and we worked with ReSpeecher and more recently HeyGen. Truly, it was not because we had a technological edge or anything else. It was just like, let's try this. Now, I will say, not surprising, Read gets asked to give speeches and go to events probably 20 times a day. and I used to just say no to all of them. And now, especially when they're in a foreign country, I say, oh, you know, read's unavailable. But if you would like Read.ai to give a three-minute speech in French, like, we can arrange that.

21:25And more often than not, people say yes. And so, you know, before they would get nothing and now they get like a little dose of like, oh, this is so cool because I'm using this technology to, again, be more human. You're not usually having someone from Silicon Valley speak in French. So it's been really fun to see all the positive use cases. Do you have a count top of mind? Because I've kind of lost count of the number of conferences. Oh, I mean like over 100. Yeah. Like literally read AI speaks way more than you do. Yes, yes. And now, actually, this is the best. No offense. I do. I've gotten at least five emails that are like, hey, can read AI speak?

22:00And I'm like, hey, it reads unavailable. And they're like, no. I said read AI. And I'm like, oh, okay. By the way, totally happy to be replaced with this content.

22:35Thank you.

22:45And a big thanks to Jennifer Whiting, Sheila Goodman, Ben Kassnoka, and the whole Village Global team, Robert Kingsley, Jerry Madlambaya, Samuel Henriquez, the Ritz-Carlton, and of course, Vincent Lucero.