An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout – Stratechery by Ben Thompson

Today’s Stratechery Interview is with OpenAI co-founder and CEO Sam Altman. I have interviewed Altman twice before, first in 2023 with Microsoft’s Kevin Scott, and again earlier this year; while OpenAI is not a public company, I am, just as I did previously, treating this like a public company CEO interview, which means this transcript is publicly available.

I actually was pretty surprised to realize it has only been seven months since I last talked to Altman; since then OpenAI has unveiled GPT-5, launched the viral AI video app Sora, and made a slew of announcements about a massive AI buildout with a variety of partners, from Nvidia to AMD to Samsung to Oracle and more. The occasion for this interview, however, was Monday’s DevDay keynote, where Altman and OpenAI launched a number of new initiatives, including Apps in ChatGPT.

In this interview, which was limited to 40 minutes or so, I did my best to cover the entire gamut: we discuss all of those infrastructure announcements, why Altman’s investor skills are particularly useful as CEO, apps in ChatGPT and Instant Checkout, Sora, and how the company thinks about user feedback. What is particularly notable, in my estimation, is the extent to which Altman and OpenAI are presenting an increasingly coherent vision about what OpenAI is building towards, which very well may be all of our future.

As a reminder, all Stratechery content, including interviews, is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.

On to the Interview:

An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout

This interview is lightly edited for content and clarity.

OpenAI’s Unifying Vision

Sam Altman, welcome back to Stratechery.

SA: Thank you.

We only have 30 minutes, I worry that if I attempted to read out all the deals and announcement OpenAI’s made over the last month, that would take up the entire time and I actually am not sure that I’m exaggerating. So to save us all some time, what is the grand unifying theory of all these announcements ranging from what you announced at DevDay to these infrastructure deals? Is there a one sentence or one paragraph answer that captures it all?

SA: Yeah, we’re trying to build very capable AI, AGI, superintelligence, whatever it’s called these days, and then be able to deploy it in a way that really benefits people and they can use it for all sorts of things and that requires quite a bit on the infrastructure side, also on the product side, obviously on the research side. That’s the one thing we’re trying to do.

I posited today on Stratechery my own version of this unifying theory. You’re positioning yourself to be the Windows of AI — and I’m the sort of analyst where that’s a compliment, not derogatory — where on one hand you provide the user interface or at least the API for all end-user interaction both in the consumer space and the enterprise, and on the other hand you’re sort of the target customer for this massive infrastructure build out. It’s sort of like all the OEMs and the processors back in the day, does that seem right to you?

SA: I always struggle with the historical analogies because I always get caught up on where there’s differences that I really care about or whatever.

The way that I think of it is that most people will want to have one AI service, and that needs to be useful to them across their whole life. And so you’ll use ChatGPT, but you’ll want it to be integrated with other services and so you need to have other apps inside of ChatGPT. We need to have an API business, because you will want to be able to sign in with OpenAI into some service that someone else has built, and you’ll want the kind of continuity of experience and you’ll want it to still know you and have your stuff and know what to share and what not to share. So we want to build this AI helper for people and that’s going to have to — there’s a few pieces that have to fit into that.

And then on the infrastructure side, this is where I’m spending most of my time now, it’s brutally difficult to have enough infrastructure in place to serve the demand we are seeing and it’s fun, it’s been an interesting new challenge for me, but there’s a lot that has to go into that.

And then on the research side, I’ve never been more optimistic than I feel right now about where our research is going and that’s why we’re trying to make such big bets to get the product infrastructure ready.

Yeah, that’s a great framing for the three themes that I want to get to.

Infrastructure Deals

We’ll talk about the infrastructure first. How does it feel to know that you can announce a deal and increase your partner’s market cap by 20% or more? This is truly the Midas Touch for anyone that gets to release a press release with OpenAI in it.

SA: So strange. This is a very new phenomenon and I think somehow we got — I don’t think it’ll last that long, I think it’s a little silly, but we somehow have to adjust to — three years ago we were like a research lab, three years ago ChatGPT had not launched and we didn’t have to think about impacts we have on markets and it’s kind of a quick adjustment for us in the recent months, but it’s very strange.

If people want exposure to the AI build out — don’t call it a bubble, we’re use another “B” word — do they basically need exposure to you? That seems to be the flip side, but in the context of if you’re going to be this unifying layer, people have an AI that’s with them and you’re going to touch and influence everything from infrastructure to apps, then aren’t you the bubble, and by bubble I mean it complimentary, but that also actually cuts both ways.

SA: We have a lot of very big, well-funded, very competent competitors. I don’t think this one will be a winner-take-all market or even close to it.

Do you think it’s going to be winner-take-all maybe though on the consumer side?

SA: I would bet not, I mean obviously it would be nice for us to get a lot of the market share, but I would bet not. We’re certainly going to try to build a very heavily used product that adds a lot of value to a lot of people, and we’ve got ambitious plans there, but I think you will see AI — my favorite analogy for AI, my favorite historical analogy is the transistor, I think we talked about this once before, but I think it will just kind of seep everywhere into every consumer product and every enterprise product too.

To the extent these dynamics are true, is actually the worst thing you can do is to not spend as much money as you can or as much money as people are willing to give you?

SA: I don’t know if it’s the worst thing you could do, we’ve been trying to explain this to the world for a while, but I think the world did not take us literally or seriously on this point until recently. We are going to spend a lot on infrastructure, we are going to make a bet, the company scale bet that this is the right time to do it. Given where we are with the research, with our business, with the product, what we see happening and is it the right decision or not? We will find out, but it is the decision we’re going to make.

Yeah. Byrne Hobart wrote a book recently, Boom, that expanded on Carlota Perez’s bubble investment dynamics, and I think his important addition was this idea that you need a lot of simultaneous parallel investments and developments, which bubbles make happen at the same time. Is that a burden you feel that you need to force? It’s not just that stuff needs to happen at the same time, it literally can’t happen sequentially over time?

SA: Yeah. You have to make the electrons, you have to build all of the physical infrastructure, you have to get the power equipment and all of the — everything that’s outside the data hall. You have to get the chip fab capacity on the racks built, you have to have the consumer demand be there and the business to pay for it. It’s a lot of stuff all at once.

Yeah. What do you think about chips? You just announced this deal with AMD. Did Jensen Huang know that Nvidia is going to become an indirect investor in AMD? (laughing)

SA: I don’t know.

Well, with this, the problem with this is both Nvidia and AMD are sourced at the same place, so there’s another solitary entity in the value chain, which is TSMC. Do you see a need and responsibility/opportunity to expand the market there as well? Is this something where when it comes to the question of Intel—

SA: I would like TSMC to just build more capacity.

What did you think I was asking, about multi-chip suppliers?

SA: Do I see a need to get TSMC to expand their rate of investment in more capacity?

Got it.

The Investor Mindset

These deals are worth an astronomical amount of money, I think a trillion dollars was what the Financial Times just calculated. Who do you expect to pay for it? Is this a matter of what these deals are about, you guaranteeing you’ll buy the output of it and you need these companies to invest?

SA: Yeah, I expect OpenAI revenue to pay for it.

But in the meantime, these companies are going to have take on debt presumably. Do you see yourself almost as a financial guarantor who’s helping them secure better interest rates? You say, “Look, there’s a guaranteed purchase agreement, this is going to be used”.

SA: Yes, but we will help with financing. We are working on plans to be able to help with the financing these companies need at this kind of scale ahead of revenue. We have some interesting ideas there, so we’ll try to sort that out.

You were long thought of as one of the all time great money raisers in Silicon Valley. Did you appreciate the extent to which that skill was going to come to bear in this space?

SA: There were all of these things early on in my career that I learned, I won’t even say got good at, but got decent, okay at, that I kind of didn’t think would be such good prep for having to run OpenAI later that have turned out to be super helpful.

Is that top of the list?

SA: It’s one of them for sure.

What else is on the list? I’m actually curious about that.

SA: Well, first of all, I think if I got to do the thing I was naturally good at, I would be an investor, not an operator. I am extremely ill-equipped in many ways to have to run a big company, but I have the training as an investor and the framing of really understanding how to think about capital allocation in a world where you can have these crazy exponentials, where you’ve seen the magic a bunch of times and how you pick projects and people and how to think about allocating capital in a world like this.

There’s some sense in which OpenAI is betting on a series of startups, but they’re all internal or maybe sometimes they’re products, like the Sora thing. That was unusual, I wouldn’t have thought that was good training for running a company. In fact, I would’ve thought it was probably bad, but it has turned out to be useful. I can do a whole episode on all of the things here, but that’s another example.

Yeah, I think that’s actually quite interesting. One of the things is to understand the dynamics of infinite upside versus capped downside, that makes a ton of sense. But in that context, what are things, if anything, to the point that actually you should spend as much money as you can get your hands on. Are there things that you do say no to or that are not the right thing to do?

SA: A lot. So that’s the thing that’s different than doing a lot of early stage startup investing, there are dozens of product ideas that I would like to say yes to, there are dozens of new things that we could go do, and we try to do a very small number of them and that’s a little sad, but I think it’s just important to do the things that we think will matter most really well.

I’m curious about hardware in this regard. You did a sit down with Jony Ive yesterday, and this is very exciting, the potential entrant of a new hardware maker.

I guess the question I have is if ChatGPT or OpenAI or whatever is going to be your friend everywhere, that needs to work across devices, it needs to be not limited to whatever device you make. Did that enter your thought process, a concern that you might get anchored on your own hardware or scare off other potential partners?

SA: It has to work everywhere. It needs to work in browsers and on everyone’s mobile phone and all that. Clearly, it has to do that as do other things. I don’t know, I’m like a real device nerd and I always have wanted to make—

I know, but this is the siren call, is everyone’s a device nerd and then they get their device and they get a little distracted from, “Actually no, we need to be everywhere”. There’s another very large company in the Valley who may have been trying to poach some of your researchers that I think can get stuck in this trap.

SA: Let me have one crack at this one, you’ve got to have a little fun.

No, I think we have the chance to do something really extraordinary, it won’t obviously replace other devices and we have to work across all of them, but the quality of thinking on what new hardware can be has been so…

Stagnant.

SA: Everyone has, yeah, stagnant is a good word for it, everyone has the same few ideas. They’re all trying them, it’s all like — the iPhone I think is the greatest piece of consumer hardware ever made and so I get why we’re in the gravity well, but I think AI, it certainly illustrates the possibility of new things.

One of the surprising deals, just to get in the weeds a little bit, was with Samsung and SK hynix. Obviously memory is a massive constraint as far as building out these chips in the future, is this sort of tied into the AMD deal? That’s one of the questions people had, “Where is AMD going to get memory for these chips?”, or are you still a building your own chip?

SA: Give us a few months and it’ll all make sense and we’ll be able to talk about the whole — we are not as crazy as it seems. There is a plan.

I think it all makes sense, to be honest, which kind of scares me about this interview. I feel like there’s actually quite a bit of coherency to what you’re doing even though it’s sort of all over the place.

SA: Why is that scary? Isn’t that better than the alternatives?

Well, no, because I think there’s maybe a confirmation bias risk. I’m talking to you and I’m like, I feel like I’ve gotten more clarity about you. I think I was probably wrong to complain about you having the API, even though you always run into capacity constraints, and that’s always one of my things that sort of bums me out is like, “Well, if you focused all your capacity on XYZ, you’d be in better shape”.

SA: We’re getting a lot more capacity!

Yeah. Well to your point, the capacity is going to come. It’s almost like, wait, it’s like complaining about processors are not being fast enough — it’s going to get faster, the capacity is going to be there and I have more appreciation for the, “Hey, why not try to be everywhere, do everything?”. When you have the opportunity, that opportunity comes along once in a lifetime.

SA: So first of all, I do feel like this is a once in a lifetime opportunity for all of us and well take the run at it.

This is the ultimate YOLO, YOLO AI, maybe that should be the name of a new article.

SA: But I think we have a vision, which I kind of talked about, and if you look at everything we do through that vision, when we are done with this new set of deals and get to explain here’s where we’re going and why, I think it will all make sense. People may say it’s too ambitious, it’s too risky, whatever. But at least why we’re doing the stack that we’re doing of stuff I think will make sense.

Apps in ChatGPT

So you’ve talked in the past, including with me, about how ChatGPT was a surprise. You actually I think referenced that just a few minutes ago. Just looking back, and I think I asked you some variation of this question before, but I’ve been dwelling on it more recently. When did it go from surprise to, “This is a thing”, to, “Actually, this is very leverageable”, because we have all this demand, we can actually, as an example, the API is good, the identity is good, people can sign in with ChatGPT and get their memory, all those things there. But just this week it’s like, “Actually, why don’t the app developers just put their apps in ChatGPT?”. Instead of building weird integrations and stuff that always falls through the cracks, it’ll just be all in one place, that’s where people will be. When did it become clear that you had the power to make that happen?

SA: There were two surprises. One was that — well, I thought ChatGPT was going to do well, but not as well as it did. So one was that ChatGPT did extraordinarily well quickly.

The second surprise was that very big tech companies that I think of as fairly competent on the whole, for whatever reason, kind of misstepped in the period after that and it let us build up some leverage before people got their acts together.

But then after those two surprises, the strategy that we wrote down that we have been executing for the two-and-a-half years since those surprises happened has kind of gone exactly like we thought it was going to. I mean, it takes a while, you’ve got to do it step-by-plodding-step, but we’ve been thinking about this infrastructure buildup for a long time. We’ve been thinking about apps and ChatGPT and how we’re going to unify the API and the consumer business for a long time, so that part’s all kind of gone like we thought.

When you make the comment about other tech companies, I think at the beginning most people would think of Google. I think Google’s gotten their act together to a decent extent, was the actual company Meta?

SA: Well, I don’t want to name specific companies here.

I thought you said you wanted to have fun! I just brought that in just for you.

SA: Google’s doing incredible work right now, but clearly they didn’t in that first period of time where ChatGPT got to market.

Yep. In the long run, for ChatGPT specifically, is consumer or enterprise a larger market? And have you been surprised at how deeply ChatGPT seems to be penetrating enterprise?

SA: There are some differences, but it’s, to go back to Google, it’s not like you use Google at home and a different company at work.

Yeah, that’s a great analogy.

SA: Enterprise needs different security stuff, different access to data. Maybe in the consumer world you care more about how good is healthcare and in the enterprise world you care more about how good it is at writing code or legal documents or whatever. But I think they’re going to converge more than people think, and you’ll have the AI tool you use.

Yeah, I think Google is the perfect analogy, there’s really no difference, you just use Google for everything. It’s at work, on your work computer, it’s at home. No employer is blocking access to Google and they can have Bing as the default, but everyone’s still going to google.com.

So what about for OpenAI as a whole? And this goes back to the API business, or do you think about the API as new companies, new startups building on it or this being an enterprise offering? Or is this the place where there’s still a real role for say Microsoft and Azure and things along those lines?

SA: Again, I think of it as we’re trying to build this one AI service that is helpful to you everywhere, and I don’t think of it as like ChatGPT Enterprise versus the API. I think it’s like, you’ll think about you have this relationship with this AI thing, sometimes you use it inside of ChatGPT, it’s doing a bunch of stuff for you. Sometimes you’re using Codex in your terminal, sometimes you’re using an API in some other service.

But I think in a couple of years it’ll look like, “Okay, I have this entity that is doing useful work for me across all of these different services”, and I’m glad there’s an API and I’m glad there’s ChatGPT and I’m glad there’s my new device and everything else, but you’ll feel like you just have this one relationship with this entity that’s helping you.

You had a great list of partners on stage with the apps in ChatGPT, you’ve taken a few swings at this. I’ve been, I think, enthusiastic about all of them, more, I think in part because I can see the overall vision. You had the connector or I can’t remember what the first one was, plugins or something, and then you had GPTs.

SA: ChatGPT plugins, that didn’t work. GPTs actually did work, GPTs get a surprising amount of usage, but inside a company or someone for their own workflows or whatever. But hopefully this works better and if it doesn’t, we’ll keep trying.

Well, with these apps in there, is there a sense on your side it’s like, just to go back to the, “We have all this usage, it’s going to be better if I can just use Zillow in the app”? This idea of going somewhere else and Zillow says, “Oh, I’d rather them be in our app, we spent so much time on it”, do you feel you have the power to dictate, “For the user, it’s a better experience and because the users are here, if you’re not there, someone else will be” — and you’re sort of able to, dictate sounds bad, but if it’s a better user experience, that’s better for everyone else?

SA: No. Here’s another place that I think my early career training was useful. There was a version of this we could have done where it was a better user experience, but terrible for the partners.

What would that look like?

SA: Well, I mean, on that Zillow example, what if you just said like, “Hey, ChatGPT, find me all of the houses that meet these things”, and we said we’re going to control the UI.

Oh, right. So there’s no even a presentation layer of the Zillow app, you’re just getting the results.

SA: Yeah, yeah. But I felt really strongly that when we do this, it’s something that the whole ecosystem benefits from, and specifically that new startups can rocket into existence because of it. So we did this in a way where you very much have the relationship with the other site. You’re calling them by name, we’re suggesting them by name, they’re taking over the UI, they’re linking their account. So I think there was something we could have done that was maybe slightly more user-friendly, but not good for the other companies, and I really didn’t want us to do that.

Instant Checkout

Yeah. I think one thing that’s interesting about Instant Checkout, there was some, speaking of analogies, there was discussion when you launched that about how some of the social networks have abandoned in-app checkout just because it didn’t convert as well, the e-commerce sellers want them on their site. Is that a matter of, that’s almost like a bit of incumbent overhang, they have certain things that work and you feel you can do it better, you’re providing so much value to the long tail that’ll be worth it and if you’re going to prioritize who offers Instant Checkout, assuming everything else is the same, you’re not going to be biased in your results, then folks have no choice but to get on board?

SA: Two things there. One, I don’t know why the other sites have abandoned instant checkout, because I think it is one of these things that is just like, as long as you let the merchant have the direct relationship with the consumer, which we do, I think it’s just better, that one’s just better for everyone, it’s a better user experience, it’s better for merchants. That one, maybe we will learn some lesson that everybody else learned and we’re just wrong about this, but that one seems really good.

On the second part of the question. I was going to say, if we ever start doing anything that is not our best effort to get the user the best answer we can give them to their question, trust in ChatGPT, which is extremely high on the whole, would fall precipitously. So we have such an incentive not to do that.

There is a reason people love ChatGPT in a way that they don’t love other large tech companies’ products and I think it is, even when ChatGPT screws up, hallucinates, whatever, you know it’s trying to help you, you know your incentives are aligned.

Yeah. I think what occurs to me, I’m extremely excited about this, I love anything long tail related. My business is long tail in many respects, and one of the reasons I was always a big Meta defender is because I think the value they provide, the long tail, is huge. It lets merchants find customers, show them products they never would’ve known they wanted otherwise. One of my surprises about moving back to America, I thought Instagram’s ads were good, now they’re really good, it’s actually unbelievable how much stuff I’ve just bought for the house and things like that.

It feels like ChatGPT provides the inverse — it lets consumers with a vague idea of what they want actually hone in on specific products that they didn’t know existed, and the way I think about this is, because you want to be trusted 100%, you’re not going to take anything really from those transactions. If I find a customer and they really want me to go to the website to transact, you’re going to be fine with that. Tech companies throw off a lot of consumer surplus, that’s fine. Where you can monetize is if it’s a product, to your point, that’s available on 10 retailers, well, if you’re a retailer that uses it to Instant Checkout and ChatGPT gets a little bit of that, then they’re the ones that are prioritized. I mean, that’s the way I’m thinking what you’re doing.

SA: First of all, on the Instagram ads point, that was actually the thing that made me think, okay, maybe ads don’t always suck. I love Instagram ads, they’ve added value to me, I found stuff I never would’ve found, I bought a bunch of stuff, I actively like Instagram ads. I think there’s many things I respect about Meta, but getting that so right was a surprisingly cool thing for me. Other than that, I viewed ads on the Internet as sort of like a tax.

Well, I think that’s the problem is I think search is mostly a tax. Usually the organic results will have what you want, and then I’m going to buy ads to be on top. I’ve always defended Meta, I’m like, I think actually this is the ad model we should be happy about.

SA: I agree with that.

So how do you think about your possibilities with business in that context?

SA: I mean, again, I believe there probably is some cool ad product we can do that is a net win to the user and a sort of positive to our relationship with the user. I don’t know what it is yet, I’m not like, “Here is our ad model” already.

But affiliate seems like a clear win, it’s not like you have to worry about cannibalizing your ad business.

SA: Yeah. That seems like a clear win.

Sora

To go back to the ChatGPT surprise angle, in your post you, meaning OpenAI broadly, I’m not sure who wrote it, you expressed much more confidence about Sora being a hit and I sort of pooh-poohed that, I was totally wrong, I totally missed that. But do you feel like you actually have a repeatable hit creation mechanism in place?

SA: First of all, the teams here are super talented and the best way to have a repeatable hit creation is to have really talented people that are going to do the work and do great research and build good products around it.

But second, Fidji Simo joined us recently and she is an incredible leader, and she’s incredible at a lot of things, product only one of them, but exceptional at product. I’m both a little sad or nostalgic or whatever to be handing over all the product stuff, but I got a lot of other things I’ve got to go do, so it’s fine. But having someone who is so much better than me thinking about the hit after hit after hit she will deliver, I’m very excited for that.

So you mentioned the brand halo bit about people trusting ChatGPT, I think that’s such a huge deal. If ChatGPT messes up, it’s like, “It’s okay, you’re trying my little friend”. How important is that in this hit production mechanism? You launched Sora, people assume it’s going to be cool, they’re going to give it a chance, they’re going to forgive any shortcomings. You have competitors that are maybe at a very different boat in that regard.

SA: I don’t think — I think it’s just a good product, I think we understood the key, I think we got the key dynamics right there. OpenAI is not a very well-known brand, ChatGPT is an extremely well-known brand, but the fact that the Sora logo is the same shape as the ChatGPT logo, probably everybody missed and slid by.

I missed that, didn’t notice.

SA: Probably not that many people outside of the tech bubble, hopefully some of them know. But no, I think that one, we just got the dynamics right and so it spread because people thought it was cool. We’re not going to launch a ton of — when we make a device, I think people will buy it because it’s cool, not because of the brand halo, but I’m sure it helps us a little bit.

Did you give any consideration to putting Sora in ChatGPT, or was it always clear it was going to be a different app?

SA: We did think about it, but it’s such a different frame of mind. ChatGPT is such an intensely personal and private thing for people.

To add in sort of social characteristics, and that would be like-

SA: I think it would’ve felt like, “Oh, should I still be telling this about my deepest problems?”.

Right. Does Sora massively accelerate the need to monetize more quickly? I mean, I can’t imagine how much it costs to do. Is it the answer to the problem it created, like a much better ad potential there?

SA: So this is another thing that’s interesting that you learn as things go. I’m not sure that, I mean, there’s probably some good ad model for Sora, I can think of a lot of cool ideas off the top of my head. But there’s so much usage where people are just making funny memes to send to their three friends and that there is no ad model that can support the cost of that kind of a world. So there will have to be like, I think, people are just going to have to pay for generations in some cases.

If we’ve been serially undervaluing — you have talked about, people in AI talk about, “AI is going to help people be more creative”, and it sounds like the boilerplate that you throw out to justify what you want to do along the way, but maybe that’s actually true. I feel like that’s one of my takeaways from Sora is, I wrote actually when DALL-E came out about this unbundling of having an idea and substantiation of the idea, and fitting that into broader framework of the printing press and oral to reading, and all those sorts of things. And I feel like, despite the fact I wrote that, I under-appreciated that. Did you realize that was the case all along?

SA: Again, another good training from the startup investing days. There is so much latent creative expression demand in the world that if you give people tools that help them do that a little bit better — maybe it’s really hard to make a great TikTok video the old fashioned way or to do whatever, but if you give people tools that let them go quickly from idea to creative output, that hits at some very deep human need and I’ve seen it work again and again and again.

Is that a human need that exists independently of sharing with other people? Sometimes you just want to create to create.

SA: Sometimes you do. But people like looking at other people’s art and people that make art like showing it to other people. So there is something about the collective appreciation or something like that. You like to create for other people.

This is another thing why I’m not worried long-term about the AI impact on jobs. We love to do stuff for other people. We love to be recognized and appreciated by other people, and so we’ll find things to do to be useful to other people.

But yeah, on Sora, there were a few dynamics that we observed that were interesting, but one thing that was clear to us from the very beginning, and one of the reasons we were excited to do this and didn’t think it was just going to feel like a mindless slop feed, is that we hoped that if we could bring the activation energy of creating down dramatically, a lot more people would do it than normally create, and I haven’t looked at the numbers in a few days, but in the first few days it was like 30-something percent of active users were active creators, it was just incredible.

That was my takeaway. There’s that old 90/9/1 rule. I’m like, actually, maybe this rule is just a function of too high of a barrier to creation, actually a lot more people than the 1% would create if they had the possibility.

SA: And I totally agree and a lot of the things we really put effort into the product was, with other AI video creation tools, it’s been hard to create something good. It’s got to do a lot of work and we tried to make the model really good at taking what you wanted and creating something good out of it and I think that really paid off.

Do you think that the shift in social networks to actually not being social networks at all, just entertainment feeds pulling from everywhere, is there almost like an umbrella to be social again?

SA: It’s funny, I always wish we could do more things, I wish we could do a great entertainment product and a great social product, and I think they’re pretty different. There’s obviously some crossover.

Which one is Sora? Is that a social product or an entertainment product?

SA: It’s mostly an entertainment product, but there is more of a social use case in these small groups sharing funny memes of each other than we thought, than I thought. But all social products have sort of drifted to entertainment products and messaging is the new social right?

Yup.

SA: I think there is a great social product to go do again, but the economic incentives of the world pulled them all to B2B entertainment products.

Right. But if you’re starting from scratch then you have a long ways to go to worry about those economic incentives.

SA: Maybe we’ll go do a great social product next.

Copyright and AI

Your last two viral moments have arguably come on the back of copyright holders. There was the Studio Ghibli image generation and then Sora complete with Pikachus and SpongeBobs and sometimes Pikachus and SpongeBobs together. You ultimately quickly clamped down on both, but is this a real validation of the, “Move fast and break things” thesis?

SA: No. We obviously talked to lots of rights holders before Sora and in image gen we still do a lot of things we won’t do in video, and this is like you’ve got to respond to the co-evolution of society and technology together, but video hits people, particularly rights owners very differently than still images, it turns out.

Why is that?

SA: I don’t know. They’re very emotionally different things. If you make a funny image of someone versus a real video, the video feels much more real and lifelike and there’s a stronger emotional resonance. Rights holders want a different approach, most of the rights holders that I’ve spoken to are actually extremely excited to get their content in here, they just want to be able to set more restrictions than they would need for images because videos feel different.

So I predict you will see right now there’s a conversation about maybe, “I don’t like content in videos”, I predict in another year, maybe less or something like that, the thing will be, “OpenAI is not being fair to me and not putting my content in enough videos and we need better rules about this”, because people want the deep connection with the fans.

Well, the other thing that occurs to me is you talked about you’re less worried about jobs because people like being social and doing things for each other. Another thing, you take these viral moments, copyright holders are worried about AI, but there’s a reason why they’re the first thing people go to, it’s a common item that is sort of a totem pole for people to touch on and it seems like as AI creates more and more customized content, things that are common and commonly recognized are going to increase in value and it seems like a big opportunity.

SA: I’ve got a biased sample here obviously, I’m sure a lot of people don’t want this, but I’ve had way more rights holders reach out to me with, “This is going to be good for me and I got to make sure that I get enough gens” than “I don’t want it”. Now they all want, not all, most want to set rules and say, “I don’t want my character to say this offensive thing or whatever”, some are just full YOLO. But yeah, I think you will see the center of gravity shift here to people who are like, “This is going to accrue value to me, not the other way around”.

Feedback and the Future

OpenAI, I think, can sometimes vaguely overhype everything on Twitter, which I think can rub some people the wrong way. I feel like maybe, I don’t know, maybe helps explain why I was so wrong about Sora. But is this really a matter of, are people overhyping, or is it just you have such a large user base that it takes so much longer to get stuff out from the lab to the real world, and when it gets to the real world, maybe it’s nerfed or not nearly as capable as it might be? I guess this is an infrastructure question if I really squint.

SA: A few things. One, we should probably hype less on Twitter, we just get excited. Two, if you look at what we have delivered relative to what most people would’ve expected from five years ago, got to say it’s been likely at least somewhat impressive.

You deserve the benefit of the doubt. I think that’s fair.

SA: I don’t want to quibble on the exact definition of the Turing Test. By the popular conception, we kind of have passed it and most people in 2020 did not think that was going to happen. So now if in the next five years we really deliver on AI, there’s discovery and important new science and we were hyping our progress and our excitement a little bit too much, I think we deserve some grace there, but it’s annoying and we should do less.

How do you think about feedback mechanisms in general? I think one of the real challenges of being a massive consumer app in the age of social media is a very small portion of user base can be unhappy but can be very loud and seem very large. How are you thinking about that versus looking at the data of what’s actually working? This seemed to happen with ChatGPT-5. ChatGPT-5, the simplification, was fantastic. You shared numbers about thinking being used by way more people, which I think was a big problem before, particularly for free users, but then there was quite a bit of backtracking and some changing stuff around in a very short amount of time.

SA: Well, first of all, I think it’s good, not bad, that we listen to users and look at data and change things quickly. We do both, we listen to users now.

Right. I think there’s a balance there between data versus listening to a limited segment of self-selected users.

SA: I care more about the data, but I believe in the principle that if the anecdotes and the data disagree, you’ve got to go look extra hard because the anecdotes are very often right.

Is online sentiment almost harder to manage these days because the way narratives form and spread on platforms like TikTok, and you don’t want to get the narrative “ChatGPT-5 is bad” established?

SA: It’s definitely a challenge. I mean, there was a real learning for me with GPT-5, which is, the way that active people on the AI corner of Twitter use AI and the way that the normies in most of the world use AI are two extremely different things.

Right.

SA: And building one product to make them both happy is possible, but not something that we have taken seriously enough.

Along those lines, do we underrate still the potential for subscriptions generally? I think tying to the creative bit, people are more creative than we realize. Are people more willing to pay than common sentiment as thought?

SA: I do think ChatGPT has shown that.

I think so too.

SA: Now, how far we can push that, I don’t know, but should be at least somewhat of an update.

Yeah. This is all amazing, are we going to have enough electricity for all this?

SA: As I said, please give me the grace of a few more months to finish. I promise to tie the whole thing together.

Okay. The next set of deals to listen for. Sam Altman, it was great to talk to you again. It was amazing month, amazing launch. Congratulations.

SA: Thank you very much. See you.

This Daily Update Interview is also available as a podcast. To receive it in your podcast player, visit Stratechery.

The Daily Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.

Thanks for being a supporter, and have a great day!