"Bubbles have almost always preceded some important technological shift”: OpenAI's COO

0:01Allie:This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome? That's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up required, compatibility and availability varies 18+. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications and more.

0:38Allie:Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a$75 sponsored job credit at Indeed.com slash podcast. That's Indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs. Bubbles have almost always preceded some important technological shift. Hello, hello. Welcome to the Termsheet Podcast. I'm Allie Garfinkel, senior writer here at Fortune. And every week on the Termsheet Podcast, we go over the latest deals, news, and insights in venture capital, private equity, and startups.

1:14Allie:Every week, this is where you can hear from some of the most exciting figures in the private markets. it's a big open ai news week for more reasons than one the company announced a data infrastructure partnership with chip maker amd that's worth billions of dollars and sent shares of amd skyrocketing in the aftermath almost simultaneously the company unveiled a new range of product announcements around coding agents apps and an api update all structured towards two things integration with partners like Canva or Zillow, and simplicity, the idea that anybody, including me, could say make an agent.

1:51Allie:And of course, a few days ago, the company also launched its video model, Sora 2. Now, speaking of OpenAI, on to this week's guest. Now, this week's guest is a surprise guest. This was a last minute opportunity. I got on a plane to come to San Francisco early just to see him. This week, we have OpenAI COO Brad Lightcap. Now, Brad is among OpenAI's earliest hires who wasn't on the research side. Brad knows Sam Altman from Y Combinator. He has been with the company through almost every step of its meteoric rise. Now, OpenAI has about 800 million weekly active users on the consumer side. But there is a sense that in order to truly succeed down the line, the company must also win on the enterprise AI side.

2:37Allie:And Brad is the guy tasked with making that happen. Brad and I sat down to talk about the past, present, and future of enterprise AI and startups. Here's Brad. Brad Lightcap, welcome to the Termsheet Podcast. Thanks for having me. We are, there are a couple firsts here. I'm at my first OpenAI Dev Day. This is our first live recording in front of an audience. This is the first time I've sat in this gaming chair, not in my apartment. So we're hitting a lot of, we're firing on all cylinders right now. Brad, you were telling me a funny story before this started. So let's go there. What was the funny story?

3:10Well, so the prompt originally was, I think, what was the craziest thing you've ever asked ChatGPT. Yes, our audience asked. And I couldn't come up with an answer fast enough. But I do remember this one thing that someone told me at a very large enterprise that shall remain nameless. There's a very senior executive at this enterprise who, this is early on, he was probably the first really true early enterprise adopter of ChatGPT that we encountered. and he had to give a speech in front of the entire company, and he thought it would be funny if he gave the entire speech in the style of Dr. Seuss.

3:43And so he wrote an entire speech for a large institution that he gave a...

3:47Allie:How long was the speech? Long. It was like addressing the company at their annual meeting, and he gave the entire speech in the style of Dr. Seuss. And he was like, I thought it was hilarious, and everyone else eventually thought it was hilarious, and I thought it was great. Everyone else eventually thought it was hilarious? I think there was confusion at first. There was initial confusion. This was before ChatGPT was like the thing it is now. So a lot of people had no idea how or why I think he had done this. But he just thought it was great. Oh, so it was the sort of thing people were like, how did he do this?

4:15Allie:Why did he do this? I'm sure this is, we'll talk about many other enterprise use cases. But let's start with DevDay here. Because implicitly what we're talking about here is enterprise, right? Which is, let's start with Dr. Seuss, why not? But what's true today that wasn't true at the last DevDay? I would say the biggest things are the models which have just gotten significantly better. And I think there's a whole lot of product engineering around the models now. There's all this kind of scaffolding and infrastructure that's been built, which is great. And I think all that stuff is important. But I think what somewhat gets lost is just the fact that every turn of the crank of the models just completely expands the aperture for the types of things that you can do and you can build.

4:56And so Dev Day for me is always this kind of cool annual check-in of like what types of things are possible now that like weren't possible a year ago. And for me, those are things that are very emblematic of what you've seen here today. So AgingKit is a great example of like, you now have ways to very dependably and reliably kind of build and construct agents for really any kind of arbitrary task that can kind of be encoded in a set of steps in a workflow on a computer. And what do you need for that to be true? You need models that have kind of a generalizable reasoning ability to be able to think through a problem.

5:28They need to be able to call tools. They need to be able to correct their own mistakes. They need to be able to, you know, use the web, write code. There needs to be kind of a mix of their kind of, you know, almost stochastic nature with kind of the way that you kind of imbued like a little bit of determinism into the business process. And so that enables something like AgentKid to get built. And I think if you tried to build that, you know, even a year ago, like I don't know that it would have been possible. I mean, people forget like 03, I think we came out with in December of last year. So not even a year ago.

5:59And now you've got models like GPT-5 that enable that type of thing to get built. So that's what's exciting for me. And I suspect next year we'll see something else amazing.

6:06Allie:Yeah. To that end, there's been a lot of news today, including the AMD deal. Why announce the AMD deal today in addition to all of these other product announcements? I don't make the common decisions. But I think for us, the infrastructure component of this is almost in some sense becoming as important as the core research. and the two go hand in hand, obviously. And so there's almost maybe something, I guess, kind of poetic about there being both of these things getting announced at the same time. One, this gigantic investment on the infrastructure side that we intend to make among many others.

6:39And then, you know, here on the other hand, this kind of, this way that we can talk about the platforms that we're building and the things that every developer in this room and many, many millions around the world can build. And this kind of almost supply and demand a little bit like real time happening today. And I think that's cool. So I didn't plan it that way. This is a little bit of like ex post, you know, revisionism. But I thought it was a, you know, it's a cool juxtaposition of like, you know, we want to invest in exactly this. And this is the afternoon in some sense is a reminder why we do the morning.

7:09Allie:Well, and it's interesting because, you know, AMD stock shot up, right? You saw this, right? And OpenAI has the power, it seems right now, to make announcements. And then the stock market reacts despite you being a private company. How do you think about that part of the equation? Yeah, look, we don't ever think about things with those types of intentions. I think what we really focus on is what can we control? We can control our models. We can control our product. We can control the amount of infrastructure that we invest in. We try and figure out ways that we can be this kind of conduit for where there's demand for AI with the supply of AI in the world.

7:46We think we're dramatically undersupplied relative to the amount of demand out there right now. I think there's probably many multiples of latent demand that have yet to be captured anywhere in the world. And so, you know, I don't I don't pay a lot of attention to the market day to day. I kind of think more on like a 10 year time scale of like, what is the thing that we actually think is going to be true in 10 years? And for us, you know, as Greg and others have alluded to, it's that the world is kind of dramatically rate limited right now by intelligence. And our job is to build intelligence and see what happens after that.

8:15Allie:So what do you think is going to be true in 10 years? I think a lot of the economy will have kind of been rebuilt on these rails that are these kind of AI rails. I think, you know, at the fundamental level, like AI really is these kind of units of intelligence that can be packaged, directed, and configured for productive output, right? Whether that's helping you individually with something or helping an enterprise achieve something or a startup that is building a new product. I mean, you know, there's like almost in some sense kind of not enough intelligence in the world. And I think the idea that we always had was like, if you could increase the amount of intelligence and make it really cheap, then something dramatically good should happen on the other end.

8:51Obviously, we want to be mindful of kind of where there are risks and harms that present. And we try and be really thoughtful about that. But I think on the whole, we're optimists. And we feel like 10 years from now, if there's some, you know, increase in the amount of intelligence by as a factor of a million, 100 million, a billion, something like that, it's going to be a good thing for the world.

9:10Allie:One of the things I find really interesting about where you sit is you have a unique vantage point on the market, right? You see how startups incumbents, developers all interact. What's the state of that landscape right now? You know, candidly, I think it's like fragmented. And I spent a lot of time here. It's kind of what I mostly spend my time on right now. And so I spent a lot of time thinking about why that is. And I think a few things are true. So one is we're still really, really early. You know, these things almost have a way of kind of being kind of weirdly chaotic at first, where everyone is trying to kind of build every piece of infrastructure.

9:44And it's like, no, use my thing. no use my thing and all these kind of intermediate layers start to to to get built and you're kind of like i don't which of these am i using which of these is even real two minutes later and then

9:54Allie:there's a layer of something you're like why is this here again yeah like it you know everything kind of gets built and rebuilt and um you know what's old is new again and so on um and so this is kind of a natural thing for i think a like a technological phase shift when you get this type of thing i mean it was true in the early internet um and you know i think is going to be equally true here. It's just the market trying to organize itself, right? So it's people responding to where they think there's demand, the kind of shape and areas of demand in the world are moving, especially in this world at high pace.

10:24And so people are just trying to kind of figure out exactly where they can build. Obviously, you've got the kind of model layer here, you've got the infrastructure layer underneath. Those have some level of consistency, although I would say even by all definitions, those have been frenetic in the way that they've changed over time. And so I think, first of all, eventually that stuff will start to settle, right? They'll start to be common layers, common rails, repeatable solutions that people use to start to build infrastructure and core kind of canonical pieces that need to get built for other things to start to emerge.

10:57So that's actually thematically been a big part of what we're trying to do here today through things like the Apps SDK, through Agent Kit, and then obviously all the kind of things that exist in our platform, things like agent of commerce, right, and the agent of commerce protocol. So what we're trying to do is actually build a lot of this kind of middle-level infrastructure that make the rules clear for how everyone can build and transact so that the applications can get built and these things can start to become useful. It takes a few years, though, so, you know, we'll need a little more time.

11:26Allie:Well, because the thing I was wondering is because a lot of the narrative right now is around, oh, it's actually easier to build than it ever has been. Like, I was actually sitting there thinking maybe I should build myself an agent. I have I cannot code. I have never. Well, I'm actually going to try. Yeah. So we'll report back. But it's the sort of thing. Does the ease of building create more fragmentation or do you think over time it's actually going to help consolidate value? I think it creates kind of more opportunity. And so I would say, you know, it's basically kind of pushing empowerment out to the individual.

11:55I think, for example, one of the great things that happened coming out of cloud was the dramatic empowerment of individual software engineers. All of a sudden now, you didn't need to raise a gigantic amount of venture capital. You didn't need to stand up your own payments infrastructure. You didn't need to stand up your own version of everything. You could just leverage a common core underlying infrastructure that can be called through APIs. It can be called on demand, paid for on demand. And so the fractionalization of the infrastructure layer is almost in some sense what enabled this kind of like innovation at the top layer.

12:26And so I think you're going to see something happen here where you've got a highly fragmented, but what should ultimately come together as a very protocol kind of infrastructure layer that ultimately enables a lot more building at the top layer. AI thematically, one of the things we see is it's very empowering people. It's very empowering individuals. And so whether that's an individual software engineer, just even an individual creative person with an idea, right? Now the time from the thing in your head to the creation of that thing basically gets condensed down to zero. And so you're going to see, I think, this explosion of stuff that gets created.

12:57And that's why we say it's like never been a better time to be a builder. You know, and it's true as true for something like Sora, where it's like you can have a creative idea in your head as it is for, you know, an entrepreneur who has an application, you know, in their head that they want to build. Both of those things are true.

13:12Allie:Or me who's dreaming of an email triage agent. Or you. Yes, we can solve your email inbox situation. Be careful what you promise. But I believe you. Because the other thing, too, amidst this fragmentation is there is the obvious question of, you know, some parts of this fragmented market are going to fare better than others. I mean, obviously, the word bubble comes up. I'm sure you've heard it once or twice. What's your take on AI bubble discourse? Well, you know, bubbles are kind of a natural part of this. And I think it's somewhat true if you kind of look at the history of it, that bubbles have almost always preceded some important technological shift.

13:51The reason for that's actually very rational. It's because there's people recognize what the opportunity is. Capital pours into the space. The landscape is, like we said, highly fragmented, highly chaotic. And so there's not kind of a fish and owl location of capital down to every kind of area where capital could be absorbed. Some valuations, you know, way, way outstrip the underlying traction of the thing.

14:15Allie:So you do think there are valuations out there that are not that are crazy? You know, I don't I don't have an opinion. It's hard to know. Like, it's very possible that like all this actually is hyper rational. Everything's undervalued by a factor of 10 or something like that. But, you know, I suspect that there's like a lot of enthusiasm. And at this point, like, you know, we just try and stay focused on the like thing in front of us. and for everything else kind of in the ecosystem, if people have opinions about it, they're welcome to invest those opinions. One of the things that really was a study pinged around the world was that MIT study.

14:51Allie:Did you see this? 95 % of generative AI pilots fail. There are a lot of questions about the methodology around the study, but you can't deny it struck a chord, right? Why do you think it struck a chord and what's actually going on here? Yeah, I mean, you know, I don't pay a lot of attention to studies. I think it's very easy. Did you see this, though? I did. I did, of course. Yeah. I think studies are, you know, it's easy to do a study that says one thing and then do another study that says another thing. Look, we take a very long dated view of this, right? These transformations don't happen overnight.

15:23You know, enterprises are gigantic, complex organisms. When we think about how we progress our research roadmap, we actually think along the lines of AIs that are capable of being, you know, impacting in a large organization. That's actually like a benchmark we look at as it's kind of a technical thing. We're not there yet. And so we're still very much in this kind of era of you're just now starting to have models that for the first time can use tools, they can take actions, they know how to kind of intelligently solve problems and also kind of like correct their own kind of like lack of problem solving in some sense.

15:59But there's still a lot of stuff that has to get built, right? Enterprises have decades of building all this infrastructure that governs access, permissions, identity, security, all of these things that kind of have to come into being to serve software in that environment. And I don't think that there's going to necessarily not be a need for those, all those things in this world. And in some sense, a lot of those things almost become more complicated, right? So if you've now got meaningful parts of your enterprise offloading work to agents that are running tasks and processes all the time, right, there has to be almost some concept of how do you manage and identify who those agents are and how do you kind of orchestrate them and facilitate transactions between them the way that you and I would in an enterprise, talk to each other, engage with each other.

16:40So there's almost this entirely new thing that kind of has to get built and we're like four seconds into this entire shift. And so in some sense, you know, going back to the MIT study, I'm like, does it really matter, I guess, kind of what it says? Like, we are early enough that I would be surprised if someone was coming out saying that, you know, we've cracked AI in the enterprise.

17:01Allie:I was going to say, because I do feel like some people, they may not say that directly, but they certainly will imply it. And it sounds like part of what you're saying is actually the technology is just getting there, the models are just getting there, that some of this is starting to be possible. Yeah, that's my view. I mean, and I work with enterprises all day. And I think there's still this kind of ingestion process. There's a learning process of what it really means to be able to integrate these things reliably into workflows and not just have something that's kind of like vaporware that it's like, oh, I shipped a thing and I'm going to call it an agent, but it doesn't really do anything useful.

17:31We're talking here about agentic systems that you can actually offload really meaningful amounts of work to. And that shift, I think, is still in process. Now, the stuff that we're announcing here today and other stuff that we have coming out down the line, those types of things will start to become these kind of core pillars and foundational pieces that enable businesses to start to make that transition. And so the kind of same way we said that everything starts is the chaos. It's the 15 cats that are running around chasing each other that kind of become now one giant cat that's able to walk in a straight line.

18:05That process is underway. And I think you are, though, starting to see kind of where if you squint, some of these agents are able to make meaningful, meaningful impact, right? Codex is a great example. You've now got, you know, software engineering agents that are able to write thousands of lines of code. They're able to go off and solve long-running tasks for hours on end, right? Those dimensions are only going to increase. And so just at least in that one narrow area, it's a glimpse into the future, I think, of what kind of every area could look like. But all that stuff still has to get built.

18:38Allie:In terms of enterprise AI adoption, How do you measure ROI? I recognize this is an existential question in a lot of ways, but what do you know to be true about that process? Yeah, we look at it in a few different ways. And so one is kind of hard to measure, one's easy to measure. The hard to measure thing is actually now what I kind of see as the leading edge of adoption of AI in the enterprise, which is kind of user level adoption. So people, when ChatGPT first came out, it was very clear people wanted to take it to work. And so one of the things we were confronted with early on is, well, okay, now we've got a work product kind of by default.

19:16And what does it really mean to implement ChatGPT in enterprise and kind of tell the story of what it does? And ironically, one of the things we've always struggled with is kind of what is the ROI story for ChatGPT? Because if you think about it, like we have incredible engagement rates at almost every enterprise we deploy into, like would defy gravity level engagement rates relative to traditional SaaS. And yet no one can perfectly articulate in one line what exactly AI does for the user in the enterprise. And the reason for that is because everyone uses it differently. Everyone's job is a little different.

19:48And ChatGPT is a product that kind of, it's like work that fills in around work, right? It's like, I need to write an email, but I like don't totally know exactly what the right tone of this is. So like, let me like drop a document in a ChatGPT and see if we can kind of model the tone and get the response right. That's not work that would have happened in any other place. It would have kind of happened in your head and you would have kind of, you know, fumbled around with that thing for maybe an hour or two. And now it kind of happens in 30 seconds. Right. And people, how do you measure that kind of compression of time?