Cognition CEO Scott Wu 谈收购 Windsurf：流程、交易与缘由

I think there’s an unspoken covenant that, as a founder, you go down with the ship. For better or for worse, I think that’s changed a bit over the last year, and I think it’s a bit disappointing, to be honest.

I think there’s some real truth to your point that often there are actually a lot of really valuable pieces that get left behind. In AI coding, to be truly honest, if there was zero progress, the world would still be entirely different.

Scott, I am so excited for this, dude. I was literally just telling you, it is a cool prep when you get to call up Vinod Khosla and Joe Lonsdale, and you call them and you’re like, “Hey, what should I ask?” And they’re just like, “Ah, I love this guy.” So thank you so much for joining me.

Yeah, thanks for having me. It’s been a crazy few days for us all.

1. How did Cognition pull off the $220M Windsurf deal in just 72 hours?

It’s been pretty insane. I want to start at the present day and work our way around, but how did you learn about the potential ability for you to acquire Windsurf, and how did that opportunity arise?

We found out honestly on Friday, at the same time everyone else did, that all of this was happening, that this was the split, that it was Google, and here’s what happens next and what happens with the team. We were talking about it, and on the Devin side, I think in a few ways it seemed to be a pretty natural fit.

One, there was obviously an incredible team left behind. If anything, what we have at Cognition is especially focused on the core engineering and product team, whereas Windsurf has built an amazing go-to-market team, marketing team, finance, and operations. Similarly, in terms of the products, we found that there was actually a very naturally complementary fit.

We reached out cold Friday evening, and our first conversation was Friday night. I’m glad—

So just take me to it, then. You reach out to Jeff and you’re like, “Hey, I figure something’s going down. Should we jump on a call?”

Yeah. We basically said, “Look, it seems like there’s enough here that it’s worth having the conversation and talking about it.” We got on the call together and talked about the ways to partner and what things would look like.

I always think about these things in terms of what the correct way is for things to happen. From Windsurf’s perspective, we saw it as: you could talk about the product, the team, and the business, and what is the right thing for each of those?

For the business, the thing that you have to call out is, sure, you could do this and have months of diligence and everything. That has apparently happened multiple times already in Windsurf’s company history. In this particular case, everyone was scrambling. All the customers wanted to know what was going on, the whole team wanted to know what was going on, and everyone in Silicon Valley was reaching out to them saying, “Hey, I heard the news. Do you want to come interview here?”

We came to the conclusion that a month-long process was not the way to do this. Even a week-long process was difficult. It was Friday night, and we wanted to have something ready to go by Monday.

I’m just intrigued when you look at scale. There’s a husk left behind, which kind of deteriorates, and you’re seeing that deteriorate in real time. Why did you believe that this was inherently different, and that the husk was not a husk but actually a treasure chest of value in a way that was different at scale?

There was definitely this conversation online about how all the best researchers left and there was just nothing left behind. We didn’t think so. We looked at it, and here’s the thing: it’s an amazing product that a lot of people use. It’s the entirety of the customer book as well, all of the code, a lot of the data and proprietary IP, and obviously an incredible team.

It’s an interesting state in the sense that you could call it a somewhat incomplete piece. As it turns out for us, we had the very nice complement of the pieces that were needed in order to go and do this.

Those pieces are the rationale that led to the decision. How do you actually execute it? Literally, how does that work in process?

2. 50% of new code is AI-written. Where does that go next?

We came to them and said, “Look, I think there’s a very natural partnership here. I think the team pieces are very complementary, and I think there’s a lot we can do on the product. The thing that seems absolutely clear from both of our sides is that, if we’re going to do something, we want to be ready to announce it by Monday morning.”

It reminds me of when a bank goes into receivership Friday night and you have to have an answer by Sunday and by Monday morning. We thought of this as a similar situation. There’s a ton of value here, but everyone is wondering what’s going on, and we need to give them a clear answer as soon as possible: We’re going to do this, we’re going to support the product, and we’re going to make sure everything is going smoothly and running smoothly.

That’s what we came to, and we gave them our honest take: here’s what we think we can do together. We were not going to have time to go as deep into the diligence and details as we would have liked to, but at a high level, I think we were at least in a position to understand the business very well.

We said, “Let’s get to a verbal agreement on Saturday, hash out all the details, terms, and legal on Sunday, get everything signed and done on Monday morning, and then we’re ready to announce.” There was no other way to make this work out.

Jeff, Graham, Kevin, and the whole crew honestly deserve huge shout-outs because they really worked around the clock for their team. Now we get to do this all together.

Were they left in the lurch by this deal?

What’s that?

Were they left in the lurch by this deal?

Oh, I see. I think they had known about the previous thing maybe a few days in advance. It was not a ton of advanced notice, and I think they were honestly really thoughtful with how they went about it.

3. Did Google overlook a goldmine in the Windsurf team and IP?

The options were: we can operate as an independent company, we can go and raise a new round of venture capital now that there are actually no investors, or we can find someone who we think will be the right partner to work with. Either way, we were going to have to do this quickly and figure out what we thought was the right long-term fit.

Most people see this as a really epic deal made by you, respectfully. Total hat tip and respect to you for it. The question I have—and I think it was Ruchi at South Park Commons who asked this—was: “Was this a Google blunder, leaving such a valuable asset up for grabs and not seeing the true value in the asset?”

I think there’s some real truth to your point that often there are actually a lot of really valuable pieces that get left behind.

Given that you’ve been through this deal structure that’s different and novel, do you think this becomes a new norm for how companies are acquired and how people join companies today?

I think there’s an unspoken covenant that, as a founder, you go down with the ship. For better or for worse, I think that’s changed a bit over the last year, and I think it’s a bit disappointing, to be honest.

Is the talent war getting a little bit out of hand when you look at Meta’s hiring spree? Just how much is it getting out of hand?

I’ll give my maybe crazy, controversial opinion, but I actually think it’s quite reasonable. That’s my view. The reason I say that is because I think AI is truly on the cusp of the greatest technology shift in our lives.

I think that’s already clear. The crazy thing is, even if you froze all the capabilities today and said there were no new research breakthroughs or new discoveries, you don’t even have to bet on that. You could just say, “All right, we’re going to grow products, figure out how to build the right experiences, get those out to users, and meet users where they are in all these different verticals all over the world.”

I think it would still be the biggest thing. I think the biggest thing in my mind—the biggest technology shift in our lives—has been the internet. I think there are some others which are obviously pretty massive as well: the mobile phone, the personal computer, and so on. But, yeah, I think even if you froze all the capabilities today, I think AI already would be bigger than that.

And the thing that's crazy about AI—and, by the way, I don't think it's going to freeze. I think it's going to keep moving even faster—is that there really are so few people who are really just determining the trajectory of AI. But I think, in aggregate, I think the view is right, which is that there are so few people—

4. Who are the 100 people that secretly shape the future of AI?

Sorry to interrupt you. I totally agree. In terms of what is it—like 100? Is it 10,000? Just what quantum is it?

It's a great question. Somewhere in between those 2 is right. I think there are at least 100 folks that really matter and are making a lot, a lot of difference, certainly less than 10,000, and probably a good bit less than that, if I had to guess.

Okay, so we've got something like 100 to 1,000 as a range there. Do you need those people unless you're working on the most cutting-edge frontier models? If you're in the application layer, which is very hard—I'm not denying that—but do you need them if you're in the application layer?

5. Can Apps Compete with Model Giants?

I think the level of talent and the fierceness of competition in many of these is actually greater even in the application layer, which is crazy to say because I think the competition at the foundation labs is extremely strong.

When you think, holy shit, anything-can-happen style, which it is today, one of the oh-shits that could happen that's very obvious to, I think, me and everyone is the dependency now around Anthropic. Obviously, Windsurf—when I had Varun on the show, they were very much reliant on Anthropic then, and then they got cut off when the OpenAI deal was done. Is the reliance on Anthropic greater than it's ever been?

Yeah, I mean, look, we work very closely with Anthropic. I think Anthropic is a great company, and I think the foundation labs in general are great companies. People always ask this question about where the value occurs in AI. Is it chips? Is it semiconductors? Is it foundation models? Is it the application layer? Is it the infra layer? I think the boring but true answer is that it occurs wherever you're able to establish real differentiation in your space. I think there will be a lot of spaces for which that's the case.

To be honest, I think the foundation labs as businesses will do extremely well. But I think the thing that we own, and that we spend all of our time on, is actually a nicely complementary piece, which is, in some ways, really thinking about how to best optimize for very particular capabilities within software engineering and how to deliver and ship a really great product experience around that.

Do you not want, though, the developer products for these foundation models to commoditize? Because then you gain leverage. If Anthropic continues to be so much better, then your leverage is gone. Do you not want that commoditization?

Yeah. Look, the folks are pushing ahead in the foundation-model space and making a ton of progress. We've seen a lot of launches even in the last week or 2, right? I mean, there's Kimi and Grok, and a ton of great progress that's been happening. I think that is the natural way of things: there will be competition in that layer, and I think some folks—especially in particular verticals and use cases—will be able to establish differentiation. There will be competition in our layer, and then some folks will be able to establish differentiation.

As this equilibrium develops, I think it is just naturally the case that folks in these layers want to collaborate and figure out the right ways to work together, as long as that holds.

Why would Anthropic want to collaborate when they could just own it? Sorry, I'm a VC, so I'm naive.

No, of course. Of course. I love it. You're hitting me with all the hard questions. Look, I think the answer at the end of the day is that what we focus on is very, very different. The kinds of questions that we think about, for example, on both the Devin and Windsurf sides, by the way, are really just: How should humans and AI work together to produce code?

It's one thing to just solve for generally smarter and smarter base models. It's another thing to teach your specific model, “All right, here's how to go to Datadog and pull up the logs for this thing, and here's how you debug a front end live, and here's a representation of the codebase,” which we're learning and iterating over time, and so on.

I think at the end of the day, there are a lot of different verticals to work in and to solve for. I think we have parts of this equation, and I think the foundation labs have parts of this equation, but I think the truth is you really need both.

You prefaced that one with, “You can ask the hard question.” So, fuck it. It's the end of the day in the UK, so I'm using that as an excuse. What percent of your revenue goes to Anthropic, do you think? Ballpark.

I have to pass on that one.

I love it. It's like with children, where they'll push and they'll see how far they can go.

When you speak about that progress and the incredible progress that we've seen, some think that it will take the same route as self-driving cars, where there is this kind of plateauing, and then we'll see another inflection point. What gives you such confidence that we will see the continual progression in models that we've seen over the last 12 months over the next 36 months?

Sure. I'll give you 2 thoughts there. The first thought is that there are very strong signs of this continued progression, largely because there's a lot of specific work—which you can call research, engineering, infra, or whatever you like—but there's a lot of work where we know the techniques, and we have a lot more to do to go and scale those.

RL, for example, is, I would say, the biggest breakthrough of the last year and a half, call it. It's really amazing. It's crazy to think that you roughly can solve any benchmark. I mean, that sounds insane to say, but that's really what RL is converging on, which is basically: if you have a clean enough set of “Here are exactly the behaviors that I want, here are the environments that you need to be able to operate in, and here's what it means to succeed or fail,” you can just train a model that does that. Naturally, that is just such a powerful capability, which I think is going to be applied to more and more spaces. I think we will see that progression.

The other thought I would give, though—and I mean this truly sincerely—is, look, there are a lot of spaces and a lot of different things going on in AI. In AI coding, to be truly honest, if there was zero progress, the world would still be entirely different. I think there are a lot of spaces that are early today where you see AI making progress: it's getting better and better, and you're going to the next step and the next step.

In code, you are just slower as a software engineer if you're not using AI. That is the truth, and it is a no-brainer already today. Look, I think it will be more of a no-brainer, and I think we'll be able to make engineers even more efficient, able to do even more with code. But I think that is already the case.

So many things to unpack there, dude. Do you think tools like Devin and others make 1x engineers 10x engineers, or 10x engineers 100x engineers, if I were to put you in one camp?

Interesting. It actually really does depend, tool by tool. I do think that the product experiences for the 1x-to-10x and the 10x-to-100x, or even the 0x-to-1x, or whatever you want to call it, are somewhat different product experiences.

I've interviewed Benioff at Salesforce. I interviewed Vlad at Robinhood, both very recently, and they both said that 50% of their net-new code is created by AI. Would you agree with that in what you see with your customers and peers, 1, and then where will that be in 3 years' time?

Absolutely. People often talk about the percent of code that's written. I think the obvious thing to call out is that it's a little odd because, for one, you have to factor in how important each line of code is, right? If it's just a ton of protobufs, then that's one thing, versus a lot of the core business logic being another thing.

And then, 2, of course, code written with the help of AI is one thing, but how much? For those reasons, we often like to think about it in terms of how much faster is an engineer using all these tools—an engineer using the best AI tools and who really understands how to get value out of them. How much are they doing in 1 hour versus how much they would do in 1 hour with no access to AI?

I would say that something in that range of 1.5 to 2x feels right to me today, in aggregate.

Look, I think in 3 years there’s no reason that shouldn’t be 10x. And I think the thing that’s really fun, by the way, is I think we’re going to have more than 10x more code. Something I’ve been saying recently is I keep track of every time in my daily life where software fails me.

We’ve become conditioned to be okay with it, but the truth is it happens all the time. You think about products out there, and the way I like to say it is that there is this top tier of products which are, in some sense, the best-made products in the world. I’m thinking of YouTube, TikTok, Instagram, and so on, and you can really feel every little detail that they’ve done with a ton of care.

They’re streaming in tons and tons of data, and it’s always super efficient. It never goes down, and it’s super reliable. The algorithm basically knows you better than you know yourself. The UX is super intuitive, and I call that hundreds of millions of hours of engineering time that went into building that piece of software.

Then you go down to the next layer, the next order of magnitude of software that has tens of millions of hours spent, then single-digit millions, and so on. You see the differences really quickly. When you’re logging into your bank, dealing with your healthcare, working with your insurance and trying to get things going, or trying to navigate your customer list or things like that, the truth is all the software can be 10x better.

I think there actually is 10x more of it to write. I think that’s one of the fun things in code, right, is that it really does have this Jevons paradox.

When you think about that developer efficiency, what are agents not able to do today that they really need to be able to do?

I think there is a real point at which agents are truly able to take over ownership, I guess is how I would describe it, of the work that is done. The way I want to say it is, I think in the future we’ll get to a point where you’re not looking at your code; you’re looking at your product.

At the end of the day, code—software engineering, this whole thing—is just telling your computer what to do. Code is the language that your computer happens to speak, and that’s why we all have to learn how to write code in order to do it. Over time, I think you get to the point where you can just say, “All right, this website—let’s add a new tab here. Let’s put this and that information, and maybe let’s collect this information from the user and save it in the database this way. This button should be a little bit rounder.”

You’re just able to make those calls and make all those decisions. I think we’ll still call it programming, but I think it’s going to transition into being more of a technical architect or a technical product manager—someone who’s really owning these decisions.

Agents have to get to the point where they’re basically that high-touch and have that kind of understanding of context that you can give them that level of instruction and they’ll just go and do it. In a world like that, what skills become more valuable and what skills become less valuable?

Everything that you do is going to be about essentially this core thing of deciding what is the solution you’re going to build. What is the problem that we’re facing? What is the solution that we want to build? How exactly do we want to architect that solution? I think that’s going to be the most important skill.

There are so many things I want to ask. I do just want to go back before I forget it. You said about the 1.5x more efficient and more productive.

Yeah.

Do you think we’re struggling now in a value-chasm gap? What I mean by that is, how much do you charge for Devin today, on average?

It’s all usage-based, but it’s essentially by the hour. We try to make things so that they’re about 10x cheaper than basically the value of your time.

Because if you think about, say, a software engineer being $300,000 a year, give or take, $150,000 then would be that 0.5x that you’re adding.

Yeah.

Are these tools going to be sufficiently paid for the value creation that they are enacting?

Yeah. Look, I think the value creation is a beautiful thing. On this point of where the value accrues in AI that we were talking about earlier, there are 30 million software engineers in the world. We’re going to make them all 10x more efficient over these coming years. We’re going to be writing 10x more code. We’re going to be doing a lot here.

We could talk about whether it’s 5%, 10%, 20%, or 30% of the value that actually gets collected by the companies doing this. But honestly, I think that the highest-order bit is actually less than that and more about just getting the technology and building the products to a point where everyone is going a lot faster with them.

When you look at something that no one sees that everyone should see, when you think about the future of AI code and the future of software engineering, what does no one talk about that you think more people should be talking about?

A focus on deep context. It’s already better than us, honestly, at these sandbox problems. A lot of the tough questions are things like, well, there was this project that we’re trying to do today, and it’s very similar to what somebody else actually asked Devin a month ago. How do we use that knowledge and improve on that in order to make Devin smarter?

Or there are these little things like wanting to test the front end for your codebase and make sure everything looks as expected. You should be able to understand what that’s supposed to look like, or why this is different from what you know. If you find a bug, you should be able to understand how you found the bug.

That’s a lot of the little detail of, honestly, what makes—I’ll call it—the difference between code and software engineering. It’s basically working in a large, complex codebase, building some intuition and some representation of all the different pieces and how they interact with each other.

Learning how to use all the various tools at your disposal to actually understand what’s going on and to debug and diagnose—that is actually the big problem in AI coding. I think that’s the big thing that folks will work on next.

The thing that’s fun, by the way, is it’s a very practical problem. You can call it a research problem—obviously, in many ways it is a research problem—in terms of how to push for these capabilities and make these things better, but it’s not the kind of thing that you can solve in a sandbox. It’s the kind of thing where you actually really just need to think about the practicality of software engineering.

You said there that coding agents were what you saw before other people saw, and what no one was discussing. It’s a hard one, and I don’t know how blunt I can be, but it felt like Devin fell out of the zeitgeist a bit, if we’re honest. Cursor and Windsurf owned consumer attention and owned the brand. Do you think that was a case that Devin just wasn’t very good at marketing, or do you think that was a product misstep?

Yeah, it’s funny because I think there’s always an external perception and then an internal picture of what’s actually going on with the numbers. Without going too much into detail, what I can say is that in the last 6 months, between January and now—even aside from the latest deal this week with Windsurf—the usage of Devin has actually grown something like 5x to 10x.

That’s been the case in both self-serve and enterprise. It’s been fun to see that. I think the pattern, for what it’s worth, is really real engineering teams that bring on Devin. They tag Devin all the time in Slack, they tag Devin in Linear, and so on, and then they use it, grow it, and share it that way.

In many ways, it is not the same kind of experience as a single non-engineer just going and signing up for an account and immediately building something really cool with it. People do use it that way, but it’s not the majority of our usage. The majority is really just real teams using it.

I think the point that you’re making does get to an important thing, which is that IDEs and the IDE experience obviously came and really started working, I think, about a year before the agent experience really did. I kind of think of agents as having taken off in the last 6 months or so, and I think of IDEs as having really started to become a no-brainer, obvious choice sometime last year, in terms of the value that they provided. What we’re seeing is an artifact of that.

Which is that folks are more familiar with IDEs because they’ve been around for a while. I think you fast-forward 6 to 12 months from now, and people will be familiar with agents to the same level that they’re familiar with IDEs today.

5–10x is mega. It’s fantastic. The hard thing is, if you compare it to Replit or Lovable, the growth is just nuts. I’m in Lovable, Scott, and I just look at the numbers and I’m like, “What? How does that work?” Is that not a fair like-for-like comparison if you’re thinking about growth rates? Actually, I don’t even know their numbers over the last 6 months. What are their numbers?

Yeah. Look, I don’t want to get into exact numbers, but even with the Windsurf deal aside—not including that at all—we’ve basically done roughly that as well in the last 6 to 8 months.

I stand corrected on my consumer SaaS guys comment. I think that’s an important point to consider. By the way, I think these are great companies, and I obviously think they’re—

But then it’s a marketing problem. Dude, I’m saying this to you now—I hope we can be friends. I’m projecting forward this relationship as a friend. I would say you guys should be better at selling yourself.

That’s nuts. That’s great.

Yeah. Well, perhaps we should be. The great news is we’ve just now inherited a great marketing team, and we get to do that.

The only thing I’d call out—I know Amjad from Replit, and I’ve met a lot of these folks before—is that I think of them as actually quite different products and businesses and so on. To the point that we were making earlier, there are a lot of different product experiences that you can solve for in code. There’s bringing a 10x engineer to 100x, there’s bringing a 1x engineer to 10x, and there’s someone who doesn’t know how to code and bringing them to 1x.

I think Replit and Lovable, to my understanding at least, are much more consumer-y, and perhaps that’s why you hear about them more on Twitter or YouTube or things like that. Whereas Devin can be anywhere from startups to some of the biggest companies in the world, but it is really focused specifically on engineers on engineering teams trying to do their work and go faster that way, right?

You can see this in how this is all set up. Most Devin sessions are started through Slack or through Linear, and Devin makes these pull requests in GitHub that you go and review and merge. Devin works with your whole development system and learns how to get onboarded to your repo and so on. I think that’s one difference there.

But with that said, point taken. I take the feedback.

Dude, I’m always here to show you how to do a good tweet. After 10 years, I now know how to do a good tweet. That’s all I know. I mean, you’re the master at it. I was going to say, after this, we’ll talk and I want to get your tips, dude.

6. IDEs & Agents: Just Training Wheels?

Honestly, I have so many for you with those numbers. I’m like, “Wow, dude. You need—” Anyway, can I ask you: my job as an investor is also to think about market makeup at an end state, because that’s where value kind of accrues in my mind.

When you think about where the developer market shakes out, is it that Cursor wins the bottom-up developer minds, and Windsurf and Cognition now win top-down, large-enterprise, super-solid blue-chip clients? Is that how you think that developer market shakes out?

I think the honest answer is that it’s far too early to call on any of these. The reason I say that is because—I’ll give you a hot take—none of us are that close to the future of software engineering.

I think the future of software engineering is going to take place over the next couple of years or so, but it really is this version where it’s not just—we call it an IDE or a coding agent or this or that. These are the terms that we use.

If we’re being real about what we’re building here and what this is all going towards, I think of this as the next generation of human-computer interface. That’s the problem that’s being solved here, right? Basically, as we said, code software—the whole point of that is just telling your computer what to do. At some point, telling your computer what to do is not going to take place with code. It’s going to take place with you just expressing your intent.

There are a lot of capability problems to solve to get there. There are a lot of interface problems to get there. But I think that’s essentially what we get to. The simple thing I’d say is, Tony Stark does not pull up his laptop. Tony Stark goes and talks to JARVIS, right? I think there is a point at which you just have a very clean connection, the ability to express your intent and do these things.

People talk about generative UI. People talk about single-use software and so on. At the end of the day, what it all really boils down to is having a very clean connection and being able to just tell your computer what to do, and it’ll do that for you.

I think if we think about where everyone is today, there are 10 or 20 levels of what the product experience looks like until we get there. The fun thing is every level is 2 or 3 months. If you just multiply that out, it means we’re going to be there in a few years if this pace keeps going.

I believe a common enemy is a very important thing within a company. If I were to push you, Scott, and say, “Who is the competitor that you most look to and think about, and respect even? Who would that be?”

Jensen Huang said this once, and it always stuck with me: when you have figured out a way for your company to win that means no one else has to lose, then you will know that you have found your path. Do you think he’s done that? I think that’s huge.

I certainly think he’s done that. NVIDIA is an incredible business. Jensen obviously has done that. It’s a freaking monopoly.

Well, I think his point is that, at the end of the day, there are so many different verticals to serve. There are so many different niches to own. There are a lot of businesses that, of course, will work in adjacent spaces, and we’ll see things and run into folks on a deal or whatever it is. But at the end of the day, people have the things that they specialize in.

I firmly believe that. I think the thing that’s been fun throughout the history of Cognition is, for better or for worse—maybe it’s because we’re insane, I don’t know—we’ve always had a pretty unique approach and a pretty unique view about, “Here’s what we think the future is going to be. Here’s what we want to build for.” And that’s different from Cursor; that’s different from OpenAI or Anthropic themselves and how they think about these things and so on.

I think the fun thing about that is there is so much to build in code, and I truly do think there will be more than 1 winner.

What have you not built that you would most like to build?

The answer I would give today—and this is the biggest thing we’ve been thinking about in the last few days since this deal—is really figuring out what that combined experience of IDE and agent is, because I think there really is something here.

To what we said earlier, I think years from now all these systems will look so different, and we may not even have the same terms for them. But in the immediate future, I think both IDEs and agents will be an important part of a developer’s workflow.

You can imagine all sorts of things: “Hey, I want to go and plan out a task in the IDE. I want to be able to use the intelligence and the retrieval and Devin Search, and read on the wiki to understand exactly what decisions I need to make, what parts of the code this is going to touch, and what things I need to plan out.” Then I want to be able to hand that off to an agent, have the agent go do the bulk of the work, and then I’m going to go and review the code.

Naturally, it’d be great to review the code locally in my IDE. If there’s any touch-up that I need to do, I can use my in-IDE tooling to go and do that.

Basically, figuring out what that combined experience is, where you can go from synchronous to asynchronous to synchronous, and be there for whatever parts need you, while being able to parallelize and do more for the parts that don’t need you—I think that’s going to be a really fun question to figure out.

In terms of what happens with the combination of Windsurf and Devin, in the immediate short term, obviously there’s a lot of Devin to run and a lot of Windsurf to run. We certainly plan on maintaining the philosophy of both of those products, but I think finding that intersection—how do you make it a really smooth experience to go between them for the folks who use both—is going to be the fun one.

7. Quick-Fire Round

Dude, I've so enjoyed this. It was—yeah, it's just been a really fun conversation. I do want to do a quick-fire round, if that's okay.

I have to say, I'm a bit underslept, so please pardon me if I need an extra 1 or 2 seconds for the fireworks.

How have you been sleeping?

Next question. No, look, we've had some crazy days. Friday, Saturday, and Sunday were obviously about figuring out if there was a deal to be done and making that happen. Monday, Tuesday, and Wednesday have been about figuring out how we bring the teams together and how we build the really great thing.

Obviously, the immediate thing is going out to the customers and letting them know, look, we're here. We've got the engineering firepower and the work to be able to support things. We're going to make sure you guys have a super-solid experience, and if anything, we're going to be able to make it better, even faster, and figure out a better way to even uplevel the engineering team. Long story short, not a lot of sleep this week.

I love it. Okay, dude, what's 1 widely held belief about AI that you think is completely wrong?

Yeah, I'll give you a take. Sam Altman had this post 10 years ago called “bubble theory.” Do you remember it? He was basically saying—this was in his YC days—everyone here says that all these companies are overvalued and we're in a bubble. It's obviously fun to talk about bubbles and all, but I don't believe that, and I'm going to give you a bet.

8. “We’ve gone from 0 to $80M ARR in 6 months. Quietly.”

It looks hilarious, by the way, when you read it in retrospect. He said, here are the top YC companies today. It's literally Uber and Airbnb and whatever. Today, these are worth this much. I predict that 5 years from now they're going to be worth 3× more.

Then it's like, here are these emerging, mid-level companies. That list was Stripe and some of these others. Today, this is worth this much, and I predict these will be up at least 3×. He basically said, everyone says we're in a bubble. If you'd like to go and bet on that and put your own money down, I am happy to put money on this.

I have to say, in AI, I really feel this. I feel this now more strongly than ever. Over the last 2 years, people have kind of thought of it as this GenAI wave, again, with lots of incremental jumps. But the truth is, RL is the master—the big story, I'd say, over the last year or 2 in capabilities. I think people have really underappreciated how much is possible with RL.

What do people not see with RL that they should see? Because I candidly don't.

Yeah, so look, again, RL is—we've had a few years of what I would call imitation learning before RL. That's what went into GPT-3 and 3.5 and so on, which is basically: you take the entire internet and read the whole internet, and you get a model that sounds like somebody on Reddit. It's crazy.

ChatGPT obviously shocked the world, and it was a huge moment for everybody. But you could kind of see why talking like the average isn't necessarily the thing that you need to do to build a great medicine specialist or a great software engineer or any of these things. I actually got a lot of the way there, to be fair—more than people might have guessed.

I think the next big thing with RL is, as we said, you can take any benchmark and solve it. It is crazy to say, but the next step of that is honestly an open question for everyone in every application: what is the benchmark?

If you're an accountant, the benchmark, in real-world terms, is roughly: all right, you submit returns on behalf of the client, and if you get audited by the IRS and something comes up, then that was a fail for the RL. If you don't, then you did a good job. Obviously, you want to be able to constrain it to shorter feedback cycles than that and think about what the ways are that you can determine whether an agent's work was a success or a failure.

But once you have that, it turns out that you can just train agents to do all sorts of things, and that's what we've seen, by the way. We've seen folks getting gold medals in these international math competitions and things like that with AI. The truth is that this is really hard stuff, right? Obviously, you have to go and apply it to every vertical, but it's going to happen.

On that point, I was just going to say, you think about the companies today: there's the foundation-layer companies, there's the application layer, and there's a lot of the tooling. I would happily bet—take all the foundation labs today, right? It's the private ones. There's OpenAI, there's Anthropic, there's xAI, let's put in SSI. We could include Thinking Machines in that list as well.

If you add all their valuations today, what is that, like $500 billion, something like that? I think that's going to go way up in the next 5 years. Then if you take all the application-layer companies, the top ones that come to mind are Perplexity, Sierra, Decagon, Harvey, us, and Cursor. Today, that list of companies is probably worth $50 billion to $100 billion, and I think that's going to go way up in aggregate.

Obviously, it doesn't mean every single company is going to go up, but I think the value that we're going to produce is so massive here that—

You can invest in OpenAI at $300 billion or Anthropic at $60 billion. Which one do you choose to invest in?

Honestly—and I mean this sincerely—I think both are great investments. I think both are great investments at the price.

Amazing. You can put money in 1.

Oh, man. Well, Anthropic has quadrupled its revenue since it was valued at $60 billion, so it seems like a fair one to pick.

Do you think we'll have—you mentioned there were, like, 6 in the foundation-model space?

Yeah.

Do you think we'll have 6? Do you not think consolidation will already—

I think consolidation happens, and obviously there's Google and there's Meta. I think consolidation happens, and in total we probably end up with something like 2 to 6 players, probably. I'm trying to hedge to be conservative. It really is probably—I think 3 to 5 seems pretty reasonable, but you're making a face at me. I want to hear what—

2. 2.

You think just 2?

You've got ChatGPT from OpenAI on consumer, and then you've got Anthropic on enterprise.

Really? You think it's just those 2?

You might have really tertiary ones. It's just search or, you know, providers. You have minor ones, like your DuckDuckGos of the world. But, yeah, the DuckDuckGo of AI. I'm sure someone's looking forward to winning that elusive title.

Look, I think the thing is funny on consumer—I totally agree—which is ChatGPT. I mean, everyone knows just ChatGPT as the term. I think there's an interesting effect, which is obviously companies want choice, and I think this is going to happen in a few spaces as well.

It isn't necessarily always a power-law distribution where number 1 is 75%, number 2 is 20%, number 3 is 4%, and everyone else combined is 1%, or things like that. But in terms of market share, let's say—in terms of people getting there on capabilities and being able to offer something that's at least competitive enough that 4% or 20% of people would consider it—I think there is reason to believe that there will be at least a few folks that get there.

For the acquisition, what was the percentage between stock versus cash?

It was a mixture of both. I probably can't comment on the exact split.

I'm so enjoying this constant game of push.

No, I know. Well, I'm very happy to—you could ask all your questions, and I'll push back where I need to push back.

Can I be honest? I think you missed a trick with your acquisition video. I watched it and I was like, no offense, it's a bit personalityless. Meeting you, you're a riot. I'd love to hang out with you, and you've got an amazing personality, charisma—you're awesome. I think the acquisition video could have been more personable.

That's fair. I agree. The thing is, dude, we were in between—the lawyers all pulled an all-nighter as well, getting this ready, because it was like, yeah, we need to get this ready to go. There were just all the various little things: this term, this term.

But, yeah, it was in between that and sorting out all the questions with the team. Then we basically had 20 or 30 minutes to go to just film something real quick.

I was going to say I was pretty happy with the output relative to the amount of effort that was put in, but I fully agree that there was more to it.

I'm 100% with you. My favorite bit was, funnily enough, we did have one competitor who we thought really highly of.

So what did you believe in the world of AI that you have subsequently changed your mind on in the last 12 to 24 months?

I think one thing which has really started to change in the space is that 24 months ago, actually, the story was all about data and just more and more and more data, the quantity of data and figuring out how to get even more. I think that has changed a lot toward figuring out a small set of highly curated data for exactly the use case that you care about.

A lot of our research work we obviously can't talk about publicly, but we did share one project: a model that we released called Kevin. I'm not sure if you saw it—Kevin 32B.

Yeah, yeah.

KernelBench is a benchmark that involves your ability to write CUDA kernels. With RL on very specific CUDA-kernel agent trajectories, you can make it way better than other models on this stuff. This is a story we're seeing in a lot of these: rather than just mass volume of data, if you have a very particular use case and behavior you're looking for, a small amount of data in exactly that vertical, with the environments and feedback loop set up correctly, is more what you need. It's a lot of compute, not a lot of data—quality of data over quantity.

Would you advise a new young student to study CS today?

Yes, absolutely. People ask me this all the time, by the way. The reason I say that is because, if anything, I'll give you my hot take: the complaint that we've had all along about schools is that they're supposed to teach you computer science and software engineering, but you go to school and learn about garbage collection, algorithms, architecture, and everything, and then you go on the job and a lot of what you have to do is debugging JavaScript stack traces, right?

I think the thing that's kind of funny is, as we said, this fundamental skill set of how you solve problems and how you think from first principles, understanding the model of a computer, and understanding a lot of these architectures—how does a database work, or things like that—those are actually going to be the important things, I think, going forward. Maybe one way to say it is that, at the end of the day, a degree in computer science is more a degree in how to think, and I think that will always be valuable.

Scott, I want to finish on one which I think is really important. I've so enjoyed this, and I watched this on TikTok. It said that if you want to build a relationship with someone, ask them about the trait within themselves that they're most proud of or like the most, because in future interactions you're able to refer to it and it makes them feel good.

I'm intrigued. What trait are you most proud of? What's your favorite trait of yourself, and why?

That's a fun one. I'm actually very curious to hear your answer to this question as well, by the way. I think my answer, funnily enough, is that I would say it's a sort of emotional calmness.

I get very salty. I'm a very competitive person. I get very frustrated and things like that. But, for better or worse, we've been in a lot of stressful situations, as you can imagine. We've been in a lot of them in the last 6 days in particular, but honestly, that's been the entire story of the company. It's probably a bit more than average than an average week, but not by that much.

I think, for better or worse, I'm proud that I'm just able to stay composed in the situation.

Dude, I love you. You're great. Seriously, I think you are the most under-discussed personality in this business. You need to do more. The Scott brand needs to be a bigger personal brand.

What do you suggest? I'm looking to you as the interim chief marketing officer at Cognition.

The thing I think that's nice with your product is actually that you fall into the Nike bucket. What do I mean by the Nike bucket? The Nike bucket is that their success was making you feel like a superhero. They tell you that everyone is an athlete, that even if you don't have the skills, you are able to do things you never could before.

I think that you and Devin fall into that similar bucket of human enhancement. I think you should tell the stories of that much better. Then I think you should also, bluntly, very clearly tell the story of your own growth much more deliberately.

People want to be part of a rocket ship, and gossip is very vicious, and it happens when you don't shape the narrative yourself. If you shape the narrative, gossip doesn't really happen, because I can't say you're at 20 million in revenue when you're not. You're much higher.

Yeah.

Don't let other people shape your narrative for you.

Yeah, no, fair enough. I think there's been a real shift. A few months ago, it was almost better if people didn't hear that agents were working. We'd rather have it that way. I think in the last few months, everyone and their mother is now trying to do agents anyway, and so we might as well be more public.

But I also think, just for talent and for funding, people also want the brand.

Yeah, yeah, yeah. No, it makes sense.