Cognition CEO Scott Wu on Acquiring Windsurf: The Process, The Deal, The Rationale

I think there's an unspoken covenant that as a founder you go down with the ship, and I think that for better or for worse it'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 code, to be truly honest, if there was zero progress, the world would still be entirely different. Ready to go.

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

We found out honestly on Friday, the same time everyone else did, that all of this was happening and that this was the split and that it was Google and here's what happens next and here's what happens with the team. And we were kind of talking about it, which is, you know, on the Devin side, I think in a few ways it seems to be a pretty natural fit. One, because I think obviously there was an incredible team left behind, and if anything, I think what we have at Cognition, I would say, is especially focused on the core engineering and product team, whereas obviously I think Windsurf has built an amazing go-to-market team, marketing team, finance, operations. And then similarly in terms of the products, you know, we found that actually there was a very naturally complementary lean. And so we reached out cold Friday evening, and our first conversation was Friday night, and it was, I mean, I'm glad.

Yeah. And so I mean, 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. We were talking about, you know, the ways to partner and what things would look like. And I think what we came to pretty quickly, you know, I always think about these things in terms of, call it like, what is the correct way that things should happen here, right? And I think from Windsurf's perspective, the way that we kind of saw it was, well, you know, you could talk about the product, the team, and the business, right? And what is the right thing for each of those. And I think for the business, the thing that you have to call out is sooner. You could do this and you could have like, you know, months of diligence and everything, and obviously that has apparently happened multiple times already, you know, in Windsurf's company history. And I think in this particular case, I think it's, look, everyone is scrambling. You know, all the customers want to know what's going on. The whole team wants to know what's going on. Everyone in Silicon Valley is reaching out to them saying, 'Hey, I heard the news. You know, you want to come interview here.' And so, you know, we kind of just came to this conclusion: look, a month long is not the way to do this. I mean, even like a week long, you know, it is Friday night now. We want to have something ready to go by Monday.

Yeah. So there was definitely this conversation online, right, about, oh, like all the best, you know, all the best researchers left and there's just nothing left behind. And we were, you know, I don't think so is basically how we felt about that. And we looked at it and I mean, it's, look, here's the thing, right? It's an amazing product that a lot of people use. It's the entirety of the customer book as well, right? It's all of the code, a lot of the data and the proprietary IP, and then obviously it's a really incredible team. And so it's an interesting, you know, I think it's an interesting state in the sense that you could call it perhaps like a somewhat incomplete piece, but I think that, I mean, as it turns out for us, I mean, we had the very nice compliment of what were the pieces that were needed, right, in order to go and do this.

Yeah. For sure. So we came to them and we said, 'Look, I think there's a very natural partnership here. I think the team pieces are very complimentary. I think there's a lot we can do on the product. The thing that just seems, you know, absolutely clear from both of our sides is if we're going to do something, we want to be ready to do this and announce it by Monday morning.' You know, it kind of reminds me of like, you know, the bank goes into receivership Friday night. You got to have an answer, you know, by Sunday and by Monday morning, right? And I think we thought of this as a similar situation, which is, look, there's a ton of value here, but everyone's wondering what's going on and we need to be able to give them a clear answer as soon as possible that, look, we are going to do this. We're going to support, you know, the product. We're going to make sure everything's going smoothly and running smoothly and so on. And so that's what we came to and we gave them, you know, our honest take of like, here's what we think we can do together. It's, you know, we are not going to have time to go as deep into the diligence and details as we would like to, but I think at a high level we at least, you know, I think are in a position to understand the business very well and let's just figure out if, you know, like get to a verbal agreement on Saturday, hash out all the details and the terms and the legal on Sunday, get everything signed and done on Monday morning, and then we're ready to announce. Like that's it. There's no other way, you know, to make this work out. And Jeff and Graham and Kevin and the whole crew, honestly, like I mean, it's a huge shout out to them because they really worked around the clock for their team. So now we get to do this all together, right?

What's that?

Oh, I see. So yeah. So I think they had known about the, I think they had known about the previous thing, you know, maybe a few days in advance. It was not a ton of advance notice and I think they were honestly really thoughtful with how they went about it, which is, look, here are the options. We can operate as an independent company. You know, 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 are going to be the right partners to work with. But either way, you know, it's going to be, we're going to have to do this quickly and we're going to have to figure out what we think is the right long-term fit for it.

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. And so...

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

Yeah, you know, I'll give my maybe crazy opinion, controversial opinion, but I actually think it's quite reasonable is my view. And the reason I say that is because I think AI... I think we are truly on the cusp of the greatest technology shift in our lives. You know, I think that's already clear. And I think this is, you know, crazy thing is even if you froze all the capabilities today, you know, and then the question was just, you know, if you said there were no new research breakthroughs, no new discoveries, you know, you don't even have to bet on that, and you just said, all right, we're going to, you know, grow products and figure out how to build the right experiences and get those out to users and, you know, meet users where they are in all these different verticals all over the world. I think it would still be the biggest thing. You know, I think the biggest thing in my mind, the biggest technology shift in our lives has been the internet. You know, I think there are some others which are obviously pretty massive as well. The mobile phone, the personal computer, and so on. But yeah, you know, 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. But the crazy thing about AI is there really are so few people that are really just determining the trajectory of AI. But I think in aggregate, I think the view is right, which is, you know, there's so few people...

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

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

Yeah, I mean look, we work very closely with Anthropic. I think Anthropic is a great company and, you know, I think the foundation labs in general are great companies. People always ask this question about, you know, where does the value occur in AI? Does it, you know, is it chips? Is it semiconductors? Is it, you know, foundation models? Is the application layer? Is the infra layer? And I think the boring but true answer is it occurs wherever you are able to establish real differentiation in your space. And I think there will be a lot of spaces for which that's the case. And, you know, to be honest, I think the foundation labs as businesses will do extremely well, right? But I think the thing that we own is, and that we spend all of our time on, I think is actually a nicely complimentary 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 like a really great product experience around that, right?

Yeah. Look, I mean I think it's, to be honest, there are folks pushing ahead in the foundation model space and making a ton of progress. We've seen a lot of launches even in the last, you know, week or two, right? I mean there's Kimi and Grok and a ton of great progress that's been happening and I think that is the natural way of things, which is, you know, there will be competition in that layer and I think some folks will, you know, especially in particular verticals and use cases, be able to establish differentiation and there will be competition in our layer and then some folks will be able to establish differentiation. As this equilibrium develops, you know, I think it is just naturally the case that folks in these layers want to collaborate, you know, and figure out the right ways to work together as long as that holds.

No, no, of course. Of course. I love it. All hidden me with all the hard questions. Look, I think the answer at the end of the day is what we focus on is very, very different, right? And the kinds of questions that we think about, for example, on both the Devin or the Windsurf side, by the way, is really just like how should humans and AI work together to produce code. It's one thing to just kind of like solve for generally smarter and smarter base models. It's another thing to teach your specific model, all right, here's how to go, you know, to Datadog and pull up the logs for this thing and here's how you, you know, debug a front end live and here's what the, here's a representation of the codebase which we're learning and iterating over time and so on. And 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.

I have to pass on that one.

Sure. Yeah. So I'll give you two thoughts there. The first thought is there's very strong signs of this continued progression largely because, by the way, there's a lot of specific work which is, you can call it research, you can call it engineering, you can call it info or whatever you like, but there's a lot which is roughly that we know the techniques and we have a lot more to do to go and scale those. You know, it's RL, for example, is I would say the biggest breakthrough of the last year and a half, call it. And it's really amazing that it's crazy to think that you roughly can solve any benchmark. I mean that is, it is almost, it sounds insane to say, but that's really what RL is converging on, right? 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, here's what it means to succeed or fail, you can just train a model that does that. And naturally that is just such a powerful capability which I think is going to be applied to more and more spaces. And so 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, you know, and a lot of different things going on in AI. In AI code, to be truly honest, if there was zero progress, you know, the world would still be entirely different. You know, I think there are a lot of spaces that are early today that 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. And look, I think it will be more of a no-brainer and I think we'll be able to make engineers even more efficient, you know, be able to do even more with code. But I think that is already the case.

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

Yeah. Absolutely. So people often talk about the percent of code that's written. I think the, you know, the obvious thing to call out is, well, 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 like a ton of protobufs, then that's, you know, that's one thing versus, you know, a lot of the core business logic is another thing. And then two, of course, is, you know, written with the help of AI is one thing, but how much? And so for those reasons we often like to think about in terms of like how much faster is an engineer using all these tools, you know, an engineer using the best AI tools and, you know, 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? And I would say that something in that range of, you know, 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 a 10x. And I think it's 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 know, and it's, you think about products out there and, you know, the way I like to say it is, you know, there are, I say this top tier of products which are in some sense like, I'll call them the best made products in the world, and I'm thinking of like YouTube, TikTok, Instagram and so on, and you can really feel every little detail that they've done with a ton of care. You know, it's, 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 that is like, you know, I call it like hundreds of millions of hours of engineering time, you know, that went into building that piece of software, right? And then you kind of go down to the next layer of, you know, the next order of magnitude of software that has tens of millions of hours spent and then single-digit millions and so on, you know, and you see the differences really quickly. I mean, when you're logging into your bank or when you're dealing with your healthcare, you know, working with your insurance and trying to get things going or when you're trying to go and like, you know, navigate your customer list or things like that. And the truth is all the software can be 10x better, you know, and I think there actually is 10x more of it to write. And I think that's one of the fun things in code, right, is that it really does have this Jevons paradox.

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. And the way I kind of want to say it is like, look, I think in the future we will get to a point where you're not looking at your code, you know, you're looking at your product, right? And you're looking at your product and at the end of the day, code, software engineering, this whole thing is just telling your computer what to do and 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, right? But I think over time, yeah, you get to the point where you can just say, all right, yeah, like this website, you know, let's add a new tab here and, you know, let's put this and that information and maybe let's go collect this info from the user and we'll save it in the database this way and, you know, this button should be a little bit rounder and you're just able to make those calls and make all those decisions and I think we'll still call it programming, you know, but I think it's going to transition into being more of like a technical architect, a technical product manager, you know, someone who's really owning these decisions and I think agents have to get to the point where they are basically that high-touch and that kind of understanding of context that you can give them that level of instruction and they'll just go and do that.

Everything that you do is going to be about essentially this core thing of deciding what is the solution you're going to build. You know, 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? And I think that's going to be the most important skill.

So it's all usage based, but it's essentially by the hour. And we try to make things so that they're about 10x cheaper than, you know, than basically the value of your time and going.

Yeah. Look, I think the value creation is a beautiful thing. And to this point, by the way, of where does the value accrue in AI that we were talking about earlier, there are 30 million software engineers in the world. You know, 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. And I think that, you know, we could talk about whether it's 5% or 10% or 20% or 30% of, you know, the value that actually gets collected by the companies doing this, but honestly, I think that the highest order bid is actually less that and more about just, you know, getting the technology and building the products to a point where everyone is going a lot faster with them.

A focus on deep context. It's already better than us honestly at these sandbox problems. A lot of the tough questions that we answered today are things like, well, you know, 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 and how do we use that knowledge and kind of improve on that in order to make Devin smarter? Or there's these little things of like, you know, you want to go test the front end for your codebase and make sure everything looks as expected. You know, you should be able to understand like what is that supposed to look like or why is this different from what you know. If you find a bug, you should be able to understand like how you found the bug. And that's a lot of the little detail of honestly of what makes, I'll call it, the difference between code and software engineering, right? Which is 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. And I think that is actually the big problem honestly in AI coding. And I think that's the big thing that folks will work on next. The thing that's fun, by the way, it's a very practical problem, you know, and so it's a, you can call it a research problem, obviously in many ways it is a research problem in terms of like 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, you know, it's a kind of thing that you actually really just need to think about the practicality of software engineering to get it to.

Yeah, so it's funny because I think there's always an external perception and then there's an internal what's actually going on with the numbers. And without going too much into detail, what I can say is that in the last 6 months, you know, between January and now, you know, like even...

Aside from this, you know, obviously the latest deal this week with Windsurf, the usage of Devin has actually grown something like 5 to 10x. And that's been the case in both self-serve and enterprise. And so it's been fun to see that. You know, it's like I think the pattern, for what it's worth, is it's really like real engineering teams that bring on Devin, and they tag Devin all the time in Slack, they tag Devin in Linear and so on, and then they kind of, you know, use it and grow it and share it that way. And so in many ways it is not the same kind of like, you know, single non-engineer just going and signing up for an account can immediately just like build something really cool with it. People do use it that way, but it's not the majority of our usage, right? I think the majority is really just kind of like real teams using it.

But I think the point that you're saying does get to an important thing, which is that look, IDEs and the IDE experience are obviously, you know, they came and they really started working I think about a year before the agent experience really did. You know, I kind of think of agents as having taken off in the last six months or so, and I think of IDEs as, you know, sometime last year as when they really started to become kind of like no-brainer obvious, let's say, in terms of the value that they provided. And I think what we're kind of seeing is an artifact of that, which is that, yeah, like folks are more familiar with IDEs because they've been around for a while. I think you fast forward six to twelve months from now, and I think people will be familiar with agents to the same level that they're familiar with today.

The reason I would say that it's not also, but actually I don't even know the number. What are their numbers over the last six months, or what are their...

Yeah. Look, I don't want to get into exact numbers, but even with the Windsurf deal aside, like not including that at all, we've done roughly that as well in the last six to eight months.