Scott Wu (Cognition AI) on Software Abundance and the Future of Software Engineering

Cool. Yeah. No, first of all, thank you to everybody for coming today. Yeah. So I guess a few thoughts. I mean, so we came out with this video in March of 2024 and arguably we were too early with it. You know, it was very reasonable that you could make a case for it, but it's very much like, okay, you're going to have full autonomous background agent. It's going to be able to test its own code. It's going to be able to do computer use and so on. A lot of things that I think folks talk about today. I think a lot of that has happened. I think there's still a few kind of shifts to come. I think obviously agents in general, people have very much adopted. I think most people don't really type code anymore as they should. But I think there's like a few levels to that still.

I think a current big one that's kind of happening right now is the whole debate of local versus remote agents, like agents that exist and work on your machine versus like in a cloud VM setup. And obviously like our view is that things will move more and more and more towards cloud. And I think that's been kind of happening. I think one of the next couple things that we see, I mean, I think one of the big ones actually is like really good event-driven and autonomous agents. And so, you know, one way to kind of put it is like it's pretty cool that you can just get like a Slack message or like a Linear ticket or something like that and then you can just like spin up an agent that responds to it. But also like you probably shouldn't even have had to do that in the first place. It's like the agent should have automatically spun up as soon as this thing got started. And you can imagine there's all sorts of different triggers that make sense. Like you could have something that like checks all of your different API routes in your product for performance and if like one of them is slower than it used to be a week ago, then it spins up the dev for that, right? You can imagine something that like, you know, goes through and does all the like, you know, the recurring like Dependabot upgrades and just like make sure all of those are in a good place. Like all this, you know, it should be your first thing on pager duty. It should be your like, you know, first thing whenever you get like a Sentry alert or something like that. So, so like setting up triggers for all of these. And then I think another big one that we kind of see coming and we're pretty excited to build for is just like moving a lot more things to English as the true like source of truth. Um, and the way I kind of describe it is like now you can write code with just English, but we all still think of code as the source of truth, right? Like the code, like the codebase is what you look at to figure out what's going on. The code is like when you're doing reviews, you're reviewing code. When you're like, what do you commit to version control? Like it's code. Um, I think we will get to the point where all of that is actually just English. It's like product specs and so on. Um, and I think some of the products that we've been working on and building like Dev Review and so on are kind of trying to push more towards that direction.

Yeah, I remember that conversation. We were talking about like, well, like it could happen, it couldn't happen. What do we think? No, I mean, I'm very, I'm very thankful for them that they did that. Okay. So, so here's what I actually think. I think it makes a lot of sense for Elon. Elon has a ton of GPUs. Obviously, he's been like figuring out his research team and everything. I think the Cursor team is a great team. I think from Cursor's perspective, I was a little bit surprised from Cursor's perspective. I think they had lots of things to continue and so on. But obviously I think Cursor was like, Cursor was burning money and, you know, they were also kind of in a crunch of trying to figure out compute. I think from our perspective, it's, I mean, it's been amazing news for us even in just like last couple weeks. Like last week, for example, our usage was about, it was up about 30% week over week, which is kind of an insane kind of growth rate that we saw. And I think there's like a few reasons for why that is. But I think at a super high level, it's we've always, you know, really valued being at the application layer, like really focusing on the product and the experience and the customer value that you can provide. And also we've always really valued being like the neutral party and being the Switzerland that actually works with all the different model providers and so on. And yeah, I mean, now I think we're like the only one in code that does that.

Yeah. Yeah, so I think a lot of stuff that's still unsolved today. And I mean, I think like in some sense the true terminal, like what is the actual end state of software engineering? It's kind of like you should just have the Neuralink plugged in, like you should just not even have to like say anything, you know? You just think your ideas and then the ideas become reality and it's like exactly what you want it to be and like it all maintains itself and everything. I think we've made a lot of progress towards that in the last couple years, but I think there's still like a long way to go in terms of the pure capabilities. Honestly, I would say like working through the whole life cycle of code, it's actually pretty, there's a lot of looking through everywhere. And so it's like I think initially what folks were most focused on was specifically the part of like writing the code. And I would say to first order it's actually like relatively perfect at that. Like if you ask any model today, okay, produce me like a single file that does X, Y, and Z. Like it just works. And now obviously it's like when you think about the different problems that exist, there's like understanding a codebase in the first place. There's like figuring out a tough bug. There's like going through and like figuring out how you communicate the information with a human and like present that to a human for review or for any of these other processes. There's like thinking about how you adhere to style or like the general guidelines of what makes sense. And all these things are actually like I think most of the base models today are really just like not optimized for. And so we see a lot of room to work more in that.

Yeah, for sure. Yeah, so on your point of, you know, being an agent lab versus like a frontier lab, the way I kind of describe it is like I think as an org, at the end of the day, you really do have like one main mission and you do a lot of the different things, right? But like the question is like what is the mission that you work towards? And I think for a lot of frontier labs, for example, it is truly like a, you know, one way to describe it is like for a frontier lab, it's very much like a just pure like pursuit of truth, like finding the base of like intelligence. Right. I think for Cognition, we've always thought about it as the thing that we want to solve for and the thing that we want to go do is we want to go actually work in the real world, help people produce a lot more with software. And so like there's like the famous line of like software eating the world, which I think is actually even more true now, you know, but like now like AI eats software. And so obviously there's just a ton to go do with actually making it and making it much easier for people to build software. And as you can imagine, those involve a lot of the same things actually. However, I think the way that you go about them is somewhat different. So, you know, versus a frontier lab, like a frontier lab will obviously like we'll do model training at the same way that a frontier lab will do model training. We'll have a product the same way a frontier lab will have product, we'll have revenue and whatever, you know. But I think the orientation and the approach is pretty different where we always think about things from a pure like solving real problems for the user first. And so when we do research and when we do model training, for example, like all of our projects are oriented around like, yeah, what do we think is actually going to concretely make the Devin experience better or what is going to unlock the next thing that allows people to, you know, to work from a lens of product specs or something like that. And so it kind of goes all through the set.

Yeah. Yeah. So now I think that there's a crazy stat. We had like in our first 50 people, I think more than half of our team had founded a company before. And I think there's like a few reasons for that. I mean, I think for us, the way that we operate is very lean. Again, maybe to a fault. Like I still wonder about like whether we should go and hire much more, but like we've always wanted to keep the team super tight and super elite. And so naturally, obviously like folks who are like good enough to go and start their own company, like it's a very exciting thing to go do. But I think the other reason too is just like I think we really want to operate from like pure first principle. We want to have people who like take ownership. And so like, you know, for a company of our size, I would say there's relatively little of the whole, you know, whole like multi-layer management structure with bureaucracy or whatever. Like we kind of want people to just like, yeah, like just do things that you think make sense and just like be an owner and act as an owner within the company. And so I think that that profile has made sense for us a lot.

Yeah. Well, you know more about these rumors than I do. You tell me, Brandon, what are the rumors, true or not? Okay. So, yeah. A few thoughts. I mean, I actually, I mean, I do think we should be bigger. One thing I'll say though is I think people underestimate the extent to which just having a really, really high bar is valuable. And of course I think you can go too far with that, but I think like that's kind of always how we've operated and thought about things. And like the number one thing that we say at Cognition all the time is like it's fun being on the Avengers. And I think that's like I think a good way to kind of describe it. Like you want to be able to just like look around and know like, okay, like I'm building this thing and like the product engineer working on this is like insane at what they do. The inference person working on this is like incredibly correct with like what they do. The like, you know, the forward deployed engineer is like super on top of things, like understands all these details really well, is like, you know, like working very tightly with the customer, like the ramp is like working, you know, it's just like it's more efficient when things are growing so quickly and when you're kind of like, you know, like managing a ton of things moving at once, it's also just like a lot more fun because you get this feeling like, like silly example is like, you know, about a year ago now we landed our first big US enterprise major which was Goldman Sachs and that whole deal, it's like some big enterprises usually when they buy a software, it's like typically like a 12 to 18 month cycle of like, oh, they have to go evaluate it and then it has to get through security and then you have to go through procurement and then whatever. We did the whole thing start to finish in like three months. And we had no business going and doing that. Like we didn't really have like customers and stuff. You know, we have a product that hopefully was like pretty interesting, but it's like, you know, there were like a lot of people that were like that the in some ways the kind of like limiting factor or the bottleneck on making sure that happened and everybody all pulled through, right? And it's like there were a ton of things that we had to go get done on the info requirements to make that work. We had to go deploy in their private cloud. We had to go solve for all these particular things. Obviously, a ton of things that we had to do on particular products to like really get, you know, this was like early 2025. This was before agents were like that accepted and like going and making full autonomous agents work for the kind of use cases that they wanted to use it for. There was like all of the work of just like really working deeply with their team and like doing the user education of showing them how to get to value or how to find the right use cases to use Devin for. There was like the actual sale itself. And it's like it's just like fun where, you know, everyone in the room is like, you know, someone that you can really, really trust.

Yeah. Yeah, for sure. No, I mean, a lot of great folks honestly like across the board. I mean, it's, so the person who we hired who was 17, well, there were a couple actually. One of them is a guy named Mox. There's another named Ryan. Both who are like incredible, incredible engineers, somewhat different backgrounds. Moxin had built a ton of different apps before and was like, you know, like a really thoughtful like product engineer, but also had done really great like infra work and came and worked with us and then ended up like pushing back going to college obviously to go work with us which was sick. Ryan came from like a competitive programming background and he was like, you know, an international gold medalist and then came to work with us on like a lot of the agent building side and that was really great. But no, I mean, I think a lot of it too is like figuring out how to make exceptions for exceptional people is also like one way to put it. Like I think there's like, you know, I can say of course it's like important to hire great people, but I think everyone really knows that and every, you know, it's like the first thing that you hear any kind of startup advice. I think in practice sometimes the harder thing is like how do you make the company the right place for like really great people and I think there's like tons and tons of little things there that like really add up. So like even just like from very early on we actually always had like chefs who made lunch and dinner at the house and it was kind of like a funny thing because it's some, you know, you could consider it wful but I think it was like a nice little touch that made it more fun. I think similarly like the way that we've always set up our culture, it's like, you know, it's great to have a team where everyone feels like they really want to pull their weight and want to make sure they're like, you know, doing their best because they see everyone around them. And then I think another obvious one that I'd call out too is like a lot of the best people like really want to start their own thing. And so naturally it's like how do we make this the right place for people who want to be founders to learn all the details and the ins and outs of like building a startup and not just like one singular skill that they're going to spend their whole time on. All these little things I think are really important.

Not that much. I don't know. I mean, like I kind of think like people always kind of ask this question of, you know, like how do I make sure that I'm like still going to have like a good job like post-AGI and I kind of think the answer like, well, like we're going to like there's going to be some interesting middle steps, but I think in the long term we're kind of going to transcend the need for jobs. I actually think and like here's one way I'd put it is like if you just consider what all of us here do today. You know, everyone's job I think is some mixture of pushing buttons and like maybe occasionally going and like calling other people or like being in the room with other people and having conversations. And so like our ancestors from like, you know, thousands of years ago if they looked at us and they saw us do this and all this work and be like, well, you know, I'm in the fields like seven days a week like this. What do you mean that's work? Like you're not even doing anything, you know? And I think and maybe my point with that is more just like I think the world will change in a lot of ways that are even kind of hard for us to appreciate. But I think one thing that probably is true is that just like providing the pure basic necessities, like I don't think that we will rely or have to rely on human, you know, economic output in order to do that which is pretty exciting.