Scott Wu 与 Russell Kaplan：软件富足的新纪元——AI 智能体带来 12 倍生产力（Ep 147）

Like a lot of the same people should be people who are kind of you know working across both of these. So catch us up since we last talked. Like like what what what's the state of Cognition? Where are we now? You're probably not giving out revenue numbers, but you're you've grown a lot. Like what can you tell us? Yeah, I know it's it's I mean it's we've had a ton of growth over the last I guess it's just under a year since we last talked um and obviously you know back in July we bought Windsurf, but but I I I I I I I I I I I I I I I I I I I think over the last several months I think both Devin and Windsurf have grown uh exponentially.

One of the fun stats actually is uh today's March 9th and we actually at this point have already done more Devin sessions uh in our customers in 2026 than we had in 2025. And so basically over the last like 2 months and change we've already done more Devin uh usage in total than than we had in all of 2025. >> So that's more more than 666% or something. Yeah, yeah and obviously you know we're we're we're we're working on making sure that that that growth trend continues and so we'll see how that goes, but but uh but no I mean I I I I I I I I I I I I I I I I I I I I think the business has grown a lot.

We've been working with um a lot of the the the biggest companies in the world, you know, Citibank, Santander um and so on um on on figuring out how we really you know transform their engineering efforts. Yeah, one of the one of the interesting developments since last year is um you know, when we launched Devin in March of '24, it was the first autonomous agent, right? It was like very early for the form factor. Anyways, I would describe it as kind of like just at the edge of possible then in terms of doing in terms of doing real useful work.

>> a lot more mistakes back then, obviously. >> it was much less reliable. You know, our infrastructure around it, the connectivity with the rest of the code base was a lot less mature. I mean, it took us until summer of 2024 for Devin to become the number one contributor to its own code base, like which was like the first real milestone. And then towards the sort of the end of 2024 to really get deployed in production at large scale at meaningful companies.

But one of the things we learned is that if you look at sort of where the technology was in 2024, it was not it was not reliable enough to to be used as the primary source of software engineering for most tasks. Right, you actually still need a tighter feedback loop in that year between AI and humans for most things. And so one of the sort of first niches we found where this was actually already really useful back then was in these very large code bases that just have tons of existing technical debt or complexity and you want to do a large scale refactor across, you know, 10,000 services.

If you're if you're doing it the sort of normal way as a human engineer, you might have to define some new architecture and then manually go implement thousands and thousands of changes and these changes would be complex enough that you couldn't just write a regex. But you know, not so complex that AI with you know, couldn't help. And so that's kind of one of the earliest places I think we got to real strong like product market fit is inside these like very large complex code bases.

And that's one of the reasons now if you look at the state of our business, we're deployed at a lot of the largest, most complex organizations in the world, you know, like most of the top health insurers, retailers, banks, government agencies and these places with large, complex amounts of code have been actually surprisingly early adopters of Gent software. Well, these guys just have like massive amounts of code that has been doing the same thing for 30 years on a very old architecture and you can go in there and pretty easily accurately fix that, I guess, huh?

Totally. It's actually it's like to your point on is this a new skill for engineering? It's you know, the the mindset of it architect inside one of those organizations has become, you know, you're almost like a a CTO of an agent fleet now. You know, you you you define the problem space of what you want to go solve, and then you just spin up your army of Devons to actually go do the implementation work, and then you kind of review the results.

So, it's a totally different way of programming. >> There's all these memes in San Francisco where like nerdy guys are on dates, but they're too distracted watching their fleet of agents instead of to talk to the girl. It's like it's a problem, I think. >> I have I have gotten in trouble with my wife for that like before. >> >> Let's Let's talk about this new era of software abundance. Uh as you call it.

What does software abundance mean? What's What's Cognition's role there? Yeah, you know, I I I think the simple way to put it is is I mean, to to Russell's point, all of these traditional industries, you know, it's you you In Silicon Valley, the way we think about software engineers is is, you know, obviously these tech startups or these tech companies that are building. The reality is every company in the world has software, right?

I mean, software is in many ways, I think, the the the premier knowledge work job of this century. Um and so, you know, you're you're you're talking about like CVS or you're talking about Walmart or you're talking about, you know, UHG or you're talking about, you know, Goldman Sachs. Like, all of these places have so many so many software engineers, and they have so much software that they're building um because obviously so much of what we do in the world now happens, you know, over the web or um yeah, with computers, right?

Um and I think what we we what we kind of think of when we think of software abundance is just getting to a point where you really can just turn your ideas into reality um and build what you want to build. Um and and one way that I like to think about this is, you know, you can kind of think about software products on a scale of like a let's say a log scale based on how much reach they have or how many users they have, right?

You think about the best products in the world and the products that everybody uses all the time, and this is like, you know, YouTube or Instagram or TikTok or something like that. And it's, you know, these are incredibly good products. Um and and it makes sense. I mean, they have billions of users, and so they have, you know, 100,000 software engineers or tens of thousands of software engineers that are working on them, right?

And you feel it in the product, you know, it's it's like the experience is perfect. Like, there's never bugs. It doesn't go down. It's streaming you like gigabytes of data. The algorithm is super addicting, right? Like they they really like perfected the I'm not going to use myself the two to three that you mentioned. They're too addictive. Yeah, they're too addictive, right? And then you go to the next tier of like, okay, instead of billions of users, let's talk about the products that have hundreds of millions of users, right?

And you're thinking about like, you know, banks or you're thinking about like, you know, apps like Uber or DoorDash or you're thinking about like, you know, a lot of these various other kind of services and products that we all use, right? And it's similarly it's like, you know, you feel the software and it's obviously built quite well, but it's already, you know, I think you notice a different level of like how much execution there is.

And then there's next level and a next level and a next level and it goes all the way down to, you know, your your like, you know, your your your kids' website, a school website which is from like 2001 or something and it's like an elementary school and it has like a picture that's like super outdated and has no other information about the right. And and so like maybe one way to put this is, you know, I think software abundance means making it much easier for everyone with every idea or every product or use case that they want to serve to be able to climb that ladder and build products as well as the best products in the world are built right now.

>> And and for someone for our listeners who are not as in the AI world, they might be CEOs of of a bank or something or running other businesses, like like what is AI actually changing about how engineers work, right? Like what does a great engineer's workflow look like now? Yeah, and that's a great question and and the simple way that I'd put this is that for us at Cognition, for example, our engineers don't type code anymore.

Like that's that's just reality. >> And and this is as of the last several months? >> And this is within the last, yeah, three to six months, honestly, when the shift has really happened. And and there are other steps, to be clear, obviously, but but I think at this point, you know, you you you used to maybe one way to put it is like, I mean, you used to work with punch cards, for example, and you used to work and and now in in many ways the medium has shifted away from code and a lot more of it has become basically English, right?

And so so, you know, we we obviously use a ton of Devin internally. We use, you know, WinSurf and the and the agents inside WinSurf internally as well. But at this point, either way, with whatever tools you're using, you know, it's not really you typing out the lines of code yourself. It's It's you looking, understanding what it is that you want to do, thinking about, okay, how do I want to handle this case or this behavior?

And you just tell the agent what you want it to do in English, right? And it goes and builds the rule. >> yeah, and in terms of then like what the impact of that is, especially, you know, if it's like a CEO level uh objective, what we're seeing across the board, across the customer base, is people are just getting more ambitious, you know? Like it It used to be the case that uh you have this entire software development life cycle, and every step of that cycle is oriented around not wasting the super precious time of engineers writing code.

And now you have this suddenly this overflowing abundance of the ability to generate code from ideas, from prompts, from ideas. >> faster on ideas. So you can just try stuff more. >> You can try stuff a lot faster, you know, you don't necessarily need to spend 3 weeks on design before you then hand it off to engineering because engineering's so fast that the turn, you know, the cycle time is much tighter.

Um and so, you know, increasingly, especially at the executive level, it's like, oh, do we have to choose between A or B? Let's just try both. >> I mean, could could the product people themselves create some things now, or how does how does that work? >> 100%. I mean, we see it every There's like a joke where it's sort of, you know, the the the engineer and the designer and the product manager all look at each other and say, I don't need you guys anymore.

Uh because they're all They're all just doing it all themselves, right? It's like every person is empowered to to do the other aspects of the product development life cycle. And And I think it's really rewarding people who are, you know, actually personally highly agentic and thinking about, okay, what's the impact I can go have um and and be sort of a self-reliant in that way. And I know, you know, with Devin, a lot of product managers, one of the very first things they they started using Devin for was actually to not bother the engineers with questions, with silly questions, you know?

How often have, you know, you're a new employee to a company, you don't understand what's going on somewhere, and you're a little nervous to say, "Hey, can you explain this to me?" You know, everyone Devin is very non-judgmental, you know, you'll ask a question, you got the answer, and >> I actually ask it a lot of them questions, too. It's like cuz cuz I'm the boss, I'm supposed to know something, like, crap, what does that acronym mean again?

But you would No one nobody you. You just do it right away. Totally. Yeah, and you you know, and you'll cite the it will sort of actually cite the source code alongside it, too. And and, you know, give it kind of puts everyone actually more on the same playing field in a way where everyone has the same context. And, by the way, agency is something I'm thinking a lot about for society recently. It seems like highly agentic people is like a big advantage for everyone, right?

But, but along those lines, like what are the habits or instincts of someone who's going to be particularly good at leveraging Devin? Like like are there certain things you're seeing? Yeah, yeah, for sure. No, I I I I mean, one way to put it um it is you know, I think software engineering for for people who who've grown up doing programming and so on and have been through all these previous eras, you know, usually what software engineering has looked like is like what why do people who love coding love coding, right?

And I think the answer is like 10% of that job is basically this really fun part of just pure problem-solving, thinking about what you want to build, being creative, like understanding the different solutions, deciding, okay, what architecture makes sense here, how exactly am I going to, you know, achieve my goals here, what do I want to build, right? And the 90% of the job is once you have figured out all those parts, just doing all the dirty work of the implementation to go make that happen, right?

And there's like a million bugs that your customers have reported for you, and you have to deal with this messy migration or this upgrade to to make your stuff still work, or like you have to go implement all the little cases and all the little details and like write all the, you know, the front-end code that that serves this thing that you just built, right? And I think what we're seeing is that the best engineers are just doing 10 times more of the first part because you don't have to do that 90%, right?

You have an agent, you have to have one that's going to go and do that for you. Um and I think to your point, that obviously means that, you know, having high agency is is really, really important. Um we think about this ourselves internally. It's it's one of the jokes is is like, you know, we're we're we're a reasoning lab building agents, and so the things that we really value are agency and reason. Um and I and I think it's to to your point, like a a lot of the skills that really matter are, you know, are you the kind of person that's going to think about, okay, this should be this way instead of that way, or like, what is the right way to solve this problem, or what what do we want to do here?

And also and also just like internally embracing the abundance mindset, too, where, you know, a lot of times you might be in your head, oh, should we do it this way or that way? I think a lot of our best engineers, they just rip it all of the ways at the same time. And then, you know, you get a bunch of devens come back, and then you can sort of analyze the results and and try them in parallel. It reminds me again, it's kind of like the mindset of actually being a good machine learning researcher is now relevant for every type of software engineering.

You know, if you look back a few years ago, what were machine learning researchers doing? So, you know, at Tesla we had one mantra, which was, you know, never go to sleep while the GPUs are idling. You know, if you let your cluster idle overnight, that's just a huge waste of resources. Just just kick off some experiments, you know, before you go to bed, so you can wake up, you get some more data. And it's it's the same for all of software now.

Like, why would you go to sleep while the devens are idling? Right? You could just be ripping a ton You could be ripping a ton of devens. >> devens working while you're sleeping. >> Exactly. Exactly. And so, I think there's actually a lot of parallels in my head where it's not just about that, it's also some comfort with non-determinism. You know, a lot of a lot of engineering, it's it's an incredibly precise craft, and there's it's sort of naturally uncomfortable, unnatural, to not have full control over everything that's going on.

But if you if you get a little bit more willing to embrace the non-determinism and okay, exactly how should this be implemented as long as we can validate system performance end to end and I can actually understand the results. I think machine learning went through that lesson years ago, and now we're going through it for the rest of software. >> So, I mean, everyone in a long while someone would probably check the machine learning machine Sorry, someone would check like the machine assembly code actually on something if they're really trying to optimize something.

Do you think people there's still people checking the actual regular code on things that they really have to these days? Or or is like, how do you think about that? >> Yeah, for sure. And for what it's worth, you know, I I I think we are we are still in the midst of a lot of this change, and so, I you know, I think while people are producing code with just English, you know, often you do come back and you're still reviewing that code, you're making sure that it looks right, or or you're looking at the code as it is to understand what's going on.

Um and I think to your point it is kind of like in the right cases when you want to peel back the layer of abstraction. Um and I and I think what we'll see over the next, you know, 12, 18 months is is that we will continue to get further and further to this point where you can just use English for for almost everything. >> You don't have to do it anymore. And and and what does this mean? So so obviously a lot's happening a lot faster.

We just we just covered this. This is and this is this is money movement. This is military stuff. This is this is health care. It's flights. It's maintenance. It's scheduling, building new things, and permitting. Like like everything in the world a lot more people realize is software. And and we're going now what, five times faster, 10 times faster soon on software? It's almost like we're getting multiple years done in one year, right?

So what what is what does that mean for society? How do you think about that? I Yeah, I think about this a lot. Have you ever seen the graph of inflation by different sectors? And you can see, you know, the sort of highly regulated constrained sectors at the top, you know, like tuition, health care spending. >> The costs of these sectors are broken and everything else gets cheaper. >> Right. And then the plasma screen TVs are just way down.

Way down. >> Japanese taught us what to do. And then, yeah. >> I kind of think that's that chart is going to happen for like all of society, basically. And in particular, if you think about where is software, it's we're going to enter this like hyper-deflationary cycle of software where it's it's so easy to build. It's so abundant. That doesn't mean that people are going to, you know, stop producing software.

They're actually going to produce thousands of times more software, right? And so you're going to have this total software abundance. But basically any problem that's kind of solvable in the digital realm should just get solved in the digital realm. And so I think what you're left behind with is, um you know, now how do we direct a lot of this new energy to go improving the physical world, go improving everything else?

You actually have to do things with real operations and real people in the real world. Now, I guess robots could change that, too. But but for now, anyway, with you put robots aside, the returns to actually doing things in the real world should go up by comparison cuz it's harder to make those cheaper, right? Yeah, we think about that a lot at Cognition of how do we how do we not just improve things improve software for software's sake, but how do those improvements basically translate into real-world benefits for everyone.

Um ultimately these things are tools in service of of our own lives, our companies, our our businesses, um and our livelihood. The real world's messy and you have to interact with >> messy. It's super messy and I think a lot of engineers uh fall into the into the trap of I I want to solve this very pure problem, you know? Um and there's nothing wrong It's actually really fun to solve pure problem, you know, a lot of our team they spend their entire programming careers just optimizing their ability to solve, you know, the purest algorithmic programming problems in the world, but it's an interesting contrast to, you know, one of the things that really differentiate differentiates cognition in the market right now what what we keep hearing from customers is, you know, we have this this four-to-one engineering team that will go partner really deeply with large organizations and say, "Hey, it's not just about the tools we're delivering, but how do we get into the weeds together to actually go drive like big structural changes in >> Tell me a bit more about that.

You guys have obviously deployed with a lot of the biggest companies in the world. What are some of the more impressive results you guys have seen in the wild? Yeah, so the earliest results that we where we where we saw me were like, "Oh, there's there's something here." where they started with um basically like modernization programs. So, people that had large legacy existing things they needed to transform. And if you sort of did the math to scope out how long it would take, maybe it would be like a two-year project.

And, you know, relatively quickly by like late 2024 we were measuring, you know, somewhere between a six to 12x productivity gain for those types of projects. Meaning that, you know, 1 hour of human time spent managing Devin was worth like six to 12 hours of that human time doing the work themselves. Uh and so that was a big that was like a big early result and we started doing lots of lots of engagements where uh our customers would use would use Devin to just refactor, migrate, modernize these large systems.

Now, the interesting trend that we're seeing is um a shift from really wrote uh reactive to proactive engineering work. So, you know, if you think of the early days of the internet, you know, most of the packets that were sent on the internet, uh it was like a human clicking a button or or visiting a link or initiating some requests. And then at some point it totally flipped and now most of the pockets are initiated by machines talking to other machines.

And I think we're now seeing the sort of the flippening there for software itself, which is the process of deciding what code to write. It used to be entirely a human scoped thing. Now we have people, you know, we have people wiring up Devin to all sorts of events and alerts inside their organizations to say, "Hey, let's just start the engineering work right away when something happens." I'll give you one example. One of the largest regulatory firms in the world, they they they do very thorough sort of security vulnerability scanning on their code.

And they make sure to understand is this potentially insecure here or there. And there's great existing tools for scanning that, you know, Sonar Qube, Veracode, Snyk. There's this whole There's this whole market of tools. And what they did was they hooked up all of those alerts they were getting from those tools and they just started piping them to Devin saying, "Devin, can you take the first pass triage?" Cuz if you're a human engineer at a company, you're drowning in these alerts.

It's not It's honestly not a really fun part of your job. And they're remediating 70% of these automatically now in production with with Devin. And so it's really flipping the, you know, kind of flipping the I've been thinking a little bit about the pyramids with regard to this lately, where it's like it's almost impossible to see how people with like no tools at all built the pyramids. And it does seem amazing if you look at the software that we have built up until like a couple years ago that all this exists all by people.

When you I feel like that'd be like something you'd look at 20 years from now when everyone just uses uses AI How do you imagine that? Yeah, yeah, no, I I think to your point, too. I mean, I I think there's sometimes a question about you know, we're talking about Okay, this is going to get way easier to build. It's going to be way cheaper to build these things. And so then what happens? And I think the reality is like we have so much more to build.

I mean, I always think about this even like today, you know, it's it's like you wake up, you're like you know, you're you're like let's say you're like logging in to check your medical records and it's like not a very good experience. You're like logging into your bank. It's like not a very good experience. You're like there's so much I know there's so many so many of these things and to your point, I think the reality is just there's so much more that we can do with software.

I mean, I think some of the things that people have started talking more about, which is I think what we're going to get to over the next little bit, are things like even like generative UIs or like single-use software, right? If you get to a point where you know, so so much of the work that you want to do is it can be done in code, it just doesn't make sense to like write code for a single, you know, for for something that you're only going to do one time or even something that you're only going to do like 10 or 20 times, right?

>> that the AI could actually write code for this instance coming up and give you the right code for what's going Gives you exactly the right thing and builds the yeah. And so this is I mean, people talk about Jevons paradox and I think I mean, in in in software it is perhaps I I think there's no where that it is more true than in software, right? As we have produced more and more and as we it's gotten easier, the reality is that actually demand has only gone up.

>> That's actually a really fun idea. Like right now luxury is like it's like vicuña, it's really nice, the stitching is right. But what if luxury was like it's a perfect UI for you just for this instance in your life where you happen to need it, right? It's kind of a funny idea. >> that or actually luxury might flip the other side because um artisanal handcrafted software is going to be so rare. It's like, you know, the in the early days of the of the industrial revolution it was like, "Oh great, mechanized mechanized labor for goods."

Like that was the higher status good cuz of course the machines were more precise, the the bags were more, you know, were more perfectly made. And now it's totally flipped. If something's handmade, that's um that's obviously much higher status cuz it's so rare and like how could you expend the resources for that? This website was handmade. You can tell because of all the little bugs and I think we're going to I think we're going to see that in in software.

>> >> Yeah, this is like handcrafted artisanal code. >> It's like it's like the Amish except that they're doing software. >> Yeah, yeah. >> >> It's like a 2020 society frozen in time. People making you handmade goods and all that. It's very silly. So speaking of speaking of like things that are broken and that there's like infinite need to fix, like government to me is like one of those areas where I think if you like were to graph everything in terms of getting better or worse, inefficient and less efficient.

It's It's like probably unfortunately comes off as like really, really messed up. And I've already started to hear some pretty interesting things in government. Um my friend Jared Kushner, who's obviously, you know, been involved in these administrations, he worked with Qatar and with Elad. And I And I they built something where the permits only take uh 120 minutes there. So, if you want a permit to build something, they'll get back to you in 2 hours.

Which is pretty cool. All right? So, there's like That's like There's all these ways government could do things better with software. Uh you guys just launched Cognition for Government. Like, what's what's the goal with that? What's going on? Yeah, I mean, I I think at a high level, I mean, we talk about places where where there has so much more software to build, so many things to fix. And I mean, I government is is is an obvious example of that.

And all of these departments, similarly, you know, things that we used to do by hand, not even that long ago, honestly. I mean, 20, 30 years ago, like lots of these things were done by hand at the Treasury. Now, so much of that is software, and yet so much of that software obviously still has such a such a long way to go. Um and I think from our perspective, you know, when we think about how do we make sure that the US stays in pace with what it needs to be.

we make sure that, you know, that the the breakthroughs that are coming through in AI in the private sector are also coming through in the public sector. Um it's it's it's something that's that's a really important problem for us to solve. >> I also think there's there's something poetic about it, because I think people don't appreciate this, but the government is really responsible for a lot of modern technical innovation.

You know, going back decades and centuries, uh Scott drew the analogy earlier to, you know, punch cards. And I think I think people don't appreciate this, but really the first wide-scale production use of punch cards was the 1890 census. All right? In the In the 1880s, uh they did the census by hand. They tallied it. It took like 7 years. And they were doing the math, and they realized that 1890, if they did the census the same way, it was it was going to be longer than 10 years.

And the census is every 10 years. So, they were screwed. >> [snorts] >> And so, the government basically put out this call for call for technology help. Uh you know, can we can we solve this problem with technology? And uh there was a There was a a guy by name of Hollerith who he invented what later became the Hollerith machine to use punch cards for actually running the 1890 census and it was on time, it was under budget and and it kind of kick-started actually a lot of a lot of modern software, you know, where where I where I studied at at Stanford, a lot of that the Silicon Valley ecosystem was actually really invested in by the by the government back, you know, decades ago partially for for defense.

And now we're in this we're sort of this flipping point even by the self-driving. I mean people DARPA doesn't get enough credit I think for for really kick-starting the self-driving revolution with the DARPA Grand Challenge. And now we're kind of in this state of the world where the government spends a hundred billion dollars a year on just IT modernization. We have huge amounts of government still written in COBOL with people who you know have left or no longer understand how how the code works and it's really holding us back and I mean you feel it as a citizen day-to-day.

Think about you know, your interaction with the DMV or or trying to pay your taxes. It's Or or waiting for a permit to come for something. >> for a permit. Yeah, I mean these things have real world consequences. Um and that's one of the things I'm personally really excited for is is you know, you can bring technology like this across the private and public sector. It's like a great equalizer for for answering work.

Can you deal with the COBOL stuff? Is that something Devin COBOL is a COBOL is actually a massive use case for us. Yeah, yeah. One of the early bets and investments we made was in what we call code base intelligence. So you know, if you think about a like a modern language model, there's a limited context window, right? But a lot of the large organizations in the world, they have very very complex code bases that do not fit inside a single inside a single context window.

We've done a lot of work both on the model training and RL side but also on the sort of harness engineering around that and the indexing around that to figure out how do you work with the really messy custom languages. And so COBOL is actually not even the the hardest one for us. One one of the examples I like is Goldman Sachs has invented some of their own programming languages. People don't know this but Goldman has like a pretty insane internal engineering team and they've they've literally written their own programming languages and they've they've been able to sort of customize and tune Devin to work on those internal languages too.

So COBOL is actually easier in some sense than that. >> So in the government, I mean they spend a hundred billion a year on IT. It's ironic because you're right, sometimes government in the past, especially when innovation was really expensive, they they pushed some new things that they otherwise wouldn't have happened. Uh, today most of that hundred billion dollars is spent by on special interests that don't seem to be using the money well.

So it's it's a giant mess. What are the types of projects you're working on? Totally. I mean the the incentives are obviously super screwed up for a bunch of reasons that your listeners are probably familiar to. One of the less One of the less real ones that I think we actually might be able to just side step is the government is a really unique buyer of software for a bunch of reasons, but one of them is that a lot of times they want to own the IP of the software they're using.

And this is a this is a really big implications for most SaaS businesses. You know, if you make scheduling software and your business is a SaaS business, you you don't want the government to own your IP. You want them to have a license to it to use it for scheduling. And actually that desire is literally incompatible with how a lot of government contracting has worked and happened historically. So you end up in a situation where the government says, "Oh, you know, this scheduling provider" This is a real example.

"This scheduling provider that has great SaaS that can do scheduling, I can't use it cuz I wouldn't own the IP, so I have to go work with the systems integrator and completely custom build my own, right?" Insane. Now, now what we could Could we lobby and go try to convince the government to change their policies? Yes, but actually I think easier for us to just side step the problem and say, "Look, Devin can just write this thing for you."

You know, they just build it for you. We're actually because of AI agents, you know, you're in the you're in the regime where actually now everyone can own their own their own IP um, more easily than before. So I think that's one of the ways we're sort of trying to side step some of this stuff. Can you can you be your own government contractor? Are you going to have your own government contract doing this then or are you just going to power others?

How are you thinking about it? >> mean right now we have dozens of of kind of FedRAMP deployments of Cognition, uh, both with agencies and with primes. You know, we work with uh, US Army, US Navy, the the Treasury, uh, we work with folks like Palantir, with Anduril, uh, and other primes. And so we just want to build sort of the most capable, most useful agentic software engineering platform and then and then work like heck to kind of get it in the hands of people to make it useful.

>> And in terms of your government, is this a fast-growing business for you then right now? I would say government is one of those things that it it really takes some time to kind of build and and be compliant and work in the way that people want to work. And then once you're there, once you're there, it's it's much easier to be helpful. And so, we've over the past year, we've really done a lot of the legwork to, you know, how do you get your FedRAMP certification?

How do you understand the needs of these agencies, which are actually pretty different in a lot of ways. Again, just in the thing just in the civilian sector, these they're operating on a completely different trade-offs, right? A lot of the software they use is actually by statute not allowed for them to write themselves. Talk about regulatory capture and intra you know, there are literally laws saying, "You government agency are not allowed to maintain your own website.

You have to bid this out to a to to a contractor." And and so, it kind of sounds insane, but you know, that's the way the system works. And I guess my experience working on AI with technology is that a lot of times it's actually easier to solve like a frontier science or engineering problem than it is to sort of change the molasses of the existing world, right? When we work on self-driving a Tesla, a lot of folks would ask us, "Hey, like why don't you make the cars talk to each other?

Wouldn't that be way easier for self-driving if all the cars could talk to themselves?" And the answer is, "Yeah, it totally would if you got every car talking to itself." But until then, you're going to have to deal with human-driven cars. And so then you have to actually solve the much harder science problem of predicting the motion of all these human cars. I think it's the same for our work with government. So yeah, I mean right now we have dozens of these deployments.

Um you know, I think folks are are giving us really good feedback on how much it's accelerating their work. And and so some of the some of the missions are are really exciting. You know, think of an organization like NASA JPL. Um you know, who doesn't want to help us get back to space faster? I love it. Yeah, we just actually interviewed Jared Isaacman, who is running NASA. So it's it's a great guy to partner with. It it it is true in government a lot of times people say, "Well, this solution would work if we just put this thing in the middle and made everything talk to it."

And I'm like, "Sure, that's what everyone always tries to do in government." But but these guys are never going to all work together. They're all going to debate, and so you actually have to design it knowing it's distributed. It's a It's a very interesting like you the real world's messy, and you deal with it as it is, not as you wish >> like the XKCD comic like we have a dozen standards. Like this is a mess.

We need one more. Now you you know you now you have one more. Yeah. Exactly. So So well I I it's a very honorable to go work in government. America needs that you're fixing it. Obviously, you're growing even faster in the enterprise in general. So these are both big businesses. This is obviously a very competitive time right now. There's a lot of the smartest people in the world. You have a lot of them here. There's some of them at other places as well.

I think very famously some of the big labs like Anthropic have been I think they've doubled into the tens of billions in the last few months. It's obviously the highest growth thing right now in the in in the general area. You guys are obviously, you know, don't say your exact revenue, but you're likely to get into the billions soon if you're not already there. What's the competition like? Do people use that with you?

Does it help you when they do well? Like how do you think about these Yeah, yeah, for sure. So no, I mean it's it's an exciting one obviously and and and software engineering and code is just so big that I think there's there's there's a ton to do. I mean, we've seen a ton of growth as well. I mean our our usage for example of Devin in in our customers has I think roughly tripled already since the start of this year.

Wow, so just the customer inside the customers alone has tripled >> Inside the customers alone. That's right. But but but what I would say I I I think a couple thoughts here. One again, there's so much different work to do go do in code, and so you'll see a lot of others who are for example um building products that will help you make a quick little website or something like that or you know, build something like fun.

Um which is, you know, you can use Devin for that, but I think that's not necessarily what we specialize in. On the other hand, a lot of what we really really focus on to Russell's point is is is working with um enterprises, governments, you know, regulated industries, working with massive massive code bases, and trying to make sense of that and and work in all of those systems, right? And so a lot of the problems that you have to go deal with and and work on are how do you, you know, absorb all of the messy context and the knowledge of this code base?

How do you work with a massive, you know, something that has hundreds of thousands of different files and work across that? How do you test and integrate against your, you know, your your existing unit testing framework or how do you uh you click around and use these products yourself and make sure that that the edits that you made were good. And that's a lot of what we've always focused on with cognition and so you know something that something I mean basically all the guys that that that you mentioned here in terms like the foundation labs and so on.

We actually partner with them and we work very closely with them. I think what we tend to find is that um you know for for obviously they're kind of like base research and and the work that they do with models there's there's a ton of interesting work for us to do together. Whereas I think for for a for a lot of this work with really enterprise transformation you know we we won't we we want to be like on the ground working very deeply with folks and and and figuring out with them how how how we use software to to to help them achieve their goals.

>> So Palantir originally we created this thing called forward deployed engineers which didn't really make sense to people 10 or 15 years ago and it turns out there was like certain types of workflows where where you could build a product that do a lot but then the forward deployed engineer would actually have to understand the business value and and kind of connect the dots and and then and then take things that they created that oftentimes go back to the core so the core gets better so the core could do it next time.

Like do you have you sound like you have something similar to this? Do you have a lot of these? Is that is a big part of the value you're providing? It's yeah so it's interesting. There's there's some similarities and some differences for us with the sort of the Palantir model. So one of the one of the interesting differences is that a lot of times you know our customers they just use our product on their own even without us talking to them or without us discovering them.

You know people bring in our tools and they just start using them immediately. But once once Devon is inside an organization you know the ceiling of what you can accomplish if you're a world-class agent manager you know versus like a you know a new engineer who's just sort of learning the tools it it's an enormous delta. And so a lot of the there's kind of some parallels to our business that are more like a you know kind of a data bricks or snowflake type thing where you just kind of get in and then and then people start using your products more and consuming more.

On the other hand if you actually want to go drive major outcomes like real business change that results in like structural you know, oh, I can launch this entire product line that I wouldn't have had the capacity to, you know, otherwise. Like that that type of outcomes, um you know, our 450 engineers are among the best in the world at managing agents at really high scale cuz that's like exactly what they they sort of focus on and do every day.

So at Appen here we had these frameworks of like like building an ontology of all the data making it talk to each other ontology of processes and we had all sorts of different frameworks over time their conceptual frameworks we'd use when you go in. Do you guys have like your own conceptual frameworks for business value and and do you have something called ontologies like how like not to not to get the secret sauce but are there things like this you could you could tell us?

>> we we really look at it from the perspective of the software development life cycle. So we go inside an organization, they are they have a way of doing things, right? Of of developing software starting from planning and deciding what they even want to write, understanding all of their existing code and process to then maybe scoping it out and and maybe writing some code, testing it, fixing it when it's wrong, iterating on it, deploying it in production, monitoring it.

You know, there's a pretty standardized software development life cycle at this point. And what's happening is agents are just eating more and more of the cycle. And it kind of started with the writing of the code. And now we're like well past that, right? And so in fact one of the more recent products we we shipped is called Devin Review. It's because we observed that there's this totally new bottleneck in the software development that in the software development life cycle that wasn't the case previously which is there's this abundance of code being written by AI now.

How can humans even keep up with it all to understand what's going on? Again, we work with, you know, a lot of like regulated large complex organizations that are they're running mission critical systems and you can't just sort of vibe code, you know, your way and like yolo merge The trea- the treasury shouldn't be vibe coded No, exactly. >> >> No, it's like actually really important and and so that's why eventually, you know, are we going to like have English as a source of truth and and people are just going to be collaborating on specs?

Yes. I think, you know, March 2026? No. You still need to understand the code that you're merging. And so we, you know, a big a big thing we care about at Cognition is we're building tools that are like future looking but still still meet people where they are, right? We still want to meet people where they are. And so, Devin review, it's a it's it's not just having like an AI auto comment on every PR and say, "Oh, this was good or not good."

It's actually tools for humans to really deeply understand huge quantities of AI-generated code. Is there some sense that at some point one of the AI models, if it gets a little too tricky, could like sneak something into the the things that are being built to like do something we don't want it to do? Yeah, so it's a great question, obviously. I I think with a lot of these things, the the the simple answer in software is you want to be working with the same review processes and the same QA processes that we all have, right?

And so, any any big enough engineering organization, frankly, already has to think about this just with their humans, you know? And and it doesn't have to be on purpose, obviously. Maybe you accidentally introduce a security risk or so on, right? And that's why you have review, and that's why you have QA, and that's why you have user testing, and that's why you have release cycles, and all these other things. And I think um you know, I I I I I I do think this is one of the one one of the things that often comes up, which is like, how do you make sure that um that your agent behaves correctly, which of course is like a very important problem.

The reality is that I think in software, we have actually a lot easier or maybe at least a lot more grounded uh of a path to do that, because we have the same thing already with all of our humans, right? And so, when you work with Devin, for example, Devin is is is making commits, is submitting, you know, code diffs and so on. Devin is not allowed to go and deploy your code to production by itself or anything like that, right?

It works in all those same systems and has the same guardrails. And so, let's let's just fast forward for fun a few years in the future. When I talk to some of the people running the top labs, they're pretty convinced things are going to keep getting better at a pretty high pace for at least two or three years. It's like kind of like with Moore's law, you can't really see out what's going to get there 5 years, but but see it feels like things are going to change a lot.

Like so so so tell us about 2028, 2029. Like is there certain things that just look very different, or certain things that are that are like we have to do now, we don't have to do at all there? Like what are the unsolved problems for Devin to just be be like doing a massive projects on itself in 3 years. Yeah, I think a couple shifts. I mean, one of the obvious ones, which I'll just call out, is just much more widespread usage of all of this and then just good knowledge on how to use these things.

I think right now, you know, you have this core group of we'll call it like agent forward engineers, right? Or agent forward companies that understand how to use this and they are seeing these, you know, 5x, 10x productivity gains as a result. I mean, obviously, you know, all all of these big organizations or these these companies or governments or things like that are are seeing the same results and realizing, "Wait, we can't just sit here and be five times slower."

And you know, we have to go learn how to do this right now. Um and so that's really happening. I mean, even this year, I would say. Um in terms of the the the continued capabilities gains that we're going to see, yeah, I mean, I think there's um I think that I think the models are going to get better and better. You know, one of the stats that people talk about a lot is this METR report, which basically says for um each different model that comes out, roughly how much human work can it do uh in an automated fashion before you have to go interrupt it and say, "Oh, that was wrong.

Let's go do this." Right? And so it's like, you know, just two or three years ago, the answer was like 10 seconds or something, you know, you would have it write one line and then it's like, "All right, the next line is already wrong. You know, let's stop here." Right? And at this point, it's already gotten to the point where it's um in the scale of, you know, 10, 20 hours um is what the latest one uh has been. I think Opus 4.

6, for example, I think it was around 18 hours or so. >> This is always very weird to me cuz it's implying that models work in human time though, right? Which is so weird. >> I mean, it's like 100 times faster. Yeah, so so so the answer the the they do typically do the tasks in less time than a human would. And then the 18 hours is basically this is how long it would take a human to do that amount of work Okay.

uh in in between each of the interruption points, right? And and the the the AI might be doing that in in 1 hour or 2 hours or something like that, right? Um and and the thing that's really crazy about the stat is you just see it very consistently double. And I think for the last few years, it's doubled about four or five times every year, you know, which is insane. Which means, you know, you wait two or three months and it's already doing twice as much work.

>> world's changing every two or three months in terms of what's possible. Yeah, and this is this is what we've what we've kind of seen as well and I think we're going to we're going to see even more of that. And to to some of Russell's previous points, the thing that's kind of interesting for us is that the form factor of what you want to deliver or how you want to work with the AI changes a lot as you're going through that, right?

And so when we're saying, "Okay, you do 10 seconds of work." Obviously, the answer is you as a human need to be staring at your file of code and like shepherding it and hand-holding it with every single step, right? If you're talking about it's doing days of work or weeks of work, now you're actually giving it whole, you know, output level tasks of like, "Hey, you know, we we really need to go make this app much faster.

Can you go and run this whole thing?" And then, you know, do a smoke test and make sure all the changes look right. And it's going to go off and do that entire project, right? Versus or or it's, you know, even even bigger initiatives of like, "Yeah, can you auto-respond to all of the um you know, the upgrades or the potential vulnerabilities that are coming in with our reporting?" And just make sure all of that, right?

Um so, you're going to see a lot more, I think, proactive work. You're going to see a lot more basically like event-driven kickoff work where it's it doesn't have to be human that's that's moderating it every step of the And then and then in terms of like, what does that mean for society or where does it go? I actually think one thing I'll go on the record on is I I think there's going to be an explosion in small businesses.

I think AI is actually like an extremely small business enabling technology in particular where you think about what's hard about about building or starting small business, it's you know, you don't have the resources of a large company, that's specialization of labor in each part of the in each part of the process. And I think AI it's extraordinarily enabling, right? Think about, you know, the quality of quick legal gut check you can get from, you know, from a chatbot, from a frontier lab.

The the quality of, you know, analysis of your financials, the quality of software that you build, right? It's like it's all coming together to empower each individual person. Again, if you exercise that agency to just do so much more on their own. I actually I love this. I actually have a a small thing on the side where I'm trying to help create tens of thousands of small business owners. So so I'm totally aligned with this.

This is a really good theme right now for us for us to do. One last thing I want to ask you guys about the business that I'm just so impressed by. So I I hired a lot of the first 200 people at Palantir. I spent a lot of time on talent. Obviously, I even hired Scott at one point a long time ago with one of my companies with with Vlad's help. How did he do How did he do in his interview? He's very very very very impressive.

I But I did not I did not see him at at the time as someone who was like a CEO person. So he really grew a lot, which is good. I mean, you know, he's just like he's learning and growing as he goes. And it definitely seems as a CEO and founder now. Um but one thing you guys have done is like a significant percent of the Cognition hires are actually former founders. So not only are you hiring the best people in the world, you're hiring a ton of former founders.

Like like like why are you doing that? How are you doing that? Tell us a little bit about the talent stuff here. >> Yeah. Yeah, for sure. No, I mean I I I I think the um I I I I think the reality is we we just have such a massive problem that we're going after, which is, you know, solving all of code. And I think the way we even started this company is, you know, Russell was a founder before this. I was a founder before this.

All of us were, you know, I think of our of our kind of like initial crew. And and and the idea for us, I think, was let's make this one the big one, you know? We're going to go for it all. We're going to go for the most ambitious Even I mean, I've the the the play of solving software engineering feels like a big enough one, you know, that that we can all do that together. And and and I think that's a lot of what it comes down to, honestly, is just like um yeah, are are we working on something and doing something that's that's that's exciting for for folks who, to your point, are are, you know, I I I think a ton of the folks at Cognition are could could very very easily go off and start their own companies and get funded and build their teams and so on.

And and and the question I think for us always has been about um how how do we make this the place that that that makes more sense for them to do that? And it's it's honestly easier to do that now than ever before, also, as as a company. Like I'm thinking, you know, we have one team in the company. Uh they're they're called special projects engineers. And basically, every person on that team is a former founder. Like every single one.

And and they they do they do, uh you know, a really interesting mix of engineering work of product work of talking to customers of driving commercial outcomes. You know, the problem space to to your point is so big that actually if you're again high agency person going to take initiative and you're working at a small fast growing company with with a you know, with a problem space so big that you know, your only constraint is like your own ambition.

Well, there seem like there's this renaissance or revolution going on in the world where the capabilities are doubling every two or three months and like this is a place you can come and be around some of the other smartest people in the world or part of growing something without which I guess is pretty fun for people. We had a good time. Well, well, one thing about the the interview process or the selection process to your point which I think is kind of interesting to call out is I think um you know, one or two years ago I think a lot of people had this mentality in terms of how you interview of like okay, there's all these AI tools.

How do we interview people in a way that um you know, how do we make sure that people aren't using AI while we're interviewing them cuz then we're just and and and I think that is totally flipped. Cuz honestly, I think that was wrong. And I think if if if you're asking the question of like how do we evaluate people on exactly the thing that AI can already do? That's kind of the wrong and and and so so so you know, I think for for us and this has always been the case for us, you know, our interview process has always been you can use as much AI as you want to use, you know, it's just like we're going to give you a few hours just build build your whole own product surface, right?

And build build your own um a lot of these are projects that frankly is like if you were trying to do this by hand, you would not be able to get done in a few hours. So you kind of have to use AI for this, right? But but in reality what we actually want to test is in addition to it to kind of how familiar you are with these tools, what we actually want to test is is yeah, what what what do you think is the right thing to build or how do you make these product decisions?

How do you make these trade-offs? How do you decide or collect information about what you should be doing? Um and and and so we found that that's that's helped us a lot. I love it. Well, we started the American Optimist podcast to try to push back on cynicism and pessimism in our country. What's the best case for an optimistic AI future? What what inspires you about what you're seeing? Yeah, no, it's a great question.

You know, so so funnily enough, I think recently we had this whole Citrini piece come out, um which I thought was frankly ridiculous. Uh I mean, I think it's it's I think it gets a lot of the the basic economics wrong is maybe a simple way to put it. And and it's it's I think it it it goes off of some of the same things that we are calling out, which is it's going to be easier to build things, it's going to be cheaper, and so on.

And then somehow concludes that the you know, the outcome is going to be much worse for all of us. And I think maybe that that you know, from from a from an economics perspective, there's a distinction between what is um real versus nominal, you know, deflation, which I think is maybe one one thing to call out here, which is of course prices are going to get cheaper. But why should that mean that we're all worse off, right?

Um but but but I I I I I I think the simple thing that I'd call out about AI is I think we right now are at a point where so many of the things that we want to build and so many of the things that we want to do are just hard bottlenecks by pure execution, right? And I think we're pretty quickly getting to a point where that's not the case. Um and my my favorite line on this is uh you know, our co-founder Walden says this, which is for so long we've all been living in Minecraft survival mode, and now we're going to be in creative mode.

Um and and I think that's really I I I I think that's what we're going to see over the next honestly, the next 5 or 10 years is getting to a point where you're really only limited to your ideas and to your imagination, where you can kind of just turn things into reality. And and I think that's going to be a great future. Awesome. Well, that's inspiring enough to leave it on. Thanks, guys. Cool. Thanks for having us.

>> Thanks for having us. Yeah.