Scott Wu on Cognition's fundraise, AI solving Navier-Stokes, and why cybersecurity is 10% of Devin's sessions and growing fast

Speaker 2: Apparently, have to ask Scott about the number 46. Do you get this reference? I don't know. Okay. We're gonna ask him because he's here in the waiting room, and let's bring in Scott Wu, founder, CEO, Cognition. Welcome to TV CNN, Scott. How are you doing? Great to you, man. Thank you. Me. Had a slow slow news day, I was gonna say. Huge news day. You guys to talk about. Congratulate. Yeah. Crazy crazy news day. We're gonna talk about math. We're gonna talk about AI. We're gonna talk about fundraising business. But first, the number 46.

Speaker 4: I was I was told to ask you about this. Why? That's funny. What is this? Well, saw the my Twitter handle is Scott with forty six. Okay. The the reason my my handle my my I'm I'm like Scott with forty six on basically all these platforms, is because when I was in elementary school, biggest thing I knew of was this middle school competition, Math Counts, which is now made a bit more famous because people have seen these videos of Math Counts and stuff. And in the math counts, there's like a written proportion, and then when you do well enough on that, you go into the actual, like, head to head. But a perfect score in the written portion is 46. 46. And so then what ended up happening is with this particular round, we actually ended up pricing it at 46, pretty partially for the meme, partially because, you know, ended up being the right kind of You did the meme. Going all that. But yeah.

Speaker 2: Yeah. You you have to imagine that certain VCs are like, okay. We've agreed to this number, but are there any insider meme references that would save us even 3% here? Let's I know. Do a deep dive to see if we can get the founder to give us a little hint. Advantage. Yeah. Yeah. No. But congratulations. Tell us about the fundraising round. You got every investor in the world on the cap table at this point. Is that right?

Speaker 4: Yeah. Look, you know, it's an exciting time for Cognition, obviously. And I mean, I think the biggest thing I would just call out is like agents are just getting really, really good, you know? And I think two years ago by the way, we're not that old, like two and a half years So so two years ago, you know, was when we kinda did the Devon launch. And back then, as you remember, it was, you know, went pretty viral and so on, but but it was really just like a prototype. Would And there was was so much skepticism. People were like There was a lot of They are coining a buzzword that will never exist. Agents, they're just hyping this up, and now it's like everything is agents. Yeah. I remember thinking about this because I was just like, man, it's it's great that Devon actually works because if it didn't, we would just look so dumb right now from Yeah. Oh, yeah. Totally. But so so that was two years ago. Yeah. And, like, we didn't have customers, you know, it's very much just like a prototype and like, here's what we're building towards and so on, right? One year ago, I would say it was very, you know, the product existed, it worked, it was for very specific use cases, that you would kind of have it work and have it do it end to end. But I think most people in the world hadn't really gotten across to the idea of using a background agent. Now I would say it's, I mean, especially in our circles, it's like, it's becoming a more and more commonplace thing. A lot of it is just that the capabilities are just so good that, obviously, you should go and delegate to something that just does the task entirely, right, and be able to manage these and so on. So so so that's been the biggest thing. We yeah. I mean, the the agentic era arrived. The importance of harnesses like Devon arrived.

Speaker 2: How how do you think about orchestration? It feels like we saw a glimpse of that with with Gastown that went viral now. Whenever you fire off a prompt in any modern AI system, you can see the harness and the lead agent sort of talking to sub agents and delegating things. Orchestration a meaningful leg up? Is it another exponential in the sense that we went from LLMs to reasoning? That was an order of magnitude. Then we went to the agentic era. Have we always been in the orchestration era, or is this a new era? Is this a new meaningful change? Are there new disciplines that need to be explored in the world of orchestration?

Speaker 4: Yeah. No. So so it's a it's a really important question. I think there's there's definitely a lot of, like, what's the word? There's there's definitely a lot of kind of, like, you know, misconceptions, I think, out there about it. I will say that there's a sense that people get of, like, oh, like, if you just say the magic words to the model, it will suddenly become 15% smarter or something like that. That's much less of a thing, especially now because the models are all RL's on very specific capabilities and these kinds of tasks and so on. I think what is much more the case is if you, number one, if if you go and bring in all of the context and the information and the systems that you need, for example, you know, in coding, a very simple example is like a model and an agent that can go and test its own code and click through its website by itself and then go and look and say, okay, was right. That part was wrong. Let me go fix that. It's obviously gonna be way more capable than something that can one shot it on its own. Right? A model that can go, you know, an agent that can go look up in the Datadog what went wrong in the logs is like much more powerful than something that can't. So that's one is just like being able to bring in all the tooling, the context and so on, all this kind of like messy real world stuff. How do you navigate a big code base? How do you get through, you know, all the secure systems that are, you know, that a company will have on its software? And then number two, would say is combining the strengths of the different models and the things that they're good at. Right? And so, you know, I I mean, everybody's talking about price performance of model. Everybody's, you know, showing the charts of the Pareto curve and so on. And this model is like really cheap, but is good enough for this percent of tasks, and this model is, like, the way expensive one, and it's, you know, this much percent better and whatever. Right? But, obviously, what that means is that by combining the the different models, as long as you know what use cases to route to each model, then then you can do a lot better than than any one alone. How permanent is that

Speaker 2: task, that job of picking the right model at the right price point? Because we see there's a lot of attention from the from the big labs, and it's very clear that they're that every big lab is going after building beautiful slides, front end, back end, cybersecurity, bio, like math. They're all working on these, and there's this trade off between a certain model might be really good at something else, and then the new big model comes out, and it's better than everything at everything at every price point. But how and and I I think the risk is that you can you you can wind up in a situation where you're like, oh, well, if I just wait two weeks, I'll just be able to use the the the major hammer because everything will look like nail. But it feels like there's also always going to be this cost optimization, speed optimization.

Speaker 4: So that optimization problem, is that sticking around forever? Is that what you want to build the core of cognition around, nailing that? I think it's sticking around. Yeah. I mean, I think it's if anything, it's gonna be a bigger thing as time goes on because for a lot of these use cases, the intelligence is is is really not the bottleneck anymore. Right? And so so obviously, it's been been been been another hot day for the last few months, but but I would I would say it's like, you know, ChatGPT, for example, new, bigger, smarter comes out, you know, smarter model comes out. It doesn't necessarily change their retention metrics all of a sudden because most of the things people ask ChatGPT, you know, it turns out the models are good. You know, they they already go and get them right. Right? And like now what you care about is is is is not just pure IQ for for any task, you know, whether it's coding or or legal or or customer service or whatever, you know, it's it's it's it's not just, okay, what is the raw logical intelligence of the model that I'm working with? It's like, okay, well, does it does it know all of the details of what I'm dealing with? Does it do things in the style of how I want it to do? Is it fast? Is it cheap? Is it effective? Is trained specifically on my subset of use cases or, you know, all of these things like that? And basically what that means is I think that there should be this frontier where people start to much more aggressively use all these different models rather than relying on the single biggest one, right? Like I think a year ago, that was more of a phenomenon of a year, year and a half ago because were a lot of use cases that were just on the cusp of possible, you know. And when that was the case, of course, you wanna go use, you know, the very smartest model that you have up there, Sonnet 3.7 or whatever it was, you know, a year and a half ago. But now because all the models are really good, you know, what that means is is you can be a lot more judicious.

Speaker 1: Kind of a question going around, I think, on different people's minds. Coding agents are now incredibly capable, but is the quality of software around the world actually increasing? And so I wanted to get your view on, you know, you guys are working with a lot of the biggest companies in the world. Do you feel like their the quality of their software is increasing or are they just doing more? Are they doing back end migrations

Speaker 2: that consumers don't even experience? Mhmm. Because you've had this, I I I guess, question for a while, which is like, okay. When I when I download an app for, like, at a big airline Yeah. I was chirping at Rune about this. I was like, oh, if the AI models are so good, why is it annoying to use the United Airlines app? And he was like, have you tried to use it recently? It's actually pretty good. And then I did try it, and I was like, yeah. Actually, maybe it is better. But I don't know. Where do you sit on this? Yeah. I think short answer is definitely yes. It is better. Mhmm. I think to the extent that that there's more to do, a lot of that is is much more like

Speaker 4: it's much more a function of these practical problems, like going out and getting distribution, going and doing that. And Yeah. I mean, I think I think people, you know, peep people in our ecosystem understand this intuitively, but it's worth kind of calling out out loud that, 50,000 person software org is not going to figure out how to use coding agents overnight the same way that like a three person YC company is going to, right? And so there are steps in the process that need to go happen. There's a lot of onboarding and education that has to be done. There's figuring out the right systems. There's, of course, getting through folks like the security guardrails and reviews and the walls that people have to make sure all of that is tight and something that they want to have operating in their ecosystem. But, you know, once that's there, we very much see that that's clearly the case across all these industries. I mean, I think, you know, it's like Yeah. So between like Yeah. Go ahead. So

Speaker 2: I can imagine if selling Cognition, there's a few different pitches I could give. One is I go to a company and say, look, you don't have an ERP or you don't have an e commerce system. We're going to come in and build that and it's going to be done by the time we're complete you're and going to have this. It's going to live forever, but it's a new capability. Others, we're going to give every person in your organization a copilot, and we will help transform the organization. And then third might just be like, look, we're just giving you access to dev and it does good stuff. Like, you know, go and you couldn't get it at this scale, but now you can because we're working together. What's resonating the most these days?

Speaker 4: Yeah. No. I mean, I think folks are, rightly, by the way, are are very focused on at at this point on on just, like, what are the actual use cases that it that it's gonna drive, and what is it gonna mean? You know? I I think there was a there was a period where it's very much, alright. We're in the world. Yeah. How my how many tokens are my engineers using? How that was you know, that lasted for all of, like, four months or so. You know, great times. But but but but but now, you know, I I think there's much more kind of, like there's much more clarity of thought, would say, from from folks in the industry about, okay. Well, AI is great. Let's talk about what use cases that we actually care about. Like, what are the top three, four priorities we care about as a company? Like you said, maybe it's this killer new feature that we want to put out. It's this app that we want to make way better or it's whatever. And let's talk about how we actually go and measure that and improve on that. Know? I think the biggest thing, if anything, I think is kind of like, yeah, like how can you show it in the concrete results, right? Like the value should be there, obviously, because AI is so good, and it's so smart. But like, unless you're tracking that, you're making sure you're using it for the things that are effective versus aren't effective, you know, you're you're looking at the the productivity on a case by case basis, obviously, it's like you you'll never actually know which things you're doing are working versus not. Right? And and I think that's been a big big theme for folks. How do you predict that the router market is gonna evolve? We've seen a lot of moves recently. Ramp has a router. Stripe, you know, bought OpenRouter. There's a bunch of other players. Everyone wants to be in that token stream.

Speaker 1: But how do you see this sort of, like, category evolving?

Speaker 4: Yeah. No. Look, I mean, I think it kind of makes sense. I mean, it's it's, you know, a lot of these companies that you're for example, are they think of themselves as like the center for finance on the Internet. Right? And, you know, what are people gonna be spending money on in the Internet in five or ten years? I think a lot of it is gonna be tokens and models and agents and whatever you call it. So I think from that perspective, I think it's super reasonable. And I think the routing part is a big piece. But to your point, I think the entire kind of like infrastructure around how you do payments, how you do spend management, all these things, I think, are still going to exist and basically need to be redone in the world of agents. And so I I I frankly think that there are a lot of products to go and build in that space.

Speaker 2: What has your reaction been to the recent progress in math? When you Yeah. I think the very first time you came on the show, you predicted correctly that the IMO gold medal would fall. Google and OpenAI both scored just barely gold, not 46. I think it was 38 or something. It was like they missed the sixth question, but they but they did achieve gold.

Speaker 4: 42, by the way. 42 is a gold in in the in the Math Olympiad, which is different. Slight different. Yes. You have to remember, it's it's different numbers for every competition. Yes. Yeah. So so I'll tell you what my my honest thoughts on it Please. Which are look. First of I think it's it's it's actually insane. I I I think there's there's a lot of controversy and there's Yeah. Yeah. Discussions, arguments, whatever. But I feel like the most important thing to call is that guys, we just solved Navier Stokes with AI. Like that is insane. It's absurd. And and I don't think it's gonna stop, obviously. I mean, I think the like, you know, it's funny. I actually have money and this is my this will be my only AI bear view that I ever expressed to you guys. I actually had money on a bet that the rebound hypothesis will not be solved by the end of twenty twenty six. Okay. We'll see as, you know, three and a half more months. But I'm pretty sure if not '26, it will be solved in '27. I think it might get done in '26. But I think it's less than 50%. We'll see. But no, I mean, I think all of this stuff is just going to get done. And I mean, think it's like it's of insane. It's hard to put into words. I think if you're not familiar with the kind of scale this, like, you know, Navier Stokes, for example, is like a very fundamental problem about fluid dynamics and such, it's been around forever, basically, and lots of people have sucked lots and lots of time into it. And the fact that this could even just be done in like a week. Eighty eight hours. Just going and working around in the system and orchestrate. It's it's just it's it's it's it's insane. Yeah. Yeah. I I think on on the point of the controversy, I mean, it's kinda funny, but but but similarly, I think a lot of people who aren't as familiar with the the math academia world might not know that. This is actually just always what happens in in in math academia. Like, literally, Newton versus Leibniz on the founding of Caligas Oh, yeah. Was the is is still, by the way, is one of the biggest arguments that people have. You know, this is back in, like, the sixteen hundreds or so. And so in practice, I think, like, you know, I'm I'm sure look. I I believe that both sides meant well. I I think the accomplish itself is gonna be the biggest thing that we all remember from this. We could debate, you know, what what exactly shapes up and and what what it means for folks. But but but but it's it's, you know, the the rest, I I would say it's kind of the the AI the AI accomplishment itself is the biggest thing by far. The rest kinda feels par for the course to be a 100% honest. Yeah. So

Speaker 2: it it seems exciting for the the world of math. It's certainly exciting as sort of like a researcher recruitment to this is where progress is happening. It's also sort of a useful benchmark. Like, I could imagine in the future, like, two weeks in the future, but just a few weeks in the future, if you're training, you know, MuSpark 1.5, you wanna just throw Navier Stokes at it and say, don't search the Internet. Don't look at the result. Can you solve it? Because that's a good benchmark. Right? But but I wanna know about your philosophy because you went viral with the Devon launch. Like, what matters to your customers, your recruitment? Because there's a world where you're like, oh, I I I wanna put my throw my hat in the ring and duke it out for different math challenges, but that doesn't seem critical path to the growth of your business. Certainly hasn't been because you're growing very quickly. But how do you how do you think about the value of when you're talking to your actual customers? Yeah. What is important to get across?

Speaker 4: Yeah. For sure. And it's kind of by the way, it's it's kind of a hilarious thought also that, you know, we're just gonna be like, I I actually still think that one of the things that's that that I've never gotten over is, like, one of the token benchmarks that people do when they're going and training models and doing runs and stuff is AIME. Yeah. AIME. And I I think people don't necessarily know what it is, but basically, the AIME was it's a high school math competition. It's one of the the hardest math competitions, and it basically selects, like, the top few 100 kids in The US Yeah. To go qualify for the next level to actually compete for the US team and everything. Like, I took those, you know, every year as as a kid. And it's kinda funny because it's like, yeah, dude, your model can't even get a 15 on the Amy. It's it's out of 15. You know? It's like, what what are you even doing? But, like, no human can do that. You know? And we're already and then someday, it'll just be like, wow. You can't you can't even solve the Riemann hypothesis. Like Yeah. What does your model even do? You must totally mess up your retraining run or something. I get it. You're trying to save money, so you're using a model to cancel Riemann.

Speaker 2: But look, not everyone has token budgets right now, so you gotta pinch pennies. It's cool. It's probably good in Photoshop.

Speaker 4: But yeah. No. So for for us, what I would say is, obviously, like, a lot of what matters is showing the real bench for don't get me wrong. It'd be sick if Devon went and solved the Riemann hypothesis. I don't currently expect that that's what's going to happen because it's much less, you know, this this kind of like fundamental, you know, basic science research is much less our focus as opposed to to kind of going and doing real world use cases. And so we see, you know, it's like when we show benchmarks, it's like, okay. Here's a benchmark on how it does at finding security vulnerabilities in real world code bases. Here's a benchmark on how it does yeah. Right. And and a lot of it is basically just, like, making sure we're speaking to the thing that Sure. That folks care about and folks need. Yeah. I mean, related to that last question,

Speaker 2: how big is the cybersecurity side of the business? How much demand is that driving? Obviously, that's been a huge story this year.

Speaker 4: Yeah. Yeah. No. I mean, it's been it's been a massive thing for us. Obviously, everyone, I think, is is really thinking about this and thinking about I I mean, everyone's kind of freaking out, I guess, is is like the honest way to put it, which is probably correct. I I think cybersecurity is going to, I think there's going to be real threats that happen. I mean, when we say it's like these big orgs take time to adapt and to use new technologies and so on. Obviously, packers out there, like small teams that are going and using the best of the models or doing their own, who knows what they're doing, they're not waiting. And a lot of these capabilities keep getting way better. And so, no, it's a small but meaningful part of our business. It's probably in the neighborhood of you know, around 10% of of the Devon sessions and or the Devon ACUs that get spent today are on security, but but we see it growing pretty quickly. I mean, our security product's only, like, two months old. So

Speaker 2: I love it. Well, congratulations on the fundraise. I gotta ring the gong. Amazing update. There

Speaker 1: we go. Great to see you, Scott. Please go please go take out some open problems just for fun. Don't don't distract the team. Like, don't don't rope them into it, but just spin up those the next prime. Yeah. Yeah. Exactly. Spin up spin up a swarm. Just you and the swarm.

Speaker 2: Great to see you. Great to see you. Cool. Cool. Thanks for having me. Alright. Cheers. Well, we were talking about security. No better time to tell you about CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. We will be joined by Greg Brockman, the cofounder and president of OpenAI in just a minute. In the meantime, The Descartes the Descartes acquisition, there were talks that Anthropic was buying Descartes, and it seems like they have walked away. This is an exclusive in Bloomberg. But Say it. They put Descartes before the horse. That's the funniest quote tweet by Shashant Raman here on the timeline.

Speaker 1: We love Stories come out that that maybe one of the is totally, you know, I I don't I don't know how true this is, but it sounds like one of the sticking points was maybe Relocation. Descartes happily building in Tel Aviv, not wanting to move over to The US. Who knows? Dean There's a lot things. Jaccard is

Speaker 2: incredibly talented. It was a very fun demo. He came on the show and used his AI image model, his video model. Yeah. His willingness to do just a live demo of technology that seemingly was better than anything else Yeah. That we have seen. I mean, it was it was it was low res, but it it showed you a glimpse into the future. He would, you know, be prompting it while he was on the call with us talking about what's behind him. Now I'm in a wizard castle. Now I'm in a sci fi, you know, cyberpunk city, and all of that was,