Q3 Review with Scott Wu (Cognition) and Hemant Taneja (General Catalyst)

I tend to think that the best poker style is tight aggressive. It's scary, you know, if your actual, you know, money is on the line, right?

AI is just one of those things where you don't want to be five years late.

Yeah. Well, first of all, thank you so much for having me and thank you for the kind words. Yeah, I know, I think for us, honestly, the way that we've always thought about things is, you know, of course you can do the super intense micromanaging and so on, but I think honestly trying to do that, even in the cases where it is effective, that's like more of a 10-20% improvement, whereas, you know, the 80 or 90% of it comes from actually just figuring out how you bring on the best people and give them the license that they need to go make the right decisions and feel ready to take risks and so on. And so, you know, what that comes down to, I think for us, is basically two things, right? One is which is recruiting itself. And similarly with recruiting, like, obviously people love to talk about like deal tactics and negotiations, but again it mostly comes down to just how do you make it a place where people really want to be? And then the second thing, of course, is just like making sure that the organization and the culture incentivizes people to — I mean, for example, like, you know, if you blow out people for every time any little thing goes wrong, then you set incentives in a way that people are not, you know, interested to go and take risks, right? Like, and I think in reality inaction sometimes in a lot of cases is the biggest risk of all and you just kind of like have to go and really show that. And so the things that we try to do are we obviously always try to push the pace and just move as fast as we can. And then, you know, on top of that we try to make it an environment where it's like, you know, I mean, you spend a lot of time here at the office and it's just like, yeah, like people are having a lot of fun, we're going and solving hard problems, you know, it's one of the reasons, I mean, we hire a lot of former founders. In fact, of our 500 people at Cognition, over 100 have founded a company before, which is, you know, one of my favorite stats. And I think part of the reason — well, the reason it makes sense for us is because the kinds of people who are founders, you know, work up and down the stack, are ready to own problems, just want to go and like, you know, build really cool stuff, obviously work well for us. But I think, you know, I think it's really important to make it the kind of culture where somebody feels like they can be an owner and can do a really good job. And so that's how we've always thought about it. My like poker analogy for this, I would say in terms of kind of like leadership and management is, you know, I tend to think that the best poker style is tight aggressive. And I think the same is true in management, which is basically like you should have a relatively small set of things where you're super, super, you know, micromanaging, super, super deep in the weeds on, you know, giving a ton of feedback on. But in those set of things, then you should actually be really forward about it. And I think as a leader, you know, your job is basically — maybe one way to kind of say it is like of the whole tree of everything that's going on, you know, you could try to get a service level understanding of everything, you could just like pick, you know, some random subset and just get really deep on those. I think neither of those is the best for a few reasons, but your job, I think as a leader, is to basically find the, you know, four or five root nodes of the tree that really matter and just be super, super on the ball and super deep on those specific things, which is tough, but it basically just means you're jumping around from crisis to crisis every day.

Yeah. Yeah, so my parents — my parents were both from Shanghai in China. They came to the US for grad school. They studied chemical engineering.

And then because they worked in a lot of the like oil and gas, you know, air emissions for oil and gas facilities, then Louisiana obviously has a lot of oil and gas, and so then we kind of ended up there. No, it was interesting. I think to a point, I mean, my parents took a huge risk in coming to the US, but I think they were always very risk averse actually with me. And I remember all these times where, yeah, they just want me to have like a super, super stable, super simple life and so on, which I never really — I never really wanted personally. And it was kind of like an interesting dichotomy to have there. And I think, no, for me, I mean, sometimes it feels like you're playing a board game, you know, it feels like you're playing like, you know, Race for the Galaxy or something like that. And I actually think that often when you're playing a board game, you actually make more thoughtful decisions about risk somehow because you're less burdened. Yeah, it feels less — it's the same thing. It's like, you know, I think most people feel it's the first time playing poker for money versus not playing poker for no money. You know, it's just like it's scary, you know, if your actual, you know, money is on the line, right? But what you realize is that at some point it's just like, yeah, you just have to make calls and you have to make decisions and there's different outcomes, but you'll navigate that too, right? And like you have to live with variance.

Yeah. One of the nice things about my parents is that they actually always disagreed basically about every major, you know, decision. So it's like they would give their thoughts and, you know, oh, should I stay in school? Should I not stay in school? Should I go to this college or that college? Should I go — you know, they would always have a different opinion. And so the reason that was nice is because it just meant that I could do whatever I wanted because it was — you know, whatever I chose, it was only — yeah, there was no consensus anyway. And so then it was — so that helped, I think, with me being able to just be more adventurous and take risks.

Look, I think safety is a super important issue and I think it's an interesting one because I think a lot of folks still don't really understand, you know, the depths of what exactly it means. And so maybe I'll try to kind of explain it a little bit, which is I think — I think hopefully we would all agree that if you had, you know, super genius level capabilities that could solve any problem or figure out or hack into any system or come up with any, you know, biological construct or whatever that you asked it to, I think we would all agree that we would need to be pretty careful about how that's rolled out, what things, you know, who's allowed to use it, what tools or what functions it has access to itself, all of these things. I think the question, of course, is that AI continues to get smarter and smarter and how do you trade off all of the benefits of what we're getting with AI versus, you know, the potential risks that come in? And, you know, I think for us one of the nice things is that so much of what Cognition is about, I'd say, is the diffusion of AI as a technology into the real world. And so, you know, we obviously do a lot of work with these models and so on, but we also work with all the — you know, we work with Anthropic, we work with OpenAI, we work with all these folks in terms of basically distributing this technology and figuring out how to build product form factor and how to deliver it and deploy it and all that in a way that it's actually useful. You know, it's one thing to say that this model can solve really hard math problems. It's another thing to say that it's actually out there going and accelerating GDP growth or helping organizations do what they're trying to do faster or, you know, ship more or build more, right? And, you know, one of the reasons that gives us, I think, a pretty unique view on safety is because in that way we're actually very aligned with our customers in what we want to go do, right? And the organizations that we work with, you know, whether it's a Citibank or a Mercedes or, you know, an Elevance or whatever, they similarly also don't want their agents to be going out of control or to have way more, you know, power and capability to go do things than they should, right? And I think the thing that people don't touch on enough is that it really is actually just in many ways a practical organization and change management problem and societal problem as much as it is a research problem. Of course, I think you can go and do the work to, you know, to teach the models to love humans and to, you know, to optimize for the behaviors that we want them to optimize for, but there's also just as much work in terms of thinking about, okay, well, how do you have the right guardrails for your agent and how do you set up a sandbox where it's not able to, you know, leak your code to the web or how do you set things up in a way where it doesn't have access to systems that would give it, you know, private or confidential information? And that's the kind of the side of safety that we most operate on and spend a lot of time thinking about.

Yeah, yeah, for sure. And so I think the current equilibrium is going to continue, you know, where there will be frontier models that are, call it, three to six months ahead, and for the kinds of use cases where you need the absolute best intelligence to go solve a problem, then that's worth a lot, obviously. And then on the other hand, there's a pretty commoditized layer where there's lots of models that hit a level of smart enough, call it, for a lot of the common everyday tasks that folks are trying to go do with AI. Look, I think this equilibrium is — no, I think it's overall healthy. You know, obviously you want it to be not a full, you know, absolute monopoly, but you also want it to be the kind of thing that incentivizes folks to be able to do the research and go and do more of it. You know, one of the big — I would say implications in the space actually is just on a kind of use case by use case or vertical by vertical basis what it means for each of these fields. And so security, I think, is one that a lot of folks are talking about and...

Yeah, no, it's been great. And I mean, I think across the board, you know, one way to say it is the attackers are already there and they're already going and using all these models, you know, whether open weight or not, you know, to go and figure out vulnerabilities, to do, you know, to find exploits or things like that. And I think similarly there it's less a raw capabilities problem. I think we all agree that the logical intelligence of the models is good enough for us to be able to secure our systems. It's actually more, you know, a real world change problem, right? And how do you get all the organizations of the world, you know, how do you get all of the — how do you get the US government and all the big banks and all the big healthcare, you know, companies and so on to clean up all of their systems and be using the AI defense force with agents to the same extent that the attackers are all using it for offense? I think that kind of interplay is, if anything, I think the bigger thing that we're seeing, yeah.

Yeah, yeah, for sure. So every single organization is going to have to figure this out. You know, maybe to just start with a few obvious points like AI is massive. It's going to get even bigger. Every single organization out there should be using AI. You know, all of these enterprises, sovereigns, etc. Like they should be using AI. And obviously it's not the kind of thing that they can just figure out for themselves overnight. And when you put those together, to your point, that, you know, I think there has to be strong neutral players that can act as real partners to these enterprises, these governments and so on, and be there to go and help them, right? And I think in that way, you know, we've seen this already, which is I think there's a lot of value actually in us not having our focus be on, you know, just training frontier models and doing pure research and so on. You know, what our focus is on is solving customer problems and figuring out how people can build more and ship faster and so on, right? And what that means is I think two things. One, to your point, every single enterprise can work with us and actually own, you know, and have a lot more control of what they're doing. You know, they can work with all the different models from all the different labs. They can have much better observability and monitoring on how they're spending and what it's going towards and what is an effective use of spend versus what's not and so on. And they have all that. But two is just plain and simple, which is we're actually — you know, we can be there to do all of the change management work with them rather than just giving them the raw technology that goes and does that, right? And so as an organization, I mean, if you think about a, you know, 40,000 person software engineering organization, having the tools that will go and write code for you or spin up software engines, super important obviously. But, you know, even once you have that, it's not like the 40,000 person org becomes five times faster overnight, right? A lot of what that looks like is figuring out first of all what initiatives do you want to be handing off and what are the playbooks and workflows that allow you to go and hand that off to agents. And then obviously figure out what you want to build next with all this additional bandwidth and then, you know, teaching all of your team how to use the tools to be much more effective. A lot of that work is just this messy change management that happens in the middle. And that's what we spend a lot of time on with our customers.