Cognition's Scott Wu: enterprise usage of Devon doubled in 6 weeks, Windsurf evolves from code editor to English-as-source-of-truth interface

Let me tell you about Cognition. They are the makers of Devon the AI software engineer. Crush your backlog with your personal AI engineering team. And we have Scott Woo from Cognition in the Restream waiting room. I want to talk about the launch today. I want to talk about AI progress. I want to talk about math and your predictions on the IMO gold medal and everything that's happening there. Um but let's start with uh the the the general update on cognition. What's the shape of the business today? And then I want to hear about the latest launch. Awesome. Yeah. What's up, guys? How's it going? Great to great to see you. It's been a little bit I feel like too long. Too long you guys have been cooking.

Yeah. U every every month feels like uh feels like a decade now in AI. So, um cool. No, so so things have been great. I mean, the business has grown a lot. You know, we shared some of our metrics today. One of which is that um our our total enterprise usage has actually more than doubled in the last 6 weeks even.

And a lot of that has just been been mass takeoff of agents. You know, I think the high level that we that we're really seeing is that as agents get more capable and you can trust them to do endto-end tasks, what you really need is the the full background cloud agent, right? And so that means, you know, being able to run your your repos and everything locally, being able to test uh being able to spin things up from, you know, Slack or Linear or GitHub or Jira or whatever it is, and just being able to have this mass parallel, uh, async workload.

Okay. Uh, and then and then the announcement today Yeah. Yeah. No, the announcement today was a was a fun one for us, it was a, you know, very near and dear to my heart, but but a lot of it, honestly, if if I were really to just describe it in one line is just

clearing through all the frictions that that we've known about and and just making it a really great experience. And so, um, you know, one of the big highlights is is automated testing and having Devon run your web app for you and send you the changes and send you screen caps of all of those things. Um, but but there's tons of little things that that that really affect the experience. And so, you know, making the the VM startup time way faster, uh, making the Slack integration way smoother, you know, showing you all of the intermediate progress of the messages and so on. And so, it's been a um I mean, it's it's changed our internal usage a lot and so that's why we're pretty excited to get this one out.

How so it seems like there's there's speed to be squeezed out from uh VM spin up time optimizations. We're also seeing some, you know, incredible progress on the custom silicon. We had the founder of Talos on generating 16,000 tokens a second. That seems like that will be really impactful when it rolls out to the broader uh code generation and software engineering world. It's still uh pretty early with that company Llama 3B at this or 3 uh 88 8B at this point. Um but where else are you seeing opportunities for speed? How do you think about the importance of speed uh for what you do? Yeah. No, there there's a ton that you can do and at some point a lot of it actually is just good old software engineering. Um, and so so you know it's it's it's it's of course like you know the the models obviously you know you can improve the tokens per second. You can you can improve the TTFT. I think those improvements will be great and we've already seen a lot of those over the last bit. We'll see many more. Uh but at some point you know your agent has to go install you know npm install. Your agent has to go UV install. Your agent has to go GP for things right? uh it has to go pull up the front end itself. A lot of that stuff is um it's good old product building and software engineering to to make that better and more efficient. And so a lot of these obviously you know you can do algorithmic tricks. You can put in you know indices right and indexes and and make those faster. Uh you can do little things uh to to kind of like cheat the loading time and and do things in parallel and do things async. But a lot of it is just building the systems around the agent to make it really fast.

No, it's a really good point. I mean, anyone who's installed Open Claw has experienced like, oh wait, I'm actually just waiting to download software because it's pulling a whole bunch of stuff together and it's not actually doing that much waiting with the LLM, at least in the setup phase. Um, but you still have to actually get this thing configured and I think a lot of people in tech went through that. How have you been processing lessons from open claw interaction patterns that you think are interesting? what it means that society more broadly is just aware of AI agents which I feel like is a term that you basically coined years ago uh and have have been running with but in a specific like enterprise context and now I'm at a bar and I'll hear somebody talking about AI agents and it's because of OpenClaw and and I feel like oh that's a I remember the Scott Woo launch video where he explained that this was going to happen. Um but but how have you been processing OpenClaw? What is interesting about that? are are there any like lessons from that open source community that project generally that that paradigm that you want to bring to Devon?

Yeah. No, I mean a lot of big changes and I think by the way I think OpenCloud gets a lot of credit for uh for for many people being the first time that people really saw Yeah.

uh what what a full you know agent would look like with access to your files, access to your computer and so on. I I think we're really getting to the point uh you know to to your previous point where I I think we're we're really starting to switch over from the early adopter cycle to the the kind of mass market cycle is my sense and and and the concrete impact of that is a lot more people are starting to hear about and really think about AI agents right and I think it used to be I mean for us for example a year and a half ago you used to go into the room and explain to people what an AI agent was and why this wasn't you know why why this was different from from just like normal autocomplete or chatgpt or something like that. Now everybody's thinking about this stuff. Everyone wants to use it. And I think one of the the kind of implications of that is just accessibility and getting people to value as soon as possible is is one of the most powerful things that you can have in your own products as a result.

You guys have had a ton of success in enterprise. The chat wants us to ask for your take on the SAS apocalypse. I imagine uh uh some of the conversations that you're having with let's say the CTO of a of a a massive uh uh company

uh are they thinking about using a Devon for things like you know big database migrations like how how are they thinking about um how agents can can impact their dependency on on sort of these like legacy tools and systems of record?

Yeah, I mean there's the whole catrini report and everything. I mean, it is honestly ridiculous. That that's that's my that's my two cents on it. I I think that like um look at at a high level, of course, yeah, AI is going to change a lot of stuff. I I I don't really understand how you go from that to saying that there's going to, you know, like take software as a good example. Software is one of the most deflationary things ever. you know, a lot of the same products that that you know, used to cost much more 10, 20 years ago got have gotten much much cheaper over time, right? Has this been terrible for software company? You know, I mean, it it seems like it's been pretty good. All the big companies in the world are still software companies, right? Um and and so I think there's like

um there there's one thing when prices go down because, you know, the demand's just not there anymore. And obviously you can get into weird cycles and all that can happen. But it's a totally different thing if prices go down because we've just gotten way better at supplying things and that's when you get Jeban's paradox and that's when you get, you know, just mass consumer surplus and so on. Um and so at a at a high level I know I mean I think there's all the customers that we work with, you know, banks and health insurers and private equity and and so on. I mean the they're um they're there obviously like a lot of these base migration modernization projects that they can go and take on immediately. that the very next thing that they say is then like, okay, how do I pull the rest of my road map forward, right? How do I build even more and get even more out to people. Um,

and and I think in reality, we just we all just have so much more software to build.

Yeah. How are you thinking about AI progress broadly? It feels like a lot of people are feeling that recursive development is on the on the horizon. People are bringing up takeoff speeds again, migrating from slow, maybe fast, I was backing off, now it's a little quicker. Um, how are you uh how are you trying to like zoom out, reset, get to reality, figure out how fast things are actually moving?

Yeah. No, I mean the the the meter report shows like the consistent doublings and everything. I I mean I think it's a very um I I think things are continuing on the exponential curve. I wouldn't say that they're going either super exponential or subexponential. I think they're they're roughly going on that exponential curve, but you know, exponential curve is a lot.

Yeah. like that's that's a very fast growth obviously. Um I I think for us um you know one one of the things that's been pretty interesting is just like noticing each of the step function changes that happen. And so for us for example it's definitely been in the last I'll call it like four or five months

where something interesting happened which is we stopped typing code you know like at some point you you just don't right like like before obviously you have all the tools and you have the combination of things and so on. Now, it's there's different experiences. There's different tools that you want to have. Obviously, between the IDE and the CLI and and and the web agent and so on, but either way, you're you're really just working in prompts and you're not really, you know, like the code that we check into GitHub, like how much of it was typed by a human at this point? I think almost none.

Yeah. Um and and maybe one of the things I would just call out is that that you know a as you kind of expose each new thing like I mean if you think of it as like a a profiler you know on on your own software engineering workflow like what is the most expensive part you you shrink that down you get to the next thing you shrink that down and you just make the whole cycle more effective. We're at the point where a lot of these other things like you know understanding the code base and review and so on are the actual bottlenecks right testing is another big one. Um, and I think what we're going to see over the next next little bit is um, you're basically going to have to solve each of those with really good product experiences, really good model capabilities and so on. So may maybe the only thing that I would say, you know, I think the exponential curve continues. I I would just kind of call out that the form factor looks very different as you continue on that exponential curve because you're actually solving different problems. Like yes, I think we will continue to to to kind of like, you know, get the doublings and the doublings, but but now it looks a lot more like how do we optimize testing and review and planning, not how do we make the AI good at writing code based on the prompt that you give because at this point it's actually frankly it's it's basically already done.

Yeah, I have a bunch more questions. Uh I'll be quick. I have two. Um first, uh what does the future of windfur look like in a world where you're not writing code? Does that become a Kindle? I mean that's a joke but does it become does it become more important for that product to be the best way to read code because even if you're not writing code you still like I've I've done terminal prompts and I'm like and then I wind up opening the files to kind of take a peek in them and I'm like ah I kind of like there there's maybe room for innovation there and like how your code reading skill improves as your code writing skill uh degrades but how do you think about the future of

windsurf Yeah, for sure. Um, I think the high level here is it's going to be a gradual thing, but I think over the next one or two years, we'll have a pretty broad transition towards what you might call having English as the source of truth.

Sure.

Um, and so so people talk about like basically I think we'll go from code to English in the same way that we went from assembly to code.

Yeah.

Right. and and so so you know one of those steps has been been you know has been mostly done at this point which is the step of figuring out how do you uh prompt in English and then have the agent produce the code but if you think about it I mean you're still you're still you know reviewing code you're still checking code into GitHub you're still reading the code to understand what's going on and I think at some point you actually want an interface that looks a lot more like a spec or a map or you know like a design doc right and that's the thing that you're iterating on that's the thing that you're reviewing for example, right? Like at some point review I mean people say, "Oh, like review is going to go away because AI is going to catch all the bugs." I think that's actually not right because what you're going to be reviewing is the decisions, right? It's like here's what we're here's what the product is doing in this case and here's what the product is doing in that case and you know here's how you know this plan works or or whatever it is, right? Um and so what you'll want to have is a very clean interface to interact basically, you know, with with your product and with your own specs and so on. And and that's what a lot of what we think winds surf evolves into over time, right? And so so again, I think it's a very gradual thing. I think I think there's a lot of value in reading code now and certainly at cognition we still do a lot of reading the code even if we are not the ones like you know writing the next line of code because instead we we write the English prompt for that. Um but but but I think what happens with Windsurf is you know at some point instead of looking at each of the files you start looking more at you know this the specs and the highle logical design of what you're building. You start looking at the you know the diagrams of your app or your your website itself and you're able to go and manipulate those and you're really just managing your agents that you kick off from there.

Okay. We blew past the IOI gold medal as you predicted correctly. Amazing. Uh I have a follow-up question about that, but it's sort of in three parts. One is uh what is the next like math or physics-based benchmark that you're excited about AI potentially unlocking? Uh when do you think that might happen? And then do you think there will be any tangible impacts of that? Because if I walk down the street and I tell some random person like they they did it, Navier Stokes is solved. I think that's the math problem that everyone talks about a lot. I don't even know. Um I I think most people would be like great like is that going to help me with my job like they're more excited about just knowledge retrieval right now. Um so so yeah uh the the the next hurdle timeline and then impact.

Yeah. Yeah. For sure. So I mean we actually I would say crossed a pretty exciting hurdle just recently. um like Alex Lupaska and some of the folks at OpenAI um had a pretty important breakthrough in physics where they used uh language models to figure out a lot of the the the key lemas and theorems for it. Um and so so you know I I would have said I think the next big breakthrough is is getting to a point where like actual science and actual discovery is happening largely powered by AI and I think we're we're effectively getting into that. I think we'll see much more of that this year. I think to your point on impact, yeah, I think it'll be some time until, you know, the the the average person feels the impact of us proving new theorems, but obviously the long term of all of this is extremely powerful, right? I mean, we're going to be discovering new medicines. We're going to be, you know, unlocking big breakthroughs in biology, material science, nutrition, um, and so on and so on. And all of this is, you know, comes from a lot of the same science. I I think I very much think of it as a

um

as as as you can call it like a you know a proof of concept or or or like an existence proof that it is possible. Um and and you know solving some of these very difficult novel math and physics and algorithms problems

is there are lots of ways that over time that itself will continue to be valuable but but even more so than that it's obviously just

you know an existence proof that that AI can do some pretty incredible things. I love it. Oh, yeah. A lot of people get abstract with the medicine science. I like the material science one because I can imagine

a much stronger, much cheaper, much lighter carbon fiber and driving a car that's pure carbon fiber for the same price as a Model 3. This is pretty attractive. That's pretty tangible. I think the average American consumer is going to get behind that.

Excited about that.

Get extreme. I'm pretty excited about the part of, you know, you you have the best pizza that you've ever tasted and except it's also the most nutritious thing for you because we've just solved taste and nutrition and everything. And I feel like AI will get us there, but that might require a few more. There'll probably there's probably a few more steps in the middle.

That's that's the new AGI benchmark.

Get the goalpost. Get the goalpost.

I'm moving the AGI will be here when I can have a pizza that tastes amazing and also is fully nutritious. Thank you, Scott. Woo. Have a great day.

Great to see you guys.

Always fun to move the goal post with you. We'll talk to you soon.

Goodbye.

Let me tell you about MongoDB.

What's the only thing faster than the AI market, your business on MongoDB?

Don't just build AI, own the data platform that powers it. And without