Cognition's Scott Wu launches agent-native IDE for Devon and says AI will beat IMO gold this year

cheese ioi 2014 problem 6 he'll know what I mean Scott welcome to the show again great to have you back we will get to the amazing cognition news but I got to know did you cheese ioi 2014 problem number six you know I I can't answer that question actually but uh but know it's great to be back and it's great to see you guys again how are you guys doing we're doing great and uh I mean congratulations on the launch can you break it down for us what what is different about Devon and cognition today yeah yeah absolutely and so what we rolled out today is um essentially an agent native IDE um and what that means is you know I think there are a lot of great products out there that are about really accelerating you in a synchronous workflow um what we've always built deadin for is to be truly asynchronous and to be able to delegate and to have your own team of devans that's taking on tasks you know and to to be able to manage several devans at once or to be doing other things while you're working with devans at the same time and um what we're rolling out essentially is is a way to really cleanly interactively work with your devons um as you're using it and so it's um it's it's a full IDE experience it's a collaborative planning experience at the start of each um at the start of each session and then you know some of these other features like Devon Devon Serge and Deon Wiki that we've been rolling out um what else do you have I want to question no no go go into it well well I I want to just get into your overall thesis for AI and why uh and how the Market's moving broadly uh almost 10 years ago to the day Sam Alman posted a blog post called Bubble talk where he said everyone was saying we're in a bubble here let's format it as a bet and he said uh the top six us companies in terms of unicorns are Uber paler Airbnb Dropbox Pinterest and SpaceX they're currently worth just over a 100 billion uh he thinks that uh in 2020 those companies would be worth over 200 billion there's a few of those companies that are worth over 200 billion alone he also said stripes zenefits instacart a bunch of other companies will be worth more than 27 billion and he was I think he was off but only by a few weeks and it was this interesting thing where everyone felt like it was a bubble he formulated into a bet he was correct but how are you how are you feeling about the overall uh like murmurs about oh is AI a bubble yeah yeah and I think he was forecasting like a 3X or something in just a few years and so whether he was like barely off or or not quite you know or or barely made it I mean I think it's like you know certainly if you just hear that list of companies you know Uber and Air BNB and SpaceX and stripe and all you know it's it's I I think the point was very clearly there that those companies did have real Merit and real value and they proved out a lot more of that actually I mean it's almost comical to it's it's it's been 10 years now since then to just think wow like these are these are some of the the great public companies right um and I think that uh you know I I we were talking about this but I was thinking about it a lot because it's been almost exactly 10 years um and it's also I mean super cool Sam was running YC at the time which is why he put that out obviously now he's running opening eye but but yeah I mean I think people talk about bubbles so much in Ai and and I it's it's almost I I I feel I I feel almost the exact opposite where the the value that we are suddenly proving out in AI is almost underrepresented in in what we're seeing and and and what is what what is going on you know and and you see this in all these different spaces and you know I think as we were saying earlier like code is one of the first ones because it's it's it's so nicely automatable there's a clean feedback loop and and all of these things that you can really train mod to get really good at code but I think we're really starting to see it basically everywhere um and it's a it's a crazy thing to think about but a lot of these businesses that we just mentioned you know these are really really amazing businesses and really powerful businesses right and and a lot of what they do arguably is you know they make particular flows of your life 1% easier you know something like that right they come in in this like particularly painful process and solve that hard part for you and and deliver this and obviously these are incredible incredible businesses this is less about the businesses and more just about just how big the world is out there you know and how much value that there is to provide right and now when we're talking about Ai and we're thinking about what AI can do you know um for consumers it's you know an AI assistant that does all of your work for you you know it does all your hobbies and takes on all these things right and and in Enterprise obviously you know we're talking about multiplying every single person's output by like 5x and so I I I think that the it's it's it's I I was thinking about it especially because you know it really does feel like across a lot of these different layers you know the foundation model layer the application layer and so on it's sure it's it's a tough question to decide which which one is going to succeed which one's not going to be you know and only a few companies get to be the Google and Amazon and then a lot of companies obviously um you know have a tougher time but if you just ask the question overall is is AI as a space going to deliver massive massive value and is it going to build these multi-trillion our companies I mean I think at this point it's pretty clear that it will yeah can you talk a little bit about uh the studio gibli moment happened last week um and I'm wondering like with the levels iio flight simulator Vibe coding on coin term took uh a term everyone kind dancing uh are you expecting a studio gibli style moment for programming because I think what was so magical about Studio gibl was that it took the it took the onus of of creativity out of the process and it made it personal because you just take any photo from your camera roll any selfie you could take your own dog or your own profile picture type one word and you get something that feels cool and different is there a world where we get something where it's like hey make a video game about my life or something and and it just does it uh what would a what would a mass adoption moment for Aid driven programming look like for sure yeah so I think there's two kind of different categories I would say within within AI coding and one is you know I'll call it kind of the more um general purpose like let's say single used software type or or more hobbyist type work right where you you know you want to build a cool website you want to do this you want to make a nice game or you want to put a lot of these elements together and I think that's going to I mean it's honestly it's already taking off um pretty pretty massively it's you know you don't have to know how to code anymore and you can put together like a really cool website and you know set things exactly as you want them and so on right um and I think there will I I'm sure there will be more exciting moments to come where you know as we cross these new tiers of capabilities you know the interface is going to change for that and and the set of things that you can do um on the other side I would say is kind of the um the the the the massive kind of just software engineering industry which is a little bit more obviously is a little bit more oriented towards um you know a lot of big companies that are spending maybe millions of hours of of you know software engineering time developing like really really great software right um and making it super resistant to to all the different things you can think of you know figuring out all the little details of architecture and so on and and I think on that side you know I think for the time being that's going to be um primarily focused on software Engineers um just because you know it's while we're still working with these layers of abstraction you still really need to have all of these kind of Concepts in your mind and the ability to to really just like reason with the computer architecture um but there it's kind of like you know we'll see like a 5x there so maybe maybe one way to describe it is like I think there will be a um a level of kind of programming and being able to interact with the world through code that everyone will have access to which is super super exciting on the one hand and then on the other hand I think for a lot of the deep technical work that we do I I also think that you know Engineers who are focused on that and working on that are going to be able to do like 5x more so we're kind of seeing both of those sides at the same time how do you uh what advice do you give Engineers that are just joining the cognition team you guys have such a ridiculous sort of you know uh sort of uh probably I would I would assume the average engineer at at cognition is a 10x engineer in any other organization and so you're looking for like you know there's uh there has to be this pressure of joining your your building the tools that build the tools you know uh how how do you think about sort of ramping people to the sort of um to the output that you want to see out of Engineers at every level of the organization yeah yeah I was going to say I mean there's very little advice that I can give because these folks are all way smarter than I am but no I mean I think you know it's it's interesting I mean one one thing that that comes to mind you know people ask me all the time like should I even be studying CS like should I should I even be you know what how how do I even think about software engineering and and I think the the main thing I would call out here is just um you know as the capabilities change the the the interface of AI product changes a lot as well right and so it's one thing you know it's for example like you know there there was a phase where where text completion was was the the main thing and that was that was where the model capabilities were at right and you could get a lot of value from just pure text completion right uh and I think we're already past that and you know pushing into a lot of new things but each New Era that we get into in terms of capabilities is going to unlock a new product experience in AI code and I think that the main thing to call out because I think a lot of this is going to happen so quickly is just it's really really important to as an engineer to stay on top of all these tools and to to understand how best to work with them I think at this point it's you know there's no shortage of things to do in code you know I I don't think we're going to run out of things you know software that we want to build anytime soon and so I think it's more a question of how do you stay maximally effective um a a at at building software and a lot of that is just you know really being up to date with everything that's happening and understanding each of the new waves and learning how to use these new technologies because it does take time to learn how do you think about AI adoption among Engineers Broad has every you know there has to be uh you know I I I think about the example of like the Japanese soldiers that like didn't know the war ended and so they were just like on an island and like you know still like in like you know War fighting mode and it would been like 20 years or something like that and so I think there has to be Engineers out there that like you know maybe they don't know about cognition or they don't know about open AI or any of these other tools and uh but it has to be some like I would imagine like 90% of people that are just obsessively writing you know know code every single day 90% plus are sort of aware of these tools but H how do you think about adoption broadly yeah yeah for sure no and I mean I think at this point there's a um yeah I mean a pretty significant fraction of all the of all the top engineers in the world are using these AI tools every day um and they're getting a lot of value out of them the main thing I would just say here is this is where you know and folks sometimes like to to get a little kind of wishy washy about you know AI is going to come in and then suddenly all of our problems are going to be solved right and and and I think one of the things that we're finding you know even in code is yeah there's just a lot of practical work to do out there in the world right and so I think I think it's one thing to um you know to be able to solve these theoretical problems in a vacuum it's another thing to you know all of these different uh all of these different code bases out there all of these different development environments all these details you know there's a lot of cobal out there in the world still you know there's a lot of all these various different languages or all these crazy idiosyncratic setups um and so I think one of the big things honestly that that uh that we're solving as a space is is actually not just the capability itself but but just solving for for a lot of these practical constraints and and getting the technology to the point where it works for everyone and not just you know the subset of people who are building the latest and greatest apps and Python and typescript or whatever it is you know so uh can you talk a little bit about the importance of multi multimodality and image diffusion and understanding images in the context of coding I think a lot of people assume that well it's just code you don't even need to think about the advantages of being able to process images effectively but at the same time it feels feels like the more data you throw at these models the better they get even if it feels completely tangential just having you know the ability to not just know python but also speak Russian and Chinese and any any and Japanese and all these different languages it seems to increase overall intelligence even though it's not the narrow set that you thought thought it was yeah and by the way it's actually a fun exercise to if you've never done this before to basically you know a lot of the the queries that we feed language models of just like raw massive Blobs of text to try to do those yourself like as a human and one of the things you find is like wow it's actually really hard if you're just looking at some massive chunk of text of just finding the thing that you want and getting the right thing like humans naturally you know there's just like a lot of systems and how we process information visually and all these things um and and what I would say yeah I mean I think in terms of code for example the um I think it's it's one thing for it to be call it like theoretically possible you know it's from the set of information I'm giving you it is theoretically possible to extract out exactly what needs to be done and you know you you could say this for example with um with let's say you're trying to build a new Fe feature and you just say all right here's the entirety of my you know 1 million line code base just tell me what I need to do to build the feature and obviously no human can do that on the first try and just tell you exactly what needs to happen and really and for most of these more complex tasks no AI can either you know um and and in practice you know the the way that we solve these things is obviously through a much more iterative process that involves using all of our different tools and you know it might involve for example running the front end locally and then clicking around on the website and seeing what happens and you know you're trying to fix this button so you go around and click on it you know and you move around and you see what errors come up in the logs when you do that and so that's how you you know solve your problem right um and so a lot of it I would describe it as almost like um setting the the stepping stone instead of the The Rock Cliff face where you can make the incremental progress towards a solution and you know this is what we do all the time in software engineering and similarly like I think you know being able to use image or being able to interact with these other data sources is is what really gives AI that same capability uh please one question that you are uniquely suited to uh answer do you think we see an AI win an IMO gold medal this year it was silver last year uh and I imagine we made some progress since then but what's your take yeah I'd be pretty surprised if we don't we have a so so the the the greatest competitive programmer of all time is a guy named tourist is his name is Gennady he's from bellarus and he works at cognition actually um you can look this up true it's true it is objectively true it's aje and and uh and so Gennady you know we we we have kind of a bet going internally on mainly the question not of whether it will be like a you know like a you know top level high schooler which is essentially what the IMO or II Gold Medal is but whether it will beat Gennady who's you know the greatest of all time and I think most folks think it's probably going to happen this year yeah and and if not this year then then next year yeah I think Gennady is uh he he he's he's pretty adamant that he'll still be able to do it uh and so I I'm kind of looking forward I mean I I feel like we should have the we should do the whole show match you know Gary kaspar of style and and do it let's do it live let's do it live on t uh well can you give us the call to action it sounds like cognition and Devon are much more accessible how can people get started how can they sign up how they can play how can they play with this tool and and see if it works for them yeah so we rolled out a whole new experience and then on top of that we you know we've would really love to hear what folks think and so we're making it a lot easier to try out um and so we have a new plan that starts at a $20 minimum um and it's all available at app.

de.

so you can you can sign up immediately it's fully generally available and you can try it out on your codebase cool well uh hope those gpus don't melt hope the servers stay online and good luck with uh with with the next couple months is I'm sure uh it's going to be a continued rocket sh yeah congratulations to the whole team appreciate you st we'll talk to you soon thanks a lot Scott talk soon bye uh and like yeah he's actually on the team yeah well I I I've met Gennady because he he worked at ramp and I interviewed him and uh a lot of what he said went over my head if I'm being honest but uh he's a beast and crazy so so on poly Market there's it's only at a 59% chance that AI wins an IMO doll for 59 folks not a financial advice but you heard it here like it seems pretty obvious anyway we got Michael from aelius coming in the studio uh welcome to the show Michael