• 1 hour ago
Backed by $200 million in funding, 28-year-old Scott Wu and his team of competitive coders at Cognition are building an AI tool that can program entirely on its own, like an “army of junior engineers.”

Read the full story on Forbes: https://www.forbes.com/sites/rashishrivastava/2024/12/02/cognition-scott-wu-devin-ai/

0:00 Introduction
2:08 New Developments For Cognition
8:38 What Is The Potential Of Cognition's AI tool?
19:03 What Does AI Look Like In 2025, 2026?

Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1

Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:

https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript

Stay Connected
Forbes newsletters: https://newsletters.editorial.forbes.com
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com

Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.

Category

🤖
Tech
Transcript
00:00Hi, my name is Katherine Schwab.
00:05I'm an assistant managing editor here at Forbes, and I'm here today with Scott Wu.
00:10He is the CEO and co-founder of a very hot AI startup called Cognition.
00:15It's valued at $2 billion, and they make an autonomous AI coding tool called Devon.
00:23So Scott, welcome.
00:25Thanks for being here with us.
00:27Tell us a bit about Devon and how this tool works and how you think it's going to change
00:33the way that computer programmers do their jobs.
00:36Absolutely.
00:37So Devon is the first fully autonomous AI software engineer.
00:42And what that means is that Devon has the ability not just to write code, but to do
00:45all of the steps of the software engineering workflow.
00:48And so every software engineer, when they're breaking down tasks, when they're fixing bugs,
00:52when they're building new features, they're making back-to-back decisions basically of
00:57all the things that they're doing.
01:00If you are looking at a bug, you're going to be investigating the report, you're going
01:04to try to reproduce it yourself, you're going to try and check the logs.
01:07Devon does all of these things and does the whole workflow.
01:09And the whole idea of working with Devon is that it's almost like working with a junior
01:14engineer co-worker.
01:15And so Devon's not perfect, it doesn't know everything.
01:19You can tag Devon on tasks and say, hey, can you help me upgrade our codebase Angular version
01:24number?
01:25Hey, can you help me do this migration?
01:27Hey, can you help me fix this small little front-end issue?
01:30Hey, can you investigate this crash and see what's going on?
01:33And Devon is going to be there to take first pass and put up something for you to review.
01:37Okay, great.
01:38So I want to start with kind of the big question I feel like with most AI companies is, is
01:46this AI coming for people's jobs, right?
01:49And you're saying here that this basically acts exactly like a junior engineer.
01:53I mean, what do you say to kind of engineers using the tool who are afraid that this is
01:58actually going to take their jobs one day?
02:02Yeah, it's a great question.
02:04It's a question that we've received a lot.
02:06I'm sure you have.
02:08What I would say to that is, I think that, you know, I'm an engineer myself.
02:14I'm a programmer by trade.
02:15And in fact, our whole team is really, you know, we're all engineers by background.
02:19And I think one of the things with software engineering, especially, is that it's just
02:23so messy, you know, and pretty much every software engineer in the world out there will
02:28relate to this, where basically 10% of your job is this really exciting part.
02:33It's what people love about engineering, which is, you know, taking a problem and breaking
02:37it down and understanding here's exactly what's going on.
02:40And here's the solution that I'm going to put together for it.
02:43Here's all the details.
02:44Here's the cases.
02:45Here's the architecture I'm going to use.
02:47Here are the flows that I'm going to set up for my website or something like that.
02:50Here's the, you know, here's the backend database design that I'm going to do, you know.
02:54And that's the fun part.
02:55You know, that's the creative problem solving and really just understanding what's going
02:59on with the product, with the users and so on.
03:01And that's 10%, right?
03:03The other 90%, I would say, is kind of just a lot of debugging, dealing with your DevOps,
03:09fixing issues that come up, adding to your test coverage, fixing up the documentation
03:13and things like that.
03:14And, you know, the way that we all have always seen it, honestly, is that there's just so
03:21much more code to write, really.
03:24And software is, you know, one of the great things about software is, you know, software
03:27has been arguably the biggest driver of change in the last 40, 50 years in the world, right?
03:32But at the same time, we've always, always only been constrained by supply for software
03:36and never been constrained by demand.
03:38So our hope is we want to do the 90% for you and let you do the 10%, just 10 times more.
03:44And we want to make every single software engineer 10 times more productive.
03:47And we think there really is 10 times more code to write, you know, and that's what we're
03:50excited about.
03:51Okay, so tell me about your coding background.
03:55You were, and maybe still are, a competitive coder, right?
03:59So can you explain what that means and your background and kind of personal relationship
04:05to coding?
04:07Yeah, yeah, I'm probably not particularly competitive up there anymore.
04:12But yeah, so I grew up in Baton Rouge, Louisiana.
04:17I always loved math and science and I started programming, I think when I was nine years
04:22old, actually, in fourth grade, and I just loved programming.
04:27And I, as you can imagine, you know, there weren't that many folks in Baton Rouge, Louisiana
04:34who had a lot of the same interests.
04:35And so one of the things that I took to was doing these math and programming competitions
04:39where I got to meet other people from across the city, across the state, across the country,
04:43across the world, eventually, who had all the same interests, basically.
04:48And so I loved those competitions as a kid, spent a lot of time on them, like my whole
04:52dream was, you know, to go there and then do that all.
04:56And actually, a lot of the folks that started Cognition together are, you know, many of
05:02us have known each other for 10 plus years because, you know, we were on the U.S. international
05:07team together or because we went to the training camps together or things like that.
05:11And so, yes, it's been very much a lifelong community for me.
05:16So what does competitive coding even mean?
05:18Like, can you give me a tangible example of what a coding competition looks like?
05:23Because I mean, you know, we know what normal sports competitions look like, or even like,
05:27you know, gaming competitions are more, you know, common.
05:31So what is a coding competition?
05:32Sure.
05:33Yeah.
05:34Sure, sure.
05:35So the biggest one is what's called the IOI, the International Olympiad of Informatics.
05:41And so it is truly in every way is the equivalent of the Olympics, but for code, you know, and
05:47for informatics.
05:48And so every country has their own selection process, they bring in the top four students
05:54from their country.
05:55And obviously, you know, there's the state level and the local level and so on.
05:58And the competition itself is an algorithmic problem solving competition.
06:02And so it's basically the full life cycle of problem solving with code.
06:06You know, you're given a few tasks and you're given some set amount of time to take them
06:09on.
06:10And your goal is to think through the tasks, figure out the algorithm or the solution you
06:14have, that's going to be the most efficient as possible for that, and then implement and
06:19code it all as well.
06:20Test the code, make sure it works, you know, make sure it runs in time, make sure it's
06:24under the memory constraints and so on.
06:26And so, yeah.
06:28I mean, it's hard to imagine someone who would be better suited to creating something that
06:33could do all of that from start to finish.
06:37So you are a previous 30 Under 30 alum from a few years ago for your previous company.
06:47Why did you kind of step away from that company and start Cognition and kind of walk me through
06:52kind of the founding, the founding story of Cognition?
06:55Yeah, yeah, yeah, absolutely.
06:57So before this, I ran Lunch Club, which was an AI company for professional networking.
07:01I ran that for about five years.
07:03And yeah, with Cognition, you know, I think, I think there were kind of a few things that
07:09came into play at the same time.
07:11And like I mentioned, it was, it was, you know, a lot of it was, you know, a tight group
07:17of friends who we've all known each other for 10 plus years, getting together and talking
07:22about what we wanted to build next.
07:24And it was from the beginning, it was always about code.
07:28And it was always about specifically building autonomous agents for code as, you know, like
07:34systems that can reason and iterate and make decisions in the real world and see the outcomes
07:39of those decisions.
07:40And, you know, it's really teaching an AI to code is for a bunch of programming nerds
07:46like us is probably the coolest single thing that we could imagine working on.
07:51And we were all really, really excited to do that together and to build that together.
07:54And it initially was, you know, very much a kind of like, like a research initiative,
08:01you know, of how do we figure out this multi-step decision making and planning.
08:05And it's obviously, you know, quite different, you know, to have an actor in the real world
08:11that's making decisions versus, for example, a chatbot that's just responding to questions
08:15that you ask.
08:16Right.
08:17And so, you know, over time, as we started building it more and more, we eventually saw
08:21that it could really actually be useful even for ourselves or for software engineers around
08:25the world.
08:26And that's what really started with Devon.
08:27And so, yeah, it's always been code and it's always been, you know, autonomous multi-step
08:32decision making.
08:33But I think over the course of the last year or so, it's changed quite a bit.
08:38Was there kind of a ha moment for you where you kind of realized the potential of what
08:44this tool could do?
08:46Yeah.
08:47Yeah.
08:48The one of the earliest fun moments for us, I still have a recording of this actually
08:52is when we when we started building what became Devon internally and, you know, at the time,
09:01again, it was very kind of theoretical almost is how I would describe it.
09:04Like a lot of it was seeing if we could get the agent to navigate and make decisions.
09:09And, you know, we were building out a lot of our own infrastructure.
09:14And part of that, what that involves was setting up our own database and doing and,
09:17you know, if you've ever done this yourself, you know that it's it's it never really gets
09:23easier.
09:24You know, it's always like running into the next error, figuring out what's going on,
09:28debugging that and so on.
09:29And, you know, a lot of this DevOps stuff.
09:31And so we just that was the first one where we said, hey, you know, Devon, just take a
09:35shot at this.
09:36Let's see.
09:37Let's see.
09:38And and, you know, we didn't really believe it at first, but but Devon actually did it.
09:40You know, Devon like went through it, figured out what was going on, what files were missing,
09:45what ports were open or all these like little things with your setup that that make that
09:50make the setup work or not work and then got it done and to end.
09:54And that was one of the kind of first really, really fun moments.
09:56And obviously, there was still a lot to do to actually build that into a real product
10:00from there.
10:01But yeah, that was that was probably the first one.
10:04So when did you actually launch Devon into the world and what was the reaction?
10:09Yeah, yeah.
10:11We launched Devon in March of this year.
10:15And yeah, I mean, it was it was it was really exciting for us to get the chance to put Devon
10:20out there and to show people what was possible.
10:22And I think that for a long time, the view, especially with researchers in the space in
10:28AI, was that, you know, these agents, these autonomous decision makers were, you know,
10:34were definitely the future.
10:35They would definitely come and they would be the kind of the dominant use case.
10:40But technology itself was still years and years away.
10:43And we had always thought from the beginning that there was actually so much more to do
10:46right now and that so much was possible right now.
10:49And so it was it was yeah, we were really lucky to get the chance to to kind of show
10:55what was possible with Devon.
10:57And since then, you know, we started working with enterprise customers basically the day
11:03after we launched, you know, we're bringing on customers and so on.
11:07And now we're getting to share a lot of the stories of what they're seeing with Devon
11:12internally.
11:13OK, so I know there was also a little bit of pushback when you first launched and some
11:19critique around kind of overselling what Devon can do in some of your marketing language.
11:25And I'm curious how you respond to that and how you've addressed those concerns in terms
11:30of, you know, hype, what's what's hype and what's kind of overstatement versus what the
11:36tool can actually do.
11:37Yeah, yeah, absolutely.
11:38Yeah.
11:39You know, I think there's a few there's there's a few components to it.
11:42And I think one of them, as we've kind of said, there really is a lot of fear.
11:45I think people with AI have always, you know, kind of asked this question of like, well,
11:52is this coming from me?
11:53You know, what's what's what happens next?
11:55I think one of the big things that has helped is honestly, for us to just get the product
12:00out there more and more, you know, and to get people to try it out themselves and to
12:04see it.
12:05And, you know, for one, obviously, it lets them see the capabilities for themselves.
12:09But also, I think, you know, we use Devon all the time when we're building Devon.
12:12Devon is the number one contributor to Devon codebase.
12:16And we love it, you know, and I think like getting to actually work with Devon and see
12:22it out there and doing work with you, it makes you really see just just how much of
12:26the human side that there actually really still is, if that makes sense.
12:29And so we've seen this all the time where, you know, in our own internal communications,
12:35for example, we have a channel where we keep track of front end feature requests, right?
12:40We have a channel where we keep track of, you know, system crashes that come up, we
12:44have a channel where we keep track of, you know, little bugs or nitpicks that get reported,
12:48right?
12:49And so in these channels, our dominant workflow now is to just tag Devon.
12:54And we just say, hey, at Devon, can you can you take a first pass, you know, so we were
12:58reporting the issue either anyway, right?
12:59And we have Devon do it.
13:00And, you know, like you said, it's, it's obviously it's not perfect, and there's still a long
13:03way to go.
13:05But Devon taking the first pass and, you know, is able to do it quite a bit of the time and
13:09the rest of the time, obviously, you work with Devon, or you give the one or two lines
13:13of feedback, or you touch up the pull request yourself, and you get it in from there.
13:17And I think a lot of it is just showing people how much can be done with human and machine
13:22combined.
13:25What is your longer term vision?
13:28If right now, this is like you said, kind of human and machine combined working together?
13:35Where do you see this going five years from now, 10 years from now?
13:39Sure, sure.
13:41Yeah, I think the, you know, we've, we've seen from customers that have done their own
13:47internal measurements for the right use cases, right, and it's for specific use cases that
13:52make a lot of sense for Devon and that are a great fit, they see something like eight
13:55to 12x productivity, great gains in their engineers.
13:59And what that basically means is that every hour of an engineer's time using Devon is
14:04worth about eight to 12 hours of an engineer without Devon.
14:08And that's, that's obviously really, really exciting.
14:10And yeah, I mean, I think in the long run, to be honest, I'm actually really excited
14:15for a world where there's even more programmers, you know, it's, it's, I've always thought,
14:19you know, software or programming or whatever we want to call it, you know, it's, it's evolved
14:23a lot over time, right?
14:24I mean, at one point, it was even punch cards, you know, and then it was assembly, and then
14:28it was these different formats.
14:29And I think the, to get to a point where we're really only constrained by our ideas is, is
14:37the thing that's really, really exciting, I think, for all of us at the Cognition team,
14:40you know, the, the idea that, because everyone has so many things that they would love to
14:45build, they have so many products that they would love to use, you know, they have ideas
14:49for things they would make on their own and the things that, you know, that people love
14:54to be creative, right?
14:55And I think that a world where people are just really unconstrained by a lot of the
15:00routine execution and are only limited by their own ideas, I think that's really exciting.
15:06I want to pivot and talk a little bit about your business strategy.
15:10So how much of this, how much of your strategy is focused on, you know, rank and file programmers
15:17coming, you know, downloading the tool and using it kind of out of the box?
15:21And how much are you focused on selling to enterprises?
15:24Like you mentioned, you already have a few customers.
15:27Yeah, yeah.
15:28So typically, what we've seen, actually, is that, you know, most, the most tech forward
15:34enterprises or the most tech forward teams are typically the ones that are, you know,
15:38able to really adopt and move with Devon the quickest.
15:41And so it's, you know, in that sense, like a lot of it is, is honestly almost like a
15:47reframe of mindset, right?
15:49Of how can you delegate tasks, like, I think the best way to describe it is it's almost
15:55like making every software engineer into their own engineering manager, you know, and you
15:58get to manage your own army of Devons, your junior Devon engineers, you know, and you're
16:03working with your Devons to do tasks.
16:05And obviously, that's a pretty, pretty big change in workflow.
16:08And so what we've seen is that, you know, we work with larger enterprises, we also work
16:13with a lot of startups and smaller teams, for example.
16:15But across the board, what we typically see is that the tech forward teams and the ones
16:20that are really excited to kind of adopt the new workflows and take these things on, they're
16:25the ones that are able to get the most value out of Devon.
16:30Great.
16:31So you're not the only company trying to do this.
16:34There's a number of companies that are building coding tools, it's kind of the number one
16:40use case for generative AI.
16:42And I imagine maybe your top competitor is even just chat GPT, which is not a coding
16:46specific tool, but is the most used tool in this space.
16:50So I'm curious how you think about, you know, Devon being different than your competitors
16:58and how does a, you know, few year old, tiny, tiny startup go up against, you know, OpenAI
17:06or like GitHub, which obviously has Copilot, which is one of the most well used coding
17:12tools as well.
17:13How do you think about going up against these giants?
17:16And what is your game plan?
17:18Yeah, yeah, absolutely.
17:20And so what I would say to that is, you know, code is so enormous.
17:27And there's just so much out there, you know, there's about 30 million software engineers
17:30in the world.
17:31And again, the demand for software engineers is maybe infinitely more than that, right?
17:36And so there's just so much more to build with code.
17:38And what we've actually found is that there's a lot of great opportunity, I think, to work
17:43together and do a lot of positive sub collaboration.
17:45And so we work closely with OpenAI, you know, we worked with them, for example, on their
17:50launch of the O1 model just a few months ago.
17:52We work very closely with Microsoft, we're partnered with Microsoft and Microsoft obviously
17:56owns GitHub and VS Code and the whole stack there.
17:59You know, they know developer experience better than anyone, right?
18:02And I think what we've typically seen is, yeah, you know, a world of code abundance
18:08is a world where just much, much, much more is possible, right?
18:13And I think that there's actually a lot of win-win scenarios there.
18:16You know, I think as far as what makes Devon unique or what is really special about the
18:20use cases of Devon, a lot of it comes down to this, you know, just a fully autonomous,
18:26deeply integrated agent experience for code, right?
18:29And so, you know, chatGPT, you gave us a great example of, you know, chatGPT will answer
18:34your code questions better than anyone, right?
18:37But for example, like to be able to look deeply into your own code base, make a few edits,
18:42test it out, see if the new code is correct, maybe you file a report or things like that
18:47or respond to an issue or respond to feedback from your software engineering team.
18:50You know, all of that whole experience, you know, that is how software is built.
18:56And we've really, really built Devon specifically with that in mind, having it be a deeply integrated
19:02co-worker.
19:03Okay.
19:04Last question for you.
19:06There's been so much hype about AI in 2023, 2024.
19:11Looking ahead into 2025, what is next for Cognition and where, you know, where do you
19:16see this field going in general?
19:19What is overhyped and what is right on the money?
19:22Yeah.
19:23Yeah.
19:24Yeah.
19:25No, I mean, I think the space as a whole, obviously it's, we're in a very exciting time
19:29for AI.
19:30And like you said, I think there is quite a bit of hype and a lot of kind of buzz about
19:35that.
19:36I think what I would say to that, I think is like, AI is not the tool that's just going
19:46to like, you know, solve all of your problems for you.
19:49It's not going to be like a magic solution or things like that or anything like that.
19:53Right.
19:54What it will be is it'll be, you know, a tool that basically allows you to figure out for
19:58yourself what are the most important problems to solve and gives you the power to solve
20:02them and to do more with them.
20:04And so I think across the board, that's generally the trend that I've seen is that a lot of
20:11it is about, a lot of it is really just about kind of like optimizing you and giving you
20:17the ability to do more or to focus on what matters, you know, rather than just taking
20:21you out of the equation and just solving all of your problems for you.
20:25And so I'm personally really excited about, you know, impacting AI across all of these
20:29different spaces.
20:30I mean, I think sales, marketing, video, finance, accounting, legal, education, I think there
20:36will be, you know, big changes in AI and I think obviously it'll take some time.
20:41But yeah, it's a very exciting period for us.
20:44Great.
20:45Scott, thank you so much for your time.
20:47It's been wonderful.
20:48Super interesting stuff.
20:50And we'll see you next time.
20:52Thank you so much, Catherine.

Recommended