Ep 39: Github CEO Thomas Dohmke on Building Copilot, Scaling to 1.2M Users and the Future of Code

It was very special to have Github CEO Thomas Dhomke on the pod for many reasons, not the least of which is that my partner Erica Brescia was COO at Gitlab just before joining Redpoint! Thomas has been at Github for almost 6 years, and has been CEO for almost 3 of those years. He has a pulse on what engineers around the world are looking for from the world’s leading developer platform, and incredible vision and empathy in making Github delivers it.

• 0:00Speaker 00

  GitHub is the world's largest source code host and at the center of the developer ecosystem. And of course GitHub also built GitHub Copilot, the first AI coding assistant, which is now used by over 1.3 million developers and over 50,000 enterprise subscribers. Today on Unsupervised Learning, we had on Thomas Domke, the CEO of GitHub, to share his perspective on the future of AI and coding and what he's learned about scaling Copilot. Full disclosure, before joining Redpoint, I was COO of GitHub, working alongside Thomas and others to scale GitHub and bring Copilot to market. So this episode was especially fun and special for me. If you're an AI founder, operator, and interested in hearing about one of the fastest growing examples of an AI product in production, this episode is for you.

• 0:45Speaker 00

  Thomas, thank you so much for joining us here today. You've had an amazing journey from being a kid growing up in East Germany before the wall fell, to being a developer, to being a founder, who was then acquired by Microsoft. Then you worked on the Microsoft acquisition of GitHub. And of course, you know, rose through the ranks. Now you're CEO of GitHub, a pretty incredible place to be. What's the coolest part of being CEO of GitHub in the age of AI?

• 1:13Speaker 01

  I mean, Copilot, right? We built Copilot now over four years ago, which seems like a really long amount of time. It's really the coolest part of the job, seeing people using it, seeing people talk about how it's helping them, how they feel. It's magical, you know, seeing our own user research coming in. You know, I remember, you know, some of the early stats where we couldn't actually believe that these numbers are true and that Copilot is already writing so much code. And we sent the team back and say, hey, verify that those numbers are actually right. Seeing my kids coding with it and solving their own problems in natural language. I think that's the coolest part of being the GitHub CEO. And of course, it's that I get to work with developers every day. I've been coding

• 1:58Speaker 01

  since the early 90s. You know, I've gone through the rise of the internet and the rise of mobile. And now we are in the age of AI. And I totally believe we are going to get, you know, we're going to enable a billion developers on this planet.

• 2:14Speaker 00

  Yeah, it's funny when you talk about we couldn't believe the metrics, you know, obviously I was there when we shipped GitHub or Copilot as well. And I think it's the only time in my life where I've been at a company, we shipped something and it was even better than we expected out of the gate. Like usually you're shipping too early, it's not quite there. Copilot, we were like, holy crap, this is powerful. I mean, it's, uh, it's pretty remarkable.

• 2:39Speaker 01

  Yeah, and I think, you know, this has kept going ever since. Obviously, there were moments, you know, in the Copilot journey where things have changed around us that opened our eyes. And one of them clearly was ChatGPT in November 2022. Right after that launch, I went on a trip to Japan, Singapore, and Australia. And all my customer conversations ever since have been around AI. Before, there was a lot of skepticism. Ever since, it's like, help us to adopt AI, help us to make AI part of our own strategy.

• 3:12Speaker 00

  Yeah, I think it's the first time in any major platform shift where you had this like consumerization of a technology happening at the same time. And so all of a sudden, instead of having to explain to people like what cloud was, if we think about that transition, instead it's like, oh no, they're, you know, kids are using it and their parents are using ChatGPT and they're understanding and seeing the power. And then they're like, wait a sec, if I can do this via, you know, some kind of consumer application, I want to be able to do this in my company as well. You know, you talked about traveling post-release of ChatGPT and you've talked a lot in your keynotes about

• 3:53Speaker 00

  lowering the barrier to software development for folks around the world. I'd love to hear a little bit about that and what you've seen, especially, you know, in some of these developing companies, you're giving people so much power. I mean, I think GitHub was doing that even before Copilot. And now it's like a whole new world of developers has opened up, which is pretty exciting. Like, what have you seen so far? What anecdotes can you share with us on that front?

• 4:20Speaker 01

  I think, you know, the first thing that comes to mind is just the power of keeping people in their creative flow. The original Copilot, you know, had no chat and it had no agents. It was just auto-completion. You know, just helped you writing text within the editor. And, you know, auto-completion has existed for quite a while, long before anyone was talking about the age of AI. Yet, you know, having a large language model doing that auto-completion, being able to, you know, show you ten lines of code, render whole test case, or how to connect to the cloud is incredibly powerful because it avoids that the developer has to switch from the editor to the browser.

• 5:05Speaker 01

  Where we all have, you know, our whole lives reflected in all the tabs that are open. And then you go on an internet search and you try to find, you know, what you're looking for. And then you're copying and pasting that back into the editor. But by the time you're actually back in the creative process, you lost the flow. And so it really isn't, it wasn't so much about generative AI when we got started. It was about keeping engineers in the flow, keeping them where they have the most fun writing code. And I think chat added another layer on top of that. And natural language plays a huge role because we're thinking, you know, in the languages that we learned as little kids, or sometimes when we immigrated to another country in that language too.

• 5:47Speaker 01

  We're not thinking in code and we're not thinking in deterministic machine language, right? And so as such, you know, that's incredibly powerful that you can now ask questions about your code. Even have it just explain your code back to you to see what you did, right? Like that's the, sometimes the model is more honest to you than you're honest to yourself. And then you bring that from English, which is predominant, you know, in the United States and in England and whatnot. But it's not, it's not predominant in Brazil and in Argentina and in South Africa and in many other countries. And so, but those six, seven-year-olds also want to learn coding. They also want to, you know, explore their creativity after they have played Super Mario for the first time.

• 6:32Speaker 01

  And so giving them the power to say, well, you know, I want to build a game in Brazil. In Portuguese is incredibly democratizing. I think it will create a whole new way of thinking about software development. Like we think about learning to play an instrument or, you know, learning to draw and all these things. It doesn't mean you have to become a professional artist. It just means you can learn it if you want to and if your creativity, you know, is pushing you into that direction.

• 6:59Speaker 02

  I mean, I would love to go a little bit deeper into how you think those changes will impact the role of the software developer in the future. You know, you mentioned at GitHub constellation last week that natural language is the most powerful tool for developers and you're launching Copilot Workspace to go from idea to production with just natural language. And so what does the role of a software developer look like in, say, five years?

• 7:21Speaker 01

  I think it will look very similar to today. We should not overestimate how fast the change will go, given how much software, how many lines of code exists, right? And, you know, every company I've met with and every company I ever worked for writes more code every single day than the leading. And many of them are still managing code bases that are, you know, decades old. You know, most of the banks still run COBOL on mainframes that was developed by folks that are hopefully retired by now. And that was developed, you know, it's easy to forget. Code developed in the seventies and eighties didn't have agile software development. It didn't have, you know, test-driven development.

• 8:02Speaker 01

  It didn't have Scrum. Nobody was writing unit tests. It's certainly not in the sense that we're doing that today. So that those code bases still need to be maintained. And I think, you know, what AI will do, it will make that much easier to go through, you know, these complex systems that software developers manage today. It will make it more accessible. You know, the hardest challenge that you have when you join a company or a new project, a new team, is to actually learn how the team is doing their stuff, learn about the code base, learn about the stack, you know, learn about the institutional knowledge of the team. And I think that will be much easier if you can just ask Copilot, you know, in natural language, questions about this. How do I deploy here? How do I add a library?

• 8:45Speaker 01

  It will, Copilot will do the not so fun stuff and so you can focus on the things that you actually enjoy doing. You know, one example that often comes to mind and that we are living ourselves, you know, here at GitHub, is that our backlogs are full, right? Like we have so many ideas ourselves, we have so many customer requests, you know, enterprise adoption blockers, uh, when the press are telling us we're not buying your product until you deliver this and that feature. Often around very boring stuff like proxy support and then policies and whatnot. And on the other side is, you know, all the fundamentals, you know, availability, scale, security, compliance, accessibility, all the things that you also have to do and it's a rising pile, you know, of work.

• 9:27Speaker 01

  Whenever, especially whenever a security vulnerability, you know, goes through the internet. And so we really think that AI, you know, agents, auto fix, as we call it Copilot auto fix, will help developers to burn down the security vulnerabilities and accessibility issues and other things. So they can focus on the innovation, on this thing that really is, you know, making them happy.

• 9:54Speaker 02

  You mentioned that agents is going to be a really core part of Copilot's strategy going forward. Can you maybe talk a little bit more about some of the different agents that you have in mind for it?

• 10:03Speaker 01

  Yeah, by the way, Copilot itself is an agent, right? If you think about the word, there's a pilot in the cockpit and a copilot. And a lot of the time, the copilot is actually flying the plane. So if you take it literally, Copilot in that sense is autonomous to a certain degree. Yeah, I think, you know, in many ways, developers have, they're already used to using agents. If you, if you, you know, simplify the concept, CICD is an autonomous process that runs whenever I push my code into repository, you know, runs all my test cases, it creates a log file and then it, you know, alerts me that I have to do something. Conceptually, you know, it works like an agent, you give it a task and it gives you a result back.

• 10:44Speaker 01

  And I think, you know, most developers are looking forward to more of these things that, that solves their problems. In the same way that, you know, we have dishwashers at home and robotic vacuums and...

• 11:02Speaker 01

  Lawn mowers and whatnot. And, you know, nobody has ever complained about having a dishwasher. You don't fear your dishwasher, you're happy that it does the work for you. And I think, you know, agents will, you know, do the same thing. It, uh, security tooling today, uh, shows you all the security issues that you have. And then you have to go and burn them down. Uh, instead, you know, we will have, um, security tools that actually helps you to, to burn down these issues. And, and so you can go back to, to, to the next feature to implement.

• 11:31Speaker 02

  Speaking of non-coding things like dishwashing, do you think that most industries, say in the next five years, once these models are even better, will still be human in the loop? Or do you think that AI will replace a lot of, a lot of humans in a lot of industries coming forward?

• 11:45Speaker 01

  I think there's certainly some industries, you know, translators come to mind that are probably going to be replaced to a certain degree by AI, because we can just use these things to translate in real time. Uh, I'm, I'm, I'm seeing a world where, you know, I can do this podcast in, in German and the listener can hear it in English and it actually sounds like my voice. And, and you can pick whether you want a German accent or an American accent and, and things like that. Um, I think in software development, um, the, the, the role won't, uh, won't change much. Um, I think we will just up level, we've marked up the step, uh, the stack as, as we always have, right? Like we, if you think to software development in the nineties was very different before open source than after open source.

• 12:27Speaker 01

  You know, uh, before the cloud and after the cloud, we had VMS and now we have containers and, and clusters. And, and so we're, we're moving up the stack, um, um, and, uh, and the, and the AI will help us doing that.

• 12:40Speaker 00

  On a related note, what does that mean for, you know, new software developers who are trying to learn and come into it? I think of, you know, we both have kids around 12, we're really into coding. And today they're using, you know, maybe Copilot or more autocomplete versions of, um, AI. Over time, as AI gets more powerful and we have these more agentic ways of building things. I, I take your point that Copilot itself is an agent, but when I think about agents, I mean like, hey, go build me this, this thing or this piece of an app, you know, this building block. You still have to kind of understand what's going on in the code in order to build the blocks and make sure you have the right blocks and then click all the right blocks together to

• 13:24Speaker 00

  build the resulting app. What do you think the learning journey is going to look like? Because if we can get a lot of help, you know, with the basic stuff, then how do we know if the more complicated things we're building are, you know, are built the right way?

• 13:40Speaker 01

  The way I think about this is that, you know, AI is a layer on top of code, but code doesn't go away. The machine at the end of the day, you know, has a processor, CPU and a GPU in it that both run a deterministic instruction set. And code is just an abstraction layer on top of that. So we don't have to write, you know, machine language or assembly. And then we move from, you know, simple coding languages to higher coding languages. And we added, you know, thousands of open source libraries. And so that's kind of like the same thing, right? Like if you use an open source library, how often do we actually look into the code base of that open source library to understand how it works? If it does the job and everything is

• 14:21Speaker 01

  fine, you're probably just using it and updating it hopefully to stay on top of security vulnerabilities. But if it doesn't work, you know, you go and dive in and maybe you forget on GitHub and then try to debug what the problem is. And then all of a sudden now you have to become an expert in that library. The same is true, you know, for Copilot. Copilot adds the human language as a higher abstraction layer. And when you see kids using it, you can see them exploring, you know, ideas and they type, you know, build me a Python game and I want to have, you know, a sprite character and then it gives you instructions. And as long as that actually works and you can just copy and paste it over to the editor's side, they might just keep going

• 15:03Speaker 01

  with natural language. But as soon as things break, they have to go into the code base and try to figure it out themselves. So they need to have the understanding on both sides. But human language, you know, in Copilot makes it easier to, makes it more accessible to get started. And it actually, I think the biggest superpower is that it makes it easier for you to unblock yourself. You know, like when I learned coding in the early 90s, all I had was books and magazines and there wasn't an internet I could dial into. I could go to computer club on Wednesdays and kind of like hope that somebody is there that has the same knowledge level. Or better knowledge level to solve my problem. And other than that, the best solution you had is go to bed and sleep and hope that in the morning

• 15:47Speaker 01

  you have that aha moment of unstuck in yourself or write a letter, you know, quite literally, the computer magazines had letters in them, letters from the readers asking questions, right? And today, you know, with the internet, you can go and do a Google search and with AI, you can actually now go stop further and explore this problem yourself, pilot the code and say, where's the bug? What can I do to fix this? You can just say it doesn't work. And because, you know, the chat window kind of like has the context of what you said before, it may tell you, well, how about you try this or you try that? And I think our kids will learn how to do that in the same way that the, you know, the kids have learned the last 10 years of how to use TikTok and YouTube and phones

• 16:32Speaker 01

  and make photos of the world in a very different fashion than, you know, we have become, you know, Erica, we have become the generation that is now old and looks at the latest generation of, you know, with very different eyes than us as kids.

• 16:50Speaker 00

  Fair, though I do have much higher hopes for AI and coding than I do for TikTok, but that's a separate podcast, I think. You know, you talked a little bit about open source and this is something that I think about a lot. Like, what do you think the impact of AI-powered coding will be on the open source ecosystem over time? You know, now, it used to be the case that maybe you were using an open source library and you didn't really speak that language, so it made it hard for you to go and contribute fixes. I think, you know, Copilot and other tools really lower that barrier, obviously. I mean, there are so many potential implications, I think, of AI and lowering the barrier to coding and being able to automate more security fixing and things like that.

• 17:35Speaker 00

  But like, what happens? Do we, do we end up with more open source? Do we have developers using AI to contribute to existing projects? And those become more popular and maybe we don't have 75 different JavaScript libraries? Like, how do you think about the evolution of the open source ecosystem once, you know, AI becomes even more mainstream than it is today in coding?

• 17:59Speaker 01

  I hope and I think, you know, that it will make open source even more dominant than it already is. You know, one anecdote I can tell is that I talked, you know, with a dad, who's also VP of Engineering at a tech company the other day. And he said, you know, he was coding with his with his son and they had to convert, I don't remember, decimal into binary or something like that. And the son had written 150 lines of code. And then I went to Copilot and asked that question and it got basically two lines of code, the import statement from open source library and the method call, right? And so it increases the usage for open source by if you, and the same happens in ChatGPT and other coding tools. They they automatically, because they were trained on so much, you know, internet information,

• 18:44Speaker 01

  they automatically suggest the use of open source projects, because that's in the training set that what's what real developers would also suggest, you know, if you ask that question. And then you can go and say, well, do that, but don't use an open source library. You still get better code than than the code that you wrote yourself. And so I think it will increase the usage as naturally these AI systems will suggest the usage of open source to solve a problem. I think it will make the life of a maintainer easier. You already mentioned, you know, burning down security issues, just managing issues, you know, summarizing, you know, a discussion in the morning. Like there's the little things in the life that AI makes already much easier. And I think it will increase the number of contributors.

• 19:29Speaker 01

  You know, today, one of the biggest challenges we have is we have to use an open source library and you find an issue or you have a feature request, now you take this and you want to learn about this code base and you want to understand how it works and you want to do the right things, because you otherwise fear that when you open that pull request and you didn't write the right test cases or you didn't, you know, follow the right structure that the maintainers have, you know. Institutionalized in the project, you're going to get rejected and then nobody wants to get rejected, especially not if they're doing free work during their hobby time. So I think it's influencing the work of software developers on the consumption, on the maintainer and the contribution side. And as such, it will create an even bigger growth for open

• 20:11Speaker 01

  source. And just the other day, Meta published the latest Llama model and Mark Zuckerberg talked about how he sees open source and how important it will be. So I think we're going to see even more open source and more sharing through the use of AI.

• 20:30Speaker 00

  I completely agree with you there. I will say, I might be a VC now, but I am forever a COO in my heart. And so I've always been all about efficiency, right? And you just use the example of, hey, instead of writing this code, it just pulled in an open source library, which is what we want. And I think how we can together jointly like build better software over time. I think the question though there is, are AI models going to start favoring specific open source projects because they're higher quality or whatever it might be, they show up more in the training data. It's probably a good thing because then maybe we don't have people who didn't find that open source project. So go and start another

• 21:12Speaker 00

  one. Like there's a lot of creativity and amazing creation and innovation in open source, but there's a lot of duplication and there's a lot of waste. And so one of the questions I've been asking myself is, do we start kind of consolidating around specific key open source projects even more so today? So when we want to go and create something new, it's truly net new instead of duplicating things because these AI systems are going to keep choosing and reusing the same components instead of going out and building something new every time?

• 21:47Speaker 01

  Maybe, I think, you know, in some ways it does exactly what humans do. Um, you know, which when you pick a framework to build a new web application, you're likely erring on the side of which are the big projects and have the most maintainers and the most longevity, you know, React or Next as examples. Um, you know, you can still use jQuery if you want to, but, um, you know, if you're up to date on the state of software development, you probably stay away from that. Um, while, you know, an open source, um, and not an open source, while large language model might actually look at what you have in your editor, um, what you have in your project already and uses that as context, you know, to suggest, you know, keep using the frameworks you already have.

• 22:30Speaker 01

  If you, you know, think about the magic behind Copilot, even the original Copilot was that it uses everything that you already have in your file, you know, above the cursor, below the cursor, that's fill in the middle, you know, use JSON tabs, you can actually intentionally open tabs uh, to give it more context and it uses that, you know, in a computational way to create embeddings and a context, um, a prompt, if you will, that you don't see. Uh, and as such, you know, it does give you, you know, answers that are in relation to the project you're working on, otherwise, you know, you wouldn't get JavaScript code. Uh, if you're in a JavaScript file, you would just get, you know, code that the model thinks is the best one. So I think it will swim somewhere in the middle of, you know, leveraging the context that you have, you know,

• 23:13Speaker 01

  uh, that you have in your IDE, in your editor, um, but it will also, you know, obviously promote the biggest libraries as do, you know, tech sites and blogs and, and, and ultimately developers themselves. I think the, you know, the consolidation is a good question of how that might work and um, if, you know, we at some point get into a place where we have specialized models, um, custom models or smaller models, you know, that are part of a company's workflow, um, or retrieval, you know, serves the same purpose where a company, you know, like Shopify says, well, you know, we have a worker on Ruby on Rails and we have certain patterns and certain ways of doing things and we customize Copilot

• 23:57Speaker 01

  to the degree that our developers also always follow those standards and don't go away and, and start using Go or some other language that is just the hottest new thing.

• 24:08Speaker 02

  Going a little bit deeper there, you're releasing fine-tuning for Copilot soon. We'd love to hear your thoughts on that and what do you think the impact will be?

• 24:16Speaker 01

  It's one of the most requested features, especially by enterprise customers, as a lot of them are just sitting on old code bases and have a certain way of writing applications and using internal libraries that the model training will never know about, given that those internal libraries, I think C++ specifically, are not becoming open source and the ecosystem there is very different than more modern languages that have package managers. And so tuning will work, that you go to your GitHub organization and pick a couple of repos and basically say, "I want a fine-tuned model for that," and the tuning process will run

• 25:00Speaker 01

  and then your projects, when you check them out in Visual Studio Code or Visual Studio JetBrains, it will automatically use that fine-tuned model. And it's actually basically getting a tailored Copilot, but you don't actually have to go into the data science and you don't have to understand parameters. It's really just customize your Copilot from the user interface and you get better results.

• 25:23Speaker 02

  We've talked a little bit about the original Copilot being so good and surprised at the traction that it received. When you think about the last few years, what has been the hardest part about evolving Copilot and the development process of that?

• 25:36Speaker 01

  The being so good piece is actually interesting because you could also say the complete opposite, which is like it often hallucinates and it often outputs code that is not perfect. And I think the interesting piece about that is that it's true for every code written out there, like whether you use a Copilot or not, naturally when you try to find code snippets that help you to solve a problem in an algorithm or API access or doing round, making round corners on iPhone app, you find code in the internet that is in almost any case not perfect and you have to go and modify it. And so the model hallucinates is actually not a bad thing per se because developers are already used to... Take code and modify it for their needs.

• 26:21Speaker 01

  And as such, you know, the original Copilot was good because it already knew your variable names and already knew what you had written above. So it was closer to the perfect answer than, you know, most GitHub, Reddit, Stackover Flow answers, answers can ever be. And I think that's the, the, the, the middle ground there that, um, it's just, it's like we're raising the average. We're not trying to be perfect. We're raising the average and as the models get better and as the systems get better, those, those averages also get better. And I think as we look forward, we will see, you know, better models and we will definitely see better customization. We haven't really talked about extensions yet, but I think, you know, that's a, a super important part because almost no project, uh, is, is built on top of a programming

• 27:05Speaker 01

  language, um, alone, right? There's lots of other services that come into play, even more so if you go into, you know, commercial, professional software development where you usually have a stack, um, a platform engineering stack, uh, that is highly custom for each individual company. Likely will not find two companies that is use exactly the same self tools unless they just all came out of a combinator and then, and somebody recommended that to, to every founder in the class, right? And so, you know, some use LaunchDarkly and some use Dazik and some use, you know, Azure and some use Aries, right? And so you want to basically pull that information also to the Copilot, um, into Copilot chat and, and, and be able to say, okay, here in my team, in my project, that's how we're doing it. And so I want all my answers to be tailored for that.

• 27:50Speaker 01

  And if I ask, you know, how to create a feature flag, I get basically a, well, you know, uh, here's, here's your LaunchDarkly connection. And if you want to create that feature flag right from chat, you can do so. And you can also ask whether the feature flag is enabled. So I think that extensibility layer on top of the better models and better reg and better fine tuning, uh, is crucially important because even if you have the best model that knows it all, it's still will not know what you're actually doing in your stack. And so you need to have these connections, the systems that you either have internally or externally.

• 28:22Speaker 02

  And along the same lines, what has surprised you in the last four years about Copilot? Are there any use cases or ways that users are using it that you didn't expect originally?

• 28:31Speaker 01

  I don't think we expected originally that things would move so fast. The project was started in something that we call GitHub Next, which is a team that sits outside of engineering and product and works on research projects, but I'm putting research into quotes because it's not research in the sense of we're trying to write papers and publish and reinvent the wheel. It's more about thinking about the future of software development, Horizon 3, if you will. And I think the Horizon 3 with Copilot really was more like a Horizon 1.8. And as time went by, it quickly became the current Horizon. And nowadays, you can't really talk about Horizon 3 anymore because everything that

• 29:15Speaker 01

  you work in terms of AI has at best a six-month horizon. And after that, you have to rethink and reconfigure what you're working on anyways. And I think that pace and also the adoption, the diffusion, Copilot has crossed the chasm. It's no longer about the early adopters. It's about whether you're in the early majority or in the late majority in getting it and using it and then thinking about what's next on that journey. And I think that's what we're really surprised about and couldn't have predicted that four years ago.

• 29:53Speaker 00

  And it's really remarkable how quickly that has happened. I mean, this is certainly, from my personal experience, the quickest I have ever seen technology go from, hey, is this even going to work, to mainstream adoption, in the enterprise no less, as opposed to consumer, where sometimes we see things take off more quickly. You know, you talked about in your TED Talk about how there have been these like three major breakthroughs in three years, thanks to AI. What are you thinking about in the next three years? Like, is the pace sustainable? What are the next major milestones that we're going to see? And what's the timeframe for those?

• 30:32Speaker 01

  I want to say the pace is not only sustainable, but accelerates because we're using now AI to build the AI systems. And so we can build much more in the same one of times. And as we, you know, hopefully get more of these agents into the process where we can just burn through our security backlog and burn through our accessibility backlog. In fact, you know, if you think about accessibility today, when you want to ship an accessible webpage, you have to go through a code and make sure that all the images have an alt tag and, and you know, and things like that, that you have like contrast and like font sizes, you know, using fixed pixel fonts and whatnot. You could also see a world where, you know, the person that has vision impairment has a device that shows them the world exactly how a person without a vision impairment

• 31:17Speaker 01

  sees the world. And so you don't have to actually take care of accessibility anymore because the device, the Geordi La Forge visor, you know, from, from Star Trek, kind of helps them see the world. And, and the question really is how do we bridge the gap between that state where we can see the world, you know, with the help of AI and where we still need to do work as developers. Same for security, you know, hopefully we burn through the security backlog. But once you have done that, then the question is how do you stay clean? And how do you, you know, leverage agents now to not reintroduce security vulnerabilities? You know, how to not downgrade on, on a version of React? Because, you know, one side of the company didn't know that the other side has already moved, you know, to the latest release to, to security vulnerability.

• 31:59Speaker 01

  And so I think we're going to see an acceleration because we make our life easier by spending more time on, on the innovation side and less time on, you know, let's be honest, the side that most developers are not excited about when they, when they look at their security dashboard and, you know, the compliance items that the lawyers just gave them because, you know, the AI Act is rolling out in Europe and DMA and DSA and whatnot. But I think that's the pace of, the pace will hopefully accelerate because we needed to accelerate. We have lots of problems to solve. And you know predicting what that will be is is hard, right? Like you couldn't have predicted, you know, where we are today three years ago.

• 32:43Speaker 01

  But I think we will see innovation across the whole stack. It's actually super exciting, you know, that we have chip companies in Silicon Valley, chip startups in Silicon Valley again. Now they're not building their own chips, they're outsourcing that to TSMC, but you know, the designing GPUs that run transformer models faster. And in similar use cases, we will see innovation across models. We will hopefully see innovation in the responsibly AI stack, making that easier, you know, to to use these models in the most optimal and secure and ethical way. And we will certainly see a lot of innovation in the app and in higher layers that we can't even think of today.

• 33:23Speaker 02

  Are there parts of the, you know, development stack that you don't think are consumed by by AI in the long term? You know, right now, for example, we have companies focused on just testing and using AI code development for that, for application security and finding vulnerabilities. You know, are there certain parts of the stack that are not going to be as impacted as much?

• 33:44Speaker 01

  I can't think of of any. I mean, I think, you know, we're going to see, you know, shifts, shifts everywhere because we need those shifts. I think the question is, the better question is, you know, which part of, you know, an organization, especially in software development, doesn't need the efficiency gains from AI, doesn't need the productivity gains? I mean, you know, we all have too many Slack messages and too many emails and we're really, what we're really, you know, fighting ultimately, and that goes back to, you know, TikTok is that we have only limited amount of time every single day and we have certainly limited amount of attention and everything wants to grab that attention and whether it's in personal life and consuming videos and that has already shifted.

• 34:29Speaker 01

  I don't think my kids understand linear TV because we never had any, you know, ever since they were born, we had YouTube and Netflix and Disney Plus, right? And so it's kind of like when they, when they go to their grandparents and they watch TV and they're like, why do we cannot pause the programming now and have dinner, right? And I think the same is true, you know, in this new world of artificial intelligence. We are going to shift, you know, and I actually, by the way, think the term artificial intelligence is scary. From a wording perspective, if you think about the gains and the better, how it makes our life better and how it makes the world better, we're going to see a dramatic shift of how we think about that.

• 35:15Speaker 02

  There's obviously a lot of companies that have raised a lot of money to go after code generation, you know, building foundation models for either general code generation or specific things like testing. Do you think this is a winner takes all market? Do you think that there will be foundation models for, you know, application security, for testing, for code generation? How do you sort of view the market?

• 35:37Speaker 01

  I don't think there's anything in tech that is winner takes all. And even those that believe they were the winner that took it all eventually got disrupted by new technology. And if you look at Microsoft, you know, which is turning 50 years old next year, is that the reason that Microsoft is still relevant is that they reinvented, or we reinvented ourselves, given that we as GitHub are part of Microsoft. But at times, in fact, GitHub to a certain degree reinvented itself with Copilot. And last year at GitHub University, I said GitHub was re-founded on Copilot. And that doesn't mean that all our legacy is no longer relevant, but we leveraged the

• 36:18Speaker 01

  new AI technology to build something that certainly changed the world with Copilot. And I think, you know, there's going to be more such companies and competition is great. You wouldn't want to watch, you know, a sports game that only has one team competing against itself. You want 18 teams or whatever league, or 10 Formula One teams with two cars each or whatnot. And your rooting, of course, for your favorite team. And then, of course, you know, I want to win the competition. And then, you know, a year or decade is over and then you have to do it over again, right? Like, just because you won, you know, the world championship this year doesn't mean you win next year again. And I think that's really how to think about this.

• 37:02Speaker 01

  You know, most companies play the infinite game, right? Like, they're playing Minecraft, not Super Mario, you know, in the long run. But in the short term, of course, they play little mini-games within Minecraft. And I think that's how you need to think about business and that's how startups need to think about their journey. Like, sure, you know, there's maybe an exit, but an exit is just you're merging into the next game and you keep playing, playing, you know, with the merge company or, you know, in the public markets.

• 37:30Speaker 00

  Your own journey is a prime example of that. And I love the F1 reference, right? Like Red Bull needed Mercedes to get to the point that they were at and McLaren certainly has needed them.

• 37:42Speaker 01

  Well, let's not forget, you know, Mercedes needed Red Bull too, right? Like it's not too long ago that Lewis Hamilton won it all for many years in a row. And then came the fantastic 21 season and ever since it's been a bit the reverse.

• 37:55Speaker 00

  Yeah, and made it a lot more fun to watch in my humble opinion. Going back to the model question, totally agree with you on this isn't a winner take all market. We certainly need several successful companies for all the reasons that you talked about. When you think about within a particular enterprise, though, how do you think about this playing out over time? Do you think companies are going to use multiple models specifically for code? Let's let's just focus on code for this. And and are they going to use specific coding models? And then will they also have other models that might be helping with documentation and testing? Or is your view that, you know, these models are getting so powerful?

• 38:40Speaker 00

  You can already write tests with Copilot as far as I know. And you can already use it to help the docs and things like that. Like, how does this play out? And how many models does an enterprise end up using specifically for code related work?

• 38:52Speaker 01

  As long as they're using Copilot, I'm fine, by the way.

• 38:57Speaker 01

  And I don't care what model they're also using. I mean, you know, Copilot today is using many models, different ones from OpenAI. You know, 3.5 Turbo for auto completion and for Turbo for chat and for for workspace. And of course, you know, we're constantly upgrading these models. And behind the scenes, there's smaller, you know, dot versions of these models that we don't even talk about. But that make the product better. But there's also, you know, a smaller model within the response to AI pipeline that scans and fixes security vulnerabilities that you never get to see. But you're using it, you know, as you're as you're using the Copilot stack. And and obviously, as soon as you're also using other tools within your engineering stack.

• 39:42Speaker 01

  You know, I already mentioned a few, you know, like Slack and their AI tooling. Or, you know, Atlassian has an AI agent. You're using the models that are behind those tools. I think the As everybody is using multiple tools, it's kind of like you're asking the question is, are the companies going to use more than one SaaS service within their IT stack or within their engineering stack? And the answer absolutely is yes. And as such, you know, they're naturally going to have multiple models for multiple use cases. And within coaching alone, you know, there's so many things that, you know, that are outside of GitHub, you know, if you think about, you know, data collection with telemetry and

• 40:24Speaker 01

  insights with DataDoc or Google Analytics, or which requires probably a time series model to predict, you know, anomalies and or, you know, that you have to scale up and things like that. I think their vision, you know, is another topic that, sure, there's multimodal models, but then there's other use cases where you might want a specific vision model to render videos or render images when you're applying. So I think this is a multimodal world and to some degrees, you know, in the, from an investor perspective, we're overthinking that a little bit of that there's the one model that rules them all, like I don't see that actually being true.

• 41:05Speaker 01

  And certainly not if you, certainly not if you go into biotech and you think about models to prevent or treat cancer as an example.

• 41:14Speaker 00

  I could have asked that question differently, which is really, is there room for multiple products? It's not really about the models, it's about the products, right? And from an investor perspective, we see companies that are entirely focused on testing or entirely focused on cleaning up tech debt, right? And so one of the topics that we discuss internally and certainly have points of view on is like, how is this going to evolve over time as the co-pilots of the world themselves get more powerful? Like, is it really going to make sense for a company to buy and use Copilot, but then go and procure a different product, regardless of underlying models, for one specific part of the software development lifecycle?

• 41:57Speaker 01

  Yeah, I think they're already doing that. I think they're already doing that today where they're combining, you know, let's just take the security landscape where they're combining multiple security vendors and they have the choice between the one that has it all and maybe a bit more shallow than the customer likes, or they pick the best of breed, you know, as the applications and combine them together. You know, most companies also are so You know, diverse that they have many teams doing different things. Even if the bosses say you can only do that, use that one stack. Whether it's on the programming language side, whether it's on the IDE side. You know, certainly within the, you know, cloud stack, there's, there's so many combinations. And as companies evolve and, you know, evolve these days means every six months, you know,

• 42:42Speaker 01

  there's the hottest new thing and that developers will jump on to. So I think there's, there's certainly a world where, um, where, where multiple AI developer tools will coexist within, within the stack.

• 42:54Speaker 02

  We've seen sort of the shift to power going to the incumbents just because they yield distribution with AI. You know, you have a really unique perspective because you founded two companies and now you're the CEO of GitHub. And so if you sort of put your founder hat on, um, what would you advise founders today in terms of being able to go up against the incumbents with AI now?

• 43:14Speaker 01

  Yeah, I already mentioned, you know, the one that I think is most important was just to play the long game, to play the infinite game and think about the, the big picture. You know, at least at certain times during, during the week, um, and, and not focus too much on, on the small battles because, you know, I think the one thing I've learned as I, as I grow older and in my career and also in life is that, you know, while we have, you know, almost a extreme sense of urgency these days, you know, to a certain degree that because Amazon delivers everything within a couple of hours, especially when you, when you live on the West Coast, right? Uh, and so we want every feature also within a couple of hours and every, every like update to your house and every, every change, you know, who wants to wait for a car for nine months these days?

• 43:57Speaker 01

  If Tesla can get you the latest model, you know, in front of your door by the next day. And even that, you know, Tesla cannot always realize, but I think we're more long-term thinking over like the short-term battles is, I think, helpful for startup founders and to even, you know, realize why they're doing what they're doing. You know, along those lines, I think it's super important to continuously reconfigure as a company. There is, there is no stable state anymore. You're only surviving if you're constantly evolving, you know. We see slack to think about the pivot and talk about that. I think there's a certain negative connotation unless the pivot resulted in a slack, you know, that became massive. But I think in reality, it's like you need to constantly adapt towards happening, especially

• 44:42Speaker 01

  in technology, but also within your product as you find product market fit. You know, there's not really a step function now you have product market fit and for the rest of your life, you're good, right? You constantly have to find new product market fit and retain or grow your position in the market. So as such, you know, your team has to change and the organization has to change along your growth journey. And then, you know, I think my favorite thing, especially since I became the GitHub CEO, is like, like you have to say no all the time. No, no to all the new ideas, no to all the investment areas. As the core of the company, you know, the biggest danger you have is that now you have a thousand people or 3,000 people or, you know, at the scale of Microsoft, hundreds

• 45:24Speaker 01

  of thousands of people, you still have to say no to most of the ideas. And, you know, I think it was Apple or even Steve Jobs that said, you know, a thousand no's for every single yes. And I think, let's not forget that. That's the true power of a startup is focus, focus, focus. But you have to keep that focus even as you grow and even as you find success.

• 45:42Speaker 02

  Yep. So one of the main things we look for on the investment side and one of the hardest things that we see founders have to deal with is saying no. The GitHub ecosystem is incredibly diverse and a lot of founders are trying to tap into that ecosystem. And you launch the extensions program, which is super cool. You know, what are certain areas that you'd love to see more extensions being published today?

• 46:03Speaker 01

  We'd love to see them across the whole stack because if you have them across the whole stack, basically all the tools that companies use, you know, within their software development teams, the more powerful Copilot becomes. And so, you know, if your software development provider and you're building tools, we'd love to invite you into the partner program. We actually have seen our customers that are directly telling developer tools providers to build a Copilot extension because they want their stack. Mapped into Copilot, they want platform engineering to be, you know, in the natural language interface, so you no longer have to go to fifteen portals and figure out what account are you using and how do you have the right access rights and all that,

• 46:45Speaker 01

  you know, if you, if you think about that, all that mapped into Copilot into the IDE is incredibly powerful and, and so we, we hope, you know, that we see more extensions and, you know, I mentioned accessibility is an example, um, uh, security, uh, compliance things, there's a lot of white space, uh, in that, in that area where, um, developers can build, build cool stuff that makes developers' lives easier.

• 47:10Speaker 00

  Are those your top kind of priority areas where you'd love to see more developed? Is there, there any other things that you'd love to see founders thinking about adding as an extension?

• 47:20Speaker 01

  Um, I think, you know, the, you know, the, the true power of the next wave of Copilot is that mix of extensions and agents of thinking about, you know, code completion to a certain degree is to solve problems, you can make it better by customizing and fine tuning, um, and writing more code, um, but eventually there is like the curve of flatten because you can't really predict what the developer wants to do until they have given more input. Um, you won't be able to predict a whole file by just typing three characters at the beginning of a file, um, and if you can do that, then you probably want to use an open source library because you have just re-writing what already exists out there. Um, so I think, you know, um, as, as order completion, you know, to a certain degree

• 48:03Speaker 01

  is a soft problem, then the next thing is, you know, the, the rest of the development stack and, and, you know, scanning a file for vulnerabilities and then fixing those vulnerabilities, you know, reviewing code, uh, you know, from a coding standards perspective or from a refactoring perspective, multi-file edits, you know, predicting the next step after the current step. And ultimately, now, understanding a codebase, you know, imagine, you know, one of your startups would hire me today, you know, I have a lot of software development experience, I would still struggle, you know, initially with their codebase, understanding it, figuring out where to change which file. Submitting the first pull request as a new employee in a company is scary.

• 48:47Speaker 01

  And they can, you know, have, hopefully they have great culture and they give you a warm welcome. But, you know, I think the more developers help Copilot to understand the context of the project they're working in, the better, you know, AI will become. And really, you know, get us to that next step change.

• 49:06Speaker 00

  Yeah, that's something that I think a lot, and when I reflect on my time at GitHub as well, you know, we were always moving around teams and moving engineers to new products, projects depending on where we wanted to focus for the quarter. And, you know, there's always this ramp up period on something new. And in particular, when you have a codebase that's as mature as GitHub's, there's just, there's a lot of surface area to cover. And I think, you know, this opens up a whole new kind of set of opportunities to move folks around, to learn on, to work on different things, to ramp much more quickly. And to your point, there's like so much work to be done, just in the codebases that exist today.

• 49:50Speaker 00

  Copilot and other tools can really help us there to play a little bit of catch up, perhaps.

• 49:57Speaker 01

  And you know, one thing that, you know, is underestimated is that just filing an issue, whether it's as a product manager, a feature request, or as a, you know, open source user, you know, a bug report. More often than not, these issues are way too little, they need to be more specific, right? They need more details, they need to be, have more context. So if you have an AI agent within the issue form, you know, along the lines of what we're doing in Copilot Workspace, that actually understands the codebase and then can tell you, well, what you described here isn't good enough. Here's a spec, you know, for what you just said, and you can edit the spec and basically iterate with the help of an agent, you know. And I don't think we would, you know, be scared of that kind of agent taking away our job.

• 50:40Speaker 01

  It's more like, oh, this is great, because now I can, you know, reflect on my own thoughts and make them better.

• 50:47Speaker 00

  Yeah, freeing up humans' time to do what humans do best and automating the things that are just toil at the end of the day. This has been fantastic. We're running out of time. We have a final section we like to do called Quick Fire. I'm going to move us into that even though I think Jordan and I could ask you questions for several more hours if we all had the time. But moving into Quick Fire, just rapid answers as they come to mind. What do you think is overhyped versus underhyped in AI today?

• 51:17Speaker 01

  Overhyped, I think AGI and also the fear of AGI. Underhyped, the impact that AI has on biotech and climate change and all the other things that are a bit further away from traditional tech.

• 51:30Speaker 02

  You just mentioned that the fear of AGI is overhyped, but does anything about Gen AI scare you?

• 51:37Speaker 01

  Nope.

• 51:38Speaker 00

  Not even as a parent?

• 51:40Speaker 01

  Nope, I couldn't be more excited.

• 51:43Speaker 00

  Huh. I'm excited, but I worry a little bit about deepfakes and just making sure that we're always keeping up with with technology around deepfakes and making sure that it doesn't have an adverse impact on our society. But I hear you on that one.

• 51:58Speaker 01

  AI will protect us from deepfakes. You know, AI, screen time, you know, is a big topic in our house and I can't wait for AI to say, well, you know, coding is good, you had enough gaming today instead I have to basically create exclusions for VS Code and Python in the screen time logic on on Mac OS and iOS. And then the same with books and podcasts, those are totally fine all day round.

• 52:25Speaker 00

  Yeah, actually, Jordan and I invested in a company recently that is helping protect us from deepfakes, so we're very much aligned there. Which incumbent besides GitHub and Microsoft have you been most impressed with the way that they've executed on their AI strategy?

• 52:42Speaker 01

  I mean, how can I not say Nvidia, which, you know, has been remarkable over the last two years.

• 52:48Speaker 02

  What is one thing that you thought would work with Copilot that didn't? And one thing that you thought wouldn't work that did?

• 52:55Speaker 01

  Well, you know, I'm a skeptic German and a developer. So when we first started, you know, looking into this, I didn't believe any of it. And then it did work. And I remember, you know, that Zoom session where we, for the first time, asked an O G P T three back then to to write code and it did. And then we were in JavaScript or whatever. And then we asked it in Python and it actually had the correct syntax. And I couldn't, I couldn't believe that it is able to put the commas and the colons and whatnot in the right position, given that the model doesn't actually have a compiler built into it. And I think on the other side of that is, you know, we didn't, we didn't believe that chat would work and it did.

• 53:38Speaker 01

  So certainly chat G P T opened our eyes that you can have this kind of like conversation with A I.

• 53:44Speaker 00

  And which A I startup are you most excited about right now?

• 53:48Speaker 01

  There's a lot of them. I really like what Ranvee is doing. I love, I mentioned silicon companies. I'm an investor in edged A I, but there's, there's many more. And then there's, you know, a bunch of them that are still in stealth mode. So I can talk about them, but it's biotech. I mentioned that as an example. It's super exciting to me.

• 54:05Speaker 02

  And last one, you have a very cool job running GitHub. But if you were not doing that, what A I application would you build?

• 54:12Speaker 01

  Well, if I not be doing GitHub, that means somebody pushed me out of GitHub. So I would build the next GitHub.

• 54:19Speaker 02

  Nice. Great answer.

• 54:20Speaker 00

  I love that. What a way to wrap up the podcast. Thomas, thanks so much for being here with us today. It's been a ton of fun. Really appreciate your time. Can't wait to see what's next from GitHub.

• 54:32Speaker 01

  Thank you so much for having me.

• 54:35Speaker 00

  Well, that was certainly a lot of fun for me. You know, obviously I've known Thomas for a long time. We worked together during my time at GitHub where he was in various roles before becoming C E O and I was C O O. So this was a bit of nostalgia for me, but how is it for you, Jordan?

• 54:52Speaker 02

  Yeah, it's always it's always cool when you've seen someone speak through various keynotes and videos online and stuff, then you finally get to actually chat with them. And I was super impressed. I mean, I thought the episode was great and then a lot of just really interesting areas that we can talk about now. You know, one of them I thought was interesting was around the future of developers. He obviously has a perspective that, you know, the developer will never go away and, you know, natural language is their most powerful tool. You know, I have a slightly different perspective, which is I do think the number of developers that you will need over time is going to decrease, particularly in the areas of, say, front-end tooling, even back-end as well. My perspective

• 55:33Speaker 02

  is, you know, infrastructure and building really sophisticated sort of next-gen products will never go away. You'll still need, you know, developers to really think through, you know, complex sort of workflows and things like that. But I do think it's going to be impacted a bit more than he thinks. So I'm curious your reaction.

• 55:50Speaker 00

  Yeah, I mean, look, I don't think he can say anything else, given his role and who they sell to. I mean, GitHub's MO has always been sell to the developer hearts first, right? And then the business will follow. So it doesn't surprise me that he says that and that to some degree he has that perspective just because of his day-to-day. I do think the number of developers needed within any particular company will absolutely go down. I think there's a lot of manual and rote tasks that'll go away. I mean, look, we've already seen this with QA, right? You see far fewer QA roles than there used to be

• 56:32Speaker 00

  just because of automation and AI takes this to an entirely different level. And so I think there's going to be a lot of roles that are subsumed by AI. I do think, you know, going back to our conversation about the learning journey, it's going to be interesting to see how we actually take people, like, you know, I love to talk about my 12-year-old, obviously, but like, I think about what his coding journey will look like over time. And right now, of course, he's using mostly autocomplete and he's getting help when he's learning a new language or something like that. But like, how do you take someone from... Learning how to code to being able to understand these more complex tasks and more

• 57:14Speaker 00

  complex coding projects or code bases. As you talked about, when a lot of the more basic stuff is going to be done for them. So the learning journey is going to be very different. The one other thing that I'd say is while I think the number of developers within a particular company or to build a particular product will go down. I think we're going to see a lot more companies, like the barrier to building software and creating apps is going to come down so much. I think we're going to see a lot more kind of micro businesses, a lot more competition in certain segments of the market. It's this whole unleashing of human creativity that's going to be really cool. You know, how that

• 57:59Speaker 00

  plays out over time, how some of those grow into big companies, I think still remains to be seen.

• 58:08Speaker 02

  Totally. I agree with all of that for what it's worth. I think we're also going to see a lot of non-traditional, you know, non-developers move into developing as a result of all of this. So definitely interesting, you know, for the future what will happen. What areas of the conversation do you think were were the most interesting?

• 58:25Speaker 00

  You know, I was I was actually just going to ask you this. I think maybe it's obvious to us and it's it's obvious to Thomas, but you know, we talked a little bit about how Copilot is not just it's not like it's just Copilot sitting on on top of OpenAI, which some people might envision it that way. It's actually a ton of different models and even versions of different OpenAI models coming together to to provide this end to end like software development experience. You know, how do you think about his comments about, you know, multimodal versus the bigger models subsuming some of these tasks over time like? Security vulnerability fixing like test writing, et cetera.

• 59:10Speaker 02

  I do agree with him. I think that we are going to have a multimodal world because depending on the use case and which part of the pipeline you're in, you will need, you know, more cost-effective models. Obviously, cost is coming down as a whole, but you know, that's one facet of it. You will need the best performing model for specific types of use cases. You may need a smaller model to be able to run effectively on on-device, for example. And so I think, like, depending on the use case and what you're aiming for and what the goal is, you'll need a model for for different, you know, different use cases.

• 59:44Speaker 00

  What else stood out to you from the conversation? Any surprises in there or anything you felt particularly interesting?

• 59:50Speaker 02

  Yeah, I guess one area, you know, is interesting talking about what scares you. And, you know, he said, you know, nothing, I'm all in on AI. And for what it's worth, like, I love what's going on in the world with AI, and I'm very, very bullish on the future from a positive perspective. I do think, you know, what you said, that deepfake detection side of things, I think is really important. You know, I already know people that have been scammed from AI. And so I think that that's one area. And so while I do agree with him, you know, whether it's even bio benefits from AI or, you know, different sort of areas in addition to just cogeneration, I do think having a little bit of skepticism and and understanding of the implications of it is also healthy, too.

• 1:00:33Speaker 00

  Yeah, I totally agree. Yeah. I mean, I think AI is going to bring more more good than harm by a very considerable margin. And I'm excited about the future and and what it may look like and how we might free up human brains to do work that humans are uniquely well-suited to do and maybe eliminate a lot of toil that is just not super productive. But I think We need to be really thoughtful about how we bring this powerful technology into the world and certainly there's a lot of great things that are going to come out of it, but also some scary things. I mean, deep fake just being one of them, but I think

• 1:01:14Speaker 00

  you gave some other examples too. So, I mean, Thomas has always been, I think, a really positive optimist, even though, you know, he talks about being a skeptical German. I think both things are true in his case. But yeah, I think my, my reaction might be a little bit more measured to him in terms of what's coming down the pipes with AI. You know, one of the things he mentioned that we didn't get to dig into, but I think is really interesting, is how much Cobol and like legacy code exists in the world. But Cobol is a unique challenge and opportunity for AI, I think, because most of the Cobol developers have retired and we have these incredibly important systems that are powering

• 1:01:59Speaker 00

  a huge part of the US government, a huge part of our financial infrastructure and hospitals, all kinds of really mission critical software. Like, what's going to happen with AI? Is AI going to learn Cobol and just maintain these systems? Or over time, are AI tools going to get so powerful that we actually can realistically re-implement these systems in more modern technology?

• 1:02:29Speaker 02

  Yeah, no, it's a, it's a good question. Cobalt does come up quite a lot as of the last few months, for sure. There's actually some startups going after just this use case of converting Cobalt to more modern code. And so I, I do think that is a use case that can be solved by AI. You know, at this point, there's so few Cobalt developers, to your point, and you know, it's sort of dying off the number of people that can actually deal with this. And so I do think it's a, a good use case for AI, provided that the banks and, you know, all the, the various institutions that are built on it are actually open to that idea and sort of touching these really, really, you know, important parts of the code base for the actual economy and, and our government. So as long as they're

• 1:03:13Speaker 02

  willing to do that, I do think it's a, it's a good use case.

• 1:03:17Speaker 00

  It feels like they're going to have to at some point. Like, for how long can we keep these? Fairly obsolete technologies powering systems and like, how do you even maintain them over time?

• 1:03:29Speaker 02

  Totally.

• 1:03:29Speaker 00

  Are they running as efficiently? And are they as powerful as they could otherwise be? It's exciting to think, I think, about being able to finally modernize some of the critical infrastructure in our country, just to help us continue to evolve that technology to meet our ever-changing needs.

• 1:03:50Speaker 02

  Well, Erica, I know that I really enjoyed this episode. I thought it was really interesting and I think our listeners will really enjoy it, too.

• 1:03:57Speaker 00

  I had a blast recording it and I sure hope so.