GitHub CEO Thomas Dohmke on Building Copilot, and the the Future of Software Development

Contents

Thomas Dohmke: The human brain is still so much more advanced than the transformer models and the diffusion models and the other types of models that we have for image recognition and whatnot that we have today. And, you know, it remains to be seen if you can kind of like add that sentience piece to it. But today I’m not seeing it, and I haven’t seen any research that’s telling me that that’s coming anytime soon.

Stephanie Zhan: Hi, everyone. Welcome to Training Data. Today we host Thomas Dohmke, CEO of GitHub, and together we talk about GitHub Copilot’s path to $100-million ARR in two years, the fastest growing product and business in generative AI after OpenAI itself.

Thomas has an ambitious vision for enabling a world of one billion developers, and bringing agents through the end-to-end developer workflow and into adjacent categories like code security. He even hints at the progress he thinks the industry will make on SWE-bench over the next few years, Some categories he’s excited about in AI outside of developer tools, and whether he thinks a new architecture will overtake the transformer.

Today we’re so excited to introduce our special guest, Thomas Dohmke, CEO of GitHub.

Thomas Dohmke: Hey, and thank you so much for having me.

Getting started with code

Stephanie Zhan: Thomas, we’re so excited to dive into the GitHub and Copilot story in particular today. Maybe to kick off, we’d love to learn a little bit about your personal background. You have a very interesting story, having grown up in East Berlin before the wall fell, and then starting your first company that brought you into the United States when Microsoft acquired the company, how did your background and upbringing really shape who you are today?

Thomas Dohmke: I think, you know, I’m living a very normal life, the North American dream now—with a wife and two kids. I think what shaped really my journey was the passion for software development early, you know, when I was an eleven year old or so. It was still East Germany and West Germany. There was the wall between the two parts of Berlin, and I saw computers for the first time. I couldn’t buy one, but in school we had one in the geography lab, and a friend of mine and I, we started playing with that, learning to code. You had to code to even do anything with the machine. You know, you needed to understand some Basic to even load a program. And then as the wall fell, I bought a Commodore 64 and later my first PC, a 386 DX-40.

And so as a teenager I spent most of my time coding and started a company. It wasn’t at that time really a startup. We just did insurance software. In the late ’90s, most of the insurance agents didn’t have software yet. Some were working on mainframes, and others just had paper in front of them. And then I moved to South Germany to work for Mercedes. And then I had the startup that got acquired by Microsoft. But it’s really this passion for doing stuff with software, being creative with software. And I think the fascinating thing back in the nineties—and it’s still true today—is that you can start very easily. There’s not a lot of capital investment required, and if you make a mistake, you can just start from scratch. And I think that’s what makes this so cool, to build software.

Stephanie Zhan: And that gave you the love for building and fixing things even today with robotic lawn mowers in your own home. [laughs]

Thomas Dohmke: [laughs] Yes. Well, and it gives me the love, you know, as the CEO of GitHub to build software for software developers. I think that’s the really cool thing about being at GitHub. We are building the tools that other developers are using. We always say we put the developers first. And that, I think, is the dream job for me. I get to speak with a lot of developers. I get to build software for developers, and I get to speak to many developers here in Silicon Valley.

Microsoft’s acquisition of GitHub

Stephanie Zhan: I think what many might not know about you outside of Microsoft is that you were actually—you played a pivotal role in sponsoring Microsoft’s acquisition of GitHub back in 2018. Can you share or take us back to that moment and share a little bit about what was your vision for GitHub back then?

Thomas Dohmke: Yeah. So in 2018, I was a product manager at Microsoft working for Nat Friedman, who at the time was the CVP for mobile developer tools. And Nat had the idea of buying GitHub and making it a part of Microsoft. And so he and a few folks in his team were kind of like strategizing of how can we pull that off and pitch it to Satya and the board? And as the deal got announced, we’re actually about six years away from that at the time of this recording. June 4, 2018, is when we announced the deal. I became the deal integration manager, which was the role within Microsoft that runs around the whole company, making sure all the pieces from legal to HR, finances and product and engineering come together to get that deal through the regulatory approval, and ultimately now getting us to a successful day zero, which then happened in October 2018. And that’s how ultimately I came to GitHub.

Stephanie Zhan: What did you envision GitHub becoming, the potential it had as part of the GitHub universe?

Thomas Dohmke: I think, you know, at the time we were thinking there’s so much potential for GitHub that’s still, you know, to be explored and to be realized. GitHub started with the first commit in late 2007 and launched in 2008 and, you know, very quickly became a new way of developers working together. You know, social coding was invented by Chris and the other founders, and it evolved into this two-sided product, the home of open source, where many open source developers were collaborating, and the place where many startups and ultimately enterprises were building their software. And often that two-sided equation was that the companies wanted to work exactly like open source developers work, which is boundaryless. You know, in the world of open source, you don’t really care where your collaborator sits, where they’re from, what their education is. They don’t sit in an org chart, and often you don’t even know their real name. All you know is their handle and the code that they want to contribute back to the project.

And I think many companies don’t have that. You have silos, and when you get an email, you’re like, “Wait, who is this? And why are they emailing me? And why do you want to be involved in my project?” And so companies want to burn down these walled gardens and have a similar collaboration model. And that’s what GitHub symbolized in 2018. But there was so much more to do to provide something like GitHub Actions that allows developers not only to manage the source code and plan, but also to build their apps. And now ultimately with Copilot, to take it even a step further, and making development a very different experience than it was 20 years or 30 years ago when I started.

The Copilot origin story

Sonya Huang: Since you mentioned Copilot, I’m dying to ask you about the behind-the-scenes view on what is not only the most successful enterprise AI application today, but I believe the first LLM-native application that was really built—built and launched. I guess, was it part of the original acquisition thesis at the time that you might be able to build something like Copilot eventually? Whose idea was it to do something like Copilot? And did everyone say, “Yes, this is gonna work?” Or was it like, “This is a crazy idea. This is a moonshot, it’s never going to work.” Like, take us back to the origin story.

Thomas Dohmke: So the original acquisition thesis had a little paragraph in it about AI, but I think that was more like a moonshot than like the proper idea at the time. What really happened in mid-2020 is that transformer models, the paper came out a few years before that, but the first really working transformer model was GPT-3 was about to launch. We got early access, we were all in lockdown, June, 2020, on a call. And one of our team members, Oege de Moor, started typing things into the model, and everybody else was just looking what Oege would be doing. And so then the natural question was: can we prompt it to write code, and can it write proper code?

And I think that was the first kind of like a-ha moment that it was actually able to write the real syntax. And then we tried different languages and we also found flaws in the model. And so we started exploring that deeper outside of that Zoom call in a research process doing analysis. We asked some of our staff and principal engineers to submit coding exercises. We looked at Python functions that were in open source repositories on GitHub, and we worked with OpenAI to take GPT-3 and fine tune the model to be better at these coding tasks. And ultimately, in August we had a model that was able to solve 92 percent of these coding exercises. And in fact of the Python bodies that we extracted from open source projects, it was like 52 percent.

Now naturally that percentage is lower because you have less specified code than in a coding exercise for an interview loop, right? But that, I think, gave us this moment of confidence saying we can build a product around this. The second moment I’d say was when we rolled it out to our internal engineers in early 2021, and they came back with saying “This is fantastic.” I think the net promoter score, NPS, was somewhere in the 70s, which is, I think, for developer tools is really, really good. Most developers are skeptical, you know, “Don’t touch my system. Never touch the running system or process.” And many folks have created their work setup not only on their physical desks but also on their virtual desk of how they want to work. And so we were really intrigued by the internal responses. And then we launched the preview, and the team came back and saying it’s writing 25 percent of Python code in those files where it was enabled. And I think we sent them back saying go and verify the telemetry, that can’t be true, right? Because we couldn’t believe it initially. And then as we saw this progressing, I think in my first keynote as CEO was in June, 2022. So now two years ago, and I said “It’s writing 40 percent of the code, and I think in five years it’s going to write 80 percent of the code in those files where it’s enabled.”

Sonya Huang: So it happened really organically. It wasn’t like you were sitting in a room whiteboarding, like “Copilot is going to be the next iteration.” It was like the model proved to you how great it was, and just seeing the model performance and seeing usage really organically built up.

Thomas Dohmke: Going back to the original thesis for GitHub, we wanted to make developers’ lives easier. And as we are a company building software, we have our own software developers, and Microsoft has 70,000 or so of them. And so we understand how software developers work because we have so many of them, and we live the life of many of our customers, which is we have way too many ideas. We are moving way too slow, at least in our feeling. Amazon delivers my package faster than we’re implementing some features. Also, the expectations have shifted significantly, but the backlog is endless, and the amount of ideas that we can brainstorm on virtual whiteboards—GitHub is a remote company, so we don’t often meet in front of real physical whiteboards. And then the other side is all the other work that we also have to do, you know, compliance, security, accessibility, enterprise requirements, privacy regulations, European AI Act, and Digital Markets Act. And all these things also cause developers work. And so everybody’s constantly struggling with the length of these backlogs. And so if you will, the intrinsic motivation was, let’s bring the effort down to write software and make it joyful again.

Evolving Copilot beyond autocomplete

Stephanie Zhan: What do you think made Copilot so good at the time, and also going forward today? Obviously GPT-3 was great even back then, but also I think a lot of people might not know all the value that you brought or GitHub brought in the both public and proprietary data around code that GitHub owned. Can you share a little bit more about that, and also how you think about that going forward?

 

Thomas Dohmke: I think the key ingredient of the original Copilot, which was only auto completion, right? Like, you would type in your editor and it would complete the next line, but it could also complete multiple lines of code, complex algorithms or simple algorithms. Some of the demos we often chose to implement—a sorting algorithm, like bubble sort or prime number detection, and it can just write those 10 lines of code by just a simple prompt, which is a comment in the code, or just writing the method declaration.

 

And so getting into the editor where developers already write code—not changing the way they work, but giving them ideas while they’re typing—I think that was the key moment, other than obviously the model being good enough, and OpenAI tuning that model on publicly-available source code from GitHub. You know, GitHub didn’t get special access to OpenAI. OpenAI was just able to access our source code in the same way that many other startups are now doing that, either through direct access through our API, or through archive programs like the Internet Archive and the Software Heritage. We actually have an established partnership with them to stream our open source code over there so it can be archived until the end of time. And so of course, the model got tuned to be good enough, but then the UI and the user experience, I think, was crucial. It wasn’t—you know, AI is now in everybody’s mind, but the truth is our cell phones have some kind of AI built in for a long time. Your keyboard is predicting the next word with some kind of machine learning algorithms. Your photo library, you can just go in there and search for a license plate and find all the photos of the cars that you took to remember what the license plate you have. And that’s AI image recognition, but that nobody perceives that as AI. It’s just a convenient feature. And so I think that also was the core ingredient of Copilot. We’re meeting developers where they are and we’re making their life better.

 

And then I think the name itself was also a stroke of genius. One of our other developers, Alex, came up with the idea to name this Copilot. Nat, he’s a hobby pilot, and so that’s where the name is coming from. So it’s like I’m thinking about, what could I name this? So my boss, the name is resonating with him.

Stephanie Zhan: Wow, that’s so interesting. I didn’t know that. If you could go back to 2020, 2021, is there anything that you might do differently?

In hindsight, you can always move faster

Stephanie Zhan: Wow, that’s so interesting. I didn’t know that. If you could go back to 2020, 2021, is there anything that you might do differently?

Thomas Dohmke: You know, in hindsight, you can always move faster and be, I think, more convicted of those ideas. I think in the beginning we kept the team intentionally small. Small teams can move fast.

Stephanie Zhan: How big was the team?

Thomas Dohmke: I think the original staffers were like three people, but that’s obviously staff researchers—sorry, or principal researchers. But I think that’s cheating a little bit in the sense that, of course there was a team at OpenAI, there was a team or multiple teams in Microsoft, both on the research side and on the inference side, the model inference side, in fact, you know, on the model training site to even enable the model. So of course, in the bigger partnership between Microsoft, OpenAI and GitHub, it was a larger team, but the original paper was written by three researchers, and then a hundred people or so are mentioned in the credits. And then we move fast with, I think, a staff of five, and then I think it increased to ten teams. And of course, today the team is much smaller, but we still have what we call GitHub Next, an incubation team that now works on Copilot workspace and iterates on new ideas. It’s almost like a startup incubator within the company. And they’re picking up ideas, and the main difference to our mainline engineering teams and product management teams are that they got to have the mindset that most of their ideas will never go into production, right? It’s kind of like this okay, ours won’t work because your key result is to throw away most of the ideas and start fresh with another idea. And I think that’s where a lot of the innovation speed is coming from.

Building on top of OpenAI

Sonya Huang: I want to go back to what you said a minute ago about how you handed over control of the model effectively to OpenAI. Was it scary because the brain of your AI application is actually being built by another company, not developers under your payroll, your control? How do you think about—was it an easy decision to kind of partner with a different external model provider? And I’m curious how you think about the value that Microsoft and GitHub provide to end users, versus the value that OpenAI provides to end users and where you seek to really bring value to users.

Thomas Dohmke: To me it wasn’t scary at all. I’m sure there were folks in the team and in the company that thought we should have our own models instead. But in reality, you know, if you look at GitHub as a company that was born in the cloud, of course we have always relied on partners to build our stack. To my memory, we have never built hardware ourselves. And where we have metal in data centers or had metal in data centers for a while, the data center itself wasn’t built by us either. Neither was the CPU and the memory and the network infrastructure. And then even if you go higher in that stack, GitHub is built on top of open source and so are the majority of software applications.

In fact, we love sharing. A statistic that says 90 percent of the stack of most applications, even if the application itself is closed source, is in fact based on the work of the open source community from the operating system—the Linux operating system is ubiquitous today on servers—to container technology like Docker and Kubernetes, to thousands, sometimes tens of thousands of open source libraries. And so the model just flows naturally in that stack. And we at Microsoft think about this as the Copilot stack with the different layers, hardware, the model, the kernel if you will, like the infrastructure, responsible AI, filtering and whatnot. And then you get into the application layer with the AI Copilot and then the extensibility on top of that.

And so if you look at these layers in the stack, Microsoft has strengths and partners. And in fact, in some form Microsoft is involved in all parts of that stack. The Copilot PC has a custom chip that Microsoft is developing with partners. We have our own models with Phi-3, and we have partnership models with OpenAI, but we’re also hosting Mistral and Llama on Azure. We, of course, have a large cloud and we have a lot of expertise in responsible AI. And then we have lots of applications we are building ourselves, and we’re enabling others to build the applications. And so two part questions are always hard for me—I forget the second part, but I think that actually describes the relationship very well. It allowed us to move really fast because we could rely on partners like OpenAI not only building GPT-3 and then Codex, but also then innovating with GPT3.5 and ChatGPT, GPT-4. Now we have GPT-4o, we have Microsoft with a large infrastructure that builds supercomputers for training, but also infrastructure to run inference. And Copilot today runs in multiple data centers spread around the world in different regions, so the developers that sit in France now connect to a GPU, to an Azure instance that’s much closer to them to enable low latency, right? And so Microsoft gives us a lot of infrastructure, a lot of expertise and responsible AI, and of course a lot of commercial distribution.

Sonya Huang: Yeah, a very complicated layer cake that makes the magic come together.

Stephanie Zhan: Would GitHub ever want to build its own models, or just use the best in class models out there?

Thomas Dohmke: Yeah, obviously there’s always the desire of engineers to build their own stuff, and I would not deny that we have played with our own ideas on models. I think the machine learning team goes back almost—I think it actually had it before Microsoft acquired GitHub. And then we obviously have fine-tuned models ourselves, and we have fine–tuned models with OpenAI and Azure. We are working with customers on them being able to customize models based on the code that they have in their repositories. And never say never what the future may bring, but today we’re really happy with the models that we have, and we’re constantly looking into the market of what not only OpenAI provides to us, but also what others have.

The latest metrics

Sonya Huang: So Copilot is already one of the most successful generative AI applications in terms of user scale, usage, et cetera. What are some of the latest metrics you can share, and what are the metrics that you’re most proud of?

Thomas Dohmke: I think the one I am most proud of is the developer happiness scores. And if you look at survey data, both the service that we have done, but also the service that now our customers either publishing themselves or bringing back to us, it’s clear that software developers, after they have tried Copilot, after they got over the initial adoption hurdle or skepticism, most of them love using Copilot. And most of them report that they are more fulfilled, they’re more satisfied, more happy. They feel like they’re requiring less mental energy to get the job done. They need to do less boilerplate. And I think that alone is really making me happy.

And there’s lots of numbers that I can throw out, but I think the general gist is no matter what developer I talk to, those that have used Copilot for a while no longer want to work without a Copilot. And then the other side is, you know, the productivity metrics of making developers more productive, which I think matters to the developers, too. But it also, of course, matters to their management chain and their leadership in the sense of getting more stuff done, delivering more value to their end customers.

And so today we are happy to say that we have more than 1.8 million paid subscribers on Copilot and more than 50,000 organizations. And that really makes us happy looking at that growth. And, you know, we’re here at Sequoia’s office, it feels like we are like a high-growth startup, and that’s a good place to be in.

Stephanie Zhan: [laughs] Awesome. Well, you’re still on our wall somewhere, actually. We know exactly where. We’ll show you.

Thomas Dohmke: Me too! The creative minds.

The surprise of Copilot’s impact

Stephanie Zhan: Was there anything that surprised you in terms of Copilot’s impact after its launch and even today?

Thomas Dohmke: I think, you know, it’s the 25 percent certainly surprised us. The quick turnaround from the skepticism after we announced this in June, 2021. I think we had a very short blog post and then just a webpage with examples and animations showing how it will work. And I think folks were looking at this and saying, “This is like a cool tech demo, but it doesn’t actually work for me.” And I think the skepticism was that people had seen how GPT-3 at the time would work, and they couldn’t understand until trying it that it has the context of whatever you wrote in the file before Copilot suggestion comes, and it considers adjacent tabs and things like that. And so it magically picks up your style and it knows about open source libraries that you’re using, not because you have opened those libraries, just because you have an import statement at the top, and because the model was trained on such a large corpus of data that it can provide the calls into these open source libraries. And so it feels to you that the Copilot understands more about your project than you thought GPT-3 can do. And I think that’s where the differentiation came from.

I think the other thing is today when I observe people using Copilot and obviously ChatGPT changed the whole game and brought in chat as a component. We have now Copilot chat. It’s this natural language component, like using not only code and comments as a trigger, but English and German and Brazilian Portuguese. And we have demos in India where Karan M.V., one of our folks there, is demoing this in Hindi. And so you can actually speak into the Copilot with voice detection and visual studio code and then gives you the response back also in Hindi.

And so it’s really cool for anyone that wants to explore coding, even if they’re not fluent in English or in programming languages. My kids are using it to find their own bugs in Python, so they’re no longer coming to me and it’s like, “Daddy, you have to find my bug.” It’s kind of like, “Well, go and find your own bug.” And that also obviously helps them to develop their skills. And it’s a bit like I often talk about natural language will democratize access to software developers, but it doesn’t mean that everybody is immediately a senior or principal software developer in the same way that just because I buy a guitar, I’m not as good as Keith Richards playing with the Rolling Stones, right? And if you look at any professional band that’s touring the world, they all still rehearse over and over again. And I think this is this idea that to be good in a craft, you have to keep doing it. Copilot does not take that away, it just gives you another tool in your toolbox.

Stephanie Zhan: I really like that analogy.

Teaching kids to code in the age of Copilot

Sonya Huang: So you still believe in teaching your kids to code? Because that’s like a debate on the internet now.

Thomas Dohmke: Oh, absolutely. I mean, first of all, the human language is not deterministic, right? You mean different things—you can mean different things by saying the same sentence. And even the …

Sonya Huang: That’s very German. [laughs]

Thomas Dohmke: Even the—well, now we can get into off topic debates about yeses and nos when you’re answering a question with a negative in it, right? Like, “You have not been to the grocery store.” Do you answer that with yes or with no, right? And the Americans expect a no, even though you mean actually yes to that question. But like, look, you know, human language is not deterministic. And so code is. With code you can very precisely describe what the machine does. Code is an abstraction layer on top of assembly language, on top of the instruction set of the CPU or GPU that the processor or manufacturer, like Intel or Nvidia has created. So it’s just another abstraction layer. Human language is something completely different. It’s creative, and that’s the power. But it also means that there will be code involved one way or another. And we’re moving up the abstraction layer, but we’re also spreading the meaning, and that’s truly powerful. But it also means that there will be some conversion to code somewhere, because the chip itself, at least today, is requiring a deterministic instruction set.

The momentum mindset

Sonya Huang: I love that. I’d love to talk about the future of Copilot. You’ve been announcing new products in rapid fire succession, Copilot X, Copilot Enterprise, something around code security, I believe, and then Workspaces most recently. Can you tell us maybe about what each of those things does, and how you see them all fitting together in the grander vision for what you hope Copilot becomes?

Thomas Dohmke: You know, what you described with all these product names is part of our mindset is that momentum is our energy. In this age of AI, moving fast and iterating fast is crucial. And so we are having developed Copilot from this original idea of autocompletion by adding chat in visual studio code. And we had this—we were developing autocompletion and then chat. And then with chat, we also announced something what we called Copilot X, which the idea was we are bringing AI features into every part of the developer lifecycle. We’re bringing Copilot wherever developers are, so while we added chat to the editor, we also added a little Copilot icon into the input field where you write your commit message.

And that might be trivial, right? Like, everybody can write a commit message, but it also means I reduce my mental workload and I reduce the bias that I have for the work that I created myself. For me, everything is obvious that I just did for the last hour. But for you, when you want to review my commit or my pull request, it’s not as obvious. And so having an AI describe that in a neutral form describing what I just did, is incredibly useful and it just keeps me in my flow. And we added it to the debugger. We added it to many different parts of the lifecycle already.

And with Copilot Enterprise, we bundled it into a higher-priced product that allows enterprises to customize Copilot based on their institutional knowledge. And Enterprise here means really any company that has gone for more than a few weeks, because they all immediately build institutional knowledge, right? How we work as a team, how our coding practices, these are the libraries and languages that we use. And so unless you’re a student in university that gets to have the free range of technologies available, at least when the professor allows that, every time you join a company or join a different project, you have to ramp up again on how they are doing things. And so Copilot Enterprise lets companies customize the Copilots to their institutional knowledge. And it makes it really easy for me to join that company because I can now ask dumb questions without you judging me. [laughs] Like, imagine I would join here my first day and, like, “Why is Thomas asking all these questions? Shouldn’t he already know all that? He’s been in professional life for a long time.” That’s the challenge that we have when we join companies, and we have the anxiety in our head that we can’t ask too many questions before Steph says, “What the hell?”

Stephanie Zhan: [laughs]

Thomas Dohmke: Right? And so that’s, I think, the power of Copilot Enterprise, and that’s the power of bringing Copilot into every part of the developer lifecycle and ultimately into every part of our lives.

Agents vs copilots?

Stephanie Zhan: You’ve also mentioned that agents are one of the most important next things for GitHub. Maybe just to set the stage, how would you describe an agent versus a copilot? And can you give us a teaser for what types of agent capabilities we should come into Copilot soon?

Thomas Dohmke: Yeah, I would say copilots and agents is kind of like the same thing. An agent is using a model to get a task done, right? Effectively, it’s looping with a model to solve something for you. And a copilot is an agent of agents, and it has multiple features available to it. You know, if you think about autocompletion, whether it’s an agent that takes every keystroke you did and the context that you have in your editor, it sends it to model inference, it gets the response back, it might pick the best prediction, and then it shows it to you. And so we are going to see more of these agents that take over more of our tasks.

And one of them that I’m most excited about is Autofix. And the way it works is so you submit a pull request, and traditionally, some security scanning feature that you have integrated into your pipeline finds security vulnerabilities, let’s say SQL injection or cross-site scripting. Well, that’s great, except now I caused more work for myself. It’s kind of like you have a Roomba, but instead of vacuuming your house, it just shows you where the dirt is, and then you have to go and vacuum yourself in that position.

So now with Autofix, we’re actually not only showing you the security vulnerability, we’re also giving you the fix. And that uses the AI model together with a vulnerability and a description and the code to basically solve that vulnerability for you. And the initial results are really impressive. With some customers we see that we can burn through like 75, 80 percent of their open alerts. Because everybody has those alerts. And if you don’t have any alerts right now, I bet you you have them by Monday. Like, that’s the challenge in this world of software security is everything around us moving so fast. There’s always a new version of an open source library, there’s a new version of Linux or a new Windows patch. There’s a new device coming along the way, or a new Nvidia GPU. And so we constantly are behind just keeping our applications up to date to the standard that is expected by our customers. And at the same time we have to build all that innovation and all that cool stuff. Yeah.

The roadmap

Sonya Huang: What else do you think is missing in the product roadmap? Like, if you could wave a magic wand, what else is there to build that you’re really excited about?

Thomas Dohmke: I mean, I think there’s still a lot of work to do on these agents. I think there’s a lot of agents you can think about—we talked a little bit, or you mentioned Workspace before. What Copilot Workspace does is it provides different agents to get you from idea to your pull request. So get you from idea to the code. And the first one is the spec agent. And what that actually does, it helps you with your thought process. So you write down an idea, implement some feature, and it looks at your existing code base and it basically helps you then to reframe that idea. Now that’s not only useful for developers, it’s actually useful for a product manager because it might tell you, “Well Thomas, this idea, there’s no way you can describe that with a single sentence. And a developer cannot implement that in just a single ticket. It needs to be an epic or multiple different user stories.”

And then the next step is the plan agent that helps to figure out where to make the changes in the code base. And again, you can see here there’s a lot of other benefits you get from that because it helps you to understand the code base, because most code bases that are older than a few days have hundreds if not thousands of files. And as developers, you have to navigate all those files, and even if you have been in a code base for a long time, you still might miss out on that one file that you haven’t touched for a while and you have to add a config statement there. So it helps your understanding the code base, and then the implement agent helps you implementing the code change.

And every step in that way you’re still in charge. You can, with natural language, modify the bullet points of each of these agents, and then obviously you can modify the code at the end. And so if you look at just these three agents, you can easily start thinking about other agents that you might have along the way, right? For example, the one that estimates the size of a ticket, the story-pointing agent as one example. Or another one is that once you have implemented the file, well, now you want to build, run and debug the file. And maybe you have an agent in the future that will just automatically fix any bugs that were introduced by the previous agents. And so I think we are going to have more of these building blocks, you know, Lego blocks, if you will, available to us. In fact, if you look at Lego, they have way more types of pieces today than they used to have, because the models are much more complex and, you know, you can buy NASA rockets and whatnot, and they need different pieces for that. And I think that’s kind of like the same way we should think about copilots. They will have more of these building blocks that enable us, in addition to more powerful models and a mix of models, these building blocks will enable us to do more.

Stephanie Zhan: Increasing modularity. Interesting. Where does your ambition for GitHub take you—or maybe even with GitHub Copilot specifically—take you as you think about what you alluded to from the perspective of deepening where you can go with just the software engineer to also expanding into potentially different personas. PMs you mentioned, maybe an SRE, maybe a security engineer. Where does that breadth also take you?

Thomas Dohmke: Yeah. I mean, first of all, I think all those roles today are already collaborating on GitHub. In fact, GitHub had always had that mantra that we are building GitHub on GitHub, and that sometimes we have pushed it a bit too far. But today, most of our GitHub employees—we call them ‘hubbers’—hubbers are engaging on GitHub, in GitHub discussions and GitHub pull requests. Our legal documentation is all on GitHub, which if you think about it, is actually much better than managing redlines in Word, because you have a version history, and you can see who made what change. And in fact, you can see soon a future where maybe your legal document is explained by a copilot in human language, in the actual understandable human language, not the lawyer language, right? And so we’re using GitHub to run our company. But yeah, it’s where all the developers and all the supporting functions collaborate on a project. So that’s, I think, number one.

Number two is we want to democratize access to software development. And, you know, I recently gave a TED talk and I talked about that. I think our goal is to get to one billion software developers in the world. Now that doesn’t mean one billion professional software developers, although that might not be a bad thing necessarily, given the demand is still very high and it’s sometimes hard to find qualified software developers. But it’s really about democratizing access to writing software on these devices that are with us.

You know, our mobile phones are a really important part of our lives today. You can’t really imagine life, urban life for sure, without a mobile phone. And so then also being able to write little applications or little scripts or just using natural language to control the phone, I think is incredibly empowering. And so bringing that into 10 percent of the world’s population by 2030 or so, assuming that then there are 10 billion inhabitants on this planet, I think is going to create a better world, and it’s going to unlock creativity everywhere. And hopefully, you know, we see cool venture-backed startups in India and in Brazil. And maybe the next big tech company is coming from one of those countries instead of the U.S. West Coast.

Making maintaining software easier

Stephanie Zhan: We hope so. Maybe zooming out of GitHub Copilot to broader GitHub itself. GitHub Copilot itself is driving so much innovation within GitHub, but what are some of the other key initiatives that you’re leading across GitHub overall as well?

Thomas Dohmke: Yeah, we already talked a little bit about Autofix and security, and I think security, securing the software supply chain, is near and dear to our hearts. There is no future of human progress with software if you’re not also able to secure the supply chain. Today, you know, there’s this XKCD comic of the internet’s infrastructure, and there’s this one building block that says the guy in Nebraska maintaining this one library alone, right? That actually is—as funny as that is, that is a reflection of the software world today. So we A) have to make it sustainable for those maintainers to build that software and keep it joyful, and then also we have to make sure that all these building blocks that are in our stacks that have become system critical to be secure. And so we’re both investing a lot in platform security and application security, and of course in our security products. And I think that’s going to be crucial, combined with Copilot and AI, to not only create all these work items, but to also enable developers to burn them down to fix all these issues.

The creative new world

Sonya Huang: Yeah. I want to zoom out from GitHub for a second and just talk about how you see the future of AI and coding overall. Like you mentioned, there’s a billion casual developers around the world. What will it look like? Will everybody be kind of coding applications for themselves to use? Will there be some number of professional developers who are super developers? Like, how do you imagine the world looks when you have so kind of democratized the craft of coding?

Thomas Dohmke: I mean, I think it’s an incredibly creative world and a world where you’re not dependent on, you know, when you’re a kid, on your parents having technical knowledge or your school having a teacher that knows how to do these things. So we’re going to have much more access to those that are interested in learning about this. You know, it’s easy today, you know, to take a sheet of paper and paint something, and every restaurant, at least, you know, in this country, gives you crayons if you come with kids and a coloring sheet.

Sonya Huang: Lifesaver.

Thomas Dohmke: Lifesaver. [laughs] Or you have your mobile phone and they use your mobile phone. You know, it’s easy to learn, easy in the sense of accessible to learn a musical instrument. And I think it should be easy to learn coding. And so first of all, I think that should be something we are excited about, and not concerned that we are inflating the number of developers, because just because kids learn it doesn’t mean they want to become a developer. I think there’s still a world where people want to do something else than software.

But then if you think, you know, about many other professions—physicists for example—they use a lot of software. You know, the first image of a black hole was with the help of open source project. you know, the Mars helicopter ran on open source, right? And so yeah, it’s space and it’s space engineering, but they’re using software and they’re building software. And so the profession itself is everywhere. Every company is a software company. Banks are software companies, you know, energy providers are software companies. Farmers are software companies predicting, you know, what seed to plant this year, what the weather is going to be like, you know, what was the soil quality from last year and things like that.

But of course, you know, the hobby scenario is also important. Like, tax season is over here in the United States, but that doesn’t mean that I couldn’t think about next year to automate a lot of that if I only had an AI agent that does all that work for me and downloads all the PDFs and extracts all the numbers. And I don’t think we are too far away from that. And I’m now in software for almost 30 years as a professional software developer and, you know, I don’t have a lot of time to code. You know, I have a company to run. I have podcasts to give and things. [laughs] But the problem today is you find an hour on a weekend and you have a project, and the first 20 minutes you’re spending with updating everything to whatever you missed. And the burden is actually—the fun is gone to a certain degree because the burden of maintaining software is so high.

So having something available to you that gets you quickly into the hobby and out of the hobby or out of that task, I think, is incredibly empowering and brings the fun back. And that’s where the Lego comparison is so useful, right? Because Lego is just incredibly accessible. And even if you—like, the best Lego is the one where you don’t have instructions, they just have a table full with Lego bricks, even in random colors, right? And then have this excitement of play. And even professional workers at their offsites or workshops often have little gadgets or bricks on the table. So you have your fingers do things while you’re thinking. So I think that’s where that world is leading us. And we will have more access to technology. We have more people that can build software. That doesn’t mean that they’re taking jobs away from professionals, professional software developers.

The AI 10X software engineer?

Sonya Huang: Very intellectually honest. On what timeframe do you think we’ll have coding agents that are as good as maybe the average professional software developer? And there’s the legend of the 10X software engineer. At what point do you think AI will be as good as the 10X software engineer in capability?

Thomas Dohmke: You know, the trick in that question is the “as good as.” And what does that mean? So I think, you know, is a model today able to write better code than the average developer? If it is prompted in the right way or given the right context, I’d say we are already there on average, because often the model just knows more about that whole space than I as a human do.

And you see that with students, if they have to implement a conversion from binary to decimal or something like that. And they might write 100 lines of code, and then they go and ask Copilot how to do that, and they get probably an open source library in one line of code, and then you can say, “Well, I’m not allowed to use open source,” and then you would probably still get a better code than they wrote. And I think the same is true for the professional software developer because look, we are not perfect, we are human. And I think that’s part of our nature, that’s part of creativity.

Now that’s the key thing, though, is that the model is not creative and the model cannot—today, the model cannot make decisions for us. Or if it does make decisions, it doesn’t actually take all the constraints into account. Like, if you think about software development other than writing code, which is, I think, the fun part, it often means I take a very complex problem and you break it, decompose the problem into small building blocks. And the block size is increasing over time. It used to be—auto completion used to be just the next word, and then it was maybe a full command, and now it’s multiple lines of code and maybe it’s whole files in the future. But along that decomposition process, you still have to make a lot of technical decisions: what database am I using? And you probably know better how many database startups Sequoia has invested in, and how many infrastructure startups, and how many serverless startups. And obviously there’s all the incumbents in all those spaces. And so there’s a thousand, if not ten thousand decisions to be made. And the engineer is the systems thinker that is making those decisions, or the team of engineers in companies. And I think, you know, that we have been building houses, you know, as humans for thousands of years, and if you ever build a house, it’s still not a solved problem.

Creativity and systems thinking in AI?

Sonya Huang: Yeah, totally. But at what point do you think that creativity and that systems thinking gets built into Copilot, or do you think it never does?

Thomas Dohmke: I mean, it’s a bit of predicting the future. I don’t know if I would say never. Never say—that’s a dangerous thing on a podcast. You know, you invite me back in three years and say, “Well Thomas, you know, last time we asked you about this, and clearly it has happened since.” No, I think, you know, we’ll see whether research goes and where the technology goes in kind of critical thinking and systems thinking and those kinds of questions. Also learning. You know, learning—you mentioned your two year old, your two year old and my kids, they learn, you know, as they mimic the humans around them. They are so good at learning language, and especially in young age, you know, they can learn multiple languages. Mine speak English and German because we speak German at home. And they don’t have an accent in English. Well, they have an American accent, but they don’t have the German accent, right? They don’t have an accent in German either. And I think this shows that the human brain is still so much more advanced than the Transformer models and the diffusion models and the other types of models that we have for image recognition and whatnot that we have today. And, you know, it remains to be seen if you can kind of like add that sentience piece to it. But today I’m not seeing it, and I haven’t seen any research that’s telling me that that’s coming anytime soon.

Sonya Huang: That’s a really clear delineation.

Stephanie Zhan: Switching gears a little bit, and we’d love to hear your thoughts on the overall ecosystem of startups right now. AI code gen is the hottest category with a lot of ambitious founders, and there are so many different attempts that they’re taking, whether it’s folks who are trying to build a better model, folks who are trying to build a better IDE, folks who are trying to build kind of an all-in full stack engineer as an agent. And, you know, but the big elephant in the room is GitHub Copilot, and all the adjacent products that you have around it, with Copilot Workspace, with Copilot Enterprise, with Autofix, and owning VS Code as well, amongst many other things. How do you think about what white space there is for existing—or sorry, for new founders? And if you were a founder yourself trying to build in the space, what would you do? [laughs]

Thomas Dohmke: I mean, I love developer tools, so I’d probably still do developer tools, and I’m not sure I would worry too much about what white space is taken by incumbents, because that can change quickly. When GitHub started, SourceForge was the big elephant in the room, and SouceForge has all the open source projects. And then came Git, and Git was accepted in that space, and then the founders of GitHub took Git and built GitHub. And all of a sudden, you know, that elephant in the room was no longer the elephant. And I think, you know, it is often fun, you know, to compete in the same space. And we love competition because it pushes us forward, you know, as much—it’s boring, you know, to do a race or a game if you don’t have an opponent, and if you don’t have other teams in the league. Who would watch the Super Bowl if there’s only one winning team every year? So I don’t think competition should hold you back, as a founder, from going into that space. And I think the software development space is wide open, and there’s lots of problems to solve. There’s lots of problems in different industries and categories to solve. We haven’t really solved modernization of source code. You know, COBOL runs still on mainframes. You can use a Copilot, and we are heavily looking into that. And because it’s such a pain point for many financial services institutions, you know, your credit cards, your bank account, Wall Street, all that runs still on COBOL. I haven’t met a single bank that doesn’t run some COBOL on some mainframe.

What about COBOL?

Stephanie Zhan: What would you do with Copilot and COBOL?

Thomas Dohmke: Well today you can explain that, which is often helpful because the code was written 60 years ago, and so the people that wrote that code are retired.

Stephanie Zhan: Yeah, exactly. Would you rewrite all the code with Copilot?

Thomas Dohmke: You can ask it to write unit tests, because nobody wrote unit tests in the ’60s and ’70s either, let alone that they were unit testing frameworks for those languages. Keep in mind the ’60s, that was before hard drives, right? Before we had personal computers. It was a very different world back then. And of course, those companies have done work to modernize to a certain degree, but it’s far behind agile software development. So you can support that transformation process today, but we’re not at that point where you can just click a button and you have it transformed.

The same is true for many more modern languages. There’s large PHP code bases, there’s lots of Java out there, and there’s lots of optimization that we can take to just make our existing stacks more efficient, where agents can help. There’s lots of things to improve in every part of the software development life cycle ecosystem around us. GitHub, you know, is—I like to think about GitHub as one planet in this universe of software development tools. And there’s smaller planets around us and equally or almost equal-sized planets in our space, and we consider them partners, and we are happy that they’re there.

Will GitHib build its own models?

Stephanie Zhan: Yeah. So interesting. Agents are—just to pull on that thread a little bit. It’s also an area of excitement for us. And I think, you know, you have a lot of the benefit of owning so much of the data and everything you can do on the post-training side. OpenAI itself is also getting better and better with each new model class, as are every other model company out there. Would you ever want to kind of invest into building your own agents, maybe from scratch by building your own agentic models, or just to partner with some of the others out there?

Thomas Dohmke: Me as GitHub CEO? Is that what you’re asking?

Stephanie Zhan: Yes.

Sonya Huang: So that’s great advice for startup founders, and I think really encouraging. What about for incumbents? I think you are such a beacon of hope for incumbent companies. I think Satya said in the last earnings call that GitHub is now growing 40 percent year over year thanks to the Copilot acceleration. And to your point earlier, you feel like you’re a young startup again, like, walking into the Sequoia offices. What advice do you have for—there’s so many incumbents that are trying to reinvent themselves around AI. What advice would you give those folks?

Thomas Dohmke: Satya actually said 45 percent, so we’re really excited about that. All up on our revenue, we’re going 45 percent. You know, I got an email from somebody last week who was looking at GitHub and losing the belief that large companies cannot move as fast as a startup can. And I think the key ingredient on this is radical focus, and basically focusing on a few small things. Just because you know you have 1,000 engineers or 50,000 engineers doesn’t mean you can do it all. That’s just a false assumption, and I think it’s in a way misleading.

As you manage larger teams, you’re losing kind of like the unit of team size that shows you how much you can actually get done and how much friction you have in the system if all these teams work on different things. So I think focus, you know, lots of nos on all the ideas that people have around you and customers. That’s, I think, the key piece to move really fast. And then obviously taking some strategic bets, and strategy means you’re thinking about it as how do I differentiate from others? What makes me specific—makes me special in this market? What lets me charge the prices I want to charge and not fight a race to the bottom? And I think that’s kind of like thinking about, okay, you know, in software we like to think about well, if I add just three more features then I’m going to win the space. Until you realize, well, everybody else can also add those three features, right? There’s almost nothing in GitHub itself as a platform that you couldn’t rebuild with significant investment and time. But I think it’s really hard to mimic our culture. It’s really hard to mimic, you know, our experience, our obsession about developers, and ultimately our focus and the way we’re approaching these things.

Sonya Huang: Yeah. I guess very tactically, do you recommend staffing a tiger team to get to the product market fit on AI? Do you recommend, like, you know, pulling half your engineers off whatever they’re working on and like, “Hey, you guys are the AI team now. We got to go big or go home.” Like, tactically, how do you recommend companies go about it?

Thomas Dohmke: Well, as in a company, you always have the challenge that you have to sustain whatever the business is you’re in, and so you can’t just pull everybody into a new topic unless you’re willing to disappoint all the existing customers. You know, no enterprise business has not made promises to enterprise customers of what’s coming next on their roadmap. Lots of conferences, including our own, is stuff they’re shipping right now and stuff we’re announcing for the next six months. And if you’re not delivering on this announcement, your customer base is not going to be happy. And so you can’t really make that drastic move right away.

And so we love this idea of a tiger team or an incubation team—we call it GitHub Next. And when we set that team up, we actually thought they’re going to work on projects that are, like, five years out or in that horizon-free space. And then it turned out well, it was more like six months ahead of the curve. The future comes really fast on us. And so as that future then comes fast, you really also need to move fast in handing things from the incubation team over to your mainline engineering team, and then it’s all about okay, so we are funding AI because we are seeing the traction there. And that means we’re leaving some of our previous bets in keep the lights on mode, KTLO, or saying goodbye to these ideas and shut them down. I think that’s the hard part of the strategic pivot. And that’s easier for a startup, at least it often looks easier from the outside when a startup pivots, right? Like, there’s lots of startups on the wall downstairs that have gone through that as well. But obviously, internally it’s also very hard. Emotionally you’re tied to ideas. You remember the workshop under sequoia trees where you had those ideas, and then six months later you’re realizing we never got to product market fit and the time is now to move to something else. So that is the same for large companies as for small companies, except that small companies have more of a forcing function to give up and move to something else.

Thomas Dohmke: Or me as angel investor?

Stephanie Zhan: You as a GitHub CEO.

Thomas Dohmke: Yeah, I think we’re probably doing both. We’re going to—in a way, we always have done both things as GitHub. We have invested into our own things, the thing that we consider as a core part of our platform and of our offering, you know, part of our primitives, and we have partners—partnered with companies. Just earlier this week we announced the partnership with JFrog, which covers binary artifacts, scanning of containers and those kind of things. And they are obviously very naturally in our space. You know, we have the source code. And what do you do with source code? Well, you either compile it into binaries or you combine it with binaries before you deploy it into the cloud. And so there’s a natural value chain there where we have things where we invested ourselves, things like releases that you see on GitHub and packages. We own NPM, the largest package registry in the world for the JavaScript ecosystem. And NuGet through our partners at Microsoft .NET, largest in the world. But that doesn’t mean that we cannot also partner with the JFrog to ultimately enable that secure software supply chain that I mentioned earlier and that I think is crucial. And that Figma, Vercel, so many other companies in our space that will play a part in the life cycle. And I don’t see a world where somebody covers all of that and can convince every developer that using all these tools is better than picking what analysts might call best of breed, the tool that I consider the best. Best is all subjective anyway, right? Does it meet the expectations of the developer in their environment?

Stephanie Zhan: Yeah, makes sense.

Rapid incubation at GitHub Next

Stephanie Zhan: The innovation and success that you’ve created around GitHub Next is amazing. How does a kernel of an idea start within GitHub Next? How do you then resource and invest into it? And then how do you—at what point, and what’s the process in which you decide whether or not something should be continued to invest in or shut down entirely?

Thomas Dohmke: So I think the start is always the employee, the Hubber, that has an idea, and we have lots of ideas in the company. We do hack weeks, hackathons, passion projects, whatever you want to call it, 20 percent time. And then the next team specifically, you know, there’s lots of demo meetings—demos and not memos. You know, it’s much more useful, just show a working prototype. It’s so easy these days, you know, to just use some design system React components or Figma and stitch something together even easier with Copilot. And then, you know, it’s strategic decisions that the leaders of these teams all the way up to me need to make.

You know, in many ways very similar to any other creative industry. You know, like at Disney or Pixar, they have to decide which movies to produce and which ones are probably not going to gain traction. And part of that is customer research, talking with developers. And that then keeps going as we go through the initial idea and the first prototype, and then going into a technical preview. And then the preview is all about that flywheel, that feedback loop with the people using it. And quickly you see, you know, are they trying it out and then they are churning? Or are they keeping the energy high and keeping the ideas flowing? It’s great in any project if people are just sending you more ideas and more feedback. And I think that’s then the decision whether we are keeping that project and making it a main project or whether we are deciding okay, the experiment is over and we learned a lot and we’re moving to the next one. And of course there’s commercial aspects as well.

The future of AI?

Stephanie Zhan: Maybe outside of the world of code generation and anything in your current purview, what else are you excited about in the world of AI in the span of one, five or ten years?

Thomas Dohmke: Oh, outside of developers. You cheated a little bit on me. I mean, I think in the context of one year within the space of developers and outside of the space of developers, I’m excited about agents. And we’re still very early in this journey. I think we have high expectations of what the agents could be, maybe too high expectations to some degree. And so things will probably take a little bit longer. But I think in the next year or so we’re going to see more of these helpers within the chat interface and out of the chat interface that will solve tasks for us.

You know, like, I look forward to that travel agent that often gets demoed by big companies to actually materialize, and I can just go into my chat interface and say, “You know, I want to have beach vacation over spring break,” and then figure out, you know, when spring break actually is, because that’s all on the internet. And you know, who I am, what my name is and what my family’s names are and their birth dates and their passport numbers. And so I don’t have to enter that into a cumbersome interface anymore. It shows me the price points, and probably we’re going to the same hotel as every year anyway. And so I think that those travel agents and those kind of agents are going to happen in the next year-plus.

I think five years is my natural language vision, and unlocking the world’s knowledge, including software developers, to everybody in any language, any human language. And maybe that’s even happening sooner. Ten years is hard. Ten years is so far away. But I think, you know, it’s the AI of things, you know, the material, like the mechanical world of AI. You know, so many things that we do in life, you know, you need to grab something or you need to push something and, you know, go and check into any hotel, there isn’t any AI involved once you get your room key. And maybe the elevator, you know, has some kind of AI, because you push a button what floor you want to go to, and then it tells you to take elevator C, and so you no longer push the button within the elevator. That’s an optimization problem in itself. But I think there’s so many physical things in life, and we are still not at really self-driving cars. We have a dishwasher, but I still have to put the dishes into the dishwasher and out of the dishwasher—it’s better with nine and twelve year olds than with two year olds. [laughs] And there’s so many other things in life where I think, whether it’s a robot or some other form of physical AI, is going to take over some of the things that we consider as chores.

Stephanie Zhan: We’re really excited to hear that. And I think we have a very similar view of what will happen one, five and ten years from now.

Sonya Huang: Who do you admire most in the world of AI?

Thomas Dohmke: I think I admire those that are building new stuff with AI and software, and those that have a dream of what they could build. I think those that obsess about a problem and not about the technology. We are talking a lot about AI, but at the end of the day, it’s what problem are you solving for the world? And, you know, there’s lots of biotech companies that use AI to try to cure diabetes or cancer. And I’m sure there’s companies trying to solve climate change with technology. And I think those builders, the founders, those that have big ideas and that can change the world, those are the ones I admire the most. And often when I meet with them in their offices or on calls, I’m like, “This is so cool.” And obviously, as GitHub, we’re enabling a small part of that. And so we feel really proud of being part of their journey, and we’re really excited about building more for them.

Advice for founders

Stephanie Zhan: I think you’re enabling a huge wave of new AI companies, especially those that are born open source. It’s amazing to see. What advice do you have for the founders listening in the audience who are building an AI today?

Thomas Dohmke: I mean, already Sonya you asked earlier, I think focus is everything. Like, it’s so easy to get lost in all the ideas that you can put on a whiteboard. And that’s the danger of a whiteboard that it has—covers so much space where you can put ideas. But focus ultimately is everything. Finding very quick market validation often in the developer space, that means growing product-led growth. Enterprise growth can come later. But from my experience, there’s nothing like a few excited developers spreading the word about your product, even though the big revenue number then later comes from big enterprises buying you. But the reverse is often much harder of going enterprise first. And you might find excited people there as well, but the feedback loop is just so different. So it’s focus, it’s trying to find that flywheel, the product market fit.

And then I think the other one is to think big. I think it’s easy to find small ideas on top of a model today that get commoditized tomorrow. And so you have to think forward and that, you know, ten years is a long period of time, but that’s a good period of time for a startup to build something that is actually meaningful in this world. And so you have to think big. You have to create a vision that may be much larger than the MVP, you know, the first thing, the prototype that you’re building right now. And I think that’s really hard as a founder to draft out that vision and then decompose, right? And go back to that small problem that you can solve right now.

Lightning round

Sonya Huang: So we’ll close with some rapid-fire questions. One word answers.

Thomas Dohmke: Oh, one word answers! You changed the rules. Okay, let’s go.

Sonya Huang: Okay. Will anyone meaningfully disrupt Nvidia in AI chips in the next, call it, five to ten years?

Thomas Dohmke: Yes.

Stephanie Zhan: In what year will we pass the 50 percent threshold for SWE-agent?

Thomas Dohmke: 2025.

Stephanie Zhan: And what about 90 percent?

Thomas Dohmke: 2028.

Stephanie Zhan: Wow.

Sonya Huang: Will GitHub promote …

Thomas Dohmke: That was too pessimistic, that last one.

Sonya Huang: [laughs] We’ll hold you to it. Will GitHub primarily be using open- or closed-source models in the next five years?

Thomas Dohmke: Both.

Stephanie Zhan: Where does the majority of value accrue in AI: models, compute, infra, applications?

Thomas Dohmke: Across the whole stack. That had a hyphen in it, so it was one word.

Stephanie Zhan: [laughs]

Sonya Huang: Is systems thinking and creativity going to get baked into the models in the next five years?

Thomas Dohmke: Maybe.

Stephanie Zhan: And will there be a new consensus architecture beyond the transformer in five years?

Thomas Dohmke: Of course, yes.

Stephanie Zhan: Wow!

Thomas Dohmke: Why would you think the other way around? It’s a much easier bet to think that there will be a new architecture because there was other architectures before transformers. And that doesn’t mean it replaces transformers. You know, your cell phone still has a CPU, even though GPUs are the hot commodity right now. So I think yeah, there will be new architectures and they might be bigger than transformers today.

Stephanie Zhan: So interesting. I mean, that would bring a lot of new oxygen for the builders in the ecosystem, and a lot of things that would have to get reworked, rebuilt and re-architected.

Thomas Dohmke: Yeah.

Stephanie Zhan: Amazing. Thomas, thank you so much for joining us today. It’s been wonderful digging into the history of GitHub, the birth of GitHub Copilot, and the ambition that you have going forward as well. Thank you.

Thomas Dohmke: Yeah. Thank you so much for having me. It was so fun to talk to you both.

Stephanie Zhan: Likewise. Thank you.

Mentioned in this episode

Mentioned in this episode:

• Nat Friedman: Former Microsoft VP (and now investor) who came up with the idea that Microsoft should buy GitHub
• Oege de Moor: Github developer (and now founder of XBOW) who came up with the idea of using GPT-3 for code and went on to create Copilot
• Alex Graveley: principal engineer and Chief Architect for Copilot (now CEO of Minion.ai) who came up with the name Copilot (because his boss, Nat Firedman, is an amateur pilot)
• Productivity Assessment of Neural Code Completion: Original GitHub research paper on the impact of Copilot on Developer productivity
• Escaping a room in Minecraft with an AI-powered NPC: Recent Minecraft AI assistant demo from Microsoft
• With AI, anyone can be a coder now: TED2024 talk by Thomas Dohmke
• JFrog: The software supply chain platform that GitHub just partnered with