Computers are getting smarter, Silicon Valley's getting weirder, and AI is taking over. At least that's what everyone keeps telling us. Welcome to Machine Gods, a new NPR podcast that will be your field guide to the AI-powered future. We're veteran tech journalists with front-row seats to the industry, and twice a week we'll make sure you get smarter about AI, so you know what's real, what's hype, and what might come next. Listen to Machine Gods from NPR on the NPR app, or wherever you get your podcasts.
Support for the show comes from ServiceNow. AI is moving fast across the enterprise, but without visibility, it's just chaos. Different tools, different models, different teams using AI in completely different ways. ServiceNow turns that chaos into control. With the AI control tower, you see all your AI across the business in one place. What it's doing, what it's done, and what it's about to do. So you stay in control. To put AI to work for people, visit servicenow.com.
Support for the show comes from CrowdStrike. It's not a huge stretch to say that AI is the next major computing platform. Every platform shift changes the way we build software, and it changes security too. CrowdStrike is defining cybersecurity in the AI era with AI detection and response, or AIDR, a solution for seeing, monitoring, and securing AI across the enterprise. Learn more about how leading companies are turning to CrowdStrike to secure AI and secure their business at crowdstrike.com slash decoder.
Welcome to Decoder. This is Alex Heath, your Thursday episode guest host and deputy editor at The Verge. It's AI coding week here at Decoder. You just heard my friend Casey Newton's interview with the CEO behind Cursor, Michael Terrell. And now I've got a conversation with GitHub CEO, Thomas Dumka. In many ways, GitHub co-pilot set off the current AI coding boom we're all living in.
But since Thomas was on the show a year ago, the rise of vibe coding has shifted the buzz to newer platforms like Cursor and Windsurf. As you'll hear in our conversation, Thomas is thinking a lot about the competition and GitHub's role in the future of software development. It's a good time to be having this conversation. Thomas has a bird's eye view of where the industry is headed. I wanted to know how he thinks the role of a software engineer will keep changing with AI and when I'll feel comfortable vibe coding myself. This is Decoder after all, so I couldn't have him on the show without asking what life is like these days running GitHub as its own company within Microsoft. Okay, GitHub CEO Thomas Dumka, here we go.
Thomas, thanks so much for being on the show. I have a lot to catch up on with you since a year ago when you were last on Decoder with Neelai. First, though, I think since last year, the biggest shift has been this rise of Vibe Coding that we're talking about nonstop at The Verge and on Decoder. And when I think about Vibe Coding, as someone who has never engineered any software in my life, is the moment where I'll be able to create an app or a website without needing to understand coding at all. That to me is very exciting, but it doesn't feel like we're...
quite there yet. Do you have a sense of when the industry will get to that point where someone like me can literally vibe code something from scratch? It all depends what you're going to build, right? Like, I think we are there for small things. I challenge you to use something like GitHub Spark or Lovable, and you can probably build like a snake game or Pong or like little scoped apps. And I think we're already at that point. And then as you get more complex, you're right, it requires you to have certain systems understanding, you know, architecture design, databases, and all those kind of things.
I think we're getting closer and closer to a point where you get more of these, as we call them, primitives, supported by the web coding platform. So if you take something like GitHub Spark, and when we launched that at GitHub Universe last October, it was only creating front-end applications in your browser with a language called JavaScript. So you could do things that work in the browser, and it had a little bit of back-end storage. Now the version of GitHub Spark actually generates a full-stack application, so it also has a back-end and a database, and it can connect to AI models and those kind of things. And so it got more and more complex over, like, what, nine months or so? And I think we will see exponential improvements to what these web coding platforms can generate.
The challenge, I think, will remain that it's one thing to create something from scratch. It's a totally different thing to take an existing software system and modify that and figure out where on the code base, which functionality, and which test case is. And as most software applications over the years become bigger, bigger, and more complex, as sometimes hundreds or thousands of engineers have worked on that, I think we're really far away from AI being able to solve all these use cases that an expert in a system can solve. That's a good segue to this blog post you recently published where you spoke to a bunch of developers. I'd be curious to know what kind of developers they were, but you interviewed a bunch of developers and kind of took their temperature on the state of AI coding and where they're at with it and what they see coming. What kind of developers did you talk to? These weren't all just people on your team, right? These were external developers? Yeah, it's external developers. Yeah. When I joined GitHub seven years ago, what's the first eye-opening moment is that when you're at GitHub, given the GitHub brand and then how widely known GitHub is in the industry, it's really easy to find people that are willing to...
jump on a 30-minute interview and answer some questions. So it was a range of developers that are certainly affiliated with GitHub or have an interest to provide feedback to us, but otherwise it was a spectrum of folks. What was the top feedback you got? And I'd like to connect that too to the role of the software engineer and how it's changing and what you basically heard from people about how they think the role of being an engineer is changing right now and then how it's going to change over the next year. If I want to summarize it really quickly, AI is here to stay. And I think most developers that we interviewed and most developers in the industry, they have realized that the profession of a software developer is going to change. It's already changing through the use of AI.
And as such, the majority of them are in the process of adjusting to that new way of working. I had another blog post earlier this year where I described this as the odyssey of developers. We moved from punch cards and mainframes to the personal computer, which drastically changed the inner loop of how fast I can dry out something. Because before that, I had to book time on a mainframe. And I remember going to my mom's office, she had punch cards on her desk. And then all of a sudden, everybody had a PC on their desk, at least the nerds had them. And they could just start coding. And then we went from assembly language and basic and Pascal to Python, Go and Ruby. We went from no internet to millions of open source libraries that I can just pull into my application from my own servers to the cloud. They've always moved up the abstraction ladder. And today, a developer that builds a full stack application...
likely does not actually know what CPU is running that application and RAM and gigahertz and what have you, the technical specification of that CPU is. And unless they're a game developer, they're likely not doing assembly either. So we're moving naturally up that stack because the amount of code that we're managing has grown exponentially over the last 50 years and will keep growing exponentially. Humans already re-backed on exponential curves. And so we need these tools to, at least for us, flatten the curve, not make it flat, but make it a little bit less steep so we can actually still handle these complex applications with now billions of lines of code. Half of the developers you spoke to for this post said that they believe that
within two years, 90% of all code will be written by AI, which is a thing I've heard from others. Dario Amadei, the CEO of Anthropic, he recently had an even more aggressive timeline. He said he thinks we'll be there in three to six months. I'm curious, where do you fall on that debate? Where do you see the timelines for a world where pretty much all code in production was at least generated in part or fully by AI?
The first thing that always comes to my mind when we have this conversation is that we need to acknowledge that 90% of our stack today is already written by somebody else, aka millions of open source developers around the world. If you look at the virtual webpage and the backend and all that, I'm sure it's 90% open source. If you count the lines of code, 90% of those are from open source libraries, open source operating system, programming languages, dependencies, all that. And the team, the engineering team is writing 10%. Same for us at GitHub and even the same for our friends at Microsoft. So we're already in a world where focusing on the thin layer on top that defines what makes our business while we're pulling in the work of millions of developers.
from around the world to accelerate and to innovate and what have you, to secure the ecosystem. And the same will happen with AI, that AI will write nine pieces of code, so I can still focus on my eight-hour workday and have my one piece of code. So it doesn't mean that I stop writing code. It just means that now I have 10 times as much code and 10 times as much functionality and features as I could produce on my own. It's an amplifier in the same way that open source is an amplifier. On timeline, I think we're going to see a huge spectrum of teams and companies, and depending on how they adopt AI, not only because they're willing to use AI, but also because...
They have designed their company, they have designed their software, their infrastructure in such a way that AI agents can leverage that infrastructure, right? Like, for example, they have a design system. So whenever you want to ask the agent to add a feature, it doesn't invent new cascading style sheets and what have you. Instead, it uses a component so all the features the agent builds actually look aligned with the design language of your product. And so you can imagine there's a whole new way of architecture and applications that make it much easier for an agent to use those Lego blocks to compose new features. And I think that will heavily influence how fast the company can get to the point that 90% of code is written.
And then vice versa, there's obviously lots of companies still that have COBOL code running on mainframes and that have PHP and Perl and all the legacy spaghetti code that sits around. Here at GitHub, we have a Ruby on Rails monolith that is now as old as GitHub is. And it's not that I can just make the decision as CEO to say, split down the monolith and we will have a nice modern architecture, let alone that often is a philosophical discussion as much as it is an engineering discussion. But getting out of this legacy state, going through the cloud transformation, digital transformation, now the AI transformation, that will take organizations way longer. than those that can be on the cutting edge where 90 or even more percent of the code is written by AI. So it sounds like you don't have a timeline. Like today, lots of folks already out there that are documenting that on their blog posts. If you know what you're doing and you started from scratch with something like Cloud Code, you can have it write 100% of the code. And the only thing the author was doing was modifying these in-human language written files to then instruct Cloud Code to build their applications.
And so that application is, by definition, 100% of the code is RhythmAI, and all the author is still writing is instruction specification. And whenever the source code that was generated was not functional, you, of course, can just go in and modify that. Or you can say, no, no, I don't want to do that because I'm breaking my system. Instead, I'm trying to figure out how do I change my instruction files to then keep going with cloud code writing on my code. Yeah, I'm kind of mad at myself, honestly, for even asking you about this, because this trend that's happening of CEOs saying, you know, random percent of our code is made by AI actually really frustrates me. Your boss, Satya Nadella, has said actually that 20 to 30 percent of your code at Microsoft is written by AI. And, you know, Sundar Pichai has put out a stat, Mark Benioff has put out a stat, Zuck has put out a stat. But I've been wondering, is this actually good code? Is this code that is something you want over time to increase? Because I was looking at this recent Stack Overflow survey where they interviewed a bunch of developers. I'm sure you've seen it. And it was really interesting because the vast majority of them were saying that they either
used or plan to use AI tools over the next year. It was like over 80%. But I thought something really interesting was that about half of them said they really distrust the accuracy of these AI coding tools and that their biggest frustration, which was cited by I think 66% of the survey, was that the code that it spits out is just not quite right, which often leads to them having to spend a lot of time debugging and that AI coding is actually more time consuming than productivity gaining. And I'm wondering, do you see that? across everything that GitHub does and also inside GitHub? Or do you think this is more isolated? It seems like a big problem that, yes, AI code output is on the rise, but maybe it's not actually good code. And now you're mad that I didn't give you a percentage? Well, I was mad that I was actually just asking you for percentages when I'm looking at my notes here and going, well, actually, people take issue with these percentage questions. Is this even the right way to be talking about this? Because, yes, AI code output is going up across companies and CEOs are using it to brag about their efforts. But are we focusing on good code? I'm curious what you make of that debate.
Those two, from my perspective, actually go together. And I think the percentage, obviously, is a bit of a marketing instrument. And what really does it matter if it's 90%, 80%, or 95%, whatever. I think what matters is that it's clear that the profession of software development is changing and that we're moving up the stack and we're moving from understanding every single line of code to switching back and forth between a specification. Human language is inherently non-deterministic. We can both say the same sentence and mean different things, let alone that different languages have different constructs of describing these things. Human language is an abstraction of the processor, right, of the CPU or GPU, what have you.
Nothing else than describing how the transistors ultimately are flipping from zero to one. And we're going to have both because the machines are running on CPUs. That's not going to change. But we're thinking in human language. So then the question really is, do we want to write more human language or do we want to write more code? And that's not AI determines that for you. That's a decision that you make as a developer. And I think this is going to be the creative freedom for me as a developer to say, I want to write code because I know what I'm doing in this moment. Or I want to use an AI agent. And in fact, if you think then about agents, the first question really is, is it faster to use an agent or is it faster to do it myself?
And if I know I can do it myself in like three seconds, I'm wasting time, energy, you know, resources. If I ask the agent to do that, it's like, you know, I hope when they're texting, you know, using JGPT to write the response to my text message. I have many reactions like this when I'm using consumer agents where it's like, yes, you can go off and book me the hotel or whatever, but it takes you an hour and it's not really right at the end. I should just go do it myself. It's more productive for me to do it myself.
And even the notification summaries in Mail or Outlook or Apple iMessage are often, for me at least, less helpful than just reading the first three lines of text, right? Yeah. Then the second question is, okay, so now AI is writing all this code. And if AI is writing 90% of the code, as in nine times as much code as I'm writing, am I reviewing all that code for quality, security, coding standards? or do I find ways for AI to help me with that as well? Because ultimately, we're going to DDoS, the human developer population, with the agent population as those agents, sorry, DDoS, distributed denial of service attack. As these agents don't sleep, they don't take time off over the weekend. And you can't run them in parallel. If you look at OpenAI Codex, they're already generating four variants of every task that you're doing. And then you're effectively helping them with their reinforcement learning by picking the one that you like most. But you could not only generate four, you could generate 40 or 400.
And then you could have another agent that looks at the quality of that code and rates that quality. And then you have a feedback loop between those agents. So actually, I think we're getting to the point that agents will generate always better quality than a human can generate because you can just run this at infinite scale. And as such, you know, find all the bugs that a human wouldn't find because they don't have the time to do that. But the challenge will remain is do we trust that code? And there, you know, they're actually coming back to GitHub. GitHub was built exactly for that reason. It was built for human to human collaboration, right? GitHub is all about me having an open source project, you fucking that project, you like something or you want to add something, you send me back what we call a pull request. And then...
It doesn't just go into my code base. Instead, I review that code and maybe give you feedback and say, hey, Alex, this is cool, but how about you rewrite this? And sometimes the tone is a little bit harsher on that as well. And then at some point, we decide, okay, this is cool, let's merge it. That is actually quality control of human-to-human collaboration. And of course, you can use the exact same process for human-to-agent collaboration. Except if you now have thousands of agents, you have to rethink that approach. And I think that's going to be the biggest differentiator for developer tools. Those who figure out how we can have agents generate so much more code than humans and have the humans still be in control and make sure that what goes into production and handles customer data and billing processes and what have you.
that that code is actually functional and has quality and doesn't have any vulnerabilities, right? That is the challenge of the years to come, is not how much code is agents writing, is how much code from agents can I actually accept into my production system. It sounds like you're saying the future of a developer is someone who is just looking at how all these agents are behaving all day long and accepting pull requests, rejecting them, or reviewing code, but not really writing it, maybe giving very high-level instructions, but really stepping back from the actual process of coding. Is that what you're saying is going to happen broadly? I think to some degree this will happen. We started this with wipe coding, and wipe coding is already happening. All you do is write a prompt. It's actually also happening with image models. If you use HTTP, you create a Studio Ghibli image, for example. You're not writing the code to do that, and you're just giving the instructions, and then there's something in the model that does that for you. And in fact, one of the hacks is that you can upload an existing image and have it decompose the image into a JSON file, like a data structure, and then you modify the data structure and use that as the prompt to generate an image that has something removed or something modified on that image, right? So you're kind of like...
acting like a developer, but you're not writing any code. You're just having the model do that. So I think there is going to be a class of developers that will use models and agents to build and verify systems. And I think there's going to be a class of developers that also still love coding and that are mixing up their workday with offloading some work to AI agents, like writing test cases, documentation, offloading the stuff that they don't want to do, so they have time for the stuff they actually love doing. Because the question then becomes, how do you spend your eight hours a day? What is it actually that lets you explore your creativity and ship innovation and so on, versus just getting stuff done? And I think it needs to be a mix of both. Most developers today in small and large companies code, you know,
maybe four hours a day, sometimes even less. The rest of the day, they do all kinds of other things. They're on meetings and they write emails and review other people's code and update servers. There's lots of other things that developers also have to do. I think most of them will preserve a certain amount of time during the day where they write code or instructions to generate code, which will feel like coding. In fact, if you think about it, I brought this example earlier. We moved from assembly language to higher programming languages and open source libraries. I haven't seen many people that are sad that they...
no longer have to write assembly. They can. It's not like assembly isn't there anymore. It's just that we have a compiler that takes my programming languages and compiles it down into assembly. Now you could see the AI model effectively like a compiler. It takes my human language instruction and compiles it down into programming language and then the real compiler takes that into assembly. Your code actually would be the human language, right? Like you wouldn't ever have to look at the programming language. Except that all the models today still have hallucinations and they still write code that doesn't exactly do what I wanted. Or maybe, you know, I do it, ask it 10 times and each of the 10 times it writes different version of their code. And you can actually see that when you watch people write coding live on stage. The snake game that I have used as a demo.
It looks different every single time. And so we don't trust the human language conversion into programming language enough to actually cut out that layer. But we were able to cut out the assembly layer and we're good about this. And I think as long as we have those two layers, we're always going to also write code and learn coding and go back and forth between the deterministic and the non-deterministic layer. Because LLMs are inherently non-deterministic and do make things up, it's actually a function, not a bug. It's how they're architected. They're designed to hallucinate. Is there ever going to be a point where I can get to what I was starting this conversation with, where someone like me who has no understanding of coding, who can't tell you the difference between languages, can...
reliably trust that something I'm building that requires access to different APIs, a database, web, all that can be built through natural language. Because if we ever get there, that seems big. That seems like it will have profound impacts on the world. But like you said, maybe it's just inherently not in this technology to be able to get there. We'll get there as long as the scope is, you know, limited to a certain degree and that scope will increase, I think. I think you're going to be able to do certain things. If you will, if you talk today to an AI agent and let it do, you know, render a chart for you, what that actually does is write a Python script and then use that Python script to render the chart. And you can, you know, go and say, show me that Python script.
And I think most users never need to understand that Python code to look at the chart. And those use cases, I think, will increase at a rapid pace. But then also what we consider as professional software systems, the products that professional software developers write will become so much more complex. Because it's now easier for professional software developers to build more functionality in the same amount of time where in the past they would have built that Python script to render a chart. That makes sense. What we consider now today state-of-the-art software will...
be so much more complex than what we can imagine a human developer can write. And in fact, you know, that has always been true also with what my iPhone can do today is unbelievably more powerful than what a Commodore 64 could do in the early 1990s, right? And that developer that was building for the Commodore 64 would look at the way of how iPhone apps are written and think about that as magic in the same way that we see AI as magic. And I think if you think about it that way, we will get to the point, but we will also get at the same time have the professional developer feel still like a magician because they can write such complex applications. We'll be right back after this short break. We're back.
Let's talk about competition. You've actually already mentioned several competitors to GitHub Copilot in this conversation. So that tells me you're open to talking about it. Coding, I think, is safe to say the fiercest competitive corner of the AI world right now. At least that's my read of it. You've got a lot of deals happening, startups being sold. Some of the fastest growing companies of all time, in fact, are AI coding companies. Cursor, Lovable, the Vibe Coding app we've talked about already in this conversation.
GitHub Copilot was really the first to the space in terms of being an AI coding assistant and still huge, obviously. But I'm curious, especially in the year that's lapsed since you've been on Decoder last, I'm wondering if you agree with the sentiment that GitHub Copilot seems to have lost mind share, at least in the Silicon Valley niche that is obsessed with this stuff and talking about it every day. I'm sure you still have a ton of users, but I'm curious if you feel that and if you feel like that's a pressure on you to evolve the product faster. And yeah, how you look at how competition has changed this year. I think the overall market of artificial intelligence, large language models and what is possible with these models is moving at a really rapid pace. I've never seen anything like that in tech and I've been around for a while. And you're right, AI code generation is at the forefront of that innovation. Logically, if you think about it, because all these companies are also employing developers. And so there's an inherent motivation for everybody to make coding easier because that means I can innovate faster. And if I have a coding tool that lets my developers move 5% faster than my competitor, I will use that. Since I was on last time at Decoder, that market has accelerated. And I want to be...
really open to this. This is amazing. I've been a developer for so long and I've never seen so much innovation in software development tools. And as GitHub, you've always understood us as a part of an ecosystem that is both competing and partnering with companies in our space. Whether it's in AI code generation, whether it's in CICD, like how applications are built, they're not making decisions of what open source library you prefer. Most people wouldn't use GitHub if it would only offer you the JavaScript ecosystem or tell you to go somewhere else to use Python. And nobody wants to watch sports if there's only one team winning at all.
Where I don't agree with you is that we lost Mindshare. I think we have won some and lost others. And sometimes it's on a week-by-week basis that you see folks posting on their X that now they're back on VS Code because the latest release and we're doing nightly builds and monthly releases has something they want. That's something we should, as an industry, be really proud of. Copilot was late on moving from a single model to multi-model choice, but we got there last year in October. We added agentic support. We built an MCP, a model context protocol integration. And now all these coding agents are using the GitHub MCP server that we have built and partnered with the topic. So it's a race. It's a race where we just announced in Microsoft earnings that we now have 20 million users on GitHub Copilot, 75% quarter over quarter growth in enterprise usage, and 90% of the Fortune 100. I can give you all the stats. Hold on, on the user number. I got to ask you about this user number. Yeah.
Thomas, there's something funny going on with this user number. This is lifetime users? This is users that have ever used it? Or is this monthly? Is this daily? Can you give me a little more granular detail here? It's 20 million users that have, 20 million GitHub users that have Copilot enabled. So it matches the 150 million accounts number, same way of measuring it. And we have 50 million users on 50, 50 million users on Visual Studio Code. So that gives you an idea of how far we are in...
in activating the users on all IDs. This post of yours where you announced that new user number stat, you wrote, the true measure of a company is never drawn during its hype wave, but by its resilience when pressure tested. And then you wrote something even more interesting, I thought, when you said, even with all the constraints we faced, we've proved that a little grit wins the game. What constraints were you referring to there in that post? Because you're the CEO of GitHub. I would think that you guys are not really constrained. You're also model agnostic, whereas the rest of Microsoft seems much more still connected to OpenAI, at least for now. But it seems like you're able to work with whatever model you want. You're still a separate part of Microsoft relative to the other parts. So what constraints were you talking about?
Any size of company is constrained, right? That's the definition of budgets and employees or headcount, what have you. And you can just always add more people and then you run into the mythical person once that adding more people to a team actually slows you down and doesn't accelerate you. It's such, you know, whether you're a 10-person startup or a 3,000-person GitHub or, you know, 200,000 or so person Microsoft, you have to make decisions. And Apple famously said, you know, a thousand knows for every yes. And so one constraint is that, of course, we have endless...
you know, number of items in our backlog. And many of the backlog items that Git has stem from a time before AI. You know, customer feedback, people wanting a feature in GitHub issues or GitHub projects and you name it. In fact, you know... there's a running joke that for every customer feedback item we get, there's already one there from months or years ago where somebody else had that same idea, right? Like, naturally, because our customer base are developers and we're developers, there's always going to be somebody having an idea that was already there or telling us how to do something better. And then, you know, we have a group on X that says, you know, well, that was a stupid decision. We have to make a default and another is the best decision ever. And so we're living in that world where we have way too many input signals and we got to, the constraints, the grid is to pick the right ones. Where you're really moving fast and where nobody in the industry is actually knowing.
what AI cogeneration or AI in general looks like in two or three years. I think that if we're all honest to ourselves, that is true. We're like, if ChatGPT was the Netscape moment of the 2020s, then we are like in 1995 now, where we have an Amazon and a Google, but we have not seen a Facebook, not seen a Shopify, and not seen the iPhone, right? Like in 1995, it would have been preposterous to predict that those things were happening, and you actually have the confidence that that is true. And I think that's the world that GitHub has been living in ever since its founding, is it was ahead of its time leveraging Git to build what nobody called DevOps back then, but what is ultimately now the largest developer platform.
And then it was swinging back and forth of what is the next big innovation and what do others in the industry are doing? And where do they actually prefer to partner instead of building it themselves? And I think with Copilot, we have navigated that well. And we had our moments where this wasn't, you know, working and strategic decision that ultimately our CEO made wasn't the one that pulled us in the right direction to win this space. And now I'd say in August 2025, we're still ahead of the curve and we're still the leader of the market. And as such, we're really proud and happy.
about the fruits of the hard work that we're earning. Speaking of partnering or not, coding has become a thing that affects huge transactions that people read about as a released AI. And I'm thinking about Google and Windsurf. Windsurf was this very ascendant AI coding tool. And they were going to sell to OpenAI. That deal got called off. And that's where, you know, Microsoft comes in here. And I know you're GitHub, but it's all kind of related. And I'm curious, you know, a big thing with that and why Google ultimately ended up hiring that team from Windsurf is that the company wasn't able to sell to OpenAI because there was a concern that the IP from Windsurf was not going to be shielded from Microsoft, that there were competitive dynamics at play given the Microsoft OpenAI connection. So I'm wondering why was that such an important issue that it actually ended up killing a huge deal that would have been a huge deal for not just the AI coding space, but Silicon Valley in general, this would have been a multi-billion dollar sale to OpenAI. And I'm curious why.
From where you said, I know you probably weren't driving that, but why? Why was that such an important issue? You have to ask the question and I can refuse to answer the question. No, look, you know, obviously I have a conflict of interest even involved in these conversations as such. I don't have any background knowledge I can share with you on this. But that actually shows you something else, which is the fear that developers are soon going to be replaced by AI is actually not justified. In fact, the opposite is true. And teams like the founders and core team behind Winsurf or, you know, other parts in the industry, you know, meta comes to mind, get offers that match.
professional basketball players. Some of them even have agents helping them to lend a new contract. And that should be something that we should be excited about. You know, how we get to that point is a good question to explore for somebody who has access to all the documents. I don't. But I think, you know, as a developer that happened in this industry, building developer tools and using them myself.
That's freaking exciting because it shows you that the market value of those at the top of their game keeps growing. And it hopefully also motivates kids, teenagers, those that are often first into gaming and then want to build their own stuff, still get into a profession. Because I do believe that we will need more developers than ever if this all plays out the way I hope it does. Let's turn to Cursor specifically, which I think has really taken the mantle as, at least in the Silicon Valley, really early adopter set, like the hottest AI coding tool right now. What have they cracked? What was it that they saw, the insight they had?
that you think got them such a quick lead in terms of just mindshare awareness? I know GitHub Copilot is still bigger, but they're growing very fast. And the CEO was actually on Dakota recently with Casey Newton, and he cited you all as inspiration for them starting the company. But what did they crack that you didn't see, and how are you responding? What cursor cracked was to realize that it's not just about adding AI into the IDE, but about changing the IDE itself and designing what you might call AI-native workflows. I was really like thinking about how will developers work if AI is the default and not an add-on. They were the first, you know, was Cloud Sonnet 3.5. So they were also, you know, ahead of the curve of saying, okay, multiple models play a role and let's give developers the choice and pick the model that works best for them instead of us making the choice, which was our philosophy, you know, a year ago when I talked with Nilay.
That's two different philosophies, right? Like we ran lots of Evo suites, we looked at these models, and then we picked the one that really was best. But that's not a binary choice because each programming language and each test case and generating functional code or testing code and so on has a different score. And so then you have different scores for different hosts and then you make a judgment call, okay, this is overall a better model, right? Well, but we know from the history of developer tools, in fact, we at GitHub know this very well, is that the best thing you can always offer to a developer is just and have them pick the model that they think is right. And then they will know from the experience or you may call it craft which to pick. And I think that has...
you know, change the market. Today, you cannot compete in AI coding if you're not offering multiple models, if you're not having the best models, quote-unquote, the best, perceived by the majority of developers, if you don't offer them, you know, to bring their own models, if you don't have an agent node that runs within the IDE, if you can't, you know, take that agent node and offload it into the cloud, what we call the copilot coding agent, that is actually integrated into the data platform. And so... You know, I'm a big Formula One fan, so I like to always think about the world like that. You know, there's teams that win whole seasons and the next season they're behind. And they're not taking that as in, well, shit, we lost and we'll never win again. But they're taking that as an inspiration to say, OK, we need to rethink of how we're doing these things. And I think that's what happened to Copilot, you know, in mid-late 2024. And, you know, I think now where we are in August 2025, we can win races again. Not every race, and neither has the competition been every race. But we certainly have innovated in AI cogeneration ourselves, while also, you know, caught up to what has made others faster.
This space does seem so dynamic because of that point you brought up, which is that it is so dependent on the models. Whatever tool is currently leading the models they're using, their way that they price for those models. And the models themselves, who is leading, changes a lot. It's almost impossible even being in it to keep up with all of these model releases. Lama 3 was a good model. Lama 4, not really as much. Anthropic right now seems to be the king of coding. Maybe that changes. Maybe it becomes OpenAI in a few months. How does that change the app layer, right? So I appreciate how quickly the space is changing. So I think to say that you all are for sure lost or cursor is for sure won is way too early. But as I think about Anthropic, and you've done a lot of work with them. You mentioned the MCP server. And obviously, you can access Cloud through GitHub Copilot. A thing I've been asking people is, why is Anthropic so good at coding?
what is this secret sauce they have? You know, Dario was on the big technology podcast recently and he did not want to talk about why they're so good at coding. But it's the billion, I mean, God, it's maybe the trillion dollar question is why is it such a good model at that? I'd be curious as someone who works with them closely, why do you think they've gotten such a lead in coding specifically? Let me, you know, for a second, go back to the first part of your question and I come to why Anthropic is good in a minute.
I think, A, in tech, we have this notion that for somebody to win, somebody else has to lose. And that has not been true for Windows and Mac OS, that has not been true for iPhone and Android, and many other technologies. In fact, in developer tools, that never has been true, because if so, we wouldn't have the dozens and dozens of programming languages today, and everybody would have moved to what is perceived as the best one. And if you look at the competitors of GitHub Copilot, like Cursor, Lovable, Windsurf, Old, Vercel, et cetera, well, where do those people that use these tools actually store their source code on GitHub? Where do they manage their issues and projects on GitHub? Where do they run their CSED on GitHub? You know where many of those competitors actually run their model inference? On Azure, right? And so as such, we are part of an ecosystem. As Microsoft, we have always been a platform company. GitHub has always been a platform company. And so we can both be competing and we can benefiting from these companies, driving the size of the overall software ecosystem. Because let's face it, software development has...
only ever grown. It hasn't really detracted. It's not getting smaller. The number of developers are getting bigger. The value generated from, look at the tech companies on the stock exchange on any given day. They're often outperforming the market.
If our competitors are winning, we are winning too. And I think that's the mindset that is really important to keep in mind when we talk about these battles. That's the benefit of being a hyperscaler, Thomas, is that when the competitors are winning, you're winning too. Not everyone can say that. Sure. That's called differentiation. And then on the topic, specifically Claude Sonnet, where Claude Sonnet, if you believe YouTubers and bloggers and experts like Simon Wilson and others, where really it outcompetes other models is tool use, like the ability of the model to pick the right tool for the next step of the agent mode. And so that shows you it's no longer just about the...
training in programming languages code and whatnot. It's also training in what are the tools that the agent should use in some fashion as a developer should use. And if that fails in a step, and you see that with some other models where the tool usage is just not as good as Cloud Sonnet, then your agent mode just falls apart because it can go through the chain of thought and actually imagine it needs to install an NPM package like a JavaScript dependency. If that doesn't work, then the agent can have the best training set in the world. If it cannot achieve that one thing, the whole thing no longer works. And I think that's where...
Anthropic had an early insight or better evolved suites, better testing that led them to have this head start and others are trying to catch up. But, you know, the funny thing is we're recording this and then, you know, by the time it actually gets published, maybe that road has already changed again. That's the fun part of that. Yeah, I think you're maybe alluding to GPT-5, which may be coming when this episode comes out. Audience's hands are in the air. We need to take another short break.
Welcome back. Right before the break, Thomas and I were talking about GitHub's competition. That gave me the perfect window to ask about how things are going with Microsoft and OpenAI. We've touched on pretty much all the main competitors. I do want to end just with, I know you can't say much because this is all being discussed, but Microsoft is currently talking with OpenAI about what the next kind of chapter of the corporate relationship looks like because OpenAI is moving from a nonprofit to a for-profit and Microsoft is obviously a huge shareholder and currently has exclusive access to the IP.
I'm curious, as a CEO of GitHub, how would you like to see the OpenAI-Microsoft relationship continue? What do you think is important for where you sit for that relationship, and what would you like to see? For GitHub and for all developers, it's always going to be important that it's a healthy relationship, one where both sides are partnering with each other. That's how the original co-pilot came to be, for a partnership between OpenAI, who had developed GP3, and then the Codex model, Microsoft, who had not only the hyperscaler infrastructure, but also years of experience in responsible AI, scaling these processes. Because look, if you think about the original co-pilot, it was a masterpiece on its own, because on every single keystroke, we're running inference against the large language model. That was unheard of at that point in time, and at that scale of millions of developers in a very short amount of time.
And so that, you know, innovation between GitHub, Microsoft, and OpenAI created this market that we're now in and that has all these competitors and probably another, you know, dozen to come in the next month. And I think that partnership going forward will create, you know, new such innovations through new models and through new ways of running these models more efficiently. I mean, a lot has been written about energy consumption of AI and cost and gross margins and those kind of things. And through GitHub, which always has put the developer first. That's what we are, you know, breathing. It has always been about human-to-human collaboration and we're moving into human-to-agent collaboration.
and more likely into agent-to-agent collaboration. And all these three layers will stack on top of each other. And OpenAI has, you know, the largest, with ChatGPT, the largest AI platform today. And so partnering with them, making sure the content that developers generate in ChatGPT with the OpenAI codecs complies with the same, because the control system, the same security compliance, you know, enterprise company standards, is going to be crucial for developers, you know, to adopt these tools and integrate them into their workflows. And I think the ultimate battle in AI is going to be fought at the platform layer. Because if all agents are insanely powerful...
then what you really care about is how can you integrate those into your other business processes? And for us, obviously, that's the developer lifecycle. Makes sense. Well, let's end on some GitHub Microsoft-specific questions. It's been a year since your last annual revenue run rate stat. I believe it was $2 billion. Why no update? What's the current number? Investor Relations team has made the decision what's being exposed in earnings and what's not. So as per usual, I can only tell you what's in the earnings script. And this time, we made the decision to not publish the number. Okay. I'm assuming it's higher than that.
I'm assuming it's higher than $2 billion. We have published other impressive numbers, and we're really happy about the Q4 results of over Microsoft. Neil, I tried this one last year, and I'm expecting to get a similar response. But a big question that we have is, does Copilot make money for GitHub? Does that even matter? Is that even a consideration? Because you're running effectively your own P&L to a degree as a separate org within Microsoft. Is this actually making money, or is it a cost center right now to gain market share? You're right. GitHub has its own P&L, and as such, we're measured against the goals of that P&L.
You know, I think in general, if you look at Microsoft as a business, there's always going to be the desire for these units to make money. The timeframe is different. And often, you know, the bets are really big. And as such, I think the better question is, is it going to make money over the lifetime of the customer contract, so-called LTV? And on that, I'm very confident that Microsoft shareholders and Microsoft's leadership team is going to be happy about GitHub's business results. Look, you know, we bought GitHub in 2018. So now, you know, a little bit over seven years ago for $7.5 billion.
at a time when the last published revenue number was 200 million a year before that. And you mentioned last year we published 2 billion. So the revenue has 10 times increased in seven years. And, you know, it's safe to assume that, you know, that number keeps growing. You mentioned the business numbers from Cursor. Given that, you know, we are very open about that we believe Copilot is bigger than Cursor, you can, you know...
forecast what our number might be. And so such, you know, we are really, really happy as Microsoft about this acquisition. I personally put it into the top three of deals that Microsoft has ever made. I think, you know, Satya Nadella and Amy Holt would agree with me on that one. And we are really happy about that we have grown the platform in a sustainable way because we haven't left open source behind. You know, I may say we haven't fucked it up. GitHub is still GitHub. GitHub is still like the most beloved brand in the developer ecosystem. It is adopted by almost any enterprise company in the world. Not all of them have it for all the developers, but most of them have it for some of the developers.
And it has, in that age of AI, reinvented itself. We have reinvented itself. Because just as we used Git to build GitHub, we used language models and GP3 together with OpenAI to build the first ever copilot and to lead that market. Yeah, that's a good response. I'm sorry there was a lot for me to unpack there. So is there a big goal where you all see this all flipping and this becomes super profitable if it's not right now, like the copilot stuff? Is it the market gets to a certain size or there's bundling that happens with other Microsoft products? Or what is the if this, then that that you're seeing that makes this all worth it?
We have a version of Moore's Law, maybe it's Jensen's Law now or whatever, where the GPUs get faster and cheaper. You mentioned the Valley a bunch of times. There's a dozen startups that are trying to invent processing units for transformers that do that more efficiently. But I believe in scale. You could have asked the same question to Jeff Bezos about 15 years ago and looked at where the hyperscalers are today in terms of profitability. And so I think we'll have the necessary patience. We will drive efficiencies. Prompt caching plays already a huge role for agents. We will see an evolution of business models for sure.
And we will see our own developers being more productive with AI. And as such, it's not only that you're optimizing those margins, whatever those may be, but you're also optimizing the way the software is built in our own company. I actually strongly believe that those companies that are still doubting the use of AI, they're going to put themselves in a position that ultimately lets them fall behind in innovation. Whoever's not using AI should have massive FOMO, fear of missed opportunity, that everybody else in their space, in their industry, is already on AI and have their developers. 10% more productive, 20%. And look, you know, these numbers are always like, oh, why is it only 20%? Couldn't it be 30%? 20% is a massive productivity gain compared to, you know, what a software developer costs, how long it costs them to go through, you know, higher education, to ramp up in a company, what have you, to move from one project to another, right? And as such, whoever is not using internally AI to make their own processes more efficient, not only for developers, for everything else, support is another great example, sales, you know, lead generation, all these things, are going to ultimately fall behind. And whether they have positive cost margins or not won't matter to them. What will matter is that they're no longer being able to compete in the market.
That brings me to a memo one of your colleagues recently wrote internally at Microsoft where it was instructing managers to evaluate employee performance based on the use of internal AI tools, saying it's, quote, no longer optional. It's core to every role and every level. I'm sure you can appreciate the angst that... people feel about this shift? People who are farther along in their careers and feel like they're set in their ways, or even younger people who are just coming in and are overwhelmed by all the change. I feel everyone's overwhelmed. And if you're a mid-level engineer at Microsoft, you're probably wondering, am I going to get left behind in this shift? And is the company actually going to be able to even evaluate this correctly? Do you feel, as the CEO of GitHub, do you feel that the management team at Microsoft has an actual measurable way to determine productivity gains internally with AI, and if that's actually going to be something that employees can know in a real way?
The memo was a bit more nuanced. The memo talked about AI learning, AI usage in what we call the Connect, which is a conversation between the employee and the manager. And at GitHub, we have a similar process. And it's about the employee, you know, writing up what they have achieved in the last year, you know, what could they have done better to have more impact? What are they going to do in the next year and how they're growing? That's kind of the framework. And I think in 2025, it's totally fair game to say you should reflect on your AI usage and you should reflect what did you learn about AI? Did you use GitHub Copilot or Microsoft Copilot?
teams co-pilot to summarize a meeting? And if not, why not? And then the manager, you know, provides their feedback on that and what they have learned, which often also is a good learning experience for the manager to say, oh, the employee is actually ahead of me in AI usage. Like, I certainly have had those moments where somebody showed me how they did something, I'm like, oh. didn't realize that that actually works. And I think that process is key to Microsoft's culture. And the term that was coined around this is growth mindset. Having the mindset of saying, okay, I did something and there's a way of doing this better. And I wasn't born with all the capabilities I have. I was born with the capability to improve myself. And I think if you see that memo in that context, it actually perfectly aligns with Microsoft's philosophy. The other thing I would say specifically to GitHub is there is no GitHub employee that cannot use GitHub.
no matter what function you have in the company. So it's not only developers and product managers, it's also HR and finance and legal and all the GNA functions, all the technical functions, all the sales functions, they're all using GitHub. There is no world where I would allow for somebody to say, well, sorry, I don't want to use GitHub. And I think, you know, that's fair game if the employee doesn't want that. And then there's tens of thousands of other tech companies out there where they can have that. But it's part of our company culture that everybody at GitHub uses GitHub. And it should be part of our company culture that everybody uses co-pilot in AI. That doesn't mean that we're looking at how many lines of code you have written today with AI in the same way that we never looked at an employee and evaluated them based on how many pull requests have they filed or how many Git comments have they done because these metrics are easily, you know, gamified, right?
But it shows a mindset that aligns with our culture if you're using our tools to build our tools. Okay, that's a good answer. I want to end on, it's kind of a heady topic. You can take it whatever direction you want. But this debate about AGI, superintelligence, how we get there. A thing I've heard, Zuckerberg has mentioned this, others have mentioned this, is that AI coding is actually the path because it gets you to self-supervised AIs building and maintaining other AIs. And that the company that wins in AI coding will probably get to...
super intelligence, whatever you want to call it. If you have a definition of super intelligence, I'd actually love to hear it. But that AI coding will actually be the way we get there. I'm curious if you agree with that or if you don't. And if you actually have a definition of super intelligence, please share it, Thomas.
I don't have a definition of superintelligence myself because I think ultimately those definitions don't matter much. Either they are encoded in contracts or like a nice marketing instrument and you can go on stage and say, we are the first to have superintelligence. And then next week somebody comes and says, well, no, no, we are the first that have actual superintelligence. These terms are like stretch goals and they are constantly changing. I think the crucial moment where you may call it AGI or ASI is that the AI can improve itself and make itself better. The model, you're basically jumping from GPT-4 to 5 without humans in the loop.
I think when we get to that stage where all of a sudden AI behaves like a kiddo, where your four-year-old moves from laughing about a joke to telling a joke, I think that's when we will talk about AGI. That's the moment when we will say it has happened. And there is now a thing that is able to improve itself and keep going. That's what we as humans really consider intelligence. It's not what you know just because you've been in trivial pursuit. It doesn't mean you're like the smartest person in the world. It just means that you have a good mapping between the questions and the answers. But the ability to constantly evolve and improve yourself, I think that's the definition that we should use. When it acts like a kiddo. I like that. I haven't heard that one yet. Thomas, thanks so much for your time. Thank you, Alex. Really enjoyed the conversation.
Thanks again to Thomas for joining the show and thank you for tuning in. If you'd like to let us know what you thought about this episode or what else you'd like us to cover, drop us a line. You can email us at decoder at theverge.com. We also have a TikTok and an Instagram. Check those out at decoderpod. If you like Decoder, please share it with your friends and subscribe wherever you get your podcasts. And if you haven't already, don't forget to subscribe to The Verge, which gets you access to all of our stories and newsletters, including the one I author called Command Line. Decoder is a production of The Verge and is part of the Vox Media Podcast Network. Our producers are Kate Cox and Nick Statt. This episode was edited by Xander Adams. The Decoder music is by Breakmaster Cylinder. See you next time.
First Citizens Innovation Banking, formerly Silicon Valley Bank, is the bank of the innovation economy. The name may have changed, but for over 40 years, their commitment to bold ideas, relentless ambition, and delivering what's next is as strong as ever. So if you're wondering where the road to innovation starts, it's here, First Citizens Innovation Banking. Head to firstcitizens.com slash innovation to learn more. Support for today's episode comes from Delta, official airline of the National Women's Soccer League. Traveling is basically magic. It creates opportunities, it fosters understanding, it expands horizons, it connects people and communities to each other and to their own potential. Those are the values that drive Delta. With up to 5,500 daily Delta and Delta Connection flights to more than 300 destinations on six continents, Delta is committed to helping you explore your world. Whatever gives you the travel bug, Delta is here to help you get there. Visit delta.com for more info. Delta, wherever the journey takes you.
Running a business shouldn't feel like surviving a software group project. One app for accounting, another for inventory, another for sales, and somehow none of them talk to each other. That's where Odoo comes in. An all-in-one business management software that brings every part of your business together. From sales and accounting to inventory and marketing, all-in-one powerful platform. No messy integrations, no bouncing between tabs, and best of all, no spreadsheets. Stop managing software and start managing your business with one unified system. Try for free today at odoo.com slash vox. That's odoo.com slash vox.