GitHub CEO Thomas Dohmke

Today we have the pleasure of hosting GitHub CEO Thomas Dohmke. He and Madrona Partner Aseem Datar talk about how Thomas got into working with computers and coding and the work he’s been doing since becoming GitHub CEO in November 2021, including the recent launch of Copilot X. But these two discuss so much more, including the rise of generative AI, talking about everything from how it is a new way for developers – everyone really – to express their creativity to how it democratizes many skills and access to those skills to the generative AI-powered developer experiences and how the constantly evolving world developers have always worked in has set them up with the perfect safety network to leverage generative AI to its fullest potential. Thomas also offers up advice for people just launching a startup. But you’ll have to listen to hear it all.

This transcript was automatically generated and edited for clarity.

Aseem: Hey, everybody. My name is Aseem Datar. I’m a partner at Madrona Ventures. Today I have my close friend and GitHub CEO Thomas Dohmke. I’m excited to chat with him on this wonderful topic of generative AI.

Thomas, welcome.

Thomas: Yeah. Hello, and thank you so much for having me, Aseem.

Aseem: We are excited more than you are, Thomas. It’s always fun to talk to somebody leading the charge on innovation in this industry. Maybe start by giving us a little bit of your story and introducing yourself.

Thomas: I’d like to say I’m Thomas, and I’m a developer. I’ve been identifying as a developer ever since the late ’80s early ’90s when I was about 12-13 years old, and I got access to computers first in the geography lab in school and then later when buying a Commodore 64. I’ve been fascinated by building software. And, obviously, as a kid also gaming and playing with all kinds of aspects of computers. And I have been working with code and being passionate about code ever since building my own applications, studying computer engineering first in Berlin. And then, doing my Ph.D. in Glasgow, I worked at Mercedes, building driver assistance systems. And then, in 2008, Steve Jobs announced the App Store. And it pulled me into the app business. I had a startup called Hockey App that was ultimately acquired by Microsoft in 2014, and that moved me from Germany all the way here to the West Coast into Microsoft and that path. Then led me first into GitHub through the acquisition, running special projects at GitHub, and since November 2021, I’ve been the CEO.

Aseem: What a fun journey. Thomas, I can’t stop myself from saying developers, developers, developers all the way from the Steve Ballmer world. And it’s so much fun to be talking to you. Clearly, a lot has changed in the world. There’s this rapid pace of innovation that we are seeing with this new capability set called generative ai. And we are all excited about talking and hearing more about generative ai. What’s your worldview? I would love to just understand that.

Thomas: If I look back over the last six months or so, we had multiple moments that you could compare to the App Store moment I described earlier. That happened in 2008. I think the biggest of those moments clearly was ChatGPT late last year. And you know, I have heard people describing that moment of ChatGPT launching and seeing fast adoption as the Mosaic moment of the 2020s. If you’re old enough, you might remember the first browser Mosaic and then quickly followed by Netscape and, actually, last night over dinner, I argued with folks — is it the Netscape moment or the Mosaic moment? I think it doesn’t really matter, but what matters is that within a rapid amount of time, people adopted ChatGPT and had seen the way they work shifting. And before that, before ChatGPT, we had already seen a shift through Midjourney and Stable Diffusion — those image models. And I think, you know, those models are great to describe what generative ai does, and part of it is really creating a new way of people expressing their creativity. And we have heard stories of folks spending, you know, their evenings rendering images instead of watching Netflix. I think that’s exciting. My example always is, you know, depending on what city I’m in, what customers I’m speaking to is like, you know, ask Stable Diffusion to render the skyline of Tel Aviv as if it were painted by the French impressionist Monet. And obviously, Monet hasn’t seen the skyline of Tel Aviv as it looks today. And yet, those models generate a picture that resembles a Monet rendering the skyline of Tel Aviv, Sydney, or San Francisco. And I think that is really the power of this new world of generative ai.

And the other thing that it brings is it democratizes a lot of skills and access to those skills. And especially if you think about students and kids that sit in class and where the teacher in front of a class of 30 kids just doesn’t have the time to be the tutor for each every single individual kid, but have it giving them an AI assistant where they can ask all the questions that they might not dare to ask in class, or where they, you know, didn’t have the time or the teacher didn’t have the time or the parents don’t have the time because they’re working three jobs. I think that is where really the power of this AI moment comes from and where we see tremendous excitement in the industry and really in, in everybody you’re talking to.

Aseem: Yeah, I mean, no question. Right? I think productivity is such a massive space where generative AI is having an impact today. It’s awesome to see these scenarios in real life, come to light, whether it’s for students, whether it’s for business workers, whether it’s for information workers, but behind it, all is the ethos of creativity in some senses in the software world are developers, right? And I think you can’t run away from the fact that there are developers creating these intelligent applications and embedding AI into it. So what does this moment really mean for developers? How do you think the generative AI-powered developer experiences will change?

Thomas: The role of developers has always changed, right? If we look back over the last 40 years, we went from punch cards and machine language and mainframes and cobalt and whatnot to modern programming languages. We went from building everything ourselves before the internet to leveraging thousands of open-source components ever since, you know, the early 2000s, I’d say.

Aseem: By the way, I thought Visual Basic was a big moment. just You know, going back to those days but, but carry on.

GitHub CEO Thomas Dohmke: And you can probably make that argument for many programming languages in their own right. I think Ruby was a great moment as well. And a lot of startups in the last decade or so were founded on Ruby on Rails because it’s just so easy to iterate with Rails. And Python, you know, unlocked a lot of the machine learning that we are now seeing. And the nice thing you know about software development is that it has been always part of the practice of software development to solve issues, right? No developer is perfect, whether we made mistakes on punch cards, we made mistakes in assembler and now we are making mistake in code. It has always been around solving issues, fixing your own bugs, or fixing your team’s bugs. And the word bug even comes from the bug on the punch card. And so, we built all this tooling, compilers, and debuggers to find, issues by writing code. We invented practices like unit testing to make sure that what we’re building is the thing we wanted to build. And in the last decade or so, we introduced DevOps practices or agile practices, code review, pull request, pair programming, continuous integration and deployment, CI/CD, code, and secret scanning. And so if you tie this now to AI, it’s actually fascinating. We’ve built the safety network within software development to leverage generative AI to its fullest potential. We all know that those models, those large language models, are not always right and that they have something called hallucinations. They think they have the answer, and they’re confident in what they’re saying, but it’s wrong. And with all these practices that software developers have, we have the safeguards in place to make sure we can work with a model suggestion and either take it and modify it or take it and then figure out in code review that is not exactly what we want to do. You could argue we built DevOps with the aspiration that in the future, there will be a moment like ChatGPT, where we can unlock more productivity, more creativity in developers to ultimately become realize even bigger ideas. I think that’s ultimately what this is all about.

And at GitHub over two years ago, now — in 2020, we started working on Copilot, which is one of the first AI pair programmers. It sits in your editor, and when you type as a developer, it suggests code to you and can complete a line, but it can also complete whole methods — multiple lines of code, lots of boilerplate, import statements and Java and whatnot, test cases, complex algorithms. And it’s not always right, but developers are used to that. They type in the editor, and it shows the suggestion. And if that’s not what I want, well, I can just keep typing, and if it’s close enough to what I want, I’d press the tab key, and I can use that and modify it. And that’s no different than copying code from Stack or from GitHub and then modifying that. Almost never, you know, you find a snippet on the internet that’s exactly what you want.

The generative AI-powered developer experiences gives them a way to be more creative. And, I mentioned DevOps earlier. I think DevOps is great because it has created a lot of safeguards, and it has made a lot of managers happy because they can monitor the flow of the idea all the way to the cloud and they can track the cycle time. And they have a certain level of confidence that developers are not just SSHing into a production server because, they are some safeguards in place, but it hasn’t actually made developers more happy. It hasn’t given them the space to be creative. And so, by bringing AI into the developer workflow by letting developers stay in the flow, we are bringing something back that got lost in the last 20 years, which is creativity, which is happiness, which is not bogging down developers with debugging and solving problems all day but letting them actually write what they want to write. I think that is the true power of AI for software developers.

Aseem: I remember my days of writing code in an Emacs editor, and that was just like slightly better than Notepad because it had a few color schemes and whatnot. Two things that you mentioned that I latched onto. One is productivity, and the second is creativity. And I think those two certainly are top of mind for developers. What are some of the things that developers should be excited about, and what are some of the areas that you guys have doubled down in and will continue to double down in?

GitHub CEO Thomas Dohmke: Yeah. I mean, let me take you on a bit of a history lesson. In the summer of 2020, GPT-3 came out, so that’s almost three years ago, and back then, you know, our GitHub next team that team within GitHub that looks into the future asked themselves can we use GPT-3 to write code? And, we looked into the model, and we came up with three scenarios.  It’s fascinating now in 2023 to look at these three scenarios because there was text to code. So that’s what Copilot does today, right? You type text, and it suggests code to you. Code to text, which is like you, you ask the model to describe what the code is doing. And we just announced that as part of Copilot X, where you can have Copilot describe a pull request to you. And if you’re a developer, you know what that’s like. You’re working all day on a feature, and you’re submitting a pull request, and now you have to fill out all these forms and its title and the body, and like, ah, I know, I know what I did today. And it’s all obvious to me because I build all this code. I don’t want to spend too much time describing that to others. And so, with copilot for pull requests, we are helping people to just do that for them. And it describes the code, but it’s not only about the pull request, it helps you to describe code that you might be reading from a coworker and the editor. It might just help you to remember what that was. And it might help people to describe old code. This old COBOL code that some banks are still running and its code that’s from the ’60s, running on mainframes where the people that wrote that code back then are long in retirement, I hope. And so, the expertise is gone. And then the last one was conversational coding. And we didn’t build that at the time because we felt the model was not good enough to have these kind of conversations. And clearly now, with ChatGPT 3.5 and, and now GPT-4, we have reached the point where those chat scenarios are useful. And more often right than wrong. Back in 2020, we explored these three scenarios, and the way we validated, that this is good enough for us and that we can build a product on top of that was we asked our staff and principal engineers to submit coding exercises, things we would use in an interview loop — a description and a method declaration and a method body. And so we got about 230 or so of these exercises, and we stripped out the body and basically gave only the declaration and the description to the model. And we gave the model 150 attempts for each exercise to solve the exercise and get close enough to the solution. And what we figured out from this experiment that 92% of those exercises could be solved by the model back then in 2020. Even then, already the model was good enough for a lot of these coding exercises. And so, we took that as inspiration to build Copilot and ship Copilot to the world.

On March 22nd, we announced Copilot X. So, then the next generation of Copilot, of really bringing the power of these AI models into all parts of the developer experience, whether it’s coding in your IDE, whether it’s chat scenarios where you can explore ideas. The example I tried first was I ask it how to build a snake game in Python. You know, the game that we were playing on cell phones before they had touchscreens. And it starts showing an explanation of how you do that, and then you can just ask it to “Tell me more on step one,” and it shows you some code, and you can start building with that. I think that’s the true power here is that you can rediscover your love for programming if you lost it. Or you can explore a new programming language, or you can just, you know, ask the chat agent to fix a bug in your code or fix the security issue, like to remove that SQL injection that you accidentally put there. We announced Pull requests. I’ve already mentioned that describing pull requests. And soon enough, we will also have test generations. So, the pull requests will check whether you actually wrote or the tests you’re supposed to write and then generate those tests for you. And then the other cool thing that we announced is Copilot for docs. And so, we built a feature that basically lets you ask questions about the documentation for React, Azure, and a couple of other projects.

And so, the model has a cutoff date until it was trained. And the training is a really expensive process. It takes, you know, weeks on a supercomputer to train the model again. The current GPT-4 has a cutoff date of September 2021. And it actually will tell you that, if you ask questions about things that happened since then. And so it doesn’t know about changes to open source projects in their documentation that happened in the meantime. And, you know, September 2021 to, we are recording this in March 2023, is a long time for APIs of open source projects. What we’re doing is basically we are collecting that data from those open source projects, and we are feeding them into the prompt, so they’re becoming part of the prompt of the part that you’re not seeing as the person asking the question, so can answer up to date questions on those projects.

Aseem: I am so excited about Docs, right? Like, I go back to my days as a developer, and so much time was spent on going and reading up docs and pulling up from different places, and it was just a productivity suck. So, congrats and kudos. And I do want to point out that I think GitHub created this notion around Copilot, which is now injected all across Microsoft, and now there’s a copilot for Office as a copilot for Teams. I couldn’t be more excited to see where this goes. Shifting gears, a little bit, Thomas, one thing that gets me excited, especially in the world of venture, is that our startup founders and teams can now go from zero to production very quickly. What advice do you have for somebody starting out, like building a business or creating a team? What should they be bullish on? What should they be worried about?

GitHub CEO Thomas Dohmke: I think you know a lot about creativity is to stay in the flow and not get distracted from all the things that are happening around you. And oftentimes, you know, we, we are like gravitating to those things, whether it’s the browser or whether it is social media and whatnot. And so, I think my first advice to startup founders is, you know, stay focused and leverage the time of the, day when you’re actually creative because that time is so limited. Like, you know, our creativity is infinite, but the time when we are actually creative during a day, when we have the energy to build cool things, is fairly limited. And, for some people, it’s early in the morning. For me, it’s usually after my first cup of coffee, that’s when I’m the most creative. And then I always want the second cup of coffee to have that same impact, and it doesn’t, right? It never works that way. am also creative at the end of the day when it’s dark outside, and I’m a bit of a night owl as well. And so I think, you know, as a founder, you have to find those moments during the day and keep that energy ultimately flowing.

We live in this, you know, world right now, whether you call it a recession or not, I think we are in a complicated macroeconomic environment, to say it more, politically correct. But I think those times are always challenging and opportunities at the same time. And we saw this in the last downturn in 2008 — many of the startups that are now part of our lives, like Airbnb, Uber, Slack, or Netflix, they were founded around that same time. And they’re now part of life. And they, or Shopify actually, is another great examples of these that was founded during a downturn, building the technology, and then as we came out of this everybody wanted to have an e-commerce store or, and buy from these stores. And I think that’s the opportunity that we have now and today or this year, it’s leveraging generative AI, as like the foundational layer. And many startups will build on top of that, and they will have to find differentiation and defensibility of their idea. And I think, you know, we’ll see a. cool ideas building on top of ChatGPT or GPT-4, and a lot of these are really cool, but they’re also probably not going to survive as a company on their own because it’s a small idea that, you know, summarizing your emails in Gmail. I would think Google will build that into the product and then you really have to push hard to make that a paid product that’s customers will pay for if they have that already built into Google.

Aseem: I couldn’t agree more. We’ve always talked about do more with less, but I think the, the AI or the capabilities that we are seeing pop up is all about doing much more with much, much less. And that’s, I think, the beauty of the pace of innovation that we are seeing all around ourselves. Thomas, I know that you are, you’re deeply plugged into the startup ecosystem. You see a lot of these open-source projects come to life. Are there any projects or startups that you are really, really excited about?

Thomas: I’m, I’m staying bullish on ChatGPT and OpenAI, and I think we are at GitHub very excited about the future of Copilot. I mentioned earlier things like Stable Diffusion and Midjourney, which make me really excited. I’m, I’m not an artist at all. I can’t draw, and I certainly cannot paint something that looks like a Monet. And if you take that a step further, I’m really bullish and excited about a startup called Runway that lets you generate videos from images, from video clips, but also from text prompts. And I think, you know, there’s going to be a moment where you can just write a script into a text field, and it generates a full animated video for you. And that will allow us to take the stories that we heard as kids from our parents or even grandparents and turn them into little video clips that we can show to our kids. And I think that will be so cool if you can basically tell the stories for me now, two or three generations ago in little videos to the next generation. I think you and I both sit on a board of a company called Spice AI that explores AI from a different perspective, which is not about large language models or image models. It’s about time series AI and finding anomalies in time series data. And it allows you to query that data, and they started with blockchain and Web3, and you can write your own queries and quickly figure out what’s Bitcoin doing. But you can also run AI on top of that and find things that are interesting, find alerts, or find price changes. And in the future, I think there’s a lot of huge space in there. You can apply this to your server data, your server monitoring maybe your Kubernetes clusters. There are all kinds of time series data that affect us every day — weather is ultimately also time series based, right? Like it’s cold in the night and warm in the day. And so, I’m excited about that. In general, you know, the AI and mL space is super exciting for me. There are so many startups I could list here. There’s Replicate, a startup that’s based in Berkeley. They’re letting you run machine learning models with just a few lines of code. And you don’t actually have to understand how machine learning works. There’s OctoML based in Seattle that uses machine learning to deploy machine learning models to the cloud and find the most efficient version, you know, the right GPU type, and the right cloud provider for your model. But I think you know the ML AI space is super exciting, and I’m sure we are going to see lots more ideas that nobody thought is possible and and nobody thought about right now. And. Similar to, ChatGPT in hindsight, seems so obvious. But until it came and conquered the world, nobody else had built it. So, I think I couldn’t be more excited about that future.

Aseem: Yeah. And I echo that sentiment. We at Moderna are really excited about being able to help Runway, OctoML, and Spice AI in their journey of building out for the future. And I think it’s always interesting to see the future getting accelerated in a way that we can, that we can’t even imagine, to be honest. And yes, there is scenarios around hallucination, et cetera, that we’ve all got to watch out for. And I think you said it well, which is, it’s a start. There’s still going to be a developer or human in the loop, at least for the short term, until it gets to a point of high confidence.

Thomas one other interest. Notion that, that I wanted to sort of also pick your brain on is if I’m a startup founder, what should I look forward to in the distant future? I mean, we talked about all these modalities, but one of the challenges that founders have is developers are hard to come by and top talent is very hard to come by. And, there’s this notion around, tools being built to go tackle the low-code, no-code space or democratize development. What’s your view on that from a GitHub perspective?

Thomas: You know, I think there’s this, slogan, fake it till you make it. And that’s true for so many founders as well. You know, you don’t have to have a perfect solution right from the start. You can combine all these AI tools that are available to you now to stitch something together really fast. Whether it’s copilot, whether it’s Stable Diffusion, whether it’s some of the other tools that help you just by AI — help you write your marketing copy. Embracing those things as much as possible and adjusting your style to it. I think what will happen to developers is that the developer will learn how to leverage AI to its best. Andrej Karpathy tweeted about this recently where he basically says I changed my programming style and by writing a bit more commentary and a bit more declarative, statements I can get Copilot or aI to actually synthesize more code for me. And I think that’s kinda like what we are going to learn and where I’m bullish on building AI in the open and having those models out there, and building with them and learning how to use them as early as possible before we get to AGI and there’s a certain amount of scare about this and what we can do. But you know, today, those models are, not sentient. They’re not actually creative. They’re predicting the next word. And if you wanna switch ’em off, you can just go, to an Azure data center and switch, switch it off. And I think, but, so we need to build this in the open and we need to learn from where’s the model good and how can we use the model to help us as humans. And we also need to learn where’s the model bad and where make does it make mistakes or makes wrong predictions.

And actually, I think, the model itself will be able to correct itself. I think, there was recently an example from Ben Thompson’s blog, Stratechery, where basically somebody on social media posted, I think, four paragraphs of a blog post from Ben into ChatGPT and then asked it who wrote this. And it basically detected that this is a blog post from Ben Thompson without telling it that information. And I think, in the same way, we will be able to use AI to detect something that was wrongly written by AI. And so, the technology works with each other. And I think by building this in the open, we are preparing for that future where AI plays a bigger role for us on this planet.

Aseem: Hey Thomas, I know we are out of time. Thanks so much. This has been a blast, and I’m sure our startup founders, our listeners, are taking so much away from this discussion with GitHub CEO Thomas Dohmke. And I couldn’t thank you enough. Thanks for being on with us. And we’re excited to be able to partner and work together.

Thomas: Yeah. Thank you so much for having me on this podcast.

Coral: Thank you for listening to this episode of Founded & Funded. If you’re interested in learning more about what’s going on at GitHub, check out their blog at Github.blog. Thanks again for listening, and tune in in a couple of weeks for our next episode of Founded & Funded.

‍