GitHub Copilot and the evolution of programming: insights from Thomas Dohmke

To close out our speaker sessions, MVL sat down with Thomas Dohmke, CEO of GitHub and visionary behind GitHub Copilot. Hailing from East Berlin, Thomas completed a Ph.D. in mechanical engineering at the University of Glasgow, founded startups, and worked at companies like Daimler Chrysler and Bosch. In 2011, he founded Hockey App, acquired by Microsoft in 2014. Thomas is now the force behind GitHub Copilot, the product transforming computer programming and developer productivity.

‍Keith: As a multiple-time founder, what do you think the most important thing our future founders ought to be doing that they're probably missing?

‍Thomas:  You never know where people are in their journey. But I think what people should not be doing [is worrying] too much about the big tech companies and the competition.  … Go build that thing and don't worry about what Microsoft, what Amazon, what Google are doing. You will find the thing that you love and that hopefully makes this world a better place.

‍Keith: At GitHub, your engineers have more experience working with Copilot than anyone else. How has GitHub Copilot transformed developer workflows and your engineering teams?

‍Thomas: Our lives are going to get so much better. A lot of [code written by developers is just boilerplate, mundane, mundane things like unit tests. [Copilot] takes away that boilerplate, that mundane task, and gives you more time to focus on the things that are really fun. [Developers] have less repetitive work. They need less mental energy to get stuff done. And they feel more fulfilled when they're done with the task. …  [Copilot] writes almost half the code.

‍Keith: GitHub Copilot introduces new ways of interacting with code. How do you see the future of programming evolving?

‍Thomas: [This is]  going to shift even further to having AI write 80 percent of that code. While AI will automate coding, programmers will still make architectural decisions. Programming languages may become more human-friendly, striking a balance between AI synthesis and human understanding.

‍Keith: Do you think AI and GitHub Copilot will change the role of programmers and their approach to coding?

‍Thomas: AI automates coding aspects, but programmers remain essential for architectural decisions, troubleshooting, and tech understanding. You have to make a lot of decisions as developers and a startup founder every single day. and those are hard because there's not one right answer. The decision is still with us and our experience and sometimes the gut feelings we have.  

‍Keith: What's next for GitHub Copilot, and what can developers look forward to? I see lots of requests for fine tuning.

‍Thomas: Copilot is continually evolving, integrating more AI, offering plugin creation, and AI-native interactions in the future. Fine tuning is only one option, and a potentially expensive and time-consuming one. Plug-ins are another way to add functionality. I think the majority of customers will live with what's called retrieval augmented generation.

‍Keith: You’ve used Copilot to code live on stage. That’s pretty bold.  

‍Thomas: Maybe, it's like a stand up comedy, right? But you never know what the audience will do and what they throw at you. And you never know what the model throws at you and what solution it will give you today. The nice thing is then you rehearse not to write exactly the same line of code over and over again, you rehearse how to react to whatever the model gives you. Also it's fun because you can also tell chat GPT and copilot that the answer is wrong.And it says, “Oh. I apologize.” Here's another way of approaching this - not because it’s really sorry, it's just the next probable thing to say.

‍Keith:  I have one last question from the audience. Can you share the story about the founding of Copilot at GitHub? How did you convince anybody that this could work?

‍Thomas:  It was right after the Microsoft build conference, which that year, was all virtual because of the pandemic. We hopped in on a zoom call [with Sam Altman and Kevin Scott] and we were dictating  prompts to render a Python method that sorts an array or write JavaScript that posts a tweet,  and we were “similar to this, a bit more interactive”. We asked internally for developers to submit programming exercises in Python, what you would give people for an interview loop. The [first] codex model was able to solve 92 percent of these 230 or so programming exercises!  And that was the moment where we are all like, okay,  we're onto something here. We started building Copilot and the rest became history. 

Thank you to Thomas and the team at GitHub!