The present, past and future of GitHub with Thomas Dohmke

The Pragmatic Engineer

The present, past and future of GitHub

0:00

-1:27:14

The present, past and future of GitHub

How GitHub evolved from a Rails monolith, embraced remote work, why it hires more junior devs than ever before, and how GH Copilot was built thanks to a happy accident. With GitHub CEO Thomas Dohmke

Gergely Orosz

Jun 18, 2025

Transcript

Stream the Latest Episode

Listen and watch now on YouTube, Spotify and Apple. See the episode transcript at the top of this page, and timestamps for the episode at the bottom.

Brought to You By

Statsig ⁠ — ⁠ The unified platform for flags, analytics, experiments, and more.
Graphite — The AI developer productivity platform.
Augment Code — AI coding assistant that pro engineering teams love.

—

In This Episode

GitHub recently turned 17 years old—but how did it start, how has it evolved, and what does the future look like as AI changes how developers work?

In this episode of The Pragmatic Engineer, I’m joined by Thomas Dohmke, CEO of GitHub. Thomas has been a GitHub user for 16 years and an employee for 7. We talk about GitHub’s early architecture, its remote-first operating model, and how the company is navigating AI—from Copilot to agents. We also discuss why GitHub hires junior engineers, how the company handled product-market fit early on, and why being a beloved tool can make shipping harder at times.

How GitHub built Copilot

An interesting quote from the episode is how and when GitHub started to build Copilot — during the pandemic, after getting access to GPT-3. And how it all started with a published paper:

Thomas (at 01:07:03): So GPT-3 came out right after the BUILD conference in 2020, which was obviously given the pandemic fully remote. But Kevin Scott and Sam Altman did a session about transformers and large language models. And then, after that, GPT-3 got into the preview, and we got access to that through the OpenAI/Microsoft partnership.
We realized with OpenAI that GPT-3 was able to write decent code in different programming languages and would not mix up the syntax between Python, Ruby, and JavaScript. Then, OpenAI finetuned a model that was called Codex that was specific for these coding scenarios.
In 2020 and August we wrote a paper with three ideas. We had:
Text to code
Code to text as in describing code
Conversational coding which today is known as Chat.
And those two latter scenarios didn't work well enough. But text to code — as in prompting the model within the editor — ultimately turned into auto completion. That worked so well that quickly we saw our internal Hubbers, adopting the tool, giving it really high scores saying "this is great, I want to keep using this."
It was not the typical “management says you have to use it” and you don't want to. In the early days, it wrote about 25% of the code in the fileswhere it was enabled. Shortly thereafter, that number got to about 46% in early 2023. That was the early days of copilot.
In June, 2021, we went into the public preview and within a few months it had gone to one million users. We saw more and more folks on social media saying "well, I was skeptical that this could ever work, but it actually is good enough that I don't want to work without it anymore.”