GitHub: “Artificial intelligence will allow us to become a pillar of Microsoft” (Thomas Dohmke, CEO)

Hello, I’m Thomas Dohmke, CEO of GitHub but above all software developer. This is how Thomas Dohmke invariably introduces himself, The gallery was able to meet in Paris. It is partly for this attachment to the qualification of developer that this forties native of East Germany was appointed to his position of leader at the end of 2021.

Already reference software from the software community open-source [des logiciels publiés sous une licence libre qui permet à n’importe qui de les utiliser et de les modifier]GitHub now aims to become “ where developers live “. Itself built on the basis of an open source project called Git, the company concretely allows developers to host their computer code and manage the different versions of their projects.

But GitHub had a real turning point in its history in 2018. Nine years after its creation, Microsoft paid, to everyone’s surprise, $7.5 billion to buy it. If the company has not denied its core business, it aims to open up a whole new market around artificial intelligence and its new flagship product, the code writing assistant. copilot. Explanations with Thomas Dohmke.

LA TRIBUNE – Microsoft’s acquisition of GitHub in 2018 shocked the free software world. Many at the time found this marriage unnatural. How have these accusations been seen on Microsoft’s side?

THOMAS DOHMKE – Microsoft has always been a software development company, it’s in its DNA. During its 45-year history, it has gone through several transformations, and one of them was opening up to open source. Microsoft has moved from an almost negative view of open source to becoming a real supporter, using it internally and contributing to it. The company now has a gigantic open source program, where tens of thousands of its developers contribute. So acquiring GitHub in 2018 was just another step, a commitment to the community.

Critics feared that Microsoft would misrepresent GitHub, including through the pursuit of profit. What about the autonomy of the company within the group today?

We want GitHub to remain independent, and we don’t want it to become a sales channel for Microsoft. If you go to GitHub today, you won’t find any Microsoft logo or IDs, and there’s no preferential treatment for Microsoft over competitors. Whether you deploy on Azure, Amazon Web Services, or Google Cloud, it’s the same for GitHub. Four years after the acquisition, we are still 100% invested in GitHub’s neutral position.

We imagine all the same that being part of Microsoft brings certain advantages…

Of course, we find commercial synergies, since we are part of a large company and it sells GitHub products to companies like Société Générale, Engie or Décathlon, among other French customers. We take advantage of Microsoft’s international sales teams, and we leverage the contracts already in place with Microsoft. Often, when selling enterprise software, one of the biggest hurdles is sorting out the details of the contract between the legal teams of both parties. However, when the contract exists between Microsoft and the customer, it is easy for him to add GitHub to the invoice.

We also find technological synergies within the group. Copilot, for example, was created in partnership with Microsoft and Open AI. We used a model of machine learning created by Open AI, which we ran on GPUs [des unités de traitement graphique, Ndlr] provided by Microsoft Azure. But we are trying to develop these synergies while maintaining our independence and our startup mentality.

In recent years, Microsoft’s growth, especially that of its profits, has been driven by its cloud division, Azure. Can we imagine GitHub taking on a similar role in the future?

Yes ! Like Windows, Office, LinkedIn, and Azure, GitHub is becoming a Microsoft mainstay. There are overlaps between these activities – Azure runs on Windows, and Office runs on Azure infrastructure for example. GitHub fits into these patterns, and can play a big role in Microsoft’s future.

Despite the $7.5 billion acquisition, GitHub remains relatively unknown outside of the IT community, and almost conveys the image of a small company. How much does it weigh in Microsoft’s business?

I can’t give you a precise turnover, but it’s becoming substantial [peu après notre interview, Microsoft a annoncé que GitHub réalise 1 milliard de dollars de revenus récurrents, contre moins de 300 millions de dollars lors de l’acquisition, ndlr]. It’s reflected in the fact that we have 3,000 employees, 94 million developer accounts, and over 4 million client organizations including 90% Fortune 100 companies. We’ve grown significantly since the acquisition and continue to grow. . In 2018, we had 28 million accounts, and more than three times that today! We have fulfilled the expectations that Microsoft had for GitHub when it was acquired, and with artificial intelligence, we believe we are entering a new era of software development, which we hope will be one of prosperity.

Copilot, the programming assistance AI that you launched in February 2022 after almost a year of testing, was your first step in this new direction. Why do you rely so much on artificial intelligence?

To understand our enthusiasm, we must retrace the history of IT development. First there was a first wave, from the 1980s until the early 1990s, where learning computer programming was only through books and magazines. You had to buy expensive books, because the libraries only had outdated references. If you got stuck with a problem, the only way to move forward was to ask other people for help, which was difficult outside the big cities. From the mid-1990s, the Internet began to develop and forums appeared. Soon enough, it was possible to exchange easily with other developers, and collaboration improved.

It’s the second wave: the Internet democratized software development, and the world of open source was born. You can share your code with other developers so they can build their own code on top of it. This ecosystem now allows startups that are launching not to start from scratch, but to have access to thousands of open source libraries. As a result, it’s easy to build an app today, thanks to many tools, including GitHub.

We believe that IT development is now entering a third wave, that of artificial intelligence. Through programs like Copilot, which help developers write their source code, developers will have access to a new way to express their creativity.

What does Copilot actually bring?

Sometimes we – developers – get stuck writing code, because we can’t remember what the programming interface looks like, or how to connect to certain servers, or how to decode an image. So you have to look online, and this is where the story gets complicated. We leave our code editor, where we were creative, to go to our browser, where 15 tabs are open with surely a TikTok video, a discussion group with the family or a news site, so many sources of distraction.

But the worst part is that we are also distracted when we look for the solution to the problem on GitHub, Google or Stack Overflow, this time by other developers who debate on the forums or in the comments on the best solution to adopt. And since there is often no single solution, you have to take the time to find what best fits your own problem. The result of this situation is that you come back to your editor 20 minutes later, and you have completely forgotten what you are working on. Copilot partially avoids this loss of time. It suggests pieces of code, which the developer chooses to accept or not. If he continues to write, the software will adjust his proposal. Once the developer is happy with it, he can press a key to insert it and modify it as he wishes. As a result, when Copilot is active, it writes 20% to 40% of the code depending on the programming languages ​​and the users.

The deployment of Copilot has raised many questions on the respect of copyrights, since certain codes are deposited in the registers of intellectual property. How do you address these concerns?

A machine cannot infringe copyright, since it is not responsible before the law. Keep in mind that Copilot has no understanding of the language, it just converts what the developer writes into a number, then does some calculations to output a new number, which will be converted into a code proposal. He only applies a method, he is not intelligent and even less gifted with feelings.

However, it is what the user does that counts. Whether he copies and pastes a piece of code he found on the Internet or whether this piece was suggested by an artificial intelligence comes to the same thing. It is always the responsibility of the developer to verify that he does not copy word for word a code deposited by someone else, and therefore that he does not violate a copyright license.

Writing code is only the first step in the code life cycle, after which it must be reviewed. Concretely, the developer sends a request to the members of his team or to the maintenance team of the open source project he is working on, and it is these other people who will accept the code in the main database. They will look at the code and compare it to what is being done. This is an important step that does not change with the wizard.

However, we are aware of the challenge ahead, and that is why we added a setting during the global launch of Copilot this year. If the developer activates it manually (it is deactivated by default), Copilot will compare the code it generates with an open source code base hosted on GitHub. If the snippet looks exactly like – we’re talking about a significant fragment, which would embed a certain level of creativity, not a simple sentence that would be too short – the assistant will not offer the snippet. The setting should prevent developers from accidentally copying something created by another developer, and therefore covers some of the risk.

What place will Copilot and future AI-powered tools take in the GitHub model?

It’s a completely new category of products, although it’s close to GitHub’s mission to put developers on the best footing. But Copilot takes us away from the heart of our business, which is to become the home of open source and devops. [l’articulation entre le développement et l’administration des systèmes, ndlr]. We are now creating a product that aims to improve developer productivity. Since the launch of the test phase last year, more than one million people have registered an account. There is clearly a market.

What are the next steps ?

We anticipate several advances. To begin with, the models will become more and more powerful and therefore precise. On the one hand, they will integrate more parameters thanks to the computing capacity of GPUs which is constantly improving. On the other hand, they will be trained on an ever larger volume of open source code, part of which will have been created thanks to Copilot. A virtuous circle sets in.

The other axis of development is the user experience layer. Copilot is successful in part because it’s an open app that fits into development environments, not a separate program. We believe a new wave of products will transform how developers work. If we manage to meet expectations, we believe that in five years, 80% of the code will come from suggestions by artificial intelligences.