GitHub CEO: In the future, programmers will be like conductors, leading a team of AIs.

With the rapid rise of generative AI, the world is undergoing an unprecedented reshaping of industries and occupations. "Will AI replace programmers?" This has become one of the most frequently discussed questions among developers. The rapid iteration of tools such as automatic writing, AI image generation, and code generation has made many in this professional group anxious about their uncertain future: Will they still have a place in the future? Especially in the field of software development, the emergence of intelligent agent tools such as GitHub Copilot and ChatGPT is regarded as a landmark signal that may "subvert" the programmer position.

Recently, Thomas Dohmke, the CEO of GitHub, the world's largest code hosting service platform, was a guest on the podcast "Eye on AI" to discuss how AI has forever changed the programming field. Dohmke said that AI is not here to replace programmers but is helping developers evolve into "intelligent agent commanders." In his view, the future of software development will be an era of in - depth collaboration between humans and intelligent agents.

The following are the core views shared by Dohmke:

1. AI technology is helping to reconstruct the programming paradigm. GitHub Copilot has helped generate 46% of the code lines, and natural language is becoming the "universal programming language".
2. The revolution of intelligent agent collaboration is coming. GitHub is shifting from "human - to - human collaboration" to "human - to - intelligent agent collaboration" and has launched three types of intelligent agents for programming, review, and repair.
3. The role of developers will evolve into "intelligent agent commanders", and their core capabilities will shift to task decomposition, requirement description, and AI collaboration decision - making.
4. GitHub's vision is to expand the number of developers from 150 million to 1 billion, enabling everyone to learn to write code and become software creators rather than consumers.
5. In the future, developers can freely switch between AI models such as GPT and Claude, and the cost of switching to another model and trying again is very low.
6. AI will not replace developers but will accelerate their work and inspire more people to participate in software creation. In the future, the open - source culture will continue to promote global developers to collaborate across borders, forming a more powerful innovation network.

The following are the highlights of Dohmke's latest dialogue:

01 GitHub is not just a code library but also a home for developers

Question: Every time I mention GitHub, I think it has long exceeded the simple definition of a "code repository (Code Repository, a systematic storage space for storing, managing, and tracking project code changes)". Can you briefly review the development history of GitHub?

Dohmke: GitHub has never been just a "repository" from the beginning. Although the repository is a basic function, it truly revolutionized the way developers collaborate. In October 2007, several founders started developing this project. Interestingly, when we later made a commemorative poster, we listed the first 10 commit records - at that time, they never imagined that this project would become so important. GitHub was officially launched in early 2008, and the earliest users mainly came from the Ruby on Rails community because the founding team itself was from this San Francisco developer circle.

Initially, GitHub was indeed a code hosting platform, but it soon launched the epoch - making "pull request" function. In fact, even the repository function was quite advanced at that time: developers could directly browse code and view file history on the web page, which was revolutionary in 2008. The simple and efficient operation attracted a large number of developers to settle in. Subsequently, GitHub successively added functions such as issue tracking, making collaboration smoother.

Looking back now, the most special thing about GitHub is that it blurs the boundaries between different types of developers. Whether they are open - source contributors, commercial developers, or start - up companies or individual projects, the development processes are actually quite similar. Many times, the same person switches between different identities at different times when working on open - source projects, amateur projects, start - up projects, and work projects. Looking at a developer's GitHub homepage, you will find that it often records various types of projects, which is a portrayal of the diversification of modern developers.

Question: Where did developers store their code before GitHub?

Dohmke: This question is quite interesting. The most primitive way was to directly create folders on the computer, such as naming them V1, V2, or adding a date. But it would get messy in no more than three days - either forgetting to update the version or not being able to find the required code when trying to roll back. To be honest, even today, many beginners still use this old - fashioned method.

In fact, the history of version control could fill a book. The earliest popular code version control software was CVS, which was later replaced by Subversion (SVN). At that time, everyone stored their code on SourceForge (an early website for hosting open - source software). The birth of Git (a distributed version control system) is thanks to the Linux kernel team, and even Linus Torvalds, the father of Linux, was involved in its development. There was also a competitor called Mercurial at that time, but in the end, Git prevailed.

Interestingly, when GitHub was first launched in 2008, no one was sure that Git would become the mainstream. The earliest batch of users were basically old users of SourceForge and SVN who switched over. Even now, we often have to help people move their old SVN (an abbreviation for Subversion, a version control system) repositories to GitHub.

Question: Can you explain the difference between Git and GitHub in plain language? Especially what are the innovative points of GitHub?

Dohmke: We can understand it this way: Git is like a smart notepad that can remember every modification you make to code files (actually text files). Simply put, imagine that when you are writing an article, Git will automatically save each version, and you can look back or restore at any time. Many office software (Google Docs) now also has a similar function, but it was a pioneering idea before Git appeared.

The most powerful feature of Git is its "decentralized" characteristic. This means that each developer's computer has a complete copy of the code library, including all modification records. And GitHub is like a "code club" that brings these scattered copies together. You can "pull" my modifications over, and I can "push" my updates to you, and then we can happily write code together.

The name GitHub implies its positioning - to provide a home for these scattered code libraries. Although we always say that decentralization is good, the code still needs a reliable storage place. So we built GitHub into a "home for developers" where everyone can collaborate and communicate.

02 In the future, programmers will be like conductors, leading an AI team

Question: How old are your sons? Did you teach them programming, or did they learn it by themselves?

Dohmke: They are 13 and 10 years old respectively. Their path to learning programming is quite interesting. During the pandemic in the previous years, I encouraged them to start practicing typing on typing software, and now their typing speeds are faster than mine. Later, they came into contact with graphical programming tools such as MIT's Scratch and Lego Mindstorms, and they could complete programming by dragging and dropping logical modules.

My eldest son learned Python at school, and my youngest son learned it entirely by himself. This reminds me of my own experience of learning programming in the 1990s. At that time, I could only rely on books and magazines, while they now have the entire Internet as a resource. Even without AI, when they encounter problems, they will go to YouTube to find tutorials, such as using Python and PyGame to make game animations.

Interestingly, they also learned how to solve the Rubik's Cube through YouTube videos. This process of continuous trial and error is actually similar to learning programming. Later, with the help of Copilot, it became even more convenient. It can explain code details in multiple languages such as English and German. This learning method that crosses language barriers makes programming education more inclusive, especially for children in non - English - speaking countries. Now, even children aged seven or eight, as long as they are interested in programming, can directly ask Copilot "how to make a Snake game" and get a complete guide from creating files to running and debugging.

I think children have more advantages in learning programming than adults because they are naturally more willing to keep trying. Adults tend to lose patience when encountering problems or give up due to lack of time, but children will persistently ask Copilot or other AI tools for help. They will repeatedly modify the code based on the AI's feedback until the program can run successfully.

Question: What is the scale of code creation on the GitHub platform?

Dohmke: Although we cannot give an exact figure, we can observe this huge code ecosystem from several dimensions. Currently, GitHub has 150 million registered developers, and the number of Pull Requests generated each year reaches hundreds of millions. Even if each Pull Request only contains 10 lines of code, the total amount easily exceeds one billion lines. In fact, most of the developers' work is to modify or refactor existing code, which makes the code activity on the platform even more significant.

From the perspective of AI participation, data from early 2023 showed that in files with Copilot enabled, 46% of the code lines were automatically generated by AI. Now, with the launch of intelligent agents, developers can even describe requirements in natural language, and AI will complete all code writing. This marks a fundamental change in the programming method - developers no longer need to directly write code but instead focus on writing precise requirement descriptions. It can be said that natural language is evolving into a new type of universal programming language, and developers are transforming into precise describers of requirements and architects of solutions.

Question: GitHub is the industry leader, but how many other platforms are you paying attention to?

Dohmke: Although there are indeed competitors like GitLab, GitHub is difficult to be shaken in the short term due to its large developer ecosystem.

However, interestingly, as I recently saw in a case - some air traffic control systems in the United States are still running Windows 95, and there are developers specifically maintaining these antique systems. This reminds us that even if a technology product dominates the market, we need to maintain a sense of crisis.

As an important part of the Microsoft ecosystem, we always maintain the anxiety of innovators. This alertness makes us continuously pay attention to industry changes, whether it is the evolution of DevOps or the revolution in development methods brought about by generative AI. Our technological assets are indeed remarkable: from the two major mainstream IDEs, Visual Studio and VSCode, to the.NET technology stack, and then to npm, the core infrastructure of the JavaScript ecosystem. But the more so, the more we need to be vigilant against the blind spots brought about by success, as warned in the book "The Innovator's Dilemma".

As a member of GitHub, I think the best thing is that everyone inside GitHub uses GitHub. From engineers to HR and even the legal team, all work processes are completed on GitHub. Our terms of service are themselves hosted in a repository on GitHub, and every modification is made through a Pull Request. For me, being able to create development tools for developers is the most ideal working state. We will continue to promote innovation on GitHub.

Question: What are you going to do next?

Dohmke: Obviously, it's intelligent agents. The current development status of programming intelligent agents is indeed both promising and challenging.

Taking the industry - recognized SWE - bench benchmark test as an example, the most optimal model currently can only achieve an accuracy rate of about 60%. Even if it is improved to 70% in the future, this figure still has obvious limitations - after all, the test set only covers 2000 problem cases from 12 Python projects. More notably, when extended to other programming languages, the accuracy rate usually drops sharply to about 30%.

Although these intelligent agents have shown great potential and can help developers with basic tasks such as task allocation, there is still a long way to go to achieve our ultimate vision - to let intelligent agents fully handle routine development work such as writing test cases, generating documentation, and fixing security vulnerabilities (true innovative R & D still needs to be led by humans).

We are also considering many other intelligent agent applications, such as the code review intelligent agent and the automatic repair intelligent agent we mentioned. At the Microsoft Build conference, we released the SIE intelligent agent, which can monitor your server and take actions based on the monitoring results, whether it is an error backlog or high server load.

Let's imagine a future scenario: developers will no longer work alone but will be like the conductor of a symphony orchestra, leading a digital team composed of various intelligent agents. This transformation not only brings efficiency improvements but also a fundamental change in the development method. Developers need to cultivate new core capabilities: first, they need to learn to make precise judgments - when it is appropriate to write code themselves and when it is more suitable to hand it over to intelligent agents. For example, sometimes assigning a micro - task of three lines of code to an intelligent agent is more efficient than writing it themselves. Second, they need to master the skills of task decomposition, breaking down complex requirements into sub - tasks that intelligent agents can understand. Most importantly, they need to learn to communicate with intelligent agents in precise development language, which will become an essential skill for future developers.

This really touched me. In actual development, the most difficult thing has never been to come up with ideas but to transform the ideas in my mind into executable code. In this process, I need to break the problem down into smaller and smaller parts. In the past few years, we have been working hard to lower the threshold for converting ideas into code; now, we are moving towards a higher level of abstraction. But no matter how technology advances, systematic thinking ability and structured design ability will always be the core competitiveness of developers. We will continue to optimize this process to enable developers to better collaborate with intelligent agents and complete more efficient development work.

Question: What is the current state of intelligent agent tools, and can ordinary people use them?

Dohmke: Nowadays, even people with no programming foundation can easily get started with tools like Copilot.

Let me first explain what a programming intelligent agent is - tools like Copilot (including similar products such as OpenAI's Codex) can directly generate code based on user descriptions. Especially in the field of web development, you can use it to quickly build a web page, develop a small web application, or even make a simple browser game.

The whole process is as simple as chatting with AI: input requirements, view the results, and if you are not satisfied, continue to adjust the prompt words until you get the ideal result, or simply start over.

This is similar to the working principle of AI painting tools: you can get better results by continuously optimizing the prompt words. But it should be noted that without basic programming knowledge, you will eventually encounter a bottleneck - relying solely on prompt words may not be able to achieve more complex requirements. When we promote these tools to hundreds of millions of users, we will ultimately face a real problem: either learn in - depth to become a professional developer or hire a professional development team to assist in achieving more complex functions.

On the other hand, there is what is called "vibe coding", which indeed makes programming easier, but there is a challenge that must be faced - developers ultimately need to understand whether the code generated by AI truly meets the requirements. This problem is particularly prominent in the enterprise - level development environment. Imagine that when you join a new project, in 99% of cases, you are taking over a code library written by others or participating in a large - scale project that has been maintained by dozens or even hundreds of developers. At this time, the programming intelligent agent shows its unique value.

Its working method is very interesting: you only need to describe in natural language the modifications you want to make to the code library, and Copilot will automatically start a virtual machine in the cloud, pull the code library, configure all the necessary development environments, and then implement these modifications based on the analysis results of the AI model. The whole process is like having a virtual development partner helping you - it will submit detailed modification requests like a human developer, clearly stating the content of each change, the implementation plan, and the technical solution.

But the key here is that these AI - generated modifications still need to be strictly reviewed by professional developers. You need to confirm that these changes are indeed what you want, ensure that they comply with the team's coding specifications, correctly use databases and cloud services, and most importantly, do not introduce new security vulnerabilities.

Indeed, for simple requirements such as building a wedding web page or quickly creating a Shopify e - commerce website for the Black Friday promotion, current AI tools are fully capable. But there is a key dividing line. When you want