GitHub CEO Thomas Dohmke discusses the transformative impact of AI on software development, drawing from his experience leading GitHub and scaling Copilot to over 1.3 million developers and 50,000 enterprise subscribers. He shares his perspective on how AI will reshape the developer role, the open-source ecosystem, and the broader technology landscape over the coming years.
The Magic of GitHub Copilot
Copilot was built over four years ago and has exceeded even internal expectations from launch, with early metrics on code generation so high the team had to verify they were accurate.
The launch of ChatGPT in November 2022 was a turning point: before it, enterprise conversations about AI were marked by skepticism; after it, customers immediately asked how to adopt AI as part of their strategy.
This represents the first major platform shift where consumerization happened simultaneously with enterprise adoption, as people experienced AI through consumer tools like ChatGPT and then demanded similar capabilities at work.
Keeping Developers in the Creative Flow
The original Copilot had no chat and no agents—it was purely auto-completion powered by a large language model, capable of showing 10 lines of code, rendering test cases, or explaining cloud connections.
The core value was keeping developers in their creative flow by eliminating the need to switch from the editor to a browser, search the internet, copy-paste code, and lose context.
Natural language adds another powerful layer: developers think in human languages, not deterministic machine language, so being able to ask questions about code or have code explained back in natural language is transformative.
This is especially democratizing for non-English-speaking developers—a child in Brazil can now say “I want to build a game” in Brazilian Portuguese, lowering the barrier to entry for millions worldwide.
The Future Role of Software Developers
Dohmke believes the developer role will look broadly similar in five years, because the world has enormous amounts of existing code that continues to grow rather than shrink.
Many critical systems still run on decades-old codebases (e.g., COBOL on mainframes in banking), written before modern practices like agile, test-driven development, or unit testing existed.
AI will make navigating complex systems far easier—the hardest challenge for any new team member is understanding the codebase, stack, and institutional knowledge, and Copilot will allow developers to simply ask questions in natural language.
AI agents like Copilot Autofix will handle the unglamorous but critical work—security vulnerabilities, accessibility issues, compliance—so developers can focus on innovation and the work they enjoy.
He draws a parallel to household appliances: nobody fears their dishwasher or complains about having a robotic vacuum, and coding agents will be received the same way.
AI’s Impact on Open Source
AI-powered coding will make open source even more dominant: when a developer asks Copilot to solve a problem, the model naturally suggests using established open source libraries because that’s what real developers do and what’s in the training data.
This was illustrated by an anecdote where a father and son needed to convert decimal to binary—the son wrote 150 lines of code, while Copilot solved it in two lines by importing an open source library.
AI will make life easier for open source maintainers by summarizing discussions, managing issues, and helping burn down security backlogs.
It will also lower the barrier to contributing: today, potential contributors are often discouraged by fear of rejection if they don’t follow the project’s conventions or write the right tests, and AI can help them get it right.
There may be consolidation around higher-quality, more popular open source projects, as AI systems consistently recommend the same well-maintained libraries, reducing the duplication and waste of having dozens of similar JavaScript libraries.
Fine-Tuning and Customization
Fine-tuning is one of the most requested features, especially from enterprise customers with old codebases and internal libraries that no public model would know about (particularly in C/C++, which lacks the package manager ecosystems of modern languages).
The tuning process allows customers to select a few repos from their GitHub organization, run the tuning process, and get a tailored Copilot that automatically activates in VS Code, Visual Studio, or JetBrains when those projects are checked out—no data science expertise required.
The Rapid Evolution of Copilot
Copilot started as a “GitHub Next” research project, originally conceived as a Horizon 3 initiative, but it quickly became a Horizon 1 product as AI adoption accelerated far faster than expected.
The pace of AI development has compressed planning horizons to roughly six months—anything beyond that requires rethinking and reconfiguring.
The adoption curve has crossed the chasm: it’s no longer about early adopters but about whether organizations are in the early or late majority.
Dohmke was surprised by how quickly Copilot went from “will this even work?” to mainstream enterprise adoption—the fastest he’s ever seen a technology make that transition.
The Competitive Landscape and Multiple Models
Dohmke does not believe the AI market is winner-takes-all, noting that even dominant companies eventually get disrupted and must reinvent themselves (pointing to Microsoft’s 50-year history and GitHub’s own reinvention through Copilot).
Enterprises will naturally use multiple models across their stack, just as they use multiple SaaS tools today—different models for different use cases, from code generation to telemetry analysis to vision tasks.
Copilot itself already uses multiple models: GPT-3.5 Turbo for auto-completion, GPT-4 Turbo for chat, GPT-4o for workspace, plus smaller models in the responsible AI pipeline for security scanning.
He sees room for multiple AI developer tools to coexist, just as companies combine multiple security vendors and best-of-breed applications today.
Advice for Founders
Play the infinite game: focus on long-term thinking rather than short-term battles, and continuously reconfigure as a company because there is no stable state.
Finding product-market fit is not a one-time event; companies must constantly find new product-market fit as they grow.
The true power of a startup is focus—saying no to most ideas—and this discipline must be maintained even as the company scales and succeeds.
He emphasizes that organizations and teams must evolve along the growth journey, and founders should resist the temptation to pursue every new opportunity.
Extensions and the Next Wave of Copilot
GitHub launched an extensions program to allow developer tool providers to integrate their platforms into Copilot, mapping the full enterprise stack into the natural language interface.
Customers are directly telling developer tools providers to build Copilot extensions because they want their entire platform engineering stack accessible from the IDE.
Priority areas for extensions include security, compliance, and accessibility, but Dohmke wants to see extensions across the entire development stack.
The next wave combines extensions and agents: auto-completion is a largely solved problem, and the frontier is now multi-file edits, codebase understanding, vulnerability scanning and fixing, code review, and predicting next steps.
A key use case is onboarding: a new employee at a company struggles to understand the codebase and submit their first pull request, and AI that deeply understands project context will be transformative.
Quickfire
Overhyped: AGI and the fear of AGI.
Underhyped: AI’s impact on biotech, climate change, and areas beyond traditional tech.
Does anything about generative AI scare him? Not as a parent, but he worries about deep fakes and their societal impact.
Most impressed incumbent besides GitHub/Microsoft: Nvidia, for its remarkable execution over the last two years.
Something he thought would work but didn’t: He was skeptical that chat would work, but it did.
Something he thought wouldn’t work but did: He didn’t believe GPT-3 could write syntactically correct code (getting commas and colons right) without a built-in compiler, but it did.
Most excited AI startup: Rove, plus his investment in Etched (a silicon startup), and several stealth-mode biotech companies.
If not running GitHub, what would he build? The next GitHub.
