From: redpointai
GitHub, the world’s largest source code host, is at the center of the developer ecosystem. They built GitHub Copilot, the first AI coding assistant, which is now used by over 1.3 million developers and more than 50,000 Enterprise subscribers. [00:00:00] Thomas Dohmke, CEO of GitHub, shares his insights on the future of AI in coding and scaling Copilot. [00:00:20]
GitHub Copilot’s Transformative Impact
The most exciting aspect of being GitHub’s CEO in the age of AI is seeing the widespread use and magical perception of Copilot. [00:01:13] Early statistics on Copilot’s code-writing capabilities were so high they were initially unbelievable. [00:01:35] The product surpassed expectations upon release, delivering more power than anticipated. [00:02:27]
A significant turning point was the launch of ChatGPT in November 2022, which shifted customer conversations entirely to AI, from skepticism to active adoption. [00:02:50] This marked the first time a major platform shift experienced such a consumerization of technology, with people using AI tools like ChatGPT at home and then seeking similar capabilities for their companies. [00:03:15]
Lowering Barriers and Enhancing Flow
AI, particularly tools like Copilot, is lowering the barrier to software development for people globally, opening up a new world for developers. [00:03:57] The core power of Copilot’s initial auto-completion feature was keeping developers in their creative flow. [00:04:24] By providing 10 lines of code or rendering entire test cases, it eliminated the need for developers to switch from their editor to a browser for internet searches, which often breaks their focus. [00:05:01]
Natural language plays a crucial role in this, as humans think in natural language, not code. [00:05:39] This allows developers to ask questions about their code or even have it explained back to them. [00:05:59] The ability to build a game in Brazilian Portuguese, for example, is incredibly democratizing, turning software development into something more akin to learning an instrument or drawing. [00:06:36]
The Future of Software Developers
The role of a software developer in five years is expected to look very similar to today, as the existing volume of code and legacy systems (e.g., banks still running COBOL on mainframes from the 70s and 80s) ensures ongoing maintenance needs. [00:07:24] AI will simplify navigating complex systems and make learning new codebases and institutional knowledge much easier. [00:08:13] Copilot will handle mundane tasks like proxy support or compliance, freeing developers to focus on innovation and enjoyable aspects of their work. [00:08:48]
Agents are a core part of Copilot’s strategy, with Copilot itself acting as an agent (like a co-pilot flying a plane). [00:10:06] Developers are already accustomed to agents through processes like CI/CD. [00:10:30] These agents will automate tasks like burning down security vulnerabilities and accessibility issues, allowing developers to focus on new features. [00:11:23]
While some industries like translation may see significant replacement of human roles, software development will likely see an “up-leveling” where developers move higher up the stack, building upon existing abstractions. [00:12:13]
The Learning Journey for New Developers
AI provides a new layer of abstraction on top of code, but the underlying deterministic machine code remains. [00:14:43] Just as developers use open-source libraries without always inspecting their internal code, AI provides suggestions. [00:14:13] Kids learning to code can use natural language to generate basic game structures. [00:14:50] However, when things break, they still need to dive into the codebase and understand the underlying logic. [00:15:06]
AI’s biggest superpower for learners is self-unblocking, allowing them to ask questions and explore solutions within the coding environment, a significant improvement over relying on books, magazines, or hoping to find a knowledgeable person. [00:15:22]
AI’s Impact on the Open Source Ecosystem
AI-powered coding is expected to make open source even more dominant. [00:18:02]
- Increased Usage: AI tools, trained on vast internet data, naturally suggest the use of existing open source projects, leading to their increased adoption. [00:18:41]
- Easier Maintenance: AI can help maintainers by automating tasks like burning down security issues, managing general issues, and summarizing discussions. [00:19:11]
- More Contributors: AI lowers the barrier to contributing to open source projects. Developers can more easily understand new codebases and adhere to project standards, reducing the fear of pull request rejection. [00:19:29]
This dynamic will lead to consolidation around high-quality open source projects, reducing duplication and waste, similar to how human developers choose established frameworks like React or Next.js. [00:21:29] AI models consider the existing project context (e.g., language, frameworks) to provide relevant suggestions, making them more effective than generic answers. [00:22:21]
Copilot Evolution and Future Outlook
Fine-tuning for Copilot is a highly requested feature, especially by enterprise customers who work with older codebases and internal libraries unknown to public models. [00:24:11] This allows for tailored Copilot experiences without requiring data science expertise. [00:25:10]
The “hallucinations” of AI models are not necessarily a bad thing in coding, as developers are accustomed to modifying imperfect code snippets from the internet. [00:25:41] Copilot’s strength lies in its ability to provide suggestions closer to the perfect answer by leveraging the context of the developer’s current code. [00:26:24] The goal is to raise the average quality and efficiency of code, not necessarily to achieve perfection from the outset. [00:26:42]
Accelerating Pace and Broader Impact
The pace of AI development is not only sustainable but accelerating because AI is now being used to build AI systems. [00:30:34] This allows for faster progress in areas like burning down security and accessibility backlogs. [00:30:49] AI could potentially handle accessibility concerns by developing devices that automatically adjust user interfaces for individuals with vision impairments, reducing the need for developers to manually implement accessibility features. [00:31:08]
Innovation is expected across the entire technology stack, from chip design to models and higher-level applications. [00:33:52] AI will bring efficiency gains to all parts of software development, helping combat the increasing volume of tasks and information developers face. [00:34:01]
Market Dynamics and Enterprise Adoption
The AI market is not a “winner takes all” scenario. [00:35:39] Companies like Microsoft and GitHub constantly reinvent themselves, and competition drives innovation. [00:35:54] Enterprises will likely use multiple models for different use cases within their IT and engineering stacks, just as they use various SaaS services today. [00:38:51]
“I think the as you as everybody is using multiple tools it’s kind of like you’re asking the question as are they are companies going to use more than one SAS service within their it stack or within their engineering stack and the answer absolutely is yes and as such you know they’re naturally going to have multiple models for multiple use cases” [00:39:54]
This will lead to a “multimodal world” where specialized models for tasks like code generation, testing, documentation, or even specific domains like biotech, coexist. [00:40:51]
Advice for Founders
For founders going up against incumbents, the advice is to:
- Play the long game: Focus on the big picture rather than short-term battles. [00:43:17]
- Continuously reconfigure: Companies must constantly adapt and evolve, as there is no stable state or single moment of “product-market fit.” [00:44:20]
- Say no: As a company grows, it’s crucial to say no to most new ideas and investment areas to maintain focus. [00:45:10]
Extensions and Future Development Areas
GitHub is actively promoting its extensions program, encouraging developers to build tools across the entire software development stack. [00:46:05] This makes Copilot more powerful by integrating with diverse platform engineering tools, allowing users to interact with complex systems via natural language. [00:46:30] Key areas for extensions include accessibility, security, and compliance. [00:46:58]
Beyond code completion, the next wave of Copilot will focus on agents that review code, refactor, perform multi-file edits, and understand entire codebases, significantly easing the onboarding process for new employees. [00:48:08] AI agents could also improve issue filing by ensuring bug reports and feature requests are more specific and context-rich. [00:50:17]
Overhyped vs. Underhyped in AI
- Overhyped: Artificial General Intelligence (AGI) and the fear associated with it. [00:51:19]
- Underhyped: The impact of AI on fields like biotech and climate change, which are further from traditional tech. [00:51:25]
While concerns about deepfakes and potential misuse exist, AI is also seen as a solution for detection and protection. [00:51:48]
The initial skepticism about AI’s ability to write functional code, particularly with correct syntax, was overcome by Copilot’s performance. [00:52:56] The surprise was not just the code generation but the ability to have a conversational interaction with the AI, heavily influenced by ChatGPT’s success. [00:53:36]
The Challenge of Legacy Code
A significant challenge and opportunity for AI is the vast amount of legacy code, such as COBOL, still powering critical systems in governments, financial infrastructure, and hospitals. [01:01:44] With most COBOL developers retired, AI could either learn to maintain these systems or become powerful enough to realistically reimplement them in more modern technologies. [01:02:15] Startups are already working on converting COBOL to modern code, which represents a significant use case for AI given the dwindling number of human experts. [01:02:37]