From: redpointai
GitHub is the world’s largest source code host and a central part of the developer ecosystem [00:00:00]. GitHub created Copilot, the first AI coding assistant [00:00:06]. It is currently used by over 1.3 million developers and more than 50,000 enterprise subscribers [00:00:10].
The Genesis and Success of GitHub Copilot
Thomas Dohmke, CEO of GitHub, views Copilot as the “coolest part” of his job in the age of AI [00:01:13]. Copilot was developed over four years ago [00:01:19]. Initial statistics on how much code Copilot was writing were so high that the team was asked to verify the numbers [00:01:36]. When it was shipped, it performed even better than expected [00:02:27].
A significant shift occurred after the launch of ChatGPT in November 2022 [00:02:52]. Prior to this, there was skepticism about AI, but afterward, customer conversations universally shifted to adopting AI [00:03:02]. This marked the first time in a major platform shift that consumerization of a technology happened simultaneously with its professional adoption [00:03:15].
Lowering the Barrier to Software Development
One of the primary powers of Copilot is its ability to keep developers in their creative flow [00:04:24]. The original Copilot, without chat or agents, offered auto-completion that used large language models to suggest multiple lines of code, render test cases, or show how to connect to the cloud [00:05:01]. This feature prevents developers from switching between their editor and browser, thus avoiding loss of flow [00:05:03].
Natural language plays a crucial role in lowering this barrier [00:05:37]. Humans think in natural languages, not in code or deterministic machine language [00:05:55]. This allows users to ask questions about their code or have it explained back to them [00:06:03]. The ability to build a game in Brazilian Portuguese, for instance, is incredibly democratizing and will create new ways of thinking about software development [00:06:45]. This makes coding more akin to learning an instrument or drawing, accessible to anyone interested, not just aspiring professionals [00:06:59].
The Role of the Software Developer in the Age of AI
Thomas Dohmke believes that the role of a software developer in five years will look “very similar to today” [00:07:24]. The change will not be as rapid as some might overestimate due to the vast amount of existing code [00:07:33]. AI will primarily make managing complex systems easier and more accessible [00:08:21].
For new team members, AI will simplify learning a new codebase, stack, and institutional knowledge by allowing them to ask Copilot questions in natural language [00:08:42]. Copilot will handle less enjoyable tasks, enabling developers to focus on innovation [00:08:51]. For example, AI agents and “Copilot Autofix” can help developers address security vulnerabilities and accessibility issues, freeing them to work on new features [00:09:51].
The Rise of AI Agents
Copilot itself can be considered an agent, as it operates somewhat autonomously like a co-pilot flying a plane [00:10:20]. Developers are already accustomed to using agents, as seen with CI/CD (Continuous Integration/Continuous Deployment) pipelines that automate tasks upon code pushes [00:10:45]. Just as people appreciate dishwashers or robotic vacuums, developers look forward to more agents solving their problems [00:11:15]. For instance, security tools will evolve from just identifying issues to automatically helping fix them [00:11:32].
Broader Impact of AI
While some industries like translation might see significant human replacement due to AI’s real-time capabilities [00:11:55], Dohmke believes that software development will primarily “up-level” [00:12:21]. This means developers will continue to move up the stack, focusing on higher-level abstractions, much like the shifts from open source to cloud, or VMs to containers [00:12:39].
For new software developers, AI serves as a layer on top of code [00:14:46]. While human language makes it more accessible to start coding, an understanding of the underlying code is still necessary when things break [00:15:13]. The biggest superpower of Copilot is its ability to help developers unblock themselves [00:15:26]. Unlike in the early 90s, when resources were limited to books or computer clubs, today’s developers can ask AI to explain or fix code, fostering exploration and problem-solving [00:16:22].
Impact on the Open Source Ecosystem
Thomas Dohmke hopes and believes that AI will make open source even more dominant [00:18:04]. AI coding tools like Copilot, trained on vast internet information, naturally suggest the use of open source projects, increasing their adoption [00:18:50]. Even when asked not to use open source libraries, AI can produce better code than human-written code [00:19:01].
AI will also simplify the life of maintainers by helping manage issues and summarize discussions [00:19:22]. More importantly, AI will likely increase the number of contributors to open source projects [00:19:29]. It lowers the barrier to contribution by assisting with understanding codebases, writing test cases, and following project structures, thereby reducing the fear of rejection for new contributors [00:20:01]. This influence on consumption, maintenance, and contribution will lead to significant growth for open source [00:20:13].
A potential outcome is the consolidation around specific, higher-quality open source projects [00:21:03]. While this might reduce duplication and waste, it mirrors human behavior where developers often choose larger, well-maintained frameworks [00:22:06]. AI models will also use context from the editor and project to suggest relevant solutions within existing frameworks [00:23:03].
Evolving GitHub Copilot
GitHub is releasing fine-tuning capabilities for Copilot, a highly requested feature, especially by enterprise customers who deal with old codebases and specific internal libraries [00:24:32]. This allows users to select GitHub repositories for fine-tuning, providing a tailored Copilot experience without needing data science expertise [00:25:24].
Despite concerns about AI “hallucinating” or outputting imperfect code, Dohmke notes that human-written code is also rarely perfect and often requires modification [00:26:22]. Copilot’s strength lies in raising the average quality of code by understanding context and variable names, providing a better starting point than generic online snippets [00:26:50].
Future improvements will focus on better models, customization, and extensions [00:27:01]. Extensions are crucial because most projects involve multiple services beyond a single programming language [00:27:12]. These extensions will allow companies to integrate their unique platform engineering stacks and internal tools with Copilot, providing tailored answers and actions [00:28:02].
The rapid pace of AI development was an unexpected surprise [00:28:35]. What was once a “Horizon 3” research project quickly became the “current Horizon,” with the timeline for AI work shrinking to at best a six-month horizon [00:29:26]. The widespread adoption of Copilot, moving beyond early adopters into the early and late majority, has also been remarkable [00:29:48].
The Accelerating Pace of Innovation
Dohmke believes the pace of AI innovation is not only sustainable but accelerating because AI is now being used to build AI systems [00:30:43]. This will allow for more rapid development and the use of AI agents to clear backlogs in security, accessibility, and compliance [00:31:01]. Eventually, AI might even handle accessibility directly on the user’s device, eliminating the need for developers to implement it [00:31:27].
The goal is to leverage AI to prevent reintroducing security vulnerabilities or inadvertently downgrading software versions [00:32:00]. This acceleration will enable developers to spend more time on innovation and less on the less exciting, compliance-driven aspects of their work [00:32:27]. Innovation is expected across the entire stack, from chip design for faster Transformer models to responsible AI development and higher-layer applications [00:33:23].
Dohmke cannot identify any part of the development stack that will remain untouched by AI, as every part of an organization needs efficiency and productivity gains [00:34:03]. He views the term “artificial intelligence” as “scary” due to its negative connotation, preferring to focus on how AI improves lives and the world [00:35:12].
Market Dynamics and AI Adoption
Dohmke believes that the AI coding market will not be a “winner-takes-all” scenario [00:35:41]. Companies, like Microsoft and GitHub, must constantly reinvent themselves [00:36:15]. Competition is essential for innovation and progress [00:36:44].
Enterprises will likely use multiple AI models for different use cases [00:39:10]. Copilot itself already utilizes various OpenAI agent development tools models (e.g., 3.5 Turbo for auto-completion, 4 Turbo for chat, 4o for workspace) and internal smaller models for tasks like responsible AI pipelines or security vulnerability fixes [00:39:39]. Just as companies use multiple SaaS services, they will adopt various models for different engineering needs, including data collection, analytics, and even specialized vision models [00:41:05].
Advice for Startup Founders in the AI Era
Dohmke, having founded two companies and now leading GitHub, advises founders to play the “long game” and focus on the big picture rather than getting bogged down in small battles [00:43:30]. He emphasizes the need for continuous reconfiguration as a company, as there is no stable state; constant evolution is necessary to survive [00:44:29]. Companies must continually seek new product-market fit and adapt their teams and organizations along their growth journey [00:45:08].
A crucial lesson for leaders, especially as companies grow, is to “say no all the time” to most new ideas and investment areas [00:45:34]. The power of a startup lies in its focus, which must be maintained even with success and growth [00:45:44].
GitHub encourages developers to build extensions across the entire development stack, as more tools integrated into Copilot make it more powerful [00:46:17]. Areas like accessibility, security, and compliance offer significant opportunities for developers to create extensions that simplify lives [00:47:09]. The next wave of Copilot’s power will come from the mix of extensions and agents, moving beyond simple code completion to address the broader development stack, including vulnerability scanning, code review, multifile edits, and understanding entire codebases [00:48:40].
Debates and Future Outlook
The hosts and Dohmke debated the future role of developers and the impact on their numbers. While Dohmke maintains that developers will always be needed, the host believes the number of developers required within a company will decrease, especially in front-end and back-end tooling [00:55:34]. Manual and rote tasks will be automated, similar to the reduction in QA roles due to automation [00:56:38].
However, the barrier to building software and creating apps will significantly decrease, potentially leading to more micro-businesses and increased competition [00:57:59]. This unleashing of human creativity will be transformative [00:58:02]. It’s also expected that more non-traditional individuals will move into development roles [00:58:16].
Regarding AI’s broader implications, Dohmke remains highly optimistic and excited, dismissing fears of AGI [00:51:42]. He highlights the positive impact of AI in fields like biotech and climate change [00:51:29]. The hosts express concerns about deepfakes and the need to keep up with technology to mitigate their negative societal impact, although Dohmke suggests AI itself will protect against deepfakes [00:52:02].
A significant challenge and opportunity for AI is the maintenance and modernization of legacy codebases, such as COBOL, which power critical government and financial infrastructure [01:02:08]. With most COBOL developers retired, AI can either learn to maintain these systems or facilitate their realistic re-implementation in more modern technologies [01:02:31]. Startups are already addressing the use case of converting COBOL to modern code [01:02:41].