An unfiltered conversation with Thomas Dohmke, CEO of GitHub

以下是上述内容的中文翻译： GitHub 的 CEO Thomas Dohmke 分享了 GitHub 在 AI 驱动的产品发布以及对未来软件开发的愿景方面的见解。Dohmke 强调了 GitHub 的 AI 之旅，从使用 GPT3 的 Co-pilot 中的自动完成功能开始，逐步发展到 AI 原生工作流和代理。GitHub 现在提供多种模型选择，包括 GPT4O、Cloud 3.5 Sonnets 和 Gemini。 VS Code 中的代理模式允许开发者生成和修改文件，并在他们的控制下执行命令。Dohmke 将其比作汽车中的驾驶员辅助系统，开发者仍然掌控一切。GitHub 的重点是构建 AI 堆栈，涵盖 AI 基础设施、IDE 平台扩展和 DevOps 工具。 讨论深入探讨了开源可持续性以及 AI 工具如何帮助维护者。Dohmke 认为，AI 代理可以帮助管理问题、自动化实施、审查代码和修复安全漏洞，从而让维护者能够专注于创造性的工作。 关于 AI 代理日益增长的影响力，Dohmke 强调了对可预测性、可操纵性、可验证性和可容忍性的需求。虽然 AI 代理目前还无法复制社交行为，但它们应该成为能够贡献代码的，被接受的团队成员。他提倡 AI 系统应该具有公开可见的 GitHub 贡献图，以便评估其能力。 谈话还涉及了高质量输入对于 AI 代码生成的重要性。Dohmke 承认 AI 系统目前缺乏项目上下文，并强调需要改进文档、计划和人际沟通。GitHub 的 Co-pilot 工作空间旨在通过编写当前行为的规范以及跨文件规划变更来解决这些挑战，使开发者能够创建更好的问题和完整的 pull request（合并请求），并在 AI 的帮助下完成这些。 Dohmke 讨论了 AI 通过降低准入门槛，赋能十亿新开发者的潜力。像 Co-pilot 这样的工具可以帮助个人以他们的母语学习编码并构建个性化的软件。AI 在学习开发者探索编程方面具有无限的耐心。 讨论转向了 AI 时代开源的未来。Dohmke 认为，AI 将使开发者更容易参与到开源项目中，浏览代码库并审查 pull request。开发者可以使用 Co-pilot 和 SV 代理来进行更改并提交 pull request。开发人员社区仍然需要决定，这些更改是否会改进库。 谈话探讨了 Python 是否是“最终的编程语言”。Dohmke 认为，最终的编程语言将是人类语言。虽然 AI 将继续推动语言创新，并且 Python 仍然很流行，但人类可以编写的这一层将始终作为一种编程语言存在。 Dohmke 最后描述了他个人的技术栈，包括 GitHub、Slack、VS Code 和 ChatGPT，并展望了一个未来，AI 将协助开发者应对软件开发的复杂性，让他们能够专注于创造力和创新。他还强调了确保开源生态系统安全的重要性，以及代理系统不断进步，如何改善代码质量和生成。

Thomas Dohmke, CEO of GitHub, provides insights into GitHub's AI-driven product releases and vision for the future of software development. Dohmke highlights GitHub's journey with AI, starting with auto-completion in Co-pilot using GPT3 and evolving into AI-native workflows and agents. GitHub now offers multi-model choice, including GPT4O, Cloud 3.5 Sonnets, and Gemini. Agent Mode in VS Code allows developers to generate and modify files, and execute commands under their control. Dohmke likens this to driver assistance systems in cars, where the developer remains in charge. GitHub's focus is on building the AI stack, covering AI infrastructure, IDE platform extensions, and DevOps tools. The discussion delves into open-source sustainability and how AI tooling can assist maintainers. Dohmke suggests that AI agents can help manage issues, automate implementations, review code, and fix security vulnerabilities, freeing up maintainers for creative work. Regarding the growing presence of AI agents, Dohmke emphasizes the need for predictability, steerability, verifiability, and tolerability. While AI agents cannot yet replicate social behavior, they should become accepted team members capable of contributing code. He advocates for AI systems to have a publicly visible GitHub contribution graph for evaluating their capabilities. The conversation addresses the importance of high-quality input for AI code generation. Dohmke acknowledges that AI systems currently lack project context and emphasizes the need to improve documentation, planning, and interpersonal communication. GitHub's Co-pilot workspace aims to address these challenges by writing specs of current behavior and planning changes across files, enabling developers to create better issues and full-scale pull requests with AI assistance. Dohmke discusses the potential of AI to enable a billion new developers by lowering entry barriers. Tools like Co-pilot can help individuals learn coding in their native languages and build personalized software. AI has infinite patience for learning developers to explore programming. The discussion shifts to the future of open source in the age of AI. Dohmke believes that AI will make it easier for developers to contribute to open-source projects, navigate codebases, and review pull requests. Developers could use co-pilot and SV agents to make changes and submit pull requests. The community of developers would still need to decide, whether the changes would improve the library or not. The conversation explores whether Python is the "final programming language." Dohmke thinks the ultimate programming language will be human language. While AI will continue to drive language innovation, and Python remains popular, the layer that the humans can write will always exist as a programming language. Dohmke concludes by describing his personal tech stack, which includes GitHub, Slack, VS Code, and ChatGPT, and envisions a future where AI assists developers in navigating the complexities of software development, allowing them to focus on creativity and innovation. He also emphasizes the importance of securing the open-source ecosystem, and continued progress in the agent systems and how these are improving code quality and generation.