著名数学家陶哲轩（Terry Tao）近期展示了如何利用现代人工智能编码代理，将早在 1999 年编写的 Java 数学可视化工具迁移至 JavaScript。这一迁移过程仅需数小时即可完成 [1]。在审查生成的代码时，团队发现了旧版应用中两个此前未被察觉的缺陷，但 AI 生成版本的总体质量与原版持平 [1]。
借助这种被称为"vibe coding"（氛围编码）的方法，陶哲轩在短时间内实现了其于 1999 年构思但未完成的相对论可视化工具 [1]。此外，他还利用该技术开发了用于补充论文的 Gilbreath 猜想交互式可视化工具 [1]。尽管大语言模型生成的代码可能存在细微缺陷，但陶哲轩认为此类工作的效率极高且风险可控，未来计划将其作为论文辅助工具使用 [1]。
Famous mathematician Terence Tao has utilized modern AI coding agents to migrate a Java mathematical visualization tool originally written in 1999 to JavaScript [1]. The migration of the application, which was over two decades old, required only a few hours using these tools [1]. During this process, two previously undetected bugs were discovered within the legacy codebase [1], yet the overall quality of the AI-generated code remained comparable to human standards [1].
Tao also employed "vibe coding" techniques with modern agents to develop new visualization tools for relativity and the Gilbreath conjecture in just a few hours [1]. These projects included an interactive tool for the Gilbreath conjecture intended as supplementary material for his paper, as well as a realization of a relativity visualizer that had been conceptualized but never completed since 1999 [1]. Although large language models (LLMs) can produce code with minor defects, Tao assessed the efficiency and risk profile to be highly favorable [1]. He plans to continue using these AI agents as assistants for his future research papers [1].