Terry Tao recounts using a large‑language‑model coding assistant to migrate a collection of Java 1.0 applets—originally built for complex analysis and linear algebra courses—into modern JavaScript. The AI completed the port in a few hours, restoring functionality and adding visual upgrades such as a colorized Besicovitch set display. During the migration of roughly two dozen applets, the agent introduced only one minor drag‑event bug and flagged two pre‑existing defects, resulting in net‑positive code quality.

Encouraged by the smooth transition, Tao tasked the same agent with creating a new special‑relativity visualization, a concept he abandoned in 1999 due to Java’s complexity. Within a couple of “vibe coding” sessions, the AI produced a working app that matches his original vision, now hosted publicly. He applied the same workflow to build an interactive tool for the Gilbreath conjecture, attaching it to a recent blog post and paper.

Tao argues that AI‑assisted development is suitable for secondary visual aids, where occasional bugs pose limited risk. The experiment suggests that LLM coding agents can accelerate legacy code modernization and prototype scientific visualizations without extensive manual effort.