Old and new apps, via modern coding agents
Points and comments are a snapshot, not live.
Terence Tao uses AI agents to port old Java applets to modern JavaScript and create new interactive visualizations.
Terence Tao used modern AI coding agents to port two dozen of his old Java 1.0 applets (from 1999) to JavaScript, making them functional again with some graphical upgrades. The porting took hours per applet, with only one minor bug introduced (a drag event issue) and two bugs found in the original code. He also created two new apps via "vibe coding": a Minkowski space visualization tool (an abandoned 1999 project) and a Gilbreath conjecture visualization to accompany a recent paper. He plans to add LLM-generated interactive supplements to future papers, considering the downside risk acceptable for non-critical visual aids.
What commenters are saying
Commenters see Tao's use of AI as a compelling tool demonstration, with several noting the humbling parity it creates. A split emerges: some view LLM outputs as low-stakes, buggy prototypes unsuitable for production, while others argue that with better workflows and understanding of limitations, AI can be a real productivity boost. Specific points: one commenter notes that most "vibe coded" apps are trivial and buggy; another cites concerns about undisclosed conflicts of interest, linking to a Reddit thread. The thread also highlights Martin Hairer as a Fields medalist who creates professional software without AI.
Several replies emphasize the importance of checking AI output and reserving trust only for low-consequence tasks, with one commenter comparing it to a hammer that works for nails but not screws.