Terence Tao is, by most accounts, the best living mathematician. Over the weekend he used an AI to finish a project he'd given up on in 1999.
In a blog post on Saturday, Tao described using an AI coding agent to port about two dozen of his old teaching applets from Java 1.0, the language he wrote them in back in 1999, into modern JavaScript. Java 1.0 stopped being widely supported years ago, so the applets had been effectively dead for a long time. In a matter of hours they were all running in a browser again, including one he originally built with the mathematician Allen Knutson that he says was especially tricky to code by hand the first time.
The scorecard is worth pausing on. Across the ported applets, Tao found one minor bug in the AI's code, a drag event issue in a complex analysis tool. The agent, in return, found two bugs in Tao's own 1999 code.
The bigger moment was a special relativity visualization he'd wanted to build back in 1999 and abandoned because the code complexity became too much for him. He describes the idea as basically "Inkscape, but for Minkowski space." A couple of hours of vibe coding later, he had a working version.
Tao is careful about where he thinks this fits. He notes that AI coding agents still produce "blatant or subtle bugs," but since these applets are meant as secondary visual aids and not as part of a mathematical argument, the downside of a bug is low. He says he might start including interactive visualizations as supplements to future papers, again with the caveat that they aren't load-bearing.
Not everyone shares his ease with the workflow. David Holz, the founder of Midjourney, posted on X (as covered by Business Insider) last week that his friends using the latest coding models are "feeling extremely productive and also extremely drained," and that something about it feels off. Mathematicians running formal proof workshops on Fermat's Last Theorem have reported spending thousands of dollars a day on model tokens with mixed feelings about the tradeoff.
What makes Tao's post worth reading is what it accidentally becomes: a quiet manual for how a serious professional should use these tools. He scoped them to work where a bug wouldn't matter. He checked the output. He noticed when the AI was better than he was in 1999, and said so plainly. The best living mathematician isn't worried about being replaced, and he isn't pretending the tools are magic. He's using them where they work and leaving them alone where they don't. That's the version of AI adoption that doesn't generate headlines, probably because it's boring, and probably because it's the one that actually works.
