Terence Tao, one of the most respected living mathematicians, used an AI coding agent to resurrect about two dozen of his 1999 Java math applets -- porting them to modern JavaScript in a matter of hours -- and reported a surprising verdict on the code quality: the agent found two pre-existing bugs in his original code that he had never known about, while introducing only one minor bug of its own. "It ended up being a net wash as far as code quality was concerned," he wrote.
Key facts
- Tao ported ~24 applets from 1999 Java to working JavaScript "in a matter of hours," with some graphical upgrades (one set is now colorized).
- Code-quality tally: the agent introduced one minor bug (a drag event misbehaving outside the main box) and surfaced two bugs in his original 1999 code he had not been aware of.
- He then built two new tools he had abandoned in 1999 as too complex -- a special-relativity visualizer and a Gilbreath-conjecture tool -- in a few hours each of "vibe coding."
- Posted July 11, 2026 on his blog, with full transcripts of the agent conversations linked. Tao's post, Hacker News.
The hook is who is saying this. Endorsements of AI coding tools are cheap; an endorsement from a Fields Medalist who carefully tallies the bugs is not. Tao did not gush -- he counted. One bug in, two bugs found, net wash. For a technology that critics dismiss as producing plausible-looking slop, a precise net-neutral verdict from a rigorous source is more persuasive than any benchmark.
Background for the non-specialist: back in 1999, interactive math demos on the web were built as Java applets -- small programs that ran in the browser. Java applets have been dead technology for years; every one of Tao's old demos had gone dark. Rewriting two dozen of them by hand in a modern language would be a tedious multi-day slog, which is exactly why they had stayed broken. He describes one, a honeycomb applet co-written with Allen Knutson, as "a particularly tricky one to code by hand" -- and it "came back to life."
The more striking half is the new work. Tao had a vision in 1999 for a special-relativity drawing tool -- he calls it "Inkscape, but in Minkowski space" -- and abandoned it because the Java was too complex. "After a couple hours of vibe coding with an AI agent, I was finally able to generate an applet that matched the vision I had back in 1999." Think of it like finding a box of half-finished blueprints in the attic and discovering you now have a contractor who can build any of them over a weekend. The projects that were never worth the time suddenly are.
Hackers News commenters converged on the same use case. One wrote that using LLMs to build "the nice-to-haves that I've always wanted but never had time for is one of their great use cases. Visualizations are a perfect use case because they don't have to be perfectly architected, maintainable code." That is the operative insight: the value is highest exactly where the code is a means to an end, not a shipped product that must be maintained for years.
Why it matters: this is the highest-authority "it actually worked" datapoint of the day, and it reframes the coding-agent conversation. The interesting capability is not writing new code from scratch -- it is finding bugs in existing code, which is precisely what happened to Tao's 27-year-old applets. That echoes a separate report the same week in which an AI agent found a long-standing numerical bug in widely-used inference software.
The honest caveat is Tao's own, and it is the responsible boundary. He accepts the risk because "these are secondary visual aids rather than critical components of a mathematical argument." A visualization that is slightly wrong is a nuisance; a proof that is slightly wrong is a disaster. Tao is endorsing AI agents for the contained, checkable, low-stakes end of his work -- and pointedly not for the load-bearing mathematics itself. He also edited the linked transcripts to remove tedious implementation chatter, so they show the collaboration, not an unfiltered log.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)