Terry Tao Let AI Coding Agents Rebuild His 1999 Math Applets
In 1999, Terry Tao spent his spare time hand-coding Java applets so students could watch honeycombs and fractal sets move on screen. This month, with those applets broken by decades of browser updates, the Fields medalist had a coding agent rewrite nearly two dozen of them and build two entirely new ones.
Tao's applets, which visualize Besicovitch sets and the honeycombs he co-designed with mathematician Allen Knutson, had stopped working on modern browsers entirely. An AI coding agent rewrote roughly two dozen of them in JavaScript, plus a new special-relativity sketchpad and a supplement for his recent Gilbreath conjecture paper.
Not everyone in the thread was charmed: one commenter said the more Tao talks about AI, the more he suspects an undisclosed conflict of interest, backed by a link to a Reddit thread.
Two dozen dead Java applets from 1999, working again
Terry Tao's interest in computer-assisted mathematics did not start with large language models. In a post on his blog, he traces it back to 1999, when he taught complex analysis and linear algebra and, in his spare time, wrote roughly two dozen Java 1.0 applets so students could see mathematical objects move instead of just reading about them: a colorized Besicovitch set (a needle-thin set that packs a line segment into every direction while covering zero area), and a honeycomb applet he built with mathematician Allen Knutson to illustrate their joint work on eigenvalues of sums of Hermitian matrices. Those applets eventually stopped running at all, killed off by a quarter-century of browser and plugin changes.
This month Tao wrote up what happened when he handed the rewrite to a modern AI coding agent instead of doing it by hand again: it ported the old applets to JavaScript, and it built two things he'd never gotten around to. One is a special-relativity diagram tool he'd abandoned in 1999 because the coordinate math was too fiddly to hand-code. The other is an interactive supplement for his recent paper on the Gilbreath conjecture — the still-open claim that if you repeatedly take absolute differences between consecutive prime numbers, the first number in each new row always comes out to 1.
The math stays his; only the pictures are outsourced
Tao is explicit that none of this touches how he checks a proof: the visualizations are illustrations, not arguments, and he treats a bug in one as a cheap failure to fix, not a research risk. It's a narrow, deliberate use of an agent, on a task he's spent parts of three decades doing by hand instead.
A reader's proof it wasn't AI-written: the post's own clunky sentence
Even that framing didn't stop the thread from getting suspicious of the framing itself. A commenter with more than 3,300 karma on HN went furthest, writing:
"The more Terry talks about AI, the more I'm starting to feel like Terry may have some undisclosed conflicts of interest."
He backed it with a link to an unrelated Reddit discussion, not evidence from the post itself, but the remark landed as a marker of how much scrutiny "a Fields medalist uses AI" now draws, independent of what he actually built.
A different reader turned the scrutiny on the writing itself, quoting the post's own prose straight back at it. A math and CS researcher who publishes paper explainers at leetarxiv.substack.com wrote:
"The article's awkward opening statement proves it wasn't written by AI."
He then pasted the sentence he meant: the post's own opening line, tracing Tao's interest in machine-assisted mathematics back to 1999 in almost exactly the words used earlier in this piece. In a thread already primed to suspect AI polish, a clunky human sentence became the alibi.
"It's a tool. Good for some things but not others"
Not every reader treated the post as a pitch for AI-assisted math. A commenter with over 7,000 karma on HN pulled out the line Tao had buried near the end and read it as the real thesis, pasting it directly:
"…as such [LLM-coded interactive] supplements are not mission-critical to the core of the paper, I again feel that the downside risk of using guided interaction with LLM agents to generate such visualizations is acceptable. It's a tool. Good for some things but not others and generally not to be trusted."
That reading got independent confirmation from someone actually running the same playbook in a classroom. recursivedoubts, creator of the htmx library, described using LLMs to build teaching visualizations he'd wanted for years but never had time to code by hand. As an example, he linked to an 8-bit computer simulator built for one of his courses:
"I wanted a simplified 8-bit computer to complement the 16-bit teaching computer I use and designed this in a few days with the help of claude"
That detail is more specific than anything Tao's own post discloses — Tao never names the agent he used, but a stranger replicating the exact pattern names Claude outright.
Not everyone bought the genre at all. An HN commenter waved off the whole category with a shrug:
"I always enjoy these 'domain expert has fun using AI to do something in their domain' articles. But it's always a hobby project, never something serious."
Tao's own post agrees with half of that: he never claims these visualizations are serious in the sense of load-bearing. The disagreement, in the end, isn't about whether AI-generated code belongs near a math paper. It's about how small the agent's job has to stay before a mathematician will hand it over.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
The Math That Powers Modern AI Was Invented Three Times, Separately
Eugenio Beltrami found it in 1873. Camille Jordan found the identical result a year later, in France, without knowing about Beltrami's paper. By 1889, James Joseph Sylvester had found it a third time: the same decomposition that now sits inside recommendation engines, image compression, and the optimizers used to train today's neural networks.