The Best Living Mathematician Just Used AI to Finish a 27-Year-Old Project

Kevin Hermes

The Best Living Mathematician Just Used AI to Finish a 27-Year-Old Project

When everyone else is arguing about whether AI coding agents are revolutionary or catastrophic, Terence Tao — widely regarded as the best living mathematician — just quietly used one to finish a project he’d abandoned in 1999.

The result is worth reading not because it proves AI is infallible, but because it’s the most honest, concrete piece of evidence about what these tools can actually do with legacy code that I’ve seen.

The setup

Back in 1999, Tao wrote roughly two dozen interactive Java applets for teaching complex analysis, linear algebra, and number theory. Visualizations of things like Besicovitch sets and honeycomb structures (the latter co-authored with mathematician Allen Knutson). They were moderately successful, and then browsers stopped supporting Java 1.0 and the applets went dark. Permanently, as far as anyone knew.

Fast forward to July 2026. Tao is migrating his website to a modern repository and hands his 27-year-old Java 1.0 code to Claude Code — Anthropic’s AI coding agent. Task: port everything to JavaScript.

It took a matter of hours. All 24 applets are now live again at his new site. The Besicovitch set applet is now colourised (it was monochrome in 1999). The honeycomb applet with Knutson — which Tao describes as “particularly tricky to code by hand” — is running again.

The bug scorecard: 1 in, 2 out

Here’s the detail that matters, and that most people discussing “AI code quality” ignore:

Across all 24 ported applets, Tao found exactly one minor bug in the AI’s output — a drag-event handler in one of the complex analysis applets that misbehaved when the cursor moved outside the applet’s bounding box. Niche, low-impact, easy to fix.

In return, the agent found two genuine bugs in Tao’s original 1999 code that he had never noticed in 27 years of those applets being publicly accessible. A faithful re-implementation, done by an agent with no domain context beyond the source it was translating, acted as an independent code review pass on mathematics code that had been sitting untouched since before most current CS undergraduates were born.

That’s not a benchmark score. That’s a real expert doing real code review on code they actually care about.

The project he couldn’t finish — until now

The porting exercise was impressive enough, but the more striking part of Tao’s blog post is what he did next.

In 1999, he’d had an idea for a special relativity visualisation tool — something he described as “Inkscape, but in Minkowski space” — and abandoned it because the code complexity was too much for him alone. Twenty-seven years later, after “a couple of hours of vibe coding” with Claude Code, he has a working version.

Tao published a curated transcript of the full Claude Code conversation, from first rough sketch to the final published tool. It’s a rare, unfiltered look at what an extended agentic coding session with a working mathematician actually looks like. Not a marketing demo — a domain expert using a tool to solve a problem they’ve had for three decades.

Why this isn’t another “AI will replace programmers” headline

Tao is careful about what he claims. He explicitly notes that “LLM-based coding agents can create various blatant or subtle bugs” and frames the applets as “secondary visual aids rather than critical components of a mathematical argument.” The downside risk of a bug, he says, is “relatively low” for visualisations that don’t carry mathematical proofs.

Contrast that with the broader discourse. Mathematicians running formal proof workshops on Fermat’s Last Theorem with AI have reported spending thousands of dollars a day on model tokens with mixed feelings about the trade-off. Midjourney founder David Holz posted last week that friends using the latest coding models are “feeling extremely productive and also extremely drained.”

Tao’s post is neither hype nor caution. It’s a domain expert using a tool where it works and being honest about where it doesn’t. He scoped the work to tasks where bugs wouldn’t matter. He checked the output. He noticed when the AI was better than he was in 1999, and said so plainly.

The version of AI adoption that doesn’t generate headlines

What makes this worth reading is what it accidentally becomes: a quiet manual for how professionals should actually use these tools.

Hand your decades-old, undocumented, self-contained code to an agent. Check the output. Use it where the risk is low. Don’t pretend the tools are magic, and don’t pretend they’re useless either.

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.

Sources:

Related Posts