Terence Tao — one of the most decorated mathematicians alive — just published a blog post describing how he used modern AI coding agents to port a couple dozen of his old, broken Java applets to JavaScript in a matter of hours, and to “vibe code” entirely new math visualization tools. Across the whole porting effort he reports finding only one minor bug in the agent’s output, and notes the agent flagged two bugs in his original 1999 code that he never knew about.
That’s the headline: agent-assisted coding has crossed a threshold where a domain expert with limited web-dev appetite can resurrect decades-old software and ship new interactive tools in an afternoon.
What Tao actually did
In a July 11, 2026 post on his blog What’s New, Tao describes migrating his old web pages and blog data to a more maintainable repository “using modern AI assistance.” As an experiment, he asked the agent to port his old applets — originally written in Java 1.0 back in 1999 for his complex analysis and linear algebra courses — to a modern, supported language. They landed on JavaScript.
The results, in his own account:
- Two dozen-ish applets ported in hours, all now functional again after years of being dead (modern browsers stopped supporting his version of Java long ago).
- A few graphical upgrades came for free — for example, his Besicovitch set applet is now colorized, versus the original monochrome version.
- A 1999 collaboration revived — a honeycomb applet he wrote with Allen Knutson, which he calls “a particularly tricky one to code by hand,” is working again.
He didn’t stop at porting. He also built new tools:
- A special-relativity visualizer he’d conceived in 1999 as “Inkscape, but in Minkowski space” — a project he’d abandoned back then because the code complexity got away from him. After “a couple hours of ‘vibe coding’” it now matches the vision he had 27 years ago.
- A Gilbreath conjecture visualization to accompany a paper and blog post he’d written earlier the same day, produced after “another few hours of conversation.”
The bug count is the interesting part
Tao is refreshingly blunt about the known failure mode. “Notoriously, LLM-based coding agents can create various blatant or subtle bugs in their code,” he writes. But in porting roughly two dozen applets, he could find only one minor issue: a drag event in a complex-analysis applet misbehaved when the user dragged outside the main box.
The more surprising line: the agent identified two bugs in his original code that he wasn’t aware of. His net assessment was that code quality ended up “a net wash.”
Here’s how the effort breaks down by his own description:
| Task | Approach | Reported outcome |
|---|---|---|
| Port ~2 dozen Java applets to JavaScript | Agent-driven port | Functional; some graphical upgrades |
| Special relativity visualizer (new) | “Vibe coding,” ~2 hours | Matches original 1999 vision; “alpha” |
| Gilbreath conjecture visualizer (new) | Guided conversation, ~few hours | Working; may become a paper supplement |
| Bugs introduced by the agent | — | One minor (drag-outside-box) |
| Bugs found in his original code | — | Two, previously unknown |
Note what’s not in that table: he doesn’t publish a formal error rate, a benchmark, or a comparison against writing the code himself. This is one expert’s lived experience across one weekend project, and he frames it that way.
Why this matters beyond one blog post
Tao is careful about scope, and that caution is the lesson. His applets are, in his words, “secondary visual aids rather than critical components of a mathematical argument,” so “the downside risk of such bugs is relatively low.” He applies the same reasoning to the new visualizations: they’re supplements, not mission-critical, so guided LLM generation is an acceptable trade.
That’s the mental model worth copying. The value isn’t “AI writes flawless code.” It’s that for a large class of low-downside, high-friction tasks — reviving dead demos, building throwaway visualizers, prototyping an idea you shelved because the plumbing was tedious — the economics have flipped. Work that wasn’t worth doing by hand is now worth doing at all.
It also lands during a moment when the community is actively re-litigating what lightweight, task-scoped agents are good for. If you’re weighing where agent-assisted coding fits your own stack, see our case for lightweight coding agents in 2026 and our take on how background agents change the way you think about coding.
The caveats Tao himself flags
He calls the relativity applet an “alpha” and explicitly invites feedback, noting that “especially given the LLM-generated nature of the code” there are likely still bugs and rough edges. He also mentions that the transcripts of his agent conversations were “edited down to remove a large number of tedious technical implementation reports” — a quiet reminder that “a couple hours of vibe coding” still involved a lot of back-and-forth, not a single magic prompt.
Tao doesn’t name which specific agent or model he used in the portion of the post describing the port, so we’re not attributing it to any one product. The takeaway isn’t about a tool brand. It’s that the workflow — expert supervises, agent grinds, expert spot-checks the low-risk output — worked well enough to bring 27-year-old code back to life.
FAQ
Q: Which AI coding agent did Terence Tao use? A: His post describes using “modern AI assistance” and “an AI agent” but the excerpt doesn’t attribute the port to a specific named product or model. We’re not going to guess — the notable part is the workflow and the outcome, not the brand.
Q: Did the AI introduce a lot of bugs? A: By Tao’s account, no. Across roughly two dozen ported applets he found one minor bug (a drag event misbehaving outside the main box), and the agent actually surfaced two previously unknown bugs in his original 1999 code. He called the overall code quality “a net wash.” That’s one expert’s experience on a low-stakes project, not a measured benchmark.
Q: Should I trust AI agents to port or write production code based on this? A: Tao’s own framing is the guide: he used agents for tasks where “the downside risk of such bugs is relatively low” — visual aids and supplements, not critical components. That’s the safe zone. For higher-stakes work, the same expert-supervision-plus-spot-checking discipline applies, just with far more rigorous review.
The smartest developers don’t pick one AI — they use them all. aiFiesta brings 9+ premium models into one chat for $12/mo. Your AI toolkit, simplified.