A Fields Medalist is now shipping code with AI agents. Terence Tao used a coding agent to port roughly two dozen of his 1999-era Java math applets to JavaScript in a matter of hours — and the agent even caught two bugs in his original code. The Hacker News thread climbing past 50 points shows the developer world is paying attention.
What Actually Happened
Terence Tao — the mathematician behind the Green–Tao theorem and a Fields Medal recipient — published a writeup on using modern coding agents to migrate and build interactive mathematics visualizations. The details matter because they push back on the usual “AI code is buggy” narrative in a specific, credible way.
The port. He asked an agent to convert about two dozen old Java applets (some written in 1999) into modern JavaScript. The work, which would have taken him a painful amount of manual effort, was completed in a matter of hours. The applets are functional again, with a few graphical upgrades — for instance, a previously monochrome visualization is now colorized.
The bug math. This is the part developers should note. Across the port of ~24 applets, Tao found only one minor bug in the generated code (a drag-event edge case). More strikingly, the agent identified two bugs in his original 1999 code that he was not aware of. Net result: code quality was a wash, arguably better.
The new builds. Emboldened by the painless port, he then “vibe coded” brand-new visualizations — a special-relativity tool he had imagined back in 1999 (“Inkscape, but in Minkowski space”) and an interactive visualization for the Gilbreath conjecture to accompany a paper. Both went from abandoned idea to working applet in a couple of hours of conversation.
Why This Resonates on Hacker News
The thread carried real weight — 51 points and active discussion — because it reframes the agentic-coding debate. Tao is not a casual user hyping a tool; he is a domain expert who did the careful thing: he playtested the output, flagged it as “alpha,” and invited feedback on rough edges. He also scoped the risk correctly: these are secondary visual aids, not safety-critical systems, so the downside of a stray bug is low.
That risk framing is the actual lesson for developers. The community reaction mirrors what we see in our broader tracking of what developers say about coding agents in 2026 — people are most enthusiastic when agents are pointed at well-bounded, non-critical tasks where iteration is cheap.
The Honest Takeaway
Two things stand out:
- Agents are strongest as porting and scaffolding engines. Moving legacy code to a modern runtime, or generating a first working version of a long-dormant idea, is exactly where the speed payoff is largest.
- The human’s job shifts to review, not authorship. Tao still playtested, still caught the one real bug, still labeled the output alpha. The agent expanded what he could ship; it did not remove the need for taste and verification.
If you treat a coding agent as a junior collaborator that drafts fast and occasionally needs correction — rather than an autonomous engineer — the results look a lot like Tao’s: a net win, with the expert’s judgment as the final layer. That mental model is more useful than either blind enthusiasm or blanket skepticism.
For a deeper lens on why the model underneath the agent matters less than the harness around it, see your coding agent is a harness, the model is the commodity.
FAQ
Did Terence Tao really use AI coding agents to write code? Yes. In a July 2026 blog post, Tao described using a coding agent to port roughly two dozen of his 1999-era Java math applets to JavaScript and to build new interactive visualizations. He documented the process and shared edited conversation transcripts.
How many bugs did the agent introduce versus catch? Across the port of about two dozen applets, Tao found one minor bug in the generated code and the agent identified two bugs in his original code — a net improvement in code quality for that body of work.
What kind of tasks did he build with the agent? He ported legacy Java math applets to JavaScript and created new visualizations, including a special-relativity tool he had conceived in 1999 and an interactive visualization for the Gilbreath conjecture tied to one of his papers.
Sometimes Claude writes better code. Sometimes GPT does. aiFiesta lets you compare both instantly in one chat for $12/mo — plus 7 more premium models.