The takeaway: In a week when OpenAI claimed its model proved a 50-year-old conjecture, the most useful datapoint about mathematicians and AI came from somewhere much quieter: Terence Tao's blog. The Fields medallist spent a few days with a coding agent resurrecting the interactive maths applets he wrote in Java in 1999 — dead for years since browsers dropped Java — and reports a precise quality tally: across roughly two dozen ported applets he found one minor new bug, while the agent found two bugs in his original hand-written code. "A net wash as far as code quality was concerned," in his words. Then he built the tool he'd abandoned 27 years ago because the code got too hard. It's the most level-headed account you'll read of what these agents are actually for — and where a working mathematician draws the risk line.

Bringing 1999 back to life

Tao has been building visual teaching tools since 1999, when he coded Java applets for his complex analysis and linear algebra courses at UCLA — visualisations of objects like Besicovitch sets and the "honeycombs" he studied with Allen Knutson. The applets were, he writes, "time-consuming to program", and when web standards abandoned that generation of Java they simply stopped working. A generation of hand-crafted teaching material went dark.

This week, as part of migrating his site to a more maintainable repository, he asked a coding agent to port the lot to JavaScript. It "managed to do so in a matter of hours, with all of my old applets now functional again", some with upgrades — the Besicovitch applet is now in colour, and the 1999 Knutson honeycomb applet, "particularly tricky" to code by hand at the time, is back to life.

The part worth clipping is his defect count. Coding agents, he notes, "can create various blatant or subtle bugs in their code" — but across the ported applets he found exactly one, a minor drag-event glitch in a complex-analysis applet. Meanwhile "the agent identified two bugs in the original code that I was not aware of, so it ended up being a net wash as far as code quality was concerned." A 27-year-old codebase by one of the world's best mathematical minds: two bugs nobody had noticed, found by the machine porting it.

Then the thing he'd given up on

Emboldened, Tao reached for a project he abandoned in 1999: a visualisation tool for special relativity — in his description, "Inkscape, but in Minkowski space", conceived before Inkscape existed. Back then, he'd started writing the Java for it and gave up: "the code complexity became too much for me." This week: "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." He's released it as an alpha and is openly soliciting bug reports, "especially given the LLM-generated nature of the code".

He also closed the loop on the same day's mathematics. Hours after posting a new analysis of Gilbreath's conjecture, he had the agent build an interactive visualisation to accompany the paper — and published a lightly edited transcript of the conversation, trimmed of the tedious implementation reports. He's minded to keep doing it: "I may add such interactive visualizations as supplements for future papers."

Where he draws the line — and why it matters this week

The through-line in Tao's post is a risk calculus, stated twice and worth quoting properly: "as these applets are meant to be secondary visual aids rather than critical components of a mathematical argument, the downside risk of such bugs is relatively low." Visual aids, teaching tools, paper supplements: yes, enthusiastically. The mathematical argument itself: that's a different risk category, and he doesn't put agent code inside it.

That distinction is exactly the one this week's bigger story keeps blurring. On Friday, OpenAI claimed its Sol Ultra setting had proved the Cycle Double Cover Conjecture — a claim that got its first qualified mathematician's endorsement this weekend, along with a citations complaint, and still awaits anything like full verification. The contrast isn't AI-sceptic versus AI-booster; Tao is plainly delighted with his week. It's about placement: one of the field's most celebrated figures using agents precisely where a bug costs little, while the industry's flagship claim sits precisely where a bug costs everything.

For everyone outside mathematics, Tao's tally is the transferable part. An agent porting a dusty, beloved, non-critical codebase in an afternoon — finding old bugs as it goes, introducing roughly as many as it fixes, cheerfully supervised by someone who understands the output — is not the singularity. It's just the best current answer to "what should I actually use these things for?" And it comes with a 27-year wait finally over: the honeycombs are moving again.