GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Terence Tao Used AI Agents to Finish a 1999 Project He Abandoned — in Two Hours

On July 11, Terence Tao — Fields Medal 2006, widely considered the world’s greatest living mathematician — published a blog post describing how he used AI coding agents to do something he could not do himself in 1999: finish a special relativity visualization tool he started and abandoned when the code complexity became unmanageable.

The post, on his WordPress blog, is methodical. No hype. Just a working mathematician documenting what actually happened.

What He Did

In the past few days, Tao migrated his old web materials to a new repository. As a test, he asked an AI coding agent to port his collection of Java 1.0 applets — written in 1999 for complex analysis and linear algebra courses — to modern JavaScript.

Result: all 24 applets functional, including two that had been broken for years. Time: hours. The agent also added color to his monochrome Besicovitch set applet without being asked.

Code quality held. Tao found one minor bug — incorrect drag behavior outside a canvas boundary. The agent independently flagged two bugs in his original 1999 code. Net change: negative one defect.

The 1999 Project

The more striking part is the spacetime tool. In 1999, Tao had an idea: Inkscape, but in Minkowski space — a visualization for special relativity that would let users manipulate spacetime diagrams interactively. He started writing Java, hit a complexity wall, and dropped it.

Twenty-seven years later, a couple of hours of “vibe coding” with an AI agent produced what he originally envisioned. The app is now live. A full transcript of the agent conversation is published alongside it.

He then applied the same approach to a Gilbreath conjecture visualization — a mathematical structure he wrote about in a paper the same day — and had a working interactive tool in another few hours.

Why This Is a Signal

Tao is not an AI booster. He has been engaged with machine-assisted mathematics since at least 1999. His assessment is direct: the agent session was “painless enough” to make him consider adding interactive visualizations to future papers as standard supplements.

His framing on acceptable risk is precise: because these visualizations are secondary aids rather than load-bearing mathematical arguments, the residual error rate from AI-generated code is acceptable. He is not claiming the agents are correct; he is claiming the quality bar fits the use case.

That distinction — domain expert using AI tools where the error tolerance matches the output stakes — is what most AI productivity claims fail to make. Tao made it explicitly.

Key Numbers

  • Applets ported: ~24 (Java 1.0 to JavaScript)
  • Bug count: agent introduced 1, found 2 pre-existing
  • Time to complete the 1999 spacetime project: “a couple of hours”
  • Time to build the Gilbreath visualization from scratch: “a few hours”
  • Years the spacetime project sat unfinished: 27

The transcripts are published. The apps are live. This is reproducible evidence, not a press release.