← all news

Terence Tao rebuilds two dozen old applets with coding agents

AI · · · source (terrytao.wordpress.com)

Terence Tao, one of the best-known working mathematicians, spent some time rebuilding software he first wrote decades ago and wrote up what happened. Using LLM coding agents, he ported about 24 of his old Java 1.0 mathematical applets into modern JavaScript, among them a Besicovitch set visualizer and a honeycomb applet he first built with Allen Knutson in 1999. He also made new interactive tools from scratch, including a special relativity spacetime diagram and a visualizer for the Gilbreath conjecture. He published edited transcripts of the sessions, so you can see exactly how each app came together rather than take his word for it.

The interesting part is the error rate. Each of the more involved apps took roughly a couple of hours of conversational work, what he calls vibe coding, and across all of the ports he found only one minor bug, a glitch in a drag event. The agent also caught two genuine bugs in his original code from the 1990s, problems that had sat unnoticed for close to thirty years. So the migration was not just faithful, it was in two small ways an improvement on the source.

None of this is a benchmark, and Tao is careful to describe it as personal projects rather than a general claim. But it is a specific, checkable account from someone with no incentive to oversell, on a task many engineers have been putting off: reviving legacy code that still works but nobody wants to touch.

Why it matters

If you have old code you assumed was too tedious to modernize, this is concrete evidence that an agent can now do that port in hours and may surface real bugs in the original as it goes. Read the transcripts before you try it, so you know what the work actually looks like on a codebase you care about.

AgentsCodingEngineering