Terry Tao on Modernizing Apps and Visualizations with Coding Agents

Coding Agents Enable Domain Experts to Build Specialized Software

Modern coding agents are significantly lowering the barrier to entry for domain experts—such as mathematicians, scientists, and educators—to create functional software and interactive visualizations without needing deep professional software engineering expertise. By reducing the "activation energy" required to start and finish projects, LLM-based agents allow experts to translate theoretical concepts into interactive tools rapidly.

Modernizing Legacy Educational Software

Coding agents are proving highly effective at porting legacy code to modern web standards. A primary example is the transition of old Java applets—once staples of mathematics and physics education—into modern JavaScript applications.

While tools like CheerpJ have historically allowed Java bytecode to run in the browser via WebAssembly, the use of AI agents enables a "proper modernization" where the logic is rewritten in a native modern language. This process not only makes the content more accessible but also brings 30-year-old educational games and tools back to life in a way that is compatible with current browser environments.

Rapid Prototyping of Mathematical Visualizations

For researchers and educators, the ability to generate interactive dashboards and visualizations is one of the most productive use cases for LLMs. These tools allow for the creation of supplements to academic papers that help visualize complex mathematical objects (such as honeycombs or Besicovitch sets) without requiring the author to manage high levels of code complexity.

However, there are clear boundaries to this utility. As Terry Tao noted, even for highly intelligent users, code complexity can eventually reach a threshold where the project becomes unmanageable, leading to abandonment. This suggests that while LLMs are powerful for prototyping and supplements, they are not yet a total replacement for structured software engineering in mission-critical systems.

The "Latent Demand" for Specialized Software

There is a vast amount of "latent demand" for software in non-software-focused fields. Many experts have ideas for tools that would benefit their field but lack the time or specific coding skills to build them.

"If LLMs stopped improving today it would take us 10 years to catch up to the new software-writing abilities that have become available."

This shift suggests that the primary "moat" for software is shifting away from the ability to write code and toward the possession of massive data storage or specialized hardware assets. For local-running applications, the ability to disassemble, rewrite, and improve existing software using agents is becoming a tangible possibility.

Risk Assessment in AI-Generated Code

When using AI agents for technical work, the level of acceptable risk depends on the criticality of the output. For interactive supplements to a research paper, the downside risk of using guided interaction with LLM agents is generally acceptable because these tools are not mission-critical to the core mathematical proofs or findings. In these contexts, the LLM acts as a productivity tool rather than a source of absolute truth.

摘要: 數學家 Terry Tao 示範了現代程式代理如何讓領域專家快速構建互動式數學可視化並現代化舊有軟體，降低了專門軟體專案的啟動門檻。

標題: Terry Tao on Modernizing Apps and Visualizations with Coding Agents