Writings

2026-07-02

From Human-Verifiable Demos to Capability Reduction

Turning model-testing experience into reusable judgment. Why the intuitive way people test new coding models — build a frontend, a compiler, a driver — is a high-bandwidth probe, and how to make the judgment it produces structured and transferable.

AIModel EvaluationBenchmarksEssays

2026-04-20

Writing and Sustainable Human/AI Collaboration

Writing — the most important piece of sustainable human/AI collaboration.

WritingAIEssaysHuman-AI Collaboration

2026-04-19

Interim notes, on AI

Knowledge as a service, LLM as the runtime. On natural-language programming, distilling knowledge out of code, and managing the human brain's context in the AI era.

AINatural Language ProgrammingOpen SourceEssays

2026-03-22

The web was built for human eyes. Maybe it's time it spoke two languages.

GUI is for humans. Text is for LLMs. What if web components could speak both languages natively—without compromise?

AIwebsemantopen-source