The code I fixed most over the past few months wasn't code that didn't work. It was code that worked fine and was wrong. Three cases, and what's left for the developer once AI takes the productivity.
The Latest Article
Benchmark scores went from 1.96% to over 70% while their meaning eroded. What SWE-bench measures, what it misses, and what I saw migrating this blog with Claude Code.
Not zip, not 7z, but a format from 1979. Why bundling first and compressing second makes npm packages 23% smaller, and why sealing them into one file is what makes integrity hashes possible.
Oxlint and Oxfmt, the Rust-based replacements for ESLint and Prettier. Measured speed on 500 components, plus two rules that caught real problems.
A year of free AWS S3 ran out and the bills started. Why I picked Cloudflare Pages, and the two things static export broke: next/image and headers().