
Nobody Knows How Much Faster AI Makes You
AI coding productivity, honestly: METR's 19% slowdown now carries an out-of-date banner, its follow-up finds no signal, and I have no number either.
Notes from building AI products and shipping software

AI coding productivity, honestly: METR's 19% slowdown now carries an out-of-date banner, its follow-up finds no signal, and I have no number either.

Context rot is real, but the 2026 study everyone quotes is weaker than its headline. What the research measured, and why the fix is architectural.

I benchmarked 15 extraction strategies on 28 real job postings. Raw markdown won 27-0 over provider schema mode, which was silently returning summaries.

I built TinyPort to measure what one JavaScript import really costs: the tree-shaken, gzipped bytes a bundler ships. The architecture, and what broke.

React Server Components flipped a default in Next.js. The mental model, the boundary that bites, and which Next.js 16 changes actually matter.
Thanks for reading
Mostly notes from building AI products and shipping software. Have a look at what I'm working on, or just say hello.