Archives

Forced TDD, CLAUDE.md behavioral guidelines, token-saving skills: recent experiments measured them all, and most didn't survive contact with the data. Measure before you adopt a vibe.

The web, docs, and CLIs are being redesigned for a new primary reader. Bot traffic passing humans, llms.txt, AXI, and why I write internal docs in English for AI.

Cognition's new benchmark FrontierCode asks not 'does it pass the tests' but 'would a maintainer merge this PR.' Why the top model scores just 13.4% on the Diamond tier, and the gap between code that runs and code that merges.

Summer isn't a season to endure. A reflection on how the sharpened senses and slowed-down time inside the heat carve our memories.

Jack Clark, who designed the GPT-2 staged release in 2019, stands at the center of Project Glasswing in 2026. Tracing how the same logic returned with different evidence.

When Claude Code does work beyond coding, how should its name change?

The easier writing becomes, the more we forget what we wrote. A look at cognitive debt in the AI era and the paradox of digital natives.

A small cold and a great cold are waiting for us. This winter, I will keep gathering beautiful sentences.

I am glad that you are my sorrow. So when you grow up, be sure to become someone's sorrow.

If you truly want to be loved purely, you had better keep a few cookie crumbs in your pocket.

Which is more beautiful, memory or love? Memory, of course. Memory lasts longer, and that is why it is more beautiful.