Notes from the work itself
Short pieces on how we think — picking the structurally-right tool over the trendy default, what agentic thinking changes, what the three frontier labs are actually playing for, and what the last mile takes.
- 27 May 2026
Three games being played in AI →
Anthropic, OpenAI and Google look like competitors in the same market, but read their pricing pages and the strategies separate cleanly. One is playing depth on a single user class. One is playing breadth on a consumer brand. One is playing defence on distribution surfaces it already owns. A year in, the scoreboard has a clear leader — but only one of these games is fast.
- 25 May 2026
When the model feels off, check the plumbing →
If a model's output is drifting — wrong register, wrong vocabulary, reaching for ideas you'd already ruled out — that uncanny feeling is almost always validated. The fix is usually upstream of the prompt, in whatever context the harness did or didn't load.
- 23 May 2026
Don't dump the meeting transcript into the model →
Modern models record what a meeting decided reliably. The cost shows up the moment you ask one to do the work — a fifty-page kickoff transcript is a worse context for drafting auth requirements than two relevant pages would be.
- 22 May 2026
The structurally-right tool beats the trendy default →
Most AI builds reach for whatever's in fashion. We start from the shape of the problem — which is why we'll put a knowledge graph where everyone else reaches for RAG.