The Log.
Field notes from the studio — what we're building, what we're seeing, and how senior engineers actually ship production software. New entries through 2026.
JEPA is not an LLM
The next serious architecture never read a word. JEPA learns from images and video of the physical world — the territory where language models genuinely can't see — and it's already driving robot arms through labs it has never visited. Meet the world model.
Read the note"AI" is a marketing name. You're using an LLM
Chess engines, spam filters and satnavs were all "AI" once. Today the label almost always means one machine: the large language model — a guessing game played a trillion times that became the most useful tool in software. How it works, without the mysticism.
Read the noteWhen your AI vendor pulls the model
Anthropic's most capable model was live for three days, then switched off worldwide by government order. If your product runs on a single frontier model, that's your risk too — and here's how we architect against it, plus a 7-question checklist.
Read the note//. more notes shipping through 2026
Got a system
to build?
Thirty minutes with a principal engineer. You leave with a clearer read on your problem — whether we work together or not.