How the daily notes are made.

Short technical notes on what moved in AI - a model, a benchmark, a tool, a licence, a paper - written for engineers who have to build on it. At most one a day. They are not news coverage and not opinion columns; each one tries to answer one engineering question about one story.

A script reads the last day and a half of Hacker News, arXiv and Google News, and a model picks the one item with the most technical substance for engineers, passing over funding rounds, executive moves, lawsuits and marketing claims, and stories this site has already covered.

The note is drafted with Claude, Anthropic’s model (currently Claude Opus 5.5), from the full text of the sources it cites and nothing else. A second pass rereads the draft against the sources and removes or corrects any number, name, version, date, licence or quotation it cannot trace to them. Every source link is checked to be live before publishing. A Hacker News thread, where one exists, is read for what practitioners are worried about and is listed as discussion, never as a source of facts.

Since 8 October 2026 every note waits as a draft until I have read it and approved it; the note then says who approved it and when. Notes published before that date went up automatically after the checks above, and say only how they were drafted.

A note never says I built, ran, benchmarked or used the thing it describes unless it says so explicitly with the evidence. It reports what the sources say, and says when they are silent.

I build Tron-1B, an open-weight decision model, and Wayne and Omnitron, AI coding and agent harnesses. Any note about decision models or coding-agent harnesses says so in a line under its title.

If a note gets something wrong, email samir.s.sengupta@gmail.com. I correct it and date the change at the top of the note.

Every note is also published as plain markdown at its own address plus ".md", in full in the RSS feed, and in /llms-full.txt, with the same disclosure.