Notes de chantier.
Des décisions d'architecture, le raisonnement qui les sous-tend, et les parties qui se sont révélées fausses. Écrit pour celles et ceux qui veulent comprendre comment une chose fonctionne avant de lui faire confiance.
Les articles sont publiés en anglais uniquement.
- Architecture6 min de lecture
Why a model has to be able to stay quiet
Fluency and correctness are separate abilities. Models got good at the first one much faster, and the gap is where the damage happens.
Lire - Models5 min de lecture
What a small model can actually do
The honest version: the line runs between what the context brings and what the model must recall unaided. Everything else follows from that.
Lire - Data4 min de lecture
Where the servers are, precisely
Most pages answer this with a flag. Here is the actual split, including the part that is not in Switzerland.
Lire - Tools5 min de lecture
An agent that works it out instead of guessing
Fifteen tools now, and every one of them exists because a language model sounds confident about something it cannot actually do.
Lire - Pricing4 min de lecture
Limits that have actually been worked out
Why our usage limits are not published as numbers, and how they are derived behind the scenes.
Lire