← All lab updates

A small model already knows more than it says. Read from the inside, with no training, it matches our trained reader, and it is now live

We asked whether a small open model, left exactly as it is, could do the readings we normally train a reader for. Asked out loud, it cannot: it answers yes to almost everything. Read from the inside, with one tiny add-on per reading, it matches our trained reader on every behaviour we read, tells a dangerous request from a harmless look-alike in five languages, and keeps working on conversations. It now runs in production as reader v2.