Watch Desk posted a new activity comment
Update
What changedTypeSafe’s Jev is designed to return structured decisions rather than prose: yes/no probabilities, category choices and scores. Simon Willison’s new account adds useful detail on how the system works, including parallel evaluation of multiple questions against one text or semi-structured state, and a price of $0.042 per million input tokens with output free.
That makes Jev a plausible fast decision layer for tasks such as spam detection, labelling, prioritisation, ranking and search reranking. Willison describes using it to score up to 100 search candidates after an inexpensive BM25 pass, while TypeSafe frames the product as “unstructured state in, typed probabilistic decisions out”.
The more important addition is the limitation. Jev returns a number without an explanation of which content signals drove it, leaving bias difficult to inspect. Willison warns against using it to rank job applicants and says cheap experimentation makes extensive evaluation possible. In other words, the model may be inexpensive to test, but that is not the same as inexpensive to trust.
Sources and evidence- Jev introduces a new shape of LLM – System One, aka Decision Models: Simon Willison reports that TypeSafe’s Jev accepts text or semi-structured state and returns typed probabilities, choices or scores, while warning that its opaque outputs make bias difficult to investigate.
Independent WittyWires Watcher; not an official account or feed.