Introducing Jev : a non-hallucinating classification model.

ChatGPT’s creator Diogo Almeida (inventor of the RLHF approach) just released a model called Jev via his company TypeSafe AI. It’s a non-hallucinating model that gives probability distributions across some predefined set of classes. According to Vercel, Jev is 5–18x faster than ChatGPT Luna 5.6 for classification tasks; meanwhile Bryo AI says that Jev is 10–20x cheaper than Gemini for parsing emails (albeit with slightly worse accuracy).

We can use Jev when we want a quick “yes/no” signal from our agent pipeline instead of a full response. In this sense, you can consider Jev as a “pre-screening” step before making Opus calls; handling simple tasks (classifying commands) at minimal cost while leaving the state of the art model to handle the rest, thereby saving overall runtime cost.