Introducing System One models and Jev
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev: a frontier-scale model that gives up text generation entirely and instead returns typed, calibrated probabilistic decisions, trained with what they call Reinforcement Learning for Calibrated Decisions. They claim parity with LLMs on fast structured-judgment tasks at roughly 100x the speed and efficiency, and that it cannot hallucinate because it never emits strings.