TypeSafe AI debuts Jev, a model that outputs calibrated probabilities instead of text

TypeSafe AI, a startup led by former OpenAI researcher and RLHF co‑inventor Diogo Almeida, unveiled a new model called Jev on September 18. Unlike conventional generative models that directly produce text or images, Jev is a transformer designed to output calibrated probabilities—numerical confidence scores over possible outcomes. The company pitches it as an engine for decision‑making systems and risk assessment tools where it is crucial not just to pick an answer, but to know how sure the model is, addressing long‑standing concerns about opaque, overconfident AI behavior.

Jev is meant to be paired with existing large language models, acting as a kind of "safety and judgment layer" in high‑stakes domains such as healthcare, finance, and safety‑critical infrastructure. If its probability estimates remain well‑calibrated in the messy real world, it could help operators set thresholds, trigger human review, or comply with emerging regulations that demand explainability and quantified uncertainty. At the same time, Jev will face scrutiny over how rigorously its reliability is validated and whether its performance justifies adding another specialized component to already complex AI stacks.