TypeSafe AI launches Jev, a transformer model that returns tagged decisions instead of text

TypeSafe AI, founded by one of the original ChatGPT developers, has released Jev, a transformer-based model that is not an LLM in the conventional sense and does not generate free-form text. Instead, developers send a state object and a map of tagged questions; the model returns tagged decisions with probabilities that code can branch on. The model is available only as a hosted API in early access behind a waitlist. Weights, parameter counts, and a self-hosting option have not been published.

The name references Daniel Kahneman's distinction between fast intuition (System One) and slow reasoning (System Two). TypeSafe argues that RLHF steered models toward human preferences, producing chat but also overconfidence and mode dropping — flaws that keep a human in the loop. Jev is built on a new stack: an undisclosed architecture, a parallel sampler, and RLCD, reinforcement learning for calibrated decisions. The architecture itself remains confidential.

A single endpoint handles everything: a POST with a body containing state, model, and a question map. The documentation defines three question types — Choice (up to 255 options), Noul (soft boolean), and Score (numeric). Questions run in parallel and in isolation against the same state, and the company says adding questions barely changes response time. SDKs are available for Python 3.10-plus via pip install typesafe-sdk, for