TypeSafe AI released Jev, a model that generates no text at all

Diogo Almeida, one of the people behind the research that taught AI to talk (RLHF, InstructGPT, ChatGPT) and his startup TypeSafe AI released Jev.

Jev lives inside software and answers preset questions: route this request, score this record, is this output a jailbreak? Every answer comes with a confidence score.

TypeSafe calls it a System One model, after Kahneman's fast, intuitive thinking, as opposed to slow deliberate reasoning.

When Jev says 90%, it should be right about 90% of the time. A model that is 95% accurate but can't tell you which 5% it doubts can't run unattended. One that knows its own uncertainty can act when confident and hand off to a human when not.

Things to try with Jev right now:
LLM-as-a-Judge evaluation
Routing for agent harnesses harness
Scaling agent orchestration by enabling smarter subagent creation with SOTA classification capabilities
Enhance dynamic harness generation where structured outputs are key
The first three deliver insane ROI in cost and efficiency.

But the idea is not magic. Within days the open-source world rebuilt the interface: state in, typed questions, probabilities out.

Laya is the one people are installing. Open weights (Apache-2.0), encoder-sized (~300–400M), no text generation, local inference. The author argues he published related decision-model work in 2025. TypeSafe has not publicly answered. Fair read: the research lineage is older than the launch; TypeSafe’s contribution is a polished API, calibration story, and distribution.