TypeSafe AI - Jev

TypeSafe AI takes the opposite research direction from mainstream chat models. Instead of optimising a language model for human preferences through RLHF — which the company argues introduces mode dropping, overconfidence and unreliability — it builds models designed to be consumed natively by machines. It calls this class System One Models and trains them with a new architecture, a new sampler and a new algorithm called Reinforcement Learning for Calibrated Decisions (RLCD). The first public member of this class is Jev, positioned for production automation.\n\nThe practical difference is in the output shape. LLMs produce text for people; Jev produces typed decisions that software can act on directly, each carrying a calibrated estimate of how confident the model is. That confidence signal is the core of the design: developers set thresholds for when the system acts autonomously and when it escalates for human review, then combine those decisions in code to build larger workflows. TypeSafe frames the results as reliable, fast, self-consistent and type-safe, and publishes benchmark comparisons showing Jev at roughly 193x faster and 444x cheaper than LLM workflows on System One tasks, with input pricing quoted at $42 per billion tokens.\n\nJev and the wider TypeSafe platform are aimed at technical teams building AI into products — screening, triage, classification and other places where a confident yes/no or judgement call is more useful than a paragraph. Access is currently via early access through the API console and documentation, and TypeSafe positions the offering as decision intelligence infrastructure for production software rather than another chat assistant.