Jev: TypeSafe's fast structured-decision model
TypeSafe AI, co-founded by a ChatGPT co-inventor, released Jev, the first 'System One' model designed for rapid, structured decisions inside software. Jev outputs calibrated, type-safe results in 70-500ms, is 20-200x faster and 40-400x cheaper than comparable LLMs, and does not generate text.
Key insights
Jev targets fast, reliable in-code decisions
TypeSafe AI launched Jev, a new category of AI model called 'System One,' designed for fast, structured decisions within software. Unlike LLMs, Jev does not generate text; it outputs type-safe, calibrated results for classification, scoring, routing, and extraction. It responds in 70-500ms, 20-200x faster and 40-400x cheaper than comparable models, with no output cost.
RLCD training method ensures reliability
Jev was built using a new training method called RLCD (Reinforcement Learning for Calibrated Decisions), developed by Diogo Almeida, a co-author of ChatGPT and GPT-4 papers. The method addresses reliability issues in automation scenarios where RLHF-trained models may be less dependable. Jev's architecture uses parallel sampling instead of sequential token generation, enabling its speed.
Jev serves as an 'intelligent if-statement' for developers
Jev is positioned as a component for developers to embed in system architectures, handling tasks like smart branching in workflows, large-scale data processing, real-time applications, and as a guardrail for LLM outputs. It is available in early access, with a waitlist for developers.