Jev and System One Models: Calibration Beats Accuracy

TypeSafe AI has unveiled Jev, a new "System One model" that is distinct from traditional language models. Unlike other models that generate text, answer queries, or engage in reasoning, Jev is designed to answer structured questions about an input in a single pass, with a probability attached to each answer. This unique approach focuses on calibration, allowing for more trustworthy decisions in production environments.

The model operates under three key principles: constraints are as important as features; it cannot generate free text; and it supports limited input options and data types. Pricing is set at $0.042 per million input tokens, with output tokens free, and access is currently limited to a waitlist. The product fills a gap in the market for faster, more accurate classification tools that can replace larger language models in certain scenarios.