Typesafe.ai Launches Jev and System One Models for Fast Typed AI Inference
Typesafe.ai has introduced Jev, the first model in its new 'System One' architecture designed specifically for rapid, typed structured inference instead of free-form text generation. Jev evaluates inputs alongside defined query types to return structured results such as probabilities, choices, and confidence scores in 70 to 500 milliseconds. By foregoing open-ended text generation, System One models trade conversational flexibility for ultra-low latency structured decisions that software applications can consume directly. This approach offers a far more efficient alternative for programmatic tasks like sentiment scoring and category classification where text generation adds unnecessary overhead. Unlike generative LLMs that predict next tokens autoregressively, Jev processes queries using specialized primitives such as Choice, Score, or Noul. The model leverages technical methodologies like parallel sampling and Reinforcement Learning from Compiler Feedback (RLCD) to optimize fast structured output.
## BACKGROUND
Standard Large Language Models (LLMs) output sequences of text tokens, which introduces latency and parsing complexity when applications only need a simple classification label or probability. The concept of 'System 1' models draws inspiration from psychologist Daniel Kahneman's framework of fast, automatic human thinking, contrasted with slow, deliberative 'System 2' reasoning.