TypeSafe AI Releases Jev System One Model for Typed Decisions

TypeSafe AI has released Jev, a novel "System One Model" that deviates from the text-generating capabilities of traditional large language models (LLMs) like those powering ChatGPT. Instead of producing human-readable text, Jev is engineered to return typed decisions accompanied by probabilities, enabling direct integration and branching logic within software applications. This approach aims to address perceived limitations of current LLMs, such as overconfidence and mode dropping, which TypeSafe AI attributes to Reinforcement Learning from Human Feedback (RLHF) tuning for human preference. The company posits that these flaws necessitate human oversight in current AI systems.

Jev operates on a distinct technological stack, incorporating a new architecture, a parallel sampler, and a specialized training methodology termed Reinforcement Learning for Calibrated Decisions (RLCD). While TypeSafe AI has not disclosed the specifics of its architecture or weights, the model is currently accessible via a hosted API in early access, requiring users to join a waitlist. The company has not yet offered a parameter count or a self-hosting option for Jev.

The Jev API is designed to handle all interactions through a single endpoint: POST https://api.typesafe.ai/v1/systemone. Users submit a request containing the current state, the model identifier, and a map of questions. The API supports three defined question types: Primitive, Asks, and Noul. Primitive asks allow for selecting one option from a predefined list. Asks enable rating against ordered levels, returning a score and probabilities. Noul questions are designed for binary true/false evaluations, returning a probability between 0 and 1. According to TypeSafe AI, these questions are processed in parallel and in isolation against the provided state, with the addition of more questions having a negligible impact on response time. A "Choice" question type supports up to 255 distinct options.

The "System One Model" concept is inspired by psychologist Daniel Kahneman's distinction between System 1 (fast, intuitive thinking) and System 2 (slow, deliberate reasoning). TypeSafe AI's Jev model is positioned as an AI that facilitates the latter, providing structured, probabilistic outputs that software can reliably act upon. This contrasts with LLMs that excel at generating human-like text but may struggle with the precision and reliability required for direct software control. The company's focus on calibrated decisions and probabilities suggests an effort to build AI systems that are more transparent and dependable for automated workflows and decision-making processes within applications.