Jev Explained: The Fast Decisions Before the AI Reply

An online store receives a message about a missing parcel, requesting assistance from the fulfillment team. To address this inquiry, the store must first determine the issue, locate the order, examine the delivery record, and decide whether a human should investigate further. While composing a helpful response is part of the process, the core of the job involves making critical judgments about the message received.

The Jev model, created by TypeSafe AI, addresses these judgment steps within workflows like this one. Jev assesses a message and returns a category, probability, or score that software can act upon. Once processed, a large language model (LLM) can explain the findings, draft a response, or handle more complex investigations. The separation of responsibilities between Jev and the LLM makes it easier to evaluate and understand the system's behavior.

Jev's functionality is built upon TypeSafe AI's System One model, which utilizes Reinforcement Learning for Calibrated Decisions (RLCD). This approach is designed for developers who supply information and questions, expecting structured answers and probabilities from Jev. Jev does not write emails, generate arbitrary code, or provide prose explanations of its reasoning.

Instead, it acts as a receptionist, routing the message to the appropriate department based on the model's judgment. This role is analogous to a smart if statement in programming, where code checks specific conditions, such as whether an order contains three products or contains expressions of disappointment. However, it's essential to remember that the model's judgment may be incorrect, necessitating additional code to handle uncertainty.

The name "System One" refers to the distinction made by Daniel Kahneman in his book "Thinking, Fast and Slow," which differentiates between fast, intuitive judgments and slower, deliberate reasoning. The "System One" label in TypeSafe's context is a product framing and analogy, without implying that the model thinks like a human brain.

The model's name also pays homage to William Stanley Jevons, who illustrated how making a resource cheaper can expand demand for it. Applied to AI, this concept could make evaluating many more messages, documents, or workflow steps practical, potentially driving broader adoption. The primary problem Jev aims to solve is the need for AI agents to make rapid decisions before performing more detailed tasks.

Traditional AI designs involve an LLM interpreting a request, selecting a tool, receiving the result, and deciding the next steps. ReAct, an influential research approach, interleaves reasoning and actions within this design. However, Jev offers a more efficient alternative by providing an interface tailored to small decision-making processes.

Instead of generating lengthy responses for every minor query, Jev offers a structured interface that addresses specific questions in parallel, saving time and resources. While parallel evaluation does not eliminate the need for retrieving order information, it does streamline the process by evaluating multiple small decisions simultaneously.

Ultimately, Jev serves as a decision-making tool that guides the LLM and application code in the appropriate direction, with the LLM handling specific tasks within the defined workflow.