# What Is Jev? The AI Model That Doesn't Generate Text

## Executive summary

Jev is a novel System 1 AI model from TypeSafe that makes fast, calibrated decisions by outputting probabilities rather than generating text. Unlike Large Language Models (LLMs) or reasoning models, Jev is designed for quick judgment calls—such as classifying support emails or determining urgency—making it faster and cheaper for structured software decisions. It complements LLMs by handling the rapid, automatic (System 1) parts of a workflow, while the LLM handles the slow, deliberate (System 2) tasks like drafting responses.

## Key takeaways

- Jev and System 1 AI: Jev is a System 1 model, which, following Daniel Kahneman's work, represents fast, automatic thinking (e.g., 2+2=4). It is built for quick judgment calls, unlike LLMs which are designed for text generation.
- Jev's Functionality: Jev takes structured input (state data and specific questions) and outputs calibrated probabilities for each option, rather than generating text. This makes it highly efficient for classification and selection tasks.
- Training and Calibration: Jev is trained using Reinforcement Learning for Calibrated Decisions (RLCD), ensuring that the model's reported probability accurately reflects its likelihood of being correct (calibration).
- Workflow Integration: Jev excels at initial triage (e.g., classifying a support email as a refund request or determining the necessary team) and can act as an AI guardrail, complementing LLMs by handling the rapid decision-making layer.

## Technical details

- Model Architecture: Jev is a System 1 AI model that outputs calibrated probabilities for predefined choices, unlike LLMs which generate tokens in a forward pass.
- Input/Output Mechanism: Jev takes 'state' (the data being analyzed, e.g., a support email) and specific questions as input. The output is a set of probabilities (e.g., 0.9 for 'Yes'), not text.
- Decision Thresholding: Probabilities can be used with a defined threshold (e.g., >0.9) to automate decisions (e.g., routing a ticket to a refund queue) or flag uncertainty (e.g., 0.1 to 0.9) for human review.
- Training Method: Jev uses Reinforcement Learning for Calibrated Decisions (RLCD), which rewards the model when its predicted probabilities are correct, ensuring accurate calibration.

## Practical implications

- Automating initial triage and classification of structured data (e.g., support tickets).
- Implementing AI guardrails to detect malicious inputs (e.g., jailbreak attempts).
- Creating efficient hybrid workflows where Jev handles fast, low-cost decisions, and LLMs handle complex, slow reasoning.
- Improving system reliability by quantifying model confidence via calibrated probabilities.

## Topics

AI Model Development, System 1 AI, Large Language Models (LLMs), Decision Making, Machine Learning Calibration, TypeSafe, Daniel Kahneman's Thinking Fast and Slow

Source: https://www.youtube.com/watch?v=YGgNBcIgI4s