What Is Jev and How Does System 1 AI Decide?

TL;DR

Jev is a system one AI that outputs probabilities rather than text to make fast routing and classification decisions. It uses calibrated probabilities for yes no and option selection and can be integrated with slower LLM workflows for complex tasks. This approach speeds up decisions while providing guardrails for automation.

Transcript

All right, I want to talk about Jev. Jev's a new AI model from a startup called TypeSafe. And it's a bit unusual because it's an AI model that doesn't actually write any text. It actually answers questions by picking from a list and then giving a probability for each one of those answers. Now type safe causes a new type of AI model. It's a system o... Read More

Key Insights

• Jev outputs probabilities instead of text to drive decision making, enabling fast and cost effective routing and classification tasks.
• The model embodies system one thinking, focusing on fast, automatic judgments rather than slow deliberation, and complements slower reasoning models like LLMs.
• Jev uses three question types in a single input: a yes no decision, a categorical team choice, and a severity score, all returned as probabilities.
• Calibration is central to Jev, with RLCD ensuring the probability outputs match real-world correctness frequencies.
• Jev can turn probabilities into actionable software decisions by applying thresholds to determine when to automate or escalate.
• Jev emphasizes a hybrid workflow where fast system one decisions feed into slower system two processes, creating an efficient AI pipeline.
• The model is designed to be degraded gracefully; uncertain cases can trigger human review rather than forcing a wrong automated path.
• Jev supports AI guardrails by providing probabilistic signals that can steer routing and response generation in broader AI systems.

Questions & Answers

Q: What is Jev and how does it differ from traditional text based models?

Jev is a system one AI model that does not generate text, instead it returns probabilities for predefined options. This enables fast, structured decisions such as yes or no, team assignment, or urgency level. Unlike text based models that generate language tokens one by one, Jev outputs a set of probabilities all at once, enabling quicker routing and automated handling with potential human review for uncertain cases.

Q: How does Jev use probabilities to guide decisions?

Jev assigns a probability to each possible option in its input questions. For a yes or no decision, it returns a single probability indicating the likelihood of yes. For categorical choices, it returns a probability per option. These probabilities are then interpreted through thresholds to decide automation versus escalation, providing a calibrated decision framework.

Q: What is system one thinking as described in the video, and how does Jev fit it?

System one thinking is fast and automatic, used for quick judgments. Jev embodies this by producing immediate probabilities rather than performing deep deliberation. It handles routine, high speed classification tasks in a workflow, leaving more complex reasoning to slower models like LLMs when necessary.

Q: What is RLCD and why is it important for Jev?

RLCD stands for reinforcement learning for calibrated decisions. It rewards the model when its probability outputs match actual correctness, ensuring that Jev’s probabilities are well calibrated. This enables reliable gating of automated actions and reduces the risk of overconfidence in incorrect decisions.

Q: How can Jev be integrated with larger AI systems?

Jev can be placed at the front of a workflow to quickly sort or classify inputs, producing probabilities that guide subsequent steps. For slower tasks like generating a full reply, an LLM can be invoked, and Jev can re classify the response or provide guardrails to ensure safe automation.

Q: What types of questions does Jev handle in a single input?

In a single input, Jev can handle a yes or no decision like refunding a charge, a team assignment decision among predefined options, and a severity or priority rating. All answers are returned as a set of probabilities corresponding to the given options.

Q: What are the potential limitations of Jev mentioned in the video?

Jev currently takes text as input and is not focused on math or counting tasks, so it may not handle numeric reasoning well. Like other AI models, it can be vulnerable to instructions hidden in data and may require complementary systems to address these edge cases.

Q: How does Jev relate to the concept of guardrails in AI systems?

Jev acts as an AI guardrail by providing calibrated probabilities that inform when to automate and when to escalate. This probabilistic signaling helps prevent over reliance on automated responses and supports safer, more controllable AI workflows in conjunction with larger language models.

Summary & Key Takeaways

• Jev is a fast, probability based AI model from TypeSafe designed to answer questions by selecting from a set of options rather than generating text. It represents a system one style of thinking and prioritizes speed and cost efficiency in decision making.

• The model handles input data and a set of questions, returning probabilities for each option, which enables calibrated decision making and thresholds for when to automate or escalate to humans.

• Jev is trained with a form of reinforcement learning called RLCD to ensure the probabilities it outputs are calibrated to reflect real correctness, allowing reliable gating of automated versus human reviewed actions.