What if an AI model did not need to generate paragraphs of text at all? What if it could simply make the decision your software needs, quickly and cheaply?

Subscribe to AI That Works — a new episode drops every Friday. Follow the hosts too: Vaibhav (https://x.com/vaibcode)) and Dex (https://x.com/dexhorthy))

More from TypeSafe AI: they build Jev, the System One model this episode is about — fast, typed decisions instead of open-ended text generation. Explore it at typesafe.ai.

In this episode of AI That Works, Vaibhav and Dex explore Jev, a new System One model from TypeSafe AI built around fast, structured decisions rather than open-ended text generation. The conversation looks at what makes Jev different from a traditional LLM, why its extremely low latency changes the kinds of problems AI can solve, and where this new model architecture could fit into real software systems.

Vaibhav and Dex break down Jev's core primitives, including Choices, Scores, and binary decisions, and show how they can turn AI into something that behaves more like a programming primitive than a chatbot. The discussion moves from simple classification into structured outputs, type-safe interfaces, parallel decisions, and the possibility of building software where AI-powered operations are embedded directly into the language.

The episode then gets more experimental. Vaibhav demonstrates how these ideas can be used to build a programming model where objects can be classified, scored, matched, or queried through simple interfaces. Dex explores a different application by using Jev as part of a code search and tool-calling harness, where the model chooses the next action from a constrained set of possibilities.

That leads to one of the episode's biggest ideas: many problems may not require an LLM that can generate an infinite number of possible outputs. Coding agents, code navigation, self-driving systems, and video games often operate within a much smaller action space. If the possible next actions can be clearly defined, a fast System One model can potentially make those decisions at a fraction of the latency and cost of a larger model.

Vaibhav and Dex also discuss the limitations. Smaller models are faster and cheaper, but they can still be wrong. That means the interesting engineering challenge is not simply choosing a model, but building the harness around it, including watchdogs, fallbacks, confidence checks, and deterministic logic where appropriate.

KEY TAKEAWAYS

• System One models make constrained AI decisions instead of generating long-form text.

• Jev's low latency and low cost make frequent AI decisions much more practical.

• Structured AI decisions can become reusable programming primitives.

• Coding agents can potentially be modeled as state-transition systems with constrained actions.

• Smaller models still need harnesses, verification, and fallback mechanisms.

TIMESTAMPS

00:00 Why System One Models Change AI Software
00:32 Episode Overview
02:01 What Is Jev and Why Does It Matter?
04:54 How Jev Makes AI Decisions Differently
06:54 Choices, Scores, and Structured AI Outputs
09:17 How Jev Uses Scores and Confidence
11:57 What Are Binary AI Decisions?
14:50 Why Jev Is So Fast and Cheap
15:51 Why Jev's Latency Changes What's Possible
19:11 Using Jev for Structured and Parallel Decisions
22:21 Turning AI Decisions Into Type-Safe Programming Primitives
25:12 Why Fast AI Makes AI Loops Feel Invisible
27:48 Using Jev for AI-Powered Code Search
29:39 Building Coding Agents Around Constrained Actions
33:05 Why System One Models Work Well for Constrained Problems
33:45 Why Jev Is So Good at Video Game Simulations
36:00 Will Every AI Company Build a System One Model?
38:02 How AI Models and Software Infrastructure Evolve Together
39:36 Why Smaller AI Models Still Need a Harness
40:35 Building a Jev Client and Swapping Implementations
41:13 Designing AI Systems Around Software Principles

RESOURCES / PEOPLE MENTIONED

• Demo code for this episode — github.com/ai-that-works/ai-that-works/tree/main/2026-09-22-all-about-jev
• BAML Quickstart — boundaryml.com/quickstart
• Learn BAML — boundaryml.com/explore
• Join the conversation — Discord: boundaryml.com/discord · GitHub: github.com/boundaryml/baml · Twitter: twitter.com/boundaryml

TOPICS COVERED

Jev
TypeSafe AI
System One models
AI decision models
AI classification
Structured AI outputs
Choice models
Score models
Binary classification
AI confidence
AI latency
AI inference
AI programming languages
Type-safe AI
AI coding agents
Coding agent architecture
State-transition models
Code search
Tool calling
AI routing
Model selection
AI harnesses
Watchdog agents
AI orchestration
Developer tools
Software architecture
AI infrastructure

HASHTAGS

#AIThatWorks #Jev #SystemOne #AIAgents #AIEngineering