Summary
Cole Medin introduces Jev, a newly released model billed as the first of a new “system one” class of AI that makes decisions rather than generating free-form text. Trained with a new algorithm called RLCD (reinforcement learning for calibrated decisions) instead of the RLHF used for generative models, Jev takes in a situation plus multiple-choice questions and returns answers with confidence scores — useful for tasks like routing customer support tickets or classifying sentiment.
The video breaks down how Jev differs from asking an LLM for structured JSON output, walking through a real example where Jev routes a billing complaint with 64% confidence and scores customer frustration. Medin also shows how he’s used Jev inside his own AI coding workflows, including a PR triage system that uses Jev for cheap classification and routing before handing deeper review work to a full LLM, and demonstrates the Firecrawl MCP server for giving coding agents efficient web search.
The video closes by addressing criticism that Jev isn’t fundamentally new, acknowledging its conceptual overlap with existing classification models while arguing its speed, cost, and reliability make it a practical addition to agentic workflows. It’s a useful primer for developers curious about decision-focused models as an alternative or complement to general-purpose LLMs.
📺 Source: Cole Medin · Published September 21, 2026
🏷️ Format: Deep Dive
