Summary
Nate Herk tests Jev, a new decision-only AI model from Typesafe co-created by a co-inventor of ChatGPT, across 12 real-world use cases, including a Chrome extension that labels X posts in real time and a Bitcoin trading bot that makes second-by-second buy/sell decisions. Unlike typical chat models, Jev doesn’t generate text or reasoning tokens; it only outputs structured decisions like yes/no confidence scores, category picks, and numeric ratings.
The video explains Jev’s underlying training method, called RLCD (reinforcement learning for calibrated decisions), and demonstrates its practical speed and cost advantages, claiming it can be 20 to 200 times faster and 40 to 400 times cheaper than models like Terra, Luna, and Soul for classification tasks. Herk runs live tests classifying thousands of emails and YouTube comments, showing exact processing times and costs, such as sorting 1,000 emails in seconds for just a few cents.
The video also covers Jev’s limitations, including a small 64,000-token context window compared to million-token models like Claude or GPT, and its inability to summarize or reason. For developers building AI automations that rely heavily on classification and routing, this offers a grounded look at where a specialized decision model like Jev might fit into existing pipelines.
📺 Source: Nate Herk | AI Automation · Published September 19, 2026
🏷️ Format: Hands On Build
