Summary
Sam Witteveen examines Jev, a new ‘system one’ model from Typesafe AI, a company founded by former OpenAI researcher Dio Almeida, a top author on the original InstructGPT paper. Unlike reasoning-heavy chatbots, Jev never generates free text at all—it takes unstructured input plus a typed question (choice, score, or null/yes-no) and returns a direct answer with a confidence probability, aimed at the huge number of simple classification decisions embedded in software.
The video explains why this matters: many real-world AI tasks, like sorting support tickets or flagging risky agent outputs, don’t need multi-minute chain-of-thought reasoning, just a fast, cheap verdict. Witteveen tests Jev’s API directly, running sentiment scoring, yes/no probability questions, and choice-based classification, noting costs as low as 0.0014 cents per call and observing how confidence shifts based on input specificity.
Viewers get a clear technical walkthrough of a genuinely new model category positioned against the reasoning-model trend from labs like OpenAI and Anthropic, including a comparison to Daniel Kahneman’s System 1/System 2 thinking framework, and a practical sense of when a classifier model like Jev could replace expensive LLM calls in production software.
📺 Source: Sam Witteveen · Published September 18, 2026
🏷️ Format: Deep Dive
