What Is Jev? The AI Model That Doesn't Generate Text

A lot of us first heard about Jev after former OpenAI researcher Diogo Almeida released it. If you have been using LLMs for a while, it was refreshing to see something genuinely new in the middle of what the last 6 to 9 months have felt like.

Most existing LLMs are optimised with Reinforcement Learning with Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR). System One and Jev use something different: Reinforcement Learning for Calibrated Decisions (RLCD), which is a mouthful of concepts right there. Jev promises to be a faster and cheaper alternative to LLMs for the cases where speed, accuracy and determinism matter most.

There is a lot already published about it, some of it exaggerated or missing the point.

This talk from IBM Technology’s Martin Keen is by far the clearest explanation of what Jev is and how it compares to regular LLMs. A must watch for anyone trying to understand this new type of model.