Sam Witteveen demonstrates TypeSafe AI's Jev model, explaining its System 1 design for rapid software decision-making across choice, score, and boolean questions. He tests classification and routing tasks, reviews claims regarding hallucination and token-free latency, and critically evaluates whether typed outputs prevent incorrect decisions.
Original by Sam WitteveenClassificationIntermediate16 min 18 secPublished Source reviewed
Before you press play
What you’ll find in the video
Jev functions as a typed System 1 model supporting three decision types: choice selection with probabilities, numeric scoring, and boolean Noul probabilities.
Because Jev produces classification outputs in a single forward pass without autoregressive token generation, responses return with low latency and zero output token fees.
TypeSafe AI's claim that Jev cannot hallucinate refers strictly to structural schema adherence rather than decision correctness, as the model can still select incorrect answers.
Worth knowing
Gemini-assisted video/transcript review. Schema adherence does not guarantee accurate decisions, and claimed training methods like RLCD lack published architectures or independent benchmark verification.