Jev introduces a new shape of LLM - System One, aka Decision Models
Jev takes text in and returns typed probability scores instead of generated text.
Willison frames TypeSafe AI’s Jev as a “decision model” suited to classification, ranking, labels, and similar tasks. Its API answers yes/no, choice, and numeric-score questions over a supplied state object, with questions evaluated in parallel. He notes the pricing is unusually low: input-only billing at $0.042 per million tokens, with no output charge. The tradeoff is opacity: Jev gives back floating point numbers, making bias and eval design especially important. Simon Willison's note
Willison frames TypeSafe AI’s Jev as a “decision model” suited to classification, ranking, labels, and similar tasks. Its API answers yes/no, choice, and numeric-score questions over a supplied state object, with questions evaluated in parallel. He notes the pricing is unusually low: input-only billing at $0.042 per million tokens, with no output charge. The tradeoff is opacity: Jev gives back floating point numbers, making bias and eval design especially important. Simon Willison's note
score 5