A new kind of AI model from a ChatGPT inventor is thrilling developers
TypeSafe AI’s Jev skips natural-language output and returns calibrated probabilities for predefined decisions.
Diogo Almeida, who worked on ChatGPT and RLHF at OpenAI, says the model is aimed at automation tasks where text generation is a poor fit. Developers cited in the piece used Jev for safety classification, email classification, agent monitoring, and model routing, with reports of lower cost and faster responses than LLM-based setups. TypeSafe says the model cannot hallucinate in the same way because users define the possible outputs ahead of time, though users still have to decide how to act on confidence scores. TechCrunch AI's note
Diogo Almeida, who worked on ChatGPT and RLHF at OpenAI, says the model is aimed at automation tasks where text generation is a poor fit. Developers cited in the piece used Jev for safety classification, email classification, agent monitoring, and model routing, with reports of lower cost and faster responses than LLM-based setups. TypeSafe says the model cannot hallucinate in the same way because users define the possible outputs ahead of time, though users still have to decide how to act on confidence scores. TechCrunch AI's note
score 6