Jev: The ChatGPT Co-Creator's System One Model Can't Talk

Diogo Almeida, a key researcher behind OpenAI's RLHF and InstructGPT, has unveiled Jev, a new 'System One' model developed by TypeSafe AI. Unlike traditional large language models, Jev is designed exclusively for decision-making rather than text generation. It operates by outputting structured choices, scores, or probabilities with calibrated confidence levels, rather than prose. The model is optimized for high-speed automation, boasting latency between 70 and 500 milliseconds and a cost-efficient pricing model of $0.042 per million tokens with no output fees. By utilizing a training method called Reinforcement Learning for Calibrated Decisions (RLCD), Jev aims to provide reliable, trust-based outputs for software pipelines. Currently in early access, Jev represents a shift in AI development, focusing on high-volume, low-latency judgment tasks that require precision over conversational ability, potentially transforming how developers integrate AI into voice agents and complex automation workflows.