Former OpenAI researcher creates an AI that 'can't chat': Jev only makes judgments, up to 200 times faster

Beating AI News Flash: TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has launched a new model called Jev. It's somewhat like a super classifier that can understand natural language — it doesn't generate text, but only returns options, scores, and probabilities, specifically designed to make judgments for software.

Ordinary large models need to generate token by token, while Jev can produce multiple results in parallel. TypeSafe also used a new RLCD method to train its confidence, allowing the model to tell the program how confident it is while giving its judgment.

TypeSafe tested four automated workflow scenarios: security alerts, Agent review, invoice processing, and customer service. In this self-built evaluation, Jev was up to about 193.6 times faster than Claude Sonnet 5 and about 444.6 times cheaper than Claude Opus 5.

TypeSafe calls Jev the first "System One Model" (fast decision-making model), with the name derived from System 1 in Kahneman's "Thinking, Fast and Slow." It is specifically designed to handle rapid judgment tasks such as classification, scoring, routing, and next-step decisions.