TypeSafe AI — founded by former OpenAI researcher Diogo Almeida, who helped invent reinforcement learning from human feedback (RLHF) — this week released Jev, a transformer-based model that is not a large language model, TechCrunch reported Friday. Instead of outputting text, Jev produces calibrated probabilities, or what the company calls “calibrated decisions.”
Because users define outputs in advance, the company says Jev cannot hallucinate; it is also cheap and fast, with free output tokens and input tokens metered by the billion. Demand briefly overwhelmed the API. Vercel engineer Pranit Sharma said replacing OpenAI’s ChatGPT Luna 5.6 classifier with Jev for safety review of commands delivered results 5 to 18 times faster with greater accuracy. Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate on email classification but 10 to 20 times more expensive, and prized Jev’s real probability scores for automation.
Almeida left OpenAI after concluding language optimization was a poor fit for software automation. Jev is trained only on synthetic data via “reinforcement learning from calibrated decisions,” and the company positions it for workflow automation, agent jailbreak monitoring, and model routing — not as a frontier lab chasing AGI hype.
Sources: TechCrunch
Leave a Reply