Ex-OpenAI Researcher's Non-Language AI Model Wins Over Developers

ChatGPT broke Diogo Almeida's heart. The OpenAI researcher helped build the chatbot and invent reinforcement learning from human feedback, the training technique behind today's AI boom, but he was disappointed. "We have lightning in a bottle, and yet it is not useful," he told TechCrunch, arguing the problem was optimizing for human language when computers speak a different one.

Two years ago he left OpenAI to start TypeSafe AI. This week it released Jev, a transformer-based model that is not a large language model. It outputs probabilities, or "calibrated decisions," rather than text. Skipping language makes it cheap, fast, and unable to hallucinate, since users define outputs in advance. Output tokens are free; input tokens are metered by the billion.

Demand briefly overwhelmed the company's API. Vercel software engineer Pranit Sharma said replacing OpenAI's Luna with Jev for a safety classifier gave results 5 to 18 times faster and more accurate. Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate at classifying business emails but 10 to 20 times pricier, and praised Jev's real probability scores.

Jev can also check misbehaving LLM agents and route workloads in real time, said Armin Ronacher, CTO of Earendil. Almeida named the model after William Stanley Jevons, the 19th-century economist whose paradox described how falling costs drive greater use. Jevons died in 1882.

Almeida calls Jev a "System One model," trained only on synthetic data via "reinforcement learning from calibrated decisions." Asked if TypeSafe is a frontier lab, he said, "the main product of Frontier Labs is fear or hype. I would like our main product to be intelligence."