TypeSafe AI's non-language model Jev is winning over developers

Diogo Almeida, an OpenAI researcher who helped build ChatGPT and co-invent reinforcement learning from human feedback (RLHF), left OpenAI two years ago to start TypeSafe AI, aiming to fix what he saw as a fundamental limitation: LLMs are optimized for human language, which he argues isn’t the language computers actually need for automation. This week, TypeSafe released a new transformer-based model called Jev that is not a large language model — it doesn’t output text, but instead produces probabilities, which the company calls “calibrated decisions.” Because users define the outputs in advance, the model cannot hallucinate, and its output tokens are free while input tokens are metered by the billion rather than the million. Demand for Jev was high enough that the company briefly lost the ability to serve users from its API.

Jev appears most useful for software automation. Pranit Sharma, a software engineer at agentic-infrastructure company Vercel, said that after his company replaced OpenAI’s ChatGPT Luna 5.6 with Jev to run a safety classifier on commands, it got results five to 18 times more quickly and with greater accuracy. Bryo AI CTO Nikhil Mudholkar tested Jev against Gemini for classifying business emails; Gemini was slightly more accurate but 10 to 20 times more expensive, and Mudholkar said he was more interested in Jev’s confidence scores, which hand back a real probability useful for automating workflows. Jev can also augment LLMs by acting as a check on their behavior — Almeida sees users deploying it to track LLM agent traces and prevent jailbreaks — and Armin Ronacher, CTO of Earendil, said it could make real-time model routing economically viable in a way an LLM-based router would not be.

Almeida is tight-lipped about Jev’s architecture, which outside observers suspect is built on top of an open-weight LLM. The company describes Jev as a “System One model” focused on intuition rather than reasoning, trained exclusively on synthetic data using a technique Almeida calls “reinforcement learning from calibrated decisions.” For now, Jev stands alone as this kind of model, though Ronacher expects competitors to emerge now that its utility is apparent.

Why it matters

If cheap, non-hallucinating probability models like Jev catch on for automation and safety-checking tasks, they could pull a meaningful slice of enterprise AI spending away from general-purpose LLMs toward narrower, far less expensive tools.