Former OpenAI researcher launches Jev for faster AI decision-making

By defining outputs in advance, Jev is designed to reduce hallucinations.

TypeSafe also says the model is faster and cheaper, with output tokens free and input tokens metered at a billion per month rather than a million.TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has launched Jev, a transformer-based AI model that does not generate text as a large language model (LLM) does.Instead, it produces probabilities, which the company calls “calibrated decisions”, TechCrunch reported.Almeida, who helped develop ChatGPT and reinforcement learning from human feedback (RLHF) at OpenAI, founded TypeSafe two years ago.

He has argued that optimising AI for human language limits its usefulness in software automation, where structured outputs are often more valuable.By defining outputs in advance, Jev is designed to reduce hallucinations.

TypeSafe also says the model is faster and cheaper, with output tokens free and input tokens metered at a billion per month rather than a million.The model has attracted interest from software developers.

According to the report, Vercel software engineer Pranit Sharma said Jev delivered results five to 18 times faster and with greater accuracy than OpenAI’s model in a command-safety classification test.Bryo AI CTO Nikhil Mudholkar found Jev 10 to 20 times cheaper than Gemini in a business-email classification test, although Gemini was slightly more accurate.

He also highlighted Jev’s probability-based confidence scores for automated workflows.Jev could also complement LLMs by monitoring AI agents, detecting potentially unsafe behaviour and supporting model routing, TechCrunch reported.

Earendil CTO Armin Ronacher said its low cost and speed could make real-time routing and automated decision-making more practical.TypeSafe has not disclosed Jev’s architecture.

The company describes it as a “System One model” focused on intuition rather than extended reasoning and says it is trained exclusively on synthetic data using a technique called “reinforcement learning from calibrated decisions”.TypeSafe plans to develop additional versions of Jev for different modalities, with Almeida arguing that cheaper AI could enable its wider use across software.