An early OpenAI researcher who helped invent reinforcement learning from human feedback has launched a new startup called TypeSafe AI. The company released a transformer-based model named Jev this week. The model departs from traditional large language models by avoiding text generation. Instead, Jev produces probabilities and calibrated decisions.
The founder, Almeida, left OpenAI two years ago to solve a perceived limitation in current AI systems. He concluded that optimizing for human language hinders automation because computers rely on a different format. Jev does not output human language. Because users define outputs in advance, the model cannot hallucinate. Its output tokens are free. Input tokens are priced per billion rather than per million.
Developer demand for the new model was immediate. Interest was so high that TypeSafe briefly lost the ability to serve users through its API. Software engineers currently view Jev as a cheaper and more robust way to add intelligence to code.
Pranit Sharma, a software engineer at Vercel, tested Jev against OpenAI's ChatGPT Luna 5.6 for reviewing commands for safety. Vercel observed results five to 18 times faster and with greater accuracy after replacing the OpenAI model with Jev. Nikhil Mudholkar, CTO at Bryo AI, tested Jev against Gemini for business email classification. Gemini proved slightly more accurate, but 10 to 20 times more expensive. Mudholkar highlighted Jev for providing real probability confidence scores.
Jev can also augment existing large language models. Armin Ronacher, CTO of Earendil, noted that users can establish probability thresholds to handle model outputs safely. Ronacher also suggested Jev could handle real-time model routing due to its speed and low cost.
The model is named after 19th-century economist William Stanley Jevons. His paradox describes how falling commodity costs lead to increased usage. Almeida hopes cheaper intelligence will lead to widespread software deployment.
Almeida keeps the architecture under wraps. Observers suspect it runs on an open-weight large language model. TypeSafe categorizes Jev as a System One model focused on intuition rather than reasoning. Training relies exclusively on synthetic data through a technique called reinforcement learning from calibrated decisions.
Ronacher expects competitors to emerge as utility becomes clear. TypeSafe plans to develop additional versions of the model in new modalities. Almeida rejects the label of a frontier lab focused on fear, hype, or building artificial general intelligence.
