According to a report in TechCrunch, Diogo Almeida, a former OpenAI researcher who helped develop ChatGPT and co-invent reinforcement learning from human feedback (RLHF), has launched a new transformer-based model called Jev through his startup TypeSafe AI. The model is not a large language model (LLM) and does not output text, but instead produces probabilities or what the company calls "calibrated decisions." Almeida left OpenAI two years ago to found the company in order to solve the gap created when models are optimized for human language rather than computer language for automation purposes.
Fast Probability Model Without Hallucinations
Eschewing natural language generation makes Jev exceptionally cheap and fast. Because users define output options in advance, the model cannot hallucinate. In terms of pricing, output tokens are provided for free, and input tokens are billed by the billions of tokens rather than by the millions. Following the launch, significant interest was recorded from developers, and the company even experienced a brief outage in its ability to serve users via its application programming interface (API) due to high demand.
According to the report, Jev appears most useful for software automation, with developers viewing it as a cheaper and more robust way to integrate intelligence into their code. Pranit Sharma, a software engineer at Vercel, which builds agentic infrastructure, stated that the company previously used OpenAI's ChatGPT Luna 5.6 as a classifier to review commands for safety. When Vercel replaced OpenAI's model with Jev, it received results 5 to 18 times faster and with greater accuracy.
In another test conducted by Nikhil Mudholkar, CTO at Bryo AI, Jev was tested against the Gemini model for classifying business emails. In that test, Gemini was slightly more accurate, but 10 to 20 times more expensive. Mudholkar emphasized that the probability scores provided by Jev return a real probability, making it ideal for workflow automation.
Agent Monitoring, Jailbreak Prevention, and Workload Routing
Beyond replacing language models in certain scenarios, the new model can also serve as a smart control layer over unwanted behavior of large language models. Almeida argues that using agents to monitor other agents can quickly become expensive, whereas using Jev for this purpose makes economic sense. According to his vision, users can deploy Jev to track the actions of LLM agents and prevent jailbreaks.
Armin Ronacher, CTO of Earendil, which develops the open-source model environment Pi, explained that the model delegates part of the hallucination problem to the user: when the model returns a result with a 50% probability, the user can treat it like a coin toss and disregard it, whereas at 95% probability, action can be taken based on the output. In addition, Ronacher noted that another potential use for the model is model routing. Predicting whether a given task requires a dedicated model is a useful tool, but using an LLM for that purpose is expensive, while Jev's low cost and speed enable real-time sorting.
The Jevons Paradox and the Distributed Software Vision
The model's name is derived from William Stanley Jevons, a 19th-century economist after whom the "Jevons paradox" is named. The paradox describes how a decrease in the cost of a basic commodity can lead to its increasingly widespread use. In the current context, Almeida hopes that the falling cost of intelligence will lead to its ubiquitous deployment. Almeida noted that the company believes smart software will spread in an emergent and distributed manner, much like the early internet, rather than in the form of mega-apps currently being built.
Training on Synthetic Data and Future Architecture
Almeida remains tight-lipped regarding the exact architecture of Jev, while outside observers suspect it is built on top of an open-weight language model. The company refers to Jev as a "System One model," focusing on intuition rather than complex logical reasoning, and directed at executing the appropriate task.
According to Almeida, Jev was trained exclusively on synthetic data using a technique he calls "reinforcement learning from calibrated decisions." Almeida noted that the early decision to generate all training data internally was one of the best decisions he ever made, and half of the company operates as a laboratory focused on the field of synthetic data. TypeSafe plans to build additional versions of the model in different modalities, while Ronacher expects competitors will develop similar models following the demonstration of the model's utility. In response to whether TypeSafe is a frontier lab, Almeida clarified that the company does not operate as a lab aimed at generating infinite wealth or building God in a server farm, but rather aims for its flagship product to be intelligence.