Last week, San Francisco startup TypeSafe previewed the first in what it claims will be a series of “System One” models to make rapid structured decisions. The model, called Jev, is not a large language model, but rather one specifically optimized for rapid classification.
The company claims that this approach can be 20 to 200x faster than general-purpose LLMs, and 40 to 1,000 times as inexpensive.
The response was immediate; a wave of buzz around Jev overtook the X AI forums over the weekend. It even spawned a new buzzword: “JevOps.” On Friday, Vercel reported that Jev was adopted faster than any other model in the history of its AI Gateway, reaching nearly 13% of its paid teams within 24 hours, more than twice as many as the GPT-5.6 family and more than six times as many as Fable 5.1.
On Sunday, TypeSafe opened Jev for general availability. Users are given a $5 credit, worth about 120 million tokens to play with the model.
Could Jev’s rapid ascent be a turning point for AI, a shift away from the general-purpose LLM?
Calibrated Decisions
Jev is the work of Diogo Almeida, a former researcher for OpenAI. Almeida was a core contributor to ChatGPT and instrumental in shaping generative AI as we know it, which was built on the concept of Reinforcement Learning from Human Feedback (RLHF).
In a talk given at the AI World Fair in July, Almeida spoke about the limitations of the RLHF model, which is optimized for human interaction. Almeida argues that because RLHF-based LLMs are geared toward human preferences, they can be prone to problems such as overconfidence and sycophancy: The model is tuned not primarily for truth but for human interaction.
Jev is built on a different approach: “Reinforcement Learning for Calibrated Decisions” (RLCD), which is not worried about appeasing its human overseers. In fact, it doesn’t even have a chat interface — developers interact with it primarily through a CLI, an SDK or curl.
Instead, it only has a limited set of predefined outputs, along with a probability score of how certain it is of the answer.
A Request, a Command, or a Query
With Jev, you enter an unstructured set of data, and then either ask a question, issue a command or run a query against the data. You then get back one of a number of predefined answers, which you can define. Think of the model as a general-purpose function call, an if-then-else statement for your data.
Unlike traditional LLMs, Jev doesn’t burn output tokens structuring a verbose answer, and because the answer types are defined ahead of time, TypeSafe says Jev returns its responses two orders of magnitude more quickly compared to a standard LLM. (Jev does not need an LLM to work, though it can serve as a fact-checking service for one.)
Jev’s TypeSafe primitives give three possible answers to any question asked. It can choose an option from a predefined list (“Choice”). Or, it can offer a score. Or it can evaluate something to be true or not (“Noul”). The first two options — Choice or Score — also come with a probability, or confidence score, allowing developers to build in more loops depending on the confidence level.
Multiple questions can be asked in parallel. Complicated questions work best when decomposed into a series of simple questions. This type safety eliminates the possibility of free-form hallucinations outside the predefined answer types, the company argues, though Jev can still make an incorrect decision.
While the output may initially appear limited, there are a wide variety of things that can be done with this format. In fact, many tasks assigned to frontier lab LLMs might be more efficiently served by an RLCD model instead, Almeida has argued.
On its site, TypeSafe offers a few examples of what can be done with the format. For instance, it can run through the game of Doom. The ultra-low latency ensures a responsiveness LLMs can’t touch. It can fact-check a document against some source material. Or it can be used to manage a smart home, where the user can just ask the system to turn on a coffee maker or secure the locks, and Jev will figure out which path to take.
Jev was named after William Stanley Jevons, who observed in the 19th century that if something becomes more efficient to use, it actually attracts more use, not less (“Jevons’ Paradox”). We’ll shortly see if Jev can be a lower-cost alternative to LLMs.
