Jev

Jev is a model from TypeSafeA San Francisco AI lab that builds models for software to use rather than for people to talk to. Its first model is Jev. built for software to use rather than for people to talk to. A language model like ClaudeAnthropic’s family of AI models, and the apps built on them. The model Macktez uses daily for operations, client work, and automation. or ChatGPTOpenAI’s AI assistant, the product that introduced most people to modern AI. reads a prompt and writes an answer. Jev doesn’t write anything. A program gives it a block of information, such as a support ticket, and a set of questions whose allowed answers are defined in advance: which of these categories does it belong to, how urgent is it on a scale, did a person write it. Jev returns one allowed answer for each question, with a probability showing how confident it is.

How it differs from a language model

• It can only answer from the options it was given, so it never returns text a program can’t read, and it can’t invent an answer that isn’t on the list. It can still pick the wrong option.
• It answers many questions about the same item at once, in a fraction of a second.
• It costs a small fraction of what a general model costs per item, which makes it practical to run on every item rather than a sample.
• It can’t explain, summarize, rewrite, or reason through an open-ended problem. That work still belongs to a general model.

TypeSafe calls this kind of model a “System One” model, as distinct from the text-generating models most people know.

How we use it

In our own helpdesk, code handles exact facts, Jev answers the narrow questions that have a right answer (what kind of ticket is this, is it routine, how urgent is it), and Claude is called only for the tickets that need open-ended judgement. We measure Jev’s answers against Claude’s before trusting them, and stop relying on any Jev question that no longer agrees.