What the Heck Is Jev?!

Put simply, Jev is a decision model with sub-second latency (70ms-500ms). It’s not like the LLMs we’re used to where you talk to it and it talks back to you in open-ended conversation. It doesn’t even write code, generate images, or talk you through a legal document.

So what can it do?

It’s much lower level than LLMs, faster than LLMs, cheaper than LLMs, and can be put to work in places you’d never dream of putting LLMs.

It can make a decision given some state and a list of possible answers, really really fast. It has been used to play Doom and shown in a Subway Surfers demo. In the Doom demo, it’s fed the game state as data, not screenshots, and decides what to do next in real time.

Much like a classifier model, you provide the state, the question, and a list of possible answers (like an enum). It evaluates all the options in parallel and assigns a probability to each answer, plus an overall confidence score. That parallelism is probably what makes it really fast.

It also can’t invent answers to questions you haven’t provided. It can still pick the wrong one though.

The three question types it supports are noul (the probability something is true), choice (select one from a list), and score (rate something along a scale you define, like how frustrated a customer is).

You always get structured JSON back. LLMs can do structured outputs too, but Jev doesn’t need to generate a response token by token to give your code an answer.

Why’s it called Jev?

Created by TypeSafe AI labs, Jev is named after the Jevons Paradox, referring to a time when steam engines becoming more efficient led to more steam engine use, not less: “when technology makes resource use more efficient, total consumption of that resource usually increases rather than decreases”.

TypeSafe calls this new class of AI a System One Model. System One, for those who have read Thinking, Fast and Slow by Daniel Kahneman, refers to the fast, intuitive judgement mode your brain uses to make split-second decisions. While System Two is the slower, deep thought kind of reasoning.

How do you talk to Jev?

For a choice question, you supply some state, a question, and a set of answers. Jev picks one and responds with a probability for each of your supplied answers, plus an overall confidence score, in structured JSON.

At a high level, the mental model goes like this:

STATE
"I was charged twice for my subscription. Can you refund me?"

QUESTION
"What is this about?"

ALLOWED ANSWERS
1. billing
2. technical
3. sales

→ Jev: billing

How that becomes an API request to Jev is:

{
  "model": "jev-latest",
  "state": "I was charged twice for my subscription. Can you refund me?",
  "questions": {
    "request_type": {
      "type": "choice",
      "instructions": "What is the customer asking about?",
      "criteria": {
        "billing": "Payments, charges, refunds, or invoices",
        "technical": "A bug or technical problem",
        "sales": "Buying or upgrading a product"
      }
    }
  }
}

A simplified example response, with made-up probabilities:

{
  "answers": {
    "request_type": {
      "type": "choice",
      "choice": "billing",
      "probabilities": {
        "billing": 0.98,
        "technical": 0.01,
        "sales": 0.01
      }
    }
  }
}

When an API call like that takes milliseconds, you can start to imagine putting this to use in regular if statements or any form of branching logic you own. A small request like this also costs a fraction of a cent. Some early adopters got $5 in credits which will probably last long enough to be passed on to their kids. You can start to see where the steam engine analogy comes from.

This feels like one of those pivotal moments in AI where speed and efficiency are being challenged, leading to new use cases where LLMs don’t fit, but their depth of knowledge does.