Jev Is an AI That Answers Yes or No Instead of Writing

The reel opened with a strange headline: a former OpenAI engineer launched an AI that cannot write. That AI is Jev, from a company called TypeSafe. You give it some text or an image and a question, and it never answers with a paragraph. It answers with a choice from a list you wrote, a yes or no with a probability, or a score.

For chatting, that would be a step backward. Inside an app, a yes, a no or one pick from a list is usually what the code wanted in the first place. Below is what Jev is in plain words, when a number beats a paragraph, and how to make your first call and read the answer.

Who made it

The claim in the reel comes from TypeSafe's launch post, written by its founder, Diogo Almeida. In his words: "At OpenAI, I helped build the methods that made language models useful at following instructions and talking with people. That work ended up as the research behind ChatGPT."

So the person who helped teach models to talk went and built one that does not. I find that more interesting than the headline itself.

It only makes decisions

TypeSafe calls Jev a System One model, after the fast, intuitive thinking Daniel Kahneman described in Thinking, Fast and Slow. Each call sends a state, which is the thing to judge, and one or more questions. Every question has a type, and the type decides the shape of the answer. The quick start lists four:

  • Choice picks one option from a list you define, with a probability for each option.
  • Score places the state on levels you describe, like calm, annoyed, furious.
  • Noul answers a yes or no question with the probability that the answer is yes, from 0 to 1.
  • Box finds where something is in an image.

Noul is the one the reel showed as yes or no. It is also the easiest place to start.

Fast, and it cannot make things up

In the reel I said Jev answers in under half a second while ChatGPT can take several. The launch post puts it at 70 to 500 milliseconds end to end, against 3 to 329 seconds for frontier models on the same kind of question. Those are TypeSafe's own numbers, and the post says its tests ran from laptops on the West Coast, so your times will depend on where you are.

"It can't make things up" needs a careful reading. Jev can only answer with options you defined in advance, so it never invents a category, never returns broken JSON and never writes a confident paragraph about something that does not exist. That part is guaranteed. It can still pick the wrong option. TypeSafe's own docs say so: its probabilities are calibrated across many answers, which "does not guarantee that an individual answer is correct." What you get in exchange is a number that tells you how sure it is.

When a yes or a number beats a paragraph

Most AI features in apps hide a decision. Is this support message about billing? Does this review mention a broken product? Is this photo a living room or a bedroom? Today a lot of apps ask a chat model, get a paragraph back, and then write code to guess what the paragraph meant.

If what your code needs is a yes, a no or one item from a list, ask for exactly that. You skip the parsing, you get the answer in a fraction of a second, and the probability lets your code decide what to do when the model is unsure. Send clear cases straight through and the unclear ones to a person.

The cards in the reel with "will they come back tomorrow" at 82% and "will they subscribe" at 34% were made up to show the idea. They are not real Jev output. Jev judges what you send it and knows nothing about what your users will do next. If you want predictions like that, TypeSafe shows a more honest route in its feature discovery cookbook: ask Jev small questions about your data, then train a regular model on those answers against what really happened.

A frame from the reel labeled ejemplo: two made up probabilities, will they come back tomorrow at 82 percent and will they subscribe at 34 percent. Illustrative, not real Jev output

Try it: your first call

These steps follow TypeSafe's quick start as it reads today.

  1. Sign in to the TypeSafe console and create an API key. If you would rather not write code yet, the console also has a Playground where you paste text and add questions in the browser.
  2. Save the key in your terminal: export TYPESAFE_API_KEY="your-api-key"
  3. With Python 3.10 or later, install the SDK: pip install typesafe-sdk
  4. Save this as first_call.py and run it with python first_call.py.
from typesafe_sdk import Noul, TypeSafeClient

message = "I have asked three times now. Can I please just talk to a real person?"

with TypeSafeClient(model="jev-latest") as client:
    response = client.system_one(
        state=message,
        questions={
            "wants_human": Noul(
                instructions="Is the customer asking for a human agent?",
            ),
        },
    )

print(response.nouls["wants_human"].noul)

It prints one number. The question id, wants_human, is just a name for your code. Jev never sees it, so the full meaning has to be in the instructions.

Reading the probability

A Noul near 1 is a strong yes, near 0 a strong no. Near 0.5 means the model finds yes and no about equally likely, which is different from a medium amount of yes. The Noul page shows recorded answers for this same question: "Thanks, that fixed it!" got 0.02, "Is there any way to speak to someone about my invoice?" got 0.84, and "Are you a bot?" landed at 0.40, a hint without a request.

TypeSafe suggests 0.7 for yes and 0.3 for no as a starting point, with anything in between going to a person. Move the line by the cost of being wrong. Raise it when a false yes is expensive, like issuing a refund. Lower it when missing a yes is worse, like a safety report.

Price is low enough that it rarely decides anything. The models page lists $0.042 per million input tokens, and output tokens are free.

Before you put it in an app, pin the version. The name jev-latest moves to each new release, so the same question can come back with a slightly different number after an update. Once you have tested thresholds on your own messages, use a fixed version ID, which the API returns with every answer, and switch when you are ready.

← All posts