“Hey Perplexity, who is Jev?”

A sentence spoken for the first time in our lifetime.

​

But, no, really, who (or more like what) is Jev?

“Jev” is the name of a new AI model from TypeSafe AI, but unlike other models, this one is quite shy.

We have become so used to communicating with LLMs, seeing their thought processes word-to-word, navigating through suggestions and options… Jev is here to avoid all that.

World meet Jev

An AI tool designed not to chat or write essays, but to make fast, structured decisions. These decisions can vary from picking an option, outputting a score, or returning a probability with a confidence rating.

The tool was released on September 15th with a goal to be presented as a fast decision component inside an application. Created by a new AI lab called TypeSafe, the idea behind Jev is not to completely replace GPT-style models, but to improve decision-making.

TypeSafe’s co-founder, Diogo Almeida, who previously worked as a researcher at OpenAI, has founded TypeSafe with the idea that models need to better themselves at automation. According to Almeida, models have gotten exceptionally good at conversation, without producing much actual automation. He believes that software needs a different kind of model for the small decisions. Something to cut through the noise.

“Build prod, not God.” is the manifesto TypeSafe runs on.

What does Jev do?

‍

We have mentioned already that Jev is quite shy. Not just with the answers. The tool is still in early access, so nobody outside TypeSafe has had a chance to test it much. Based on the inspiration behind the tool and the news about it, we get the general idea of what Jev does.

​

To put it simply, it gives you a yes or no answer.

​

Again, Jev is created to be different from other text-generating LLMs. This is not a chatbox tool, and it’s not a back-and-forth conversation process. Jev evaluates structured input and prefers things put simply, so it can provide the best possible answer.

​

Our task is to provide a state. You do provide a text or structured data to describe the situation you want Jev’s input on. Once you provide your explanation, you also ask questions, one or several, demanding answers which could be a choice, score, or a noul.

​

So given the right context, Jev can make a selection from given options, rate something against a rubric, or determine the validity of a statement.

​

Jev evaluates all of the questions in parallel and returns direct answers which software can use right away. The process is much simpler than the standard - asking LLMs to generate prose from which the users extract answers about next steps.

​

In the words of TypeSafe, Jev is a “frontier-intelligence function call”. The process should be intuitive - you put in unstructured data to receive a structured probabilistic decision.

Inspiration and references behind Jev

‍

We know him as a Congolese‑Canadian rapper and songwriter, a nickname to a French Formula E champion and the word for "phenomenon" in Czech and Slovak, the world behind Jev is grand.

​

And let’s not forget the iconic “My name is Jeff” quote from the movie “21 Jump Street".

​

This Jev is named after a coal economist.

​

William Stanley Jevons is a 19th-century English economist who observed that when steam engines grew more efficient, Britain burned more coal, since cheaper power made new uses for it worth pursuing.

​

TypeSafe’s vision with Jev is similar. They expect AI to follow a similar pattern where, as answers get cheaper, companies will find more questions which they consider worth asking.

Do we really need another AI model?

‍

To put all things in perspective, this one is different. This one gets into the nitty-gritty routine work, somewhere where you need the simplest, but often the hardest, decision.

​

Jev’s architecture and training approach are designed for automation rather than conversation. TypeSafe has named this method “Reinforcement Learning from Calibrated Decisions”

​

Considering all that, Jev becomes useful for:

​

• Choosing which tool or workflow step to invoke.
• Ranking and classification.
• Controlling real-time software agents.
• Approval or escalation decisions.
• Routing support tickets.

‍

Looking at the bigger picture, the key idea is to simplify the workflow. Break complicated workflows into small judgements. And then combine them in ordinary program code. The goal is to build a system that is easy to analyze, inspect, and judge, instead of creating one prompt scenario that takes all the logic with it.

System 1 vs. System 2

‍

Many connect these two worlds.

​

The world of Daniel Kahneman’s book “Thinking Fast and Slow” and his beliefs on the work of two mental systems, and the world of Jev. As Jev delivers something that has been missing in the world of LLMs - movement in one direction without novelizing every action.

​

“Thinking Fast and Slow” is a best-selling book that takes the readers on a tour through a living brain. The book explores the work of “two brains” in the head: System 1- representing fast intuitive thinking and System 2 - which showcases slow, logical reasoning. Both thinking systems shape every decision that we as humans have.

​

Through his life’s work, the Nobel-awarded scientist wanted to better understand how people make economic decisions. With his research, he focused on cognitive psychology in relation to the mental process in order to increase understanding of how people make economic decisions.

​

Acknowledging the results, he drew on cognitive psychology in relation to the mental process used in forming judgements and making choices. His research on decision-making under uncertainty resulted in the formulation of a new branch of economics, prospect theory.

​

Besides being a great read, the book can easily explain how Jev fits in the puzzle of our existence. Jev trailblazes a “System 1” workflow into the  “System 2” setup we have all become so comfortable in.

​

The future that includes Jev

‍

There is a place for Jev in your workday; there is no doubt about it. Is it just that if you are looking for a novelist to draft your blogs and emails, don’t look this way. Jev does not generate explanations, essays, code, guides, or general conversational replies.

​

Jev operates best when the roadmap to a decision is supported by several smaller questions. The tool guarantees schema-conforming output, avoiding free-form text and hallucinations. And although it strives to be straight to the point, the human touch to finalize the process is still very much needed. At the end of the day, just like every tool, its probabilities and decisions can still be wrong if the output is ambiguous or you begin with a poorly designed question.

​

Jev is here to crash the rule that you need the most recent, priciest model for every single question you have. Sometimes, a simple yes/no answer is just the step you need.

Interested in learning more about the AI world? Read our previous blogs!  

‍