Why developers are giving Jev a shot

Welcome to The Prompt, your guide to the biggest issues and developments shaping AI, with Asia front and center. Want full access to all our AI reporting? Subscribe here.

Hello reader,

Not every AI task needs a model to think hard and write an essay. Sometimes, you just need it to make a call: Is this customer complaining, asking for a refund, or trying to buy something?

That seemingly boring job helps explain why developers have been buzzing about a new model called Jev over the past two weeks. Vercel, a popular cloud platform for developers, says Jev became its fastest-adopted model based on the number of paying teams who used it within the first 24 hours of its release.

So what’s all the hype about?

Jev is built to make decisions, an approach shaped by its creator and TypeSafe founder, Diogo Almeida, who helped build ChatGPT at OpenAI.

Provide some information, ask a question, give a set of possible solutions, and it picks an answer for you. It even assigns probabilities to the options provided.

Some billionaires famously wear the same outfit every day for a reason: Fewer choices mean fewer decisions. With Jev, instead of asking a model to generate from a near-limitless universe of possible answers, you narrow the field and ask it to choose.

Large language models (LLMs) can do this, too, but they are built to generate answers token by token, which can be an overkill when all you need is a quick pick among several options.

To be clear, Jev isn’t trying to replace LLMs. It can decide which tool to use or what action to take next, unlike an LLM, which steps in when the job requires deeper reasoning.

Image credit: Made by Ulla/Tech in Asia with the help of AI

Some companies are already finding uses for Jev.

Ivan Lee, founder and CEO of Datasaur, says his team has been experimenting with the model for the company’s annotation platform. He describes Jev as a “universal classifier” that offers an alternative to both LLMs and traditional classification models.

LLM inference costs can quickly add up when labeling data at scale, according to Lee. And while traditional classification models are cheaper, they require additional training for each new task.