Shut up and calculate: Jev's new AI primitives for coders

TypeSafe's Jev, a new AI technology launched last week, is capturing the attention of developers as they explore its potential. Unlike standard frontier models known for their chatty approach, Jev behaves like a classifier with built-in intelligence. When provided with data, it generates predefined "typed" decisions, complete with probability distributions and, for certain query types, a confidence level. This structured output allows developers to make decisions and automate processes based on the AI's responses.

Developers have enthusiastically embraced Jev, quickly creating tools and applications to test its capabilities. From translating plain English into fancy prose to adding urgency columns to spreadsheet rows, Jev demonstrates a wide range of possible applications. Gaming enthusiasts have even leveraged Jev to optimize performance in popular games like Doom, Tetris, and chess.

Some developers have coined the term "JevOps" to describe the potential for running all code on Jev-based virtualization, although this remains a meme for now.

Jev's simplicity is one of its key advantages. Developers must define the schema and candidate options beforehand, unlike the free-form prompting required for other large language models (LLMs). This structured approach enables quick and efficient decision-making, with answers provided in as little as 150 milliseconds. Additionally, Jev's service is significantly more affordable than standard frontier LLMs, with input tokens costing $0.042 per million tokens and no charges for output tokens.

Potential workloads that could benefit from Jev include job recruiting, scientific paper screening, and automating software that currently requires human oversight. However, its accuracy remains uncertain, as there are limited benchmarks and insights into its intelligence. While Jev may be well-suited for tasks involving static, self-contained datasets, such as email spam filtering or online shopping recommendations, it may not be ideal for predictive tasks like suggesting stock picks.

Experts in the field recognize the potential of Jev, pointing to its faster and more affordable nature as a step forward from the often-deceptive and unstructured responses of current LLMs. By fostering smarter systems rather than merely smarter models, Jev represents a promising direction for future AI developments.