What Is Jev? I Tested TypeSafe's Decision Model on My Comments and Email

Jev is a System One model released by TypeSafe, which is a component that makes decisions based on text input rather than an assistant that can respond or summarize. Unlike chat models like Claude or GPT, Jev does not write, summarize, or provide explanations for its decisions. It can only pick from predefined answer options and assign a probability and confidence score to each option.

Jev uses one HTTP request to process data, where you provide text (state), questions with fixed answer options (questions), and the model to use (jev-latest). You can ask Jev to classify text into categories such as genuine, promo, or bot. It can also determine if the text asks a question or assign a score on a defined scale.

In the author's tests, Jev resolved various issues such as identifying genuine comments, detecting promotional content, and sorting security questions. For instance, Jev was able to correctly identify if comments were genuine, promotional, or bot-like with 97% confidence. However, Jev did make some mistakes, such as initially classifying a bot as genuine and failing to accurately classify certain self-promotional comments.

Despite its limitations, Jev proved to be a cost-effective solution in the author's tests, costing under one cent for processing emails and comments. The model's effectiveness largely depended on the quality of the descriptions provided for each answer option. If the descriptions were vague or not well-defined, Jev's accuracy decreased. Therefore, it is crucial to provide clear and specific descriptions to ensure Jev makes accurate decisions.