How to use Jev, the fast, cheap, doesn't speak, system 1 decision model from TypeSafe AI

Sep 28, 2026 · Lenny's Podcast

🎧 PodShort 26 min squeezed to 3 New

Full episode from Lenny's Podcast

Quotable Moments

this fast, cheap, doesn't speak, system 1 decision model

Jev only charges you on input tokens because it barely outputs anything.

Jev is really good at classifying data, that's a lot of what I'm going to talk to you about today.

Key Insights

Jev is a new type of AI model that acts as a 'system 1' decision-maker, providing fast, cheap, and effective classifications and choices based on pre-defined values.
Jev offers a cost-effective solution, with input tokens costing as low as 4 cents per million, making it significantly cheaper than traditional LLMs for certain tasks.
Jev's primary strength lies in classifying and clustering unstructured data, enabling tasks like categorizing PRs, sorting emails, and analyzing user feedback with high efficiency.
By combining Jev for initial classification and routing with larger, more capable LLMs for deeper analysis, users can create powerful and efficient AI workflows.
Jev is capable of performing real-time applications, such as analyzing user comments to determine sentiment or matching incoming data points to specific categories quickly.
The host used Jev to analyze YouTube comments, successfully categorizing sentiment, identifying episode ideas, and even spotting potential spam, demonstrating its practical application in content analysis.
Jev can be used to build real-time applications like a color-matching tool for text or a real-time search function for comments, showcasing its speed and decision-making capabilities.
The 'system 1' nature of Jev, combined with its low cost and speed, makes it ideal for tasks requiring immediate decisions and filtering of large datasets, complementing 'system 2' models for deeper analysis.

Metrics Mentioned

4 cents per million input tokens (The cost of Jev's input tokens, highlighting its affordability compared to other models.)
1,100 PRs analyzed (The scale of data processed by Jev for categorizing pull requests in the host's personal GitHub.)
9 cents (The cost of analyzing 1,100 PRs using Jev on the host's local machine.)
4,400 pairs analyzed (The number of comparisons Jev made when analyzing pull requests for thematic similarity.)
4,400 comments processed (The volume of YouTube comments analyzed by Jev for sentiment and topic extraction.)
4,400 identified episode ideas (The number of potential future podcast episode topics extracted by Jev from YouTube comments.)
4,400 identified spam comments (The number of spam comments Jev identified within the YouTube comment dataset.)
1000+ vs 30% (A comparison of the effort required to categorize data before Jev (1000+ hours) versus using Jev and other models (30% of effort).)