Building Decision-Based Apps with TypeSafe Jev
Lukas Margeriego watch the original →
the gist
Jev is a System One AI model designed for fast, low-cost probabilistic decision-making rather than text generation, enabling real-time classification and UI control tasks.
The Core Mechanism of Jev
Jev functions as a specialized System One model that outputs structured, probabilistic decisions instead of generating natural language tokens. By bypassing text generation, the model operates with latencies between 70 and 500 milliseconds per pass. It is designed to evaluate input states to perform tasks like classification, ranking, or truth estimation, making it significantly cheaper and faster than traditional LLMs for software-integrated decision logic.
Implementation and Remixing Workflows
Developers access Jev via the TypeSafe platform, where they integrate the model into workspaces using an API key and a base skill prompt. The model excels at remixing existing open-source projects into new utility tools:
- Chrome Extension Classification: By adapting existing classifiers, developers can build browser extensions that label web content in real-time. For example, a YouTube scanner can categorize videos by hovering over thumbnails, identifying content as specific or vague based on Jev's evaluation.
- UI Component Extraction: Jev can scan video frames from product launches to identify and isolate UI components. These components can then be exported to tools like Diffusion Studio to recreate or animate designs.
- Real-time Environment Control: Jev can be integrated into game loops to process voice input and manipulate game states. By mapping voice commands to environmental variables, users can dynamically alter game difficulty or aesthetics while playing.
- Pixel-Level Generation: Rather than using diffusion models, Jev can predict pixel colors in parallel to render images using pure JavaScript, allowing for rapid, programmatic visual generation.