Developers introduced the Laya vs Jev platform for comparing AI agents

The development team has released Laya vs Jev — a tool for direct comparison of the performance of two AI models in a dynamic environment. The system uses the classic Chrome T-Rex game as a testing ground for testing the reaction speed and decision-making of agents in real-time.

The Laya model works fully locally on devices with Apple Silicon architecture, using the MLX framework to optimize calculations. In turn, Jev functions through a cloud API from TypeSafe company, which allows evaluating the difference between local inference and network requests.

The platform's interface provides users with tools for in-depth analysis of agent performance. The functionality includes live latency graphs that display data transmission delays, as well as a crash replays system that allows frame-by-frame analysis of error or character "death" moments in the game.

The project is aimed at demonstrating the differences in architectural approaches to AI agents. Using a game with a high frame rate update allows you to clearly see how hardware acceleration on Apple chips and network delays of the API affect the final result and stability of the model's behavior in conditions of limited decision-making time.