Inference engine in game scripting language using macros (dasllama.io)

A new inference engine, dasllama, has been developed using daslang, a game scripting language that is significantly faster than existing libraries like llama.cpp, achieving up to 6.8 times faster performance on speech recognition tasks. This advancement is particularly relevant for the AI/ML community as it demonstrates the capabilities of a specialized scripting language in optimizing machine learning workloads without relying on traditional dependencies like C++ or separate shader languages. Daslang was originally designed for efficient game logic execution, featuring strong static typing and a fast compilation process. Its macro system allowed developers to create a powerful inference engine that supports both speech recognition and text-to-speech applications. By optimizing kernel generation through compile-time code manipulation, dasllama achieves high performance on various CPU architectures, including Graviton and Apple M1/M5 Max, and provides a self-tuning capability that adapts to the specific hardware it runs on. This development not only streamlines the integration of AI models into game environments but also presents a promising avenue for more efficient AI processing in diverse applications.