Larql: LLMs Are Databases. Query neural network weights like a graph database (github.com)

Larql introduces an innovative approach to managing transformer model weights by transforming them into a queryable format known as a vindex (vector index). This groundbreaking development allows users to interact with these neural networks as if they were graph databases, enabling querying, editing, and recompiling of the model's knowledge without needing a GPU. Users can execute queries using the Lazarus Query Language (LQL) to explore and manipulate model data succinctly, exemplified by commands like DESCRIBE and INSERT that provide insights into model facts and relations. This technology holds profound significance for the AI/ML community as it democratizes access to powerful transformer models, allowing real-time modifications and knowledge updates without extensive resources or retraining, thus enhancing model usability and adaptability. Key technical details include lightweight "patch" files for updates, multi-layer tracing for inference insights, and considerable performance efficiencies—demonstrated by fast querying times. Overall, Larql redefines how model knowledge is stored, accessed, and modified, paving the way for more flexible applications in natural language processing and beyond.