PAW AI: Offline LLM Tasks with 23MB Files

A new AI method allows large language models to compile tasks into tiny files that can run offline. This system, called Program-as-Weights, or PAW, uses a 23-megabyte compiled adapter to match the accuracy of a model over fifty times larger. An interpreter runs offline on a MacBook M3 at 30 tokens per second, using roughly one-fiftieth the memory. Researchers from the University of Waterloo, Cornell, and Harvard developed PAW. It addresses "fuzzy functions" like fixing JSON or ranking search results, which are common but costly when using large language model APIs for every call. The bottom line: this innovation could significantly reduce costs and improve efficiency for many everyday AI tasks, allowing them to run locally and offline.