Many common programming tasks, like log analysis or search ranking, currently rely on expensive, remote LLM APIs. Researchers propose "fuzzy-function programming" to compile natural language specifications into compact, executable neural artifacts that run locally. This approach aims to reclaim locality, reproducibility, and cost-efficiency by moving computation from cloud APIs to the developer's machine.

Their implementation, Program-as-Weights (PAW), uses a 4B parameter compiler trained on a 10M-example dataset called Fuzzy Bench. PAW generates parameter-efficient adapters for a lightweight, frozen interpreter. A 0.6B Qwen3 interpreter running PAW programs matches the performance of direct prompting of a much larger Qwen3-32B model, while consuming significantly less memory and achieving 30 tokens/s on a MacBook M3.

PAW reframes foundation models from per-input problem solvers into tool builders. Instead of repeatedly querying large models for each task, PAW compiles a function once, producing a small, reusable artifact. Subsequent calls to this compiled function are fast and operate offline, fundamentally changing how developers can integrate AI capabilities into their applications.