The reliance on large language model APIs for complex, rule-resistant programming tasks like log analysis or data parsing introduces significant overheads in locality, reproducibility, and cost. A new paradigm, fuzzy-function programming, emerges as a solution.
Related startups
Compiling Functions, Not Just Prompts
This research introduces Program-as-Weights (PAW), a system that compiles natural-language specifications into compact, locally executable neural artifacts. Instead of treating a foundation model as a black box for every query, PAW leverages a compiler to generate parameter-efficient adapters for a frozen, lightweight interpreter. This reframes the LLM's role from a per-input problem solver to a tool builder, invoked once during function definition to create a reusable, inexpensive artifact.
Unlocking Local, Efficient AI Execution
The instantiation of fuzzy-function programming via PAW demonstrates remarkable efficiency. A 0.6B Qwen3 interpreter executing PAW programs matches the performance of direct prompting of a much larger 32B Qwen3 model. Crucially, this comes at a fraction of the computational cost, using approximately one-fiftieth of the inference memory and achieving 30 tokens/s on consumer hardware like a MacBook M3. This breakthrough is powered by a 4B compiler trained on FuzzyBench, a 10M-example dataset released by the authors.
