Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Program-as-Weights compiles natural-language fuzzy functions (JSON repair, log filtering) into adapters for a frozen 0.6B interpreter — matching Qwen3-32B prompting at ~1/50th the memory, 30 tok/s on an M3.
**The gist** A **4B compiler model** trained on **FuzzyBench (10M examples)** turns a natural-language spec into a parameter-efficient adapter that runs on a frozen **0.6B Qwen3** interpreter. The compiled function matches direct prompting of **Qwen3-32B** at about **1/50th the inference memory**, hitting **30 tokens/s on a MacBook M3**.
**Why it matters** The fuzzy glue you currently route to an LLM API — repairing malformed JSON, flagging important log lines, ranking by intent — could become a compile-once, run-locally artifact: reproducible, cheap, offline. It reframes big models as **tool builders** invoked once per function definition rather than once per call.
**The gist** A **4B compiler model** trained on **FuzzyBench (10M examples)** turns a natural-language spec into a parameter-efficient adapter that runs on a frozen **0.6B Qwen3** interpreter. The compiled function matches direct prompting of **Qwen3-32B** at about **1/50th the inference memory**, hitting **30 tokens/s on a MacBook M3**. **Why it matters** The fuzzy glue you currently route to an LLM API — repairing malformed JSON, flagging important log lines, ranking by intent — could become a compile-once, run-locally artifact: reproducible, cheap, offline. It reframes big models as **tool builders** invoked once per function definition rather than once per call. **Watch out** Results come from tasks resembling the **FuzzyBench** training distribution; how compilation holds for messier or novel specs, and how you validate a compiled function's behavior before trusting it, is untested here.