Program-as-Weights compiles specs to LoRA weights

Compiling natural language into local model weights rather than calling runtime APIs is a major architectural shift that could finally make offline, edge-based fuzzy computing practical.

* Hot-swapping task-specific LoRA adapters on a tiny 0.6B local base model completely bypasses network latency and API query costs.

* Shifting the LLM's role from a runtime executor to a compile-time weight builder enables predictable, local execution of non-deterministic functions.

* The companion benchmark, FuzzyBench, provides 10 million synthetic training examples across 800 tasks to help generalize compiler models.

* The paradigm is currently optimized for single-purpose, pure fuzzy functions, and scaling it to complex, stateful multi-step agentic tasks remains an open question.