Program-as-Weights compiles fuzzy functions into local neural artifacts
PAW reframes LLMs as one-time compilers for reusable local fuzzy functions, cutting the per-call dependency on large hosted models.
Links: Original source · Shared link · Related link 1 · Related link 2
Logged at IST: 2026-09-17 23:48 IST
What it is: Yuntian Deng points to Program-as-Weights, a paper and toolchain for compiling natural-language function descriptions into small local neural programs.
Gist: PAW targets “fuzzy functions”: tasks that are easy to specify in English but awkward to implement as rules, such as log triage, malformed JSON repair, semantic filters, alert routing, or intent-based ranking. Instead of calling a large model on every input, a 4B compiler is invoked once for the function definition and emits a compact adapter for a frozen lightweight interpreter.
The paper reports that a 0.6B Qwen3 interpreter running PAW programs can match direct prompting of Qwen3-32B while using roughly one fiftieth of the inference memory and running locally on a MacBook M3. The public site makes the product framing explicit: define a function in English, compile it into a .paw artifact, then run it as a local Python function without internet access or per-call API fees. The released FuzzyBench dataset grounds the compiler training and evaluation with spec/input/output examples for many small text-transformation and classification tasks.
Newsletter angle: Strong fit for the “AI as software primitive” lane: a foundation model becomes a compiler that manufactures cheap, deterministic-ish, local micro-models for repeatable fuzzy work.