Waterloo Open-Sources ProgramAsWeights: Compile English Descriptions into Local Neural Programs

A University of Waterloo team led by Yuntian Deng released and open-sourced ProgramAsWeights (PAW), proposing the "Program-as-Weights" paradigm: a developer describes a "fuzzy function" in English (e.g., log alert triage, fixing malformed JSON, email urgency classification), and the system uses a large model to generate a LoRA adapter that "compiles" the description into a small neural program runnable locally on CPU and callable directly from Python—no internet, no inference API, zero per-call cost. The paper claims a compiled 0.6B model rivals a 32B model with prompting. Code and models are open-sourced, making AI functions distributable like ordinary software packages.

Confirmed

• The project is open-sourced with a Python SDK: after pip install programasweights, paw.compile("fix malformed JSON…") turns a natural-language description into a deterministic program
• Trained "neural programs" are distributed as LoRA adapters and can be installed and reused like ordinary packages
• The paper "Program-as-Weights: A Programming Paradigm for Fuzzy Functions" was released by Yuntian Deng et al.
• The GitHub repository has 304 stars
• Yuntian Deng demonstrated an end-to-end workflow with a coding agent given a single prompt—"read the project docs and build/evaluate a local CPU email urgency classifier"—which automated setup, compilation, and evaluation

Why it matters

• The core idea separates "understanding the task" from "repeated execution": the LLM acts as a compiler once, after which execution is offline, local, and cheap
• Turning per-call APIs into compile-once offline execution cuts cost and latency, and addresses privacy and offline scenarios
• Solidifying natural-language descriptions into deterministic, distributable small weights gives AI functions an engineering path analogous to traditional software compilation
• Combined with coding agents, the description-to-deployment pipeline can be highly automated, lowering the barrier for ordinary developers

2026-09-18 ~ 2026-09-20 · 10 related posts

• [source] ProgramAsWeights: compile English-described AI functions into tiny CPU-only neural nets — yuntiandeng · 2026-09-18
• Let your coding agent build and evaluate a local CPU email-urgency classifier — yuntiandeng · 2026-09-18
• ProgramAsWeights compiles plain-English specs into tiny neural programs that run locally — yuntiandeng · 2026-09-18
• [source] Program-as-Weights: 0.6B Model Matches Qwen3-32B Prompting With 1/50 Memory — yuntiandeng · 2026-09-18
• ProgramAsWeights: describe an AI function in English, run it offline on CPU — yuntiandeng · 2026-09-19
• [source] ProgramAsWeights: compile English descriptions into local CPU-runnable neural functions — yuntiandeng · 2026-09-19
• ProgramAsWeights: compile AI functions from English and run them offline on your CPU — yuntiandeng · 2026-09-19
• ProgramAsWeights compiles English function specs into local neural programs — yuntiandeng · 2026-09-20
• Program-as-weights: pip install a LoRA adapter to run neural programs locally — yuntiandeng · 2026-09-20

1 near-duplicate retellings: yuntiandeng