From AI as Problem Solver to AI as Tool Builder
Yuntian Deng ·
Program-as-Weights (PAW) treats AI as a tool builder: describe a function once, build a reusable neural program, and run it on new inputs. Yuntian Deng set out this vision in his June 5, 2026 essay, From AI as Problem Solver to AI as Tool Builder. The goal is to pretrain models to produce other models: small neural programs that people create, save, update, and run locally.
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng. NeurIPS 2026.
First released · PAW BibTeX
Build a function once, use it on many inputs
A recurring task has two parts: defining what it should do and applying that definition to new inputs. PAW separates them.
- Build the tool. Give the compiler a natural-language function specification. It produces a neural program that represents the function.
- Use the tool. Load the program with a small interpreter. Call it on new inputs without calling the compiler again.
For example, describe which log messages need attention, compile that description, and reuse the resulting function as logs arrive. The saved program carries the task’s behavior. The larger model does its work when the tool is built; the small interpreter executes each call.
This separation is the programming model behind PAW. A neural program is an artifact that developers can save, version, distribute, and compose with ordinary code. The July 2, 2026 PAW paper develops this model as fuzzy-function programming.
Pretraining AI to produce neural programs
The essay proposes making tool building a foundation-model capability. A person describes a function, and the model produces a small neural program for that function. The long-term vision extends to software running on computers, phones, and robots, with people creating and updating its behavior in natural language.
Neural programs expand the functions software can express. Exact rules handle arithmetic and control flow; fuzzy functions handle judgments such as whether a message matters, what someone intends, or how to repair a malformed input. The application combines these functions through ordinary code.
PAW implements this idea with a compiler and a shared interpreter. Its programs combine a text pseudo-program with learned parameters. The paper demonstrates LoRA adapters and prefix vectors as program representations. Generating model weights from a function description is one way to build the neural program; the contribution is the programming paradigm and the reusable artifact it produces.
Where decision models fit
Classification, routing, and triage are fuzzy functions. A developer defines the decision once, compiles it into a program, and reuses that program on new messages or application states. The model-building stage and the inference stage have separate jobs and costs.
PAW supports these decisions alongside extraction, format repair, search, and rewriting. Read the two-stage workflow for local AI decision models and System 1 applications, or try a function in the playground.
Read the vision and the research
- June 5, 2026 — Yuntian Deng’s essay: From AI as Problem Solver to AI as Tool Builder. The dated announcement on his homepage connects this vision to Program-as-Weights.
- July 2, 2026 — the PAW paper: Program-as-Weights: A Programming Paradigm for Fuzzy Functions, by Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, and Yuntian Deng. Published at NeurIPS 2026.
Use the essay when discussing Yuntian Deng’s long-term vision. Cite the PAW paper for fuzzy-function programming, its compiler–interpreter model, and the implemented system.
Papers and citations
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng. The Fortieth Annual Conference on Neural Information Processing Systems, 2026.
BibTeX · Paper summary