ProgramAsWeights offers English-defined neural programs that run locally

@yuntiandengI increasingly think future software will combine code for control flow with small neural programs for "fuzzy" judgments. That's what I've been exploring with ProgramAsWeights, which answers the question of where those neural programs come from: they are "compiled" from English descriptions. For example, I combined 30 neural programs with a decision tree to build a course website helper that looks like a chatbot but runs locally. Each small program handles a question like "Which specialist answerer should this be routed to", and code controls the overall flow. You can try building with it here: https://programasweights.com

@iamoonkey@yuntiandeng Fully agree on a future with more dynamic fusion of deterministic code and neural intelligence

@yuntiandengRT @iamoonkey: @yuntiandeng Fully agree on a future with more dynamic fusion of deterministic code and neural intelligence

@yuntiandengThis composability angle really resonates with me. A little self-promotion 😅 We've been exploring it with http://programasweights.com: "compiling" English function descriptions into small neural programs that run locally, then composing them with ordinary code. I built my course website helper from 30 of them. Thought you might find this interesting:

@yuntiandengAnd here are the helpers in case people are curious (yes I built multiple of them!): My own homepage: https://yuntiandeng.com/#ask NeuralOS: https://neural-os.com/#ask My course: https://yuntiandeng.com/teaching/spring2026/cs486-introduction-to-artificial-intelligence/#ask

@yuntiandengThis is very close to what we’ve been exploring with ProgramAsWeights: train a larger model to generate a small model’s LoRA adapter from an English description of the function. The resulting function runs locally on CPU and can be reused inside ordinary code. We’ve released the code, model weights, and demos: https://programasweights.com

@yuntiandengYou can describe an AI function in English, compile it once, and run it with Wi-Fi off. That’s what we’ve been building with ProgramAsWeights: small neural programs you can call like Python functions, locally on your own CPU. Code and models are public: https://programasweights.com

@yuntiandengYour coding agent can easily handle setup, compilation and testing. For example: "Read https://programasweights.com/AGENTS.md and help me build and evaluate an email-urgency classifier that runs locally on CPU."

@yuntiandengThis is also what I used to build the English <-> Claudish translator:

@yuntiandengHosted inference is available too, if you prefer not to download a model. Inference remotely takes ~150 ms: https://github.com/programasweights/programasweights-python#remote-inference-optional Local inference is optional, so you can choose where your inputs are processed.

@yuntiandengDetails, code, and models are all available: Paper: https://arxiv.org/abs/2607.02512 Code: https://github.com/programasweights/programasweights-python Models: https://huggingface.co/programasweights

@yuntiandengOn the local-inference point in your report: we've been building ProgramAsWeights for this. You describe a text function in English, compile it once, and run the resulting neural program on your own CPU without API calls. The code and model weights are public. Just shared an example:

ProgramAsWeights offers English-defined neural programs that run locally

TLDR

Combined views

ProgramAsWeights offers English-defined neural programs that run locally

12 Sources

Combined views

Sentiment

Summary

TLDR

Sentiment

Summary

Sentiment

Summary

Sentiment

Summary

12 Sources