ProgramAsWeights: Natural language specs compiled to tiny neural programs (programasweights.com)

ProgramAsWeights (PAW) has launched a groundbreaking tool that allows users to define functions in plain English, which are then compiled into compact neural programs capable of running locally on their machines. This innovative compiler utilizes a 4B model based on Qwen3 to generate customizable programs for various applications, such as fuzzy searching, format repair, and email classification, without requiring internet connectivity or external services. Programs are lightweight, typically ranging from 5 to 22 MB, and offer fast inference times of around 100 milliseconds. This development is significant for the AI/ML community as it democratizes access to coding functionalities, enabling individuals without programming expertise to create and run AI-driven applications effortlessly. The seamless integration with local environments ensures data privacy and cost-effectiveness, appealing to developers aiming to streamline workflows without incurring per-call fees. With both Python and JavaScript interfaces available, PAW may catalyze a new wave of innovation across industries, promoting local AI solutions that prioritize usability and efficiency.