AI Functions is an open-source Python experiment with a delightful premise: never write the function body again — describe what a function should do, give it a signature, and let a GPT model produce the result at runtime. Published on GitHub with about 936 stars and last updated in 2024, the project was heavily inspired by Ask Marvin and became a small viral classic of the early GPT-4 era.
The library exposes a single core call, ai_function, which takes a function signature as a string, a list of arguments, a natural-language description of the intended behavior, and a model name. The model then returns output matching the described contract — for example, generating a list of fake people with names and ages from nothing but a type signature and a one-line description. It works with GPT-4 or other model versions, and setup is minimal: clone the repository, install the requirements, and provide an OpenAI API key. While nobody should ship runtime-generated logic to production carelessly, AI Functions is a brilliant prototyping and exploration tool, and a clean illustration of how LLMs can act as interpreters for intent.
The examples show why the idea caught on: a function that ‘generates n examples of fake data representing people, each with a name and an age’ simply works, returning a clean list of dictionaries that matches the declared return type. Under the hood, the library formats the signature, arguments, and description into a prompt and parses the model’s reply back into Python values. It is a small codebase — an afternoon read — and a historically interesting glimpse of LLM-as-runtime thinking.
AI Functions is free and open source on GitHub. The only cost is your own OpenAI API usage.
Python developers and LLM experimenters who want to prototype ideas at the speed of thought, generate structured data from descriptions, or explore programming-by-intent patterns.