The AI supply chain

AI products and services are no longer developed in-house. Multiple actors contribute to their development and deployment. One company may source the data, while another company trains the base model, and yet another company fine-tunes the base model to sell a specialized AI product. The result is a complex network of interdependent AI actors, products, and services known as the AI supply chain (see here). Open questions include:

1. Suppose one company builds on the product of another (e.g., startup builds their product on OpenAI’s GPT). What information should be passed from one entity to another to ensure safe and reliable behavior downstream?
2. How do model attributes propagate down the AI supply chain? For example, if model A has attribute X, will model B fine-tuned on A also possess attribute X?
3. When an upstream model is changed, can it be detected downstream?
4. How should AI supply chains be audited to ensure compliance with regulations?
5. Where do AI developers source their compute from? What are the implications on global dynamics on the chip supply? What are the implications of increasing compute demands on energy and water pricing?

As such, some projects can focus on machine learning aspects of the AI supply chain, while others study the economic or policy implications of the AI supply chain. One recent tool we
may build on is: https://aisupplychains.org/. The goal is to write a paper and policy report by the end of the year, aiming for a top ML venue.