Proposal in one sentence:
I propose to research whether decentralised ownership can reduce the risk of misaligned or misdirected Artificial Intelligence.
Description of the project and what problem is it solving:
Powerful AI models that are either misaligned with human values, or used to enforce a particular moral view, represent a significant threat to our near and long-term future. This issue has recently emerged as a key priority from research on long-termism and existential risk; it is the focus of Stuart Russel’s latest book “Human Compatible”, and is one of the core focus areas of FTX-founder Sam Bankman-Fried’s recently launched “Future Fund”.
AI safety researchers have yet to focus on decentralisation, because prior to the emergence of blockchain technology and its integration with AI development, it was not feasible to own or control an AI system in a truly decentralised manner. However, progress in this field is now rapid, raising important questions about the future. Is an AI with decentralised ownership more or less likely to be developed in-line with a representative view of human values? Would decentralised control make it harder for an AI to be directed to cause harm, or conversely would it be harder to regulate?
Research on decentralisation may uncover significant opportunities to de-risk future AI systems, and thereby direct resources towards accelerating development in this area. It may also identify and help mitigate new risks that might otherwise go unnoticed.
This project aims to demonstrate that decentralisation has significant implications for AI safety, which should be studied and then factored into future research and development efforts. This will be achieved through a literature review and theoretical discussion. The project will conclude with the design of an experiment that tests whether an AI is safer, and more easily controlled, if its ownership is decentralised. This will likely involve a game in which centralised and decentralised groups compete through control of AI agents in a simulated environment.
Grant Deliverables:
- Literature review of AI risks that relate to centralised ownership and control
- Theoretical discussion of how decentralisation mitigates or enhances such risks, and what news risk it might introduce
- Design for an experiment which tests the theory in real world scenarios
Squad
Casey (BarefootDev), ML Researcher & Engineer, Co-founder and Lead Developer at Spectra.art