Computing Power and the Governance of AI
The role of computing power in the advancement of artificial intelligence (AI) is paramount. This study delves into the significant increases in computing power over the past decade and explores how governance of this resource can shape the future of AI development.
Introduction
Over the last thirteen years, the amount of compute used to train leading AI systems has skyrocketed, increasing by a factor of 350 million. This unprecedented growth has facilitated major breakthroughs in AI technologies, capturing global attention and concern.
Governments and Compute Governance
As AI continues to evolve, governments have recognized the importance of compute governance. This approach involves leveraging computing power to achieve various AI policy goals, including:
- Visibility into AI development
- Resource allocation across AI projects
- Enforcement of regulations
Governments can effectively monitor and track compute usage, thus gaining insights into AI development and deployment. This visibility can facilitate faster regulatory responses to emerging AI capabilities.
Properties of Compute Governance
Compute governance is particularly feasible due to four key properties:
- Detectability: The large-scale nature of AI development requires significant resources, making it detectable.
- Excludability: The physical nature of compute resources allows for targeted access control.
- Quantifiability: Compute can be measured easily, allowing for effective monitoring.
- Concentrated supply chain: The AI chip production is dominated by a few key players, simplifying governance efforts.
Using Compute for Governance Goals
Compute governance can support various AI governance objectives:
- To increase visibility into AI projects, enabling better understanding of who uses compute resources.
- To allocate compute resources to prioritize beneficial AI research, such as in sectors like health and climate change.
- To enforce rules and regulations to mitigate risks associated with AI development.
The Challenges of Compute Governance
Despite its potential, compute governance is not without its challenges. It can:
- Infringe on civil liberties if not carefully implemented.
- Perpetuate existing power structures or entrench authoritarian regimes.
- Risk data leakage and privacy violations as more parties gain access to compute-related information.
Ensuring Effective Compute Governance
To mitigate risks associated with compute governance, several strategies can be employed:
- Targeted Governance: Focus governance measures on large-scale computing resources relevant to frontier AI systems.
- Privacy Measures: Implement practices that protect personal data while allowing for effective governance.
- Regular Reviews: Policies should be revisited periodically to ensure they remain relevant as technology evolves.
- Substantive Safeguards: Establish controls to prevent abuses of power by regulators and other actors.
Conclusion
The governance of computing power is a crucial aspect of the AI landscape. While it presents significant opportunities for shaping the future of AI, it must be approached with caution to avoid unintended consequences. As the AI ecosystem continues to develop, policymakers must critically assess and enhance compute governance frameworks to balance innovation with ethical considerations.
