Tags
report
Date
Aug 29, 2023 2:03 PM
A response to the OSTP Request for Information on National Priorities for Artificial Intelligence by Jonas Schuett, Markus Anderljung, Lennart Heim, and Elizabeth Seger.
1. Risks from frontier AI models:
- Foundation models already cause significant harm.
- Further integrating foundation models into society might lead to systemic risks.
- As foundation models become more capable, more extreme risks might emerge.
2. Frontier AI regulation:
- We need specific regulation for frontier AI models.
- Defining the scope of frontier AI regulation is challenging.
- Regulators need more visibility into frontier AI development.
- Frontier AI developers should be required to:
- Conduct thorough risk assessments informed by evaluations of dangerous capabilities and controllability
- Engage external experts to scrutinize frontier AI models
- Follow shared guidelines for how frontier AI models should be deployed based on their assessed risk
- Monitor and respond to new information on model capabilities
- Comply with cybersecurity standards.
- In the future, the deployment and potentially even the development of frontier AI models may require a license.
- The US Government should support the creation of standards for the development and deployment of frontier AI models.
3. Compute governance:
- Compute is a particularly promising node to govern frontier AI models.
- The US Government should grant the Bureau of Industry and Security (BIS) a larger budget and empower it with the tools to effectively enforce the October 7th export controls.
- Frontier AI developers should be required to report training runs above a certain threshold.
- Compute providers should be required to have “Know Your Customer (KYC)” processes for compute purchases above some very large size.
- If companies want access to more compute, they should be subject to additional review requirements (“more compute, more responsibility”).
4. AI and democracy:
- AI might threaten democracy.
- “Democratizing AI” does not mean that frontier AI developers should open-source models.
- “Democratizing AI” is ultimately about ensuring benefits of AI are distributed widely and fairly.