SDS 597: A.I. Policy at OpenAI — with Miles Brundage

Safety Strategies

Miles Brundage emphasizes the multifaceted nature of AI safety, which includes alignment, robustness, and security. He explains that AI policy translates ethical considerations into actionable practices, shaping decision-making to mitigate risks identified in AI safety and ethics research 1. Brundage notes, "There's something for everyone and lots of opportunities to make sure that AI helps rather than hurts people." 2.

   

Existential Risks

Miles Brundage discusses the evolving threats posed by AI, including voice spoofing and weaponized drones. He reflects on the rapid advancements in language generation, which have outpaced initial predictions 3. Jon Krohn highlights the use of AI in warfare, such as the machine vision in Javelin missiles, underscoring the importance of understanding AI's role in modern conflicts 4.

   

Policy and Safety

Brundage explains the importance of integrating safety measures into AI models, noting that these mitigations can be both commercially and research beneficial. He stresses the need for government involvement to ensure widespread adoption of these safety practices 5. He also discusses the concept of "model cards" for documenting AI model limitations, comparing them to nutrition facts on cereal boxes 6.