How could the US and China verifiably slow AI? (www.lesswrong.com)

A recent discourse highlights the complex dynamics between the U.S. and China regarding the acceleration of AI development, emphasizing the need for a verifiable slowdown. An October poll indicated that 64% of Americans believe AI is evolving too quickly, driven by concerns over labor displacement and potential AI dangers. Advocates for a slowdown, including leaders from AI firms like OpenAI, Meta, and Anthropic, are pushing for international governance tools to manage AI capabilities in a manner similar to nuclear arms control treaties. The discussion underscores the dilemma both countries face: while a mutual agreement to slow down AI progress could mitigate risks, the potential for one party to gain a significant geopolitical edge by secretly advancing their AI capabilities complicates matters. Successful implementation of such a treaty would likely hinge on governing compute resources—monitorable and controllable metrics that reflect AI training intensity. However, as AI technology evolves exponentially, the urgency for developing effective governance mechanisms grows, highlighting the tension between rapid technological advancement and the political capacity to ensure societal safety.