For this debate, “Plan A,” proposed by the AI Futures Project, is an international agreement intended to delay the emergence of artificial superintelligence (ASI) until around 2040. It would begin between the United States and China and eventually include other countries controlling major artificial intelligence (AI) development projects or critical parts of the semiconductor supply chain. Under Plan A:
• the use of new software advances would be restricted to prevent rapid, AI-driven acceleration of AI research and development, while AI capabilities would continue to improve gradually through expanding computing capacity;
• research activity at frontier training data centers would be largely open to the public, while information at customer-serving data centers would remain private; and
• AI computing infrastructure, semiconductor fabrication facilities, and related industrial capacity built after the agreement began would be subject to mechanisms allowing them to be disabled or destroyed in response to violations of the agreement, largely reversing the expansion of computing capacity since the agreement began.
Tom Davidson, an AI progress researcher at Forethought, has proposed an alternative that retains Plan A’s broad framework, with two primary differences:
• rather than expanding computing capacity while restricting software advances, the alternative would prohibit the construction of new AI computing infrastructure and semiconductor fabrication facilities, while allowing AI capabilities to improve gradually through software advances using roughly fixed computing capacity; and
• software advances could then be slowed or paused as necessary to maintain gradual capability growth.