AI 2040

Summary of Plan A

Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, Daniel Kokotajlo

AI companies are racing to build AIs that are smarter than humans in every way. In AI 2027, we predicted this poses an unacceptable risk of extinction or irreversible concentration of power.

Plan A is our positive vision for what should happen instead. It involves an international coalition to delay the development of superintelligence until we are confident that it won’t destroy humanity. Read the complete plan at ai-2040.com.

Plan A’s core principles

  • Buy time: Slow down whenever is needed to have high confidence in safety.

  • Total research transparency: Make almost all AI research fully visible to the public.

  • Diffuse AI broadly: Many companies in many countries at the frontier.

  • Reversibility: Maintain self-destruct switches on datacenters, so that if the agreement breaks down we avoid a catastrophe.

Key steps to implement Plan A

  1. Track down the world’s AI-relevant compute. Audit the compute supply chain to bound the amount of “dark compute” that could be used to build existentially dangerous AI.

  2. Temporarily pause AI training while developing better verification technology. During this time, build new R&D datacenters with improved security and verification.

  3. Resume R&D under safety-case-based regulation and total research transparency. Companies must argue that capability improvements involve very low risk before proceeding.

We recommend that the following actions are taken as soon as possible to prepare for Plan A and better AI futures more generally:

  • Transparency to the public and governments: It’s important that there is a small gap between internal and publicly deployed models so that society can understand and respond to frontier capabilities. Other transparency targets include model specifications, whistleblowers, and information about internal usage.

  • Government AI capacity: The US government does not have adequate AI talent, this should be urgently fixed.

  • AI compute tracking: The US should prioritize highly collecting and analyzing AI-relevant intelligence, especially on the compute supply chain and AI datacenters.

  • Enforce export controls: Existing US export controls are poorly enforced; a third of Chinese compute is acquired via smuggling. Smuggled chips make agreements like Plan A more difficult to enforce because they are hard to trace.

  • Verification R&D: For example, developing an inference-only verification solution would enable the US and China to agree to stop doing new frontier AI training runs while allowing the public to maintain access to existing AI models.

Contact

Daniel Kokotajlo, Executive Director of AI Futures Project; daniel@ai-futures.org