AI 2040: Plan A for Avoiding Superintelligence Catastrophe

The Core Proposal: A Verified Slowdown to Superintelligence

Plan A is a policy recommendation to avoid AI-driven existential catastrophe by delaying the development of superintelligence until 2040. The strategy advocates for an international deal—primarily between the US and China—to replace the current secretive race for AI dominance with a regime of total research transparency and mutually assured compute destruction.

By making all AI research public and allowing multiple global companies to scale slowly and safely together, Plan A aims to prevent two primary risks: the loss of human control over superintelligent systems (misalignment) and the unprecedented concentration of power in the hands of a few individuals or entities.

The Proposed Timeline for Plan A

To stress-test their policy recommendations, the authors outline a concrete scenario where Plan A is implemented successfully:

• 2029: The US and China agree to avoid a reckless race to superintelligence.
• 2030: The world avoids the "default" path of fully automated AI R&D, which the authors predict would have led to superintelligence by the end of the year.
• 2030–2035: AI capabilities scale within the human range, reaching the level of top human experts.
• 2035: A strategic pause is implemented at the top-human-expert level to maintain human control.
• 2040: The pause is lifted, and scaling to superintelligence begins.

Key Policy Interventions

Plan A relies on several critical interventions to ensure the slowdown is verifiable and effective:

Total Research Transparency

AI companies would be required to limit the gap between internal and external deployment. This prevents the risk of "covert" recursive self-improvement. Requirements include publicly reporting model specifications (goals and values), internal usage statistics, and qualitative impressions of internal AI orchestration.

Compute Governance and Verification

Because frontier AI training requires massive amounts of hardware, compute serves as the primary lever for enforcement. Plan A proposes:

• Supply Chain Tracking: Requiring chip fabs and datacenter owners to publicly declare major purchases and sales.
• Verification R&D: Developing "inference-only" verification solutions that allow the public to use existing models while verifying that no new frontier training runs are occurring.
• Export Control Enforcement: Strengthening the enforcement of existing export controls to prevent smuggled chips from enabling covert projects.

Government Capacity Building

The authors argue that the US government currently lacks the top-tier AI talent necessary to oversee these policies, making the acquisition of high-quality AI expertise an urgent priority for the state.

Alternative Scenarios and the "Race" Risk

The authors contrast Plan A with other potential paths, most notably Plan D (Race to ASI). In Plan D, companies automate AI R&D to reach superintelligence as fast as possible to beat geopolitical rivals.

"We think Plan D is atrocious... we don’t expect the AI companies to retain control over their AIs through the intelligence explosion, if they race approximately as fast as they can... [and] the risk of World War III is too high."

Critical Analysis and Counterpoints

Community discussion surrounding the AI 2040 proposal highlights several technical and economic skepticism points:

• The "Inevitability" Assumption: Critics argue the proposal assumes LLMs will inevitably lead to ASI, ignoring the possibility that current architectures may hit a plateau (a "sigmoid" curve rather than an exponential one) or that they lack true intelligence beyond token prediction.
• Economic Feasibility: Some observers point out that the projected GPU build-out costs (reaching trillions of dollars) would exceed the GDP of major nations, questioning where the capital for such expansion would originate.
• Enforcement Challenges: Skeptics question the plausibility of a deal with China, citing the difficulty of verifying compliance in a Stalinist one-party state and the historical tendency for research to go underground rather than stop.
• Labor Market Speculation: Some critics find the scenario's predictions of mass unemployment (up to 74%) and the speed of physical robot deployment by 2035 to be wildly speculative and disconnected from economic reality.