All Engine, No Brakes: Preparing for the Superintelligence Era

Imagine driving a high-performance car where every available dollar has been poured into upgrading the engine, but nothing has been spent on the brakes, headlights, or seatbelts. According to some AI policy experts, this is the current state of artificial intelligence development: a relentless push for speed with little regard for the safety infrastructure needed to survive the ride.

As AI models grow increasingly sophisticated, the conversation in policy circles is shifting from regulating chatbots to preparing for "Superintelligence"—AI systems that vastly outperform human cognitive abilities. But how does a nation prepare for a technological leap that is entirely unprecedented? The RAND Corporation, a prominent US think tank, recently attempted to answer this by outlining a comprehensive strategy to secure geopolitical advantage and, quite literally, ensure humanity's survival.

The core of RAND's proposal is a "Freedom of Action" strategy. Because the path to superintelligence is fraught with unknowns—such as when AI might become dangerously autonomous or whether humans can effectively restrain it—the best approach is to maximize optionality. In other words, the US government needs to build a versatile toolkit now, so it isn't caught off-guard later.

RAND categorizes the potential future strategies into three broad families. The first is Acceleration, which closely mirrors today's reality. It relies on market forces and rapid iteration, operating on the belief that trying to constrain AI is actually more dangerous than letting it evolve.

On the other end of the spectrum is Denial. This family of strategies kicks in if AI proves fundamentally unsafe. It ranges from enforcing a verifiable global moratorium on advanced AI development to the stark "Continuity of Society" scenario. In this extreme case, assuming all preventive measures fail, the objective shrinks to pure survival—building geographically distributed, biologically self-sustaining settlements hardened against AI-enabled threats.

Between these extremes lies Coexistence. This includes scenarios where the US maintains dominance, but also fascinating collaborative models. For instance, RAND explores the idea of "Co-Development," where the US might lead an international consortium—potentially including rivals like China—to jointly develop safe superintelligence through shared governance, pooled computing resources, and strict monitoring.

To make any of these strategies viable, RAND argues that massive, immediate investments are required outside of just raw compute power. Nations need to build an "AI-security architecture" capable of tracking underlying compute resources globally. They must overhaul legacy national security institutions and create "break-glass" emergency plans for unexpected AI crises.

The overarching message is clear: winning the AI race isn't just about crossing the finish line first. It's about ensuring the track, the vehicle, and the driver survive the journey.

Key Points

• The RAND Corporation proposes a 'Freedom of Action' strategy to maintain geopolitical flexibility as superintelligence approaches.
• Potential strategies range from 'Acceleration' (market-driven growth) to 'Denial' (global moratoriums and survival settlements).
• A 'Co-Development' scenario suggests global powers could form a consortium to pool compute and jointly govern safe AI.
• Experts warn that current AI development focuses too heavily on capabilities ('the engine') and not enough on safety infrastructure ('the brakes').

Why It Matters

The transition to superintelligence will redefine global power dynamics. Understanding how major think tanks plan for this transition reveals the high stakes of AI safety, international cooperation, and national security.

Sources:

• Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics — Import AI (Jack Clark)