0% found this document useful (0 votes)
19 views21 pages

Superintelligence Strategy: Expert Version

The document outlines a strategic framework for managing superintelligence, focusing on deterrence, nonproliferation, and competitiveness in AI technology. It highlights the risks of strategic competition, terrorism, and loss of control associated with AI advancements, and emphasizes the need for robust measures to prevent rogue actors from exploiting AI capabilities. The document also discusses the balance between offense and defense in the proliferation of dual-use technologies and the importance of maintaining stability in military and economic contexts.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
19 views21 pages

Superintelligence Strategy: Expert Version

The document outlines a strategic framework for managing superintelligence, focusing on deterrence, nonproliferation, and competitiveness in AI technology. It highlights the risks of strategic competition, terrorism, and loss of control associated with AI advancements, and emphasizes the need for robust measures to prevent rogue actors from exploiting AI capabilities. The document also discusses the balance between offense and defense in the proliferation of dual-use technologies and the importance of maintaining stability in military and economic contexts.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

SU PER INTE LLIGENCE

S T R AT E G Y
EXPERT V ER SION
arXiv:2503.05628v1 [[Link]] 7 Mar 2025

DA N HENDRYC KS ERIC SCH MIDT ALEXANDR WANG


Strategy

Deterrence Nonproliferation Competitiveness


Detect and deter Reliably know the location Guarantee access to AI
destabilizing AI projects with of AI chips and prevent rogue chips through domestic
(cyber) espionage and sabotage actors from smuggling them manufacturing
Problems

Strategic Competition Terrorism Loss of Control


States may race on AI capabilities Rogue actors could use AI systems An autonomous AI system
in destabilizing bids for economic to create bioweapons or launch could slip from
and military dominance cyberattacks on critical infrastructure meaningful human control
When should a dual-use technology be proliferated
without restrictions?
Dual-Use Technology

Offense-Defense
Balance?

Defense Dominant Offense Dominant

Proliferate Widely Catastrophic?

No Yes
(localized harm) (scalable harm)

Debateable Limit Proliferation


Escalation MAD vs MAIM
U.S. Nuclear U.S. AI
Recursion Not Controlled Omnicide
China Maims Project Aggression Aggression
U.S. Begins
Do Superintelligence China Outpaces U.S. Create Superweapon US Crushed U.S. Develops
Manhattan Project Mutual Mutual Developing
Recursion Controlled Disarmament Nuclear Weapons Disarmament Powerful AI
China Does Not Maim No Superweapon Escalation (Unstable) (Unstable) U.S. AI Project
Sabotaged
Recursion Controlled
Create Superweapon Crush China
U.S Outpaces China

U.S. Bid for


Soviet Nuclear Mutual Assured U.S. Nuclear Chinese AI Mutual Assured Strategic
Recursion Not Controlled Omnicide
Destruction First Strike Aggression AI Malfunction
Aggression Soviets China Begins
Dominance
(Stable) (Unstable) (Stable) (Unstable)
Develop Developing
Nuclear Powerful AI
Weapons
Chinese AI
Soviet Nuclear Project Chinese Bid
Mutual Mutual
First Strike Sabotaged for Dominance
Destruction Destruction
(Unstable) (Unstable)
Nonproliferation

Compute Security Information Security AI Security


Reliably know the location Limit proliferation of Prevent rogue actors from
of AI chips and potentially catastrophic dual-use accessing potentially catastrophic
prevent their smuggling AI weights to rogue actors dual-use AI capabilities
Know-Your-Customer Input/Output Filters Refusal Training Circuit Breakers Monitoring
Competitiveness
Military Economy
Manufacture drones and Guarantee access to
carefully integrate AI AI chips through
into military command domestic manufacturing
and control and export controls

Law Politics
Extend legal requirements Maintain political stability
to AI agents in the face
to facilitate commerce of mass automation

Common questions

Powered by AI

MAD traditionally refers to the stable deterrence stemming from nuclear capabilities, where both sides are deterred from launching nuclear weapons due to assured mutual destruction. In contrast, MAIM involves preventing aggressive AI development from leading to catastrophic malfunctions or escalations, promoting stability through proper control and disarmament measures across stakeholders .

Unilateral progression may lead to strategic dominance but risks triggering competitive escalations akin to the nuclear arms race. It could destabilize existing geopolitical balances, incite similar developments in adversaries, and increase the likelihood of international tensions and conflicts, undermining collective global security .

Political stability amidst automation requires extending legal requirements to AI agents, ensuring transparent and fair commerce facilitation. It includes adaptive governance structures that can manage mass employment shifts due to automation, fostering an environment of economic and social stability .

The proliferation of dual-use AI technologies poses risks of scalable harm when in the wrong hands, such as terrorism through bioweapon creation or severe economic and military destabilization. Mitigation includes strict nonproliferation measures, reliable location tracking of AI chips, export controls, and ensuring defensive dominance through legal and strategic frameworks .

Strategies proposed include ensuring robust compute security by reliably tracking the location of AI chips and preventing their smuggling. Implementing input/output filters, refusal training, and circuit breakers are also key measures. Additionally, monitoring through know-your-customer protocols aims to prevent rogue actors from accessing potentially catastrophic AI capabilities .

To ensure competitiveness, the document suggests domestic manufacturing of AI chips and strictly controlling their export. It also emphasizes military and economic integration of AI technologies, guarantee of continuous access to critical computing resources, and strategic investments in AI infrastructure to maintain leadership in the global race .

Wide proliferation could either stabilize or destabilize global security based on the balance of offense and defense capabilities. If defense is dominant, it could mitigate localized harms effectively. However, if offense is dominant, it can lead to scalable catastrophic harms, raising tensions and potential conflicts among nations, possibly triggering an AI arms race .

The document proposes careful integration of AI through strategic manufacturing of drones and AI systems, reinforced by secure chips domestically produced and controlled. This integration targets enhancing command and control features, establishing a robust military AI ecosystem without succumbing to the risks of destabilizing proliferation .

Detecting destabilizing AI projects involves utilizing cyber espionage and sabotage, employing advanced monitoring systems, and establishing international collaborations for intelligence sharing. Preemptive strategies emphasize neutrality in aggressive project development and strategic countermeasures to prevent rogue usage .

Controlling recursion involves strict monitoring of AI development processes, implementing protocols like refusal training and circuit breakers to halt dangerous autonomous actions, and fostering global cooperation to manage technological escalations proactively. Emphasizing shared safety principles can mitigate risks of uncontrolled recursive improvements .

You might also like