Former Google CEO Eric Schmidt has published a paper, “Superintelligence Strategy,” co-authored with Dan Hendrycks, director of the Center for AI Safety, and Alexander Wang, founder and CEO of Scale AI, which suggests that the U.S. should refrain from pursuing a Manhattan Project-style effort to achieve AI superiority because it would invite preemptive cyber responses from countries like China, The Register reports.
The authors of the article argue that any state that creates super-AI will pose a direct threat to other countries, and they, seeking to ensure their own survival, will be forced to sabotage such AI projects. Any “aggressive attempt at unilateral dominance in the field of AI will lead to preventive sabotage by competitors,” which can be implemented in the form of espionage, cyberattacks, covert operations to degrade model training, and even a direct physical attack on AI data centers.
The authors believe that in the field of AI, we are already close to the Cold War doctrine of Mutual Assured Destruction (MAD). The authors have given the current situation the name “Mutual Assured AI Malfunction” (MAIM), in which AI projects developed by states are limited by mutual threats of sabotage.
Image source: Igor Omilaev/unsplash.com
However, AI, like nuclear programs in their time, can benefit humanity in many areas, from breakthroughs in drug development to automation of manufacturing processes, the use of AI is important for economic growth and progress in the modern world. According to the article, states can choose one of three strategies.
- Complete non-interference in the activities of AI developers, chips and models.
- A global voluntary moratorium designed to halt further AI development, either immediately or after certain dangers are discovered, such as AI being able to be hacked or operate completely autonomously.
- A global monopoly on development by creating an international consortium like CERN that will lead global AI development.
Commenting on the US-China Economic and Security Review Commission’s (USCC) proposal for US government funding for a kind of “Manhattan Project” to create a superintelligence in some remote corner of the country, the authors of the article warned that China would respond to this move in a way that would only lead to a long-term imbalance of power and ongoing instability.
Source: Superintelligence Strategy: Expert Version
The authors of the article believe that states should prioritize the doctrine of containment over winning the race for artificial superintelligence. MAIM implies that any state’s attempts to achieve a strategic monopoly in the field of AI will face countermeasures from other countries and will also lead to agreements aimed at limiting the supply of AI chips and open-source models, which in essence will be similar to agreements on nuclear arms control.
To protect against nation-state attacks aimed at slowing down AI development, the article suggests building data centers in remote locations to minimize potential damage, writes Data Center Dynamics. Those looking to harm other countries’ AI efforts could start with cyberattacks: “Nations could poison data, corrupt model weights and gradients, disrupt software that handles accelerator errors and manages power and cooling…”
Source: Superintelligence Strategy: Expert Version
Transparency of developments will also help reduce the likelihood of attacks. AI can be used to assess the security of other AI projects, which will help avoid attacks on “civilian” data centers. At the same time, transparency of supply chains will not hurt. Since AI accelerators exist in the real, not virtual world, it is not that difficult to track their movement. Thus, even outdated or deemed unusable chips will not be able to end up on the black market – they are proposed to be disposed of with the same degree of responsibility as chemical or nuclear materials.
However, following all these recommendations will not solve the main problem – dependence on Taiwan for the production of advanced chips, which is critical for the United States, the authors of the article say. Therefore, Western countries should develop guaranteed supply chains for AI chips. This will require significant investment, but it is necessary to ensure competitiveness.