Even the most advanced artificial intelligence (AI) algorithms may be limited by the hardware on which they run. However, researchers are exploring the possibility of using this hardware constraint to prevent the development of dangerous AI. By encoding rules directly into computer chips, researchers suggest that AI systems can be governed and controlled to mitigate potential harm. This approach could prove to be a more effective way of regulating AI than relying solely on conventional laws or treaties. Trusted components already exist in some computer chips to safeguard sensitive data, and the idea is to expand on this concept by etching additional controls into future chips, such as graphics processing units (GPUs), to limit computing power and who can build the most powerful AI systems. Through the issuance of licenses by governments or international regulators, access to AI training can be monitored and limited, potentially preventing rogue nations or irresponsible companies from developing dangerous AI. This article explores the potential of this approach and the challenges it may face.
Building Limitations into Crucial Chips
In the realm of artificial intelligence (AI), there is growing concern about the potential dangers associated with unchecked AI power. To address this, researchers are exploring the idea of building limitations directly into crucial chips, such as GPUs, that power AI systems. By encoding rules into the computer chips themselves, it may be possible to cap the power of AI algorithms and prevent them from causing harm. This approach could offer a more effective and evasive way to regulate AI compared to traditional laws or treaties. A recent report from the Center for New American Security (CNAS) outlines the potential of this approach in preventing the secret development of dangerous AI.
Harnessing Trusted Components in Existing Chips
Some computer chips already incorporate trusted components designed to protect sensitive data and prevent misuse. For example, iPhones feature a “secure enclave” that safeguards a person’s biometric information. Similarly, Google utilizes a custom chip in its cloud servers to ensure the integrity of its systems. The CNAS report suggests leveraging these existing features in GPUs, or even creating new ones, to restrict access to computing power for AI projects. By implementing licensing protocols, governments or international regulators can control who has the ability to train the most powerful AI models. Refreshing these licenses periodically can allow for added oversight and the ability to deny access to AI training.
Etching AI Controls into GPUs
To impose limitations on AI systems, it is crucial to restrict access to computing power. GPUs are commonly used for AI training due to their high processing capabilities. By implementing controls within GPUs, it becomes possible to regulate the amount of computing power available for training AI models. Licensing protocols can ensure that only authorized entities have access to the necessary computing resources. Moreover, evaluation protocols can be established to assess the safety and effectiveness of AI models before deployment. This approach ensures that AI models meet certain standards and mitigate potential risks.
Addressing Concerns about AI Power and Misuse
There is growing apprehension regarding the power and potential misuse of AI. Some individuals worry that AI systems may become unruly and pose a dangerous threat. Additionally, there are concerns that AI could be used to develop chemical or biological weapons or facilitate cybercrime. To address these concerns, export controls on AI chips have been implemented to limit access to advanced AI technologies. However, these controls can be circumvented through smuggling and other means. To effectively address these concerns, it is necessary to develop AI chip export controls that are robust and difficult to bypass.
Drawing Parallels with Nuclear Nonproliferation
To understand the potential for regulating AI through hardware controls, it is helpful to draw parallels with nuclear nonproliferation efforts. In the field of nuclear technology, extensive infrastructure has been established to monitor and control important technology. Verification technologies, such as seismometers, play a crucial role in detecting underground nuclear tests and ensuring compliance with treaties. Similarly, hardware controls in AI chips can provide the necessary infrastructure to monitor and regulate the development and deployment of AI systems. By adopting verification technologies and protocols inspired by nuclear nonproliferation efforts, it may be possible to ensure responsible and safe use of AI.
Existing Examples and Proof of Concept
While the concept of hardware controls for AI is relatively new, there are already some examples and proof of concept. Nvidia, a prominent player in the AI industry, incorporates secure cryptographic modules in its AI training chips. These modules help protect sensitive information and restrict unauthorized access to AI models. Researchers at the Future of Life Institute and Mithril Security have also demonstrated the use of security modules in Intel CPUs to enforce cryptographic schemes that restrict the unauthorized use of AI models. These examples showcase the potential for implementing robust hardware controls in AI systems.
Barriers to Implementing Hardware Controls
Implementing hardware controls for AI poses both technical and political challenges. Developing cryptographic software schemes that can effectively safeguard AI models while being acceptable to the tech industry is no simple task. Additionally, new hardware features may need to be introduced in future AI chips to support these controls. Overcoming these technical challenges requires collaboration between industry experts and policymakers to find effective solutions. Furthermore, opposition from the tech industry, which has a history of resisting hardware interventions, can impede the implementation of hardware controls. Balancing the interests of various stakeholders is crucial to overcoming these barriers.
Government Interest in Hardware Controls
The US government has expressed interest in exploring the potential of hardware controls for regulating AI. The Bureau of Industry and Security has requested technical solutions that would allow chips to restrict AI capabilities, emphasizing the national security need to prevent the development of large AI models without proper safeguards. This government interest highlights the recognition of the importance of microelectronic controls in AI chips for maintaining national security. Collaboration between government agencies, researchers, and industry stakeholders is essential to drive progress in this domain.
The Role of WIRED in Technology Transformation
WIRED has played a significant role in covering emerging technologies, including AI, and the potential impact they have on our lives. Through its reporting and analysis, WIRED sheds light on breakthroughs and innovations in the field of AI. The exploration of etching AI controls into silicon aligns with WIRED’s commitment to uncovering transformative technologies and discussing their technical and political feasibility. By examining the potential of hardware controls for AI and addressing the challenges and opportunities they present, WIRED contributes to the conversation surrounding the responsible development and deployment of AI.
Conclusion
The idea of harnessing hardware controls to limit the power and potential harm of AI is a thought-provoking approach. Encoding rules into computer chips and regulating access to computing power can provide a more evasive and effective means of preventing the development of dangerous AI. However, implementing these controls comes with technical and political challenges, necessitating collaboration and innovative solutions. Government interest in hardware controls demonstrates the recognition of their national security implications. As technology continues to evolve, it is crucial to explore the potential of etching AI controls into silicon while carefully considering the technical and political feasibility.
