by Sam Altman, Greg Brockman, and Ilya Sutskever
In the foreseeable future, artificial intelligence (AI) systems have the potential to surpass the expertise of human professionals in various fields and rival the productivity of today’s largest corporations. This advancement in AI, known as superintelligence, presents both immense benefits and significant risks that surpass those associated with previous technologies. While we can envision a highly prosperous future, it is crucial to proactively manage the risks involved. An analogy can be drawn to technologies like nuclear energy and synthetic biology, which possess similar characteristics and necessitate careful risk mitigation.
To navigate the development of superintelligence successfully, coordination among leading AI development initiatives is paramount. Establishing mechanisms for collaboration is vital to ensure that superintelligence is developed in a manner that prioritizes safety and facilitates smooth integration into society. Governments of major nations could initiate a project that incorporates existing efforts, or a collective agreement could limit the rate of AI capability growth at the frontier per year. Additionally, individual companies must adhere to exceptionally high standards of responsible conduct.
Furthermore, it is likely that we will eventually require an international authority akin to the International Atomic Energy Agency (IAEA) to oversee superintelligence projects. Any initiative surpassing a certain capability or resource threshold should be subject to inspection, audits, compliance testing with safety standards, deployment restrictions, and security protocols mandated by this authority. Monitoring and regulating compute and energy consumption could serve as initial steps towards realizing this idea. Voluntary compliance by companies and subsequent implementation by individual countries would be instrumental. The focus of such an agency should primarily revolve around mitigating existential risks rather than regulating the content generated by AI systems, which should be left to individual countries.
Moreover, developing the technical capability to ensure the safety of superintelligence is a pressing research question that demands significant effort. Researchers and organizations are actively engaged in addressing this challenge, recognizing the criticality of creating robust safeguards.
It is important to note that regulation and oversight should not impede the development of AI models that fall below a certain capability threshold. Companies and open-source projects should be allowed to innovate freely without burdensome regulations, licenses, or audits. While these systems carry risks, they align with the level of risks associated with other Internet technologies, and society’s existing approaches appear appropriate for managing them.
However, for the governance and deployment of the most powerful AI systems, strong public oversight is essential. Decisions concerning these systems should be subject to democratic processes, allowing people worldwide to collectively define the boundaries and defaults for AI systems. The design of such a mechanism remains an ongoing challenge, but efforts are being made to experiment with its development. Within these broad boundaries, individual users should have significant control over the behavior of the AI they utilize.
Given the risks and challenges involved, it is essential to consider why we are pursuing the development of superintelligence. OpenAI identifies two fundamental reasons for their commitment to this technology. Firstly, they believe that superintelligence will lead to a significantly improved world, as evidenced by early examples in education, creative work, and personal productivity. These advancements can help tackle societal problems, enhance creative abilities, and generate astounding economic growth and quality of life improvements.
Secondly, OpenAI argues that halting the creation of superintelligence would be highly challenging and counterintuitively risky. The cost of development decreases annually, the number of actors engaged in this endeavor is rapidly increasing, and it aligns with the trajectory of technological progress. Stopping superintelligence development would require a global surveillance regime, which itself offers no guarantee of success. Therefore, getting the development and deployment of superintelligence right becomes imperative in order to maximize its potential benefits while minimizing the associated risks.