Leopold Aschenbrenner, a pivotal member of OpenAI's Superalignment Team, was dedicated to ensuring the safety and control of advanced AI systems, closely collaborating with team leaders Ilya Sutskever and Jan Leike. However, in May 2024, following leadership changes and restructuring at OpenAI, Aschenbrenner, along with other team members including William Saunders and Pavel Izmailov, left the company. Following his departure, Aschenbrenner founded his own investment firm focusing on advanced AI systems (AGI), attracting notable investors such as Patrick Collison, John Collison, Nat Friedman, and Daniel Gross. Previously, he conducted research on economic growth at the Global Priorities Institute at the University of Oxford, establishing himself as an expert in AI safety and control.
Aschenbrenner predicts that Artificial General Intelligence (AGI) will be achieved by 2027-2028, operating at a level comparable to the best human experts. Through continuous training and optimization, these models could evolve into superintelligence, surpassing human capabilities in almost every domain, potentially reshaping the global balance of power significantly.
A critical aspect of Aschenbrenner's analysis is the substantial technological and infrastructural investment required for developing superintelligence. He explains that by 2030, the size of training clusters will increase exponentially, necessitating enormous computational power. Such clusters could consume up to 100 gigawatts of electricity, more than 20% of the total US power production. This development will profoundly impact how businesses and governments operate, with early investors in these technologies gaining substantial competitive advantages.
From a geopolitical perspective, controlling superintelligence will become a crucial factor in the global power balance. Countries that dominate these technologies will significantly enhance their economic and military power. Aschenbrenner emphasizes that China has made considerable progress and might surpass the US in the long term due to its impressive industrial capacity, enabling rapid mobilization of large resources. This could be decisive in a prolonged technological arms race.
However, the development of superintelligence poses significant security risks. Aschenbrenner warns that authoritarian regimes could exploit these technologies to consolidate and expand their power, potentially leading to new weapon systems and other dangerous technologies that are difficult to control. Of particular concern is the possibility that China, through state-supported espionage, could acquire critical technological information. The Chinese government is making substantial efforts to infiltrate American AI labs and steal key algorithms and model weights, allowing China to quickly close the technological gap or even take the lead.
To address these challenges, Aschenbrenner calls for stringent security measures and increased international cooperation. American companies and research institutions must significantly enhance their security precautions to resist state-supported espionage, including intensive surveillance measures, air-gapping sensitive systems, and strict security controls for employees. Simultaneously, it is crucial to build and protect large data centers and computational capacities within the US to maintain technological supremacy. Establishing such infrastructure in authoritarian states is considered a significant security risk.
Kommentare
Kommentar veröffentlichen