OpenAI presents the concept of super alignment, a significant step forward in managing the risks associated with superintelligence. Superintelligence, billed as the most impactful technology ever invented by mankind, could help solve many of the world's most important problems. However, the vast power of the superintelligence it could also be very dangerous, leading to the disempowerment of humanity or even human extinction.
OpenAI presents Superalignment: the attempt to "align" artificial intelligence with human intent
While superintelligence may seem far off, OpenAI believes it could be coming this decade. Managing these risks will require, among other things, new institutions for governance and resolution of superintelligence alignment problem: How can AI systems much smarter than humans be ensured that they follow human intent?
Currently, there is no solution to guide or control a superintelligence system. Current AI alignment techniques they rely on humans' ability to oversee AI, which will not be scalable to superintelligence. The goal of OpenAI is to build a automated human-level alignment seeker. To achieve this goal, scalable training methods need to be developed.
Read also: AI and AGI: Google CEO DeepMind predicts an imminent breakthrough
OpenAI is assembling a team of top machine learning researchers and engineers to work on this problem. 20% of available computing power over the next four years will be dedicated to solving this problem. The goal is to solve the key technical challenges of superintelligence alignment in four years. OpenAI invites all those who have been successful in machine learning to focus their attention on alignment. The fruits of this effort will be shared widely, and alignment and security of non-OpenAI models are important. It should be noted that the company will also work to improve the security of current models and address other risks from AI.
Solving the superintelligence alignment requires the best minds in the world. Open AI encourages interdisciplinary collaboration and active engagement with sociotechnical issues related to superintelligence alignment.
