We believe superintelligence could arrive within the next 10 years. These AI systems would have vast capabilities—they could be hugely beneficial, but also potentially pose large risks.
Today, we align AI systems to ensure they are safe using reinforcement learning from human feedback (RLHF). However, aligning future superhuman AI systems will pose fundamentally new and qualitatively different technical challenges.
Superhuman AI systems will be capable of complex and creative behaviors that humans cannot fully understand. For example, if a superhuman model generates a million lines of extremely complicated code, humans will not be able to reliably evaluate whether the code is safe or dangerous to execute. Existing alignment techniques like RLHF that rely on human supervision may no longer be sufficient. This leads to the fundamental challenge: how can humans steer and trust AI systems much smarter than them?
This is one of the most important unsolved technical problems in the world. But we think it is solvable with a concerted effort. There are many promising approaches and exciting directions, with lots of low-hanging fruit. We think there is an enormous opportunity for the ML research community and individual researchers to make major progress on this problem today.