Surveys of leading AI researchers suggest a significant probability of human-level machine intelligence being achieved in this century. Machines already outperform humans on several narrowly defined tasks, but the prospect of general machine intelligence would introduce novel challenges. Superintelligent machines would need to be carefully designed to ensure that their actions would be safe and beneficial.
Present-day machine learning algorithms (if scaled up to very high levels of intelligence) would not reliably preserve the things we value, such as sentient beings living worthwhile lives. We therefore face a ‘control problem’: how to create advanced AI systems that we could deploy without risk of unacceptable side effects. If we can crack the control problem and build superintelligent systems that are aligned with our aspirations, it is likely the human condition will be dramatically improved.
The Alexander Tamas Initiative on Artificial Intelligence Safety supports three AI safety researchers and a programme and policy manager in pushing forward the state of the art in control problem solutions. Eric Drexler’s work focuses on the safe application, through carefully engineered capability control and monitoring, of very powerful narrow AI and superintelligent oracle (question-answering) AI systems. Stuart’s work, in collaboration with researchers at the Machine Intelligence Research Institute and Google DeepMind, explores methods for rendering superintelligent agents ‘corrigible’, allowing their designers and users to correct potentially dangerous behaviors at run-time.