If AI was potentially dangerous, why couldn’t we just build it on virtual machines to separate it from other computer systems? Surely that would stop the AI from infiltrating our physical infrastructure and shutting down an electrical grid or manipulating financial markets. Yudkowsky immediately had an answer. “It wouldn’t really be virtual,” he replied, sipping his drink. Electrons could flow in all sorts of different directions, which meant there was always going to be a way for powerful AI systems to touch and change the configuration of hardware.
