Professor Stuart Russell, computer science professor at University of California, Berkeley, gave a clear and powerful talk on the promise and peril of artificial intelligence at the CSER in Cambridge on 15th May.
Professor Russell has been thinking for over 20 years about what will happen if we create an AGI – an artificial general intelligence, a machine with human-level cognitive abilities. The last chapter of his classic 1994 textbook Artificial Intelligence: A Modern Approach was called “What if we succeed?”
Although he cautions against making naive statements based on Moore’s Law, he notes that progress on AI is accelerating in ways which cause “holy cow!” moments even for very experienced AI researchers. The landmarks he cites include Deep Blue beating Kasparov at chess, Watson winning Jeopardy, self-driving cars, the robot which can fold towels, video captioning, and of course the Deep Mind system which learns how to play Atari video games at superhuman level within a few days of being created.
Until fairly recently, most people did not notice the improvements in AI because they did not render it good enough to impact every day life. That threshold has been crossed. AI is now performing at a level where small improvements can add millions of dollars to the bottom line of the company which introduces them. After self-driving cars, he thinks that domestic robots will be the Next Big Thing.
Professor Russell claims it is no exaggeration to say that success in creating AGI would be the biggest event in human history. He argues that pressing ahead without paying attention to AI safety on the grounds that AGI will not be created soon is like driving headlong towards a cliff edge and hoping to run out of petrol before we get there. The arrival of AGI, he says, is not imminent, and he won’t be drawn on a date: we can’t predict when the breakthroughs which will get us there will happen, he insists. But they might not be many decades away. Facilities like Amazon‘s Elastic Compute Cloud (Amazon EC2) keep changing the landscape.
The risk in superintelligence, he thinks, is less from spontaneous malevolence than from competent decision making which is not wholly based on the same assumptions that we make. His hunch is that achieving Friendly AI by constraining a superintelligence will not work, and instead we should work on directing its motivations – solving the value misalignment problem, he calls it. He is hopeful about techniques based on the idea of inverse reinforcement learning.
Professor Russell argues that AI researchers need to expand the scope of their work to embrace the Friendly AI project. Civil engineers don’t fall into two categories: those who erect structures like buildings and bridges, and those who make sure they don’t fall down. Similarly, nuclear fusion research doesn’t have a separate category of person who studies the containment of the reaction. So AI researchers should not just be working on “AI”, but on “provably beneficial AI”.
He urges the whole AI community to adopt this approach, and hope that AAAI’s willingness to debate autonomous weapons in January means it is relaxing its opposition to involvement in any kind of ethical or political debate.
