ai.bythebay.io: Stuart Russell Interview
Recording: ai.bythebay.io: Stuart Russell Interview
you [Music] you so I cover a pretty broad range of things and the definition is pretty broad it means making machines do smart things and smart means roughly speaking more likely to achieve the Machine objectives or the objectives that we put into the machine I think there's a lot of progress happening with machine learning techniques and those have been very successful mainly at the perceptual level there's also a lot of other stuff going on in reasoning planning search game playing so when you look at alpha go which feed the human has many of the best players in the world what alpha go does is a combination of offline machine learning but actually a lot of online search so it's a you know arguably a natural successor to techniques that were explored starting the 1950s so I think one of the things I really hope people understand is that AI is a very broad and deep discipline with roots going back decades and the current interest in deep learning is a is a facet of the field which can help when you combine it with lots of the other techniques but by itself at the moment isn't enough to solve a lot of the problems we care about so I have sort of two two threads of my web one is actually making AI SS improving its capabilities and and there I work on long-range decision-making so how do we make decisions of a long time scale particularly under uncertainty and also the combination of knowledge representation of machine learning so we developed what we call open universe probability models which are very expressive formal languages for writing probability models that have they have full expressive power relative to Turing machines and so you can use those to express lots of prior knowledge and then combine that with data and get really good prediction performance and then the other threat of my work has to do with the long-term consequences of AI so the main issue that I see for the human race when we develop systems that are more intelligent than us is how do we make sure that the decisions those systems make are actually beneficial for us and the most obvious failure mode that we give the machines and objective that sounds reasonable to us like you're in cancer but when you actually optimize that objective using much more intelligence than we have access to then you may see unintended consequences that could be irreversible for us so that to me is a very serious issue it's not an immediate issue but it's an issue that we don't know how to solve and we don't know how long it's going to take to solve and if we haven't got a solution by the time that super intelligent capabilities come on stream then we may face actually a very unpleasant set of consequences so my goal is to figure out how to how to make a I provably beneficial which means that you know even though we may not be able to explicate our own objective the AI system should behave in such a way that it is consistent with those objectives and essentially makes attacking so that that's the goal and it's I think it's a goal that solvable I really believe that with the right conceptual framework and one of the main underpinnings of that framework is uncertainty and objectives that the machine should be explicitly uncertain about what its objectives are and those objectives are really to make the human happy it explicitly doesn't know exactly what that means and so it's job is to learn what it means and in the meantime to be cautious and only to act in ways that it's fairly certain will be calculus so I got some great audience questions I think so it looks like the people in the audience are very current on what's going on the field and very interested in the topic and I'm looking forward to the rooftop party all right thank you