Can a better understanding of human intelligence make for smarter machines?
We explored that topic this week with Rafael Reif, president of the Massachusetts Institute of Technology, following the announcement of the MIT Intelligence Quest, an effort to “discover the foundations of human intelligence” to develop better technology, especially artificial intelligence.
Reif has been MIT’s president since 2012. An electrical engineer by training, he has been outspoken in his defense of funding for basic scientific research. He was in Seattle this week to talk with alumni about MIT’s plans for the future of education, research, and innovation. We spoke about all of those topics, plus diversity in the tech industry and Boston’s bid for Amazon HQ2, on this episode of the GeekWire Podcast.
Listen to our full conversation in the player below, and continue reading for edited excerpts. Subscribe to the GeekWire Podcast on Apple Podcasts or wherever you listen.
Bishop: You’ve said the MIT Intelligence Quest will seek to better understand how human intelligence works, in engineering terms, and then take that understanding and figure out how to build wiser and more useful machines to benefit society. Why is this so important to MIT and to you?
Reif: It’s very clear that we have advanced those fields so much, that everything is moving in that direction. AI will be everywhere and will power everything. So, in talking to my colleagues at MIT, I learned not only that quite a few of them are working on understanding human intelligence '” it’s the only model we have for intelligence that we can study from '” and then many of my colleagues are using machine learning tools to advance their disciplines. But none of them are machine learning experts, they are electrical engineers or material scientists or climate action people. But they’re all using it because they need it to figure out all the massive information in the data they are dealing with.
So, clearly we needed to do a couple of things: Really power the advance of human intelligence, so that we know what kind of new algorithms we can use in the future. Right now, all the machine learning algorithms we have and those tools are terrific, are very powerful, but are really based on fundamental ideas that we came up with '” we collectively '” decades ago. We just have to double down on that. We have to replenish the well of ideas on AI and actually replenish the well of talent on AI. So, one point was to advance human intelligence so that we learn how to come up with intelligent machines. That’s one part of the initiative I launched.
And then the other part is: Can we create an interface, a group of people who can figure out how to use those advanced tools and customize them to be used in biology, in medicine, in engineering and so forth. That’s the essence of it. I think the goal is, by doing it that way at MIT, we’re basically doing what society is going to be doing eventually sooner or later '” and I think much more sooner than later '” and we’re basically creating or educating the people that will power that kind of evolution.
Bishop: There’s a lot of talk about machine learning and artificial intelligence across the tech industry, those are the big buzz words. People are using Alexa and Siri on a daily basis. But how close are we to real breakthroughs in generalized AI? Because that really is the holy grail of this field.
Rafael: Absolutely. As with every kind of breakthrough, that’s hard to predict. A spark of a new idea can happen at any moment. If I were to look at the evolution of these fields, maybe we are a couple of decades away from really coming up with a breakthrough idea, but how to put it together and implement it? So, it’s not here anytime real soon. But there is a fundamental question that is also important: Do we want to build AI tools that replace us, or do we want to build AI tools that understand us and work for us? So, conceptually there is a great deal of work to be done even defining what kind of algorithm we want to put together. The way we view it is: We want to create tools that can predict our behavior, can predict human behavior so that we can play together or work together with a robot, or another machine that does and says and practices exactly what we need them to do. That’s a kind of thinking of AI. So there are different branches of what AI means and how to use it. Of course, generalized AI’s the ultimate goal, but again, to what purpose? That part is a very important question.
Bishop: In a lot of ways, this leverages an interdisciplinary approach, like you can have at MIT, with life sciences and computer science, engineering, all coming together. If the MIT Intelligence Quest succeeds, what will that look like? What will be the outcomes? What are you driving for here ultimately?
Rafael: Well, let me just say, to answer your question … it will succeed. But let me make a point: It’s a new way of thinking. So, take science right now, whether it’s biology or physics or chemistry. The way we think today, we do one experiment at a time in which we work on a particular cell and understand that cell tremendously, or a molecule and understand it tremendously, or an atom. Machine learning tools allow us to do many experiments because we don’t have to worry about a very precise control of how a particular cell behaves and then do the next experiment because we learned something new. We can do a massive number of experiments with a massive number of variables, because we have the tools that can give us the information extracted. But how to design an experiment like that? That is basically a revolutionary way of thinking, of changing the way we think. What I’m trying to accomplish is just basically a paradigm shift in the pace of advancement of science by empowering them using this kind of new tool.
Bishop: Let’s talk a little bit about the political landscape. You’ve been very outspoken on the immigration changes and the threats to scientific funding at the national level. How much of this is bluster from the Trump administration and how