Andrew Moore's career path at Carnegie Mellon has become emblematic of the way the University fosters its star talent. He became a tenured professor at Carnegie Mellon in 2000. In 2006, Moore joined Google, where he was responsible for building a new engineering office. As a vice president of engineering, Andrew was responsible for Google Shopping, the company's retail segment. Moore returned to Carnegie Mellon in 2014 as the Dean of the Computer Science department. In that role and given his experience, Moore is among the most influential people in the fields of computer science and artificial intelligence.
In the past, Moore has described the poaching problem that the Computer Science department has, given its stable of extraordinary talent in fields such as artificial intelligence, machine learning, robotics, and others that are high in demand. Of course, he recognizes himself in those professors and students who would choose to follow their passions into lucrative positions in the private sector. The department allows professors to leave and come back in many cases, and is hiring at higher rates in anticipation of this trend continuing.
In this interview, Moore offers insights into the evolving field of artificial intelligence, what is likely to be the factors to determine the companies who will win or lose in this space, as well as insights into what makes Carnegie Mellon specifically and Pittsburgh more generally a hot test bed for cutting edge technology.
(To listen to an unabridged audio version of this interview, please visit this link. This is the 24th interview in the IT Influencers series. To listen to past interviews with the likes of former Mexican President Vicente Fox, Sal Khan, Sebastian Thrun, Steve Case, Craig Newmark, Stewart Butterfield, and Meg Whitman, please visit this link. This is the 15th interview in my artificial intelligence series. Please visit these links to interviews with Mike Rhodin of IBM Watson, Sebastian Thrun of Udacity, Scott Phoenix of Vicarious, Antoine Blondeau of Sentient Technologies, Greg Brockman of OpenAI, Oren Etzioni of the Allen Institute for Artificial Intelligence, Neil Jacobstein of Singularity University, Geoff Hinton of Google, and Nick Bostrom of Oxford University, Jeff Dean of Google, and Andrew Ng of Stanford. To read future posts from either series, please click the link above to follow me on Twitter @PeterAHigh.)
Peter High: Andrew, you are the Dean of the School of Computer Science at Carnegie Mellon University. Please describe your purview.
Andrew Moore: Carnegie Mellon's School of Computer Science has a couple of hundred strong faculty members who are working on every aspect of computer technology. We also have a few thousand amazing students. My role as Dean is to make sure the whole organization gets to move forward. I see my role as helping to clear the way for these geniuses to get to do what they want to do.
High: You have said that being at CMU is like being at Hogwarts Academy. That when you walk around the School of Computer Science, the College of Engineering, and the university at large, you see a great number of smart people working on a variety of things that will change the technology landscape, and ultimately our lives. What was the origin of Carnegie Mellon's influence?
Moore: It all comes down to two visionaries, Allen Newell and Herbert Simon. They were two of the four people who, in 1956, took part in the Dartmouth Artificial Intelligence Conference, where they discussed what might be possible with computers in the future. These two gentlemen were in the business school at Carnegie Tech, which later became Carnegie Mellon University. There was, of course, not a computer science school in the 1960s. Newell and Simon used their passion and extreme intellect to speculate and bring together a team of people who looked at, not what computers would do in the next five to 10 years, but what it would mean to live in a world where there are thinking machines. They inspired so many other thinkers through that period that it snowballed over the decades. Today, we have 250 faculty members in the School of Computer Science. They work on everything from the lowest level details of how photons move and how you count them up, to the highest level details of what it means to have an emotional relationship with a talking machine. It was Newell's and Simon's initial interest that sparked this and shaped our computer science department.
High: As you pointed out, from its genesis, artificial intelligence has been a topic of great relevance. There are a number of related topics, like machine learning and robotics, which you and the School of Computer Science are involved with. Can you explain how they relate to each other?
Moore: Artificial intelligence is the science and engineering of making computers behave in ways that, until recently, we thought required human intelligence. In the 1990s, computers began to beat humans. In 1997, they beat the chess world champion. Until then, we had thought of chess as a uniquely human intelligence type of thing. In the following years, we discovered we were able to write computer algorithms to accomplish more things that we had thought needed human intelligence.
When I am running a software team that is building an AI, we always break the problem down and assign people to three sections: perception, decisions, and actions. Usually, when you are building an AI system, you have to figure out the perception: How is it going to understand the world around it? The folks working on decisions have to figure out: Given its ultimate goals, how is it going to search among lots of possible things that it could hypothetically do, to find the one which is best for the goals? The people implementing the actions figure out: How do you make it so the AI is able to interface with the world? That could be through talking to humans or through controlling valves in a factory. We look at this as a big loop. You perceive the current world, you decide, you act, you look and see what happened, and then you go around again.
That is the top-level architecture. I will jump into an extra detail regarding the decision part, which is where more mathematical types, like myself, are usually involved. It has an interesting history. Allen Newell and Herb Simon rightly thought about how to tell computers to make decisions. The successful approach to this is: The robot thinks to itself, hypothetically, “If I do this, I will predict what happens next, and see if that is any good.” And thinks, “But if I do this other thing, I will predict what happens next, and see if that is any good.” The general form of a decision making system is to imagine different things that could happen, predict their results, and then choose the one that is best. That worked extremely well, but only for problems where folks could write down, in software, prediction rules. This is why there was so much success in chess, the rules are completely clear. You know that if you move this pawn onto that square, it will take this bishop, and so forth.
The success with chess made us excited because we thought we could do this elsewhere, maybe for scheduling production in a factory or for figuring out how to do all the little pieces of work that need to be done to run a space shuttle launch. However, we quickly got into trouble, and AI took a few steps back in the late 1990s, just after the chess success. The problem was that there were so many things where we could not write down the prediction programs. For example, if you are trying to show good results in a search engine, things like AltaVista at the time, and you want to say, “If I show this, let me predict how happy the user is going to be with that answer,” everyone had trouble figuring out how to write down those predictions.
This is when the field of machine learning appeared on the scene and quickly exploded. It was the idea that you do not, in general, ask human programmers to write predictions programs about what will happen as a result of our actions. Instead, you let the system use data from its previous existence to program itself, if you like, and come up with a prediction thing that best matches historical data. In a nutshell, the history of artificial intelligence is: perceive, decide, act. The decide part of the systems looked great, but we get into trouble because we did not know how to write the ability to predict what happens next. During the 2000s, there was a massive explosion in technologies that helped the computers write their own programs to predict what happened next. That is where we are now.
High: Is the pace of change dependent upon computer power versus human ingenuity? Of course, there are numerous ingredients that come together to determine the pace of change, but what parts can you allocate to the aforementioned?
Moore: That is a timely question. We are at a stage where there are things that we want to accomplish, and are able to, with artificial intelligence. It becomes a question of how. You could quickly bring in an army of 10,000 people to work with the AIs to train them up. However, just when you are about to do that, you suddenly get paranoid because perhaps if you thought about it more carefully, you could just get 20 super smart people together and have them automate the thing that you are going to get 10,000 people to do. A great example is cyber warfare. Countries and governments have teams of cyber warriors, people who defend their nation's interests, and in some cases are ready to attack if they are attacked. The approach taken with cyber warriors has varied. Some societies say, "Let's train up 10,000 cyber warriors to manually go and do a big attack or a big defense." Versus the approach, which is more common in the United States, "Let's have a smaller group of folks who are writing the systems that automate what those 10,000 people would be doing."
High: In hearing you talk about the history and evolution of AI, it seems to be a topic where the goalposts keep getting pushed farther back. As things advance, it almost appears as though the practitioners, and the general public for that matter, take for granted the advances that have been made. Can you share your insights on where we currently stand with AI?
Moore: That is perceptive. It is almost like the process of charting out untraveled lands to find out which tasks can be automated. Interestingly, it is not happening uniformly. Not every human ability is falling prey to this. We can make computers good at things that mostly involve prediction and running over lots of options. The other things we do, especially anything that touches upon a concept like lateral thinking or analogical reasoning, where humans often solve problems by completely transcribing them into some other part of our existence, are areas where there has been little progress. I have no idea if it will be five years or 55 years until those sorts of advanced bits of reasoning fall. In the meantime, there is so much low hanging fruit around pieces of human intelligence that we do know how to automate, it is not clear if we need to be working on those more advanced pieces of cognition.
High: Given what you have described, what are some of the near-term advances that will impact our lives?
Moore: One thing that will change many consumer products and consumer experiences over the next five years is the understanding of emotions and emotional intent. Up until three or four years ago, the advances in computer vision and speech processing were around recognizing people, recognizing objects, and transforming spoken words into underlying written sentences. Now we realize we can go farther than that. For example, the cameras in modern cell phones have such high resolution that they can see little imperfections in the face and use them to track all the parts of skin as they move around the face. From tracking all the bits of the skin, you can work out what the muscles are doing under the face. From what the muscles are doing, using previous knowledge from psychology, you can detect facial action units and micro expressions to get information that we as humans are not even consciously aware of. This means that when in dialogue with a person, we can capture when they are excited, when they are happy, when there are fearful, or when there is a showing of contempt. It is a wonderful possibility, but the idea that we may be able to build computers that can do a good job of assessing aspects of our emotional state, can also feel like a Pandora’s box.
High: Much has been written about jobs that are likely to be impacted by automation, such as truck drivers and personal assistants. You and others have pointed out that it is not only those sorts of jobs that are at risk, but also some that require a tremendous amount of training, like jobs in the legal and medical professions. What sorts of skills do you think are not going to be replaced soon? Also, which ones will rise in importance, beyond those that require training in the subjects related to artificial intelligence, robotics, machine learning, et cetera?
Moore: There is a large class of jobs where, even if we could automate them, it would not make sense. A perfect example is a kindergarten teacher. The experience of being a child beginning to learn is all about the human-to-human interaction. It is not about soulless training. I do not see any reason that we would decide to replace humans in jobs where people interact directly with people. For the next 20 years, we will not have the technology to even remotely simulate the sort of caring and person-to-person interactions you get with people working together. Roles like kindergarten teachers, social workers, nurses, community policemen, and other folks whose jobs involve being part of a community, helping folks out, checking on things, those kinds of roles will continue to be important.
The most interesting thing is, if we have a huge productivity boost out of our GDP because of automation in many parts of the economy. Then, if we choose to, our societies can afford to aggressively fund things like elder care, social work, and education. Depending on how the economics and politics go, we could use the savings from the productivity increase to train more people for human-to-human interaction types of roles. We might be able to afford to have classrooms with one teacher for every five kids, instead of one teacher for every 40 kids.
High: In addition to your time in academics, you also worked for Google in their Pittsburgh operation. You have spoken about the AI race between some of the major technology companies like Google, Apple, Facebook, and Microsoft. As someone who has had experience within one of those companies and has developed partnerships with some of the others, what are some of the determining factors for who will win?
Moore: An essential part of that race is the business problem of who is going to develop the personal assistant that best helps you in your daily life. I am impressed with Amazon's work with Echo and Alexa, and with how Google, Apple, and Microsoft have all put impressive resources into their own versions of those two systems. When one of them becomes useful enough that you realize you are relying on it to get advice and to organize things in the same way that you currently do for the GPS navigation system in your car, the game will change. When that happens, if it happens for one of those companies before the other three, that company will, to some extent, have won. Because the next thing that will happen is the rest of industry, people who provide services like travel, health care, restaurants, and entertainment will want to be hooked up with the one AI system that is giving advice to everyone in the country. That is why the stakes are so high. Right now, there is not much clear evidence that one of those folks is miles ahead of the others. I do know a little bit about some of the technical differences they are using, and there will be a front-runner coming out soon. At the moment, they are all in the game. The winner depends on which group within those companies tries the right experiment at the right time and gets the breakthroughs in precision and recall that are needed.
High: One of the determinants is how much talent they can accrue. Each of those companies are gobbling up tremendous talent. Your department has not been immune to this. There has been a fair amount of poaching of strong academicians from CMU. You are in fact a representative of this, from your time at Google. This presents a challenge for the university, but on the other hand, the fact that companies are so attracted to talent from CMU must be a wonderful tool to recruit additional talented people to become students and professors. How do you think about the benefits versus the risks of the union between academics and the private sector?
Moore: We can and should help each other. One of the responsibilities that we at the School of Computer Science have as being part of this big ecosystem is to create the amazing people that will revolutionize the future world. There is a dark possibility that all of the great educators in AI and computer science will suddenly go work for the big tech companies and there will be no one left to educate the next generation.
We are trying to do something about that. Our approach is, No. 1, we bend with the wind on this. Rather than tell our incoming faculty they have to sign a pledge of loyalty to be an academic for the next 40 years, many of us try to show by doing. They see that a great career, and one where you can have the most impact, involves spending some time in university, and then taking your ideas out to the world. Carnegie Mellon provides a landing pad that they can come back to as they are rejigging their career or getting into a new emerging area. Faculty can go through those cycles in their career subsequently. The main cost of doing that, if we are careless about it, is that we end up with corridors of empty offices. We are aggressively responding to that concern. Last year, we hired 25 new faculty members into the college. This year, we have got an additional 22 faculty members. In our new reality, we have a large faculty, but they are not here all of the time. By being a part of Carnegie Mellon, they are free to pursue absolutely wild and crazy research ideas, and go safely off to make sure their ideas are delivered to the world, rather than just living in a journal or an academic publication. If they want to, and we are at about 50/50 on this, they can return later. When they return, they are enriched with new knowledge about what is going on [in the business world].
High: There is an analogy here with broader society. It would be short sighted to put roadblocks in the faculty’s way to try to keep them because as a result they probably would leave, or not come in the first place. Likewise, there are many who advocate for tighter regulations for artificial intelligence due to safety concerns, which in turn may slow the pace of advances. Is that a myopic argument? On a broader level, how real are the safety concerns with artificial intelligence and its advances?
Moore: There are two different kinds of safety concerns with artificial intelligence. It is important not to confound them into one concern. The practical engineering of safety critical systems is a huge and important thing that Carnegie Mellon and some of the other strongest computer science departments are taking action on. Examples of safety critical systems are modern AI technology that drives autonomous cars, helps make sure an aircraft lands safely, or guides a fine sensor into a human body to do surgery. Each of those safety critical systems performs much better if they are using an artificial intelligence type of approach that has machine learning in the middle of it. In theory, each of those things can help us save lives. We hope and aspire to a world in which the number of road deaths goes down by a factor of two, but a factor of three, or even four, is possible if the cars are able to take care of managing accidents, rather than relying on human reactions that are not fast enough. Good can come from this. Engineers know that when they engineer a safety critical system they have an ethical responsibility to perform large amounts of testing and formal proofs, like mathematical proofs, which show that what they are building is going to be safe in a wide number of circumstances. It is exactly like what a good architectural firm or construction firm does before it constructs a building. The firm goes through the proof that the building is going to survive whatever earthquake or other natural disasters come along. Engineers have an ethical responsibility to do that.
When it comes to systems that are learning, we currently only have weak mathematical theory to help the engineers prove the safety. Strengthening the theories is one of the important growth areas for academic artificial intelligence research. Like an architect can, we need to be able to write down a proof that under this whole big space of possible changes in the environment, the system will still work. Even if a system, on average, is going to save a large number of lives, you still have an ethical responsibility to make sure there are not bugs that will cause an accidental deployment that injures or kills someone. That is the primary concern and is a strong area of faculty hiring for us, and an area where we are growing our research and education.
The second type of safety concern is interesting for those of us in the software world since as the industry has grown, most of our early successes with artificial intelligence have been in non-safety critical industries. For example, when you are making a search engine or a social network more intelligent, it could have a serious impact on a person's life, but it is almost never going to directly kill someone if their email system goes down for a few hours. Take Facebook, they have a fun development methodology to keep all of their engineers moving fast: launch quickly, fix later. That is a perfectly good thing to do for entertainment or video games. It means that your engineers are innovative, trying out crazy new things, and having a lot of fun. However, no one would ever want to use that philosophy when you have a safety critical system. You have to launch carefully.
High: You are originally from the UK, but chose to become a U.S. citizen and to be a longtime Pittsburgh resident. You have said, “I am adamant the Pittsburgh region in general, and Carnegie-Mellon more specifically, are right in the center of all of this change.” We have talked about Carnegie Mellon and some of the things that make it special, what about the Pittsburgh region? It famously went through some dark times in the late ‘70s and early ‘80s with the decimation of the steel industry, which dominated the region and jobs. Through forward thinking from some of the people you already mentioned from CMU, in addition to strong governance from a succession of mayors of the city, the region has been working to transform itself into a technology hub, which has been realized, to some degree, in recent years. You could live anywhere in the world, and you have chosen to live in Pittsburgh. What makes the region so attractive to you and why are you bullish on it?
Moore: The first thing is, Pittsburgh has a positive and friendly culture. During its renaissance, the city’s old established neighborhoods welcomed new Pittsburghers to become a part of them. Many of us living in Pittsburgh live the lifestyle where our children can just wander around. They can walk through the high-density housing areas that we live in to visit their friends. It is the type of place where people just drop around each other’s houses randomly. Rather than the West Coast lifestyle where activities and social interactions are planned out because you have to schedule in the extra time to drive across town. Pittsburgh is high density living in an environment that has lots of academic, research, medical, and manufacturing types of people. Pittsburgh's pleasant lifestyle and the culture of being tight knit and raised around a community helps us recruit people.
The second thing is that Pittsburgh still has echoes from when it was one of the world's centers for massive big iron manufacturing. There is a culture of building stuff here. Building and creating big things is respected. Recently, when I was talking to some undergraduates about why they chose CMU, they said that when they visited the campus the professors and the students were talking about building things, as opposed to learning about or writing about things.
The third aspect is our city government. It is supportive of the kinds of high-tech things that need to happen next. For example, 15-20 years ago, a big turnoff was Pittsburgh's poor air quality. A significant number of people who we tried to recruit would not come because they were concerned about the poor air quality. The city, as a whole, has worked on that, and it has improved. It also helps that everyone sees us working together on making Pittsburgh a great place to live.
Peter High is President of Metis Strategy, a business and IT advisory firm. His latest book is Implementing World Class IT Strategy. He is also the author of World Class IT: Why Businesses Succeed When IT Triumphs. Peter moderates the Forum on World Class IT podcast series. He speaks at conferences around the world. Follow him on Twitter @PeterAHigh.
