AI & Humans Working Together

Welcome to Happens Next. My name is Larry Bernstein. What Happens Next is a podcast which covers economics, education, and culture. Today's topic is AI and humans working together, not independently. Our speaker is Tom Malone, who is professor of management at MIT's Sloan School and the director of the MIT Center for Collective Intelligence. Tom is the author of the book Supermines, the surprising power of people and computers thinking together. AI is the rage. And I want to learn from Tom about how we will interact with AI-based algorithms to become a supermind by combining the best that humans can do with the awesome computing power of AI working symbiotically together. Let's begin today's podcast with Tom's opening six-minute remarks. So let me summarize the two main messages in my book Supermines. The first is that I think we should be spending a lot less time thinking about people or computers, and a lot more time thinking about people and computers. Less time for us to think about how computers are going to take away jobs, and more time thinking about how people and computers together can do things that could never be done before. This second message is about how to see ghosts. Now, I don't mean real ghosts, of course, but I do mean powerful entities that are all around us and that are often invisible, unless you know how to look. These ghosts are what I call Supermines, which I define as groups of individuals acting together in ways that seem intelligent. Now by this broad definition, Supermines really are all around us all the time. For instance, every hierarchical company is a kind of supermine. A group of individuals acting together in ways that seem intelligent. Every democracy is a supermine, whether it's in a club or a company or some other kind of organization. A very important kind of Supermines are the markets for goods and services, and communities, whether it's a neighborhood, a scientific community, or some other kind of group. Now, here's something a lot of people don't realize. Almost everything we humans have ever accomplished was done by groups of people working together often over time and space. Computers have the potential to make these Supermines much smarter. Think of Wikipedia, where thousands of people and computers all over the world have created the largest encyclopedia our planet has ever known. I think that's a good example about how computers can make Supermines smarter, but it's also possible to use computers in ways that make Supermines stupid, like when fake news influences voters in a democracy. As fashionable these days to be pessimistic about how AI and other kinds of computers will affect society, but I have a more optimistic view. Computers have the potential to help us create human computer Supermines that are smarter than anything we've ever seen before. Computers can do the things they do better than people, like arithmetic and pattern recognition, and humans can do the rest. Perhaps even more importantly, we can also use the computers, the internet, and zoom to help create what I call hyper-connectivity, that is connecting people to other people and often to computers at much larger scales. Now, I think we often overestimate the potential of AI in all this, perhaps because it's easy to imagine computers as smart as people, because our science fiction is full of it. But unfortunately, it's much easier to imagine such computers than to actually create them. And I think it's likely to be at least many, many decades before we reach full human-level artificial intelligence. On the other hand, I think we often underestimate the potential of hyper-connectivity. Perhaps that's because in a certain sense, it's probably easier to create hyper-connectivity than to imagine it. For instance, we've already created the most massively hyper-connected group of people our planet has ever known, with billions of people connected to the internet. But I think it's still hard for us to imagine how to use this hyper-connectivity we already have. I think we need to move from thinking about humans in the loop to thinking about computers in the group. That's my message. A super-mind, our individuals working together to make an intelligent group. And in your book, you use Xerox Repairmen that interact using a message board to solve problems as an example of a successful super-mind group. Tell us about that study. Well, you might think that fixing copying machines is a one-person job, that the person who goes out there should have the knowledge needed to diagnose and repair these machines. But a former colleague of mine at Xerox Parkinet, Julian Orr did do an ethnographic study that you were referring to where he talked to these repair people and found that, in fact, the work they did was often much more cooperative, collaborative, collective than you might think. They actually ended up creating an online knowledge base of different kinds of problems and ways of solving them and gave awards to the people that came up with the most useful examples in that knowledge base. So that's an example of how it can make your organization, in this case, the group of repair people much more intelligent than they would otherwise have been. In your book, you describe five different superminds, hierarchies, communities, democracies, markets, and ecosystems. Tell us about that. I think one of the big points of the book is five different kinds of superminds for making group decisions. The hierarchy is our group decisions are essentially made by delegating them to individuals, lower in the hierarchy, and those decisions made lower can always be overruled by people higher in the hierarchy. That's a pretty effective way of organizing a lot of activities and making group decisions in a way that's pretty robust. Certainly worked well for us in the last century or so. The Xerox repair people were an example of what I call a community, a group of individuals that make decisions through informal consensus, often based on shared norms and reputations. That's also a very robust, in fact, that's probably the oldest form of supermind that we humans have used. But there are democracies where you actually have a formal voting process and the individuals or the decisions that get the most votes are the ones that are the decisions of the whole group. Another important kind of supermind is the markets that exchange all goods and services. Markets have an interesting way of making group decisions, which is the group decision is really just the combination of pairwise agreements between individual buyers and sellers. So they don't all have to agree on any one thing. If two people get together and agree they want to exchange things, they do. And the total of all those is the group decision of the market. So those are the four human forms of cooperation that do require at least some amount of cooperation to work. But if you don't have any cooperation among group members, then you have the fifth kind of supermind, which I call ecosystems. So in an ecosystem, the group decision is made by the law of the jungle. Whoever has the most power gets what they want and the survival of the fittest. So I would say those five types of superminds for making group decisions account for the vast majority of things we see in the world around us. Almost everything we see, what happens in a company, what happens in a country, those are all combinations of those different kinds of superminds. I spent my career trading fixing come securities. And the amazing thing is all you need to know is the price and the description of the security to invest. You don't need to know the seller, why it's for sale or any other aspects. It's just price. Tell us about the role of prices and markets. Absolutely, you've just described the magic of the market or what Adam Smith called the invisible hand of the market. And it is hard to believe and almost magical if you think about it, that this very, very decentralized form of decision making can do a pretty good job of allocating resources, goods, people's time, et cetera, in a pretty efficient way. So if those are commodities, then those descriptions are well known. But if you're trading things that aren't commodities, each thing has some individual characteristics like the services of a particular human with expertise of a particular kind of automobile with all kinds of special options, you need to know what it is that's being sold at that price. But I think that combination of prices and product descriptions does allow a very information efficient and decentralized way of making lots and lots of very complicated decisions with pretty good outcomes. Wikipedia is a fantastic example of a community supermind. I heard that one day one of the random contributors downloaded all the US census data for every town in America to the Wikipedia sites for each individual town. He's in a community that applies knowledge and technology to make a superb final product that is constantly evolving. Well, I think that's a great example, one that's been a favorite of mine for a long time. Wikipedia is primarily a community. So it operates with a kind of informal consensus. Essentially, anybody on Wikipedia can change anything anytime they want to. And if somebody else thinks it should be changed, they can change it too. And so it keeps changing until everyone who cares enough to look is satisfied with what's there. So that's a literal consensus. Part of the secret of Wikipedia was that they used those well-known aspects of community decision making like norms about what's good, what kind of articles we need, what things we shouldn't have in articles. In fact, Wikipedia codifies many of those norms very explicitly on the site. So people who edit Wikipedia have repetitions. It seems that within the Wikipedia community, part of what motivates people is their desire to have a reputation among people they care about in the community of Wikipedia editors. What's unusual about Wikipedia is the degree of which they've used modern communication and computational technology to let a community operate at a scale and in a way that would have been unimaginable before the number of active Wikipedia contributors is something like 50,000, some big number of people. So imagine you had say 50,000 people in a giant football stadium and they all had paper and pencil and they were trying to write an encyclopedia by consensus. It would have just been impossible for enough people to see all the things people wrote. They would have probably resorted to something more like a hierarchical structure with editors and sub editors and so forth. And it seems to me almost impossible that they could have done anything remotely as good and as fast as Wikipedia. Years ago, I was watching the Chicago Bears in a playoff game and the starting quarterback, Jay Cutler got injured on a play and his back up entered the game. I didn't know anything about his replacement. So I looked up on Wikipedia and I had always danced from college and pro football, but most incredibly, it seemed that he had just entered that playoff game 15 seconds ago. This truly was a real time encyclopedia. What are the implications of being up to date to the minute? Yes, absolutely. So it makes it possible to make so many things so much more widely available so quickly. Another point here is that if you didn't have the communication editing technology that Wikipedia is built upon, you probably couldn't have anything like that larger scale of community operating effectively. You would probably have had to resort to some kind of hierarchical structure to reduce the amount of communication that's needed. One of the big lessons here is that the cheap communication and also computation makes it possible to organize collective human activity in ways that are not only faster or wider or bigger, but often it makes it possible to organize them using a community instead of a hierarchy, using a market instead of democracy. We can do things faster and really different, more decentralized ways. Some organizations, like Wikipedia, are successful because they have rules and norms that allow for growth and use decentralized expertise to solve problems. Occupy Wall Street failed maybe because it lacked a hierarchical structural and community norms to succeed. What norms were missing from Occupy Wall Street that undermined its efforts? I used the example of Occupy Wall Street in my book and it's a cool example in part because I was there personally to witness part of what went on. I was on sabbatical at NYU at the time of Occupy Wall Street and a couple of weekends I went down and listened to the protesters discussing things. Most interesting session I went to was one where several dozen people were trying to come up with a mission statement for Occupy Wall Street. So anybody could walk in off the street and be part of it. In fact, Michael Moore, the movie director was one of the people sitting there in the room with me and a bunch of other people. When I came in, they were very near the beginning of the proposed mission statement. So far they had agreed on saying, we believe in a fair and just society. Somebody said, I think it should really be, we believe in a truly fair and just society. So they were following a set of fairly rigid norms for consensus decision making. That included things like almost everyone has to agree. And if there's anyone in the room who feels so strongly about it that they would leave the group, if this decision is made, then it's not a consensus decision that you have to keep going until you find something that everyone will agree to or at least be willing to not leave the group. So given those norms of consensus decision making that the occupied members were following, they started discussing this question of whether we should add the word truly to the mission statement. And there was one person there who said, if you do do this, I would leave the group. And people went back and forth for at least two hours trying to decide whether to use the word truly in there. So a person who was originally blocking the consensus, eventually said, oh, well, never mind. It's okay with me, you can go ahead. But then there was a long debate about whether the rules allowed that to happen. Very little of the discussion was actually about, well, is the word truly a good thing here or not. So I left there feeling very pessimistic about the possibility that this group would ever come up with a coherent mission statement. Years later, when I was writing my book, I scanned the web as well as I could to see if I could find any mission statement for occupied Wall Street and couldn't find one anywhere. So I think it's likely that they didn't manage to do that. Which I think was a pity. They were certainly very well-intentioned. And then I thought later, as I was writing my book, if this group, instead of trying to do everything in a face-to-face group, had been able to use the technology of Wikipedia, they might well have come up with a very good mission statement. That's an example of how new technology can make new ways of organizing human groups. Life is complicated. And institutions work for reasons that are not obvious. And they get the kinks out. Parliamentary procedures are a fast and easy way to manage a group meeting, instead of coming up with something new that seems better or fair. The point you're making is when I completely agree with that norms can be very useful, very valuable. And we need them. The norms are, in a sense, the key components of communities as decision-making mechanisms. Humans, ever since our days of hunting and gathering bands, have used norms as a very important way of organizing our new activity. Some norms are kind of arbitrary. It doesn't matter that much what they are as long as everyone agrees on them. Like, how many days should there be in the week? There's no single right answer to that. It just should be something that everyone agrees on and similar for at times. If you had to choose between parliamentary procedure and consensus decision-making, and if you were in a large group, especially a group of people who may not have been kind of aligned in terms of what they wanted, then it's almost certain that the parliamentary procedure methods would work more effectively than the group consensus methods. But if you're in a group where deep dedication to the group decision is critical, like if you're about to go into war and everybody's life is going to be on the line, one way of getting deep commitment of the group members is if everyone essentially had to agree on something before the group did it. There are trade-offs. I'm not saying that either one is always better. Depends on the situation. Depends on how much. Dedication is needed from the people in the group. Depends on how quickly you need to make the decision. In the case of the occupied protesters, they were using a very formal and rigid form of consensus. Most communities have a more flexible, informal kind of consensus decision-making. That seems to work pretty well in a lot of cases. How AI changed the nature of work? The main way that computer technology has changed things is by reducing the cost of communication. It made it possible for people all over the world to communicate much faster, much more cheaply, much more easily. Now, with a new generation of generative AI, it will be possible for computers to participate in ways that weren't ever possible before. Computers will be able to play a role like what used to be played by humans. For instance, in Wikipedia, the computer technology will allow the use of little bots, little programs to find curse words or other objectionable content and remove it from Wikipedia articles automatically. They've also had simple tools that would correct grammar at various places automatically. So I think it will now become possible for more intelligent bots based on generative AI technology to understand English and respond in English in ways kind of like what a human would do. Moore's law is an observation that the number of transistors in an integrated circuit doubles every 18 months of two years. And this doubling also increases processing speeds and database sizes and reduces the cost of computing. How will Moore's law impact the power of AI in the future? So there are several things going on here. One is Moore's law was in fact a remarkably stable and accurate prediction of progress in computer technology for many decades long after people thought it would last. In other words, people were always saying, well, now we've reached the end of Moore's law. But then some new technology would make it possible to keep going. It's not, by the way, a law of nature or a law of physics. It's more a description of a very complicated human economic technological process of how these things get made and how they get priced and so forth. It's been a remarkable phenomenon. I'm not a deep expert on this, but my understanding is that Moore's law isn't operating as reliably in the last few years as it didn't for many decades before. One way we've dealt with that, however, is by going to parallel processing in a much bigger way. Moore's law is about how fast a single processor can process things or how much memory a single processor can have. But it's always been possible to have multiple processors working in parallel. They can't do everything that a single processor can do as fast. In other words, if you make a single processor five times faster than it can do anything five times faster. If you have five parallel processors, for some tasks, they can do things five times as fast, as long as those tasks are all parallelizable. In other words, you can do things without worrying about the things going on in the other four processors. In general, you can't do everything that way. But there are some things like graphic processing, which can be done in a highly parallel way. And the new AI chips that Nvidia and other companies are selling originally were created or designed for graphical processing. But now they can be used for many of these AI programs, which are highly parallelizable. So that's a way of getting around Moore's Law. Even if you can't make the individual processors twice as fast to every 18 months, you can just create many more of them and use software algorithms that allow things to be done in parallel to a large degree. And that's a big part of how a lot of these recent generative AI programs have been able to do what they do. How will these more powerful computing machines change AI's performance? Nobody knows for sure. There are two possibilities. One is that to go from the AI we have today to human level of AI, all we need is more processing power. So basically, the deep learning algorithms that these systems use depend on adjusting billions of parameters through learning from many more billions of examples of text and image and so forth. What they found is that when you use the same algorithms, but you just have more parameters, you get better performance. So again, I'm not a deep expert on this. I don't think this is likely, but it's possible that just getting enough more parameters will make these current algorithms so smart that you can't tell them apart from humans. I think some other much deeper algorithmic advances will be needed, but at least to some degree, just having more and more processors that can manage more and more parameters will almost certainly continue to increase the intelligence of these AI algorithms. How should we think abstractly about how AI works? Yeah, well, here is a kind of intuition for how these current, generative AI systems work. A lot of people, when they see a program that you type something in English into it, and it comes back with a flexible, often very articulate answer in English, they think, wow, that sounds just like a human. It must be a human, but I think a better model for how these things work is not that they're like humans underneath, and they have human emotions and human motivations and stuff. It's like what they really are is just much bigger versions of the autocomplete that you have when you type in something in the Google search bar. So basically, what these algorithms are trying to do is, given the words that are here so far, they're basically just trying to predict what's the most likely next word. So it turns out that when they have these billions of parameters that are learned to recognize various kinds of patterns, they can do an amazingly good job of predicting the next words. They must have learned a lot of things about how the world works and how humans talk about things by capturing those very complex patterns in probabilistic form. But they're not thinking like humans do, or at least, they're not thinking like humans do consciously. An interesting possibility is that the way these current generative AI systems work may be more like human unconsciousness, or unconscious thinking, may work more like those algorithms, and our conscious thinking, where we're doing logical reasoning that we can explain in words, is more like what was called classical AI. That is, if Julius Caesar was a Roman and all Romans drink wine, then Julius Caesar must have drunk wine. So that would be an example of a logical syllogism that classical AI was all about. The recent progress in AI has been driven largely by these machine learning systems, very deep neural nets, that have these billions of parameters that learn things, but in a way that's very hard, even for the human programmers who create these systems to understand. The reason you'd said this word at that point is because these billion parameters had a greater value for that word than any other word. That's not a very satisfying explanation to humans. It's not the way we think we think. But much of what we humans do really occurs at a kind of unconscious level that we can't describe at all. For instance, how do you recognize your mother's face? You can say something about it. Well, she had dark hair and a short nose or whatever it is. But it's very difficult to describe another person's face in a way that someone who's never seen the person before would recognize them just from your description. So the facial recognition we do occurs at a kind of unconscious level that we can't articulate explicitly in a way that other people can understand. And these new systems are maybe capturing that human intelligence better than their capturing the intelligence that we can think about and talk about consciously. Some of my colleagues, like Josh Tenenbaum and MIT, have talked about what's called neuro-symbolic computing, a combination of the neural net approach, which is what today's most advanced systems are using. And the logical, symbolic kind of reasoning that classical AI used, maybe what we need is some combination of those two kinds of computing. And that might be more like what we humans actually do. We can't talk very much about the things we do in the purely neural way. But we must have some kind of combination of those two kinds of reasoning going on in our minds. How can AI help us make better decisions and make us more productive? The much more powerful changes will result not just from simple substitution of automated processing for human labor, but from rethinking how we do things in the first place. Take Wikipedia as an example in the non-economic sphere. Before Wikipedia, the way Encyclopedia's got done was through a hierarchical process of editors and reviewers and expert contributors. Wikipedia made it possible to do a far better, far bigger, far more accessible Encyclopedia in a completely different way. So how can we invent new organizations that do things in very different ways, create very different kinds of products, but take advantage of the potentials that these new technologies make possible. We need to invent new technologies that do new things far better, far cheaper, or whatever. I think it may be just as important for us to think of innovative new ways of organizing human work, innovative new ways of producing the same old products and services better and producing new products and services that couldn't even be done before. Another example of that would be Google Search. Google Search is a service that would have been completely impossible 30 years ago. It's now one of the biggest companies because it figured out how to do something that was very valuable for people in a way that couldn't have been remotely done before, but it now is. So I think it's figuring out how to do things like that that will be needed to get the big payoffs from these new technologies in terms of productivity. As I get older, I find it more difficult to adapt new technologies. What does this mean for society, for the adaptation of AI? Yeah, so I think you put your finger on another one of the factors that's needed for us to take advantage of the real potential of these new technologies. It's not just reinventing organizations and products and services, it's also changing people's ability to work with and take advantage of the new potentials on an individual basis. So that's one reason for optimism about how it may not take forever for some kinds of changes, but new generations may be needed before they become widely adopted. I end each episode with a note of optimism. What are you optimistic about as it relates to AI and technology? I'm optimistic that we will figure out how to do new and better things with AI, not just cheaper than what we used to do, but new ways that people and computers together can do things that were never possible before. I think 10 years from now, it will be hard for us to imagine how we ever survived without AI blah blah blah. I can't tell you exactly what AI blah blah will be, but that's my optimistic expectation for the future. Thanks, Tom, for joining us today. If you missed last week's show, check it out. The topic was gerrymandering, as well as economic liberty. Our first speaker, retired federal judge Gary Feynerman, discussed the recent case in North Carolina related to gerrymandering, and whether the state courts can be the final arbiter for state redistricting maps. Our second speaker was Renee Fletherty, who is an attorney who successfully argued a case in the Georgia Supreme Court. This is about the limits of state authority to regulate occupations. Georgia demanded that women who teach new mothers how to breastfeed meet minimum education requirements that would prevent otherwise qualified women from counseling lactation care. Renee is an attorney with the Institute for Justice, a not-for-profit that challenges government overreach on licensing and regulation, as well as infringement on individual rights. You can find our previous episodes and transcripts on our website, what happens next in six minutes.com. Please subscribe to our weekly emails and follow us on our podcast or Spotify. Thank you for joining me. Goodbye.