How do computers think, and how is that changing? For a peek into the ethics and governance surrounding tomorrow’s advanced computing models, join us with Anthony Scriffignano, Chief Data Scientist from Dun & Bradstreet, and Stephen Wolfram, the creator of the powerful computing system Wolfram Alpha. Their deep perspectives have implications for both policymakers and corporate strategists.
How do computers think, and how is that changing? For a peek into the ethics and governance surrounding tomorrow’s advanced computing models, join us with Anthony Scriffignano, Chief Data Scientist from Dun & Bradstreet, and Stephen Wolfram, the creator of the powerful computing system Wolfram Alpha. Their deep perspectives have implications for both policymakers and corporate strategists.
Anthony Scriffignano has over 35 years experience in information technologies, Big-4 management consulting, and international business. Sciffignano leverages deep data expertise and global relationships to position Dun & Bradstreet with strategic customers, partners, and governments. A key thought leader in D&B’s worldwide efforts to discover, curate, and synthesize business information in multiple languages, geographies, and contexts, he has also held leadership positions in D&B’s Technology and Operations organizations. Dr. Scriffignano has extensive background in linguistics and advanced computer algorithms, leveraging that background as primary inventor on multiple patents and patents pending for D&B.
Stephen Wolfram is the creator of Mathematica, Wolfram Alpha and the Wolfram Language; the author of A New Kind of Science; and the founder and CEO of Wolfram Research. Over the course of nearly four decades, he has been a pioneer in the development and application of computational thinking—and has been responsible for many discoveries, inventions and innovations in science, technology and business.
Michael Krigsman: Welcome to Episode #227 of CxOTalk. I’m Michael Krigsman. I’m an industry analyst and the host of CxOTalk. I want to thank LiveStream for supporting us with just really amazingly great video infrastructure and streaming. Today, we have … Oh! And I forgot to say that if you go to LiveStream.com/cxotalk, they will give you a discount. So, do that. Livestream.com/cxotalk.
What an amazing show we are going to have today! When we talk about the future of computing, there’s a risk that it’s going to sound like a pretentious topic, but today, with the two guests we have, it’s actually a realistic and very fascinating topic to discuss. We’re going to be speaking today with Anthony Scriffignano, who is the Chief Data Scientist of Dun & Bradstreet. Anthony has been a guest here a number of times in the past. Anthony, how are you? How are you doing?
Anthony Scriffignano: I’m doing very well, Michael. Thank you very much.
Michael Krigsman: Well, thank you for being here again on CxOTalk. And, we’ll also be speaking with Stephen Wolfram, who is truly one of the fathers of modern computer science. And, I don’t think that’s an exaggeration. Stephen Wolfram, thank you! This is your first time here. Thank you for being with us!
Stephen Wolfram: Thank you!
Michael Krigsman: So, just to jump in very quickly with some brief background introductions, I’ll ask Anthony just to tell us who you are and what you do.
Anthony Scriffignano: So, very, very quickly: I’m the Chief Data Scientist at Dun & Bradstreet, and in my role, I’m responsible for looking at technologies and capabilities that are sort of on the edge of computer science – hence, our interest in the topic today – not necessarily things that are common practice, and sometimes not even things that have words to describe them. And then also, I work with governments around the world as they develop legislation around things like data privacy and data localization, and so forth, to share best practices and make sure that we're doing the right thing for the community at large, and the business community.
Michael Krigsman: Fantastic! And, Stephen Wolfram, please tell us a little bit about yourself. In a way, you need no introduction, but I think an introduction is always good.
Stephen Wolfram: Well, I’ve kind of alternated between doing basic science and developing technology for the last many decades. I run Wolfram Research, a company I started 30 years ago, and we’ve mainly done three things: Make a product called “Mathematica” that gets widely used in research and development and education; about 90% of US universities now have licenses for it. We make a thing called Wolfram Alpha, which is a system that answers questions, and provides, for example, the knowledge system for things like Siri; and most recently, we’ve been sort of owing all that technology into a thing we call “Wolfram Language,” which is kind of a new generation computer language that has the main objective of sort of building in a lot of knowledge right into the language so that it provides the highest level possible platform from which people can build things. And the exciting thing for us, in most recent times, is kind of the deployment of Wolfram Language not just in our traditional research and development, and consumer spaces, but also very much in the enterprise space, and for software development purposes, and so on.
Michael Krigsman: Okay. So, thank you so much. To begin, I think if we’re going to talk about the future of computing and as Stephen and Anthony were saying just before we came live, it’s a very big topic to talk about in 45 minutes. Maybe, we can begin with artificial intelligence, because it’s such a popular topic today. And, Anthony, how about you take a stab initially to define some terms so we have common ground. When you think about AI, how do you define it and think about it?
Anthony Scriffignano: Well, first, I would say that there isn’t one commonly-accepted definition. Second, I would say that there’s often very little intelligence in artificial intelligence. There are different types of technologies and capabilities. Many of them share the fact that there’s some type of goal that they’re trying to reach using different methodologies. Some of them are what we call "regressive methods," things like machine learning where they look backward at data that has pre-existed – the incidence of the application and trying to project that forward into what happened. Some of the non-regressive methods are what we call "neuromorphic methods," these are methods that are designed to mimic how we think the brain works. And then, more recently, we've seen things like cognitive computing, which are methods that work alongside an intelligent user to help that intelligent user to reach that goal in a more efficient way and learning from that behavior watching that user.
The other thing I would say about artificial intelligence is that there is a lot of technology that underpins it that people normally lump in; for example, Natural Language Processing. Some would argue that there are elements of artificial intelligence in that, and I would agree with them, but that is not an inclusive definition. So overall, some sort of goal is usually not one that the machine will modify. Be very afraid when they start doing that. And then, either a regressive method or a non-regressive method, and then either working alongside a use, which is sometimes called “heuristics,” or working separately and giving the user the answer. That’s my shortest possible definition that I’m comfortable giving in this form.
Michael Krigsman: Stephen Wolfram, how can we make this understandable to the average person who’s not a computer scientist, but who wants to understand what all of this actually means and the implications of it?
Stephen Wolfram: Well I think, artificial intelligence, as now discussed, and I’ve watched its evolution over the course of nearly 40 years now, it’s really an extension of a long-running story of technology now, which is “How do we take things which we humans have known how to do, and make machines do them for us?” And typically, the pattern is, in the end, we humans define the goals that we’re trying to achieve, and then we want to automate as much as possible getting those goals done. And over the course of the last few decades, there have been all kinds of things where people have said, “Gosh! When machines can do this or that particular thing, then we’ll know that we’ve really achieved artificial intelligence.” It’s always a bit disappointing because, in the end, when one thinks it’s sort of something that’s truly a special human thing, in the end when it gets done by machines, it’s just code underneath.
So, a great example of that is what we did with Wolfram Alpha, where we're able to answer all sorts of general questions about the world; you know, "What was last year's revenue of IBM," or let's take the GDP of such-and-such a country and compare it with the GDP of some other country, or figure out if your uncle's, uncle's, uncle's son, what relation to you is that? These kinds of natural language questions that one asks and then being able to answer those questions on the basis of knowledge that our civilzation has accumulated, that was one of the kinds of characteristic, "When you can do this, you've got AI," kinds of directions. And then, when we brought out Wolfram Alpha eight years ago, it was like, "Okay, we can now do a pretty good job of this," and if we look at how it was done, part of it is we're kind of leveraging all of the knowledge that our civilization has accumulated and turned it into something that a computer can deal with. Making it compute-able. That's a large part of it.
There were some other things that did come more from a sort of basic science point of view of being able to understand natural language. It wasn't obvious it was going to be possible to understand the typical random things that people ask their phones and so on. that had been a problem people had been working on for a long time, turned out - I hadn't really realized this – but it turned out the key extra ingredient that one needed to do good natural language understanding was not just being able to pick apart pieces of English or whatever language one's dealing with, but also having a lot of actual knowledge about the world, because that's what allows one to determine if somebody says "Springfield," for example. You have to realize, "Well, which Springfield are they probably talking about?" Well if you know about all the different Springfields and you know what their populations are, how popular they are on Wikipedia, where they are relative to the person asking the question, then you can figure out what they're talking about. But without that underlying knowledge, you can't really do that.
Anthony Scriffignano: Maybe, yeah, if I could just jump in on what you just said, Steven, because there are some really interesting things in there that I would love to unpack. The first thing is, you know, last year’s revenue at IBM. At Dun & Bradstreet, we obviously look at things like that and it’s easy enough to answer for a public company that files 10-Q’s and 10-K’s, and when you start asking about Bob’s Hat Company, or you start asking about a company in China and you maybe want to ask the question in English, and the answer to the data is in Chinese, it starts to get complex for reasons that transcend Natural Language Processing. We have laws in different countries about what data can cross borders. Those laws are constantly changing. There are privacy laws; there are considerations about data sovereignty, where the data lives, and so forth.
And so, it’s no small feat that you’re talking about. You’re making sound easy, but the orchestration of curating all of that data in different places, and then orchestrating an answer in, I’ll use the phrase “real-time” in quotes, because there’s no such thing, but orchestrating the answer in an amount of time that we’re comfortable with, is no small feat. None of us would believe that the truth is out there on the internet all the time, and yet sometimes we behave that way. And so, just adjudicating the truth is another challenge in there. It’s sort of as you unpack this thing, you get more and more surprises, and it becomes a more curiouser and curiouser world. And so, part of it is making it look more intelligent, and part of it is giving it an intelligent, empirical answer that you can scale and reproduce and learn from. Those all kind of boil together in what you’re talking about.
Stephen Wolfram: You know, maybe one thing I might discuss is the question … In the last few years, one of the big excitements has been the whole kind of deep learning/neural networks business. And maybe we should talk a little bit about how that compares with other things that really are the mainline artificial intelligence that people like Anthony and myself make use of all the time in the systems that we build.
So, what’s happened; the story of neural networks, which are kind of idealized models of how brains might work, that story starts in the 1940’s. And the models that we’re using today are pretty much models that were invented in the 1940’s. For example, as part of sort of basic science that I did, I worked on these things back in the early 1980’s, and I tried to make neural nets that would do interesting things. It completely failed. And, then over the course of time, it’s just a few years ago, it finally got to the point where it was possible to have powerful enough computers, large enough training sets that all the things that we’ve been trying to do for years actually started to work.
And I think that the thing that's worth understanding about a neural-net-type story for computing things is it's all a question of did a programmer visibly write the code, or did the code somehow automatically get produced? And actually, as a result of some of the basic science that I've done, I got very interested in the question of, "If you want to find a program to do something, how do you do it?" Do you have to get a human to write the program, and could you just go and search in a very huge space of possible programs, and just discover a program out there in this computational universe of possible programs.
And actually, one of the things that made Wolfram Alpha possible was a bunch of developments of being able to search the space of possible programs to pull in surprising things that no human will probably have ever have come up with from this computational universe. But there’s sort of a version of that that’s happened recently with neural networks and so on, where it’s possible now to kind of give a large training set.
So, for example, a couple of years ago, we built an image identification system where you can show it, you can find it on the web, imageidentify.com. You show it an object, one of about ten thousand kinds of objects, and it will tell you, "Yes, that's a teacup," or "That's an elephant," or whatever else. How is that done? About 30 million training images, and about maybe a quadrillion GPU operations to actually do the training, but basically one is showing examples to the system. And what it's doing is basically building a model on how to make distinctions between things so that in the end, it will decide "This is an elephant" and "This is a teacup."
But what’s notable about that is when you say, “Well what’s the code that does that?” It isn’t something where you can identify, “Yes, the code works this-or-that way,” it’s something where the code has emerged from looking at all of these examples, it’s not something where you can say “A human built all those little pieces, and this is exactly how it works.” I think that’s one of the things that’s sort of a recent feature of the current wave of artificial intelligence enthusiasm are these cases where, essentially, there’s functionality that got produced that no human was involved step-by-step in making it happen.
Anthony Scriffignano: And of a way of making that feel a little bit real. If I asked you to name five red things, you could rattle off a clown’s nose, and a stop sign, and a kid’s ball, and a post-it note, and an apple, and you’ve never connected the stop sign and the clown’s nose before, and now, all of a sudden you’ve done it. You don’t know how you did it; you know that the answer is right, and you’re comfortable with that answer. You’ve effectively written a synaptic program in your brain. Your brain doesn’t understand what it, itself just did; and that’s a lot of what it’s like to use this type of technology that you’ll get to an answer. And one of the big criticisms is it’s difficult to understand the provenance of that decision. If you have to defend that decision say, in court, or you have to make sure that decision didn’t include any inappropriate bias like race or gender, and you have to make sure that decision is consistent with other decisions you’ve made, it’s difficult to do that with this type of technology.
Some of the newer modalities are trying to address that, effectively having the technology take notes while it’s forming its own neural network paths. But then, you get into issues with performance and really understanding the answers. So, I would very much appreciate the journey that you’re talking about, and I watched at least part of that journey happen over time. I’m still rooting for the day when just Iike I could ask you as a person, “Well, what were you thinking about when you did that? What was your thought process?” I’d like to be able to do a better job with this neuromorphic technology of asking it that same question. I think it’s still a shortcoming.
Stephen Wolfram: So, you know, one of the things I think is interesting about that, it's one of the places where we, as humans, potentially have a dramatic shortcoming. Because, when you train up one of these neural networks to, for example, do physiologic recognition, what's happening is that it's effectively as the neural network runs. It's effectively asking itself a bunch of questions: "Is this a very vertical-looking kind of thing? Is this a very green looking kind of thing? Is this a…" It's making a bunch of distinctions. It's identifying distinctions about things in the world.
Well, we, as humans, will sometimes have words for those distinctions. We'll say, "It's green," or "It's blue," or whatever else. "It's round," or "It's square." Those kinds of things. It's effectively looking at the world and figuring out what are the best distinctions to make to be able to distinguish things in the world? And it's coming up with potentially millions of those distinctions. We, as humans, we have about five thousand words for picturable nouns, for example, in English. We have a very limited number of these kinds of distinctions that we make as part of our human language. What's happening inside these kinds of AI systems is that they are learning distinctions for which there could be words. They just aren't words in our human languages. These are a kind of post-linguistic emergent concepts that exist in systems that we’re building and over time, some of those concepts could turn into words that we, as humans, understand.
Anthony Scriffignano: But, one of the difficulties is that sort of what’s going on in the inner life of the kind of modern AI is something somewhat beyond what I think humans are going to be able to wrap their brains around. We have three categories that we think about, and I think it’s a cousin of what you’re talking about. So we make a distinction between things, and behaviors, and relationships.
And, so as an example, some of the things that we’re looking for are abstract, like “malfeasance”. We use the word “malfeasance” as opposed to the word “fraud” because when someone lies to us, they are doing it in anticipation of some future gain, so at the time, legally, that they lie to us, it’s questionable whether or not it’s fraud, but it’s often a precursor to fraud. Now, if we just used standard regressive methods and modeled the way fraudsters have behaved in the past, and we know that the best fraudsters will change their behavior when we know they’re being watched, then we’ll be modeling how the best fraudsters are no longer behaving, which is exactly the opposite of what we want to do.
So, if we tried to use some sort of learning technology; I’ll just broaden the term from AI to some sort of learning technology; one of the things we want to learn is sort of the opposite of the space that we’re observing: what’s not happening, or how is what is happening changing in quality and character over time so that we can understand when a new behavior has emerged, and then use some sort of discrimination to determine whether it’s more or less likely to be more malfeasant, and then hand it off to a human agent to do that last, really hard part of the work.
So, artificial intelligence in our world is sometimes reducing the complexity of a problem that’s very hard to see, and making the work of those human brains much more valuable by letting them work on the really hard stuff and not waste their time on the more obvious things. So, really tricky dance.
Michael Krigsman: I have a question for either of you. Right now, there is so much hype around AI. Is that hype justified and if the implications are so profound for society, governments, businesses, and so forth, how do we get what the pathway is from here to there; to that point of, shall we say, the “flowering” or AI, machine intelligence, and all of its various forms so it becomes worthy of that hype?
Anthony Scriffignano: I hate to say this, and I'll probably get, you know, attacked for saying this by somebody. Just like anything else, AI is a tool. And, just like any other tool, you should always understand the problem before you pick up the tool. So, I think that AI can become very important when you think about autonomous self-driving vehicles, […] drones, […] anything where we might have something thinking for us because we're not there; I think it's pretty important that we get better at that. But, I think we should also be careful when we anthropomorphize. Learning isn't really learning, it's curation, organization, and attaching of that should be … [and] so forth.
If we really want to be careful about the terminology, I think AI can be very important in our future. The question is whether it’s very important because it becomes our worst nightmare and because we forgot to think about these things, or whether it really improves our human condition. And honestly, the jury’s out right now. I can see bad guys doing bad things just with this amazing technology just as well as I can see good guys solving real problems. I think that we’re really at a cusp.
Stephen Wolfram: So, I mean, the way I see it, what is happening with kind of AI is a continuous line from this really important idea of computation. I mean, back before the 1930’s and so on, people imagined that if you wanted to have machines that did several different things, you would need several different machines. Then, there was this kind of notion that arose in the 1930s of universal computers, the idea that you could have a single machine that could just be programmed to do all these different kinds of things. That idea, which initially was kind of an abstract, kind of mathematical idea – [a] logical kind of idea – that’s the idea that led to the modern computer revolution, and so on.
That idea is the only way through getting locked out. What we are seeing is kind of the computationalization of everything. One of the things I like to say about different fields of human endeavor is to pick a field X from sort of archeology to zoology. They're either is, now, or soon will be a field called computational X, and that will be the future of that field. Some of those fields already exist: computational biology, computational linguistics, others are just emerging.
What's happening is there is this kind of way of thinking about things in computational terms that really is extremely powerful. It's kind of, I think the defining idea of the century, is this idea of thinking about things in computational terms. Now, once you are thinking about things in computational terms, you get to automate a lot of stuff that you couldn't automate before. We happen to be kind of in the middle of a moment when a particular kind of automation that's made possible by neural nets, and so on, is in rapid growth. So, I've probably seen, in my career, I've probably seen a dozen or two fields that have gone into this kind of hypergrowth period. Typically, what happens in fields of human endeavor, whether it's areas of physics, or whether it's biology or lots of other kinds of areas is there will be long periods – decades, maybe even a century – of fairly slow growth. And then, some methodological advance will occur typically, and then there’s this period of hypergrowth where there’s a new result every two weeks.
Now, I happen to be lucky enough when I was a kid, basically, to get involved with particle physics at a moment in the late 1970s when it was in the hypergrowth phase, and where important new results … You know, every week or two – New Paper with Important New Thing. That lasted for about five years. Since the late 1970s, particle physics has been in a pretty flat state. We are, right now, kind of in the middle of the hypergrowth phase for machine learning and neural network-type techniques. There’s a lot of low-hanging fruit to be picked. You know, we’ve been having a great time because we’ve been using this Wolfram language system that we’ve been building for thirty years now, and kind of integrating what’s new with the possibilities of neural networks with the kind of large-scale language that we’ve developed, and we’ve now got this nice, symbolic way to develop a high-level development mechanism for neural networks we’ve just released a few weeks ago.
But, it’s always exciting to see in these periods of hypergrowth for a field. And what we’re seeing right now for neural networks is a lot of low-hanging fruit being picked. There’s a lot of things where people have said for years, “Oh, computers will never be able to do that!” For example, physiologic recognition was one of those things. And then suddenly, it’s like, “Yes, computers can now do it!”
What we’ve seen over the last couple of years is with physiologic recognition: basically, the Big Thing happened a couple of years ago. You know, we got to the point where we can basically recognize that more-or-less human level of performance, a whole bunch of different kinds of objects. What happens from here on out, with that kind of particular problem, is kind of “slow-growth, hard work.” But, there are a whole bunch of these things. We’re probably about halfway through, I estimate, under the hypergrowth period for this.
Anthony Scriffignano: Yeah, I was at a conference last week where someone was talking about the autonomous self-driving vehicles, and the importance, obviously, of object recognition and having an algorithm drive a car. And, I won't mention the neighborhood, but the neighborhood near the organization that is developing the cars, apparently, the local youths have found an amusing pastime of making homemade stop signs and holding them up in front of the autonomous self-driving vehicles and make them stop in the middle of traffic. A human being wouldn't be confused by a kid holding up a fake stop sign, but an algorithm designed to recognize a stop sign, until it realizes it's being tricked that way because someone told it, will stop every time.
So, I think that I totally agree with what you’re saying, Stephen, that there’s this sort of constant evolution and we are certainly in an explosion of evolution. I think we’re also, from the way I stand, sometimes in this rush to get to the market, and these sort of unintended consequences don’t get thought about; sometimes, they’re funny like the stop sign; sometimes, they’re not so funny, and people find a way to take down half the internet with a denial of service attack on security cameras because people don’t update their software.
I think that people could do a better job of thinking about unintended consequence, and I worry a little bit about rush to market. This has nothing to do with you; I don’t think you’ve ever rushed to market; but I think that people who make some of these things are in a hurry to make the next really clever thermostat or the next really clever car, and there are really clever people out there just waiting for them to do that.
Stephen Wolfram: So, you know, one thing I might comment that I think both of us are somewhat involved in is the whole question about how do you put knowledge into compute-able form? How do you go from what's out there in the world, to something where knowledge is organized; where you know systematically that companies and their characteristics, things like this, are about chemicals and their characteristics, all these kinds of things. And one of the questions is always, "Can you just go to the web, forage the web, have a machine intelligence figure out what the right answer is," and that's been a long story. And, the basic answer is you can get 85% of the way there. You know, it's pretty easy to get; to use automated methods, to forage Wikipedia and find out 85% of the correct facts.
The problem is, you don’t know which 15% are completely wrong. And, to know that is really something where one needs a process of curation that is a mixture of really good automation, with human activity. And I think both of us in our rather different ways have been deeply involved in this curation process. It’s something which is not widely understood in the industry at large. It’s something where people say, “We’ve got this AI! Now we can just automatically ingest knowledge and get it organized.” My observation has been, as I say, you’ve got a certain fraction of the way there, but in the end, us humans aren’t useless after all, and you actually need to inject that sort of moment of human expertise into the thing.
Michael Krigsman: Can I…
Stephen Wolfram: It's such a difficult thing for companies to understand because it's kind of like, companies tend to be either "We're a tech company, or a people company," so-to-speak. And the tech companies just sort of say, "We'll just attach the magic AI and it's just going to work," and the people companies are like, "Oh, we don't know about this technology stuff, we'll just have people all the way." Curation is an interesting thing that is a complicated management problem, where you kind of have to use automation, inject human judgment when it's appropriate, figure out how to move the process of judgment through the organization in the right way to actually be sure that you're getting the right answer.
Anthony Scriffignano: I would add about that 15% that often, it's that 15% that ends in that really apocryphal tale of how someone just failed miserably. There's some really important significance. The reason that 15% became sublimated and wasn't so easily discoverable is exactly the reason why there was something really valuable in it. You think about, to use an example of knowing all there is to know about companies; I happen to be in that be in that business, right? It's not that hard. We all fail to go on the internet to do research and find out about a company. How do you know what you're looking at is real? How do you know it's current? All true information is no simultaneously true. All information that you can get to in real-time wasn't created a minute ago.
You know, being able to curate, to understand, to put like with like, to triangulate, to test for veracity, to have some experience; things get new and things change; when the environment changes, to understand how it’s changing. These are the critical moments where I think there’s still hope for the need for our human brains, and I think you’re not going to program ourselves out of business here. I think that we’re going to get to solve bigger and better problems.
If I look at progressive decomposition; taking those really big problems that are not solved yet and breaking them down into smaller and smaller problems that are still not solved, I think there's great hope for being able to focus on the more important parts of those problems with our human brains. And, these technologies will help get everything else out of the way, if we let it.
Michael Krigsman: So what are the implications of all of this for society? The way that you are talking about it is in, let’s say, mechanistic, computer science terms as opposed to the way the software industry as a whole, in its marketing, talks about AI; which is in magical results terms…
Anthony Scriffignano: Yeah. I think it’s fair to say, and I won’t speak for both of us, but I’ll speak first for both of us and then I’ll see if Stephen agrees with me. I think we would both say that it’s very important to have technologies. It’s very important to advance those technologies. But, there’s never going to be a substitute for understanding the problem, for humans to continue to advance the art. The machines can help convince the art. But for the foreseeable future, I think we still get to conduct the orchestra.
Stephen Wolfram: I think that the main question is, what can be automated in the world? And, the fundamental thing to realize is that what can’t be automated is what you’re trying to do. That is the definition of the goal. There is no abstract sort of ultimate, automatic goal. The goal is something that’s defined by us humans on the basis of our history, and culture, and characteristics, and so on. The real picture of how we interact with technology and nowadays with AI, is we, as humans, define the goals. We say what should happen and what we want to achieve, and then it’s a question of, “Can we make automated systems that do the best possible job of achieving those things in the best possible way?”
So, one of the big issues, then, is how do you tell the machines what you want them to achieve? So, in some cases, it’s very straightforward. But when it gets to be sort of a bigger picture of … Well, one of the things I’ve been interested in, in recent times, is kind of how do we communicate with AIs? You know, one thing we can do is just say something with natural language. That's good when it comes to short things like asking a knowledge-seeking question or telling some device to do one particular thing. That works pretty well with simple, natural language.
When it gets more complicated, natural language doesn’t work very well. We have kind of a model of that right now when we look at things like legal contracts. Legal contracts are trying to define what should happen. They start by being written in natural language, but it turns out we need to invent legalese because we need to take natural language and make it a bit more precise. Well, I think there's an end point of that direction, which is to have a kind of code, a kind of computer language, which could say the kinds of things we would want to say in a legal contract; but can do so in a precise fashion.
And with our Wolfram Language System, this is very much the direction that we’ve been going in. I mean, what we’ve mostly been dealing with is things like; we can talk about cities; we can talk about distances between places on the Earth; we can talk about all sorts of things about genes and biology, those kinds of things. We can – I like to use the example – “I want a piece of chocolate.” We can talk about a piece of chocolate; we know a great deal of deal of detail about different brands of chocolate and their nutrition content, and so on. The “I want” part we can’t yet talk about, but we’re working towards being able to have a precise language for talking about those kinds of things, and I see that as being an important intermediary between the way we think about things and the way that we can have machines do things.
Anthony Scriffignano: There’s an interesting nuance in the corners of this problem that we’re each working. So, in my world, the problem that we have is the parts of language that we’re interested in mostly, are talking about proper nouns. And proper nouns tend not to be in the dictionary, and they tend also to have meaning that is very context-specific. So, if I say, “Apple announced the new iPhone, whatever, 19,” I know that a fruit doesn’t announce anything because a fruit is an inanimate object. So, I’m immediately down the path that this is more likely to be a proper noun, not only because it’s capitalized, but because it’s the operator of a transitive verb like that.
So, what our problem is, is not so much to understand what happened, but to understand whether this thing that happened over here, and this thing that happened over here, are likely to be talking about the same entity, and whether or not that entity is likely to be a business. So, we don’t have to get down to the meaning of “What did they really do?”, we just have to know that this is something that Apple did, and that apple that did it is more likely to be the computer company than the fruit. That gets notoriously difficult when we start crossing language barriers, because the way you talk about Apple, in Chinese, for example, happens to be the word for “apple.” The way you talk about Dun & Bradstreet, for example, there’s no word for “Dun” or “Bradstreet,” so you have to use some other words that either sound like it and mean something not offensive, etc. So that understanding, it’s called “semantic disambiguation” of what are you likely to be talking about here; not necessarily “What does it mean?”, but what is it likely to be, is the tricky part for us.
In your world, you actually didn't know what the Apple is, and you need to know if it's a Macintosh apple or a delicious apple – much more difficult problem on the meaning part, I would guess; much less of a problem in terms of a business side of, "Is it an acquisition or is it some sort of an LLC," you know, what kind of businessy thing happened here?
[…] And these are good examples of different types of problems you have in AI. In one case, I want to know what it means, and in another case, I want to know that these two pieces of information are talking about the same entity and I don’t necessarily care what it means. One of the things that we look at is, we actually look at confounding characteristics in language. So, we look at sarcasm and neologism. You talked about the legal documents and the fact that if I talk about a “lean piece of meat,” or “leaning into something,” or a “lien” in a legal document; I know those are spelled differently, but those are three different things. The difference is there and it’s in context.
When we start making up words, when we start using Twitter handles, when we start talking about tweeting things, when we start using sarcasm; when I say, “This is a great company if you don’t mind destroying the environment,” we have to figure out that there’s an independent and a dependent clause there, and they have opposite sentiment, and destroying the environment is bad because the environment is good and destroying it is bad. Tricky problem. A different tricky problem. I think we’re going to separate schools together on this, and that’s exactly the nature of AI.
Stephen Wolfram: I think one of the challenges is that the capabilities that computation provides and that AI, which is sort of a thing that sits on top of computation provides; there are all sorts of impressive things that can be done. The issue is how do we direct them for human purposes? You know, one of the things I’ve been interested in for a long time and we’re done for a long time is this business of algorithm discovery in the computational universe. If you look out there in the space of all possible programs, there are programs that do all sorts of remarkable things. The issue is can we really mine that space of possible programs for ones that are useful to us? It’s very much analogous to what happens in physical technology. It’s like, “Okay, there’s an iron mine somewhere. There’s a tantalum mine somewhere.” You know, we find the material like, let’s say, tantalum, and we say, or gadolinium: is this useful for anything? And the answer is, “Yes. It’s discovered it’s useful to make magnets out of.” There’s this similar kind of issue in the computation universe of possible programs. There are a lot of them that do things which look interesting, but can they be “mined” for human purposes?
So this, again, puts the pressure on, “Okay, so what do we actually want to do?” Computation and now AI provide amazing capabilities. The issue is what do we want to do with them, and how do we make something where both we, as humans, can sort of understand what we’re asking for, and where the machine can understand what we’re asking for?
So, for example, in my life, I’ve spent a huge amount of time developing very high-level computer languages that let one express things in a way that are the highest possible level of expression for humans or what they want to do that can also be understood by machines. I think this is … As we look towards the smart contract world of telling our machines in general terms, “What do you want to achieve?”, we need a language to do that. I think the ultimate smart contract that us humans have to think about is the whole constitution of what do we want to AIs to do? We would like to say, “Okay, AIs, we’re going to make AIs in charge of all kinds of things in the world. We’re going to have all kinds of systems be done automatically.” We want to give some overall guidelines. You call them “ethical guidelines” for how the AIs should behave with respect to us. “Okay, AIs, be nice to us.” How do we express that? How do we define what that means? How do we specify the constitution of the AIs?
I’ve been interested in this problem of what is the language that we can use to write the constitution for the AIs. And the next question is, “What should the constitution actually say?”
Michael Krigsman: Yeah.
Stephen Wolfram: You’re mentioning, Anthony, the whole question of looking backward; the regressive approach to things. You know, one of the things you might say is, “Okay, AIs. Just do things like us humans do.”
Michael Krigsman: So, Anthony Scriffignano, this question that Stephen just raised on the ethical dimensions: is it simply a matter of, “Do what we would want. Be nice to us.” Your company, Dun & Bradstreet deals with financial matters, so dive into the ethical implications. And, I think this relates back to an earlier part of the discussion that [you] brought up regarding the unintended consequences.
Anthony Scriffignano: Yes, so I want to first just respond to the somewhat rhetorical ending of what Stephen said because I loved it. If we actually asked our algorithms to behave as we have, you might be surprised at how badly they behave, because we correct our memory and make our prior accomplishments greater than they were. We tend to sublimate our prior failures. We tend to amplify in our minds how successful certain things were, and I like to say that we should make new mistakes, right? So if I just write algorithms that behave the way I have behaved in the past, then that means warts and all. That means foibles. And so, hopefully, I have learned in my life and I don’t even behave the way I have behaved in the past. So, what I probably best would want them to do is behave the way I would like to behave in the future, and that’s a whole …
Michael Krigsman: So you want AIs to be an idealized version of humanity?
Anthony Scriffignano: I don't know. Maybe. Maybe I do. But, there is this problem. There are a couple problems with this. One is that we don't all want the same thing. We don't define success the same way. We don't define winning the same way. We don't define being nice the same way. Sometimes being nice to someone is tough love. Sometimes, you have to do something that isn't nice to get a quote-on-quote "better outcome." And then there's this whole question of sub-optimality that sometimes, what’s better for me isn’t necessarily better for everyone else or better for you. And so, we get into the whole “best alternative to a negotiated agreement” kind of argument of, “What’s better for the common good, and is what’s better for the common good what we should do […]?”
So, these are all really philosophical questions. But, it’s so easy to go down that road. You can imagine how hard it would be to automate this, and to do this in AI. When Stephen was talking, I was also thinking about language. So, I do a lot of work in computational linguistics across languages, and within the sphere of two or more languages, something that you think is pretty obvious to say can’t easily be transformed or translated into a language maybe doesn’t have a word for a “yes” or a “no,” or a language that doesn’t have a subjunctive. In a legal context, there’s a very vast distinction between “You should” and “You must.” It’s critical in English. If you say, “You should do to the following things,” or “You must do the following things,” those are … There are multiple words for “should” and “must” in Chinese and some of them are used synonymously. So, you have to understand the context to understand if it’s a “should” or a “must” and sometimes native speakers disagree as to the interpretation of that. Think about what that does to a contract.
Now, try to imagine talking to a machine that set its goals at a level of ambiguity that we were talking about with this ethics, and we can’t even get “should” and “must” right. That’s a very big problem. And, learning is important.
Stephen Wolfram: You know, I think you're making the case for what I just spent thirty years trying to do, which is we need an actual, precise language for expressing what we want. When we look at sort of the future of legal contracts, and so on, today, legal contracts are written in English, in legalese, whatever else. And, you know, people say, "Well, it's important. There's a little bit of wiggle-room in the language, and so on." But, in many cases, it would actually be a lot better if everybody knew this is exactly what the contract means. So, so long as we can express those concepts in a precise language, it's much better if both parties can just agree, "This is what we meant." And some part of the contract will be written in something which is almost like logic, some part of it might be written using some machine learning classifier. It might say, "We both agree that it's a Grade A orange if this machine learning classifier that we both agreed on looks at it and says it's a Grade A orange." And that might be a good way to write the contract.
And of course, a big advantage of having a contract that's written in compute-able form is that, then, you can just say, "Okay, computer. Figure out if the contract was satisfied or set things up so that it is satisfied." And I think it becomes … We're able to kind of automate a lot of the process of doing these things. I think that's a really interesting … "If both parties agree this is what we really meant." It's written in code, not in legalese, then we have quite an interesting thing going on and a place where we can get a lot of transactions to happen in a much more efficient way because machines can interact with each other using automatic contracts rather than humans having to interpret this-or-that thing going in.
Anthony Scriffignano: I completely … First of all, and I mean this with a complete absence of sarcasm, I look forward to your success. I think that this is something for that 85% of contracts that are pretty straightforward, “Here’s what you’re going to do, here's what I'm going to do, here's how we'll measure the results, what happens if there's a breach of the contract," all the elements of a contract, most of the time, are pretty straightforward; certainly in commercial terms; if you look at ECC, things like that. There are commercial codes that codify these things. Great.
When we have to make a contract to describe what we're going to do for something in the future that has never been done before; and maybe doing it with things that we don't have exact terminology. Now, it gets tricky. I think a good argument would be, "Great! Let's let the jurists and the legal minds ... Let's let everybody focus on that part of the hard stuff and not worry about the contract of whether you shipped me the apples I ordered […]”
Michael Krigsman: I hate to cut you off. We’re just out of time, and I would ask each of you to spend a moment and share your advice to businesspeople, policymakers in the government, and technologists regarding AI. And again, I apologize, Stephen, for cutting you off. We’re just out of time.
Stephen Wolfram: Well, I guess … Let me say something first, perhaps. The most important thing, I think is this kind of “think computationally.” Use this kind of computational paradigm that we are slowly beginning to understand and think about those things in those terms. Define things in those terms. Once you’ve defined things in those terms, then it’s sort of a much easier problem to say, “Okay, AIs. Now we’ve defined what we want in computational terms. Now help us achieve this, so to speak.” That’s my main point. Think in computational terms. Get your thoughts organized in a way that you could imagine explaining them to a computer, so to speak. That, I think is the important direction.
Anthony Scriffignano: So, I always give my advice in threes, and one of my three is being almost verbatim what you just said. I thought it first…just kidding.
The first thing I would say is, be humble. I think that there is always a difference between the theoretical, science fiction; what's possible and what's real. Make an honest assessment of how much of what you're trying to do – especially if it's a commercial application or law about the future. Being realistic about what's possible so that you constrain this to a space that you can understand.
The second one is what I think you were saying, Stephen, which is to be as clear as you possibly can by about your goals. And if that means you have to invent a whole new language to do it, then have at it! But, we’ve got to get this ambiguity and squishiness out of this to really get better at it, and I think we’re making great strides in that regard.
And then, the third thing I would say is to learn; to make new mistakes; to make sure that we’re constantly observing our behavior with this amazing technology and make sure we’re solving incrementally more important problems and more difficult problems; that we’re not rushing to market and making the same mistakes over again with different names.
Michael Krigsman: Okay. What an amazing discussion we’ve had. We have been talking with Stephen Wolfram, who is truly one of the fathers of modern computer science and the founder of Wolfram Research and other companies. And, Anthony Scriffignano, who is an amazing thinker, is the Chief Data Scientist at Dun & Bradstreet. Gentlemen, thank you so much! Stephen, thank you for being here, and I hope you'll come back and do it another time. And, Anthony, thank you for being here and you're scheduled to come back. So, I know you'll be back to do it another time! And, my hope is that we can get the two of you back together again to continue this conversation because as you both pointed out right at the start, 45 minutes is not enough time.
And, everybody, thank you for watching! Come back next week. We will have another great show. And, subscribe to our YouTube channel. Click the YouTube button on your screen and subscribe. Thanks, everybody. Take care. Bye-bye!
Published Date: Apr 14, 2017
Author: Michael Krigsman
Episode ID: 427