AI CXO

Trusting Artificial Intelligence for Big Business Decisions

Trusting Artificial Intelligence for Big Business Decisions

Artificial Intelligence and related analytics capabilities offer the promise of better, more accurate decision making. As humans, we are still not ready to sit in a driverless, speeding car. Have we come to trust the machines to a point where we are willing to let Artificial Intelligence drive big decisions in business?

Contributor

    • Pedro Domingos, Professor of Computer Science, University of Washington

Download Podcast

Apple PodcastGoogle PodcastSpotifyPandoraiHeartRadioSoundCloudTuneIn, and Stitcher. Find other syndication channels here or search CIO Talk Network podcast on any other app.

Top 3 Learning Points

  1. How to leverage the power of technology to add to your bottom-line
  2. Is AI really replacing humans or just primitive computer systems?
  3. Will the human brain that creates the AI programs ever be able to catch up with its learning abilities?

Show Notes

  • At the end of the day, the computers are just extensions of us. They’re just there to do what we want them to do. It’s for us to figure out what it is that we want to do.
  • Once businesses grow beyond a certain point, they have to use computers to help them make decisions for everything, from customers to production
  • People often think of AI as replacing humans, but it is replacing computer systems that were much worse before.
  • Do you think AI is going to be only as good as the data that you provide, or there is some other unique way by which it can profile me beyond this, even if I am not ready to share every part of my life?
  • Maybe in the very long term computers will become better than humans at everything but that’s going to pick up. In the foreseeable future, what we’re going to need is a very close combination of people and computers doing most things.
  • The CEO isn’t going to be replaced by an algorithm anytime soon because while algorithms are very good at some things, there are also many things which humans are much better and we need both.

Summary

While artificial intelligence is able to perform most business functions, deductions and tasks perfectly, and sometimes even more perfectly than humans, the truth is for 99.9% of companies out there, the CEO isn’t going to be replaced by an algorithm anytime soon because, while the algorithms are very good at some things but there are many things that which humans are much better than the algorithm, and we need both. So while there’s more promising machine learning happening in one year today, than it used to happen in a decade, we still have a million more miles to go. It might not take a huge amount of time to get there, but nobody knows if it will be decades or more. What will finally work is a grand unified theory of machine learning to actually have all the intelligence that human beings have, in one system. Once we have that, it is possible that the majority of jobs will end up being done by computers.

Transcript

Sanjog: Our topic for today is Trusting Artificial Intelligence for Big Business Decisions. Our guest today is Pedro Domingos, Professor of Computer Science, University of Washington.

We shall be talking about something which is our reality today, but it is futuristic for many. Enterprise businesses are at all times making decisions, sometimes very strategic. We always thought humans would bring complex processing into the decision-making, but today artificial intelligence is also claiming it can either do equal or better than humans. Do we need to explore if we can trust artificial intelligence for big business decision-making? It will not necessarily replace humans but at least come close. But why is someone even looking at the need for using AI to drive such business decisions?

Pedro: There are a number of reasons, but the biggest driver is growth. Once businesses grow beyond a certain point, they have to use computers to help them make decisions for everything, from customers to production. The benefit of AI is that decisions can actually be made in a way that is much closer to or better than what humans would do. People often think of AI as replacing humans, but it is replacing computer systems that were much worse before.

And in many other cases, like in companies like Google and Amazon, it is making decisions that simply weren’t made by anyone before, because it didn’t even exist. There’s this whole new spectrum within which AI is being used today.

People often think of AI as replacing humans, but it is replacing computer systems that were much worse before.

Sanjog: You said the computers were not able to churn data quickly or as accurately but we cannot compare AI to a regular database management system. It’s a different animal, isn’t it?

Pedro: They are different animals, but if you have to market a new product, you can rely on your gut feeling on whom you want to market it. Traditionally, companies would rely on these broad demographic categories, which were very coarse. They did this partly because there wasn’t any better data than that earlier. But now we have so much data in such detail, as well as the computing power, that we can leverage AI to use that data to make better decisions. 20 years ago, the state of the algorithm, of machine learning and making the predictions and recommendations was much more primitive. All of these things have progressed enormously, and we’re now at the point where they’re actually ready to make the vast majority of the decisions that it didn’t make.

20 years ago, the state of the algorithm that to the machine learning and that make the predictions and recommendations was just much more primitive.

Sanjog: You’ve mentioned that the data wasn’t available, so that means somebody collected the data versus machine doing it by itself. Because some form of human intervention would be required to collect that data. Secondly, when you speak about churning that data and getting some insights out of it, that’s an analytics function. Was this analytics produced in the recent past, but was not allowing us to dig deeper? If that’s the gap, how are you defining artificial intelligence? Is it some machine logic which is going, to sum up, come up with an insight, where even a computer-driven analytics would not be able to find?

Pedro: Yes. Two parts to this. The first one is that data used to be gathered by telephone survey-calling a few thousand people on the phone and even that was expensive. In the 90s, there were databases of projects, and that’s when people actually started to do real data mining. But it wasn’t that much, and it was expensive together. These days your cell phone is continuously collecting data value. It’s everything that you do on social networks is being captured by somebody or other.

For example, when you’re on the Amazon website, it’s not just the part that you buy that generates data value; it’s everything that you click on. Amazon can actually see what the sequence of things that you did was. The data is available in magnitudes greater than before. It’s also much cheaper to collect. In many cases, these days the data doesn’t even have to be collected. It’s naturally generated by people as they go about living their lives which is done more and more online. And a lot of data gathering happens from the things that people in the physical world do. This is one aspect. But the other aspect that is equally important is as you mentioned, the machine learning.

The whole name of the game in machine learning is to generalize from your past behavior to your future behavior. And this is actually a very hard thing to do.

The data that people generate doesn’t actually necessarily say anything about the future. This is just something that they did in the past. The whole question is, how can I use that to predict what you want in the future? Machine learning is used to generalize from your past behavior to predict your future behavior. Over the last few decades, but much faster over the last decade, machine learning has become better things and now to a point where it can actually do things amazingly well in many cases.

Sanjog: Essentially what you’re saying is the fuzzy logic which artificial intelligence would use to compute and predict behavior based on what they have done in the past is just an expanded version of multiple humans sitting together for years trying to figure the same thing out. Is the name of the game the speed, or the quality of decision?

Pedro: It’s both. It’s the speed and quantity, and quality of the decisions. You mentioned fuzzy logic. It is something that was used a lot back in the 80s, and it was called knowledge-based system. It was an expert’s system which responded to the program that you would write down to, for example, make a medical diagnosis. You have rules that said if this patient had these symptoms, then what you have is the flu.

True logic came to this very well, which is why in anything like the fuzzy logic you have information there’s more or less confident. But the problem with that and why AI failed, is that these systems were too brittle and they didn’t know enough. There was always more knowledge that they needed to have and that affected their ability to deal with uncertainty. It was very limited. As soon as you went outside the narrow scope of what you had taught, they failed.

Today, the big reason why AI is not really taking off is that we’ve replaced things like fuzzy logic and knowledge base systems with things like machine learning, where you use things like neural networks. And the neural network is essentially a simulation of a human brain that learns from experience from data to some degree in the same way that a human brain does. Another type of machine learning, for example, is based on simulating evolution, except that what you are doing is evolving programs, instead of evolving animals or plants. But there’s a same general idea of genetics and natural selection as if everything is on the computer. So there are a number of these caveats in machine learning that let us do things that you could do back in the 80s with the so-called classic AI.

Sanjog: To that end, my question to you would be the change in complexity that you programmed it, as you mentioned neural networks and many other types of logic. At the end of the day, the logic of how it would function, how would it learn itself is still coming back to some human beings who were working at it. Now the way the complexities are coming, or we’re facing the complexities in the business world as well as in our lives is about a whole set of parameters that were not even be dreamed of in the past. We’re continually evolving. So how does a human brain, which in turn is supposed to program that AI logic, ever is able to catch up, so you can make a decision for today?

Pedro: The big difference is that in computer science, it’s humans that write the programs. In machine learning the computers write their own programs, the computers program themselves. What makes them so powerful is that the computer is looking at that from the data, from the input and the desired output, and it’s figuring out what the program should be. The more data you have with machine learning, they’re correspondingly better, the programs get more output with no additional human work. It’s a little bit like a human being as a child that learns by playing around by herself. And to a large extent, this is what machine learning is able to do.

The big difference is that in computer science, it’s humans that write the programs. In machine learning the computers write their own programs, the computers program themselves.

At the end of the day, the machine learning is going to learn themselves were written by a human being. But a machine learning algorithm could be a hundred lines of code that generates millions upon millions of lines of code that are modeling specific people and specific quests. It would have been completely hopeless to try to write that code by hand; it wouldn’t have been economically impossible. For example, suppose that programmers at Amazon had to write a different program to model each one of their customers, well Amazon will go out of business. They wouldn’t even find enough programmers to do it. With machine learning, they do have a model of you, which will learn from the data.

Of course, the action comes becomes where and how can you learn to model, what people do. Then you unleash those algorithms on a lot of data, and then you derive a very detailed and very rich model. That way you make decisions with a level of quality that was not achievable before.

If the programmers at Amazon had to write a different program to model each one of their customers, well Amazon will go out of business. They wouldn’t even find enough programmers to do it. With machine learning, they do have a model of you, which will learn from the data.

Sanjog: Pedro, on this front you mentioned about the machine learning which is a fascinating area. Now it all still is dependent on data. That means you’re expecting the human to divulge their whims and fancies and uncertainties and quirks if you will, on a larger scale or at that very advanced level for your machine to understand. Suppose they are trying to profile me, I have to tell things way beyond what I want in a product. I have to tell them about my family, my background and everything which I am not willing to give out.

Do you think AI is going to be only as good as the data that you provide, or there is some other unique way by which it can profile me beyond this, even if I am not ready to share every part of my life? Pedro, as I mentioned it seems like still, whatever advances that AI technologies and programs and logic may have made, it’s dependent on data. I’m just putting myself in front of Amazon or any other way the data can be collected. But they only know my behavior in terms of how I shop versus everything else about my background and how I do other things in life and how I make decisions in other junctures of life. With that said, you are still dependent on data and if you don’t have a full profile of an individual, how could you really predict accurately enough what the person is going to be wanting next?

…there are the enormous amount that can be inferred from you indirectly. That’s one of the things that machine learning is very good for – that it can infer your shopping preferences, for example, from things that are not directly a shopping behavior.

Pedro: Yes. This is a crucial point. If the machine learning systems don’t have data, they will not learn that much about you. Today that data is being gathered about people in quantity and at a pace that is astonishing. There’s a potential conflict of interest here; people don’t necessarily want to give up some of the data that may be without their complete knowledge. Having said that, there’s a couple of important things to note here. One of them is that there is the enormous amount that can be inferred from you indirectly. That’s one of the things that machine learning is very good for – that it can infer your shopping preferences, for example, from things that are not directly a shopping behavior. Thus what a lot of machine learning does is – from the data that is available, inferring the things that are actually relevant to the company. You’d be often surprised how correlated things are. For example, Facebook can tell all sorts of things about your tastes from things that are maybe talking about something else. This is one aspect.

The other aspect is that in practice most of the time people are quite happy to give up their data in return for what they get because they get better service and more personalized offerings from the companies. In essence companies like Google and Facebook give you a lot of stuff completely for free, but what you’re actually implicitly paying them with is the data that you generate. That allows them and the advertisers to target things better for you. At the end of the day, that is what makes things like online advertising better than things like TV advertising, at least for some things.

It is very important, and people should be more conscious of how the data is being used. I think they should have more control over that too. But to a large extent the interests of all the different parties that actually aligned in this.

From that data that you generate and they mine, everybody benefits, that’s what sustains their business at Google and Facebook. The advertisers are able to better target their products. You get advertisements that are more relevant to you and more interesting, more useful. You are right that it is very important, and people should be more conscious of how the data is being used. I think they should have more control over that too. But to a large extent, it is the interests of all the different parties that actually aligned in this.

Sanjog: All the examples and the things that we discussed so far are where you get a consumer to act in a certain way, collect data and get them to buy. About the business decision-making within an organization, you could have some geographic areas you want to expand into, some new markets or products that you want to introduce, some new capabilities you would like to create or build buy around. Then you could have many other strategic decisions that a business would want to make. Do you think AI has the chops to be able to deliver on this front?

Pedro: Yes, in fact, the B to C applications are better known to the public because they set our daily life. But the B to B applications are just as important, if not more important. In fact, here’s an example. Amazon, the best-known use of machine learning at Amazon is the recommended system. But they apparently they derive a third of their business from the recommended system, and we all see the recommendations that it makes for us. But an application that Amazon uses that is much less known but is also extremely important, is demand forecasting. That Amazon wants to predict who is going to want to buy how much of what were, in essence, every minute of the day, and this is a big use of machine learning. They’re predicting the demand for the millions of products that they sell every day and continuously trying to come up with better machine learning algorithms to do this.

They’ve got so good at this that one thing that Amazon is now doing trials with is what they call predictive delivery. Predictive delivery is when Amazon sends you something to buy before you order it so that you will have it minutes after you’ve ordered it. This is one of those things that sounds like science fiction, but it’s actually not. They don’t reveal the details of this. My guess is that they know that for example, in a certain neighborhood someone or other is going to want to buy Harry Potter the first book. They may not know that it’s exactly you, but they know that somebody will, so they put it on the truck, and then by the time that the truck gets there like the order will have been placed, and then they can immediately deliver it to the person.

They actually say that if you ordered something, if you received something that you didn’t order, then you can actually have it for free. Let’s see how confident they are that they can predict that well. The same thing is true for companies, and the most remarkable thing about the companies that use a lot of machine learning is that they don’t just use it, for one thing, they use it in every nook and cranny of what they do. For example, another important use of machine learning that people don’t see is that companies like Google and Microsoft and Amazon, they use machine learning to optimize their data centers. They have these huge server fronts that cost a lot of money and spend a lot of energy and they continuously optimize and then try and minimize, for example, the energy consumption, which is a huge cost.

Machine learning is penetrating these industries at different rates, but a lot of important examples of the use of machine learning are in industries that do not have anything to do with tech, they’re in finance, they’re in telecom, they’re in retail, and they’re in medicine and health.

Again, they can do this very successfully to the point where they can even reduce the energy consumption by knowing what to turn on and what to turn off at what times by close to a half. There are many more examples of things like this. By the way, I’ve been using mostly examples of these tech companies like Google and Facebook and Amazon. Machine learning is penetrating these industries at different rates, but a lot of important examples of the use of machine learning are in industries that do not have anything to do with tech, like in finance, telecom, retail, in medicine and health and so on.

Sanjog: With all examples that you gave are transactional or related to transactions and operational optimization. If we were to roll this up or maybe go at a level higher upstream, there are some decisions to be made about maybe, which geography should this business expand into? Or which markets to tap? You’ve got a CEO with a bunch of consulting firms supporting him or her, and then also their staff or like the senior executives who are all working, strategizing, sitting in a room and trying to figure this out. Would you think AI has a place in improving or replacing the decision-making which is happening in a boardroom and deliver a better result?

Pedro: Yes, definitely. AI is used a lot of these things today. For the most part, it’s in support to the humans who make the final decision. There are cases however where the AI itself makes the final decision. Or something that is very common is for the AI to make the triage among the thousands of things that the company could do to just maybe 10 or 20 that’s going to consider carefully. One very good example of this, actually one of the original usages of machine learning is in finance is. Here you have 3000 neural networks, each one modeling a stock in the Russell 3000 and the machine learning takes from the 3000 stocks, the 20 that are good for human analysts to look at.

The machine learning actually puts the buy and sell orders in, and of course, things like high-frequency trading are almost entirely done by machine learning because they have to be because no human can function that fast.

And then out of those, the humans take the ones that they actually are going to invest or maybe the ones that they held and are going to sell. But there are also some hedge funds that are completely run by machine learning. The machine learning actually puts the buy and sell orders in. Things like high-frequency trading are almost entirely done by machine learning because no human can function that fast. But they also do things like for a venture capital firm that has actually named learning algorithm to its board of directors. It has seven directors that vote on investment decisions and what firms to invest in. One of those is actually an algorithm. It has as much of a vote is as it should be a human investor. I can easily see if this works out well, the algorithm is having more and more votes on the board until maybe they have a majority or they have all the votes.

To give another example that is very consequential, a lot of hiring decisions and promotion decisions these days are made by machine learning algorithms. Companies get thousands and thousands of resumes and applications, and it’s the machine learning that actually sifts through them to find the more promising candidate. Again, typically, the machine learning doesn’t make the final decision. But from 10,000 candidates, maybe the 50 that the firm is going to look at. Most of the selection is actually made by machine learning.

Interestingly, this has been studied. People often prefer this because they find that the algorithm is fair. But the algorithm tends to focus on just the things that really are relevant as opposed to things that aren’t relevant or the preconceptions that the human beings have.

A famous example of this, of course, is Moneyball. You’re selecting baseball players for a team, but the human scouts used to use all sorts of things that they thought were meaningful but weren’t. Using this kind of statistical learning, what the Oakland A’s found was that you could actually do very well at selecting the right individuals and the right combination of the individual. Like the player that they were missing to complete their team, they could select using machine learning and then pay a much lower price for them. They were the pioneers of this activity, and the same kind of thing is happening in the job market at large.

If you’re a company with a subscription to LinkedIn, one of the things that LinkedIn will do for you is mind the resumes of the people who are on LinkedIn for the people who are likely good candidates for the job that you have. This is something that happens on a routine basis today.

So these are just a couple of examples of how machine learning is being used for these higher-level decisions in companies today.

Sanjog: Why do we then even need to have humans in a leadership position? If AI promises the outcome which could incrementally or significantly better than an individual, so that means you could have a CEO which is an algorithm. As you already mentioned, a board of directors could be a bunch of algorithms, and soon you are having any and every decision being made within an organization by a set of software algorithm which has got support from machine learning, and you just have worker bees. Is that what we are looking at as a future for an enterprise? Then what are humans going to do in an enterprise?

I imagine a CEO that’s an algorithm, and for all I know there’s probably some clever startup somewhere, they really have something like this. Maybe in the very long term computers will become better than humans at everything but that’s going to pick up. In the foreseeable future, what we’re going to need is a very close combination of people and computers doing most things.

Pedro: No, actually I wouldn’t say that, but I imagine a CEO that’s an algorithm, and for all I know there’s probably some clever startup somewhere who really have something like this. But the truth is for 99.9% of companies, the CEO isn’t going to be replaced by an algorithm anytime soon and nor are the CFOs nor are the board members. The reason is very simple. The algorithms are very good at some things today and for some things they are way better than humans are. But there are also many, many things that which humans are much better than the algorithm, and we need both. Maybe in the very long term computers will become better than humans at everything but that’s going to pick up. In the foreseeable future, what we’re going to need is a very close combination of people and computers doing most things. In fact, that people will be successful are the ones that who know how to use computers to complement what they do.

Automation is a little bit like having a horse. If you have a horse, you don’t try to outrun it, you ride the horse, and as a result of which you can go farther.

Here’s an example, chess. At one point Deep Blue beat Kasparov, and now chess programs are the world champions, computers play the best chess in the world, right? Actually wrong. The best chess players in the world today are not computers, they’re what is called in the community Centaur. A Centaur is a mythical creature with half man half horse, but in this case, it’s half men half computer. The best chess players in the world today are a team of a human and the computer because the human can see things that the computer can’t, then the computer can also do a lot of work that is completely beyond what the human unaided would do. I think this thing that happens in chess that a combination of human and machine does better than just a machine is going to be the pattern in the great majority of occupation.

I think this thing that happens in chess that a combination of human and machine does better than just a machine is going to be the pattern in the great majority of occupation.

That’s how people should be thinking. The best way to protect your job from automation is to automate it yourself. How can I automate the parts of my job that can be done by a computer that is probably more routine? Then I can maybe do the more interesting ones and get to do more of the stuff that I wished I had time to do all along. For example, anything that is written, that is structured, they excel, and they have very clear rules that a computer can probably do. But things that involve integrating a lot of information, rather a type of reasoning, cannot be done by computers.

By the way, another thing that we in this field find very surprising is that people used to think, computers are going to automate the blue-collar jobs, and then the white-collar ones will be the harder ones to automate. This is actually not the case. Some blue-collar jobs are easy to automate, but some are hard, like a construction worker, it’s hard to automate. We don’t have robots that can walk around the construction site without tripping over. On the other hand, there are many white-collar jobs like a financial advisor or the direct market that are very and in fact, increasingly being replaced by computers. The question is not so much white collar versus blue collar, as routine versus non-routine if you will.

Sanjog: If you were to qualify and say, okay, this is where AI is very good at as we see it today and this is where humans should be proud of, in terms of where they are even better than the most sophisticated AI that exists.

Pedro: What it’s missing is often the context. What the human beings in the hedge fund can do that a computer can’t do. They can take a huge variety of other factors into account that the computer doesn’t even know, exist. They understand politics; they understand what’s happening globally, in between countries and regulation and so on and so forth. There’s all flow of things that the algorithms just have no idea about, and it’s the human being that has an idea about that.

I would say that, and again it’s important to realize that the frontier between what is done by the algorithm and what is done by people is continuously shifting. It doesn’t stay in the same place. But I would say, going back to the hedge fund example, what the computer can do that the people can’t do very well is that it can look at thousands of stocks very quickly or in the case of millions of consumers. It can just do things on a scale that people can’t. Having said that the extent to which Amazon or their customer which one of these algorithms know the company, that might be something less and is very limited. It’s not very deep; it’s based on, really for each individual customer, not that much information for each company.

Also, what it’s missing is often the context. What the human beings in the hedge fund can do that a computer can’t do. They can take a huge variety of other factors into account that the computer doesn’t even know exist. They understand politics; they understand what’s happening globally, in between countries and regulation and so on and so forth. There’s all flow of things that the algorithms just have no idea about, and it’s the human beings have an idea about that.

…what the computer can do that the people can’t do very well is that it can look at thousands of stocks very quickly or in the case of millions of consumers. It can just do things on a scale that people can’t.

Sanjog: You’re saying essentially the effective way is machine learning but leaving the softer, the human-specific issues like politics and change of behavior. Basically looking at that fuzzy side of how humans are able to connect the dots that still is a privilege of humans versus coming to machines.

Pedro: Yeah, the fuzzy side, I might say more the higher-level thing. What machine learning algorithms typically do in one of these companies? Let’s say automated for example in Google, or Amazon, or Facebook, is- they make the day-to-day decisions, the point by point decisions. Every time you see an ad on Google or Facebook actually, there was an option to decide what to put there. They were machine learning algorithms predicting the probability of each of the candidate ads to be clicked on. These millions or even billions of continuous decisions about what to do are made by machines. But what the humans are doing is a very important part which is they vet this at a higher level. The captain of the ship is still a human. The human beings, unlike the algorithms can say- here is the big market that we should be heading to next. That type of decision can’t really be made by learning algorithms. They don’t know or understand enough to do that. But they can gather a lot of data, and you can make, at the end of the day, some things for which that it can replace insight and intuition. But there are things for it can’t, and those decisions are still being made by humans. They do the higher-level steering and the computers, you know, somebody I know says that machine learning is like having an army of camps. Each camp by itself is not very powerful, but you have a lot of them together being used for all these different things, it adds up to a lot. But these individual things are still fairly small, the really big ones, for the most part, are still the problem to humans.

Sanjog: Let’s take examples of the type of business decisions such someone would make, in terms of the approach or the quality of those decisions. Sometimes those decisions in an enterprise could be made fast, or in a sense, you want to do it like yesterday if that’s possible? Or you would do it in the small incremental type of decisions for short, mid or long-term. You could even make some decisions to say, let’s get started on this journey, and we will play by the ear and figure out as we go. Now, this is also being done by people using their gut partly and partly some data. Do you think such decision-making which doesn’t have really set boundaries or set parameters are good candidates and if you put AI, they will do a better job at it?

Pedro: Well what happens a lot today with this type of decisions is that human will generate the list of things to do. But then it’s not set by you. The machine learning then takes over- either it gets tried out and measured and the machine learning figures out what it works on it. Something that uses a lot today is what I call AB tests. An AB test is the companies currently doing things one way, and then somebody had the idea of trying things, doing things in A way, and then some of the things, doing things the B way. Then we just try them both online. We continue to do it as before, but then we pick out random 100,000 customers, and we try the B way on those, and we see if the results are better. These types of AB test are being done continuously by the thousand every day by these companies. You’ve participated in thousands of AB test, yet you just don’t know, you’re just not aware of it.

Well what happens a lot today with this type of decisions is that human will generate the candidate things to do. But then it’s not set by you. The machine learning then takes over- either it gets tried out and measured and the machine learning figures out what it works on it.

And the AB test can be for large things, like offering this new product or getting into a market. More commonly, however, they are for things that are smaller, but they can make a lot of difference. For example, Google at one point did an AB test for what shade of blue to make the links on the ad. They tried 50 different shades of blue for the links that you can click on the ad. This is a very straightforward thing to do for Google. As a result, they picked the shade of blue that made people most likely to click on it, and as a result of that, they are now making $200 million a year more. $200 million is not a large percentage of the money that you make which is, I don’t know, in the tens of billions but considering that this was probably the work of one person over a week, then this is giving a return of $200 million, it’s an amazing result for a simple AB test. Again, these kinds of AB tests are being done for all sorts of things all the time by a lot of different companies.

I think increasingly more, and more of this information that was not accessible to algorithms is accessible, and people will develop the algorithms to make use of it.

Sanjog: This is very fascinating the way you say. Now, coming to those areas where you mention that humans can see the politics across countries or across regimes and things which machine learning is the yet not able to do. So are we trying to go in that direction that you can completely remove the reasons why a human should make any decision?

Pedro: Well, yes. For example, hedge funds these days are using things like mining Twitter data to figure out what are things that people are talking about and how those things might affect stocks or that fits a particular company. People are enormously creative in the way that they can use machine learning and more and more. Again, information about people’s sentiment and about what’s happening in politics used to be hard come by but now there are things like Twitter that you can mine. Or there are things like satellite data, for example. There are hedge funds that use satellite images of the parking lots of the Walmart of this world to see how good the business is. Are there a lot of cars parked in the parking lot? I think increasingly more, and more of this information that was not accessible to algorithms is accessible, and people will develop the algorithms to make use of it.

But then what happens is that the human beings go up one level. They can now, based on these things, still make decisions that the algorithms wouldn’t be able to. I think eventually to have algorithms that do everything that humans can; we’re just going to need the really deep breakthroughs in AI that haven’t happened yet. State of the art in AI is far beyond what it was 50 years ago, so, in my book, The Master Algorithm, I talk about the different types of machine learning that exist and what they can do. There are these five major paradigms that I call the five paradigms of machine learning, but the truth is each of these paradigms can do some things very well, these are man’s work, evolution, automating the scientific method to have something like the equivalent of a scientist.

I think eventually to have algorithms that do everything that humans can; we’re just going to need the really deep breakthroughs in AI that haven’t happened yet.

We are finally going to need a grand unified theory of machine learning in the same way that the standard model is the grand unified theory of physics, to actually have all the intelligence that human beings have in one system. Once we have that, I think it is possible that the majority of jobs will end up being done by computers. People vary a lot in their estimate of how far along that road we are. My personal opinion is that, yes, we’ve come a thousand miles, but there are a million more miles to go. But these days because machine learning is so economically important, the amount of investment that is being done in it is enormous. There’s more promising machine learning happening in one year today than there used to happen in a decade. Even though we have a million more miles to go, it might not take a huge amount of time to get there. But it’ll probably still be in the decades. But the truth is nobody knows for sure.

Progress in science and technology is not known yet; we are making very rapid progress in AI today. You have these long periods of slow progress and then, who knows, what’s going to happen. All we can say is that people are going to be making progress and companies are going to be investing in this. But exactly what will happen when is very hard to predict.

My personal opinion is that, yes, we’ve come a thousand miles but there are a million more miles to go, so we’re still quite a distance from it. But these days because machine learning is so economically important, the amount of investment that is being done is enormous.

Sanjog: Let’s take a quick break listeners, and we will be right back after the break and talk about leadership. Someone who makes a good decision for the business may not really be as good as they need to be with people and be able to gain credibility through what they do, how do they develop people? Now, I’m going on the softer side of what makes a good leader. Now, in this case, when we’re talking an enterprise context, we are making decisions, but then also we have to work on culture, also take care of the people so that they follow you as you come up with the decisions so that it gets adopted and you realize the dream. With all of that, do you think AI has a role to play or if AI is going to take the credit for even be able to do that, then are we proposing the end of organizational leadership? Making decisions, whether they’re good or bad, developing culture, getting to know people and have them follow you is what makes a leader. If at each juncture, the leader is not gaining credibility, there will be a problem in getting the decisions which otherwise could have been made by humans. Or by AI to get adopted and realize the very dream that you are chasing. That said, do you think AI is proposing or is it leading us to a stage where most decisions which are of course the people who are working within the organizations are aware of is being made by, are being made by some machine learning, more and more so. Then which is less and less emotional, which is less and less human, do you think that’s going to undermine or eliminate or destroy. Not rather eliminate but destroy or undermine the organizational leadership that people look up to other people and not to an algorithm.

Pedro: Increasingly leaders that just make their decisions with some gut feeling showing will have less credibility, the leaders that make their decisions based on, on analysis of data and so on, will have more credibility.

Well, I would say that what’s going to happen is that people will trust their leaders more if the leaders made their decisions with the help of AI machine learning that if they didn’t. But it’s one thing to say like, “I made a decision based on my gut instinct,” it’s another thing to say, “I made the decision based on this data and based on these results.” A lot of, the more data-based companies what happened, it’s not that the leadership has become…I mean look at Amazon, like Jeff Bezos, part of what he gives to Amazon is this culture of being very data-driven – of saying – we have this disagreement about what to do, let the data decide, let’s go out and try things out in the real world. I think it would be more the other way around is that increasingly leaders that just make their decisions with some gut feeling showing will have less credibility, the leaders that make their decisions based on, on analysis of data and so on, will have more credibility.

Having said that some decisions would be made by algorithms, the experience far seems to be that people at first are resistant to that. People don’t like the idea of algorithms making decisions for them, and they tend to hold algorithms to an impossibly high standard. A human being with doing the job, they were good at 80% of the time, but an algorithm will not be adopted unless it’s good a 100% of the time. People often have these unrealistic expectations for the algorithm. But then what happens is they start to use the algorithms that they quickly see that, well, this algorithm actually makes a very good decision and then they actually like it.

For example, there was a survey not long ago where they ask people, would you prefer to have a machine as your God? You would think that the vast majority people would say absolutely no, and actually it turns out, a large number of people said yes. The thing is that if you have a great boss, then you don’t want a machine instead of the boss. But if your boss sucks and some bosses do suck, then a machine maybe is actually better. When they ask people, why is it? They’ll say, “The machine will make decisions more consistently, and it will be fairer. I’ll be less dependent on their good and bad moods because it won’t have good and bad moods, and I will have a clearer picture of what it is that I need to do to perform well.” People want to know, well what is good versus bad performance, and often with algorithms, it’s very clear, it’s actually given explicitly to people what are the parameters, and they actually feel more comfortable doing that.

Marc Andreessen, who is the founder of Netscape and is also a famous venture capitalist, says that in the future there will be two kinds of jobs. There will be the jobs where you tell a computer what to do, and there will be jobs where a computer tells you what to do. I think this captures something. It’s not necessarily the case that being told what to do by a computer is a bad thing. Certainly, have a computer doing the things that you tell it to can often be – can make up your life a lot easier than having to have those things being done by people and have to explain that people of what you need to do and what not. But actually, we think that the deeper reality is that all of us already are doing a mix of telling computers what to do and being told what to do by them.

For example, let’s say you want to go on a date, you can use a dating site to find a date. The decision to go there was yours, but then the dates are selected to you by the site. Then you decide to go a restaurant but then Yelp is what selects the restaurants to go to, and then you drive there, and then it’s the car GPS system that figures up the route to take. In some sense, you are following its instructions, and you’re just doing the driving. But then when you doing the driving, the car is a bunch of computers that are basically helping you drive where you want to go. There’s a really intricate mesh of things that are done by the human and things that are done by the computer. But at the high level, the control is in your hands. At the end of the day, the computers are just extensions of us. They’re just there to do what we want them to do. It’s for us to figure out what it is that we want to do.

Sanjog: One last question I have for you, 30 seconds. What do you think should be the change in mindset of the leaders and how do they get the people to look at AI a little differently like after this show many people should look at it differently, but otherwise, in organizations, what do you want leaders to do so that AI could be introduced to people trusting it?

Pedro: I think the key idea is that an organization with AI is different from a traditional organization. It’s more flexible; it’s more adaptable, it can move more quickly. Everybody has more visibility into the company. What really AI can do for companies to make its intelligence more than the sum of the parts. Today, the intelligence of a large organization is less than the sum of the parts because people can only communicate so much and the organization doesn’t know what it knows. With AI, you can actually have the intelligence of the whole be greater than the intelligence of the part. This is I think what leaders want to be driving. There’s going to be a new type of company in the 21st century which we already see some of them becoming, and it’s the job very much of the leaders of the company to drive their companies in that direction.

The key idea is that an organization with AI is different from a traditional organization. It’s more flexible; it’s more adaptable, it can move more quickly. Everybody has more visibility into the company. What really AI can do for companies to make its intelligence more than the sum of the parts.

Sanjog: On behalf of the show and our listeners, I really like to thank you, Pedro. You really opened some eyes today, giving people a perspective of how AI can actually be working shoulder to shoulder with humans to create value, so it is important for leaders to, of course, convey this message to their people that if you introduce AI into an organization, it will be of a benefit to them and organization, and thus get a better adoption, and get people to trust AI as part of their lives. Thank you so much, again.

Pedro: Thanks for having me.

Sanjog: Listeners, hope you enjoyed this fascinating conversation. Like us on Facebook search for CTN and please be sure to follow us on Twitter and join our LinkedIn community. Thank you again for listening to this segment on CTN. This is Sanjog Aul, your talk show host. Until next week, take care, and God bless.

Explore More

Contributors

Pedro Domingos

Pedro Domingos, Professor of Computer Science, University of Washington

Pedro Domingos is a professor of computer science at the University of Washington and the author of "The Master Algorithm", the best-selling popular-science introduction to machine learning. He is a winner of the SIGKDD Innovation Award, th... More   View all posts
Add Comment
Click here to post a comment

Advertisement

Citrix DaaS MPU 300X250
Pedro Domingos

Login


Not Member Yet?
Register

Register

  • Name

  • Contact Info

  • About Yourself

  • Minimum length of 8 characters
  • Upload
  • Location

  • Professional Background

  • Other Social Profiles

  • Areas of Interest