KGOU

Jeremy Howard: Will Artificial Intelligence Be The Last Human Invention?

Apr 21, 2017
Originally published on April 21, 2017 12:27 pm

Part 4 of the TED Radio Hour episode The Digital Industrial Revolution.

About Jeremy Howard's TED Talk

Data Scientist Jeremy Howard has studied machine learning for 25 years. He says artificial intelligence can help achieve amazing things. But he warns the impact on jobs may cause a great deal of social instability.

About Jeremy Howard

Jeremy Howard is a data scientist and the founding researcher at fast.ai — a company dedicated to making deep learning accessible.

Previously, Jeremy was the CEO of Enlitic, an advanced machine learning company. He was also the president and chief scientist at Kaggle, a community and competition platform of over 200,000 data scientists.

In addition to his research, Jeremy is a faculty member at Singularity University and a researcher in residence at the University of San Francisco.

Copyright 2017 NPR. To see more, visit http://www.npr.org/.

GUY RAZ, HOST:

It's the TED Radio Hour from NPR. I'm Guy Raz. And on the show today, ideas about a new industrial revolution.

JEREMY HOWARD: I mean, it's not just a new phase of the Industrial Revolution. It's a - it's an entirely new revolution.

RAZ: This is data scientist Jeremy Howard.

HOWARD: So we went through the process of replacing hunting and gathering with domestication. We went through the process of replacing animal energy with mechanical energy. We're now going through the process of replacing human intelligence with artificial intelligence.

(SOUNDBITE OF MUSIC)

RAZ: So for the past 25 years, Jeremy has been working on a technology called deep learning, and it's based on the way the human brain and nervous system work.

HOWARD: Deep learning relies on a particular kind of function called a neural network. It is heavily inspired by neuroscience and can actually compute anything.

RAZ: Anything because these machines can learn and perceive. They can see, hear, read, write. They can make decisions all while being able to process billions of data points.

HOWARD: It's creepy.

RAZ: Yeah.

HOWARD: And it's possibly about to get creepier.

RAZ: (Laughter) Oh, no. But before we get to the creepy part, we should point out we're already using these neural networks for a lot of pretty cool things.

HOWARD: So today, we have a thousand-layer neural networks doing things like Skype translation. I don't know if you've tried that.

RAZ: Yeah. We just actually tried it.

(SOUNDBITE OF SKYPE LOADING)

CASEY HERMAN, BYLINE: Hey, you can hear me?

COMPUTER-GENERATED VOICE #1: (Speaking French).

MATHILDE: (Speaking French).

COMPUTER-GENERATED VOICE #2: Yes. Hello. We hear you very well.

RAZ: OK. Let me explain what's going on here. Our producer, Casey, is speaking into Skype...

HERMAN: OK, great. So you're recording.

RAZ: ...Obviously in English. And Skype is translating that in real time into French.

COMPUTER-GENERATED VOICE #1: (Speaking French).

RAZ: And on the other end is Mathilde (ph)...

MATHILDE: (Speaking French).

RAZ: ...Except she's speaking in French, and Skype turns that into English.

COMPUTER-GENERATED VOICE #2: Yes, we are recording.

RAZ: And how does this work?

HOWARD: The basic approach is that you download lots of sentences that are written both in English and French. You basically write three or four lines of code, and you then tell that neural network where to find the input and output data, leave it overnight and come back in the morning and see if it works. So a sentence could be please enter your username here.

COMPUTER-GENERATED VOICE #1: (Speaking French).

HOWARD: It learns to map at a very deep and subtle way such that the next day I can then put in a different sentence. What is the world's largest country?

COMPUTER-GENERATED VOICE #1: (Speaking French).

RAZ: And Jeremy says that's all it takes. You start with simple sentences.

MATHILDE: (Speaking French).

COMPUTER-GENERATED VOICE #2: What is the distance between the Earth and the sun?

RAZ: As you keep going, the program keeps learning on its own...

MATHILDE: (Speaking French).

COMPUTER-GENERATED VOICE #2: What is the population of the city of Paris?

RAZ: ...To the point where you can have entire conversations translated...

HERMAN: Hey, good morning, how are you doing?

COMPUTER-GENERATED VOICE #1: (Speaking French).

RAZ: ...In real time....

MATHILDE: (Speaking French).

RAZ: ...By a machine.

COMPUTER-GENERATED VOICE #2: I'm fine, thank you.

HERMAN: How is your family? Have you talked to them?

COMPUTER-GENERATED VOICE #1: (Speaking French).

RAZ: I mean, this concept is absolutely incredible - right? - because with this technology, once you hit go, it's unstoppable. I mean, it just continues to get better and better and better, right?

HOWARD: Right. So you set up your problem, and you run it overnight, and you come back the next day, and hopefully, it solved it. Now, it's pretty hard to describe exactly what it's done, and often - very, very often - I don't really understand at all how my programs work. A few years ago, I built a system for diagnosing lung cancer which could beat a panel of four of the world's best radiologists. But I know nothing whatsoever about what lung cancer looks like or doesn't look like and could tell you nothing about how it worked because I have no background in medicine. So that's the nature of setting up these things. You fire them off and they come back with a model.

RAZ: And Jeremy says diagnosing lung cancer is just the beginning.

HOWARD: And so you can imagine the power that this provides for good. Every time a piece of medical imaging is done, automatically and instantly an alert appears saying this person has an aneurism or this person has a malignant nodule in their left lung. You know, that would be saving millions of lives and billions of dollars.

RAZ: Here's how Jeremy explained it from the TED stage.

(SOUNDBITE OF TED TALK)

HOWARD: This kind of technique could allow us to fix a major problem, which is that there's a lack of medical expertise in the world. The World Economic Forum says that there's between a 10x and a 20x shortage of physicians in the developing world. And it would take about 300 years to train enough people to fix that problem. So imagine if we can help enhance their efficiency using these deep learning approaches. So I'm very excited about the opportunities. I'm also concerned about the problems. The problem here is that every area in blue on this map is somewhere where services are over 80 percent of employment. What are services?

RAZ: OK. So at this point in his talk, Jeremy pulls out a large world map, and he explains that in most of the developed world, and even in a lot of developing countries, their economies depend mainly on service jobs, people who prep food and drive cars and file documents, do legal research, even diagnose disease.

(SOUNDBITE OF TED TALK)

HOWARD: These are also the exact things that computers have just learned how to do. So 80 percent of the world's employment in the developed world is stuff that computers have just learned how to do. What does that mean? Well, it'll be fine. It'll be replaced by other jobs. For example, there'll be more jobs for data scientists. Well, not really. It doesn't take data scientists very long to build these things. For example, these four algorithms were all built by the same guy.

So if you think, oh, it's all happened before. You know, we've seen the results in the past of when new things come along, and they get replaced by new jobs. What are these new jobs going to be? It's very hard for us to estimate this because human performance grows at this gradual rate. But we now have a system, deep learning, that we know actually grows in capability exponentially.

So currently, we see the things around us and we say oh, computers are still pretty dumb, right? But in five years time, computers will be off this chart. The better computers get at intellectual activities, the more they can build better computers to be better at intellectual capabilities. So this is going to be a kind of change that the world has actually never experienced before. So your previous understanding of what's possible is different.

RAZ: I mean, the fact is is that machines are getting better and better as we speak, second by second, in being able to process billions and billions of pieces of data, and they're just going to get better and better next month and in a year and in five years and 20 years. And they are going to be able to make decisions in ways that will be confounding to us.

HOWARD: I mean, so I did my TED talk two and a half years ago, I guess, and at this point, computers are better at recognizing what is in a photo than humans are. They are better at understanding Chinese and English speech than Chinese and English native speakers. We now have deep learning algorithms that are better at building the neural networks that create deep learning algorithms...

RAZ: (Laughter).

HOWARD: ...Than humans are.

RAZ: God.

HOWARD: So two and a half years since I got up and said, hey, there's this new technology called deep learning which is about to surpass human capabilities in these fundamentally human areas. Between then and now, it happened.

RAZ: Do you think we can even articulate or imagine what this technology will do to change our world, our lives, our species?

HOWARD: Not only is it impossible for us to imagine, there is a great many things which are stopping us from being able to imagine it.

RAZ: Like what?

HOWARD: So people often talk about exponential technologies, but the fact is, every technology that's come so far is actually an S-shaped technology. Before electricity, nearly everything that required an input of energy was done with human energy or horse energy or something like that. Then we learnt how to electrify nearly every energy-requiring process until eventually, we did them all. And so there was initially exponential growth as the newly electrified processes allowed us to improve other processes, but then it flattened off again.

Now, on the other hand, think about replacing intellectual power, all right? There's no S-curve here. There's no drop-off. There's no point where you go, OK, we've now used all of the intellectual power that could be used. So our ability to actually understand the outcome of a truly exponential technology cannot be based on anything that has been observed in history because there's never been a technology like this before.

RAZ: Yeah, I mean, it never stops. It doesn't reach a point where we say, you know, right, OK, next, you know, we've reached that milestone, and now we can just move on to the next thing.

HOWARD: Right. Well, that's the thing, right? This is kind of the last human input is kind of perception...

RAZ: Yeah.

HOWARD: ...And intelligence. So a lot of people then say, oh, yeah, so there'll be new jobs. But then you say, like what?

RAZ: Yeah.

HOWARD: And that's where people come up short. Like, you could tell, even in the industrial revolution, you could say OK, electricity can't do all of these things - look at things or listen to things or react to things. That's not where we are anymore. We're now in the process of saying, OK, this is the last bastion of stuff that only humans can do.

RAZ: Yeah, I mean, I've been asking everyone on the show this question, and, you know, they seem pretty convinced that, you know, we humans will continue to have jobs. Like, you and I will continue to do what we do and will continue to be necessary as a species, but I'm not really so sure.

HOWARD: Yeah. There's no question that the vast majority of things that humans are doing today are going to be replaced. And it's also important to remember, though, that the vast majority of people have jobs that kind of suck, right?

RAZ: Yeah.

HOWARD: There's far more people who are chopping celery for a living than there are people running TED Radio Hour for a living. So given that we both agree that the jobs that we have today are going to disappear probably quicker than anybody expects, that means that we need to have ways of supporting people...

RAZ: Yeah.

HOWARD: ...Economically because you can't just say, you're not adding value, so therefore, you're not worthwhile to society. You actually have to say OK, we think everybody deserves to live in dignity, so let's have a basic guaranteed income which is enough to ensure that every human can have a life of dignity.

RAZ: I mean, do you think we're, like, designing our replacement? Because you could imagine a point when, you know, we won't be the most intelligent species on the planet. Like, we would still live on it. We would still reproduce, but...

HOWARD: Right.

RAZ: ...But, I mean, intelligent machines will create a new economic model and a system of governments for us on our behalf. We're going to be, like, zoo animals.

HOWARD: Yeah, I don't think that's too hard to imagine. You know, if we continue down this path where more and more of the economy looks like the inside of a Amazon logistics warehouse - where every human being, every move they're making, is being supervised by and rewarded or penalized by machines until we're all at the beck and call of the machinery and we could well look at each other and go, none of us signed up for this.

RAZ: Yeah.

HOWARD: None of us designed this.

RAZ: It's just hard for me to wrap my head around this idea that, you know, we have consciousness. I mean, even though we're not sure what exactly animates us - right? But I just have trouble believing that, you know, one day, we're going to develop machines that think and feel and act like we do.

HOWARD: The more I use this technology, the less my own consciousness seems surprising or inexplicable.

RAZ: Really?

HOWARD: Oh, very much so.

RAZ: Wow.

HOWARD: Once you've used these kinds of tools for long enough, you just keep seeing them behave in certain ways. It just looks very familiar.

RAZ: But you're basically saying that the ones and zeros in a machine are very similar to the ones and zeros in your brain.

HOWARD: Well, clearly. I mean, it has to be true. The alternative is if you believe that there's a God...

RAZ: Yeah?

HOWARD: ...And there's a soul and it's this ineffable thing we can't see or touch...

RAZ: Yeah.

HOWARD: ...OK, that's fine. I can't, and I'm not going to argue with that because it's based on faith, not science. Unless you believe those things, you have to believe that the chemical and electrical signals and connectivity in our brain is us. You know, the idea that these kinds of massively parallel, massively connected functions inside huge computer systems can display intelligent behavior - well, we're already there. We just keep redefining intelligent behavior so that, you know, we're always outside it.

RAZ: But, I mean, as a human, I can emote. I can feel things. You know, like, I understand my mortality. I love. I feel grief (laughter). I mean, how are those things replicated by - you know, in a machine by ones and zeros?

HOWARD: Very, very easily. All of the things you are describing are rational, evolutionary responses to a fitness function that attempts to keep your genes alive. The reason you love is because your genes want you to have sex and have more of those genes. The reason that you grieve is because your genes want you to avoid doing things that cause people with similar genes to yours to die. All of the things that you emote are, at some level, evolutionary responses to this fitness function of keeping your genes going for longer. So as our complex, computer-based functions get better at figuring out what causes their fitness functions to be successful, they will have all of these very complex interactions.

RAZ: I'm not walking away from this conversation optimistic. I'm troubled.

HOWARD: Good.

RAZ: Yeah, I'm troubled.

HOWARD: Good. Be troubled.

RAZ: Yeah.

HOWARD: You know, being troubled is the correct response - but also excited. Be both. Because all of these people who say don't worry, everything's fine - there will always be jobs. Just feed the robots; feed the machines. Don't worry - to me, they're just the same as climate-change deniers. You know, they're ignoring the basic science that says this is something different. This is something we can't control. This is something we don't know where it's going. And this is something which can definitely have terrible societal outcomes.

On the other hand, anybody who says the government should step in and regulate and control and stop all of this from happening - people shouldn't be doing this kind of research - they're just as bad. You know, they're saying the billions of people that don't have access to modern medicine should never be given it even though we have the technology today to provide it. So we need to be both. We need to be aware of the opportunities and aware of the threats.

(SOUNDBITE OF MUSIC)

RAZ: Jeremy Howard - he's a data scientist and the founding researcher at fast.ai. It's a company dedicated to making deep learning accessible to everyone. You can see his entire talk at ted.com.

(SOUNDBITE OF SONG, "ROBOT")

THE FUTUREHEADS: (Singing) I am a robot living like a robot, talk like a robot in the habitating way. Look up to the sky...

RAZ: Hey, thanks for listening to our show on the Digital Industrial Revolution this week. If you want to find out more about who was on it, go to ted.npr.org. To see hundreds more TED Talks, check out ted.com or the TED app. Our production staff at NPR includes Jeff Rogers, Sanaz Meshkinpour, Jinae West, Neva Grant and Rachel Faulkner with help from Ramtim Arablouei and Daniel Shukin. Our intern is Thomas Lu. Our partners at TED are Chris Anderson, Kelly Stoetzel, Anna Phelan and Janet Lee. I'm Guy Raz, and you've been listening to ideas worth spreading right here on the TED Radio Hour from NPR.

(SOUNDBITE OF SONG, "ROBOT")

THE FUTUREHEADS: (Singing) I am a robot, living like a robot, talk like a robot in the habitating way. Transcript provided by NPR, Copyright NPR.