Ray Kurzweil: Future of Intelligence | MIT 6.S099: Artificial General Intelligence (AGI)

00:00:00.000 | - Welcome to MIT course 6S099,

00:00:04.240 | Artificial General Intelligence.

00:00:06.020 | Today we have Ray Kurzweil.

00:00:09.960 | He is one of the world's leading inventors,

00:00:12.040 | thinkers, and futurists,

00:00:13.480 | with a 30-year track record of accurate predictions.

00:00:16.800 | Called the restless genius by the Wall Street Journal

00:00:19.640 | and the ultimate thinking machine by Forbes magazine.

00:00:22.960 | He was selected as one of the top entrepreneurs

00:00:25.840 | by Inc. Magazine, which described him

00:00:28.280 | as the rightful heir to Thomas Edison.

00:00:31.520 | PBS selected him as one of the 16 revolutionaries

00:00:34.360 | who made America.

00:00:35.380 | Ray was the principal investigator

00:00:38.940 | of the first CCD flatbed scanner,

00:00:41.060 | the first omni-font optical character recognition,

00:00:44.320 | the first point-to-speech reading machine for the blind,

00:00:47.000 | the first text-to-speech synthesizer,

00:00:49.100 | the first music synthesizer capable of recreating

00:00:52.000 | the grand piano and other orchestral instruments,

00:00:55.040 | and the first commercially marketed

00:00:56.920 | large vocabulary speech recognition.

00:00:59.400 | Among Ray's many honors, he received a Grammy Award

00:01:02.620 | for Outstanding Achievements in Music Technology.

00:01:05.000 | He is the recipient of the National Medal of Technology,

00:01:07.720 | was inducted into the National Inventors Hall of Fame,

00:01:10.920 | holds 21 honorary doctorates and honors

00:01:14.320 | from three US presidents.

00:01:16.940 | Ray has written five national best-selling books,

00:01:20.240 | including the New York Times bestsellers

00:01:23.160 | The Singularity is Near from 2005,

00:01:25.640 | and How to Create a Mind from 2012.

00:01:28.840 | He is co-founder and chancellor of Singularity University

00:01:33.120 | and a director of engineering at Google,

00:01:36.260 | heading up a team developing machine intelligence

00:01:38.720 | and natural language understanding.

00:01:40.600 | Please give Ray a warm welcome.

00:01:42.640 | (audience applauding)

00:01:45.800 | - It's good to be back.

00:01:52.640 | I've been in this lecture hall many times

00:01:55.800 | and walked the infinite corridor.

00:01:57.760 | I came here as an undergraduate in 1965.

00:02:02.200 | Within a year of my being here,

00:02:05.100 | they started a new major called computer science.

00:02:08.900 | It did not get its own course number.

00:02:12.600 | It's 6.1.

00:02:15.000 | Even biotechnology recently got its own course number.

00:02:18.600 | How many of you are CS majors?

00:02:22.560 | Okay, how many of you do work in deep learning?

00:02:25.680 | How many of you have heard of deep learning?

00:02:29.680 | I came here first in 1962 when I was 14.

00:02:36.880 | I became excited about artificial intelligence.

00:02:42.000 | It had only gotten its name six years earlier,

00:02:45.440 | the 1956 Dartmouth Conference

00:02:48.280 | by Marvin Minsky and John McCarthy.

00:02:52.660 | So I wrote Minsky a letter.

00:02:56.500 | There was no email back then.

00:02:58.780 | And he invited me up.

00:03:00.540 | He spent all day with me as if he had nothing else to do.

00:03:03.100 | He was a consummate educator.

00:03:05.740 | And the AI field had already bifurcated

00:03:11.580 | into two warring camps,

00:03:13.220 | the symbolic school which Minsky was associated with

00:03:18.740 | and the connectionist school was not widely known.

00:03:22.780 | In fact, I think it's still not widely known

00:03:24.380 | that Minsky actually invented the neural net in 1953.

00:03:29.300 | But he had become negative about it,

00:03:32.100 | largely 'cause there was a lot of hype

00:03:34.120 | that these giant brains could solve any problem.

00:03:37.680 | So the first popular neural net, the perceptron,

00:03:45.040 | was being promulgated by Frank Rosenblatt at Cornell.

00:03:49.280 | So Minsky said, "Oh, where are you going now?"

00:03:50.920 | And I said to see Rosenblatt at Cornell.

00:03:54.920 | He said, "Don't bother doing that."

00:03:56.920 | And I went there and Rosenblatt was touting the perceptron

00:04:01.920 | that it ultimately would be able to solve any problem.

00:04:04.980 | So I brought some printed letters that had the camera

00:04:08.120 | and it did a perfect job of recognizing them

00:04:10.960 | as long as they were carrier 10,

00:04:14.080 | different type style didn't work at all.

00:04:16.840 | And he said, "But don't worry.

00:04:18.320 | "We can take the output of the perceptron

00:04:20.560 | "and feed it as the input to another perceptron

00:04:22.560 | "and take the output of that and feed it to a third layer.

00:04:25.460 | "And as we add more layers,

00:04:26.660 | "it'll get smarter and smarter and generalized."

00:04:29.520 | And I said, "That's interesting.

00:04:30.960 | "Have you tried that?"

00:04:31.960 | Well, no, but it's high on our research agenda.

00:04:35.240 | Things did not move quite as quickly back then

00:04:38.800 | as they do now.

00:04:39.720 | He died nine years later, never having tried that idea.

00:04:44.160 | Turns out to be remarkably prescient.

00:04:46.040 | I mean, he never tried multi-layer neural nets

00:04:49.520 | and all the excitement that we see now about deep learning

00:04:52.440 | comes from a combination of two things,

00:04:58.320 | many layer neural nets and the law of accelerating returns,

00:05:04.480 | which I'll get to a little bit later,

00:05:07.120 | which is basically the exponential growth of computing

00:05:10.440 | so that we can run these massive nets

00:05:12.580 | and handle massive amounts of data.

00:05:15.180 | It would be decades before that idea was tried.

00:05:20.600 | Several decades later, three-level neural nets were tried.

00:05:24.980 | They were a little bit better.

00:05:25.940 | They could deal with multiple type styles,

00:05:28.280 | still weren't very flexible.

00:05:30.400 | Now, it's not hard to add other layers.

00:05:33.580 | It's a very straightforward concept.

00:05:36.260 | There was a math problem,

00:05:37.800 | the disappearing gradient or the exploding gradient,

00:05:41.700 | which I'm sure many of you are familiar with.

00:05:43.900 | Basically, you need to take maximum advantage

00:05:50.000 | of the range of values in the gradients,

00:05:56.260 | not let them explode or disappear and lose the resolution.

00:06:00.260 | That's a fairly straightforward

00:06:02.600 | mathematical transformation.

00:06:04.760 | With that insight, we could now go to

00:06:06.780 | 100-layer neural nets.

00:06:09.680 | And that's behind sort of all the fantastic gains

00:06:13.880 | that we've seen recently.

00:06:16.140 | AlphaGo trained on every online game

00:06:22.560 | and then became a fair Go player.

00:06:27.920 | It then trained itself by playing itself

00:06:30.960 | and soared past the best human.

00:06:33.380 | AlphaGo Zero started with no human input at all.

00:06:37.160 | Within hours of iteration, soared past AlphaGo,

00:06:42.160 | also soared past the best chess programs.

00:06:47.480 | They had another innovation.

00:06:49.080 | Basically, you need to evaluate the quality of the board

00:06:53.400 | at each point, and they used another 100-layer neural net

00:06:58.000 | to do that evaluation.

00:06:59.880 | So, there's still a problem in the field.

00:07:07.220 | Which is, there's a motto that life begins

00:07:09.980 | at a billion examples.

00:07:11.460 | One of the reasons I'm at Google

00:07:14.100 | is we have a billion examples.

00:07:16.140 | For example, there's pictures of dogs and cats

00:07:19.220 | that are labeled, so you got a picture of a cat

00:07:21.460 | and it says cat, and then you can learn from it,

00:07:23.420 | and you need a lot of them.

00:07:24.780 | AlphaGo trained on a million online moves.

00:07:29.780 | That's how many we had of Master Games.

00:07:33.540 | And that only created a sort of fair Go player,

00:07:37.740 | a good amateur could defeat it.

00:07:39.660 | So, they worked around that in the case of Go

00:07:43.700 | by basically generating an infinite amount of data

00:07:48.700 | by having the system play itself.

00:07:50.980 | Had a chat with Demis Hassabis.

00:07:55.500 | What kind of situations can you do that with?

00:07:58.700 | You have to have some way of simulating the world.

00:08:01.740 | So, Go or Chess are, even though Go

00:08:04.740 | is considered a difficult game,

00:08:07.020 | the definition of it exists on one page.

00:08:11.780 | So, you can simulate it.

00:08:13.700 | That applies to math.

00:08:15.140 | I mean, math axioms can be contained on a page or two.

00:08:19.860 | It's not very complicated.

00:08:22.340 | Gets more difficult when you have real life situations,

00:08:26.020 | like biology.

00:08:27.660 | So, we have biological simulators,

00:08:29.460 | but the simulators aren't perfect.

00:08:30.980 | So, learning from the simulators

00:08:33.100 | will only be as good as the simulators.

00:08:35.580 | That's actually the key to being able

00:08:38.140 | to do deep learning on biology.

00:08:40.460 | Autonomous vehicles, you need real life data.

00:08:45.980 | So, the Waymo systems have gone three and a half

00:08:50.460 | million miles.

00:08:52.380 | That's enough data to then create a very good simulator.

00:08:56.660 | So, the simulator is really quite realistic

00:08:58.500 | because they had a lot of real world experience

00:09:01.660 | and they've gone a billion miles in the simulator.

00:09:06.100 | But we don't always have that opportunity

00:09:09.140 | to either create the data or have the data around.

00:09:12.820 | Humans can learn from a small number of examples.

00:09:17.820 | Your significant other, your professor, your boss,

00:09:21.580 | your investor can tell you something once or twice

00:09:24.860 | and you might actually learn from that.

00:09:27.020 | Some humans have been reported to do that.

00:09:29.780 | And that's kind of the remaining advantage of humans.

00:09:34.780 | Now, there's actually no back propagation

00:09:38.340 | in the human brain.

00:09:39.580 | It doesn't use deep learning.

00:09:41.780 | It uses a different architecture.

00:09:43.580 | That same year, in 1962, I wrote a paper,

00:09:47.100 | How I Thought the Human Brain Worked.

00:09:49.500 | There was actually very little neuroscience to go on.

00:09:52.420 | There was one neuroscientist, Vernon Mountcastle,

00:09:55.340 | that had something relevant to say, which is he did...

00:09:59.260 | I mean, there was the common wisdom at the time,

00:10:01.820 | and there's still a lot of neuroscientists that say this,

00:10:04.060 | that we have all these different regions of the brain,

00:10:06.380 | they do different things, they must be different.

00:10:10.580 | There's V1 in the back of the head

00:10:12.180 | where the optic nerve spills into,

00:10:14.140 | that can tell that that's a curved line,

00:10:18.460 | that's a straight line,

00:10:20.140 | does these simple feature extractions on visual images.

00:10:23.820 | That's actually a large part of the neocortex.

00:10:26.140 | There's a fusiform gyrus up here,

00:10:28.580 | which can recognize faces.

00:10:31.340 | We know that because if it gets knocked out

00:10:33.460 | through injury or stroke, people can't recognize faces.

00:10:36.500 | They will learn it again

00:10:37.500 | with a different region of the neocortex.

00:10:39.900 | There's the famous frontal cortex,

00:10:41.740 | which does language and poetry and music.

00:10:46.340 | So these must work on different principles.

00:10:49.220 | He did autopsies on the neocortex

00:10:51.340 | and all these different regions

00:10:52.540 | and found they all looked the same.

00:10:53.780 | They had the same repeating pattern,

00:10:55.900 | same interconnections.

00:10:57.780 | He said neocortex is neocortex.

00:10:59.780 | So I had that hint.

00:11:01.980 | Otherwise, I could actually observe human brains in action,

00:11:05.780 | which I did from time to time.

00:11:07.980 | And there's a lot of hints that you can get that way.

00:11:11.100 | For example, if I ask you to recite the alphabet,

00:11:14.140 | you actually don't do it from A to Z,

00:11:15.980 | you do it as a sequence of sequences,

00:11:18.580 | A, B, C, D, E, F, G, H, I, J, K.

00:11:22.420 | So we learn things as forward sequences of sequences.

00:11:26.180 | Forward, because if I ask you to recite the alphabet

00:11:29.380 | backwards, you can't do it unless you learn that

00:11:31.980 | as a new sequence.

00:11:33.580 | So these are all interesting hints.

00:11:36.220 | I wrote a paper that the neocortex is organized

00:11:40.300 | as a hierarchy of modules,

00:11:41.860 | and each module can learn a simple pattern.

00:11:44.660 | And that's how I got to meet President Johnson.

00:11:48.620 | And that initiated a half century

00:11:52.380 | of thinking about this issue.

00:11:54.420 | I came to MIT to study with Marvin Minsky.

00:11:58.060 | Actually, I came for two reasons.

00:11:59.340 | One, that Minsky became my mentor,

00:12:02.260 | which was a mentorship that lasted for over 50 years.

00:12:05.900 | The fact that MIT was so advanced,

00:12:07.780 | it actually had a computer,

00:12:09.260 | which the other colleges I considered didn't have.

00:12:14.540 | It was an IBM 7094, 32K of 36-bit words,

00:12:19.540 | so it's 150K of core storage,

00:12:22.420 | two microsecond cycle time,

00:12:24.180 | two cycles for instruction, so a quarter of a MIP.

00:12:26.660 | And thousands of students and professors

00:12:30.460 | shared that one machine.

00:12:32.060 | In 2012, I wrote a book about this thesis.

00:12:36.500 | It's now actually an explosion of neuroscience evidence

00:12:39.780 | to support it.

00:12:40.980 | The European Brain Reverse Engineering Project

00:12:43.900 | has identified a repeating module of about 100 neurons.

00:12:47.420 | It's repeated 300 million times,

00:12:50.100 | so it's about 30 billion neurons in the neocortex.

00:12:53.140 | The neocortex is the outer layer of the brain.

00:12:55.340 | That's the part where we do our thinking.

00:12:58.820 | And they can see in each module,

00:13:00.420 | axons coming in from another module.

00:13:04.660 | And then the output, the single output axon of that module

00:13:08.020 | goes as the input to another module.

00:13:10.260 | So we can see it organized as a hierarchy.

00:13:13.300 | It's not a physical hierarchy.

00:13:16.340 | The hierarchy comes from these connections.

00:13:18.660 | The neocortex is a very thin structure.

00:13:20.740 | It's actually one module thick.

00:13:22.980 | There's six layers of neurons,

00:13:25.340 | but it constitutes one module.

00:13:28.060 | And we can see that it learns a simple pattern.

00:13:31.540 | And for various reasons, I cite in the book,

00:13:34.780 | the pattern recognition model it's using

00:13:37.580 | is basically a hidden Markov model.

00:13:40.380 | How many of you have worked with Markov models?

00:13:43.820 | Okay.

00:13:44.660 | That's usually no hands go up when I ask that question.

00:13:49.860 | But a Markov model is not, it is learned,

00:13:54.540 | but it's not back propagation.

00:13:56.780 | It can learn local features.

00:13:58.700 | So it's very good for speech recognition.

00:14:00.500 | And the speech recognition work I did in the 80s

00:14:03.900 | used these Markov models that became the standard approach

00:14:08.220 | because it can deal with local variations.

00:14:11.540 | So the fact that a vowel is stretched,

00:14:15.460 | you can learn that in a Markov model.

00:14:18.100 | It doesn't learn long distance relationships.

00:14:21.740 | That's handled by the hierarchy.

00:14:24.340 | And something we don't fully understand yet

00:14:26.340 | is exactly how the neocortex creates that hierarchy.

00:14:30.460 | But we have figured out how it can connect

00:14:33.740 | this module to this module.

00:14:36.020 | Does it then grow?

00:14:37.340 | I mean, there's no virtual communication

00:14:39.580 | or wireless communication.

00:14:41.180 | It's an actual connection.

00:14:42.940 | So does it grow an axon from one place to another,

00:14:47.060 | which could be inches apart?

00:14:50.020 | Actually, all these connections are there from birth,

00:14:53.780 | like the streets and avenues of Manhattan.

00:14:57.700 | There's vertical and horizontal connections.

00:14:59.620 | So if it decides and how it makes that decision

00:15:03.860 | is still not fully understood,

00:15:05.740 | but it wants to connect this module to this module,

00:15:08.060 | there's already a vertical,

00:15:09.620 | horizontal and a vertical connection.

00:15:11.300 | It just activates them.

00:15:12.740 | We can actually see that now.

00:15:15.500 | And it can see that happening in real time

00:15:17.940 | on non-invasive brain scans.

00:15:21.460 | So there's a tremendous amount of evidence

00:15:23.140 | that's in fact, the neocortex is a hierarchy of modules

00:15:28.140 | that can learn, each module learns

00:15:31.900 | a simple sequential pattern.

00:15:34.980 | And even though the patterns we perceive

00:15:38.100 | don't seem like sequences,

00:15:39.700 | they may seem three-dimensional or even more complicated,

00:15:42.980 | they are in fact represented as sequences,

00:15:46.220 | but the complexity comes in with the hierarchy.

00:15:48.620 | So the neocortex emerged 200 million years ago with mammals.

00:15:54.460 | All mammals have a neocortex.

00:15:57.020 | That's one of the distinguishing features of mammals.

00:16:00.740 | These first mammals were small.

00:16:02.740 | They were rodents,

00:16:04.980 | but they were capable of a new type of thinking.

00:16:07.780 | Other non-mammalian animals had fixed behaviors,

00:16:11.220 | but those fixed behaviors were very well adapted

00:16:14.100 | for their ecological niche.

00:16:16.180 | But these new mammals could invent a new behavior.

00:16:20.340 | So creativity and innovation

00:16:21.980 | was one feature of the neocortex.

00:16:24.340 | So a mouse is escaping a predator,

00:16:27.020 | its usual escape path is blocked,

00:16:29.380 | it will invent a new behavior to deal with it.

00:16:32.700 | Probably wouldn't work,

00:16:33.540 | but if it did work, it would remember it

00:16:35.180 | and it would have a new behavior.

00:16:36.980 | And that behavior could spread virally through the community.

00:16:40.180 | Another mouse watching this would say to itself,

00:16:43.020 | hmm, that was really clever going around that rock,

00:16:45.060 | I'm gonna remember to do that.

00:16:47.140 | And it would have a new behavior.

00:16:51.260 | Didn't help these early mammals that much

00:16:54.020 | because as I say, the non-mammalian animals

00:16:56.700 | were very well adapted to their niches

00:17:01.620 | and nothing much happened for 135 million years.

00:17:05.580 | But then 65 million years ago, something did happen.

00:17:09.660 | There was a sudden violent change to the environment.

00:17:12.900 | We now call it the Cretaceous extinction event.

00:17:15.980 | There's been debate as to whether it was a meteor

00:17:18.460 | or an asteroid, I mean a meteor or a volcanic eruption.

00:17:23.460 | The asteroid or meteor hypothesis is in the ascendancy.

00:17:29.500 | But if you dig down to an area of rock

00:17:32.380 | reflecting 65 million years ago,

00:17:34.460 | the geologists will explain that it shows

00:17:37.060 | a very violent sudden change to the environment.

00:17:39.860 | And we see it all around the globe.

00:17:42.420 | So it was a worldwide phenomenon.

00:17:44.780 | The reason we call it an extinction event

00:17:48.060 | is that's when the dinosaurs went extinct.

00:17:51.300 | That's when 75% of all the animal

00:17:54.700 | and plant species went extinct.

00:17:56.660 | And that's when mammals overtook their ecological niche.

00:18:00.340 | So to anthropomorphize biological evolution

00:18:03.380 | said to itself, this neocortex is pretty good stuff

00:18:06.420 | and it began to grow it.

00:18:07.940 | So now mammals got bigger, their brains got bigger

00:18:10.740 | at an even faster pace, taking up a larger fraction

00:18:13.220 | of their body.

00:18:14.500 | The neocortex got bigger even faster than that

00:18:17.700 | and developed these curvatures that are distinctive

00:18:20.380 | of a primate brain basically to increase its surface area.

00:18:25.540 | But if you stretched it out,

00:18:27.700 | the human neocortex is still a flat structure.

00:18:29.900 | It's about the size of a table napkin, just as thin.

00:18:33.580 | And it basically created primates

00:18:39.740 | which became dominant in their ecological niche.

00:18:45.900 | Then something else happened 2 million years ago.

00:18:49.620 | Biological evolution decided to increase

00:18:53.260 | the neocortex further and increase the size

00:18:56.140 | of the enclosure and basically filled up the frontal cortex

00:19:00.260 | with our big skulls with more neocortex.

00:19:03.900 | And up until recently it was felt, as I said,

00:19:06.940 | that this was, the frontal cortex was different

00:19:09.860 | 'cause it does these qualitatively different things.

00:19:12.540 | But we now realize that it's really

00:19:18.980 | just additional neocortex.

00:19:22.220 | So remember what we did with it.

00:19:27.220 | We're already doing a very good job of being primates.

00:19:30.100 | So we put it at the top of the neocortical hierarchy

00:19:33.580 | and we increased the size of the hierarchy.

00:19:36.700 | It was maybe 20% more neocortex,

00:19:39.020 | but it doubled or tripled the number of levels

00:19:41.940 | 'cause as you go up the hierarchy,

00:19:43.180 | it's kind of like a pyramid.

00:19:44.340 | There's fewer and fewer modules.

00:19:47.380 | And that was the enabling factor for us

00:19:50.500 | to invent language and art, music.

00:19:53.860 | Every human culture we've ever discovered has music.

00:19:57.300 | No primate culture really has music.

00:19:59.780 | There's debate about that, but it's really true.

00:20:03.020 | Invention, technology.

00:20:07.500 | Technology required another evolutionary adaptation,

00:20:10.620 | which is this humble appendage here.

00:20:13.940 | No other animal has that.

00:20:15.220 | If you look at a chimpanzee,

00:20:16.220 | it looks like they have a similar hand,

00:20:17.420 | but the thumb is actually down here.

00:20:19.100 | Doesn't work very well if you watch them

00:20:20.620 | trying to grab a stick.

00:20:22.740 | So we could imagine creative solutions.

00:20:26.100 | Yeah, I could take that branch and strip off the leaves

00:20:29.460 | and put a point on it.

00:20:31.060 | We could actually carry out these ideas and create tools

00:20:36.020 | and then use tools to create new tools

00:20:38.420 | and started a whole other evolutionary process

00:20:41.260 | of tool making.

00:20:42.740 | And that all came with the neocortex.

00:20:46.620 | So Larry Page read my book in 2012 and liked it.

00:20:51.620 | So I met with him and asked him for an investment

00:20:55.460 | in a company I'd started actually a couple of weeks earlier

00:20:58.820 | to develop those ideas commercially

00:21:01.660 | 'cause that's how I went about things

00:21:03.140 | as a serial entrepreneur.

00:21:05.380 | And he said, "Well, we'll invest,

00:21:08.380 | "but let me give you a better idea.

00:21:09.540 | "Why don't you do it here at Google?

00:21:11.500 | "We have a billion pictures of dogs and cats

00:21:13.780 | "and we've got a lot of other data

00:21:15.420 | "and lots of computers and lots of talent,

00:21:17.660 | "all of which is true."

00:21:19.540 | And I says, "Well, I don't know.

00:21:20.380 | "I just started this company to develop this."

00:21:23.860 | He says, "Well, buy your company."

00:21:25.140 | And I said, "How are you gonna value a company

00:21:26.940 | "that hasn't done anything?

00:21:28.580 | "It just started a couple of weeks ago."

00:21:30.020 | And he said, "We can value anything."

00:21:32.380 | So I took my first job five years ago

00:21:38.460 | and I've been basically applying this model,

00:21:41.740 | this hierarchical model to understanding language,

00:21:46.620 | which I think really is the holy grail of AI.

00:21:50.500 | I think Turing was correct in designating

00:21:54.700 | basically text communication

00:21:56.740 | as what we now call a Turing-complete problem

00:22:01.260 | that requires, there's no simple NLP tricks

00:22:04.420 | that you can apply to pass a valid Turing test

00:22:08.380 | with an emphasis on the word valid.

00:22:10.860 | Mitch Caper and I had a six-month debate

00:22:13.740 | on what the rules should be

00:22:15.020 | 'cause if you read Turing's 1950 paper,

00:22:18.820 | he describes this in a few paragraphs

00:22:20.700 | and doesn't really describe how to go about it.

00:22:22.380 | But if it's a valid Turing test,

00:22:25.180 | meaning it's really convincing you

00:22:27.620 | through interrogation and dialogue that it's a human,

00:22:32.620 | that requires a full range of human intelligence.

00:22:37.820 | And I think that test has stood the test of time.

00:22:41.460 | We're making very good progress on that.

00:22:44.260 | I mean, just last week, you may have read

00:22:46.180 | that two systems passed a paragraph comprehension test.

00:22:51.180 | It's really very impressive.

00:22:55.940 | When I came to Google,

00:22:57.980 | we were trying to pass these paragraph comprehension tests.

00:23:01.740 | We aced the first grade test.

00:23:05.820 | Second grade test, we kind of got average performance.

00:23:08.820 | And the third grade test had too much inference.

00:23:12.100 | Already you had to know some common sense knowledge

00:23:15.420 | as it's called and make implications of things

00:23:19.260 | that were in different parts of the paragraph.

00:23:21.380 | And there's too much inference and it really didn't work.

00:23:24.380 | So this is now adult level,

00:23:27.500 | just slightly surpassed average human performance.

00:23:30.660 | But we've seen that once something,

00:23:33.700 | an AI does something at average human levels,

00:23:37.180 | it doesn't take long for it to surpassed average human levels.

00:23:41.180 | I think it'll take longer in language

00:23:42.700 | than it did in sort of simple games like Go,

00:23:44.980 | but it's actually very impressive

00:23:47.140 | that it surpasses now average human performance.

00:23:50.660 | It's used in LSTM, long, short temporal memory.

00:23:55.940 | But if you look at the adult test,

00:23:58.500 | in order to answer these questions,

00:24:00.060 | it has to put together inferences and implications

00:24:03.260 | of several different things in the paragraph

00:24:05.660 | with some common sense knowledge that's not explicitly stated.

00:24:08.980 | So that's, I think, a pretty impressive milestone.

00:24:14.460 | So I've been developing, I've got a team of about 45 people,

00:24:19.300 | and we've been developing this hierarchical model.

00:24:24.060 | We don't use Markov models

00:24:26.060 | 'cause we can use deep learning for each module.

00:24:30.220 | And so we create an embedding for each word

00:24:32.340 | and we create an embedding for each sentence.

00:24:35.180 | This, we have a, I can talk about it

00:24:38.340 | 'cause we have a published paper on it.

00:24:40.420 | It can take into consideration context.

00:24:43.780 | If you use Smart Reply on, if you use Gmail on your phone,

00:24:49.620 | you'll see it gives you three suggestions for responses.

00:24:53.100 | That's called Smart Reply.

00:24:55.500 | They're simple suggestions,

00:24:58.980 | but it has to actually understand

00:25:01.180 | perhaps a complicated email.

00:25:03.460 | And the quality of the suggestions is really quite good,

00:25:06.540 | quite on point.

00:25:08.140 | That's for my team using this kind of hierarchical model.

00:25:12.380 | So instead of Markov models, it uses embeddings

00:25:15.060 | 'cause we can use backpropagation, we might as well use it.

00:25:20.780 | But I think what's missing from deep learning

00:25:25.540 | is this hierarchical aspect of understanding

00:25:29.100 | 'cause the world is hierarchical.

00:25:30.580 | That's why evolution developed a hierarchical brain structure

00:25:35.300 | to understand the natural hierarchy in the world.

00:25:39.180 | And there's several problems with big, deep neural nets.

00:25:44.780 | One is the fact that you really do need a billion examples

00:25:48.100 | and we don't, sometimes we can generate them

00:25:50.260 | as in the case of Go,

00:25:52.500 | or if we have a really good simulator

00:25:54.660 | as in the case of autonomous vehicles,

00:25:56.500 | not quite the case yet in biology.

00:25:59.980 | Very often you don't have a billion examples.

00:26:02.820 | We certainly have billions of examples of language,

00:26:05.420 | but they're not annotated.

00:26:06.660 | And how would you annotate it anyway

00:26:08.220 | with more language that we can't understand

00:26:09.980 | in the first place?

00:26:10.820 | So it's kind of a chicken and an egg problem.

00:26:13.780 | So I believe this hierarchical structure is needed.

00:26:17.100 | Another criticism of deep neural nets

00:26:19.340 | is they don't explain themselves very well.

00:26:21.180 | It's a big black box that gives you

00:26:25.780 | pretty remarkable answers.

00:26:27.380 | I mean, in the case of these games,

00:26:29.500 | Demis described it's playing in both Go and chess

00:26:32.940 | as almost an alien intelligence

00:26:34.580 | 'cause we do things that were shocking to human experts

00:26:37.940 | like sacrificing a queen and a bishop at the same time

00:26:42.020 | or in close succession, which shocked everybody,

00:26:46.100 | but then went on to win,

00:26:47.700 | or early in a Go game,

00:26:48.940 | putting a piece at the corner of the board,

00:26:52.180 | which is kind of crazy to most experts

00:26:55.220 | 'cause you really wanna start controlling territory.

00:26:57.740 | And yet on reflection, that was the brilliant move

00:27:00.940 | that enabled it to win that game.

00:27:02.940 | But it doesn't really explain how it does these things.

00:27:07.620 | So if you have a hierarchy,

00:27:09.940 | it's much better at explaining it

00:27:11.380 | 'cause you could look at the content

00:27:13.180 | of the modules in the hierarchy

00:27:16.300 | and they'll explain what they're doing.

00:27:20.340 | And just to end on the first application

00:27:25.340 | of applying this to health and medicine,

00:27:28.340 | this will get into high gear

00:27:29.660 | and we're gonna really see us break out

00:27:31.500 | of the linear extension to longevity

00:27:35.540 | that we've experienced.

00:27:37.340 | I believe we're only about a decade away

00:27:39.260 | from longevity escape velocity.

00:27:41.380 | We're adding more time than is going by,

00:27:44.580 | not just to infant life expectancy,

00:27:46.340 | but to your remaining life expectancy.

00:27:49.340 | I think if someone is diligent,

00:27:50.740 | they can be there already.

00:27:52.340 | I think I've at longevity escape velocity.

00:27:56.060 | Now, a word on what life expectancy means.

00:28:01.060 | It used to be assumed that not much would happen.

00:28:03.780 | So whatever your life expectancy is

00:28:07.900 | with or without scientific progress,

00:28:09.780 | it really didn't matter.

00:28:11.260 | Now it matters a lot.

00:28:12.500 | So life expectancy really means,

00:28:14.740 | how long would you live?

00:28:17.380 | What's the, in terms of a statistical likelihood,

00:28:20.940 | if there were not continued scientific progress?

00:28:24.340 | But that's a very inaccurate assumption.

00:28:26.260 | The scientific progress is extremely rapid.

00:28:29.220 | I mean, just as an AI in biotech,

00:28:31.460 | there are advances now every week.

00:28:33.860 | It's quite stunning.

00:28:35.980 | Now, you could have a computed life expectancy,

00:28:41.260 | let's say 30 years, 50 years, 70 years from now,

00:28:46.180 | you could still be hit by the proverbial bus tomorrow.

00:28:49.420 | We're working on that with self-driving vehicles.

00:28:52.540 | But we'll get to a point,

00:28:56.540 | I think if you're diligent,

00:28:57.580 | you can be there now in terms of basically advancing

00:29:01.180 | your own statistical life expectancy,

00:29:04.580 | at least to keep pace with the passage of time.

00:29:08.860 | I think it will be there for most of the population,

00:29:12.780 | at least if they're diligent within about a decade.

00:29:17.060 | So if you can hang in there,

00:29:19.260 | we may get to see the remarkable century ahead.

00:29:21.700 | Thank you very much.

00:29:23.340 | (audience applauding)

00:29:26.500 | - A question, please raise your hand,

00:29:29.140 | we'll get you a mic.

00:29:30.180 | - Hi, so you mentioned both neural network models

00:29:37.780 | and symbolic models.

00:29:39.460 | And I was wondering how far have you been thinking

00:29:42.780 | about combining these two approaches,

00:29:45.020 | creating a symbiosis between neural models

00:29:47.940 | and symbolic ones?

00:29:49.140 | - I don't think we wanna use symbolic models

00:29:54.380 | as they've been used.

00:29:56.780 | How many are familiar with the Psych Project?

00:30:00.980 | That was a very diligent effort in Texas

00:30:06.220 | to define all of common sense reasoning.

00:30:09.060 | And it kind of collapsed on itself.

00:30:13.460 | And became impossible to debug

00:30:15.740 | 'cause you'd fix one thing

00:30:17.820 | and it would break three other things.

00:30:19.940 | That complexity ceiling has become typical

00:30:24.620 | of trying to define things through logical rules.

00:30:29.620 | Now it does seem that humans can understand logical rules.

00:30:34.220 | We have logical rules written down for things like law

00:30:37.100 | and game playing and so on.

00:30:41.780 | But you can actually define a connectionist system

00:30:46.580 | to have such a high reliability on a certain type of action

00:30:51.580 | that it looks like it's a symbolic rule

00:30:55.780 | even though it's represented in a connectionist way.

00:31:00.780 | And connection systems can both capture the soft edges

00:31:05.500 | 'cause many things in life are not sharply defined.

00:31:09.580 | They can also generate exceptions.

00:31:12.060 | So you don't wanna sacrifice your queen in chess

00:31:15.340 | except certain situations that might be a good idea.

00:31:18.980 | So you can capture that kind of complexity.

00:31:21.260 | So we do wanna be able to learn from accumulated human wisdom

00:31:27.580 | that looks like it's symbolic.

00:31:29.700 | But I think we'll do it with a connectionist system.

00:31:33.300 | But again, I think that connectionist systems

00:31:37.020 | should develop a sense of hierarchy

00:31:40.420 | and not just be one big massive neural net.

00:31:43.820 | - So I understand how we wanna use the neocortex

00:31:48.420 | to extract useful stuff and commercialize that.

00:31:51.460 | But I'm wondering how our middle brain

00:31:54.820 | and the organs that are below the neocortex

00:31:57.140 | will be useful for turning that into what you wanna do.

00:32:01.780 | - Well, the cerebellum is an interesting case in point.

00:32:05.940 | It actually has more neurons than the neocortex.

00:32:10.220 | And it's used to govern most of our behavior.

00:32:15.140 | Some things, if you write a signature,

00:32:18.180 | that's actually controlled by the cerebellum.

00:32:20.020 | So a simple sequence is stored in the cerebellum.

00:32:25.020 | But there's not any reasoning to it.

00:32:28.340 | It's basically a script.

00:32:29.980 | And most of our movement now

00:32:33.900 | has actually been migrated from the cerebellum

00:32:36.380 | to the neocortex.

00:32:37.940 | Cerebellum is still there.

00:32:39.660 | Some people, entire cerebellum is destroyed through disease.

00:32:45.660 | They still function fairly normally.

00:32:48.340 | Their movement might be a little erratic

00:32:50.860 | 'cause our movement is largely controlled

00:32:53.500 | by the neocortex.

00:32:54.380 | But some of the subtlety is a kind of pre-programmed script.

00:32:58.500 | And so they'll look a little clumsy,

00:33:01.380 | but they actually function okay.

00:33:03.980 | A lot of other areas of the brain

00:33:06.940 | control autonomic functions like breathing.

00:33:09.940 | But our thinking really is controlled by the neocortex.

00:33:14.420 | In terms of mastering intelligence,

00:33:18.780 | I think the neocortex is the brain region we wanna study.

00:33:23.780 | - I'm curious what you think might happen

00:33:29.220 | after the singularity is reached

00:33:31.300 | in terms of this exponential growth of information.

00:33:34.420 | Yeah, do you think it will continue

00:33:37.860 | or will there be a whole paradigm shift?

00:33:40.060 | What do you predict?

00:33:41.220 | - Well, in the singularity's near,

00:33:44.140 | I talk about the atomic limits.

00:33:47.100 | Based on molecular computing as we understand it,

00:33:50.100 | and it can actually go well past 2045

00:33:53.540 | and actually go to trillions of trillions of times

00:33:57.220 | greater computational capacity than we have today.

00:34:01.300 | So I don't see that stopping any time soon

00:34:05.540 | and will go way beyond what we can imagine.

00:34:09.500 | And it becomes an interesting discussion

00:34:13.780 | what the impact on human civilization will be.

00:34:18.780 | So to take a maybe slightly more mundane issue

00:34:22.580 | that comes up is, oh, it's gonna eliminate

00:34:25.060 | most jobs or all jobs.

00:34:28.180 | The point I make is it's not the first time

00:34:30.140 | in human history we've done that.

00:34:32.180 | How many jobs circa 1900 exist today?

00:34:35.500 | And that was the feeling of the Luddites,

00:34:39.700 | which was an actual society that formed in 1800

00:34:42.820 | after the automation of the textile industry in England.

00:34:47.340 | They looked at all these jobs going away

00:34:48.780 | and felt, oh, employment's gonna be

00:34:51.340 | just limited to an elite.

00:34:52.900 | Indeed, those jobs did go away, but new jobs were created.

00:34:57.980 | So if I were a prescient futurist in 1900,

00:35:00.820 | I would say, well, 38% of you work on farms

00:35:04.500 | and 25% work in factories.

00:35:06.260 | That's 2/3 of the working force.

00:35:09.060 | But I predict by 2015, 115 years from now,

00:35:16.260 | it's gonna be 2% on farms and 9% in factories

00:35:19.500 | and everybody would go, oh my God,

00:35:21.260 | we're gonna be out of work.

00:35:22.660 | And I said, well, don't worry.

00:35:23.580 | For all these jobs we eliminate through automation,

00:35:25.700 | we're gonna invent new jobs.

00:35:27.620 | And people say, oh, really, what new jobs?

00:35:29.540 | And I'd say, well, I don't know.

00:35:30.940 | We haven't invented them yet.

00:35:32.380 | That's the political problem.

00:35:34.860 | We can see jobs very clearly going away fairly soon,

00:35:38.060 | like driving a car or a truck.

00:35:39.860 | And the new jobs haven't been invented.

00:35:42.860 | I mean, just look at the last five or six years.

00:35:45.220 | A lot of the increase in employment

00:35:47.980 | has been through mobile app-related

00:35:51.900 | types of ways of making money

00:35:53.380 | that just weren't contemplated even six years ago.

00:35:56.460 | If I really prescient, I would say,

00:35:59.460 | well, you're gonna get jobs creating mobile apps

00:36:01.620 | and websites and doing data analytics

00:36:04.860 | and self-driving cars.

00:36:07.820 | Cars, what's a car?

00:36:09.020 | Nobody would have any idea what I'm talking about.

00:36:12.700 | Now, the new job, some people say,

00:36:16.740 | yeah, we created new jobs, but it's not as many.

00:36:19.020 | Actually, we've gone from 24 million jobs in 1900

00:36:23.220 | to 142 million jobs today,

00:36:25.420 | from 30% of the population to 45% of the population.

00:36:28.900 | The new jobs pay 11 times as much in constant dollars.

00:36:32.940 | And they're more interesting.

00:36:34.260 | I mean, as I talk to people starting out their career now,

00:36:37.820 | they really want a career that gives them some

00:36:40.500 | life definition and purpose and gratification.

00:36:43.620 | We're moving up Maslow's hierarchy.

00:36:46.060 | 100 years ago, you were happy if you had a backbreaking job

00:36:49.340 | that put food on your family's table.

00:36:52.220 | And we couldn't do these new jobs

00:36:55.180 | without enhancing our intelligence.

00:36:57.700 | So we've been doing that, well,

00:36:59.900 | for most of the last 100 years through education.

00:37:03.020 | We've expanded K through 12 in constant dollars tenfold.

00:37:07.020 | We've gone from 38,000 college students in 1870

00:37:11.380 | to 15 million today.

00:37:13.740 | More recently, we have brain extenders.

00:37:17.860 | They're not yet connected directly in our brain,

00:37:20.300 | but they're very close at hand.

00:37:21.900 | When I was here at MIT, I had to take my bicycle

00:37:23.980 | across campus to get to the computer

00:37:26.340 | and show an ID to get in the building.

00:37:28.460 | Now we carry them in our pockets and on our belts.

00:37:33.460 | They're gonna go inside our bodies and brains.

00:37:36.780 | I think that's not a really important distinction.

00:37:41.380 | So we're basically gonna be continuing

00:37:43.100 | to enhance our capability through merging with AI.

00:37:46.780 | And that's the, I think, ultimate answer

00:37:49.740 | to the kind of dystopian view we see in future movies

00:37:54.500 | where it's the AI versus a brave band of humans

00:37:57.180 | for control of humanity.

00:37:58.740 | We don't have one or two AIs in the world today.

00:38:01.340 | We have several billion,

00:38:03.140 | three billion smartphones at last count.

00:38:05.900 | It'll be six billion in just a couple of years

00:38:08.460 | according to the projections.

00:38:10.460 | So we're already deeply integrated with this.

00:38:13.540 | And I think that's gonna continue.

00:38:16.100 | And it's gonna continue to do things

00:38:18.100 | which you can't even imagine today.

00:38:19.860 | Just as we are doing today things we couldn't imagine

00:38:23.260 | even 20 years ago.

00:38:24.580 | - You showed many graphs that go through exponential growth

00:38:29.700 | but I haven't seen one that isn't.

00:38:31.140 | So I would be very interested in hearing--

00:38:33.660 | - You haven't seen one that, what?

00:38:35.060 | - That is not exponential.

00:38:36.820 | So tell me about regions that you've investigated

00:38:40.420 | that have not seen exponential growth

00:38:43.100 | and why do you think that's the case?

00:38:45.980 | - Well, price performance and capacity

00:38:49.180 | of information technology invariably follows exponential.

00:38:53.580 | When it impacts human society it can be linear.

00:38:58.500 | So for example, the growth of democracy

00:39:00.860 | has been linear but still pretty steady.

00:39:06.060 | You could count the number of democracies

00:39:07.660 | on the fingers of one hand a century ago.

00:39:10.380 | Two centuries ago you could count the number of democracies

00:39:12.900 | in the world on the fingers of one finger.

00:39:15.980 | Now there are dozens of them

00:39:17.380 | and it's become kind of a consensus

00:39:20.060 | that that's how we should be governed.

00:39:22.460 | So the, and I attribute all this to the growth

00:39:29.060 | in information technology, communication in particular

00:39:32.500 | for progression of social and cultural institutions.

00:39:37.500 | But information technology,

00:39:42.580 | because it ultimately depends on a vanishingly small

00:39:47.460 | energy and material requirement,

00:39:51.580 | grows exponentially and will for a long time.

00:39:54.620 | There was recently a criticism that,

00:39:57.180 | well, chess scores have,

00:39:58.820 | it's actually a remarkably straight linear progression.

00:40:03.820 | So humans, I think it's like 2,800

00:40:08.260 | and it just soared past that in 1997 with Deep Blue

00:40:11.300 | and it's kept going.

00:40:12.580 | And remarkably straight and saying,

00:40:15.620 | well, this is linear, not exponential.

00:40:17.300 | But the chess score is a logarithmic measurement.

00:40:20.940 | So it really is exponential progression.

00:40:26.620 | - So philosophers like to think a lot about

00:40:31.060 | the meaning of things, especially in the 20th century.

00:40:33.420 | So for instance, Martin Heidegger gave a couple of speeches

00:40:37.580 | and lectures on the relationship of human society

00:40:40.340 | to technology and he particularly distinguished

00:40:43.980 | between the mode of thinking which is calculating thinking

00:40:47.540 | and a mode of thinking which is reflective thinking

00:40:49.900 | or meditative thinking.

00:40:52.180 | And he posed this question,

00:40:54.020 | what is the meaning and purpose

00:40:56.340 | of technological development?

00:40:58.140 | And he couldn't find an answer.

00:40:59.820 | He recommended to remain open to what he called,

00:41:03.740 | he called this an openness to the mystery.

00:41:07.100 | I wonder whether you have any thoughts on this.

00:41:09.540 | Is there a meaning of purpose to technological development

00:41:12.980 | and is there a way for us humans to access that meaning?

00:41:17.260 | - Well, we started using technology to shore up weaknesses

00:41:26.460 | in our own capabilities.

00:41:28.860 | So physically, I mean, who here could build this building?

00:41:33.260 | So we've leveraged the power of our muscles with machines.

00:41:38.740 | And we're in fact very bad at doing things

00:41:41.100 | that the simplest computers can do,

00:41:44.140 | like factor numbers or even just multiply

00:41:48.060 | to eight digit numbers.

00:41:52.420 | Computers can do that trivially, we can't do it.

00:41:54.780 | So we originally started using computers

00:41:57.940 | to make up for that weakness.

00:42:00.540 | I think the essence of what I've been writing about

00:42:05.180 | is to master the unique strengths of humanity,

00:42:09.340 | creating loving expressions in poetry and music

00:42:14.220 | and the kinds of things we associate

00:42:17.100 | with the better qualities of humanity with machines.

00:42:20.740 | That's the true promise of AI.

00:42:23.140 | We're not there yet,

00:42:25.460 | but we're making pretty stunning progress.

00:42:28.220 | Just in the last year, there's so many milestones

00:42:31.140 | that are really significant, including in language.

00:42:34.900 | But I think of technology as an expression of humanity.

00:42:39.900 | It's part of who we are.

00:42:43.500 | And the human species is already

00:42:45.620 | a biological technological civilization.

00:42:50.180 | And it's part of who we are.

00:42:51.620 | And AI is part of humans.

00:42:54.180 | So AI is human and it's part

00:42:59.700 | of the technological expression of humanity.

00:43:02.660 | And we use technology to extend our reach.

00:43:06.940 | I couldn't reach that fruit

00:43:07.980 | at that higher branch a thousand years ago.

00:43:09.540 | So we invented a tool to extend our physical reach.

00:43:12.420 | And we now extend our mental reach.

00:43:14.820 | We can access all of human knowledge with a few keystrokes.

00:43:19.380 | And we're gonna make ourselves literally smarter

00:43:25.540 | by merging with AI.

00:43:31.460 | - Hi, first of all, honored to hear you speak here.

00:43:36.460 | So I first read The Singularity

00:43:38.700 | as near nine years ago or so.

00:43:42.340 | And it changed the way I thought entirely.

00:43:45.380 | But something I think it caused me

00:43:48.140 | to over steeply discount was tail risk in geopolitics,

00:43:53.140 | in systems that span the entire globe.

00:44:01.100 | And my concern is that there is obviously

00:44:06.100 | the possibility of tail risk,

00:44:11.020 | existential level events,

00:44:14.900 | swamping all of these trends that are otherwise war proof,

00:44:19.900 | climate proof, you name it.

00:44:22.500 | So my question for you is what steps do you think

00:44:26.860 | we can take in designing engineered systems,

00:44:31.460 | in designing social and economic institutions

00:44:35.380 | to kind of minimize our exposure to these tail risks

00:44:39.940 | and survive to make it to a beautiful mind filled future?

00:44:44.940 | - Yeah, well, the world was first introduced

00:44:54.140 | to a human made existential risk

00:44:57.900 | when I was in elementary school,

00:45:00.100 | we would have these civil defense drills,

00:45:02.300 | get under our desk and put our hands behind our head

00:45:04.620 | to protect us from a thermonuclear war.

00:45:07.700 | And it worked, we made it through.

00:45:11.700 | But that was really the first introduction

00:45:15.420 | to an existential risk.

00:45:17.140 | And those weapons are still there, by the way,

00:45:21.100 | and they're still on a hair trigger.

00:45:23.980 | And they don't get that much attention.

00:45:26.940 | There's been a lot of discussion,

00:45:28.660 | much of which I've been in the forefront of initiating

00:45:32.140 | on the existential risks of what's sometimes referred to

00:45:35.420 | as GNR, G for genetics, which is biotechnology,

00:45:39.180 | N for nanotechnology, and gray goo, robotics, which is AI.

00:45:43.620 | And I've been accused of being an optimist.

00:45:48.540 | And I think you have to be an optimist

00:45:49.940 | to be an entrepreneur.

00:45:51.180 | If you knew all the problems you were gonna encounter,

00:45:53.500 | you'd never start any project.

00:45:55.460 | But I've written a lot about the downsides.

00:46:00.460 | I remain optimistic.

00:46:01.980 | There are specific paradigms, they're not foolproof,

00:46:06.740 | that we can follow to keep these technologies safe.

00:46:09.740 | So for example, over 40 years ago,

00:46:13.940 | some visionaries recognized the revolutionary potential,

00:46:18.140 | both for promise and peril, of biotechnology.

00:46:22.500 | Neither the promise nor peril was feasible 40 years ago.

00:46:25.900 | But they had a conference

00:46:28.340 | at the Asilomar Conference Center in California

00:46:30.900 | to develop both professional ethics and strategies

00:46:36.700 | to keep biotechnology safe.

00:46:38.820 | And they've been known as the Asilomar Guidelines.

00:46:41.300 | They've been refined through successive Asilomar conferences.

00:46:46.100 | Much of that's baked into law.

00:46:48.380 | And in my opinion, it's worked quite well.

00:46:51.180 | We're now, as I mentioned, getting profound benefit.

00:46:54.140 | It's a trickle today, it'll be a flood over the next decade.

00:46:58.300 | And the number of people who have been harmed,

00:47:00.740 | either through intentional or accidental abuse

00:47:03.420 | of biotechnology, so far is zero.

00:47:06.100 | Actually, I take that back.

00:47:07.180 | There was one boy who died in gene therapy trials

00:47:11.060 | about 12 years ago.

00:47:12.420 | And there was congressional hearings

00:47:13.860 | and they canceled all research for gene therapy

00:47:18.820 | for a number of years.

00:47:20.700 | You could do an interesting master's thesis

00:47:22.980 | and demonstrate that 300,000 people died

00:47:25.580 | as a result of that delay, but you can't name them.

00:47:28.420 | They can't go on CNN, so we don't know who they are.

00:47:30.740 | But that has to do with the balancing of risk.

00:47:34.340 | But in large measure, virtually no one has been hurt

00:47:38.260 | by biotechnology.

00:47:39.500 | Now, that doesn't mean you can cross it off our list.

00:47:41.780 | Okay, we took care of that one

00:47:43.940 | because the technology keeps getting more sophisticated.

00:47:46.660 | CRISPR's a great opportunity.

00:47:49.300 | There's hundreds of trials of CRISPR technologies

00:47:53.260 | to overcome disease, but it could be abused.

00:47:56.500 | You can easily describe scenarios,

00:47:58.860 | so we have to keep reinventing it.

00:48:00.580 | January, we had our first Asilomar Conference on AI ethics.

00:48:06.020 | And so I think this is a good paradigm.

00:48:08.100 | It's not foolproof.

00:48:09.540 | I think the best way we can assure a democratic future

00:48:18.660 | that includes our ideas of liberty

00:48:20.260 | is to practice that in the world today

00:48:22.580 | 'cause the future world of the singularity,

00:48:24.820 | which is a merger of biological

00:48:26.940 | and non-biological intelligence,

00:48:28.660 | is not gonna come from Mars.

00:48:31.100 | I mean, it's gonna emerge from our society today.

00:48:34.460 | So if we practice these ideals today,

00:48:36.340 | it's gonna have a higher chance of us practicing them

00:48:39.580 | as we get more enhanced with technology.

00:48:41.980 | That doesn't sound like a foolproof solution.

00:48:45.140 | It isn't, but I think that's the best approach.

00:48:48.460 | In terms of technological solutions,

00:48:50.740 | I mean, AI is the most daunting.

00:48:52.500 | You can imagine there are technical solutions

00:48:56.100 | to biotechnology and nanotechnology.

00:49:00.260 | There's really no subroutine you can put

00:49:02.020 | in your AI software that will assure that it remains safe.

00:49:07.020 | Intelligence is inherently not controllable.

00:49:10.940 | If there's some AI that's much smarter than you

00:49:13.060 | that's out for your destruction,

00:49:15.180 | the best way to deal with that

00:49:16.380 | is not to get in that situation in the first place.

00:49:19.700 | If you are in that situation,

00:49:22.380 | find some AI that will be on your side.

00:49:24.700 | But basically, it's going to,

00:49:29.100 | I believe we have been headed through technology

00:49:33.220 | to a better reality.

00:49:35.660 | I go around the world

00:49:37.580 | and people really think things are getting worse.

00:49:39.940 | And I think that's 'cause our information

00:49:42.380 | about what's wrong with the world

00:49:43.580 | is getting exponentially better.

00:49:46.060 | They say, "Oh, this is the most peaceful time

00:49:47.620 | "in human history."

00:49:48.460 | And people say, "What are you, crazy?

00:49:49.980 | "Didn't you hear about the event yesterday and last week?"

00:49:52.860 | Well, 100 years ago, there could be a battle

00:49:56.540 | that wiped out the next village

00:49:57.780 | and you wouldn't even hear about it for months.

00:50:00.100 | I have all these graphs on education and literacy

00:50:04.980 | has gone from like 10% to 90% over a century

00:50:09.340 | and health, wealth,

00:50:15.340 | poverty has declined 95% in Asia over the last 25 years,

00:50:20.340 | it's documented by the World Bank.

00:50:22.660 | All these trends are very smoothly getting better

00:50:24.820 | and everybody thinks things are getting worse.

00:50:27.620 | But you're right, like on violence,

00:50:30.380 | that curve could be quite disrupted

00:50:34.100 | if there's an existential event.

00:50:36.660 | As I say, I'm optimistic,

00:50:39.140 | but I think that is something that we need to deal with

00:50:43.300 | and a lot of it is not technological,

00:50:45.020 | it's dealing with our social, cultural institutions.

00:50:48.540 | - So you mentioned also exponential growth of software

00:50:54.620 | and ideas, I guess, related to software.

00:50:56.540 | So one of the reasons for which you said

00:50:59.420 | that information technology costs is exponential

00:51:02.420 | is because of fundamental properties of matter and energy.

00:51:05.900 | But in the case of ideas,

00:51:07.420 | why would it have to be exponential?

00:51:10.020 | - Well, a lot of ideas produce exponential,

00:51:12.900 | exponential gains.

00:51:14.700 | They don't increase performance linearly.

00:51:18.820 | There was actually a study during the Obama administration

00:51:22.020 | by the Scientific Advisory Board on assessing this question,

00:51:27.020 | how much gains on 23 classical engineering problems

00:51:31.540 | were gained through hardware improvements

00:51:36.740 | over the last decade and software improvements.

00:51:38.940 | And there's about a thousand to one improvements,

00:51:41.460 | it's about doubling every year from hardware.

00:51:44.100 | There was an average of something like 26,000 to one

00:51:47.580 | through software improvements, algorithmic improvements.

00:51:50.500 | So we do see both, and apparently,

00:51:56.060 | if you come up with an advance,

00:51:57.380 | it doubles the performance or multiplies it by 10.

00:52:01.420 | We see basically exponential growth from each innovation.

00:52:05.060 | So, and we certainly see that in deep learning,

00:52:11.060 | the architectures are getting better

00:52:13.500 | while we also have more data and more computation

00:52:16.220 | and more memory to throw at these algorithms.

00:52:19.660 | - Thank you very much.

00:52:20.500 | Let's give a very big hand.

00:52:21.900 | (audience applauding)

00:52:25.060 | Thank you for being here.

Ray Kurzweil: Future of Intelligence | MIT 6.S099: Artificial General Intelligence (AGI)

Chapters