vtt/episode_001_large.vtt

WEBVTT

00:00.000 --> 00:04.200
 As part of MIT course 6S099, Artificial General Intelligence,

00:04.200 --> 00:06.600
 I've gotten the chance to sit down with Max Tegmark.

00:06.600 --> 00:08.680
 He is a professor here at MIT.

00:08.680 --> 00:11.920
 He's a physicist, spent a large part of his career

00:11.920 --> 00:16.960
 studying the mysteries of our cosmological universe.

00:16.960 --> 00:20.680
 But he's also studied and delved into the beneficial

00:20.680 --> 00:24.000
 possibilities and the existential risks

00:24.000 --> 00:25.800
 of artificial intelligence.

00:25.800 --> 00:29.040
 Amongst many other things, he is the cofounder

00:29.040 --> 00:33.080
 of the Future of Life Institute, author of two books,

00:33.080 --> 00:35.160
 both of which I highly recommend.

00:35.160 --> 00:37.260
 First, Our Mathematical Universe.

00:37.260 --> 00:40.160
 Second is Life 3.0.

00:40.160 --> 00:44.080
 He's truly an out of the box thinker and a fun personality,

00:44.080 --> 00:45.480
 so I really enjoy talking to him.

00:45.480 --> 00:47.980
 If you'd like to see more of these videos in the future,

00:47.980 --> 00:50.640
 please subscribe and also click the little bell icon

00:50.640 --> 00:52.720
 to make sure you don't miss any videos.

00:52.720 --> 00:56.840
 Also, Twitter, LinkedIn, agi.mit.edu

00:56.840 --> 00:59.600
 if you wanna watch other lectures

00:59.600 --> 01:01.080
 or conversations like this one.

01:01.080 --> 01:04.000
 Better yet, go read Max's book, Life 3.0.

01:04.000 --> 01:07.940
 Chapter seven on goals is my favorite.

01:07.940 --> 01:10.480
 It's really where philosophy and engineering come together

01:10.480 --> 01:13.440
 and it opens with a quote by Dostoevsky.

01:14.400 --> 01:17.940
 The mystery of human existence lies not in just staying alive

01:17.940 --> 01:20.520
 but in finding something to live for.

01:20.520 --> 01:23.920
 Lastly, I believe that every failure rewards us

01:23.920 --> 01:26.560
 with an opportunity to learn

01:26.560 --> 01:28.360
 and in that sense, I've been very fortunate

01:28.360 --> 01:30.960
 to fail in so many new and exciting ways

01:31.840 --> 01:34.020
 and this conversation was no different.

01:34.020 --> 01:36.160
 I've learned about something called

01:36.160 --> 01:40.840
 radio frequency interference, RFI, look it up.

01:40.840 --> 01:42.960
 Apparently, music and conversations

01:42.960 --> 01:45.480
 from local radio stations can bleed into the audio

01:45.480 --> 01:47.080
 that you're recording in such a way

01:47.080 --> 01:49.360
 that it almost completely ruins that audio.

01:49.360 --> 01:52.060
 It's an exceptionally difficult sound source to remove.

01:53.240 --> 01:55.520
 So, I've gotten the opportunity to learn

01:55.520 --> 02:00.200
 how to avoid RFI in the future during recording sessions.

02:00.200 --> 02:02.680
 I've also gotten the opportunity to learn

02:02.680 --> 02:06.240
 how to use Adobe Audition and iZotope RX 6

02:06.240 --> 02:11.240
 to do some noise, some audio repair.

02:11.720 --> 02:14.380
 Of course, this is an exceptionally difficult noise

02:14.380 --> 02:15.220
 to remove.

02:15.220 --> 02:16.280
 I am an engineer.

02:16.280 --> 02:18.240
 I'm not an audio engineer.

02:18.240 --> 02:20.180
 Neither is anybody else in our group

02:20.180 --> 02:21.880
 but we did our best.

02:21.880 --> 02:25.040
 Nevertheless, I thank you for your patience

02:25.040 --> 02:27.960
 and I hope you're still able to enjoy this conversation.

02:27.960 --> 02:29.320
 Do you think there's intelligent life

02:29.320 --> 02:31.360
 out there in the universe?

02:31.360 --> 02:33.480
 Let's open up with an easy question.

02:33.480 --> 02:36.240
 I have a minority view here actually.

02:36.240 --> 02:39.440
 When I give public lectures, I often ask for a show of hands

02:39.440 --> 02:42.920
 who thinks there's intelligent life out there somewhere else

02:42.920 --> 02:45.440
 and almost everyone put their hands up

02:45.440 --> 02:47.360
 and when I ask why, they'll be like,

02:47.360 --> 02:50.900
 oh, there's so many galaxies out there, there's gotta be.

02:51.840 --> 02:54.560
 But I'm a numbers nerd, right?

02:54.560 --> 02:56.640
 So when you look more carefully at it,

02:56.640 --> 02:58.040
 it's not so clear at all.

02:59.080 --> 03:00.680
 When we talk about our universe, first of all,

03:00.680 --> 03:03.040
 we don't mean all of space.

03:03.040 --> 03:04.040
 We actually mean, I don't know,

03:04.040 --> 03:05.440
 you can throw me the universe if you want,

03:05.440 --> 03:07.280
 it's behind you there.

03:07.280 --> 03:11.440
 It's, we simply mean the spherical region of space

03:11.440 --> 03:15.360
 from which light has a time to reach us so far

03:15.360 --> 03:17.040
 during the 14.8 billion year,

03:17.040 --> 03:19.320
 13.8 billion years since our Big Bang.

03:19.320 --> 03:22.320
 There's more space here but this is what we call a universe

03:22.320 --> 03:24.040
 because that's all we have access to.

03:24.040 --> 03:25.960
 So is there intelligent life here

03:25.960 --> 03:28.920
 that's gotten to the point of building telescopes

03:28.920 --> 03:29.960
 and computers?

03:31.160 --> 03:34.540
 My guess is no, actually.

03:34.540 --> 03:37.800
 The probability of it happening on any given planet

03:39.240 --> 03:42.620
 is some number we don't know what it is.

03:42.620 --> 03:47.620
 And what we do know is that the number can't be super high

03:48.480 --> 03:50.300
 because there's over a billion Earth like planets

03:50.300 --> 03:52.880
 in the Milky Way galaxy alone,

03:52.880 --> 03:56.280
 many of which are billions of years older than Earth.

03:56.280 --> 04:00.600
 And aside from some UFO believers,

04:00.600 --> 04:01.880
 there isn't much evidence

04:01.880 --> 04:05.600
 that any superduran civilization has come here at all.

04:05.600 --> 04:08.440
 And so that's the famous Fermi paradox, right?

04:08.440 --> 04:10.180
 And then if you work the numbers,

04:10.180 --> 04:13.440
 what you find is that if you have no clue

04:13.440 --> 04:16.880
 what the probability is of getting life on a given planet,

04:16.880 --> 04:19.680
 so it could be 10 to the minus 10, 10 to the minus 20,

04:19.680 --> 04:22.960
 or 10 to the minus two, or any power of 10

04:22.960 --> 04:23.800
 is sort of equally likely

04:23.800 --> 04:25.480
 if you wanna be really open minded,

04:25.480 --> 04:27.600
 that translates into it being equally likely

04:27.600 --> 04:31.800
 that our nearest neighbor is 10 to the 16 meters away,

04:31.800 --> 04:33.880
 10 to the 17 meters away, 10 to the 18.

04:35.400 --> 04:40.400
 By the time you get much less than 10 to the 16 already,

04:41.080 --> 04:45.960
 we pretty much know there is nothing else that close.

04:45.960 --> 04:47.280
 And when you get beyond 10.

04:47.280 --> 04:48.680
 Because they would have discovered us.

04:48.680 --> 04:50.360
 Yeah, they would have been discovered as long ago,

04:50.360 --> 04:51.440
 or if they're really close,

04:51.440 --> 04:53.560
 we would have probably noted some engineering projects

04:53.560 --> 04:54.640
 that they're doing.

04:54.640 --> 04:57.880
 And if it's beyond 10 to the 26 meters,

04:57.880 --> 05:00.000
 that's already outside of here.

05:00.000 --> 05:05.000
 So my guess is actually that we are the only life in here

05:05.800 --> 05:09.040
 that's gotten the point of building advanced tech,

05:09.040 --> 05:10.720
 which I think is very,

05:12.680 --> 05:15.360
 puts a lot of responsibility on our shoulders, not screw up.

05:15.360 --> 05:17.240
 I think people who take for granted

05:17.240 --> 05:20.120
 that it's okay for us to screw up,

05:20.120 --> 05:22.760
 have an accidental nuclear war or go extinct somehow

05:22.760 --> 05:25.960
 because there's a sort of Star Trek like situation out there

05:25.960 --> 05:28.360
 where some other life forms are gonna come and bail us out

05:28.360 --> 05:30.400
 and it doesn't matter as much.

05:30.400 --> 05:33.400
 I think they're leveling us into a false sense of security.

05:33.400 --> 05:35.200
 I think it's much more prudent to say,

05:35.200 --> 05:36.400
 let's be really grateful

05:36.400 --> 05:38.720
 for this amazing opportunity we've had

05:38.720 --> 05:43.720
 and make the best of it just in case it is down to us.

05:44.080 --> 05:45.680
 So from a physics perspective,

05:45.680 --> 05:48.800
 do you think intelligent life,

05:48.800 --> 05:51.360
 so it's unique from a sort of statistical view

05:51.360 --> 05:52.560
 of the size of the universe,

05:52.560 --> 05:55.840
 but from the basic matter of the universe,

05:55.840 --> 05:59.040
 how difficult is it for intelligent life to come about?

05:59.040 --> 06:01.280
 The kind of advanced tech building life

06:03.120 --> 06:05.720
 is implied in your statement that it's really difficult

06:05.720 --> 06:07.640
 to create something like a human species.

06:07.640 --> 06:11.560
 Well, I think what we know is that going from no life

06:11.560 --> 06:15.720
 to having life that can do a level of tech,

06:15.720 --> 06:18.720
 there's some sort of two going beyond that

06:18.720 --> 06:22.200
 than actually settling our whole universe with life.

06:22.200 --> 06:26.560
 There's some major roadblock there,

06:26.560 --> 06:30.880
 which is some great filter as it's sometimes called,

06:30.880 --> 06:33.520
 which is tough to get through.

06:33.520 --> 06:37.160
 It's either that roadblock is either behind us

06:37.160 --> 06:38.720
 or in front of us.

06:38.720 --> 06:41.080
 I'm hoping very much that it's behind us.

06:41.080 --> 06:45.960
 I'm super excited every time we get a new report from NASA

06:45.960 --> 06:48.480
 saying they failed to find any life on Mars.

06:48.480 --> 06:50.080
 I'm like, yes, awesome.

06:50.080 --> 06:51.680
 Because that suggests that the hard part,

06:51.680 --> 06:54.240
 maybe it was getting the first ribosome

06:54.240 --> 06:59.240
 or some very low level kind of stepping stone

06:59.520 --> 07:00.400
 so that we're home free.

07:00.400 --> 07:01.720
 Because if that's true,

07:01.720 --> 07:03.640
 then the future is really only limited

07:03.640 --> 07:05.200
 by our own imagination.

07:05.200 --> 07:07.360
 It would be much suckier if it turns out

07:07.360 --> 07:11.440
 that this level of life is kind of a dime a dozen,

07:11.440 --> 07:12.760
 but maybe there's some other problem.

07:12.760 --> 07:16.160
 Like as soon as a civilization gets advanced technology,

07:16.160 --> 07:17.000
 within a hundred years,

07:17.000 --> 07:20.320
 they get into some stupid fight with themselves and poof.

07:20.320 --> 07:21.760
 That would be a bummer.

07:21.760 --> 07:26.160
 Yeah, so you've explored the mysteries of the universe,

07:26.160 --> 07:29.000
 the cosmological universe, the one that's sitting

07:29.000 --> 07:30.000
 between us today.

07:31.080 --> 07:35.960
 I think you've also begun to explore the other universe,

07:35.960 --> 07:38.000
 which is sort of the mystery,

07:38.000 --> 07:40.960
 the mysterious universe of the mind of intelligence,

07:40.960 --> 07:42.840
 of intelligent life.

07:42.840 --> 07:45.280
 So is there a common thread between your interest

07:45.280 --> 07:48.760
 or the way you think about space and intelligence?

07:48.760 --> 07:51.040
 Oh yeah, when I was a teenager,

07:53.040 --> 07:57.280
 I was already very fascinated by the biggest questions.

07:57.280 --> 08:00.560
 And I felt that the two biggest mysteries of all in science

08:00.560 --> 08:05.000
 were our universe out there and our universe in here.

08:05.000 --> 08:08.120
 So it's quite natural after having spent

08:08.120 --> 08:11.040
 a quarter of a century on my career,

08:11.040 --> 08:12.680
 thinking a lot about this one,

08:12.680 --> 08:14.320
 that I'm now indulging in the luxury

08:14.320 --> 08:15.960
 of doing research on this one.

08:15.960 --> 08:17.720
 It's just so cool.

08:17.720 --> 08:20.120
 I feel the time is ripe now

08:20.120 --> 08:25.120
 for you trans greatly deepening our understanding of this.

08:25.120 --> 08:26.640
 Just start exploring this one.

08:26.640 --> 08:29.560
 Yeah, because I think a lot of people view intelligence

08:29.560 --> 08:33.520
 as something mysterious that can only exist

08:33.520 --> 08:36.120
 in biological organisms like us,

08:36.120 --> 08:37.680
 and therefore dismiss all talk

08:37.680 --> 08:41.160
 about artificial general intelligence as science fiction.

08:41.160 --> 08:43.200
 But from my perspective as a physicist,

08:43.200 --> 08:46.680
 I am a blob of quarks and electrons

08:46.680 --> 08:48.360
 moving around in a certain pattern

08:48.360 --> 08:50.080
 and processing information in certain ways.

08:50.080 --> 08:53.600
 And this is also a blob of quarks and electrons.

08:53.600 --> 08:55.360
 I'm not smarter than the water bottle

08:55.360 --> 08:57.880
 because I'm made of different kinds of quarks.

08:57.880 --> 08:59.640
 I'm made of up quarks and down quarks,

08:59.640 --> 09:01.400
 exact same kind as this.

09:01.400 --> 09:05.080
 There's no secret sauce, I think, in me.

09:05.080 --> 09:08.560
 It's all about the pattern of the information processing.

09:08.560 --> 09:12.240
 And this means that there's no law of physics

09:12.240 --> 09:15.600
 saying that we can't create technology,

09:15.600 --> 09:19.960
 which can help us by being incredibly intelligent

09:19.960 --> 09:21.680
 and help us crack mysteries that we couldn't.

09:21.680 --> 09:23.560
 In other words, I think we've really only seen

09:23.560 --> 09:26.480
 the tip of the intelligence iceberg so far.

09:26.480 --> 09:29.960
 Yeah, so the perceptronium.

09:29.960 --> 09:31.280
 Yeah.

09:31.280 --> 09:33.200
 So you coined this amazing term.

09:33.200 --> 09:35.760
 It's a hypothetical state of matter,

09:35.760 --> 09:38.360
 sort of thinking from a physics perspective,

09:38.360 --> 09:40.080
 what is the kind of matter that can help,

09:40.080 --> 09:42.920
 as you're saying, subjective experience emerge,

09:42.920 --> 09:44.280
 consciousness emerge.

09:44.280 --> 09:46.640
 So how do you think about consciousness

09:46.640 --> 09:48.160
 from this physics perspective?

09:49.960 --> 09:50.800
 Very good question.

09:50.800 --> 09:55.800
 So again, I think many people have underestimated

09:55.800 --> 09:59.120
 our ability to make progress on this

09:59.120 --> 10:01.320
 by convincing themselves it's hopeless

10:01.320 --> 10:05.840
 because somehow we're missing some ingredient that we need.

10:05.840 --> 10:09.560
 There's some new consciousness particle or whatever.

10:09.560 --> 10:12.720
 I happen to think that we're not missing anything

10:12.720 --> 10:16.320
 and that it's not the interesting thing

10:16.320 --> 10:18.560
 about consciousness that gives us

10:18.560 --> 10:21.400
 this amazing subjective experience of colors

10:21.400 --> 10:23.320
 and sounds and emotions.

10:23.320 --> 10:26.320
 It's rather something at the higher level

10:26.320 --> 10:28.800
 about the patterns of information processing.

10:28.800 --> 10:33.160
 And that's why I like to think about this idea

10:33.160 --> 10:34.480
 of perceptronium.

10:34.480 --> 10:36.920
 What does it mean for an arbitrary physical system

10:36.920 --> 10:41.920
 to be conscious in terms of what its particles are doing

10:41.920 --> 10:43.560
 or its information is doing?

10:43.560 --> 10:46.080
 I don't think, I hate carbon chauvinism,

10:46.080 --> 10:47.960
 this attitude you have to be made of carbon atoms

10:47.960 --> 10:50.160
 to be smart or conscious.

10:50.160 --> 10:53.520
 There's something about the information processing

10:53.520 --> 10:55.360
 that this kind of matter performs.

10:55.360 --> 10:57.840
 Yeah, and you can see I have my favorite equations here

10:57.840 --> 11:00.720
 describing various fundamental aspects of the world.

11:00.720 --> 11:02.560
 I feel that I think one day,

11:02.560 --> 11:04.360
 maybe someone who's watching this will come up

11:04.360 --> 11:07.280
 with the equations that information processing

11:07.280 --> 11:08.760
 has to satisfy to be conscious.

11:08.760 --> 11:11.800
 I'm quite convinced there is big discovery

11:11.800 --> 11:15.400
 to be made there because let's face it,

11:15.400 --> 11:18.720
 we know that so many things are made up of information.

11:18.720 --> 11:21.960
 We know that some information processing is conscious

11:21.960 --> 11:25.520
 because we are conscious.

11:25.520 --> 11:27.600
 But we also know that a lot of information processing

11:27.600 --> 11:28.440
 is not conscious.

11:28.440 --> 11:30.040
 Like most of the information processing happening

11:30.040 --> 11:32.680
 in your brain right now is not conscious.

11:32.680 --> 11:36.040
 There are like 10 megabytes per second coming in

11:36.040 --> 11:38.080
 even just through your visual system.

11:38.080 --> 11:40.480
 You're not conscious about your heartbeat regulation

11:40.480 --> 11:42.120
 or most things.

11:42.120 --> 11:45.680
 Even if I just ask you to like read what it says here,

11:45.680 --> 11:48.040
 you look at it and then, oh, now you know what it said.

11:48.040 --> 11:51.560
 But you're not aware of how the computation actually happened.

11:51.560 --> 11:53.680
 Your consciousness is like the CEO

11:53.680 --> 11:56.680
 that got an email at the end with the final answer.

11:56.680 --> 12:01.000
 So what is it that makes a difference?

12:01.000 --> 12:05.120
 I think that's both a great science mystery.

12:05.120 --> 12:07.080
 We're actually studying it a little bit in my lab here

12:07.080 --> 12:10.920
 at MIT, but I also think it's just a really urgent question

12:10.920 --> 12:12.080
 to answer.

12:12.080 --> 12:14.880
 For starters, I mean, if you're an emergency room doctor

12:14.880 --> 12:17.160
 and you have an unresponsive patient coming in,

12:17.160 --> 12:19.600
 wouldn't it be great if in addition to having

12:22.360 --> 12:25.320
 a CT scanner, you had a consciousness scanner

12:25.320 --> 12:27.920
 that could figure out whether this person

12:27.920 --> 12:30.960
 is actually having locked in syndrome

12:30.960 --> 12:32.440
 or is actually comatose.

12:33.360 --> 12:37.000
 And in the future, imagine if we build robots

12:37.000 --> 12:41.480
 or the machine that we can have really good conversations

12:41.480 --> 12:44.840
 with, which I think is very likely to happen.

12:44.840 --> 12:47.760
 Wouldn't you want to know if your home helper robot

12:47.760 --> 12:51.320
 is actually experiencing anything or just like a zombie,

12:51.320 --> 12:53.520
 I mean, would you prefer it?

12:53.520 --> 12:54.360
 What would you prefer?

12:54.360 --> 12:56.200
 Would you prefer that it's actually unconscious

12:56.200 --> 12:58.560
 so that you don't have to feel guilty about switching it off

12:58.560 --> 13:02.120
 or giving boring chores or what would you prefer?

13:02.120 --> 13:06.520
 Well, certainly we would prefer,

13:06.520 --> 13:08.960
 I would prefer the appearance of consciousness.

13:08.960 --> 13:11.720
 But the question is whether the appearance of consciousness

13:11.720 --> 13:15.040
 is different than consciousness itself.

13:15.040 --> 13:18.200
 And sort of to ask that as a question,

13:18.200 --> 13:21.760
 do you think we need to understand what consciousness is,

13:21.760 --> 13:23.520
 solve the hard problem of consciousness

13:23.520 --> 13:28.240
 in order to build something like an AGI system?

13:28.240 --> 13:30.440
 No, I don't think that.

13:30.440 --> 13:34.520
 And I think we will probably be able to build things

13:34.520 --> 13:36.080
 even if we don't answer that question.

13:36.080 --> 13:37.720
 But if we want to make sure that what happens

13:37.720 --> 13:40.960
 is a good thing, we better solve it first.

13:40.960 --> 13:44.960
 So it's a wonderful controversy you're raising there

13:44.960 --> 13:47.960
 where you have basically three points of view

13:47.960 --> 13:48.800
 about the hard problem.

13:48.800 --> 13:52.800
 So there are two different points of view.

13:52.800 --> 13:55.160
 They both conclude that the hard problem of consciousness

13:55.160 --> 13:56.840
 is BS.

13:56.840 --> 13:59.320
 On one hand, you have some people like Daniel Dennett

13:59.320 --> 14:01.480
 who say that consciousness is just BS

14:01.480 --> 14:05.000
 because consciousness is the same thing as intelligence.

14:05.000 --> 14:06.440
 There's no difference.

14:06.440 --> 14:11.080
 So anything which acts conscious is conscious,

14:11.080 --> 14:13.480
 just like we are.

14:13.480 --> 14:15.960
 And then there are also a lot of people,

14:15.960 --> 14:18.400
 including many top AI researchers I know,

14:18.400 --> 14:19.920
 who say, oh, consciousness is just bullshit

14:19.920 --> 14:22.760
 because, of course, machines can never be conscious.

14:22.760 --> 14:24.520
 They're always going to be zombies.

14:24.520 --> 14:27.880
 You never have to feel guilty about how you treat them.

14:27.880 --> 14:30.880
 And then there's a third group of people,

14:30.880 --> 14:34.920
 including Giulio Tononi, for example,

14:34.920 --> 14:37.440
 and Krzysztof Koch and a number of others.

14:37.440 --> 14:39.520
 I would put myself also in this middle camp

14:39.520 --> 14:41.880
 who say that actually some information processing

14:41.880 --> 14:44.160
 is conscious and some is not.

14:44.160 --> 14:46.960
 So let's find the equation which can be used

14:46.960 --> 14:49.080
 to determine which it is.

14:49.080 --> 14:52.040
 And I think we've just been a little bit lazy,

14:52.040 --> 14:54.960
 kind of running away from this problem for a long time.

14:54.960 --> 14:57.840
 It's been almost taboo to even mention the C word

14:57.840 --> 15:00.520
 in a lot of circles because,

15:00.520 --> 15:03.520
 but we should stop making excuses.

15:03.520 --> 15:07.920
 This is a science question and there are ways

15:07.920 --> 15:11.960
 we can even test any theory that makes predictions for this.

15:11.960 --> 15:13.640
 And coming back to this helper robot,

15:13.640 --> 15:16.080
 I mean, so you said you'd want your helper robot

15:16.080 --> 15:18.160
 to certainly act conscious and treat you,

15:18.160 --> 15:20.880
 like have conversations with you and stuff.

15:20.880 --> 15:21.720
 I think so.

15:21.720 --> 15:22.560
 But wouldn't you, would you feel,

15:22.560 --> 15:23.920
 would you feel a little bit creeped out

15:23.920 --> 15:27.680
 if you realized that it was just a glossed up tape recorder,

15:27.680 --> 15:31.560
 you know, that was just zombie and was a faking emotion?

15:31.560 --> 15:34.560
 Would you prefer that it actually had an experience

15:34.560 --> 15:37.000
 or would you prefer that it's actually

15:37.000 --> 15:39.120
 not experiencing anything so you feel,

15:39.120 --> 15:42.200
 you don't have to feel guilty about what you do to it?

15:42.200 --> 15:45.040
 It's such a difficult question because, you know,

15:45.040 --> 15:47.280
 it's like when you're in a relationship and you say,

15:47.280 --> 15:48.120
 well, I love you.

15:48.120 --> 15:49.760
 And the other person said, I love you back.

15:49.760 --> 15:52.640
 It's like asking, well, do they really love you back

15:52.640 --> 15:55.360
 or are they just saying they love you back?

15:55.360 --> 15:58.120
 Don't you really want them to actually love you?

15:58.120 --> 16:03.120
 It's hard to, it's hard to really know the difference

16:03.520 --> 16:08.520
 between everything seeming like there's consciousness

16:09.000 --> 16:10.640
 present, there's intelligence present,

16:10.640 --> 16:13.840
 there's affection, passion, love,

16:13.840 --> 16:16.200
 and it actually being there.

16:16.200 --> 16:17.720
 I'm not sure, do you have?

16:17.720 --> 16:19.400
 But like, can I ask you a question about this?

16:19.400 --> 16:20.760
 Like to make it a bit more pointed.

16:20.760 --> 16:22.920
 So Mass General Hospital is right across the river, right?

16:22.920 --> 16:23.760
 Yes.

16:23.760 --> 16:26.720
 Suppose you're going in for a medical procedure

16:26.720 --> 16:29.320
 and they're like, you know, for anesthesia,

16:29.320 --> 16:31.000
 what we're going to do is we're going to give you

16:31.000 --> 16:33.160
 muscle relaxants so you won't be able to move

16:33.160 --> 16:35.040
 and you're going to feel excruciating pain

16:35.040 --> 16:35.880
 during the whole surgery,

16:35.880 --> 16:37.600
 but you won't be able to do anything about it.

16:37.600 --> 16:39.200
 But then we're going to give you this drug

16:39.200 --> 16:40.760
 that erases your memory of it.

16:41.960 --> 16:43.440
 Would you be cool about that?

16:44.960 --> 16:47.600
 What's the difference that you're conscious about it

16:48.600 --> 16:51.640
 or not if there's no behavioral change, right?

16:51.640 --> 16:54.520
 Right, that's a really, that's a really clear way to put it.

16:54.520 --> 16:57.400
 That's, yeah, it feels like in that sense,

16:57.400 --> 17:01.080
 experiencing it is a valuable quality.

17:01.080 --> 17:04.800
 So actually being able to have subjective experiences,

17:05.840 --> 17:09.120
 at least in that case, is valuable.

17:09.120 --> 17:11.240
 And I think we humans have a little bit

17:11.240 --> 17:13.600
 of a bad track record also of making

17:13.600 --> 17:15.480
 these self serving arguments

17:15.480 --> 17:18.040
 that other entities aren't conscious.

17:18.040 --> 17:19.160
 You know, people often say,

17:19.160 --> 17:21.800
 oh, these animals can't feel pain.

17:21.800 --> 17:24.040
 It's okay to boil lobsters because we ask them

17:24.040 --> 17:25.960
 if it hurt and they didn't say anything.

17:25.960 --> 17:27.400
 And now there was just a paper out saying,

17:27.400 --> 17:29.320
 lobsters do feel pain when you boil them

17:29.320 --> 17:31.040
 and they're banning it in Switzerland.

17:31.040 --> 17:33.560
 And we did this with slaves too often and said,

17:33.560 --> 17:34.680
 oh, they don't mind.

17:36.240 --> 17:39.480
 They don't maybe aren't conscious

17:39.480 --> 17:41.160
 or women don't have souls or whatever.

17:41.160 --> 17:43.200
 So I'm a little bit nervous when I hear people

17:43.200 --> 17:46.360
 just take as an axiom that machines

17:46.360 --> 17:48.960
 can't have experience ever.

17:48.960 --> 17:51.560
 I think this is just a really fascinating science question

17:51.560 --> 17:52.400
 is what it is.

17:52.400 --> 17:54.720
 Let's research it and try to figure out

17:54.720 --> 17:56.000
 what it is that makes the difference

17:56.000 --> 17:58.880
 between unconscious intelligent behavior

17:58.880 --> 18:01.120
 and conscious intelligent behavior.

18:01.120 --> 18:04.680
 So in terms of, so if you think of a Boston Dynamics

18:04.680 --> 18:07.680
 human or robot being sort of with a broom

18:07.680 --> 18:11.920
 being pushed around, it starts pushing

18:11.920 --> 18:13.320
 on a consciousness question.

18:13.320 --> 18:17.040
 So let me ask, do you think an AGI system

18:17.040 --> 18:19.720
 like a few neuroscientists believe

18:19.720 --> 18:22.320
 needs to have a physical embodiment?

18:22.320 --> 18:25.720
 Needs to have a body or something like a body?

18:25.720 --> 18:28.280
 No, I don't think so.

18:28.280 --> 18:30.560
 You mean to have a conscious experience?

18:30.560 --> 18:31.640
 To have consciousness.

18:33.160 --> 18:36.080
 I do think it helps a lot to have a physical embodiment

18:36.080 --> 18:38.440
 to learn the kind of things about the world

18:38.440 --> 18:41.480
 that are important to us humans, for sure.

18:42.560 --> 18:45.600
 But I don't think the physical embodiment

18:45.600 --> 18:47.120
 is necessary after you've learned it

18:47.120 --> 18:48.760
 to just have the experience.

18:48.760 --> 18:51.400
 Think about when you're dreaming, right?

18:51.400 --> 18:52.600
 Your eyes are closed.

18:52.600 --> 18:54.240
 You're not getting any sensory input.

18:54.240 --> 18:55.960
 You're not behaving or moving in any way

18:55.960 --> 18:58.160
 but there's still an experience there, right?

18:59.720 --> 19:01.400
 And so clearly the experience that you have

19:01.400 --> 19:03.320
 when you see something cool in your dreams

19:03.320 --> 19:04.800
 isn't coming from your eyes.

19:04.800 --> 19:08.640
 It's just the information processing itself in your brain

19:08.640 --> 19:10.920
 which is that experience, right?

19:10.920 --> 19:13.640
 But if I put it another way, I'll say

19:13.640 --> 19:15.120
 because it comes from neuroscience

19:15.120 --> 19:18.280
 is the reason you want to have a body and a physical

19:18.280 --> 19:23.280
 something like a physical, you know, a physical system

19:23.920 --> 19:27.040
 is because you want to be able to preserve something.

19:27.040 --> 19:30.840
 In order to have a self, you could argue,

19:30.840 --> 19:35.840
 would you need to have some kind of embodiment of self

19:36.400 --> 19:37.960
 to want to preserve?

19:38.920 --> 19:42.400
 Well, now we're getting a little bit anthropomorphic

19:42.400 --> 19:45.200
 into anthropomorphizing things.

19:45.200 --> 19:47.280
 Maybe talking about self preservation instincts.

19:47.280 --> 19:50.560
 I mean, we are evolved organisms, right?

19:50.560 --> 19:53.520
 So Darwinian evolution endowed us

19:53.520 --> 19:57.120
 and other evolved organism with a self preservation instinct

19:57.120 --> 20:00.560
 because those that didn't have those self preservation genes

20:00.560 --> 20:02.960
 got cleaned out of the gene pool, right?

20:02.960 --> 20:06.880
 But if you build an artificial general intelligence

20:06.880 --> 20:10.040
 the mind space that you can design is much, much larger

20:10.040 --> 20:14.440
 than just a specific subset of minds that can evolve.

20:14.440 --> 20:17.280
 So an AGI mind doesn't necessarily have

20:17.280 --> 20:19.880
 to have any self preservation instinct.

20:19.880 --> 20:21.600
 It also doesn't necessarily have to be

20:21.600 --> 20:24.040
 so individualistic as us.

20:24.040 --> 20:26.080
 Like, imagine if you could just, first of all,

20:26.080 --> 20:27.960
 or we are also very afraid of death.

20:27.960 --> 20:29.920
 You know, I suppose you could back yourself up

20:29.920 --> 20:32.000
 every five minutes and then your airplane

20:32.000 --> 20:32.840
 is about to crash.

20:32.840 --> 20:36.680
 You're like, shucks, I'm gonna lose the last five minutes

20:36.680 --> 20:39.520
 of experiences since my last cloud backup, dang.

20:39.520 --> 20:41.520
 You know, it's not as big a deal.

20:41.520 --> 20:45.680
 Or if we could just copy experiences between our minds

20:45.680 --> 20:47.640
 easily like we, which we could easily do

20:47.640 --> 20:50.360
 if we were silicon based, right?

20:50.360 --> 20:54.040
 Then maybe we would feel a little bit more

20:54.040 --> 20:56.560
 like a hive mind actually, that maybe it's the,

20:56.560 --> 20:59.960
 so I don't think we should take for granted at all

20:59.960 --> 21:03.000
 that AGI will have to have any of those sort of

21:04.880 --> 21:07.360
 competitive as alpha male instincts.

21:07.360 --> 21:10.160
 On the other hand, you know, this is really interesting

21:10.160 --> 21:13.840
 because I think some people go too far and say,

21:13.840 --> 21:16.680
 of course we don't have to have any concerns either

21:16.680 --> 21:20.800
 that advanced AI will have those instincts

21:20.800 --> 21:22.680
 because we can build anything we want.

21:22.680 --> 21:26.280
 That there's a very nice set of arguments going back

21:26.280 --> 21:28.560
 to Steve Omohundro and Nick Bostrom and others

21:28.560 --> 21:32.280
 just pointing out that when we build machines,

21:32.280 --> 21:34.680
 we normally build them with some kind of goal, you know,

21:34.680 --> 21:38.520
 win this chess game, drive this car safely or whatever.

21:38.520 --> 21:40.960
 And as soon as you put in a goal into machine,

21:40.960 --> 21:42.760
 especially if it's kind of open ended goal

21:42.760 --> 21:44.640
 and the machine is very intelligent,

21:44.640 --> 21:47.000
 it'll break that down into a bunch of sub goals.

21:48.280 --> 21:51.280
 And one of those goals will almost always

21:51.280 --> 21:54.200
 be self preservation because if it breaks or dies

21:54.200 --> 21:56.120
 in the process, it's not gonna accomplish the goal, right?

21:56.120 --> 21:58.040
 Like suppose you just build a little,

21:58.040 --> 22:01.000
 you have a little robot and you tell it to go down

22:01.000 --> 22:04.040
 the store market here and get you some food,

22:04.040 --> 22:06.200
 make you cook an Italian dinner, you know,

22:06.200 --> 22:08.400
 and then someone mugs it and tries to break it

22:08.400 --> 22:09.480
 on the way.

22:09.480 --> 22:12.920
 That robot has an incentive to not get destroyed

22:12.920 --> 22:14.720
 and defend itself or run away,

22:14.720 --> 22:17.720
 because otherwise it's gonna fail in cooking your dinner.

22:17.720 --> 22:19.560
 It's not afraid of death,

22:19.560 --> 22:22.960
 but it really wants to complete the dinner cooking goal.

22:22.960 --> 22:25.040
 So it will have a self preservation instinct.

22:25.040 --> 22:27.920
 Continue being a functional agent somehow.

22:27.920 --> 22:32.920
 And similarly, if you give any kind of more ambitious goal

22:33.720 --> 22:37.000
 to an AGI, it's very likely they wanna acquire

22:37.000 --> 22:39.840
 more resources so it can do that better.

22:39.840 --> 22:42.720
 And it's exactly from those sort of sub goals

22:42.720 --> 22:43.800
 that we might not have intended

22:43.800 --> 22:47.160
 that some of the concerns about AGI safety come.

22:47.160 --> 22:50.600
 You give it some goal that seems completely harmless.

22:50.600 --> 22:53.360
 And then before you realize it,

22:53.360 --> 22:55.480
 it's also trying to do these other things

22:55.480 --> 22:56.920
 which you didn't want it to do.

22:56.920 --> 22:59.160
 And it's maybe smarter than us.

22:59.160 --> 23:01.000
 So it's fascinating.

23:01.000 --> 23:05.680
 And let me pause just because I am in a very kind

23:05.680 --> 23:08.720
 of human centric way, see fear of death

23:08.720 --> 23:11.840
 as a valuable motivator.

23:11.840 --> 23:16.440
 So you don't think, you think that's an artifact

23:16.440 --> 23:19.120
 of evolution, so that's the kind of mind space

23:19.120 --> 23:22.120
 evolution created that we're sort of almost obsessed

23:22.120 --> 23:24.400
 about self preservation, some kind of genetic flow.

23:24.400 --> 23:29.400
 You don't think that's necessary to be afraid of death.

23:29.480 --> 23:32.920
 So not just a kind of sub goal of self preservation

23:32.920 --> 23:34.920
 just so you can keep doing the thing,

23:34.920 --> 23:38.720
 but more fundamentally sort of have the finite thing

23:38.720 --> 23:43.080
 like this ends for you at some point.

23:43.080 --> 23:44.160
 Interesting.

23:44.160 --> 23:47.440
 Do I think it's necessary for what precisely?

23:47.440 --> 23:50.920
 For intelligence, but also for consciousness.

23:50.920 --> 23:55.040
 So for those, for both, do you think really

23:55.040 --> 23:59.120
 like a finite death and the fear of it is important?

23:59.120 --> 24:04.120
 So before I can answer, before we can agree

24:05.160 --> 24:06.960
 on whether it's necessary for intelligence

24:06.960 --> 24:08.360
 or for consciousness, we should be clear

24:08.360 --> 24:09.800
 on how we define those two words.

24:09.800 --> 24:11.960
 Cause a lot of really smart people define them

24:11.960 --> 24:13.320
 in very different ways.

24:13.320 --> 24:17.080
 I was on this panel with AI experts

24:17.080 --> 24:20.080
 and they couldn't agree on how to define intelligence even.

24:20.080 --> 24:22.000
 So I define intelligence simply

24:22.000 --> 24:24.760
 as the ability to accomplish complex goals.

24:25.640 --> 24:27.280
 I like your broad definition, because again

24:27.280 --> 24:29.040
 I don't want to be a carbon chauvinist.

24:29.040 --> 24:30.400
 Right.

24:30.400 --> 24:34.600
 And in that case, no, certainly

24:34.600 --> 24:36.480
 it doesn't require fear of death.

24:36.480 --> 24:40.120
 I would say alpha go, alpha zero is quite intelligent.

24:40.120 --> 24:43.080
 I don't think alpha zero has any fear of being turned off

24:43.080 --> 24:46.320
 because it doesn't understand the concept of it even.

24:46.320 --> 24:48.440
 And similarly consciousness.

24:48.440 --> 24:52.240
 I mean, you could certainly imagine very simple

24:52.240 --> 24:53.920
 kind of experience.

24:53.920 --> 24:57.200
 If certain plants have any kind of experience

24:57.200 --> 24:58.560
 I don't think they're very afraid of dying

24:58.560 --> 25:00.920
 or there's nothing they can do about it anyway much.

25:00.920 --> 25:04.560
 So there wasn't that much value in, but more seriously

25:04.560 --> 25:09.200
 I think if you ask, not just about being conscious

25:09.200 --> 25:14.200
 but maybe having what you would, we might call

25:14.320 --> 25:16.400
 an exciting life where you feel passion

25:16.400 --> 25:21.400
 and really appreciate the things.

25:21.480 --> 25:24.440
 Maybe there somehow, maybe there perhaps it does help

25:24.440 --> 25:27.880
 having a backdrop that, Hey, it's finite.

25:27.880 --> 25:31.200
 No, let's make the most of this, let's live to the fullest.

25:31.200 --> 25:33.800
 So if you knew you were going to live forever

25:34.880 --> 25:37.400
 do you think you would change your?

25:37.400 --> 25:39.560
 Yeah, I mean, in some perspective

25:39.560 --> 25:43.960
 it would be an incredibly boring life living forever.

25:43.960 --> 25:47.360
 So in the sort of loose subjective terms that you said

25:47.360 --> 25:50.480
 of something exciting and something in this

25:50.480 --> 25:53.240
 that other humans would understand, I think is, yeah

25:53.240 --> 25:57.120
 it seems that the finiteness of it is important.

25:57.120 --> 25:59.560
 Well, the good news I have for you then is

25:59.560 --> 26:02.120
 based on what we understand about cosmology

26:02.120 --> 26:05.120
 everything is in our universe is probably

26:05.120 --> 26:07.960
 ultimately probably finite, although.

26:07.960 --> 26:11.560
 Big crunch or big, what's the, the infinite expansion.

26:11.560 --> 26:13.840
 Yeah, we could have a big chill or a big crunch

26:13.840 --> 26:18.440
 or a big rip or that's the big snap or death bubbles.

26:18.440 --> 26:20.040
 All of them are more than a billion years away.

26:20.040 --> 26:24.600
 So we should, we certainly have vastly more time

26:24.600 --> 26:27.920
 than our ancestors thought, but there is still

26:29.160 --> 26:32.360
 it's still pretty hard to squeeze in an infinite number

26:32.360 --> 26:36.560
 of compute cycles, even though there are some loopholes

26:36.560 --> 26:37.720
 that just might be possible.

26:37.720 --> 26:41.960
 But I think, you know, some people like to say

26:41.960 --> 26:44.760
 that you should live as if you're about to

26:44.760 --> 26:46.720
 you're going to die in five years or so.

26:46.720 --> 26:47.960
 And that's sort of optimal.

26:47.960 --> 26:50.560
 Maybe it's a good assumption.

26:50.560 --> 26:54.680
 We should build our civilization as if it's all finite

26:54.680 --> 26:55.680
 to be on the safe side.

26:55.680 --> 26:56.960
 Right, exactly.

26:56.960 --> 26:59.720
 So you mentioned defining intelligence

26:59.720 --> 27:02.960
 as the ability to solve complex goals.

27:02.960 --> 27:05.440
 Where would you draw a line or how would you try

27:05.440 --> 27:08.200
 to define human level intelligence

27:08.200 --> 27:10.680
 and superhuman level intelligence?

27:10.680 --> 27:13.280
 Where is consciousness part of that definition?

27:13.280 --> 27:16.640
 No, consciousness does not come into this definition.

27:16.640 --> 27:20.280
 So, so I think of intelligence as it's a spectrum

27:20.280 --> 27:21.960
 but there are very many different kinds of goals

27:21.960 --> 27:22.800
 you can have.

27:22.800 --> 27:24.000
 You can have a goal to be a good chess player

27:24.000 --> 27:28.520
 a good goal player, a good car driver, a good investor

27:28.520 --> 27:31.160
 good poet, et cetera.

27:31.160 --> 27:34.320
 So intelligence that by its very nature

27:34.320 --> 27:36.680
 isn't something you can measure by this one number

27:36.680 --> 27:37.960
 or some overall goodness.

27:37.960 --> 27:38.800
 No, no.

27:38.800 --> 27:40.320
 There are some people who are more better at this.

27:40.320 --> 27:42.360
 Some people are better than that.

27:42.360 --> 27:45.440
 Right now we have machines that are much better than us

27:45.440 --> 27:49.040
 at some very narrow tasks like multiplying large numbers

27:49.040 --> 27:53.200
 fast, memorizing large databases, playing chess

27:53.200 --> 27:56.280
 playing go and soon driving cars.

27:57.480 --> 28:00.080
 But there's still no machine that can match

28:00.080 --> 28:02.720
 a human child in general intelligence

28:02.720 --> 28:05.720
 but artificial general intelligence, AGI

28:05.720 --> 28:07.880
 the name of your course, of course

28:07.880 --> 28:12.880
 that is by its very definition, the quest

28:13.400 --> 28:16.000
 to build a machine that can do everything

28:16.000 --> 28:17.800
 as well as we can.

28:17.800 --> 28:21.960
 So the old Holy grail of AI from back to its inception

28:21.960 --> 28:25.560
 in the sixties, if that ever happens, of course

28:25.560 --> 28:27.320
 I think it's going to be the biggest transition

28:27.320 --> 28:29.040
 in the history of life on earth

28:29.040 --> 28:33.200
 but it doesn't necessarily have to wait the big impact

28:33.200 --> 28:35.400
 until machines are better than us at knitting

28:35.400 --> 28:39.160
 that the really big change doesn't come exactly

28:39.160 --> 28:41.800
 at the moment they're better than us at everything.

28:41.800 --> 28:44.120
 The really big change comes first

28:44.120 --> 28:45.840
 there are big changes when they start becoming better

28:45.840 --> 28:48.800
 at us at doing most of the jobs that we do

28:48.800 --> 28:51.160
 because that takes away much of the demand

28:51.160 --> 28:53.200
 for human labor.

28:53.200 --> 28:55.640
 And then the really whopping change comes

28:55.640 --> 29:00.640
 when they become better than us at AI research, right?

29:01.040 --> 29:03.760
 Because right now the timescale of AI research

29:03.760 --> 29:08.400
 is limited by the human research and development cycle

29:08.400 --> 29:10.160
 of years typically, you know

29:10.160 --> 29:13.480
 how long does it take from one release of some software

29:13.480 --> 29:15.720
 or iPhone or whatever to the next?

29:15.720 --> 29:20.720
 But once Google can replace 40,000 engineers

29:20.920 --> 29:25.920
 by 40,000 equivalent pieces of software or whatever

29:26.400 --> 29:29.680
 but then there's no reason that has to be years

29:29.680 --> 29:31.840
 it can be in principle much faster

29:31.840 --> 29:36.040
 and the timescale of future progress in AI

29:36.040 --> 29:39.320
 and all of science and technology will be driven

29:39.320 --> 29:40.960
 by machines, not humans.

29:40.960 --> 29:45.960
 So it's this simple point which gives right

29:46.520 --> 29:48.720
 this incredibly fun controversy

29:48.720 --> 29:51.880
 about whether there can be intelligence explosion

29:51.880 --> 29:54.400
 so called singularity as Werner Vinge called it.

29:54.400 --> 29:57.040
 Now the idea is articulated by I.J. Good

29:57.040 --> 29:59.480
 is obviously way back fifties

29:59.480 --> 30:01.040
 but you can see Alan Turing

30:01.040 --> 30:03.640
 and others thought about it even earlier.

30:06.920 --> 30:10.080
 So you asked me what exactly would I define

30:10.080 --> 30:12.800
 human level intelligence, yeah.

30:12.800 --> 30:15.680
 So the glib answer is to say something

30:15.680 --> 30:18.520
 which is better than us at all cognitive tasks

30:18.520 --> 30:21.800
 with a better than any human at all cognitive tasks

30:21.800 --> 30:23.080
 but the really interesting bar

30:23.080 --> 30:25.760
 I think goes a little bit lower than that actually.

30:25.760 --> 30:27.920
 It's when they can, when they're better than us

30:27.920 --> 30:31.760
 at AI programming and general learning

30:31.760 --> 30:35.360
 so that they can if they want to get better

30:35.360 --> 30:37.240
 than us at anything by just studying.

30:37.240 --> 30:40.560
 So they're better is a key word and better is towards

30:40.560 --> 30:44.120
 this kind of spectrum of the complexity of goals

30:44.120 --> 30:45.680
 it's able to accomplish.

30:45.680 --> 30:50.360
 So another way to, and that's certainly

30:50.360 --> 30:53.040
 a very clear definition of human love.

30:53.040 --> 30:55.240
 So there's, it's almost like a sea that's rising

30:55.240 --> 30:56.800
 you can do more and more and more things

30:56.800 --> 30:58.640
 it's a geographic that you show

30:58.640 --> 30:59.880
 it's really nice way to put it.

30:59.880 --> 31:01.560
 So there's some peaks that

31:01.560 --> 31:03.280
 and there's an ocean level elevating

31:03.280 --> 31:04.800
 and you solve more and more problems

31:04.800 --> 31:07.720
 but just kind of to take a pause

31:07.720 --> 31:09.000
 and we took a bunch of questions

31:09.000 --> 31:10.240
 and a lot of social networks

31:10.240 --> 31:11.720
 and a bunch of people asked

31:11.720 --> 31:14.480
 a sort of a slightly different direction

31:14.480 --> 31:19.480
 on creativity and things that perhaps aren't a peak.

31:23.560 --> 31:24.720
 Human beings are flawed

31:24.720 --> 31:28.720
 and perhaps better means having contradiction

31:28.720 --> 31:30.200
 being flawed in some way.

31:30.200 --> 31:34.960
 So let me sort of start easy, first of all.

31:34.960 --> 31:36.600
 So you have a lot of cool equations.

31:36.600 --> 31:39.760
 Let me ask, what's your favorite equation, first of all?

31:39.760 --> 31:42.760
 I know they're all like your children, but like

31:42.760 --> 31:43.680
 which one is that?

31:43.680 --> 31:45.560
 This is the shirt in your equation.

31:45.560 --> 31:48.640
 It's the master key of quantum mechanics

31:48.640 --> 31:49.880
 of the micro world.

31:49.880 --> 31:52.800
 So this equation will protect everything

31:52.800 --> 31:55.840
 to do with atoms, molecules and all the way up.

31:55.840 --> 31:58.560
 Right?

31:58.560 --> 31:59.760
 Yeah, so, okay.

31:59.760 --> 32:02.080
 So quantum mechanics is certainly a beautiful

32:02.080 --> 32:05.160
 mysterious formulation of our world.

32:05.160 --> 32:08.760
 So I'd like to sort of ask you, just as an example

32:08.760 --> 32:12.160
 it perhaps doesn't have the same beauty as physics does

32:12.160 --> 32:16.960
 but in mathematics abstract, the Andrew Wiles

32:16.960 --> 32:19.360
 who proved the Fermat's last theorem.

32:19.360 --> 32:22.040
 So he just saw this recently

32:22.040 --> 32:24.160
 and it kind of caught my eye a little bit.

32:24.160 --> 32:27.960
 This is 358 years after it was conjectured.

32:27.960 --> 32:29.960
 So this is very simple formulation.

32:29.960 --> 32:32.640
 Everybody tried to prove it, everybody failed.

32:32.640 --> 32:34.800
 And so here's this guy comes along

32:34.800 --> 32:38.640
 and eventually proves it and then fails to prove it

32:38.640 --> 32:41.320
 and then proves it again in 94.

32:41.320 --> 32:43.480
 And he said like the moment when everything connected

32:43.480 --> 32:46.040
 into place in an interview said

32:46.040 --> 32:47.880
 it was so indescribably beautiful.

32:47.880 --> 32:51.040
 That moment when you finally realize the connecting piece

32:51.040 --> 32:52.800
 of two conjectures.

32:52.800 --> 32:55.280
 He said, it was so indescribably beautiful.

32:55.280 --> 32:57.040
 It was so simple and so elegant.

32:57.040 --> 32:58.760
 I couldn't understand how I'd missed it.

32:58.760 --> 33:02.080
 And I just stared at it in disbelief for 20 minutes.

33:02.080 --> 33:05.240
 Then during the day, I walked around the department

33:05.240 --> 33:07.880
 and I keep coming back to my desk

33:07.880 --> 33:09.840
 looking to see if it was still there.

33:09.840 --> 33:10.680
 It was still there.

33:10.680 --> 33:11.760
 I couldn't contain myself.

33:11.760 --> 33:12.880
 I was so excited.

33:12.880 --> 33:15.880
 It was the most important moment on my working life.

33:15.880 --> 33:18.960
 Nothing I ever do again will mean as much.

33:18.960 --> 33:20.800
 So that particular moment.

33:20.800 --> 33:24.640
 And it kind of made me think of what would it take?

33:24.640 --> 33:27.960
 And I think we have all been there at small levels.

33:29.480 --> 33:32.240
 Maybe let me ask, have you had a moment like that

33:32.240 --> 33:34.880
 in your life where you just had an idea?

33:34.880 --> 33:37.040
 It's like, wow, yes.

33:40.000 --> 33:42.480
 I wouldn't mention myself in the same breath

33:42.480 --> 33:44.760
 as Andrew Wiles, but I've certainly had a number

33:44.760 --> 33:52.200
 of aha moments when I realized something very cool

33:52.200 --> 33:56.000
 about physics, which has completely made my head explode.

33:56.000 --> 33:58.320
 In fact, some of my favorite discoveries I made later,

33:58.320 --> 34:01.080
 I later realized that they had been discovered earlier

34:01.080 --> 34:03.240
 by someone who sometimes got quite famous for it.

34:03.240 --> 34:05.480
 So it's too late for me to even publish it,

34:05.480 --> 34:07.440
 but that doesn't diminish in any way.

34:07.440 --> 34:09.760
 The emotional experience you have when you realize it,

34:09.760 --> 34:11.320
 like, wow.

34:11.320 --> 34:15.520
 Yeah, so what would it take in that moment, that wow,

34:15.520 --> 34:17.320
 that was yours in that moment?

34:17.320 --> 34:21.440
 So what do you think it takes for an intelligence system,

34:21.440 --> 34:24.520
 an AGI system, an AI system to have a moment like that?

34:25.640 --> 34:26.760
 That's a tricky question

34:26.760 --> 34:29.200
 because there are actually two parts to it, right?

34:29.200 --> 34:33.920
 One of them is, can it accomplish that proof?

34:33.920 --> 34:37.640
 Can it prove that you can never write A to the N

34:37.640 --> 34:42.760
 plus B to the N equals three to that equal Z to the N

34:42.760 --> 34:45.320
 for all integers, et cetera, et cetera,

34:45.320 --> 34:48.720
 when N is bigger than two?

34:48.720 --> 34:51.360
 That's simply a question about intelligence.

34:51.360 --> 34:54.120
 Can you build machines that are that intelligent?

34:54.120 --> 34:57.280
 And I think by the time we get a machine

34:57.280 --> 35:00.840
 that can independently come up with that level of proofs,

35:00.840 --> 35:03.360
 probably quite close to AGI.

35:03.360 --> 35:07.240
 The second question is a question about consciousness.

35:07.240 --> 35:11.760
 When will we, how likely is it that such a machine

35:11.760 --> 35:14.240
 will actually have any experience at all,

35:14.240 --> 35:16.160
 as opposed to just being like a zombie?

35:16.160 --> 35:20.560
 And would we expect it to have some sort of emotional response

35:20.560 --> 35:24.640
 to this or anything at all akin to human emotion

35:24.640 --> 35:28.320
 where when it accomplishes its machine goal,

35:28.320 --> 35:31.920
 it views it as somehow something very positive

35:31.920 --> 35:39.160
 and sublime and deeply meaningful?

35:39.160 --> 35:41.440
 I would certainly hope that if in the future

35:41.440 --> 35:45.120
 we do create machines that are our peers

35:45.120 --> 35:50.160
 or even our descendants, that I would certainly

35:50.160 --> 35:55.480
 hope that they do have this sublime appreciation of life.

35:55.480 --> 35:58.840
 In a way, my absolutely worst nightmare

35:58.840 --> 36:05.760
 would be that at some point in the future,

36:05.760 --> 36:07.400
 the distant future, maybe our cosmos

36:07.400 --> 36:10.600
 is teeming with all this post biological life doing

36:10.600 --> 36:12.880
 all the seemingly cool stuff.

36:12.880 --> 36:16.480
 And maybe the last humans, by the time

36:16.480 --> 36:20.120
 our species eventually fizzles out,

36:20.120 --> 36:21.920
 will be like, well, that's OK because we're

36:21.920 --> 36:23.600
 so proud of our descendants here.

36:23.600 --> 36:26.680
 And look what all the, my worst nightmare

36:26.680 --> 36:30.360
 is that we haven't solved the consciousness problem.

36:30.360 --> 36:32.880
 And we haven't realized that these are all the zombies.

36:32.880 --> 36:36.200
 They're not aware of anything any more than a tape recorder

36:36.200 --> 36:37.840
 has any kind of experience.

36:37.840 --> 36:40.040
 So the whole thing has just become

36:40.040 --> 36:41.520
 a play for empty benches.

36:41.520 --> 36:44.640
 That would be the ultimate zombie apocalypse.

36:44.640 --> 36:47.200
 So I would much rather, in that case,

36:47.200 --> 36:52.240
 that we have these beings which can really

36:52.240 --> 36:57.000
 appreciate how amazing it is.

36:57.000 --> 37:01.080
 And in that picture, what would be the role of creativity?

37:01.080 --> 37:04.960
 A few people ask about creativity.

37:04.960 --> 37:07.080
 When you think about intelligence,

37:07.080 --> 37:09.840
 certainly the story you told at the beginning of your book

37:09.840 --> 37:15.200
 involved creating movies and so on, making money.

37:15.200 --> 37:17.240
 You can make a lot of money in our modern world

37:17.240 --> 37:18.600
 with music and movies.

37:18.600 --> 37:20.880
 So if you are an intelligent system,

37:20.880 --> 37:22.960
 you may want to get good at that.

37:22.960 --> 37:26.280
 But that's not necessarily what I mean by creativity.

37:26.280 --> 37:29.640
 Is it important on that complex goals

37:29.640 --> 37:31.600
 where the sea is rising for there

37:31.600 --> 37:33.800
 to be something creative?

37:33.800 --> 37:37.400
 Or am I being very human centric and thinking creativity

37:37.400 --> 37:41.880
 somehow special relative to intelligence?

37:41.880 --> 37:47.240
 My hunch is that we should think of creativity simply

37:47.240 --> 37:50.760
 as an aspect of intelligence.

37:50.760 --> 37:57.840
 And we have to be very careful with human vanity.

37:57.840 --> 37:59.520
 We have this tendency to very often want

37:59.520 --> 38:01.560
 to say, as soon as machines can do something,

38:01.560 --> 38:03.560
 we try to diminish it and say, oh, but that's

38:03.560 --> 38:05.920
 not real intelligence.

38:05.920 --> 38:08.400
 Isn't it creative or this or that?

38:08.400 --> 38:12.200
 The other thing, if we ask ourselves

38:12.200 --> 38:14.320
 to write down a definition of what we actually mean

38:14.320 --> 38:18.840
 by being creative, what we mean by Andrew Wiles, what he did

38:18.840 --> 38:21.880
 there, for example, don't we often mean that someone takes

38:21.880 --> 38:26.000
 a very unexpected leap?

38:26.000 --> 38:29.680
 It's not like taking 573 and multiplying it

38:29.680 --> 38:33.840
 by 224 by just a step of straightforward cookbook

38:33.840 --> 38:36.520
 like rules, right?

38:36.520 --> 38:39.680
 You can maybe make a connection between two things

38:39.680 --> 38:42.640
 that people had never thought was connected or something

38:42.640 --> 38:44.480
 like that.

38:44.480 --> 38:47.720
 I think this is an aspect of intelligence.

38:47.720 --> 38:53.000
 And this is actually one of the most important aspects of it.

38:53.000 --> 38:55.520
 Maybe the reason we humans tend to be better at it

38:55.520 --> 38:57.840
 than traditional computers is because it's

38:57.840 --> 38:59.640
 something that comes more naturally if you're

38:59.640 --> 39:04.120
 a neural network than if you're a traditional logic gate

39:04.120 --> 39:05.720
 based computer machine.

39:05.720 --> 39:08.640
 We physically have all these connections.

39:08.640 --> 39:13.800
 And you activate here, activate here, activate here.

39:13.800 --> 39:16.560
 Bing.

39:16.560 --> 39:21.040
 My hunch is that if we ever build a machine where you could

39:21.040 --> 39:29.200
 just give it the task, hey, you say, hey, I just realized

39:29.200 --> 39:32.320
 I want to travel around the world instead this month.

39:32.320 --> 39:34.600
 Can you teach my AGI course for me?

39:34.600 --> 39:35.960
 And it's like, OK, I'll do it.

39:35.960 --> 39:37.920
 And it does everything that you would have done

39:37.920 --> 39:39.760
 and improvises and stuff.

39:39.760 --> 39:43.360
 That would, in my mind, involve a lot of creativity.

39:43.360 --> 39:45.680
 Yeah, so it's actually a beautiful way to put it.

39:45.680 --> 39:52.640
 I think we do try to grasp at the definition of intelligence

39:52.640 --> 39:56.360
 is everything we don't understand how to build.

39:56.360 --> 39:59.360
 So we as humans try to find things

39:59.360 --> 40:01.240
 that we have and machines don't have.

40:01.240 --> 40:03.800
 And maybe creativity is just one of the things, one

40:03.800 --> 40:05.480
 of the words we use to describe that.

40:05.480 --> 40:07.200
 That's a really interesting way to put it.

40:07.200 --> 40:09.520
 I don't think we need to be that defensive.

40:09.520 --> 40:11.560
 I don't think anything good comes out of saying,

40:11.560 --> 40:18.080
 well, we're somehow special, you know?

40:18.080 --> 40:21.040
 Contrary wise, there are many examples in history

40:21.040 --> 40:27.840
 of where trying to pretend that we're somehow superior

40:27.840 --> 40:33.120
 to all other intelligent beings has led to pretty bad results,

40:33.120 --> 40:35.960
 right?

40:35.960 --> 40:38.440
 Nazi Germany, they said that they were somehow superior

40:38.440 --> 40:40.080
 to other people.

40:40.080 --> 40:42.440
 Today, we still do a lot of cruelty to animals

40:42.440 --> 40:44.440
 by saying that we're so superior somehow,

40:44.440 --> 40:46.440
 and they can't feel pain.

40:46.440 --> 40:48.480
 Slavery was justified by the same kind

40:48.480 --> 40:52.200
 of just really weak arguments.

40:52.200 --> 40:57.120
 And I don't think if we actually go ahead and build

40:57.120 --> 40:59.440
 artificial general intelligence, it

40:59.440 --> 41:01.360
 can do things better than us, I don't

41:01.360 --> 41:04.080
 think we should try to found our self worth on some sort

41:04.080 --> 41:09.760
 of bogus claims of superiority in terms

41:09.760 --> 41:12.120
 of our intelligence.

41:12.120 --> 41:18.080
 I think we should instead find our calling

41:18.080 --> 41:23.360
 and the meaning of life from the experiences that we have.

41:23.360 --> 41:28.720
 I can have very meaningful experiences

41:28.720 --> 41:32.920
 even if there are other people who are smarter than me.

41:32.920 --> 41:34.400
 When I go to a faculty meeting here,

41:34.400 --> 41:36.520
 and we talk about something, and then I certainly realize,

41:36.520 --> 41:39.080
 oh, boy, he has an old prize, he has an old prize,

41:39.080 --> 41:40.800
 he has an old prize, I don't have one.

41:40.800 --> 41:43.760
 Does that make me enjoy life any less

41:43.760 --> 41:47.560
 or enjoy talking to those people less?

41:47.560 --> 41:49.560
 Of course not.

41:49.560 --> 41:54.160
 And the contrary, I feel very honored and privileged

41:54.160 --> 41:58.760
 to get to interact with other very intelligent beings that

41:58.760 --> 42:00.680
 are better than me at a lot of stuff.

42:00.680 --> 42:02.840
 So I don't think there's any reason why

42:02.840 --> 42:06.080
 we can't have the same approach with intelligent machines.

42:06.080 --> 42:07.320
 That's a really interesting.

42:07.320 --> 42:08.920
 So people don't often think about that.

42:08.920 --> 42:10.600
 They think about when there's going,

42:10.600 --> 42:13.320
 if there's machines that are more intelligent,

42:13.320 --> 42:15.080
 you naturally think that that's not

42:15.080 --> 42:19.080
 going to be a beneficial type of intelligence.

42:19.080 --> 42:23.000
 You don't realize it could be like peers with Nobel prizes

42:23.000 --> 42:25.120
 that would be just fun to talk with,

42:25.120 --> 42:27.560
 and they might be clever about certain topics,

42:27.560 --> 42:32.240
 and you can have fun having a few drinks with them.

42:32.240 --> 42:35.880
 Well, also, another example we can all

42:35.880 --> 42:39.320
 relate to of why it doesn't have to be a terrible thing

42:39.320 --> 42:42.560
 to be in the presence of people who are even smarter than us

42:42.560 --> 42:45.600
 all around is when you and I were both two years old,

42:45.600 --> 42:48.360
 I mean, our parents were much more intelligent than us,

42:48.360 --> 42:49.040
 right?

42:49.040 --> 42:51.960
 Worked out OK, because their goals

42:51.960 --> 42:53.960
 were aligned with our goals.

42:53.960 --> 42:58.680
 And that, I think, is really the number one key issue

42:58.680 --> 43:02.280
 we have to solve if we value align the value alignment

43:02.280 --> 43:03.080
 problem, exactly.

43:03.080 --> 43:06.520
 Because people who see too many Hollywood movies

43:06.520 --> 43:10.000
 with lousy science fiction plot lines,

43:10.000 --> 43:12.200
 they worry about the wrong thing, right?

43:12.200 --> 43:16.320
 They worry about some machine suddenly turning evil.

43:16.320 --> 43:21.480
 It's not malice that is the concern.

43:21.480 --> 43:22.880
 It's competence.

43:22.880 --> 43:27.440
 By definition, intelligent makes you very competent.

43:27.440 --> 43:31.920
 If you have a more intelligent goal playing,

43:31.920 --> 43:33.680
 computer playing is a less intelligent one.

43:33.680 --> 43:36.120
 And when we define intelligence as the ability

43:36.120 --> 43:38.600
 to accomplish goal winning, it's going

43:38.600 --> 43:40.560
 to be the more intelligent one that wins.

43:40.560 --> 43:43.560
 And if you have a human and then you

43:43.560 --> 43:47.720
 have an AGI that's more intelligent in all ways

43:47.720 --> 43:49.520
 and they have different goals, guess who's

43:49.520 --> 43:50.720
 going to get their way, right?

43:50.720 --> 43:57.120
 So I was just reading about this particular rhinoceros species

43:57.120 --> 43:59.200
 that was driven extinct just a few years ago.

43:59.200 --> 44:02.280
 Ellen Bummer is looking at this cute picture of a mommy

44:02.280 --> 44:05.080
 rhinoceros with its child.

44:05.080 --> 44:09.320
 And why did we humans drive it to extinction?

44:09.320 --> 44:12.800
 It wasn't because we were evil rhino haters as a whole.

44:12.800 --> 44:14.920
 It was just because our goals weren't aligned

44:14.920 --> 44:16.000
 with those of the rhinoceros.

44:16.000 --> 44:17.680
 And it didn't work out so well for the rhinoceros

44:17.680 --> 44:19.560
 because we were more intelligent, right?

44:19.560 --> 44:21.240
 So I think it's just so important

44:21.240 --> 44:27.120
 that if we ever do build AGI, before we unleash anything,

44:27.120 --> 44:31.840
 we have to make sure that it learns

44:31.840 --> 44:36.000
 to understand our goals, that it adopts our goals,

44:36.000 --> 44:37.920
 and that it retains those goals.

44:37.920 --> 44:40.520
 So the cool, interesting problem there

44:40.520 --> 44:47.040
 is us as human beings trying to formulate our values.

44:47.040 --> 44:51.360
 So you could think of the United States Constitution as a way

44:51.360 --> 44:56.680
 that people sat down, at the time a bunch of white men,

44:56.680 --> 44:59.680
 which is a good example, I should say.

44:59.680 --> 45:01.480
 They formulated the goals for this country.

45:01.480 --> 45:03.760
 And a lot of people agree that those goals actually

45:03.760 --> 45:05.360
 held up pretty well.

45:05.360 --> 45:07.160
 That's an interesting formulation of values

45:07.160 --> 45:09.440
 and failed miserably in other ways.

45:09.440 --> 45:13.320
 So for the value alignment problem and the solution to it,

45:13.320 --> 45:19.560
 we have to be able to put on paper or in a program

45:19.560 --> 45:20.400
 human values.

45:20.400 --> 45:22.400
 How difficult do you think that is?

45:22.400 --> 45:24.040
 Very.

45:24.040 --> 45:25.880
 But it's so important.

45:25.880 --> 45:28.000
 We really have to give it our best.

45:28.000 --> 45:30.120
 And it's difficult for two separate reasons.

45:30.120 --> 45:33.440
 There's the technical value alignment problem

45:33.440 --> 45:39.120
 of figuring out just how to make machines understand our goals,

45:39.120 --> 45:40.440
 adopt them, and retain them.

45:40.440 --> 45:43.200
 And then there's the separate part of it,

45:43.200 --> 45:44.200
 the philosophical part.

45:44.200 --> 45:45.920
 Whose values anyway?

45:45.920 --> 45:48.320
 And since it's not like we have any great consensus

45:48.320 --> 45:52.040
 on this planet on values, what mechanism should we

45:52.040 --> 45:54.120
 create then to aggregate and decide, OK,

45:54.120 --> 45:56.520
 what's a good compromise?

45:56.520 --> 45:58.440
 That second discussion can't just

45:58.440 --> 46:01.560
 be left to tech nerds like myself.

46:01.560 --> 46:05.720
 And if we refuse to talk about it and then AGI gets built,

46:05.720 --> 46:07.160
 who's going to be actually making

46:07.160 --> 46:08.480
 the decision about whose values?

46:08.480 --> 46:12.080
 It's going to be a bunch of dudes in some tech company.

46:12.080 --> 46:17.240
 And are they necessarily so representative of all

46:17.240 --> 46:19.400
 of humankind that we want to just entrust it to them?

46:19.400 --> 46:23.000
 Are they even uniquely qualified to speak

46:23.000 --> 46:25.240
 to future human happiness just because they're

46:25.240 --> 46:26.480
 good at programming AI?

46:26.480 --> 46:30.200
 I'd much rather have this be a really inclusive conversation.

46:30.200 --> 46:32.560
 But do you think it's possible?

46:32.560 --> 46:37.560
 So you create a beautiful vision that includes the diversity,

46:37.560 --> 46:40.960
 cultural diversity, and various perspectives on discussing

46:40.960 --> 46:43.600
 rights, freedoms, human dignity.

46:43.600 --> 46:46.520
 But how hard is it to come to that consensus?

46:46.520 --> 46:50.400
 Do you think it's certainly a really important thing

46:50.400 --> 46:51.880
 that we should all try to do?

46:51.880 --> 46:54.240
 But do you think it's feasible?

46:54.240 --> 47:00.160
 I think there's no better way to guarantee failure than to

47:00.160 --> 47:02.840
 refuse to talk about it or refuse to try.

47:02.840 --> 47:05.320
 And I also think it's a really bad strategy

47:05.320 --> 47:08.560
 to say, OK, let's first have a discussion for a long time.

47:08.560 --> 47:11.040
 And then once we reach complete consensus,

47:11.040 --> 47:13.360
 then we'll try to load it into some machine.

47:13.360 --> 47:16.560
 No, we shouldn't let perfect be the enemy of good.

47:16.560 --> 47:20.600
 Instead, we should start with the kindergarten ethics

47:20.600 --> 47:22.120
 that pretty much everybody agrees on

47:22.120 --> 47:24.360
 and put that into machines now.

47:24.360 --> 47:25.880
 We're not doing that even.

47:25.880 --> 47:31.000
 Look at anyone who builds this passenger aircraft,

47:31.000 --> 47:33.000
 wants it to never under any circumstances

47:33.000 --> 47:35.600
 fly into a building or a mountain.

47:35.600 --> 47:38.480
 Yet the September 11 hijackers were able to do that.

47:38.480 --> 47:41.800
 And even more embarrassingly, Andreas Lubitz,

47:41.800 --> 47:43.960
 this depressed Germanwings pilot,

47:43.960 --> 47:47.360
 when he flew his passenger jet into the Alps killing over 100

47:47.360 --> 47:50.640
 people, he just told the autopilot to do it.

47:50.640 --> 47:53.200
 He told the freaking computer to change the altitude

47:53.200 --> 47:55.040
 to 100 meters.

47:55.040 --> 47:58.160
 And even though it had the GPS maps, everything,

47:58.160 --> 48:00.640
 the computer was like, OK.

48:00.640 --> 48:05.320
 So we should take those very basic values,

48:05.320 --> 48:08.400
 where the problem is not that we don't agree.

48:08.400 --> 48:10.120
 The problem is just we've been too lazy

48:10.120 --> 48:11.480
 to try to put it into our machines

48:11.480 --> 48:15.520
 and make sure that from now on, airplanes will just,

48:15.520 --> 48:16.920
 which all have computers in them,

48:16.920 --> 48:19.720
 but will just refuse to do something like that.

48:19.720 --> 48:22.160
 Go into safe mode, maybe lock the cockpit door,

48:22.160 --> 48:24.480
 go over to the nearest airport.

48:24.480 --> 48:28.080
 And there's so much other technology in our world

48:28.080 --> 48:31.320
 as well now, where it's really becoming quite timely

48:31.320 --> 48:34.120
 to put in some sort of very basic values like this.

48:34.120 --> 48:39.240
 Even in cars, we've had enough vehicle terrorism attacks

48:39.240 --> 48:42.040
 by now, where people have driven trucks and vans

48:42.040 --> 48:45.480
 into pedestrians, that it's not at all a crazy idea

48:45.480 --> 48:48.680
 to just have that hardwired into the car.

48:48.680 --> 48:50.280
 Because yeah, there are a lot of,

48:50.280 --> 48:52.240
 there's always going to be people who for some reason

48:52.240 --> 48:54.800
 want to harm others, but most of those people

48:54.800 --> 48:56.760
 don't have the technical expertise to figure out

48:56.760 --> 48:58.520
 how to work around something like that.

48:58.520 --> 49:01.760
 So if the car just won't do it, it helps.

49:01.760 --> 49:02.840
 So let's start there.

49:02.840 --> 49:04.960
 So there's a lot of, that's a great point.

49:04.960 --> 49:06.800
 So not chasing perfect.

49:06.800 --> 49:10.840
 There's a lot of things that most of the world agrees on.

49:10.840 --> 49:11.840
 Yeah, let's start there.

49:11.840 --> 49:12.680
 Let's start there.

49:12.680 --> 49:14.560
 And then once we start there,

49:14.560 --> 49:17.240
 we'll also get into the habit of having

49:17.240 --> 49:18.520
 these kind of conversations about, okay,

49:18.520 --> 49:21.760
 what else should we put in here and have these discussions?

49:21.760 --> 49:23.920
 This should be a gradual process then.

49:23.920 --> 49:28.600
 Great, so, but that also means describing these things

49:28.600 --> 49:31.240
 and describing it to a machine.

49:31.240 --> 49:34.200
 So one thing, we had a few conversations

49:34.200 --> 49:35.640
 with Stephen Wolfram.

49:35.640 --> 49:37.080
 I'm not sure if you're familiar with Stephen.

49:37.080 --> 49:38.360
 Oh yeah, I know him quite well.

49:38.360 --> 49:42.040
 So he is, he works with a bunch of things,

49:42.040 --> 49:46.560
 but cellular automata, these simple computable things,

49:46.560 --> 49:47.960
 these computation systems.

49:47.960 --> 49:49.880
 And he kind of mentioned that,

49:49.880 --> 49:52.480
 we probably have already within these systems

49:52.480 --> 49:54.680
 already something that's AGI,

49:56.120 --> 49:58.720
 meaning like we just don't know it

49:58.720 --> 50:00.400
 because we can't talk to it.

50:00.400 --> 50:04.800
 So if you give me this chance to try to at least

50:04.800 --> 50:06.720
 form a question out of this is,

50:07.600 --> 50:10.880
 I think it's an interesting idea to think

50:10.880 --> 50:12.680
 that we can have intelligent systems,

50:12.680 --> 50:15.600
 but we don't know how to describe something to them

50:15.600 --> 50:17.360
 and they can't communicate with us.

50:17.360 --> 50:19.840
 I know you're doing a little bit of work in explainable AI,

50:19.840 --> 50:22.040
 trying to get AI to explain itself.

50:22.040 --> 50:25.520
 So what are your thoughts of natural language processing

50:25.520 --> 50:27.640
 or some kind of other communication?

50:27.640 --> 50:30.120
 How does the AI explain something to us?

50:30.120 --> 50:33.640
 How do we explain something to it, to machines?

50:33.640 --> 50:35.320
 Or you think of it differently?

50:35.320 --> 50:39.960
 So there are two separate parts to your question there.

50:39.960 --> 50:42.440
 One of them has to do with communication,

50:42.440 --> 50:44.440
 which is super interesting, I'll get to that in a sec.

50:44.440 --> 50:47.280
 The other is whether we already have AGI

50:47.280 --> 50:49.240
 but we just haven't noticed it there.

50:49.240 --> 50:50.080
 Right.

50:51.800 --> 50:53.000
 There I beg to differ.

50:54.280 --> 50:56.480
 I don't think there's anything in any cellular automaton

50:56.480 --> 50:59.040
 or anything or the internet itself or whatever

50:59.040 --> 51:03.560
 that has artificial general intelligence

51:03.560 --> 51:05.520
 and that it can really do exactly everything

51:05.520 --> 51:07.000
 we humans can do better.

51:07.000 --> 51:11.600
 I think the day that happens, when that happens,

51:11.600 --> 51:15.600
 we will very soon notice, we'll probably notice even before

51:15.600 --> 51:17.440
 because in a very, very big way.

51:17.440 --> 51:18.840
 But for the second part, though.

51:18.840 --> 51:20.720
 Wait, can I ask, sorry.

51:20.720 --> 51:24.400
 So, because you have this beautiful way

51:24.400 --> 51:29.400
 to formulating consciousness as information processing,

51:30.360 --> 51:31.360
 and you can think of intelligence

51:31.360 --> 51:32.280
 as information processing,

51:32.280 --> 51:34.320
 and you can think of the entire universe

51:34.320 --> 51:38.720
 as these particles and these systems roaming around

51:38.720 --> 51:41.360
 that have this information processing power.

51:41.360 --> 51:44.840
 You don't think there is something with the power

51:44.840 --> 51:49.040
 to process information in the way that we human beings do

51:49.040 --> 51:54.040
 that's out there that needs to be sort of connected to.

51:55.400 --> 51:57.880
 It seems a little bit philosophical, perhaps,

51:57.880 --> 52:00.080
 but there's something compelling to the idea

52:00.080 --> 52:01.920
 that the power is already there,

52:01.920 --> 52:05.440
 which the focus should be more on being able

52:05.440 --> 52:07.360
 to communicate with it.

52:07.360 --> 52:11.960
 Well, I agree that in a certain sense,

52:11.960 --> 52:15.360
 the hardware processing power is already out there

52:15.360 --> 52:19.000
 because our universe itself can think of it

52:19.000 --> 52:21.000
 as being a computer already, right?

52:21.000 --> 52:23.800
 It's constantly computing what water waves,

52:23.800 --> 52:26.120
 how it devolved the water waves in the River Charles

52:26.120 --> 52:28.440
 and how to move the air molecules around.

52:28.440 --> 52:30.480
 Seth Lloyd has pointed out, my colleague here,

52:30.480 --> 52:32.920
 that you can even in a very rigorous way

52:32.920 --> 52:35.480
 think of our entire universe as being a quantum computer.

52:35.480 --> 52:37.680
 It's pretty clear that our universe

52:37.680 --> 52:40.320
 supports this amazing processing power

52:40.320 --> 52:42.160
 because you can even,

52:42.160 --> 52:44.920
 within this physics computer that we live in, right?

52:44.920 --> 52:47.040
 We can even build actual laptops and stuff,

52:47.040 --> 52:49.000
 so clearly the power is there.

52:49.000 --> 52:52.040
 It's just that most of the compute power that nature has,

52:52.040 --> 52:54.240
 it's, in my opinion, kind of wasting on boring stuff

52:54.240 --> 52:56.520
 like simulating yet another ocean wave somewhere

52:56.520 --> 52:58.040
 where no one is even looking, right?

52:58.040 --> 53:00.880
 So in a sense, what life does, what we are doing

53:00.880 --> 53:03.880
 when we build computers is we're rechanneling

53:03.880 --> 53:07.200
 all this compute that nature is doing anyway

53:07.200 --> 53:09.360
 into doing things that are more interesting

53:09.360 --> 53:11.440
 than just yet another ocean wave,

53:11.440 --> 53:13.200
 and let's do something cool here.

53:14.080 --> 53:17.080
 So the raw hardware power is there, for sure,

53:17.080 --> 53:21.080
 but then even just computing what's going to happen

53:21.080 --> 53:23.520
 for the next five seconds in this water bottle,

53:23.520 --> 53:26.000
 takes a ridiculous amount of compute

53:26.000 --> 53:27.920
 if you do it on a human computer.

53:27.920 --> 53:29.920
 This water bottle just did it.

53:29.920 --> 53:33.440
 But that does not mean that this water bottle has AGI

53:34.760 --> 53:37.040
 because AGI means it should also be able to,

53:37.040 --> 53:40.160
 like I've written my book, done this interview.

53:40.160 --> 53:42.080
 And I don't think it's just communication problems.

53:42.080 --> 53:46.760
 I don't really think it can do it.

53:46.760 --> 53:49.280
 Although Buddhists say when they watch the water

53:49.280 --> 53:51.240
 and that there is some beauty,

53:51.240 --> 53:53.720
 that there's some depth and beauty in nature

53:53.720 --> 53:54.840
 that they can communicate with.

53:54.840 --> 53:56.480
 Communication is also very important though

53:56.480 --> 54:01.200
 because I mean, look, part of my job is being a teacher.

54:01.200 --> 54:06.200
 And I know some very intelligent professors even

54:06.200 --> 54:09.800
 who just have a bit of hard time communicating.

54:09.800 --> 54:12.640
 They come up with all these brilliant ideas,

54:12.640 --> 54:14.520
 but to communicate with somebody else,

54:14.520 --> 54:16.920
 you have to also be able to simulate their own mind.

54:16.920 --> 54:18.360
 Yes, empathy.

54:18.360 --> 54:20.640
 Build well enough and understand model of their mind

54:20.640 --> 54:24.400
 that you can say things that they will understand.

54:24.400 --> 54:26.480
 And that's quite difficult.

54:26.480 --> 54:28.280
 And that's why today it's so frustrating

54:28.280 --> 54:32.600
 if you have a computer that makes some cancer diagnosis

54:32.600 --> 54:34.120
 and you ask it, well, why are you saying

54:34.120 --> 54:36.120
 I should have this surgery?

54:36.120 --> 54:37.960
 And if it can only reply,

54:37.960 --> 54:40.800
 I was trained on five terabytes of data

54:40.800 --> 54:45.080
 and this is my diagnosis, boop, boop, beep, beep.

54:45.080 --> 54:49.120
 It doesn't really instill a lot of confidence, right?

54:49.120 --> 54:51.120
 So I think we have a lot of work to do

54:51.120 --> 54:54.320
 on communication there.

54:54.320 --> 54:58.040
 So what kind of, I think you're doing a little bit of work

54:58.040 --> 54:59.320
 in explainable AI.

54:59.320 --> 55:01.320
 What do you think are the most promising avenues?

55:01.320 --> 55:05.240
 Is it mostly about sort of the Alexa problem

55:05.240 --> 55:07.200
 of natural language processing of being able

55:07.200 --> 55:11.600
 to actually use human interpretable methods

55:11.600 --> 55:13.160
 of communication?

55:13.160 --> 55:16.000
 So being able to talk to a system and it talk back to you,

55:16.000 --> 55:18.640
 or is there some more fundamental problems to be solved?

55:18.640 --> 55:21.160
 I think it's all of the above.

55:21.160 --> 55:23.520
 The natural language processing is obviously important,

55:23.520 --> 55:27.600
 but there are also more nerdy fundamental problems.

55:27.600 --> 55:31.640
 Like if you take, you play chess?

55:31.640 --> 55:33.040
 Of course, I'm Russian.

55:33.040 --> 55:33.880
 I have to.

55:33.880 --> 55:34.720
 You speak Russian?

55:34.720 --> 55:35.560
 Yes, I speak Russian.

55:35.560 --> 55:38.040
 Excellent, I didn't know.

55:38.040 --> 55:39.160
 When did you learn Russian?

55:39.160 --> 55:41.800
 I speak very bad Russian, I'm only an autodidact,

55:41.800 --> 55:44.560
 but I bought a book, Teach Yourself Russian,

55:44.560 --> 55:47.720
 read a lot, but it was very difficult.

55:47.720 --> 55:48.560
 Wow.

55:48.560 --> 55:49.960
 That's why I speak so bad.

55:49.960 --> 55:51.960
 How many languages do you know?

55:51.960 --> 55:53.840
 Wow, that's really impressive.

55:53.840 --> 55:56.320
 I don't know, my wife has some calculation,

55:56.320 --> 55:58.400
 but my point was, if you play chess,

55:58.400 --> 56:01.040
 have you looked at the AlphaZero games?

56:01.040 --> 56:02.600
 The actual games, no.

56:02.600 --> 56:05.000
 Check it out, some of them are just mind blowing,

56:06.320 --> 56:07.720
 really beautiful.

56:07.720 --> 56:12.400
 And if you ask, how did it do that?

56:13.760 --> 56:16.520
 You go talk to Demis Hassabis,

56:16.520 --> 56:18.240
 I know others from DeepMind,

56:19.120 --> 56:20.600
 all they'll ultimately be able to give you

56:20.600 --> 56:23.920
 is big tables of numbers, matrices,

56:23.920 --> 56:25.720
 that define the neural network.

56:25.720 --> 56:28.080
 And you can stare at these tables of numbers

56:28.080 --> 56:29.600
 till your face turn blue,

56:29.600 --> 56:32.520
 and you're not gonna understand much

56:32.520 --> 56:34.520
 about why it made that move.

56:34.520 --> 56:37.640
 And even if you have natural language processing

56:37.640 --> 56:40.280
 that can tell you in human language about,

56:40.280 --> 56:42.520
 oh, five, seven, points, two, eight,

56:42.520 --> 56:43.560
 still not gonna really help.

56:43.560 --> 56:47.480
 So I think there's a whole spectrum of fun challenges

56:47.480 --> 56:50.520
 that are involved in taking a computation

56:50.520 --> 56:52.240
 that does intelligent things

56:52.240 --> 56:56.240
 and transforming it into something equally good,

56:57.760 --> 57:01.840
 equally intelligent, but that's more understandable.

57:01.840 --> 57:03.240
 And I think that's really valuable

57:03.240 --> 57:07.440
 because I think as we put machines in charge

57:07.440 --> 57:09.760
 of ever more infrastructure in our world,

57:09.760 --> 57:12.680
 the power grid, the trading on the stock market,

57:12.680 --> 57:14.320
 weapon systems and so on,

57:14.320 --> 57:17.760
 it's absolutely crucial that we can trust

57:17.760 --> 57:19.400
 these AIs to do all we want.

57:19.400 --> 57:21.520
 And trust really comes from understanding

57:22.520 --> 57:24.400
 in a very fundamental way.

57:24.400 --> 57:27.560
 And that's why I'm working on this,

57:27.560 --> 57:29.160
 because I think the more,

57:29.160 --> 57:31.840
 if we're gonna have some hope of ensuring

57:31.840 --> 57:33.520
 that machines have adopted our goals

57:33.520 --> 57:35.800
 and that they're gonna retain them,

57:35.800 --> 57:38.800
 that kind of trust, I think,

57:38.800 --> 57:41.200
 needs to be based on things you can actually understand,

57:41.200 --> 57:44.240
 preferably even improve theorems on.

57:44.240 --> 57:46.080
 Even with a self driving car, right?

57:47.040 --> 57:48.680
 If someone just tells you it's been trained

57:48.680 --> 57:50.640
 on tons of data and it never crashed,

57:50.640 --> 57:54.200
 it's less reassuring than if someone actually has a proof.

57:54.200 --> 57:55.960
 Maybe it's a computer verified proof,

57:55.960 --> 57:58.800
 but still it says that under no circumstances

57:58.800 --> 58:02.320
 is this car just gonna swerve into oncoming traffic.

58:02.320 --> 58:04.640
 And that kind of information helps to build trust

58:04.640 --> 58:08.080
 and helps build the alignment of goals,

58:09.400 --> 58:12.200
 at least awareness that your goals, your values are aligned.

58:12.200 --> 58:13.840
 And I think even in the very short term,

58:13.840 --> 58:16.360
 if you look at how, you know, today, right?

58:16.360 --> 58:19.320
 This absolutely pathetic state of cybersecurity

58:19.320 --> 58:21.720
 that we have, where is it?

58:21.720 --> 58:25.960
 Three billion Yahoo accounts we can't pack,

58:27.200 --> 58:31.720
 almost every American's credit card and so on.

58:32.800 --> 58:34.120
 Why is this happening?

58:34.120 --> 58:37.960
 It's ultimately happening because we have software

58:37.960 --> 58:41.200
 that nobody fully understood how it worked.

58:41.200 --> 58:44.800
 That's why the bugs hadn't been found, right?

58:44.800 --> 58:47.480
 And I think AI can be used very effectively

58:47.480 --> 58:49.640
 for offense, for hacking,

58:49.640 --> 58:52.320
 but it can also be used for defense.

58:52.320 --> 58:55.360
 Hopefully automating verifiability

58:55.360 --> 59:00.360
 and creating systems that are built in different ways

59:00.680 --> 59:02.920
 so you can actually prove things about them.

59:02.920 --> 59:05.240
 And it's important.

59:05.240 --> 59:07.680
 So speaking of software that nobody understands

59:07.680 --> 59:10.640
 how it works, of course, a bunch of people ask

59:10.640 --> 59:12.160
 about your paper, about your thoughts

59:12.160 --> 59:14.680
 of why does deep and cheap learning work so well?

59:14.680 --> 59:15.520
 That's the paper.

59:15.520 --> 59:18.320
 But what are your thoughts on deep learning?

59:18.320 --> 59:21.880
 These kind of simplified models of our own brains

59:21.880 --> 59:26.440
 have been able to do some successful perception work,

59:26.440 --> 59:29.560
 pattern recognition work, and now with AlphaZero and so on,

59:29.560 --> 59:30.880
 do some clever things.

59:30.880 --> 59:33.880
 What are your thoughts about the promise limitations

59:33.880 --> 59:35.680
 of this piece?

59:35.680 --> 59:40.680
 Great, I think there are a number of very important insights,

59:43.080 --> 59:44.640
 very important lessons we can always draw

59:44.640 --> 59:47.120
 from these kinds of successes.

59:47.120 --> 59:48.960
 One of them is when you look at the human brain,

59:48.960 --> 59:51.480
 you see it's very complicated, 10th of 11 neurons,

59:51.480 --> 59:53.320
 and there are all these different kinds of neurons

59:53.320 --> 59:55.040
 and yada, yada, and there's been this long debate

59:55.040 --> 59:57.200
 about whether the fact that we have dozens

59:57.200 --> 1:00:00.160
 of different kinds is actually necessary for intelligence.

1:00:01.560 --> 1:00:03.360
 We can now, I think, quite convincingly answer

1:00:03.360 --> 1:00:07.640
 that question of no, it's enough to have just one kind.

1:00:07.640 --> 1:00:09.920
 If you look under the hood of AlphaZero,

1:00:09.920 --> 1:00:11.080
 there's only one kind of neuron

1:00:11.080 --> 1:00:15.000
 and it's ridiculously simple mathematical thing.

1:00:15.000 --> 1:00:17.280
 So it's just like in physics,

1:00:17.280 --> 1:00:20.320
 it's not, if you have a gas with waves in it,

1:00:20.320 --> 1:00:23.240
 it's not the detailed nature of the molecule that matter,

1:00:24.240 --> 1:00:26.040
 it's the collective behavior somehow.

1:00:26.040 --> 1:00:30.720
 Similarly, it's this higher level structure

1:00:30.720 --> 1:00:31.760
 of the network that matters,

1:00:31.760 --> 1:00:34.080
 not that you have 20 kinds of neurons.

1:00:34.080 --> 1:00:37.040
 I think our brain is such a complicated mess

1:00:37.040 --> 1:00:41.720
 because it wasn't evolved just to be intelligent,

1:00:41.720 --> 1:00:45.840
 it was involved to also be self assembling

1:00:47.000 --> 1:00:48.760
 and self repairing, right?

1:00:48.760 --> 1:00:51.920
 And evolutionarily attainable.

1:00:51.920 --> 1:00:53.560
 And so on and so on.

1:00:53.560 --> 1:00:54.720
 So I think it's pretty,

1:00:54.720 --> 1:00:57.040
 my hunch is that we're going to understand

1:00:57.040 --> 1:00:59.520
 how to build AGI before we fully understand

1:00:59.520 --> 1:01:02.600
 how our brains work, just like we understood

1:01:02.600 --> 1:01:05.560
 how to build flying machines long before

1:01:05.560 --> 1:01:07.800
 we were able to build a mechanical bird.

1:01:07.800 --> 1:01:08.640
 Yeah, that's right.

1:01:08.640 --> 1:01:13.280
 You've given the example exactly of mechanical birds

1:01:13.280 --> 1:01:15.680
 and airplanes and airplanes do a pretty good job

1:01:15.680 --> 1:01:18.560
 of flying without really mimicking bird flight.

1:01:18.560 --> 1:01:20.920
 And even now after 100 years later,

1:01:20.920 --> 1:01:23.880
 did you see the Ted talk with this German mechanical bird?

1:01:23.880 --> 1:01:25.040
 I heard you mention it.

1:01:25.040 --> 1:01:26.520
 Check it out, it's amazing.

1:01:26.520 --> 1:01:27.760
 But even after that, right,

1:01:27.760 --> 1:01:29.360
 we still don't fly in mechanical birds

1:01:29.360 --> 1:01:32.720
 because it turned out the way we came up with was simpler

1:01:32.720 --> 1:01:33.840
 and it's better for our purposes.

1:01:33.840 --> 1:01:35.280
 And I think it might be the same there.

1:01:35.280 --> 1:01:36.280
 That's one lesson.

1:01:37.520 --> 1:01:42.520
 And another lesson, it's more what our paper was about.

1:01:42.640 --> 1:01:45.800
 First, as a physicist thought it was fascinating

1:01:45.800 --> 1:01:48.240
 how there's a very close mathematical relationship

1:01:48.240 --> 1:01:50.800
 actually between our artificial neural networks

1:01:50.800 --> 1:01:54.560
 and a lot of things that we've studied for in physics

1:01:54.560 --> 1:01:57.520
 go by nerdy names like the renormalization group equation

1:01:57.520 --> 1:01:59.800
 and Hamiltonians and yada, yada, yada.

1:01:59.800 --> 1:02:04.360
 And when you look a little more closely at this,

1:02:05.720 --> 1:02:06.560
 you have,

1:02:10.320 --> 1:02:12.360
 at first I was like, well, there's something crazy here

1:02:12.360 --> 1:02:13.520
 that doesn't make sense.

1:02:13.520 --> 1:02:18.520
 Because we know that if you even want to build

1:02:19.200 --> 1:02:22.560
 a super simple neural network to tell apart cat pictures

1:02:22.560 --> 1:02:23.400
 and dog pictures, right,

1:02:23.400 --> 1:02:25.400
 that you can do that very, very well now.

1:02:25.400 --> 1:02:27.520
 But if you think about it a little bit,

1:02:27.520 --> 1:02:29.080
 you convince yourself it must be impossible

1:02:29.080 --> 1:02:31.920
 because if I have one megapixel,

1:02:31.920 --> 1:02:34.160
 even if each pixel is just black or white,

1:02:34.160 --> 1:02:36.960
 there's two to the power of 1 million possible images,

1:02:36.960 --> 1:02:38.960
 which is way more than there are atoms in our universe,

1:02:38.960 --> 1:02:41.000
 right, so in order to,

1:02:42.040 --> 1:02:43.200
 and then for each one of those,

1:02:43.200 --> 1:02:44.640
 I have to assign a number,

1:02:44.640 --> 1:02:47.080
 which is the probability that it's a dog.

1:02:47.080 --> 1:02:49.440
 So an arbitrary function of images

1:02:49.440 --> 1:02:54.440
 is a list of more numbers than there are atoms in our universe.

1:02:54.440 --> 1:02:57.360
 So clearly I can't store that under the hood of my GPU

1:02:57.360 --> 1:03:00.640
 or my computer, yet somehow it works.

1:03:00.640 --> 1:03:01.480
 So what does that mean?

1:03:01.480 --> 1:03:04.960
 Well, it means that out of all of the problems

1:03:04.960 --> 1:03:08.200
 that you could try to solve with a neural network,

1:03:10.120 --> 1:03:12.880
 almost all of them are impossible to solve

1:03:12.880 --> 1:03:14.560
 with a reasonably sized one.

1:03:15.480 --> 1:03:17.440
 But then what we showed in our paper

1:03:17.440 --> 1:03:22.360
 was that the fraction, the kind of problems,

1:03:22.360 --> 1:03:23.800
 the fraction of all the problems

1:03:23.800 --> 1:03:26.520
 that you could possibly pose,

1:03:26.520 --> 1:03:29.480
 that we actually care about given the laws of physics

1:03:29.480 --> 1:03:32.480
 is also an infinite testimony, tiny little part.

1:03:32.480 --> 1:03:35.440
 And amazingly, they're basically the same part.

1:03:35.440 --> 1:03:37.560
 Yeah, it's almost like our world was created for,

1:03:37.560 --> 1:03:39.000
 I mean, they kind of come together.

1:03:39.000 --> 1:03:42.800
 Yeah, well, you could say maybe where the world was created

1:03:42.800 --> 1:03:44.960
 for us, but I have a more modest interpretation,

1:03:44.960 --> 1:03:46.680
 which is that the world was created for us,

1:03:46.680 --> 1:03:48.040
 but I have a more modest interpretation,

1:03:48.040 --> 1:03:50.360
 which is that instead evolution endowed us

1:03:50.360 --> 1:03:53.120
 with neural networks precisely for that reason.

1:03:53.120 --> 1:03:54.640
 Because this particular architecture,

1:03:54.640 --> 1:03:56.040
 as opposed to the one in your laptop,

1:03:56.040 --> 1:04:01.040
 is very, very well adapted to solving the kind of problems

1:04:02.480 --> 1:04:05.560
 that nature kept presenting our ancestors with.

1:04:05.560 --> 1:04:08.120
 So it makes sense that why do we have a brain

1:04:08.120 --> 1:04:09.280
 in the first place?

1:04:09.280 --> 1:04:11.880
 It's to be able to make predictions about the future

1:04:11.880 --> 1:04:12.880
 and so on.

1:04:12.880 --> 1:04:16.440
 So if we had a sucky system, which could never solve it,

1:04:16.440 --> 1:04:18.280
 we wouldn't have a world.

1:04:18.280 --> 1:04:23.280
 So this is, I think, a very beautiful fact.

1:04:23.680 --> 1:04:24.520
 Yeah.

1:04:24.520 --> 1:04:29.000
 We also realize that there's been earlier work

1:04:29.000 --> 1:04:32.040
 on why deeper networks are good,

1:04:32.040 --> 1:04:34.680
 but we were able to show an additional cool fact there,

1:04:34.680 --> 1:04:38.360
 which is that even incredibly simple problems,

1:04:38.360 --> 1:04:41.080
 like suppose I give you a thousand numbers

1:04:41.080 --> 1:04:42.720
 and ask you to multiply them together,

1:04:42.720 --> 1:04:46.680
 and you can write a few lines of code, boom, done, trivial.

1:04:46.680 --> 1:04:49.520
 If you just try to do that with a neural network

1:04:49.520 --> 1:04:52.440
 that has only one single hidden layer in it,

1:04:52.440 --> 1:04:53.400
 you can do it,

1:04:54.320 --> 1:04:57.360
 but you're going to need two to the power of a thousand

1:04:57.360 --> 1:05:00.920
 neurons to multiply a thousand numbers,

1:05:00.920 --> 1:05:02.520
 which is, again, more neurons than there are atoms

1:05:02.520 --> 1:05:03.360
 in our universe.

1:05:04.600 --> 1:05:05.480
 That's fascinating.

1:05:05.480 --> 1:05:09.960
 But if you allow yourself to make it a deep network

1:05:09.960 --> 1:05:13.240
 with many layers, you only need 4,000 neurons.

1:05:13.240 --> 1:05:14.520
 It's perfectly feasible.

1:05:16.400 --> 1:05:17.960
 That's really interesting.

1:05:17.960 --> 1:05:18.800
 Yeah.

1:05:18.800 --> 1:05:21.040
 So on another architecture type,

1:05:21.040 --> 1:05:22.720
 I mean, you mentioned Schrodinger's equation,

1:05:22.720 --> 1:05:26.360
 and what are your thoughts about quantum computing

1:05:27.240 --> 1:05:32.240
 and the role of this kind of computational unit

1:05:32.400 --> 1:05:34.880
 in creating an intelligence system?

1:05:34.880 --> 1:05:39.520
 In some Hollywood movies that I will not mention by name

1:05:39.520 --> 1:05:41.040
 because I don't want to spoil them.

1:05:41.040 --> 1:05:44.240
 The way they get AGI is building a quantum computer.

1:05:45.480 --> 1:05:47.600
 Because the word quantum sounds cool and so on.

1:05:47.600 --> 1:05:48.440
 That's right.

1:05:50.040 --> 1:05:52.880
 First of all, I think we don't need quantum computers

1:05:52.880 --> 1:05:54.920
 to build AGI.

1:05:54.920 --> 1:05:59.240
 I suspect your brain is not a quantum computer

1:05:59.240 --> 1:06:00.640
 in any profound sense.

1:06:01.600 --> 1:06:03.200
 So you don't even wrote a paper about that

1:06:03.200 --> 1:06:04.560
 a lot many years ago.

1:06:04.560 --> 1:06:08.120
 I calculated the so called decoherence time,

1:06:08.120 --> 1:06:10.320
 how long it takes until the quantum computerness

1:06:10.320 --> 1:06:13.400
 of what your neurons are doing gets erased

1:06:15.320 --> 1:06:17.960
 by just random noise from the environment.

1:06:17.960 --> 1:06:21.320
 And it's about 10 to the minus 21 seconds.

1:06:21.320 --> 1:06:24.600
 So as cool as it would be to have a quantum computer

1:06:24.600 --> 1:06:27.320
 in my head, I don't think that fast.

1:06:27.320 --> 1:06:28.360
 On the other hand,

1:06:28.360 --> 1:06:33.040
 there are very cool things you could do

1:06:33.040 --> 1:06:34.200
 with quantum computers.

1:06:35.240 --> 1:06:37.480
 Or I think we'll be able to do soon

1:06:37.480 --> 1:06:39.360
 when we get bigger ones.

1:06:39.360 --> 1:06:40.960
 That might actually help machine learning

1:06:40.960 --> 1:06:43.160
 do even better than the brain.

1:06:43.160 --> 1:06:45.640
 So for example,

1:06:47.040 --> 1:06:50.760
 one, this is just a moonshot,

1:06:50.760 --> 1:06:55.760
 but learning is very much same thing as search.

1:07:01.800 --> 1:07:03.160
 If you're trying to train a neural network

1:07:03.160 --> 1:07:06.240
 to get really learned to do something really well,

1:07:06.240 --> 1:07:07.280
 you have some loss function,

1:07:07.280 --> 1:07:10.360
 you have a bunch of knobs you can turn,

1:07:10.360 --> 1:07:12.080
 represented by a bunch of numbers,

1:07:12.080 --> 1:07:12.920
 and you're trying to tweak them

1:07:12.920 --> 1:07:15.080
 so that it becomes as good as possible at this thing.

1:07:15.080 --> 1:07:19.680
 So if you think of a landscape with some valley,

1:07:20.720 --> 1:07:22.120
 where each dimension of the landscape

1:07:22.120 --> 1:07:24.120
 corresponds to some number you can change,

1:07:24.120 --> 1:07:25.640
 you're trying to find the minimum.

1:07:25.640 --> 1:07:26.760
 And it's well known that

1:07:26.760 --> 1:07:29.040
 if you have a very high dimensional landscape,

1:07:29.040 --> 1:07:31.840
 complicated things, it's super hard to find the minimum.

1:07:31.840 --> 1:07:35.840
 Quantum mechanics is amazingly good at this.

1:07:35.840 --> 1:07:38.240
 Like if I want to know what's the lowest energy state

1:07:38.240 --> 1:07:39.720
 this water can possibly have,

1:07:41.720 --> 1:07:42.560
 incredibly hard to compute,

1:07:42.560 --> 1:07:45.400
 but nature will happily figure this out for you

1:07:45.400 --> 1:07:48.000
 if you just cool it down, make it very, very cold.

1:07:49.800 --> 1:07:50.880
 If you put a ball somewhere,

1:07:50.880 --> 1:07:52.240
 it'll roll down to its minimum.

1:07:52.240 --> 1:07:54.280
 And this happens metaphorically

1:07:54.280 --> 1:07:56.320
 at the energy landscape too.

1:07:56.320 --> 1:07:59.280
 And quantum mechanics even uses some clever tricks,

1:07:59.280 --> 1:08:02.520
 which today's machine learning systems don't.

1:08:02.520 --> 1:08:04.160
 Like if you're trying to find the minimum

1:08:04.160 --> 1:08:06.960
 and you get stuck in the little local minimum here,

1:08:06.960 --> 1:08:08.760
 in quantum mechanics you can actually tunnel

1:08:08.760 --> 1:08:11.840
 through the barrier and get unstuck again.

1:08:13.480 --> 1:08:14.320
 That's really interesting.

1:08:14.320 --> 1:08:16.120
 Yeah, so it may be, for example,

1:08:16.120 --> 1:08:19.160
 that we'll one day use quantum computers

1:08:19.160 --> 1:08:22.840
 that help train neural networks better.

1:08:22.840 --> 1:08:23.680
 That's really interesting.

1:08:23.680 --> 1:08:27.040
 Okay, so as a component of kind of the learning process,

1:08:27.040 --> 1:08:27.880
 for example.

1:08:27.880 --> 1:08:29.440
 Yeah.

1:08:29.440 --> 1:08:33.080
 Let me ask sort of wrapping up here a little bit,

1:08:33.080 --> 1:08:36.880
 let me return to the questions of our human nature

1:08:36.880 --> 1:08:40.000
 and love, as I mentioned.

1:08:40.000 --> 1:08:41.640
 So do you think,

1:08:44.280 --> 1:08:46.000
 you mentioned sort of a helper robot,

1:08:46.000 --> 1:08:48.640
 but you could think of also personal robots.

1:08:48.640 --> 1:08:52.480
 Do you think the way we human beings fall in love

1:08:52.480 --> 1:08:54.680
 and get connected to each other

1:08:54.680 --> 1:08:58.040
 is possible to achieve in an AI system

1:08:58.040 --> 1:09:00.360
 and human level AI intelligence system?

1:09:00.360 --> 1:09:03.720
 Do you think we would ever see that kind of connection?

1:09:03.720 --> 1:09:06.160
 Or, you know, in all this discussion

1:09:06.160 --> 1:09:08.520
 about solving complex goals,

1:09:08.520 --> 1:09:10.760
 is this kind of human social connection,

1:09:10.760 --> 1:09:12.560
 do you think that's one of the goals

1:09:12.560 --> 1:09:16.280
 on the peaks and valleys with the raising sea levels

1:09:16.280 --> 1:09:17.360
 that we'll be able to achieve?

1:09:17.360 --> 1:09:20.040
 Or do you think that's something that's ultimately,

1:09:20.040 --> 1:09:21.760
 or at least in the short term,

1:09:21.760 --> 1:09:23.640
 relative to the other goals is not achievable?

1:09:23.640 --> 1:09:25.120
 I think it's all possible.

1:09:25.120 --> 1:09:27.600
 And I mean, in recent,

1:09:27.600 --> 1:09:30.840
 there's a very wide range of guesses, as you know,

1:09:30.840 --> 1:09:33.720
 among AI researchers, when we're going to get AGI.

1:09:35.120 --> 1:09:37.640
 Some people, you know, like our friend Rodney Brooks

1:09:37.640 --> 1:09:41.040
 says it's going to be hundreds of years at least.

1:09:41.040 --> 1:09:42.200
 And then there are many others

1:09:42.200 --> 1:09:44.040
 who think it's going to happen much sooner.

1:09:44.040 --> 1:09:45.520
 And recent polls,

1:09:46.840 --> 1:09:48.640
 maybe half or so of AI researchers

1:09:48.640 --> 1:09:50.920
 think we're going to get AGI within decades.

1:09:50.920 --> 1:09:52.720
 So if that happens, of course,

1:09:52.720 --> 1:09:55.040
 then I think these things are all possible.

1:09:55.040 --> 1:09:56.840
 But in terms of whether it will happen,

1:09:56.840 --> 1:10:00.600
 I think we shouldn't spend so much time asking

1:10:00.600 --> 1:10:03.240
 what do we think will happen in the future?

1:10:03.240 --> 1:10:05.160
 As if we are just some sort of pathetic,

1:10:05.160 --> 1:10:07.040
 your passive bystanders, you know,

1:10:07.040 --> 1:10:09.280
 waiting for the future to happen to us.

1:10:09.280 --> 1:10:11.640
 Hey, we're the ones creating this future, right?

1:10:11.640 --> 1:10:15.520
 So we should be proactive about it

1:10:15.520 --> 1:10:16.920
 and ask ourselves what sort of future

1:10:16.920 --> 1:10:18.240
 we would like to have happen.

1:10:18.240 --> 1:10:19.920
 We're going to make it like that.

1:10:19.920 --> 1:10:22.720
 Well, what I prefer is just some sort of incredibly boring,

1:10:22.720 --> 1:10:24.320
 zombie like future where there's all these

1:10:24.320 --> 1:10:26.040
 mechanical things happening and there's no passion,

1:10:26.040 --> 1:10:28.040
 no emotion, no experience, maybe even.

1:10:29.600 --> 1:10:32.040
 No, I would of course, much rather prefer it

1:10:32.040 --> 1:10:35.240
 if all the things that we find that we value the most

1:10:36.240 --> 1:10:40.680
 about humanity are our subjective experience,

1:10:40.680 --> 1:10:43.000
 passion, inspiration, love, you know.

1:10:43.000 --> 1:10:48.000
 If we can create a future where those things do happen,

1:10:48.000 --> 1:10:50.840
 where those things do exist, you know,

1:10:50.840 --> 1:10:54.560
 I think ultimately it's not our universe

1:10:54.560 --> 1:10:57.960
 giving meaning to us, it's us giving meaning to our universe.

1:10:57.960 --> 1:11:01.840
 And if we build more advanced intelligence,

1:11:01.840 --> 1:11:03.680
 let's make sure we build it in such a way

1:11:03.680 --> 1:11:08.680
 that meaning is part of it.

1:11:09.120 --> 1:11:11.400
 A lot of people that seriously study this problem

1:11:11.400 --> 1:11:13.600
 and think of it from different angles

1:11:13.600 --> 1:11:16.880
 have trouble in the majority of cases,

1:11:16.880 --> 1:11:19.160
 if they think through that happen,

1:11:19.160 --> 1:11:22.520
 are the ones that are not beneficial to humanity.

1:11:22.520 --> 1:11:25.560
 And so, yeah, so what are your thoughts?

1:11:25.560 --> 1:11:29.400
 What's should people, you know,

1:11:29.400 --> 1:11:32.040
 I really don't like people to be terrified.

1:11:33.440 --> 1:11:35.040
 What's a way for people to think about it

1:11:35.040 --> 1:11:39.600
 in a way we can solve it and we can make it better?

1:11:39.600 --> 1:11:42.960
 No, I don't think panicking is going to help in any way.

1:11:42.960 --> 1:11:44.840
 It's not going to increase chances

1:11:44.840 --> 1:11:45.880
 of things going well either.

1:11:45.880 --> 1:11:48.400
 Even if you are in a situation where there is a real threat,

1:11:48.400 --> 1:11:51.080
 does it help if everybody just freaks out?

1:11:51.080 --> 1:11:52.680
 No, of course, of course not.

1:11:53.640 --> 1:11:56.600
 I think, yeah, there are of course ways

1:11:56.600 --> 1:11:58.440
 in which things can go horribly wrong.

1:11:59.560 --> 1:12:03.680
 First of all, it's important when we think about this thing,

1:12:03.680 --> 1:12:05.280
 about the problems and risks,

1:12:05.280 --> 1:12:07.160
 to also remember how huge the upsides can be

1:12:07.160 --> 1:12:08.440
 if we get it right, right?

1:12:08.440 --> 1:12:12.360
 Everything we love about society and civilization

1:12:12.360 --> 1:12:13.400
 is a product of intelligence.

1:12:13.400 --> 1:12:15.320
 So if we can amplify our intelligence

1:12:15.320 --> 1:12:18.760
 with machine intelligence and not anymore lose our loved one

1:12:18.760 --> 1:12:21.080
 to what we're told is an incurable disease

1:12:21.080 --> 1:12:24.800
 and things like this, of course, we should aspire to that.

1:12:24.800 --> 1:12:26.680
 So that can be a motivator, I think,

1:12:26.680 --> 1:12:29.120
 reminding ourselves that the reason we try to solve problems

1:12:29.120 --> 1:12:33.520
 is not just because we're trying to avoid gloom,

1:12:33.520 --> 1:12:35.760
 but because we're trying to do something great.

1:12:35.760 --> 1:12:37.680
 But then in terms of the risks,

1:12:37.680 --> 1:12:42.680
 I think the really important question is to ask,

1:12:42.680 --> 1:12:45.480
 what can we do today that will actually help

1:12:45.480 --> 1:12:47.320
 make the outcome good, right?

1:12:47.320 --> 1:12:49.880
 And dismissing the risk is not one of them.

1:12:51.240 --> 1:12:54.800
 I find it quite funny often when I'm in discussion panels

1:12:54.800 --> 1:12:55.960
 about these things,

1:12:55.960 --> 1:13:00.960
 how the people who work for companies,

1:13:01.200 --> 1:13:03.120
 always be like, oh, nothing to worry about,

1:13:03.120 --> 1:13:04.760
 nothing to worry about, nothing to worry about.

1:13:04.760 --> 1:13:09.600
 And it's only academics sometimes express concerns.

1:13:09.600 --> 1:13:11.880
 That's not surprising at all if you think about it.

1:13:11.880 --> 1:13:12.880
 Right.

1:13:12.880 --> 1:13:15.200
 Upton Sinclair quipped, right,

1:13:15.200 --> 1:13:18.040
 that it's hard to make a man believe in something

1:13:18.040 --> 1:13:20.120
 when his income depends on not believing in it.

1:13:20.120 --> 1:13:24.080
 And frankly, we know a lot of these people in companies

1:13:24.080 --> 1:13:26.240
 that they're just as concerned as anyone else.

1:13:26.240 --> 1:13:28.480
 But if you're the CEO of a company,

1:13:28.480 --> 1:13:30.280
 that's not something you want to go on record saying

1:13:30.280 --> 1:13:33.440
 when you have silly journalists who are gonna put a picture

1:13:33.440 --> 1:13:35.720
 of a Terminator robot when they quote you.

1:13:35.720 --> 1:13:39.040
 So the issues are real.

1:13:39.040 --> 1:13:41.920
 And the way I think about what the issue is,

1:13:41.920 --> 1:13:46.920
 is basically the real choice we have is,

1:13:48.040 --> 1:13:50.840
 first of all, are we gonna just dismiss the risks

1:13:50.840 --> 1:13:54.480
 and say, well, let's just go ahead and build machines

1:13:54.480 --> 1:13:57.560
 that can do everything we can do better and cheaper.

1:13:57.560 --> 1:14:00.200
 Let's just make ourselves obsolete as fast as possible.

1:14:00.200 --> 1:14:01.720
 What could possibly go wrong?

1:14:01.720 --> 1:14:03.440
 That's one attitude.

1:14:03.440 --> 1:14:05.440
 The opposite attitude, I think, is to say,

1:14:06.400 --> 1:14:08.800
 here's this incredible potential,

1:14:08.800 --> 1:14:11.960
 let's think about what kind of future

1:14:11.960 --> 1:14:14.640
 we're really, really excited about.

1:14:14.640 --> 1:14:18.480
 What are the shared goals that we can really aspire towards?

1:14:18.480 --> 1:14:19.960
 And then let's think really hard

1:14:19.960 --> 1:14:22.000
 about how we can actually get there.

1:14:22.000 --> 1:14:24.160
 So start with, don't start thinking about the risks,

1:14:24.160 --> 1:14:26.720
 start thinking about the goals.

1:14:26.720 --> 1:14:28.200
 And then when you do that,

1:14:28.200 --> 1:14:30.480
 then you can think about the obstacles you want to avoid.

1:14:30.480 --> 1:14:32.840
 I often get students coming in right here into my office

1:14:32.840 --> 1:14:34.120
 for career advice.

1:14:34.120 --> 1:14:35.560
 I always ask them this very question,

1:14:35.560 --> 1:14:37.920
 where do you want to be in the future?

1:14:37.920 --> 1:14:40.640
 If all she can say is, oh, maybe I'll have cancer,

1:14:40.640 --> 1:14:42.480
 maybe I'll get run over by a truck.

1:14:42.480 --> 1:14:44.280
 Yeah, focus on the obstacles instead of the goals.

1:14:44.280 --> 1:14:46.880
 She's just going to end up a hypochondriac paranoid.

1:14:47.920 --> 1:14:49.920
 Whereas if she comes in and fire in her eyes

1:14:49.920 --> 1:14:51.840
 and is like, I want to be there.

1:14:51.840 --> 1:14:53.960
 And then we can talk about the obstacles

1:14:53.960 --> 1:14:55.760
 and see how we can circumvent them.

1:14:55.760 --> 1:14:58.880
 That's, I think, a much, much healthier attitude.

1:14:58.880 --> 1:15:03.880
 And I feel it's very challenging to come up with a vision

1:15:03.880 --> 1:15:08.120
 for the future, which we are unequivocally excited about.

1:15:08.120 --> 1:15:10.320
 I'm not just talking now in the vague terms,

1:15:10.320 --> 1:15:12.360
 like, yeah, let's cure cancer, fine.

1:15:12.360 --> 1:15:14.720
 I'm talking about what kind of society

1:15:14.720 --> 1:15:15.840
 do we want to create?

1:15:15.840 --> 1:15:20.360
 What do we want it to mean to be human in the age of AI,

1:15:20.360 --> 1:15:21.720
 in the age of AGI?

1:15:22.840 --> 1:15:25.360
 So if we can have this conversation,

1:15:25.360 --> 1:15:28.200
 broad, inclusive conversation,

1:15:28.200 --> 1:15:31.400
 and gradually start converging towards some,

1:15:31.400 --> 1:15:34.240
 some future that with some direction, at least,

1:15:34.240 --> 1:15:35.400
 that we want to steer towards, right,

1:15:35.400 --> 1:15:38.160
 then we'll be much more motivated

1:15:38.160 --> 1:15:39.960
 to constructively take on the obstacles.

1:15:39.960 --> 1:15:43.560
 And I think if I had, if I had to,

1:15:43.560 --> 1:15:46.640
 if I try to wrap this up in a more succinct way,

1:15:46.640 --> 1:15:51.480
 I think we can all agree already now

1:15:51.480 --> 1:15:56.160
 that we should aspire to build AGI

1:15:56.160 --> 1:16:05.160
 that doesn't overpower us, but that empowers us.

1:16:05.160 --> 1:16:08.560
 And think of the many various ways that can do that,

1:16:08.560 --> 1:16:11.000
 whether that's from my side of the world

1:16:11.000 --> 1:16:12.720
 of autonomous vehicles.

1:16:12.720 --> 1:16:14.720
 I'm personally actually from the camp

1:16:14.720 --> 1:16:16.800
 that believes this human level intelligence

1:16:16.800 --> 1:16:20.480
 is required to achieve something like vehicles

1:16:20.480 --> 1:16:23.880
 that would actually be something we would enjoy using

1:16:23.880 --> 1:16:25.120
 and being part of.

1:16:25.120 --> 1:16:27.040
 So that's one example, and certainly there's a lot

1:16:27.040 --> 1:16:30.920
 of other types of robots and medicine and so on.

1:16:30.920 --> 1:16:33.880
 So focusing on those and then coming up with the obstacles,

1:16:33.880 --> 1:16:35.920
 coming up with the ways that that can go wrong

1:16:35.920 --> 1:16:38.160
 and solving those one at a time.

1:16:38.160 --> 1:16:41.520
 And just because you can build an autonomous vehicle,

1:16:41.520 --> 1:16:42.800
 even if you could build one

1:16:42.800 --> 1:16:45.080
 that would drive just fine without you,

1:16:45.080 --> 1:16:46.720
 maybe there are some things in life

1:16:46.720 --> 1:16:48.400
 that we would actually want to do ourselves.

1:16:48.400 --> 1:16:49.240
 That's right.

1:16:49.240 --> 1:16:51.400
 Right, like, for example,

1:16:51.400 --> 1:16:53.040
 if you think of our society as a whole,

1:16:53.040 --> 1:16:56.320
 there are some things that we find very meaningful to do.

1:16:57.200 --> 1:16:59.640
 And that doesn't mean we have to stop doing them

1:16:59.640 --> 1:17:02.000
 just because machines can do them better.

1:17:02.000 --> 1:17:04.080
 I'm not gonna stop playing tennis

1:17:04.080 --> 1:17:07.360
 just the day someone builds a tennis robot and beat me.

1:17:07.360 --> 1:17:09.600
 People are still playing chess and even go.

1:17:09.600 --> 1:17:14.600
 Yeah, and in the very near term even,

1:17:14.600 --> 1:17:18.880
 some people are advocating basic income, replace jobs.

1:17:18.880 --> 1:17:20.840
 But if the government is gonna be willing

1:17:20.840 --> 1:17:24.040
 to just hand out cash to people for doing nothing,

1:17:24.040 --> 1:17:25.840
 then one should also seriously consider

1:17:25.840 --> 1:17:27.640
 whether the government should also hire

1:17:27.640 --> 1:17:29.480
 a lot more teachers and nurses

1:17:29.480 --> 1:17:32.160
 and the kind of jobs which people often

1:17:32.160 --> 1:17:34.440
 find great fulfillment in doing, right?

1:17:34.440 --> 1:17:36.320
 We get very tired of hearing politicians saying,

1:17:36.320 --> 1:17:39.320
 oh, we can't afford hiring more teachers,

1:17:39.320 --> 1:17:41.480
 but we're gonna maybe have basic income.

1:17:41.480 --> 1:17:44.000
 If we can have more serious research and thought

1:17:44.000 --> 1:17:46.200
 into what gives meaning to our lives,

1:17:46.200 --> 1:17:48.960
 the jobs give so much more than income, right?

1:17:48.960 --> 1:17:50.520
 Mm hmm.

1:17:50.520 --> 1:17:53.320
 And then think about in the future,

1:17:53.320 --> 1:17:58.320
 what are the roles that we wanna have people

1:18:00.000 --> 1:18:03.040
 continually feeling empowered by machines?

1:18:03.040 --> 1:18:06.120
 And I think sort of, I come from Russia,

1:18:06.120 --> 1:18:07.240
 from the Soviet Union.

1:18:07.240 --> 1:18:10.160
 And I think for a lot of people in the 20th century,

1:18:10.160 --> 1:18:14.080
 going to the moon, going to space was an inspiring thing.

1:18:14.080 --> 1:18:18.080
 I feel like the universe of the mind,

1:18:18.080 --> 1:18:20.880
 so AI, understanding, creating intelligence

1:18:20.880 --> 1:18:23.240
 is that for the 21st century.

1:18:23.240 --> 1:18:24.400
 So it's really surprising.

1:18:24.400 --> 1:18:25.640
 And I've heard you mention this.

1:18:25.640 --> 1:18:27.400
 It's really surprising to me,

1:18:27.400 --> 1:18:29.240
 both on the research funding side,

1:18:29.240 --> 1:18:31.760
 that it's not funded as greatly as it could be,

1:18:31.760 --> 1:18:34.760
 but most importantly, on the politician side,

1:18:34.760 --> 1:18:36.520
 that it's not part of the public discourse

1:18:36.520 --> 1:18:40.800
 except in the killer bots terminator kind of view,

1:18:40.800 --> 1:18:44.880
 that people are not yet, I think, perhaps excited

1:18:44.880 --> 1:18:46.680
 by the possible positive future

1:18:46.680 --> 1:18:48.120
 that we can build together.

1:18:48.120 --> 1:18:51.520
 So we should be, because politicians usually just focus

1:18:51.520 --> 1:18:53.320
 on the next election cycle, right?

1:18:54.480 --> 1:18:57.160
 The single most important thing I feel we humans have learned

1:18:57.160 --> 1:18:59.320
 in the entire history of science

1:18:59.320 --> 1:19:02.040
 is they were the masters of underestimation.

1:19:02.040 --> 1:19:07.040
 We underestimated the size of our cosmos again and again,

1:19:08.480 --> 1:19:10.200
 realizing that everything we thought existed

1:19:10.200 --> 1:19:12.240
 was just a small part of something grander, right?

1:19:12.240 --> 1:19:16.640
 Planet, solar system, the galaxy, clusters of galaxies.

1:19:16.640 --> 1:19:17.560
 The universe.

1:19:18.440 --> 1:19:23.120
 And we now know that the future has just

1:19:23.120 --> 1:19:25.160
 so much more potential

1:19:25.160 --> 1:19:27.640
 than our ancestors could ever have dreamt of.

1:19:27.640 --> 1:19:32.360
 This cosmos, imagine if all of Earth

1:19:33.600 --> 1:19:35.440
 was completely devoid of life,

1:19:36.640 --> 1:19:38.520
 except for Cambridge, Massachusetts.

1:19:39.560 --> 1:19:42.680
 Wouldn't it be kind of lame if all we ever aspired to

1:19:42.680 --> 1:19:45.560
 was to stay in Cambridge, Massachusetts forever

1:19:45.560 --> 1:19:47.160
 and then go extinct in one week,

1:19:47.160 --> 1:19:49.760
 even though Earth was gonna continue on for longer?

1:19:49.760 --> 1:19:52.800
 That sort of attitude I think we have now

1:19:54.200 --> 1:19:57.800
 on the cosmic scale, life can flourish on Earth,

1:19:57.800 --> 1:20:00.840
 not for four years, but for billions of years.

1:20:00.840 --> 1:20:02.920
 I can even tell you about how to move it out of harm's way

1:20:02.920 --> 1:20:04.840
 when the sun gets too hot.

1:20:04.840 --> 1:20:09.520
 And then we have so much more resources out here,

1:20:09.520 --> 1:20:12.480
 which today, maybe there are a lot of other planets

1:20:12.480 --> 1:20:14.960
 with bacteria or cow like life on them,

1:20:14.960 --> 1:20:19.880
 but most of this, all this opportunity seems,

1:20:19.880 --> 1:20:22.440
 as far as we can tell, to be largely dead,

1:20:22.440 --> 1:20:23.560
 like the Sahara Desert.

1:20:23.560 --> 1:20:28.480
 And yet we have the opportunity to help life flourish

1:20:28.480 --> 1:20:30.280
 around this for billions of years.

1:20:30.280 --> 1:20:32.680
 So let's quit squabbling about

1:20:34.080 --> 1:20:36.480
 whether some little border should be drawn

1:20:36.480 --> 1:20:38.440
 one mile to the left or right,

1:20:38.440 --> 1:20:41.080
 and look up into the skies and realize,

1:20:41.080 --> 1:20:44.040
 hey, we can do such incredible things.

1:20:44.040 --> 1:20:46.640
 Yeah, and that's, I think, why it's really exciting

1:20:46.640 --> 1:20:49.440
 that you and others are connected

1:20:49.440 --> 1:20:51.880
 with some of the work Elon Musk is doing,

1:20:51.880 --> 1:20:54.480
 because he's literally going out into that space,

1:20:54.480 --> 1:20:57.000
 really exploring our universe, and it's wonderful.

1:20:57.000 --> 1:21:02.000
 That is exactly why Elon Musk is so misunderstood, right?

1:21:02.000 --> 1:21:05.000
 Misconstrued him as some kind of pessimistic doomsayer.

1:21:05.000 --> 1:21:07.640
 The reason he cares so much about AI safety

1:21:07.640 --> 1:21:12.080
 is because he more than almost anyone else appreciates

1:21:12.080 --> 1:21:14.280
 these amazing opportunities that we'll squander

1:21:14.280 --> 1:21:16.640
 if we wipe out here on Earth.

1:21:16.640 --> 1:21:19.680
 We're not just going to wipe out the next generation,

1:21:19.680 --> 1:21:23.320
 all generations, and this incredible opportunity

1:21:23.320 --> 1:21:25.400
 that's out there, and that would really be a waste.

1:21:25.400 --> 1:21:30.080
 And AI, for people who think that it would be better

1:21:30.080 --> 1:21:33.600
 to do without technology, let me just mention that

1:21:34.680 --> 1:21:36.320
 if we don't improve our technology,

1:21:36.320 --> 1:21:39.320
 the question isn't whether humanity is going to go extinct.

1:21:39.320 --> 1:21:41.160
 The question is just whether we're going to get taken out

1:21:41.160 --> 1:21:44.800
 by the next big asteroid or the next super volcano

1:21:44.800 --> 1:21:48.280
 or something else dumb that we could easily prevent

1:21:48.280 --> 1:21:49.840
 with more tech, right?

1:21:49.840 --> 1:21:53.160
 And if we want life to flourish throughout the cosmos,

1:21:53.160 --> 1:21:54.760
 AI is the key to it.

1:21:56.120 --> 1:21:59.840
 As I mentioned in a lot of detail in my book right there,

1:21:59.840 --> 1:22:04.840
 even many of the most inspired sci fi writers,

1:22:04.880 --> 1:22:08.120
 I feel have totally underestimated the opportunities

1:22:08.120 --> 1:22:11.240
 for space travel, especially at the other galaxies,

1:22:11.240 --> 1:22:15.360
 because they weren't thinking about the possibility of AGI,

1:22:15.360 --> 1:22:17.520
 which just makes it so much easier.

1:22:17.520 --> 1:22:18.440
 Right, yeah.

1:22:18.440 --> 1:22:23.440
 So that goes to your view of AGI that enables our progress,

1:22:24.080 --> 1:22:25.760
 that enables a better life.

1:22:25.760 --> 1:22:28.320
 So that's a beautiful way to put it

1:22:28.320 --> 1:22:29.960
 and then something to strive for.

1:22:29.960 --> 1:22:31.440
 So Max, thank you so much.

1:22:31.440 --> 1:22:32.560
 Thank you for your time today.

1:22:32.560 --> 1:22:33.560
 It's been awesome.

1:22:33.560 --> 1:22:34.400
 Thank you so much.

1:22:34.400 --> 1:22:35.240
 Thanks.

1:22:35.240 --> 1:22:40.240
 Have a great day.