Making Sense - Sam Harris - July 01, 2017


#84 — Landscapes of Mind


Episode Stats

Length

45 minutes

Words per Minute

161.91957

Word Count

7,396

Sentence Count

298

Hate Speech Sentences

6


Summary

In this episode of the Making Sense Podcast, I speak with Kevin Kelly about his new book, The Inevitable, about the 12 technological forces that will shape our future. We discuss AI, the nature of consciousness, and the role of technology in shaping the world, and how we need to embrace these things in order to steer the many ways in which we do have control and influence over them. We don t run ads on the podcast and therefore, therefore, are made possible entirely through the support of our listeners. If you enjoy what we re doing here, please consider becoming a supporter of what we're doing here by becoming a subscriber. You'll get access to all kinds of premium features, including ad-free versions of the podcast, as well as access to our most popular podcasting platform, The Huffington Post, where you can read and subscribe to all sorts of news and discussion about the happenings around the world of culture, politics, technology, and culture. And, of course, there's plenty of time to catch up on the latest episodes of Making Sense! Subscribe to the podcast on your favorite podcast platform, wherever you get your podcasts, if you're listening to the thing you care about things that matter. Thanks for listening! and Happy Listening! Make sense! Sam Harris and Kevin Kelly Make Sense - The Making Sense podcast. - Sam Harris Music: "Space Traveler" by Jeff Kaale (ft. John Singleton ( ) Art: "Out of Control" by Ian Dorsen ( ) "Goodbye" by Kevin Kelly ( ) and "Space Junk (feat. by by The Good Morning and ( ) by The Lonely Planet ( ) - "The Good Morning America" by John Rigsby ( ) is out on the Good Morning Podcast ( ) (featuring John Rocha ( ) & & by Shadydave ( ) . by Jeff Perla ( , "Good Morning" ( ) , . , and , "Good Luck" by , & ( by James Gray ( ) (feat., ) ( ) with is out of New York Magazine ( . , ) and (c) & , Thank You by Bill Simmons ( ) -- Thank You, Kevin Kelly? ( ), and .


Transcript

00:00:00.000 Welcome to the Making Sense Podcast.
00:00:08.820 This is Sam Harris.
00:00:10.880 Just a note to say that if you're hearing this, you are not currently on our subscriber
00:00:14.680 feed and will only be hearing the first part of this conversation.
00:00:18.420 In order to access full episodes of the Making Sense Podcast, you'll need to subscribe at
00:00:22.720 samharris.org.
00:00:24.060 There you'll find our private RSS feed to add to your favorite podcatcher, along with
00:00:28.360 other subscriber-only content.
00:00:30.520 We don't run ads on the podcast, and therefore it's made possible entirely through the support
00:00:34.640 of our subscribers.
00:00:35.800 So if you enjoy what we're doing here, please consider becoming one.
00:00:46.340 Today I'm speaking with Kevin Kelly.
00:00:49.020 Kevin helped launch Wired Magazine and was its executive editor for his first seven years.
00:00:54.660 So he knows a thing or two about digital media.
00:00:58.440 And he's written for the New York Times, The Economist, Science, Time Magazine, The Wall
00:01:05.360 Street Journal, and many other publications.
00:01:08.920 His previous books include Out of Control, New Rules for the New Economy, Cool Tools, and
00:01:16.680 What Technology Wants.
00:01:18.100 And his most recent book is The Inevitable, Understanding the Twelve Technological Forces
00:01:24.700 That Will Shape Our Future.
00:01:27.220 And Kevin and I focused on this book, and then spent much of the conversation talking
00:01:32.840 about AI, the safety concerns around it, the nature of intelligence, the concept of the
00:01:40.780 singularity, the prospect of artificial consciousness, and the ethical implications of that.
00:01:48.120 And it was great.
00:01:49.020 We don't agree about everything, but I really enjoyed the conversation.
00:01:53.120 And I hope you enjoy it as much as I did.
00:01:56.600 And now I bring you Kevin Kelly.
00:02:05.060 I am here with Kevin Kelly.
00:02:06.840 Kevin, thanks for coming on the podcast.
00:02:08.200 Oh, man, I'm enjoying this right now.
00:02:11.960 Listen, so many people have asked for you, and obviously, you know, I've known you and
00:02:17.140 about you for many years.
00:02:18.960 I'll talk about how we first met at some point.
00:02:21.900 You are so on top of recent trends that are subsuming everyone's lives that it's just great
00:02:28.340 to get a chance to talk to you.
00:02:29.860 Well, thanks for having me.
00:02:30.540 So before we jump into all these common topics of interest, how would you describe what you
00:02:37.480 do?
00:02:38.200 I package ideas.
00:02:41.860 And they're often visual packages, but I like to take ideas, not necessarily my ideas,
00:02:49.880 but other people's ideas, and present them in some way.
00:02:53.740 And that kind of is what I did with magazines, beginning with the Holworth Review, formerly
00:03:00.200 called Co-Evolution Coralie, the Holworth Catalogues, Wired, websites like Cool Tools, and my books.
00:03:07.800 So you've written these two recent books on technology, what technology wants, and your
00:03:14.080 most recent one, The Inevitable.
00:03:15.360 How would you summarize the arguments you put forward in those books?
00:03:19.520 At one level, I'm actually trying to devise a proto-theory of technology.
00:03:28.240 So before Darwin's theory of biology, the evolutionary theory, there was a lot of naturals and they
00:03:36.540 had these curiosity cabinets where they would just collect biological specimens and there
00:03:42.740 was just one weird creature after another.
00:03:45.660 There was no framework for understanding how they were related or how they came about.
00:03:51.760 And in many ways, technology is like that with us.
00:03:55.260 We have this sort of parade of one invention after another and there's really no theory about
00:04:00.900 how these different species of technology are related and how they come together.
00:04:09.000 So at one level, my books were trying to devise a rough theory of their origins and perhaps
00:04:19.500 no surprise, cutting to the punchline, I see these as an extension and acceleration of the
00:04:28.200 same forces that are at work in natural evolution, or cosmic evolution for that matter.
00:04:34.480 And that if you look at it in that way, this system of technology that I call the tech name
00:04:43.340 is in some ways the extension and acceleration of the self-organizing forces that are running
00:04:49.100 through the cosmos.
00:04:50.400 So that's one thing that I'm trying to do.
00:04:53.640 And the second thing I'm trying to do is to say that there is a deterministic element
00:04:59.700 in this, both in evolution and in technological systems.
00:05:06.120 And a lot of, at the very high level, a lot of what we're going to see and have seen is
00:05:11.500 following kind of a natural progression.
00:05:13.940 And so therefore is inevitable and that we as humans, individuals and corporately need
00:05:20.160 to embrace these things in order to be able to steer the many ways in which we do have
00:05:27.840 control and choice, the character of these.
00:05:30.800 So I would say like the, once you invented electrical wires and you invented switches and stuff, you'd
00:05:36.000 have telephones.
00:05:37.140 And so the telephone was inevitable, but the character of the telephone was not inevitable.
00:05:41.900 You know, iPhone was not inevitable and, and we have a lot of choices about those, but
00:05:48.040 the only way we make those choices is by embracing and using these things rather than prohibiting
00:05:52.560 them.
00:05:53.800 So now you start the book, The Inevitable, with some very amusing stories about how clueless
00:05:59.780 people were about the significance of the internet in particular.
00:06:02.860 I was vaguely aware of some of these howlers, but you just wrap them all up in one paragraph
00:06:08.500 and it's, it's amazing how blind people were to what was coming.
00:06:13.580 So you, you, you cite Time and Newsweek saying that, that more or less the internet would amount
00:06:18.360 to nothing.
00:06:19.160 One network executive said it would be the CB radio of the nineties.
00:06:23.220 There was a wired writer who bought the domain name for McDonald's, McDonald's.com and couldn't
00:06:29.540 give it away to McDonald's because they couldn't see why it would ever be valuable to them.
00:06:33.740 Now I don't recall being quite that clueless myself, but I, I'm, I'm continually amazed
00:06:40.900 at my inability to see what's coming here.
00:06:44.540 And I mean, if you had told me five years ago that I would soon be spending much of my
00:06:49.320 time podcasting, I would have said, what's a podcast.
00:06:51.920 And if you had told me what a podcast was, essentially describing it as on demand radio,
00:06:57.640 I would have been absolutely certain that there was no way I was going into radio.
00:07:01.500 Just, it would not apply.
00:07:03.320 I feel personally, no ability to see what's coming.
00:07:07.680 How, why do you think it is so difficult for most people to see into the, even the very
00:07:12.260 near future here?
00:07:13.920 Yeah, it's a really good question.
00:07:16.580 It's, I don't, I don't think I have a good answer about why we find it hard to imagine the
00:07:22.040 future, but it is true that the more we know about the, in other words, the experts in a
00:07:30.200 certain field are often the ones who are most blinded by the changes.
00:07:34.640 We did this thing at Wired called reality check and we would poll different experts and non-experts
00:07:41.900 in some future things like, you know, whether they're going to use like laser drilling in
00:07:47.160 dentistry or, you know, flying cars and stuff like that.
00:07:53.480 And, and they would have dates.
00:07:54.860 And when these came around later on in the future, it was the experts who were always
00:08:01.460 underestimating, who, who, who are, I guess, overestimating when things were going to happen.
00:08:07.120 There was, they were more pessimistic and it was sort of the people who, so the people who
00:08:11.000 knew the most about things were often the ones that were most wrong.
00:08:16.700 And, um, so I think, I, I think it's kind of like we know too much and we, um, find it
00:08:24.540 hard to release and believe things that seemed impossible.
00:08:30.380 Um, so, so, uh, the other observation that I would make about the things that have surprised
00:08:36.520 me the most in, in the last 30 years, and I think the things that will continue to surprise
00:08:42.220 us in the next 30 years all have to do with the fact that the things that are most surprising
00:08:49.820 are actually things are done in collaboration at a scale that we've not seen before, like
00:08:55.920 things like Wikipedia, Facebook, or even cell phones and smartphones to some extent, that
00:09:02.120 basically we are kind of organizing work and collaboration at a scale that was just really
00:09:09.480 unthinkable before.
00:09:11.920 And that's where a lot of these surprises have been originating is this, this, the, the,
00:09:17.840 our ability to collaborate in real time and scales that, that were just unthinkable before.
00:09:26.360 And so they seemed impossible.
00:09:27.780 Um, and, um, for me, most of these surprises have, have been in, have had that connection.
00:09:35.320 Well, I, I know you and I want to talk about AI because I think that's, that's an area where
00:09:39.760 we'll find some, I think, significant overlap, but also some disagreement.
00:09:44.460 And I want to spend most of our time talking about that, but I do want to touch on some of
00:09:50.020 the, the issues you raise in, in the inevitable, because you, you divide the book into these,
00:09:55.500 these 12 trends.
00:09:57.920 I'm sure some of those will come back around in our discussion of AI, but take an example
00:10:02.260 of, I mean, let's say this podcast.
00:10:04.540 I mean, one, one change that a podcast represents over radio is that it's, it's on demand.
00:10:10.720 You can listen to it whenever you want to listen to it.
00:10:12.980 It's instantly accessible.
00:10:14.880 In this case, it's free.
00:10:16.440 So there's no, there's no barrier to listening to it.
00:10:19.580 People can slice it and dice it in any way they want.
00:10:23.360 They, they, people remix it, that people have taken snippets of it and put it behind music.
00:10:28.740 So it becomes the basis for other people's creativity.
00:10:32.780 Ultimately, I would imagine all the audio that exists and all the video that exists will
00:10:37.740 be searchable in a way that text is currently searchable, which is, that's a real weakness
00:10:42.280 now.
00:10:42.740 But eventually you'll be able to search and get exactly the snippet of audio you want.
00:10:48.100 And this change in just this one domain of, of how people listen to a conversation that
00:10:54.700 captures some of these trends, right?
00:10:56.720 Exactly.
00:10:57.460 So there was the, the flow or the, the, the, the verb of the remixing was to your point,
00:11:03.200 the fact that, um, that was the big, the big change in music, which the music companies
00:11:08.620 didn't kind of understand that they thought that the free aspect of downloads of these files
00:11:14.140 was because people wanted to cheat them and get things for free.
00:11:17.820 But the, the, the, the chief value was that the, the freeze and freedom is that people
00:11:22.760 could take these music files, they could get less than an album, they could kind of remix
00:11:27.200 them into singles.
00:11:27.940 They could then manipulate them, make them into playlists.
00:11:31.740 They could do all these things that make it much more fluid and liquid, um, and manipulable
00:11:38.380 and fungible.
00:11:39.360 And, um, and that was the great attraction for people.
00:11:43.420 The fact that it was, doesn't cost anything was sort of a bonus that wasn't the main event
00:11:48.620 and that all the other things that you mentioned about this aspect of podcasts, of getting them
00:11:54.300 on demand, the shift from owning things to having access to things.
00:11:58.880 If you have instant, um, access, uh, anytime, anywhere in the world, um, that's part of the
00:12:06.020 shift there, the, the, the shift, um, away from things that are static and, um, monumental
00:12:14.280 to things that are incomplete and always in the process.
00:12:19.040 This, the movement from centralized to decentralized is also made possible when you have things
00:12:25.760 in real time, you know, when you're in a world of like the Roman error, when you, uh,
00:12:31.920 where it's very little information flows, the best way to organize an army was to have people
00:12:38.200 give a command at the top and everybody below would, would follow it because the commander
00:12:43.900 had the most information, but in a world in which information flows liquidly and pervasively
00:12:51.100 everywhere, then a decentralized system is much more powerful, uh, because you can actually,
00:12:57.900 um, have the edges and steer as well as the center and center becomes less important.
00:13:05.060 And so all these things are feeding into it.
00:13:07.500 And your example of, of the podcast is just a perfect example where all these trends in
00:13:13.460 general conspire to make this a new genre.
00:13:17.540 And I would say in the future, we would continue to remix the elements inside a podcast and that
00:13:27.360 we would, you know, have, um, podcasts within VR that will have, um, podcasts, as you said,
00:13:34.100 that are searchable and have AI, um, remix portions of it, or that we would, you know, begin to do
00:13:42.900 all the things that we've done with texts and annotations and footnoting would be brought to
00:13:48.500 this as well.
00:13:49.160 So if you just imagine what we've done with podcasts and now multiply that by every other
00:13:54.900 medium from GIFs to YouTube, um, we're entering into an era where we're going to have, um, entirely
00:14:06.180 brand new genres of art, expression, and media.
00:14:13.120 And we're just, again, at the beginning of this process.
00:14:16.560 What do you think about the, the new capacity to fake media?
00:14:20.480 So now I think you, you must've seen this, I think it was a TED talk initially where I saw it,
00:14:24.460 but it's been unveiled in various formats now where they can fake audio so well that given
00:14:32.080 the sample that we've just given them, they could, someone could produce a fake conversation
00:14:36.080 between us where we said all manner of reputation destroying things.
00:14:41.000 And it wouldn't be us, but it would be, I think by current technology, undetectable as a fraud.
00:14:47.440 And I think there, there are now video versions of this where you can get someone's mouth to move
00:14:52.400 in the appropriate way.
00:14:53.860 So it looks like they're delivering the fake audio, although the facial display is not totally
00:14:58.940 convincing yet, but presumably it will be at some point.
00:15:02.080 What do you think about that?
00:15:03.380 There's, I've, I've, in a hand-waving way, not really knowing what I'm talking about,
00:15:07.480 I've imagined there must be some blockchain-based way of ensuring against that.
00:15:13.380 But, uh, where are we going with that?
00:15:15.100 So, so, um, in, I don't know, 1984 or something, I did a cover story for the Holworth Review of CQ,
00:15:24.340 I think it was called at the time.
00:15:26.360 Um, it was called Photography as the End of Evidence for Anything.
00:15:29.900 And we were, we used a very expensive, um, Cytex machine.
00:15:36.220 It was like multi-million dollar machine, which cost, uh, tens of thousands of dollars an hour
00:15:41.460 to basically what we would now call Photoshop.
00:15:44.680 This is the early Photoshop.
00:15:46.160 So, you know, National Geographic and Time and Life magazine had access to things
00:15:51.080 and they would do little retouching stuff.
00:15:52.680 But we decided to Photoshop, uh, flying saucers arriving in San Francisco.
00:15:59.900 And, um, the point of this article was that, okay, this was the beginning of using photography
00:16:05.480 as the evidence of anything.
00:16:06.920 And what I kind of, uh, concluded back then was that the only, well, there's two things.
00:16:13.660 One was, um, the primary evidence of believability was simply going to be the reputation of the source.
00:16:21.580 So, for most people, you wouldn't be able to tell.
00:16:25.000 And that we already have that thing with text, all right?
00:16:28.060 I mean, it's like words, you know, you could quote somebody, you can say, put some words
00:16:32.520 and say, Sam Harris says this and it would look just like it was real.
00:16:36.920 Yes, it's been done.
00:16:37.320 Yeah.
00:16:37.720 Exactly.
00:16:38.260 So how would you know?
00:16:39.040 Well, the only way you could know was basically you have to trust the, the, the source and
00:16:43.960 the same thing was going to happen with photography.
00:16:46.040 And now it'll be with video and audio.
00:16:49.160 And so they're coming up to this place where text is, which is basically you can only rely
00:16:54.680 on the source.
00:16:55.340 The second thing we discovered from this was that, and this also kind of applied to this
00:17:00.240 question of like, when you have AI and agents, how would you be able to tell if they're human
00:17:04.340 or not?
00:17:04.820 And the thing is, is that for most cases, like in a movie right now, you can't tell whether
00:17:12.440 something has been CGI, whether it's real actor or not, we're, we've already left that
00:17:18.560 behind and, but we don't care in a certain sense.
00:17:22.360 And when we call up on a phone and there's a robot, an agent there, and we're trying to
00:17:28.820 do a service problem, in some ways we don't really care whether it's a human or not.
00:17:32.980 If they're giving us good service, but in the cases where we do care, there will always
00:17:38.980 be ways to tell, and they may cost money.
00:17:42.600 There's forensic ways to, to really come decide whether this photograph has been doctored,
00:17:50.380 whether, um, a CGI is actually been used to, to make a frame, whether this audio file has
00:17:57.880 been altered, there, there will always be some way if you really, really care, but in most
00:18:03.280 cases we won't care.
00:18:05.080 And we will just have to rely on the reputation of, of the source.
00:18:10.140 And so, um, I think we're going to kind of get to this, to the place where text is already,
00:18:15.440 which is the same thing.
00:18:17.540 If, if someone's making it up, then you have no way to tell by looking at the text, you
00:18:21.620 have to go back to the, to the source.
00:18:23.760 But that doesn't address the issue of fake news.
00:18:26.500 And for, for, for, for there, I think what we're going to see is a, like a truth signaling
00:18:32.200 layer added on somewhat, maybe using AI, but mostly to devise what I would think is going
00:18:39.440 to be kind of like a, a probability index to a statement that would be made in a networked
00:18:44.620 way rather than it'll, it'll involve Wikipedia and Snopes and, and in places, you know, maybe
00:18:49.980 other academics, but it would be like page rank, meaning that you'll have a statement, you
00:18:55.800 know, um, London is the capital of England.
00:18:58.840 They'll be like, that's, that statement has a 95% probability or 98% probability of being
00:19:04.800 true.
00:19:05.960 And then other statements will have a 50% probability of being true and others will have a 10% probability.
00:19:11.580 And that will come out of a networked analysis of these, these sites or these, you know, the
00:19:17.840 Encyclopedia Britannica or whatever it says.
00:19:19.880 So, so these other sources have a high reliability because in the past they had been true.
00:19:24.620 And this, this, this network of, of, uh, corresponding sources, which are ranked themselves by other
00:19:34.100 sources in terms of their reliability will generate some index number to a statement.
00:19:41.180 And as the statements get more complex, that's a, becomes a more difficult job to do.
00:19:45.760 And that's where the AI could become involved in trying to detect the pattern out of, um, all
00:19:52.480 these sources.
00:19:53.240 And so, um, you'll get a probability score of, of this statement is likely truthfulness.
00:20:03.460 That's kind of like a prediction market for epistemology.
00:20:06.160 Yes.
00:20:06.900 That's interesting.
00:20:07.820 So in, in light of what's happening and the trends you discuss in, in the inevitable, if
00:20:14.320 you had a child going to college next year, what would you hope that he or she study and
00:20:20.680 or ignore in light of what opportunities will soon exist?
00:20:25.800 One of the things I talk about in the book is this idea that we're all in going to be
00:20:29.080 perpetual newbies, no matter whether we're 60 or 16 or six, that, um, we're feeling very
00:20:35.880 good that we've mastered, you know, smartphones and we know laptops, but the gestures and how
00:20:41.340 things work, this kind of literacy, but, you know, in five years from now, there'll be a
00:20:45.780 new platform, virtual reality, whatever it might be.
00:20:48.660 And we'll have to learn another set of gestures and commands and logic.
00:20:54.120 And so the, the digital natives right now have a past because they, uh, are doing with technology
00:21:03.380 that was invented, um, after they were born, but, but eventually, um, they're going to have
00:21:08.920 to learn new things too.
00:21:09.860 And they're going to be in the same position as the old folks of having to learn these things.
00:21:14.020 They're going to be newbies again, too.
00:21:15.660 So we're all going to be perpetual newbies.
00:21:18.380 And I think the really only literacy or skill that should be taught in schools is so that
00:21:26.300 when you graduate, you have learned how to learn.
00:21:29.920 So learning how to learn is, is the, the, the, the meta skill that you want to have.
00:21:35.540 And really, I think the only one that makes any difference because whatever language you're
00:21:39.980 going to learn is not necessarily going to be the one that you are going to get paid for
00:21:44.100 knowledge.
00:21:45.340 If you want an answer, you ask a machine.
00:21:47.360 So I, I, I think this idea of, of learning how to learn is the real skill that you should
00:21:56.480 graduate with.
00:21:57.300 And for extra bonus for, for, for the ultimate golden pass, if you can learn how you learn
00:22:05.880 best yourself, if you can optimize your own style of learning, that's the superpower that
00:22:11.240 you want that I think almost takes a lifetime to get to.
00:22:15.440 And some people like Tim Ferriss are much better at dissecting how they learn and understanding
00:22:20.180 how they can optimize their self-learning.
00:22:22.040 But if you can get to that state where you have really, really understand how you personally
00:22:29.220 learn best, then, then you're golden.
00:22:32.500 And I think that's what we want to aim for is that every person on the planet today will
00:22:39.480 learn how to learn and will optimize how they learn best.
00:22:43.780 And that, I think, is what schools should really be aiming for.
00:22:48.520 Yeah, I was going to say our mutual friend, Tim, seems well poised to take advantage of
00:22:52.740 the future.
00:22:53.720 I'm just going to have to keep track of him.
00:22:55.040 Tim Ferriss Let's talk about AI.
00:22:57.260 I want to, I'll set this up by just how this, this podcast got initiated because though I, I,
00:23:02.900 I long knew that I wanted you on the podcast.
00:23:05.440 You recently sent me an email after hearing my podcast on robot ethics with Kate Darling.
00:23:12.360 And in that email, you, you sketched ways where you think you and I disagree about the implications
00:23:18.960 and, and safety concerns of AI.
00:23:21.520 You were also reacting to my TED talk on the topic and also a panel discussion that you saw
00:23:28.500 where I was on stage with, with Max Tegmark and Elon Musk and Jan Talon and other people
00:23:33.860 who were at this conference on, on AI at Asilomar earlier this year.
00:23:38.500 In, you, you wrote in the setup to this email, and now I'm quoting you, there are at least
00:23:43.080 five assumptions the super AI crowd hold that I can't find any evidence to support.
00:23:48.340 In contradistinction to this orthodoxy, I find the following five heresies to have more evidence.
00:23:54.220 One, intelligence is not a single dimension.
00:23:57.420 So, quote, smarter than humans is a meaningless concept.
00:24:01.420 Two, humans do not have general purpose minds and neither will AIs.
00:24:06.080 Three, emulation of human thinking will be constrained by cost.
00:24:11.840 Four, dimensions of intelligence are not infinite.
00:24:15.880 And five, intelligences are, are only one factor in progress.
00:24:20.820 Now, I think these are all interesting claims, and I think I agree with several of them, but
00:24:26.000 most of them don't actually touch what concerns me about AI.
00:24:31.520 So, I think we should talk about all of these claims, because I think they get at interesting
00:24:35.720 points.
00:24:36.920 But I think I should probably start by just summarizing what my main concern is about AI.
00:24:42.020 So, we can, as we talk about your points, we can also just make sure we're, we're hitting
00:24:46.540 that.
00:24:47.480 And, you know, you, when you talk about AI and when you talk about this one trend in your
00:24:53.180 book, perhaps the most relevant, cognifying, you know, essentially putting intelligence into
00:24:57.760 everything that can be made intelligent, you can sound very utopian, and I can sound very
00:25:03.640 dystopian in, in how I talk about it.
00:25:06.060 So, but I actually think we, we overlap a fair amount.
00:25:09.400 I guess my main concern can be summarized under the, the heading of the alignment problem,
00:25:15.960 which is now kind of a phrase of jargon among those of us who are worried about AI gone wrong.
00:25:22.400 And there are really two concerns here with AI, and, and, and I think that they're, they're
00:25:29.760 concerns that they're visited on any powerful technology.
00:25:33.540 And the first is just the obvious case of people using it intentionally in ways that cause
00:25:39.800 great harm.
00:25:40.500 So, it's just the kind of the bad people problem.
00:25:42.460 And that's, that's obviously a real problem.
00:25:45.040 It's a problem that probably never goes away, but it's not the interesting problem here.
00:25:49.840 I think that the, the interesting problem is the unintended consequences problem.
00:25:53.360 So, it's the situation where even good people with the best of intentions can wind up committing
00:25:59.220 great harms because the technology is such that it's not, it won't reliably conform to
00:26:05.740 the best intentions of good people.
00:26:08.040 So, for, for a powerful technology to be safe or, you know, or, or to be operating within our
00:26:14.980 risk tolerance, it has to be the sort of thing that good people can reliably do good things
00:26:20.320 with it rather than accidentally end civilization or, or do something else that's terrible.
00:26:26.060 And for that, for this to happen with AI, it's going to have to be aligned with our values.
00:26:33.560 And so, again, this is often called the, the alignment problem.
00:26:36.120 When you have autonomous systems working in ways and increasingly powerful systems and
00:26:42.160 ultimately systems that are more powerful than any human being and even any collection
00:26:46.520 of human beings, you need to solve this, this alignment problem.
00:26:51.120 But at this point, people who haven't thought about this very much get confused or, or at least
00:26:57.880 they wonder, you know, why on earth would an AI, however powerful, fail to be aligned with
00:27:04.640 our values, because after all, we, we built these things or we will build these things.
00:27:09.200 And they imagine a kind of silly Terminator style scenario where just, you know, robot armies
00:27:15.900 start attacking us because for some reason they have started to hate us and, and want to kill
00:27:20.680 us.
00:27:21.060 And that, that really isn't the issue that, that even the most dystopian people are, are thinking
00:27:27.000 about.
00:27:27.440 And it's not, it's not the issue I'm thinking about.
00:27:29.160 It's, it's not that our machines will become spontaneously malevolent and want to kill us.
00:27:34.640 The issue is that they, they can become so competent at meeting their goals that if their
00:27:41.740 goals aren't perfectly aligned with our own, then the unintended consequences could be so
00:27:47.620 large as to be catastrophic.
00:27:50.000 And, and there are, there are cartoon versions of this, as you know, which more clearly dissect
00:27:56.160 the fear.
00:27:56.780 I mean, they're, they're as cartoonish as the Terminator style scenarios, but they're, they're
00:28:00.480 different.
00:28:00.820 I mean, it's something like Nick Bostrom's paperclip maximizer to review.
00:28:05.600 I think many people are familiar with this, but so Nick Bostrom imagines a machine whose
00:28:10.420 only goal is to maximize the number of paperclips in the universe, but it's a super powerful,
00:28:16.920 super competent, super intelligent machine.
00:28:19.040 And given this goal, it could quickly just decide that, you know, every atom accessible,
00:28:24.860 including the atoms in your own body are, are best suited to be turned into paperclips.
00:28:30.520 And, you know, obviously we wouldn't build precisely that machine, but the point of, of
00:28:35.620 that kind of thought experiment is to point out that these machines, even super intelligent
00:28:40.640 machines will not be like us and they'll lack common sense or they'll, or they'll only have
00:28:47.220 the common sense that we understand how to build into them.
00:28:51.920 And so the bad things that they might do might be very counterintuitive to us and therefore
00:28:57.660 totally surprising.
00:28:59.500 And just, you know, kind of the final point I'll make to set this up.
00:29:01.900 I think we're misled by the concept of intelligence.
00:29:05.720 Because when we talk about intelligence, we assume that it includes things like common
00:29:11.420 sense.
00:29:12.160 In the space of this concept, we insert something fairly anthropomorphic and, and, and familiar
00:29:19.200 to us.
00:29:19.880 But I think intelligence is more like competence or effectiveness, which is just an ability
00:29:26.780 to meet goals in an environment or across a range of environments.
00:29:32.060 And given a certain specification of goals, even a superhumanly competent machine or system
00:29:41.320 of machines might behave in ways that would strike us as completely absurd.
00:29:47.340 And yet we, we will not have closed the door to those absurdities, however dangerous, if we
00:29:53.060 don't anticipate them in advance or, or, or figure out some generic way to, to solve this alignment
00:29:58.300 problem.
00:29:58.860 Um, so I think a good place to start is where we agree.
00:30:03.060 And, um, I think where we, the first thing I think we both agree on is, is that we have
00:30:08.720 a very poor understanding of what our own intelligence is as humans.
00:30:12.880 Um, and I would, um, make a further statement that I think the common conception that we have
00:30:20.420 of IQ is a very misleading notion of intelligence and humans that, that we can kind of rank intelligences
00:30:29.340 in a relative scale, a single dimension of, you know, and this is the taken from Nick Bostrom's
00:30:34.940 own book that, you know, you have a single dimension and you have, uh, the, the intelligence
00:30:39.800 of a mouse say, or the IQ of a mouse, and then a rat's a little bit more and then that
00:30:44.240 chimpanzees a little bit more.
00:30:45.400 And then you have the kind of a really dumb human and average human, and then a super
00:30:48.900 genius like Albert Einstein.
00:30:50.320 And then there's the, the super AI, which is kind of off the charts in terms of, uh,
00:30:54.640 how much smarter along this IQ it can be.
00:30:57.920 And that, I think is a very, very misleading idea of what intelligence is.
00:31:03.960 It's obviously the human intelligence is, um, a suite, a symphony, a portfolio of dozens,
00:31:11.660 20, maybe, who knows how many different modes or nodes of, of thinking there's perception,
00:31:18.060 there's symbolic reasoning, there's a deductive reason, inductive reasoning, and emotional intelligence,
00:31:24.280 spatial navigation, long-term memory, short-term memory.
00:31:28.020 There's, there's, there's, there's many, many different nodes of thinking.
00:31:32.700 And of course, that complex varies person by person.
00:31:37.820 And, um, when we get into the animal kingdom, we have a different mixture of, of these.
00:31:44.760 And most of them are maybe simpler complexes.
00:31:47.260 Uh, but in some cases, um, they're, uh, a particular node that we might have may actually
00:31:54.740 be higher in, um, maybe superior in, in, in an animal in terms of, uh, I mean, if you've
00:32:01.100 seen some of these, um, the chimpanzee, yeah, chimpanzees, remembering the locations of numbers,
00:32:06.220 it's like, oh my gosh, obviously it's like, we're just, they're, they're, they're, they're
00:32:10.600 smarter than us and sort of in that dimension.
00:32:13.100 We should just describe that so that people are aware of what, because they should find that
00:32:16.500 video online.
00:32:17.420 What it is, is a chimpanzee has a screen and there's a series of, of, uh, numbers in sequence
00:32:23.760 or numbers that appear in different positions on the screen very, very briefly.
00:32:29.240 It's like a checkerboard with, with a suddenly illuminates with, let's say 10 different digits
00:32:33.480 and you have to select all the digits in order and select all the, you know, the, you have
00:32:38.800 to hit the right squares and the numbers then disappear.
00:32:40.780 And then you just have a blank checkerboard.
00:32:42.740 Right.
00:32:43.000 And you have to remember, you see, it sees it for like a split second and you have
00:32:46.480 to remember where they are and you have to go back and hit the locations in order.
00:32:50.620 And no human can, can do this, but this, for some reason, chimps seem to be able to do
00:32:56.900 this very easily.
00:32:57.680 So, so they have some kind of a, uh, a short-term memory or a long-term memory.
00:33:02.020 I'm not sure what kind of memory of spatial memory that does, that, that really, um, would
00:33:06.120 amaze us and we would find superhuman.
00:33:08.920 And so, um, so I think we both agree that, that, that the human intelligence is very complex.
00:33:14.600 And, and, um, my suggestion about thinking about AI is, is, is, is always to use plural,
00:33:22.080 to try to talk about AIs, because I think as we make these synthetic types of minds, we're
00:33:29.020 going to make thousands of different species of them with different combinations of these
00:33:34.860 primitive, these kind of primitive, uh, modes of thinking.
00:33:39.520 And that what we think of ourselves, our own minds, we think of our, that, that we think
00:33:45.900 of kind of as a singular intelligence.
00:33:49.580 It's very much like this, the illusion of us having an eye or being center.
00:33:54.660 There's an illusion that we have a kind of a unified universal intelligence.
00:33:58.560 But in fact, we have a, we've evolved a very, very specific, um, combination of elements
00:34:10.040 in, in thinking that are not really general purpose at all.
00:34:14.540 They're, they're, they're, they've, they've, it's a very specific purpose to survive on this
00:34:18.960 planet and in this regime of biology.
00:34:22.200 When we compare our intelligence to the space of possible intelligences, we're going to see
00:34:29.580 that we're not at the center of some universal, but we're actually at the edge, like we are
00:34:33.920 in the real galaxies of, of, of possible minds.
00:34:38.200 And what we're doing with AI is actually going to make a whole zoo of possible ways of thinking,
00:34:44.480 including inventing some ways of thinking that don't exist in biology at all today.
00:34:50.620 Just as we did with flying.
00:34:53.380 So, so, so when, the way we made, uh, artificial flying is we looked at natural flight and mostly
00:34:59.720 birds and bees and bats is, is flapping wings.
00:35:03.460 And we tried to, to, to artificially fly by flapping wings.
00:35:06.720 It just didn't work.
00:35:07.780 The way we made artificial flying is we invented a type of flight that does not exist in nature
00:35:12.400 at all, which was a fixed wing and a propeller.
00:35:14.520 And we are going to do the same thing of, of inventing ways of thinking that cannot really
00:35:21.480 occur in biology, biological tissue that will be different, a different way of thinking.
00:35:29.280 And, um, we'll combine those into maybe many, many new complexes of, of, of types of, of
00:35:37.300 thinking to do, um, and achieve different, different things.
00:35:41.400 And there may be, uh, problems that are so difficult in science or business that human
00:35:49.220 type thinking alone cannot reach that we will have to work with a two-step process of inventing
00:35:55.820 a different kind of thinking that we can then together work to solve some of these problems.
00:35:59.840 So I think just like there's a kind of a misconception in thinking that humans are sort of on this ladder
00:36:08.400 of evolution where we are superior to the animals that are below us, in reality, the way evolution
00:36:17.700 works is that it kind of radiates out from a common ancestor of 3.7 billion years ago.
00:36:24.580 So we're all equally evolved and you, the way, the proper way to think about it is like,
00:36:29.760 are we superior to the, uh, to the starfish, to the giraffe?
00:36:35.680 They have all enjoyed the same amount of evolution as we have.
00:36:39.020 The proper way to kind of map this is to map this in a possibility space and saying these
00:36:45.080 creatures excel in this niche and these creatures excel in this niche and they aren't really superior
00:36:51.200 to us in, in that way. It's even hard to determine whether they're more complicated than us or more
00:36:57.040 complex. So I think a better vision of AIs is to have a possibility space of all the different
00:37:04.960 possible ways you can think. And some of these complexes will be greater than what humans are,
00:37:11.860 but we can't have a complex of, of intelligence that maximize everything. That's just the engineering
00:37:21.180 principle. The engineering maximizes you cannot optimize all, everything you want to do. You're
00:37:28.260 always bound by, by resources and time. So you have to make trade-offs. And if you want to have a
00:37:35.700 Swiss army knife version of intelligence that has all the different things, then they're
00:37:41.840 going to be kind of mediocre in all the things that they do. Um, you can always excel in another
00:37:48.560 version, another dimension by just specializing in that particular node of, of thinking and thought.
00:37:55.620 And so, um, this idea that we're going to make this super version of human intelligence that somehow
00:38:07.500 excels us in every dimension, I think is, I don't see any evidence for that.
00:38:14.220 Let me try to map what you just said on to the way I think about it, because I agree with
00:38:20.380 most of what you said. I think the last bit I don't agree with, but I certainly, and I, I come to a
00:38:26.620 different conclusion or I have a different, at least I have a very vivid concern that survives
00:38:32.940 contact with all the things you just said. I certainly agree that IQ does not map on to
00:38:39.580 the way we think about the intelligence of other species. To ask, you know, what is the IQ of an
00:38:44.700 octopus doesn't make any sense. And it's fine to think about human intelligence, not as a, a single
00:38:51.980 factor, but as, as a constellation of things that we care about. And our, our notion of intelligence could
00:38:58.620 be fairly elastic or that we could suddenly care about other things that, that we haven't cared
00:39:03.580 about very much. And we would want to wrap that up in, in terms of assessing a person's intelligence.
00:39:09.260 And as you mentioned, emotional intelligence, for instance, I think that's a, a discrete capacity that,
00:39:15.580 that, you know, doesn't segregate very reliably with something like mathematical intelligence, say,
00:39:20.860 and, you know, it's, it's fine to talk about it. I think there are reasons why you might want to test
00:39:27.020 it separately from IQ. And, and I think the notion of general intelligence as measured by IQ is, is more
00:39:33.820 useful than, than many people let on. But I definitely take your point that we're this constellation of
00:39:39.180 cognitive capacities. So putting us on a spectrum with a, with a chicken, you know, as I did in,
00:39:45.660 in my Ted talk is more or less just saying that you can issue certain caveats, which, which I didn't
00:39:51.740 issue in that talk, but issuing those caveats still makes this a valid comparison, which is that of the
00:39:57.900 things we care about in cognition, of the things that make us able to do the extraordinarily heavy
00:40:06.460 lifting and unique things we do, like, you know, building a global civilization and producing science
00:40:13.340 and art and mathematics and music and everything else that is making human life both beautiful and
00:40:19.420 durable. There are, there are not that many different capacities that we need to enumerate in order to
00:40:25.580 capture those abilities. It may be 10, it's not a thousand, and a chicken has very few of them. Now,
00:40:33.660 a chicken may be good at other things that we can't even imagine being good at, but for the purposes of
00:40:38.460 this conversation, we don't care about those things, and those things are clearly not leading to chicken
00:40:43.660 civilization and chicken science and the chicken version of the internet. So of the things we care
00:40:49.980 about in cognition, and again, I think the list is, is small, and it's possible that there are things on
00:40:55.980 the list that we really do care about that we haven't discovered yet. Take something like emotional
00:41:00.700 intelligence. Let's say that we, we roll back the clock, you know, 50 years or so, and there's very few
00:41:06.700 people thinking about anything like emotional intelligence, and then put us in the presence of,
00:41:13.500 you know, very powerful artificial intelligent technology, and we don't even think to build
00:41:19.180 emotional intelligence into our systems. It's clearly possible that we could leave out something
00:41:24.140 that is important to us just because we haven't conceptualized it. But of the things we know that
00:41:29.580 are important, there's not that many of them that lead us to be able to, you know, prove mathematical
00:41:34.780 theorems or invent scientific hypotheses or propose experiments, you know, and then if you add things
00:41:42.940 like even emotional intelligence, the ability to detect the emotions of other people in their tone
00:41:49.900 of voice and in their facial expressions, say. These are fairly discrete skills, and here's where I begin
00:41:57.580 to edge into potentially dystopian territory. Once the ground is conquered in artificial systems,
00:42:05.740 it never becomes unconquered. Really, the preeminent example here is something like chess, right? So,
00:42:10.940 for the longest time, chess playing computers were not as good as the best people, and then suddenly they
00:42:17.820 were as, you know, more or less as good as the best people, and then, you know, more or less 15 minutes
00:42:22.460 later, they were better than the best people, and now they will always be better than the best people.
00:42:27.660 And I think we're living in this bit of a mirage now where you have human computer teams, you know,
00:42:34.700 cyborg teams, you know, much celebrated by people like Gary Kasparov, who's been on the podcast talking
00:42:40.700 about them, which are for the moment better than the best computers. So, you know, having the ape still
00:42:46.460 in the system gives you some improvement over the best computer, but ultimately, the ape will just
00:42:53.100 be adding noise, or so I would predict. And once computers are better at chess and better than any
00:43:01.020 human computer combination, that will always be true, but for the fact that we might merge with
00:43:07.260 computers and cease to be merely human. And when you imagine that happening to every other thing we care
00:43:14.460 about in the mode of cognition, then you have to imagine building systems that escape us in their
00:43:23.100 capacities. They could be highly alien in terms of what we have left out in building them, right? So,
00:43:31.980 again, if we had forgotten to build in emotional intelligence, or we didn't understand emotional
00:43:38.700 intelligence enough to build everything in that humans do, we could find ourselves in the presence of
00:43:43.980 you know, say, the most powerful autistic system, you know, the universe has ever devised, right? So
00:43:50.860 we've left something out, and it's only kind of quasi familiar to us as a mind, but, you know, godlike in
00:43:57.900 its capacities. I think it's just the fact that once the ground gets conquered in an artificial system,
00:44:05.340 it stays conquered. And by definition, you know, the resource concerns that you mentioned at the end,
00:44:12.220 you know, if you build a Swiss army knife, it's not going to be a great sword, and it certainly isn't
00:44:16.700 going to be a great airplane. Well, then, I just think that doesn't actually describe what will
00:44:23.100 happen here. Because when you compare the resources that a superhuman intelligence will have, especially
00:44:30.140 if it's linked to the internet, you compare that to a human brain or any collection of human brains,
00:44:36.380 I don't know how many orders of magnitude difference that is. And in terms of the time frame of
00:44:41.660 operation, I mean, you're talking about systems operating a billion times faster than a human brain,
00:44:47.340 there's no reasonable comparison to be made there. And that's where I feel like the possibility of
00:44:53.580 something like the singularity or something like an intelligent explosion is there and worth worrying
00:44:59.740 about. So, again, I'd like to go where we agreed. So, do you use the term?
00:45:06.540 If you'd like to continue listening to this conversation, you'll need to subscribe at
00:45:10.220 samharris.org. Once you do, you'll get access to all full-length episodes of the Making Sense podcast,
00:45:15.820 along with other subscriber-only content, including bonus episodes and AMAs,
00:45:20.860 and the conversations I've been having on the Waking Up app. The Making Sense podcast is ad-free
00:45:25.900 and relies entirely on listener support. And you can subscribe now at samharris.org.