Making Sense - Sam Harris - #255 — The Future of Intelligence

00:00:00.000 Welcome to the Making Sense Podcast.

00:00:08.860 This is Sam Harris.

00:00:10.900 Just a note to say that if you're hearing this,

00:00:13.120 you are not currently on our subscriber feed

00:00:15.180 and will only be hearing the first part of this conversation.

00:00:18.460 In order to access full episodes of the Making Sense Podcast,

00:00:21.600 you'll need to subscribe at samharris.org.

00:00:24.180 There you'll find our private RSS feed

00:00:25.960 to add to your favorite podcatcher,

00:00:28.000 along with other subscriber-only content.

00:00:30.560 We don't run ads on the podcast,

00:00:32.420 and therefore it's made possible entirely

00:00:33.960 through the support of our subscribers.

00:00:35.840 So if you enjoy what we're doing here,

00:00:37.700 please consider becoming one.

00:00:46.660 Today I'm speaking with Jeff Hawkins.

00:00:49.740 Jeff is the co-founder of Numenta,

00:00:52.680 a neuroscience research company,

00:00:54.660 and also the founder of the Redwood Neuroscience Institute.

00:00:58.600 And before that, he was one of the founders

00:01:00.180 of the field of handheld computing,

00:01:03.620 starting Palm and Handspring.

00:01:07.160 He's also a member of the National Academy of Engineering.

00:01:10.600 And he's the author of two books.

00:01:12.640 The first is On Intelligence,

00:01:14.340 and the second and most recent is A Thousand Brains,

00:01:19.760 a new theory of intelligence.

00:01:22.040 And Jeff and I talk about intelligence

00:01:24.520 from a few different sides here.

00:01:27.140 We start with the brain.

00:01:29.640 We talk about how the cortex creates models of the world,

00:01:33.420 the role of prediction in experience.

00:01:37.240 We discuss the idea that thought is analogous to movement

00:01:41.100 in conceptual space.

00:01:42.700 But for the bulk of the conversation,

00:01:46.240 we have a debate about the future of artificial intelligence,

00:01:49.820 and in particular, the alignment problem,

00:01:53.260 and the prospect that AI could pose

00:01:56.120 some kind of existential risk to us.

00:01:59.280 As you'll hear, Jeff and I have very different takes

00:02:03.580 on that problem.

00:02:05.420 Our intuitions divide fairly sharply.

00:02:08.140 And as a consequence, we have a very spirited exchange.

00:02:14.120 Anyway, it was a lot of fun.

00:02:15.780 I hope you enjoy it.

00:02:17.480 And now I bring you Jeff Hawkins.

00:02:26.560 I'm here with Jeff Hawkins.

00:02:28.340 Jeff, thanks for joining me.

00:02:30.480 Thanks for having me, Sam.

00:02:31.560 It's a pleasure.

00:02:32.740 I think we met probably just once,

00:02:35.360 but I feel like we met about 15 years ago,

00:02:38.140 at one of those Beyond Belief conferences

00:02:40.440 at the Salk Institute.

00:02:41.640 Does that ring a bell?

00:02:42.880 You know, I was at one of the Beyond Belief conferences,

00:02:45.960 and I don't recall meeting you there,

00:02:48.100 but it's totally possible.

00:02:49.560 And I just...

00:02:50.020 Yeah, it's possible we didn't meet,

00:02:51.720 but I remember, I think we had an exchange

00:02:54.200 where one of us was in the audience,

00:02:56.520 and the other was...

00:02:57.400 I mean, so we had an exchange over 50 feet or whatever.

00:03:00.380 Yeah, oh, that makes sense.

00:03:02.020 Yeah, I was in the audience, and I was speaking up.

00:03:04.560 Yeah, okay, and I was probably on stage

00:03:06.960 defending some cockamamie conviction.

00:03:09.980 Well, anyway, nice to almost meet you once again.

00:03:13.340 And you have a new book,

00:03:15.280 which we'll cover part of,

00:03:17.340 by no means exhausting its topics of interest,

00:03:20.940 but the new book is A Thousand Brains,

00:03:23.420 and it's a work of neuroscience

00:03:26.000 and also a discussion about the frontiers of AI

00:03:30.100 and where all this is heading.

00:03:32.780 But maybe we should start with the brain part of it

00:03:36.260 and start with the really novel

00:03:40.060 and circuitous and entrepreneurial route

00:03:43.820 you've taken to get into neuroscience.

00:03:47.500 This is the non-standard course

00:03:50.040 to becoming a neuroscientist.

00:03:52.000 Give us your brief biography here.

00:03:55.000 How did you get into these topics?

00:03:56.380 Well, I fell in love with brains

00:03:59.840 when I just got out of college.

00:04:01.640 So I studied electrical engineering in college.

00:04:04.500 And right after I started my first job at Intel,

00:04:08.100 I read an article by Francis Crick

00:04:10.660 about brains and how we don't understand they work.

00:04:14.160 And I just became enamored.

00:04:15.780 I said, oh my God, we should understand this.

00:04:17.320 This is me.

00:04:17.940 I am my brain.

00:04:19.300 No one seems to know how this thing is working.

00:04:21.740 And I just couldn't accept that.

00:04:23.420 And so I decided to dedicate my life

00:04:26.400 to figuring out what's going on when I'm thinking

00:04:28.340 and who we are, basically, as a species.

00:04:32.600 And it was a difficult path.

00:04:34.780 So I quit my job.

00:04:37.460 I essentially applied to become a graduate student,

00:04:41.200 first at MIT in AI,

00:04:43.000 but then I settled at Berkeley in neuroscience.

00:04:46.600 And I said, okay, we're going to spend my life

00:04:49.040 figuring out how the neocortex works.

00:04:51.240 And I found out very quickly

00:04:53.380 that that was a very,

00:04:54.880 not difficult thing to do scientifically,

00:04:56.680 but difficult to do from the practical aspects of science,

00:05:00.460 that you couldn't get funding for that.

00:05:02.200 It was considered too ambitious.

00:05:03.980 You know, there was theoretical work

00:05:05.140 and then people didn't fund theoretical work.

00:05:07.480 So after a couple of years

00:05:08.740 as a graduate student at Berkeley,

00:05:10.620 I set a different path.

00:05:12.060 I said, okay, I'm going to go back

00:05:13.220 to work in industry for a few years

00:05:15.400 to mature,

00:05:16.940 to figure out how to make institutional change

00:05:19.460 because I was up against an institutional problem,

00:05:21.320 not just a scientific problem.

00:05:23.400 And that turned into a series of successful businesses

00:05:26.660 that I was involved with and started,

00:05:28.880 including Palm and Handspring.

00:05:30.720 These are some of the early handheld computing companies.

00:05:34.180 And we were having a tremendous amount of success with that.

00:05:36.920 But it was never my mission

00:05:38.640 to stay in the handheld computing industry.

00:05:42.360 I wanted to get back to neuroscience.

00:05:43.700 And everybody who worked for me knew this.

00:05:45.240 In fact, it was, you know,

00:05:48.000 I told the investors,

00:05:48.920 I'm only going to do this for four years.

00:05:50.340 And they said, what?

00:05:51.180 I said, yeah, that's it.

00:05:52.620 But it turned out to be a lot longer than that

00:05:54.340 because all the success we had.

00:05:55.460 But eventually, I just extracted myself from it.

00:05:58.900 And I said, I'm going to go.

00:06:00.120 And I have so many years left in my life.

00:06:02.440 So after having all that success

00:06:03.800 in the mobile computing space,

00:06:05.080 I started a neuroscience institute.

00:06:07.100 This is at the recommendation

00:06:08.180 from neuroscience friends of mine.

00:06:09.840 So they helped me do that.

00:06:11.500 And I ran that for three years.

00:06:12.880 And now I've been running sort of a private lab

00:06:14.740 just doing pure neuroscience for the last 17 years.

00:06:19.540 That's Numenta, right?

00:06:21.220 That's Numenta, yeah.

00:06:22.980 And we've made some really significant progress

00:06:26.300 in our goals.

00:06:28.680 And the book documents

00:06:30.080 some of the recent,

00:06:31.320 really significant discoveries we've made.

00:06:34.160 So am I right in thinking

00:06:35.400 that you made enough money

00:06:37.520 at Palm and Handspring

00:06:39.820 that you could self-fund

00:06:41.180 to your first neuroscience institute?

00:06:43.420 Or did you, or is that not the case?

00:06:45.040 Did you have to go raise money?

00:06:46.480 It was, well, it was a bit of both.

00:06:48.660 It was, certainly, I was a major contributor.

00:06:51.160 I wasn't the only one.

00:06:52.540 But I didn't want the funding

00:06:54.480 to be the driver of what we did

00:06:56.480 and how we spent all our time.

00:06:58.080 So at the institute,

00:06:59.160 we had collaborations

00:07:00.700 with both Berkeley and Stanford.

00:07:03.160 We didn't get funds from them,

00:07:04.280 but we did work with them on various things.

00:07:07.180 And then we had,

00:07:09.020 but that was mostly funded by myself.

00:07:11.820 Numenta is still,

00:07:13.400 I'm a major contributor to it,

00:07:14.440 but there are other people

00:07:15.240 who've invested in Numenta.

00:07:16.520 We have one outside venture capitalist

00:07:18.440 and several people,

00:07:20.880 but I'm still a major contributor to it.

00:07:23.320 I don't,

00:07:24.000 I just view that as a sort of a necessary thing

00:07:26.780 to get onto the science

00:07:28.340 and not have to worry about it.

00:07:29.620 Yeah.

00:07:30.240 Because when I was at Berkeley,

00:07:31.540 what I was told over and over again,

00:07:33.300 and I really came to understand this.

00:07:34.840 In fact,

00:07:35.820 I went and eventually,

00:07:37.740 after that,

00:07:38.860 when I was running

00:07:39.320 the Redwood Neuroscience Institute,

00:07:41.660 I went to Washington

00:07:42.500 to talk about

00:07:43.280 to the National Science Foundation

00:07:45.000 and the National Institute of Health

00:07:46.220 and also to DARPA,

00:07:48.320 who were the funders of neuroscience.

00:07:50.140 And everyone thought

00:07:50.940 what we were doing,

00:07:51.660 which is sort of big theory,

00:07:53.160 large scale theories

00:07:54.000 of neocortical function,

00:07:55.860 that this was like

00:07:56.460 the most important problem to work on,

00:07:58.280 but everyone said

00:07:59.200 they can't fund it

00:08:00.080 for various reasons.

00:08:01.100 And so over the years,

00:08:03.380 I've come to appreciate

00:08:04.060 that it's very difficult

00:08:05.400 to be a scientist

00:08:06.420 doing what we do

00:08:07.320 with traditional funding sources.

00:08:10.180 But we don't work outside of science.

00:08:12.900 We partner with labs

00:08:14.220 and we go to conferences

00:08:15.420 and we publish papers.

00:08:16.360 We do all the regular stuff.

00:08:17.880 Right, right.

00:08:19.060 Yeah, it's amazing

00:08:19.640 how much comes down to funding

00:08:21.280 or lack of funding

00:08:22.460 and the incentives

00:08:23.880 that would dictate

00:08:25.300 whether something gets funded

00:08:26.720 in the first place.

00:08:27.440 It's by no means

00:08:29.600 a perfect system.

00:08:31.260 It's a kind of

00:08:32.000 an intellectual market failure.

00:08:34.080 Yeah, it is fascinating

00:08:35.560 and we could have

00:08:36.560 a whole conversation

00:08:37.180 about that sometimes perhaps

00:08:38.320 because I ask myself,

00:08:39.960 why is it so hard?

00:08:40.920 Why do people can't fund this?

00:08:42.580 And there's reasons for it

00:08:43.820 and it's a complex,

00:08:45.820 strange thing

00:08:46.760 when people were telling me

00:08:47.960 this is the most important thing

00:08:49.100 anyone could be working on

00:08:50.400 and we think your approaches are great,

00:08:52.900 but we can't fund that.

00:08:54.460 And why is that?

00:08:55.620 You know, but it's, you know,

00:08:57.700 I just accepted the way it was.

00:08:59.380 I said, okay,

00:08:59.720 this is the world I'm living in.

00:09:00.960 I'm going to get one chance here.

00:09:02.820 If I can't do this through like,

00:09:04.540 you know, being, you know,

00:09:06.020 working my way as a graduate student

00:09:07.620 to getting a position in university,

00:09:09.960 how am I going to do it?

00:09:11.460 And I said, okay,

00:09:12.340 it's not what I thought,

00:09:13.120 but this is what it's going to be.

00:09:15.640 Nice.

00:09:16.100 Well, let's jump into

00:09:17.840 the neuroscience side of it.

00:09:19.860 Generally speaking,

00:09:20.480 we're going to be talking about

00:09:21.780 intelligence

00:09:23.060 intelligence and

00:09:24.120 how it's accomplished

00:09:25.540 in physical systems.

00:09:27.440 So let's,

00:09:28.260 let's start with a,

00:09:29.260 a definition,

00:09:30.840 however loose.

00:09:33.420 What is intelligence

00:09:35.180 in your view?

00:09:37.220 So I didn't know

00:09:38.220 and didn't have any

00:09:39.100 pre-ideas about

00:09:40.080 what this would be.

00:09:40.940 It was a mystery to me.

00:09:42.540 And, but we've learned

00:09:43.900 what a good portion

00:09:45.460 of your brain is doing.

00:09:46.480 And so we started

00:09:47.080 the neocortex,

00:09:47.680 which is about 70%

00:09:49.360 of the volume

00:09:49.980 of a human brain.

00:09:50.640 And I now know

00:09:52.480 what that does.

00:09:53.200 And so I'm going to

00:09:54.080 take that

00:09:54.820 as my definition

00:09:56.160 for intelligence here.

00:09:57.900 What's going on

00:09:58.820 in your neocortex

00:09:59.600 is it's learning

00:10:00.520 a model of the world,

00:10:01.960 an internal

00:10:02.560 recreation

00:10:03.440 of all the things

00:10:05.080 in the world

00:10:05.840 that you know of

00:10:06.540 and how it does.

00:10:08.120 That's the key

00:10:09.140 in what we've discovered,

00:10:10.500 but it's this internal model.

00:10:12.340 And intelligence

00:10:12.860 is,

00:10:13.620 requires having

00:10:14.500 an internal model

00:10:15.440 of the world

00:10:16.480 in your head.

00:10:17.500 And it allows you

00:10:18.360 to recognize

00:10:19.020 where you are.

00:10:19.540 It allows you

00:10:20.060 to act on things.

00:10:21.060 It allows you

00:10:21.460 to plan

00:10:21.960 and think about

00:10:22.540 the future.

00:10:23.060 So if I'm going to say

00:10:23.680 what happens

00:10:24.420 when I do this,

00:10:25.340 the model tells you that.

00:10:27.380 So to me,

00:10:28.140 intelligence is just

00:10:29.100 about having a model

00:10:29.780 in your head

00:10:30.400 and using that

00:10:31.420 for planning and action.

00:10:33.040 It's not about

00:10:33.860 doing anything particular.

00:10:35.440 It's about

00:10:36.100 understanding the world.

00:10:38.360 Yeah, that's interesting.

00:10:39.060 I think most people would,

00:10:40.400 that's kind of

00:10:40.700 an internal

00:10:41.600 definition

00:10:43.020 of intelligence,

00:10:44.160 but I think

00:10:44.680 most people would

00:10:45.720 reach for

00:10:46.860 an external one

00:10:48.240 or a functional

00:10:50.000 one that has

00:10:50.900 to take in

00:10:51.640 the environment.

00:10:53.340 I mean,

00:10:53.460 it's something about

00:10:54.080 being able to

00:10:55.240 flexibly

00:10:56.620 meet your goals

00:10:58.160 under a range

00:10:59.420 of conditions,

00:11:00.260 you know,

00:11:00.460 more flexibly

00:11:01.760 than rigidly.

00:11:02.660 I guess there's

00:11:03.060 rigid forms

00:11:04.540 of intelligence,

00:11:05.280 but when we're

00:11:06.340 talking about

00:11:06.740 anything like

00:11:07.620 general intelligence,

00:11:09.100 we're talking about

00:11:10.080 something that is

00:11:11.320 not merely

00:11:12.980 hardwired

00:11:14.040 and reflexive

00:11:15.260 but flexible.

00:11:16.560 Well, yes,

00:11:16.820 but if you have

00:11:17.920 an internal model

00:11:18.900 of the world,

00:11:19.460 you had to learn it.

00:11:20.480 I mean,

00:11:20.680 at least from a human

00:11:21.560 point of view,

00:11:22.020 there's some things

00:11:22.600 we have built in

00:11:23.420 when we're born,

00:11:24.520 but the vast majority

00:11:25.940 of what you and I know,

00:11:27.420 Sam,

00:11:27.720 is learned.

00:11:29.220 You know,

00:11:29.400 we didn't know

00:11:29.800 what a computer was

00:11:30.540 when you're born.

00:11:31.080 You don't know

00:11:31.340 what a coffee cup is.

00:11:32.260 You don't know

00:11:32.600 what a building is.

00:11:33.780 You don't know

00:11:33.960 what doors are.

00:11:34.700 You don't know

00:11:35.040 what computer codes are.

00:11:36.540 None of this stuff.

00:11:37.700 Everything that,

00:11:38.740 almost everything

00:11:39.240 we interact with

00:11:39.920 in the world today,

00:11:41.600 in language,

00:11:42.300 we don't know

00:11:42.800 any particular language

00:11:43.800 when we're born.

00:11:45.000 We don't know

00:11:45.580 mathematics.

00:11:46.100 So we had to learn

00:11:46.680 all these things.

00:11:47.600 So if you want to say

00:11:48.820 there might be

00:11:49.320 an internal model

00:11:50.080 that wasn't learned,

00:11:50.980 well,

00:11:51.080 that's pretty trivial.

00:11:52.160 But I'm talking

00:11:52.860 about models

00:11:53.380 that are learned

00:11:53.920 and you have to

00:11:54.620 interact with the world

00:11:55.540 to learn it.

00:11:56.060 You can't learn it

00:11:56.760 without being present

00:11:58.420 in the world,

00:11:58.880 without having an embodiment,

00:11:59.980 without moving about,

00:12:00.920 touching and seeing

00:12:01.680 and hearing things.

00:12:03.020 So a large part

00:12:04.160 of what people think about,

00:12:05.200 like you brought up,

00:12:05.960 is, okay,

00:12:07.060 you know,

00:12:07.320 we are able to solve a goal.

00:12:09.220 But that's what

00:12:10.180 a model lets you to do.

00:12:11.340 It's not the,

00:12:11.920 that is not

00:12:12.820 what intelligence itself is.

00:12:14.660 Intelligence is having

00:12:15.440 this ability

00:12:16.020 to solve any goal,

00:12:17.240 right?

00:12:17.560 Because you have,

00:12:18.360 if your model

00:12:19.000 covers that part

00:12:20.320 of the world,

00:12:20.700 you can figure out

00:12:21.440 how to manipulate

00:12:22.140 that part of the world

00:12:22.920 and achieve what you want.

00:12:25.060 So it's,

00:12:25.660 I'll give you

00:12:26.080 a little further analogy.

00:12:27.740 It's a little bit

00:12:28.080 like computers.

00:12:28.860 When we talk about

00:12:29.420 like a universal

00:12:30.080 Turing machine

00:12:30.840 or what a computer is,

00:12:32.220 it's not defined

00:12:33.200 by what the computer

00:12:34.640 is applied to do.

00:12:36.160 It's like,

00:12:36.760 a computer isn't something

00:12:37.580 that solves

00:12:38.040 a particular problem.

00:12:38.900 A computer is something

00:12:39.540 that works

00:12:39.900 on a set of principles

00:12:40.740 and that's how

00:12:42.400 I think about intelligence.

00:12:43.460 It's a modeling system

00:12:45.200 that works

00:12:45.540 on a set of principles.

00:12:47.040 Those principles

00:12:47.660 can exist in a mouse

00:12:49.140 and a dog

00:12:49.700 and a cat

00:12:50.120 and a human

00:12:50.520 and probably birds,

00:12:52.460 but don't focus

00:12:54.300 on what those animals

00:12:55.080 are doing.

00:12:55.520 Yeah, I think it's important

00:12:58.220 to point out

00:12:58.680 that a model

00:12:59.580 need not be

00:13:00.360 a conscious model.

00:13:01.640 In fact,

00:13:01.940 most of our models

00:13:02.940 are not conscious

00:13:04.320 and might not even be

00:13:06.840 in principle

00:13:07.420 available to consciousness,

00:13:09.740 although I think

00:13:10.400 at the boundary,

00:13:11.920 something that you'd say

00:13:14.120 is happening entirely

00:13:15.560 in the dark

00:13:16.140 does have a kind of,

00:13:17.620 or can have a kind of

00:13:18.600 liminal conscious aspect.

00:13:21.440 So I mean,

00:13:21.760 to take, you know,

00:13:22.520 the coffee cup example,

00:13:23.620 this leads us into

00:13:25.040 a more granular discussion

00:13:26.940 of what it means

00:13:28.000 to have a model

00:13:28.720 of anything

00:13:29.340 at the level

00:13:29.980 of the cortex.

00:13:31.600 But, you know,

00:13:31.980 if I reach for my coffee cup

00:13:33.400 and grasp it,

00:13:35.480 the ordinary experience

00:13:37.220 of doing that

00:13:37.940 is something

00:13:38.620 I'm conscious of.

00:13:40.080 I'm not conscious

00:13:41.260 of all of the prediction

00:13:43.580 that is built

00:13:44.840 into my accomplishing that

00:13:47.480 and experiencing

00:13:49.440 what I experience

00:13:50.460 when I touch a coffee cup,

00:13:51.580 and yet it's prediction

00:13:53.820 that is required

00:13:55.800 having some ongoing expectation

00:13:58.060 of what's going to happen there

00:13:59.140 when I, you know,

00:14:00.160 when each finger

00:14:00.960 touches the surface

00:14:02.620 of the cup

00:14:03.220 that allows for me

00:14:04.940 to detect

00:14:05.700 any error there

00:14:07.840 or to be surprised

00:14:08.920 by something

00:14:09.600 truly anomalous.

00:14:10.680 So if I reach

00:14:11.120 for a coffee cup

00:14:11.900 and it turns out

00:14:13.300 that's a, you know,

00:14:13.860 it's a hologram

00:14:14.540 of a coffee cup

00:14:15.560 and my hand

00:14:16.400 passes right through it,

00:14:17.860 the element of surprise

00:14:19.440 there seems predicated

00:14:20.820 on some ongoing

00:14:22.740 prediction processing

00:14:25.120 to which the results

00:14:26.800 of my behavior

00:14:27.540 is being compared.

00:14:29.120 So maybe you can talk

00:14:30.080 about what you mean

00:14:31.140 by having a model

00:14:33.460 at the level

00:14:34.020 of the cortex

00:14:34.420 and how prediction

00:14:35.960 is built into that.

00:14:38.780 Yeah.

00:14:39.260 Well, my first book,

00:14:40.480 which I published

00:14:40.960 like 14 years ago,

00:14:42.300 called On Intelligence,

00:14:43.180 was just about that topic.

00:14:45.680 It was about

00:14:46.160 how it is

00:14:47.160 the brain is making

00:14:47.920 all these predictions

00:14:48.760 all the time

00:14:49.520 and all your sensory modalities

00:14:50.980 and you're not aware of it.

00:14:52.940 And so that's

00:14:54.000 sort of the foundation

00:14:54.780 and you can't make

00:14:55.740 a prediction

00:14:56.220 without a model.

00:14:57.000 I mean, a model,

00:14:57.800 to make a prediction,

00:14:58.760 you had to have

00:14:59.120 some expectation,

00:15:00.500 the expectation,

00:15:01.340 whether you're not aware

00:15:01.900 of it or not,

00:15:02.540 but you have an expectation

00:15:03.660 and that has to be driven

00:15:05.480 from some internal

00:15:06.600 representation of the world

00:15:07.660 that says,

00:15:08.360 hey, you're about

00:15:09.720 to touch this thing,

00:15:10.460 I know what it is,

00:15:11.200 it's supposed to feel

00:15:11.920 this way.

00:15:13.040 And even if you're not aware

00:15:14.480 that you're doing that.

00:15:16.160 One of the key discoveries

00:15:17.700 we made,

00:15:18.720 and this was maybe

00:15:19.560 about eight years ago,

00:15:21.220 we had to get to the bottom,

00:15:23.320 like how do neurons

00:15:24.120 make predictions?

00:15:25.360 What is the physical manifestation

00:15:27.140 of a prediction

00:15:28.160 in the brain?

00:15:29.680 And most of these predictions,

00:15:30.820 as you point out,

00:15:31.420 are not conscious,

00:15:32.400 you're not aware of them.

00:15:33.520 They're just happening

00:15:34.240 and if something is wrong,

00:15:36.320 then your attention

00:15:36.940 is drawn to it.

00:15:38.320 So if you felt the coffee cup

00:15:39.480 and there was a little burr

00:15:40.260 on the side or a crack

00:15:41.240 and you didn't know

00:15:41.920 that was expected,

00:15:43.020 that you'd say,

00:15:43.580 oh, there's a crack.

00:15:44.340 What was the brain doing

00:15:46.680 when it was making

00:15:47.240 that prediction?

00:15:48.560 And we have a theory about this

00:15:51.020 and I wrote about it

00:15:51.660 in the book a bit

00:15:52.220 and it's a beautiful,

00:15:55.160 I think it's a beautiful theory,

00:15:57.440 but it's basically

00:15:59.180 most of the predictions

00:16:00.120 that are going on

00:16:00.800 in your brain,

00:16:01.220 most of them,

00:16:01.820 not all of them,

00:16:02.500 but most of them,

00:16:03.680 happen inside individual neurons.

00:16:06.920 It is internal

00:16:08.420 to the individual neurons.

00:16:09.680 Now, not a single neuron

00:16:11.480 can predict something,

00:16:12.580 but an ensemble of neurons

00:16:13.900 do this,

00:16:15.040 but it's an internal state

00:16:16.460 and we wrote a paper

00:16:18.000 that came out in 19,

00:16:20.980 2016, excuse me,

00:16:22.740 2016,

00:16:23.800 which is,

00:16:24.980 it's called,

00:16:25.520 why do neurons

00:16:26.040 have so many synapses?

00:16:27.920 And what we posited

00:16:29.920 in that paper,

00:16:30.940 and I'm pretty sure

00:16:31.460 this is correct,

00:16:32.680 is that neurons

00:16:33.920 have these thousands

00:16:34.680 of synapses.

00:16:35.940 Most of those synapses

00:16:37.120 are being used

00:16:37.620 for prediction.

00:16:38.720 And when a neuron

00:16:39.440 recognizes a pattern

00:16:40.600 and says,

00:16:41.240 okay, I'm supposed

00:16:42.020 to be active soon,

00:16:43.020 I should be,

00:16:43.620 I should be becoming

00:16:45.220 active soon.

00:16:45.860 If everything is according

00:16:46.740 to our model here,

00:16:47.920 I should be becoming

00:16:48.480 active soon,

00:16:49.540 and it goes into

00:16:50.200 this internal state,

00:16:51.420 the neuron itself

00:16:52.160 is saying,

00:16:52.600 okay, I'm expecting

00:16:53.740 to become active.

00:16:55.220 And you can't detect

00:16:57.280 that consciously.

00:16:58.040 It's internal to the,

00:16:59.060 it's essentially

00:16:59.800 just a depolarization

00:17:00.980 or a change

00:17:01.480 of the voltage

00:17:01.940 of the neuron.

00:17:03.700 And so when,

00:17:04.980 but it,

00:17:06.140 we showed how

00:17:06.840 the network

00:17:07.380 of these neurons,

00:17:08.080 what will happen

00:17:08.620 is if your prediction

00:17:09.800 is correct,

00:17:11.020 then a small subset

00:17:12.440 of the neurons

00:17:12.840 become active.

00:17:13.780 But if the prediction

00:17:14.460 is incorrect,

00:17:15.440 a whole bunch

00:17:15.980 of neurons

00:17:16.340 become active

00:17:16.980 at the same time.

00:17:18.280 And then that draws

00:17:19.420 your attention

00:17:19.900 to the problem.

00:17:21.140 So it's a fascinating

00:17:22.080 problem,

00:17:22.560 but most of the predictions

00:17:23.520 going on in your brain

00:17:24.320 are not accessible

00:17:25.380 outside of individual neurons.

00:17:27.620 So there's no way

00:17:28.240 you could be conscious

00:17:28.800 about it.

00:17:29.760 Hmm.

00:17:31.020 I guess most people

00:17:32.160 are familiar

00:17:32.820 with the general anatomy

00:17:34.160 of a neuron

00:17:34.660 where you have

00:17:35.160 this spindly

00:17:36.680 looking thing

00:17:38.140 where there's

00:17:39.460 a cell body

00:17:40.160 and there's

00:17:40.600 a long process,

00:17:41.880 the axon

00:17:42.600 leading away,

00:17:44.140 which carries

00:17:45.000 the action potential

00:17:46.680 if that neuron

00:17:47.700 fires to the synapse

00:17:49.520 and communicates

00:17:50.960 neurotransmitters

00:17:52.200 to other neurons.

00:17:54.040 But on the other

00:17:55.100 side of,

00:17:56.400 in the standard case,

00:17:57.900 on the other side

00:17:58.760 of the cell body,

00:18:00.680 there's this really,

00:18:01.920 often really profuse

00:18:03.960 arborization

00:18:04.860 of dendrites,

00:18:06.420 which is kind of

00:18:07.420 the mad tangle

00:18:08.240 of processes

00:18:08.900 which receive

00:18:10.960 information

00:18:12.160 from other neurons

00:18:14.320 to which this neuron

00:18:15.700 is connected.

00:18:16.700 And it's the integration

00:18:18.860 of information

00:18:19.660 on that side,

00:18:20.720 but before

00:18:21.400 that neuron fires,

00:18:23.800 the change,

00:18:24.500 the probability

00:18:25.140 of its firing,

00:18:26.520 that that's the place

00:18:28.280 you are locating

00:18:29.960 in the full set

00:18:31.420 of predictive changes

00:18:32.720 or the full set

00:18:33.460 of changes

00:18:34.120 that constitute

00:18:35.120 prediction

00:18:35.760 in the case

00:18:36.340 of a system

00:18:37.620 of neurons.

00:18:38.480 Yeah.

00:18:38.880 It was interesting

00:18:39.880 for many years,

00:18:40.960 people looked

00:18:41.420 at the connections

00:18:42.820 on the dendrites,

00:18:44.240 on that bushy part

00:18:45.200 called synapses,

00:18:46.980 and when they

00:18:47.920 activated a synapse,

00:18:49.360 most of the synapses

00:18:50.500 were so far

00:18:51.800 from the cell body

00:18:52.700 that they didn't

00:18:54.040 really have much

00:18:54.660 of an effect.

00:18:55.420 They didn't seem

00:18:56.180 like they could

00:18:56.720 make anything happen.

00:18:58.700 And so,

00:18:59.440 but there are

00:19:00.020 thousands and thousands

00:19:00.980 of them out there,

00:19:02.260 but they don't seem

00:19:03.400 powerful enough

00:19:04.100 to make anything occur.

00:19:06.060 And what was discovered

00:19:07.180 basically over the last

00:19:08.560 20 years,

00:19:09.660 that there are,

00:19:11.080 there's a second

00:19:11.800 type of spike.

00:19:12.840 So you mentioned

00:19:13.620 the one that goes

00:19:14.220 down the axon,

00:19:15.060 that's the action potential,

00:19:16.980 but there are spikes

00:19:17.900 that travel along

00:19:18.840 the dendrites.

00:19:20.440 And so basically

00:19:21.560 what happens is

00:19:22.600 the individual sections

00:19:24.200 of the dendrite,

00:19:24.940 like little branches

00:19:25.640 of this tree,

00:19:26.340 each one of them

00:19:27.560 can recognize

00:19:28.140 patterns on their own.

00:19:29.500 They can recognize

00:19:30.280 hundreds of separate

00:19:31.860 patterns on these

00:19:32.580 different branches,

00:19:33.660 and they can cause

00:19:34.480 this spike to travel

00:19:35.900 along the dendrite.

00:19:37.560 And that lowers,

00:19:39.720 changes the voltage

00:19:41.020 of the cell body

00:19:41.880 a little bit.

00:19:42.900 And that is what we

00:19:43.960 call the predictive state.

00:19:45.060 The cell is like prime.

00:19:46.500 It says,

00:19:46.900 oh,

00:19:47.580 if I fire,

00:19:49.280 I'm ready to fire.

00:19:50.820 And it's not actually

00:19:52.140 a probability change,

00:19:53.680 it's the timing.

00:19:55.460 And so a cell

00:19:56.180 that's in this

00:19:56.700 predictive state

00:19:57.340 that says,

00:19:57.740 I think I should

00:19:58.520 be firing now

00:19:59.420 or very shortly,

00:20:01.260 if it does generate

00:20:03.080 the regular spike,

00:20:04.140 the action potential,

00:20:04.920 it does it a little bit

00:20:05.840 sooner than it would

00:20:07.040 have otherwise.

00:20:07.700 And it's the timing

00:20:08.480 that is the key

00:20:09.480 to making the whole

00:20:10.020 circuit work.

00:20:11.280 We're getting pretty

00:20:11.840 down in the weeds

00:20:12.400 here about neuroscience.

00:20:13.320 I don't know

00:20:14.240 if all your readers

00:20:15.340 or your listeners

00:20:16.480 will appreciate that.

00:20:17.860 Yeah,

00:20:18.100 no,

00:20:18.260 I think it's useful though.

00:20:19.900 More weeds here.

00:20:21.120 One of the novel things

00:20:22.680 about your argument

00:20:24.360 is that it was inspired

00:20:26.900 by some much earlier

00:20:28.440 theorizing.

00:20:29.380 You mark your debt

00:20:30.960 to Vernon Mountcastle,

00:20:33.080 but the idea is

00:20:34.320 that there's a common

00:20:36.320 algorithm operating

00:20:38.300 more or less everywhere

00:20:39.620 at the level

00:20:40.540 of the cortex

00:20:41.100 that is,

00:20:42.540 it's more or less,

00:20:43.560 the cortex is doing

00:20:45.060 essentially the same thing,

00:20:47.520 whether it's producing

00:20:49.280 language or vision

00:20:52.080 or any other sensory channel

00:20:54.520 or motor behavior.

00:20:56.580 So talk about

00:20:57.540 the general principle

00:20:59.640 that you spend

00:21:01.360 a lot of time on

00:21:02.200 in the book

00:21:02.580 of just the organization

00:21:04.080 of the neocortex

00:21:05.440 into cortical columns

00:21:06.920 and the implications

00:21:08.460 this has

00:21:09.880 for how we view

00:21:11.620 what the brain is doing

00:21:13.180 in terms of sensory

00:21:15.160 and motor learning

00:21:16.800 and all of its consequences.

00:21:20.160 Vernon Mountcastle

00:21:21.100 made this proposal

00:21:21.800 back in the 70s

00:21:22.860 and it's just

00:21:24.780 a dramatic idea

00:21:26.120 and it's an incredible idea

00:21:27.840 and so incredible

00:21:28.500 that some people

00:21:29.080 just refuse to believe it,

00:21:30.260 but other people

00:21:30.860 really think

00:21:31.700 it's a tremendous discovery.

00:21:34.240 But what he noticed

00:21:35.040 was if you look

00:21:35.680 at the neocortex,

00:21:36.760 if you could take one

00:21:37.580 out of your head

00:21:38.300 or out of a human's head,

00:21:39.820 it's like a sheet.

00:21:41.300 It's about

00:21:41.740 two and a half millimeters thick.

00:21:43.560 It is about

00:21:44.720 the size of a large

00:21:45.520 dinner napkin

00:21:46.240 or 1,500 square centimeters

00:21:47.820 and if you could fold it,

00:21:50.320 lay it flat

00:21:50.820 and the different parts of it

00:21:52.720 like that do different things.

00:21:54.620 There's parts that do vision,

00:21:55.760 there's parts that do language

00:21:56.740 and parts that do hearing

00:21:57.760 and so on.

00:21:59.000 But when you,

00:21:59.680 if you cut into it

00:22:00.760 and you look at

00:22:02.280 the structure

00:22:03.560 in any one of these areas,

00:22:05.720 it's very complicated.

00:22:07.360 There are dozens

00:22:08.160 of different cell types

00:22:09.220 but they're very

00:22:10.220 prototypically connected

00:22:11.760 and they're arranged

00:22:13.220 in certain patterns

00:22:14.160 and layers

00:22:14.700 and different types of things

00:22:15.620 so it's a very

00:22:16.240 it's a very complex structure

00:22:18.040 but it's almost

00:22:19.060 the same everywhere.

00:22:20.760 It's not the same everywhere

00:22:21.940 but almost the same everywhere

00:22:23.340 and so this is not just true

00:22:25.040 in a human neocortex

00:22:26.020 but if you look

00:22:26.900 at a rat's neocortex

00:22:27.840 or a dog's neocortex

00:22:29.020 or a cat

00:22:29.720 or a monkey,

00:22:31.060 this same basic structure

00:22:32.560 is there

00:22:33.060 and what Werner Malkus

00:22:34.780 said

00:22:35.380 is that

00:22:36.240 all the parts

00:22:37.320 of the neocortex

00:22:38.020 are actually,

00:22:39.480 we think of them

00:22:40.140 as doing things,

00:22:40.960 different things

00:22:41.460 but they're actually

00:22:42.060 all doing

00:22:42.720 some fundamental algorithm

00:22:44.000 which is the same.

00:22:45.440 So hearing

00:22:45.960 and touch

00:22:46.440 and vision

00:22:46.800 are really

00:22:47.260 the same thing.

00:22:48.780 He says

00:22:49.060 if you took part

00:22:49.820 of the cortex

00:22:50.240 and you hook it up

00:22:50.840 to your eyes

00:22:51.260 you'll get vision.

00:22:51.920 If you hook it up

00:22:52.340 to your ears

00:22:52.760 you'll get hearing.

00:22:53.940 If you hook it up

00:22:54.460 to other parts

00:22:55.000 of the neocortex

00:22:55.520 you'll get language

00:22:56.280 and so

00:22:57.620 he spent many years

00:22:59.700 giving the evidence

00:23:01.020 for this.

00:23:02.100 He proposed further

00:23:03.040 that this algorithm

00:23:04.720 was contained

00:23:05.500 in what's called

00:23:05.960 a column

00:23:06.440 and so

00:23:07.360 if you would take

00:23:08.640 a small area

00:23:10.360 of this neocortex,

00:23:11.520 remember it's like

00:23:12.060 two and a half

00:23:13.800 millimeters thick,

00:23:14.840 you take a very

00:23:15.500 sort of skinny

00:23:16.160 little one millimeter

00:23:18.100 column out of it

00:23:19.680 that that is

00:23:20.740 the processing element

00:23:21.740 and so

00:23:22.760 our human neocortex

00:23:24.560 we have about

00:23:25.420 150,000 of these columns

00:23:27.240 other animals

00:23:28.900 have more or less.

00:23:30.240 People should picture

00:23:30.780 something resembling

00:23:31.940 a grain of rice

00:23:32.940 in terms of scale here.

00:23:34.260 Yeah, yeah.

00:23:34.920 I sometimes say

00:23:35.660 take a piece

00:23:36.180 of skinny spaghetti

00:23:36.980 like you know

00:23:37.520 angel hair pasta

00:23:38.300 or something like that

00:23:39.040 and cut it into

00:23:39.620 two little

00:23:40.060 two and a half

00:23:40.920 millimeter lengths

00:23:41.640 and stack them

00:23:42.480 side by side.

00:23:43.240 Now the funny thing

00:23:44.720 about columns

00:23:45.320 is you can't see them.

00:23:46.340 They're not visual

00:23:47.040 things.

00:23:47.760 You can't look

00:23:48.460 under a microscope

00:23:49.020 you won't see it

00:23:49.760 but he pointed out

00:23:51.440 why they're there.

00:23:53.900 It has to do

00:23:54.540 with how they're connected

00:23:55.640 so all the cells

00:23:57.240 in one of these

00:23:57.820 little millimeter

00:23:58.480 pieces of rice

00:23:59.660 or spaghetti

00:24:00.280 if you will

00:24:00.720 are all processing

00:24:02.080 the same thing

00:24:02.800 and the next

00:24:03.400 piece of rice

00:24:04.600 over processing

00:24:05.320 something different

00:24:06.060 and the next piece

00:24:06.620 of rice over

00:24:07.040 processing

00:24:07.480 something different

00:24:08.140 and so he didn't

00:24:10.320 know what was

00:24:11.220 going on

00:24:11.860 in the cortical

00:24:12.680 column.

00:24:13.840 He articulated

00:24:15.560 the architecture

00:24:16.360 he talked about

00:24:17.920 the evidence

00:24:18.580 that this exists

00:24:19.400 he said

00:24:20.200 here's the evidence

00:24:20.820 why these things

00:24:22.000 are all doing

00:24:22.480 the same thing

00:24:23.360 but he didn't

00:24:25.120 know what it was

00:24:26.520 and it's kind of

00:24:26.980 hard to imagine

00:24:27.540 what it is

00:24:28.320 that this algorithm

00:24:29.820 could be doing

00:24:30.520 but that was

00:24:31.440 essentially the core

00:24:32.380 of our research

00:24:33.000 that's what we've

00:24:33.660 been focused on

00:24:34.340 for close to 20 years.

00:24:35.400 It's also hard

00:24:36.920 to imagine

00:24:37.340 the microanatomy

00:24:38.500 here because

00:24:39.040 in each one

00:24:39.580 of these little

00:24:41.220 columns

00:24:41.720 there's something

00:24:42.760 like 150,000

00:24:44.260 neurons on average

00:24:45.640 and if you

00:24:46.780 could just unravel

00:24:47.920 all of the

00:24:49.500 connections there

00:24:51.160 you know

00:24:52.020 the tiny filaments

00:24:53.700 of nerve endings

00:24:55.540 what you would

00:24:56.440 have there

00:24:57.100 is on the order

00:24:57.960 of kilometers

00:24:58.820 in length

00:24:59.920 you know

00:25:00.680 all wound up

00:25:01.640 into that

00:25:02.100 tiny structure

00:25:03.100 so it's

00:25:04.060 this is a

00:25:04.680 strange juxtaposition

00:25:06.540 of simplicity

00:25:07.880 and complexity

00:25:08.660 but it's

00:25:09.460 there's certainly

00:25:10.580 a mad tangle

00:25:12.220 of processes

00:25:12.820 in there.

00:25:14.080 Yeah this is why

00:25:14.620 brains are so hard

00:25:15.360 to study

00:25:15.700 you know if you

00:25:16.040 look at another

00:25:16.600 organ in the body

00:25:17.520 whether it's

00:25:18.040 the heart

00:25:18.720 or the liver

00:25:19.240 or something like

00:25:19.820 that

00:25:20.080 and you take

00:25:21.160 a little section

00:25:21.760 of it

00:25:22.000 it's pretty uniform

00:25:22.920 you know what I'm

00:25:23.420 saying

00:25:23.560 but here

00:25:25.140 if you take

00:25:25.560 a teeny piece

00:25:26.200 of the

00:25:26.480 teeny teeny piece

00:25:27.620 of the cortex

00:25:28.100 it's got this

00:25:28.860 incredible complexity

00:25:30.020 in it

00:25:30.400 which is not

00:25:30.960 just a

00:25:31.400 it's not random

00:25:32.380 it's very specific

00:25:34.420 and so

00:25:35.720 yeah it's hard

00:25:37.120 to get

00:25:37.480 wrap your heads

00:25:38.180 around how complex

00:25:38.960 it is

00:25:39.500 but we need

00:25:40.540 to be complex

00:25:41.220 because what we do

00:25:42.440 as humans

00:25:42.860 is extremely complex

00:25:44.080 and you know

00:25:45.580 we shouldn't be fooled

00:25:46.520 that we're just

00:25:47.100 a bunch of neurons

00:25:47.680 that are doing

00:25:48.040 some mass action

00:25:48.960 no there's a very

00:25:50.040 complex processing

00:25:51.560 going on

00:25:52.200 and in your brain

00:25:54.300 that it's

00:25:55.380 it's not just

00:25:55.960 a blob of neurons

00:25:57.200 that are pulsating

00:25:58.160 you know

00:25:58.740 very detailed

00:26:00.120 mechanisms

00:26:01.100 that are undergoing it

00:26:02.460 and we figured out

00:26:03.300 what some of those are

00:26:04.080 so describe to me

00:26:06.320 what you mean

00:26:07.840 by this phrase

00:26:08.840 a reference frame

00:26:10.740 what does that mean

00:26:12.340 at the level of

00:26:13.040 the cortex

00:26:14.080 and cortical columns

00:26:15.520 yeah

00:26:16.560 so

00:26:17.480 we're jumping

00:26:18.760 to the end point

00:26:19.660 because that's not

00:26:20.560 where we started

00:26:21.220 we were trying to figure

00:26:22.500 out how cortical columns

00:26:23.380 work

00:26:23.740 and what we realized

00:26:25.880 is that

00:26:26.880 they're little

00:26:27.800 modeling engines

00:26:28.600 they

00:26:28.880 each one of these

00:26:29.740 cortical columns

00:26:30.400 is able to build

00:26:31.220 a model

00:26:31.900 of its input

00:26:32.880 and

00:26:33.860 that model

00:26:34.820 is what we would

00:26:35.460 call a sensory

00:26:36.180 motor model

00:26:36.800 that is

00:26:37.280 it's getting

00:26:38.600 let's assume

00:26:39.360 it's getting

00:26:39.700 from your finger

00:26:40.840 right

00:26:41.460 a tip of your finger

00:26:42.300 one of the columns

00:26:43.020 is getting input

00:26:43.620 from the tip

00:26:44.100 of your finger

00:26:44.600 and as your finger

00:26:46.300 moves and touches

00:26:47.020 something

00:26:47.460 the input changes

00:26:48.600 but it's not

00:26:50.500 just sufficient

00:26:51.020 to how the input

00:26:51.640 changes

00:26:52.100 for you to build

00:26:53.320 a model

00:26:53.820 of the object

00:26:54.640 you're touching

00:26:55.120 and I use

00:26:55.560 the coffee cup

00:26:56.060 example quite a bit

00:26:56.960 because that's

00:26:57.680 how we did it

00:26:58.320 if you move

00:26:59.140 your finger

00:26:59.520 over the coffee cup

00:27:00.360 and you're not

00:27:01.040 even looking

00:27:01.420 at the coffee cup

00:27:01.920 you could learn

00:27:02.480 a model

00:27:02.840 of the coffee cup

00:27:03.420 you could feel

00:27:03.860 just with one finger

00:27:05.320 you could feel

00:27:05.700 like oh

00:27:06.000 this is what

00:27:06.600 its shape is

00:27:07.320 but to do that

00:27:08.900 your brain

00:27:09.760 that cortical column

00:27:10.780 your brain as a whole

00:27:11.600 but that cortical column

00:27:12.500 individually

00:27:13.100 has to know something

00:27:14.300 about where your finger

00:27:15.220 is relative

00:27:16.160 to the cup

00:27:16.720 it's not just

00:27:17.780 a changing pattern

00:27:18.660 that's coming in

00:27:19.360 it has to know

00:27:20.580 how your finger's moving

00:27:21.480 and where your finger

00:27:22.220 is as it touches it

00:27:23.580 so the idea

00:27:25.040 of a reference frame

00:27:25.960 is a way of

00:27:26.680 noting a location

00:27:28.240 you have to have

00:27:28.780 a location signal

00:27:30.060 you have to have

00:27:30.440 some knowledge

00:27:31.080 about where things

00:27:32.540 are in the world

00:27:33.500 relative to other things

00:27:34.760 in this case

00:27:35.280 where's your finger

00:27:36.060 relative to the object

00:27:36.960 you're trying to touch

00:27:37.820 the coffee cup

00:27:38.500 and we realize

00:27:40.200 that for you

00:27:40.740 your brain

00:27:42.140 to make a prediction

00:27:42.900 of what you're going

00:27:43.700 to feel

00:27:44.220 when you touch

00:27:44.800 the edge of the cup

00:27:45.620 and again

00:27:46.400 you mentioned earlier

00:27:47.100 you're not conscious

00:27:47.760 of this

00:27:48.100 you'd reach the cup

00:27:48.860 and you just

00:27:49.300 but your brain's predicting

00:27:50.360 what all your fingers

00:27:51.540 are going to feel

00:27:52.120 it needs to know

00:27:53.820 where the finger's

00:27:54.280 going to be

00:27:54.720 and it has to know

00:27:56.500 what the object is

00:27:57.120 it's a cup

00:27:57.580 it needs to know

00:27:57.980 where it's going to be

00:27:58.600 and that requires

00:28:00.360 a reference frame

00:28:00.980 a reference frame

00:28:01.860 is just a way

00:28:02.580 of noting a location

00:28:03.960 it's saying

00:28:04.940 relative to this cup

00:28:06.440 your finger's over here

00:28:07.860 not over there

00:28:08.880 not on the handle

00:28:09.620 up at the top

00:28:10.540 whatever it is

00:28:11.260 and this is a deduced property

00:28:14.000 we can say for certainty

00:28:15.020 that this has to exist

00:28:16.240 if your finger's

00:28:17.100 going to make a prediction

00:28:17.800 when it reaches

00:28:18.360 and touches the coffee cup

00:28:19.200 it needs to know

00:28:19.820 where the finger is

00:28:20.520 that location

00:28:21.840 has to be relative

00:28:22.460 to the cup

00:28:22.940 so we can just say

00:28:24.620 for certainty

00:28:25.260 that there need

00:28:26.680 to be reference frames

00:28:27.380 in the brain

00:28:27.760 and this is not

00:28:28.260 a controversial idea

00:28:29.980 what we

00:28:31.120 perhaps is novel

00:28:31.980 is that we realize

00:28:32.680 that these reference frames

00:28:33.460 exist in every

00:28:34.120 cortical column

00:28:34.800 and it's the structure

00:28:36.460 of knowledge

00:28:37.100 it applies to not just

00:28:38.380 what your finger feels

00:28:39.220 on a coffee cup

00:28:39.960 and what you see

00:28:40.580 when you look at it

00:28:41.520 but also how you arrange

00:28:42.980 all your knowledge

00:28:43.680 in the world

00:28:44.200 is stored

00:28:45.360 in these reference frames

00:28:46.320 and so when

00:28:47.440 this we're jumping ahead

00:28:48.760 here many steps

00:28:49.840 but when we think

00:28:51.260 and when we posit

00:28:53.380 when we try to

00:28:54.380 you know

00:28:54.720 reason in our head

00:28:55.900 when even my language

00:28:57.200 right now

00:28:58.240 is where

00:28:59.300 the neurons

00:29:00.200 are walking

00:29:01.080 through locations

00:29:01.820 in reference frames

00:29:02.600 recalling the information

00:29:03.660 stored there

00:29:04.380 and that's what

00:29:05.500 comes into your head

00:29:06.320 or that's what you say

00:29:07.180 so it becomes

00:29:08.600 the core reference

00:29:10.020 the reference frame

00:29:10.700 becomes the core structure

00:29:11.740 for the entire

00:29:12.480 everything you do

00:29:13.540 it's knowledge

00:29:14.120 about the world

00:29:14.680 is in these reference frames

00:29:15.600 yeah you make

00:29:17.000 a strong claim

00:29:17.920 about the primacy

00:29:19.500 of motion

00:29:20.280 right

00:29:20.700 because there's

00:29:21.660 everyone knows

00:29:22.540 that there's part

00:29:23.140 of the cortex

00:29:23.800 devoted to

00:29:25.340 motor action

00:29:26.620 we refer to it

00:29:27.420 as the motor cortex

00:29:28.760 and distinguish it

00:29:30.340 from sensory cortex

00:29:31.280 in that way

00:29:31.980 but it's also true

00:29:33.800 that other regions

00:29:35.040 of the cortex

00:29:35.600 and perhaps

00:29:37.240 every region

00:29:37.980 of the cortex

00:29:38.460 does have some

00:29:40.220 connection

00:29:41.240 to lower structures

00:29:42.400 that can affect

00:29:44.160 motion

00:29:45.000 right

00:29:45.380 so it's not

00:29:45.860 it's not that

00:29:46.980 it's just motor cortex

00:29:48.640 that's in the

00:29:49.340 motion game

00:29:50.240 and by analogy

00:29:51.720 or by direct

00:29:53.160 implication

00:29:53.680 you think of

00:29:55.120 thought

00:29:56.420 as

00:29:57.440 itself

00:29:58.660 being a kind

00:29:59.440 of movement

00:30:00.120 in conceptual

00:30:01.420 space

00:30:02.080 right

00:30:02.260 so there's a mapping

00:30:02.860 of the

00:30:03.260 the sensory world

00:30:04.740 that can really

00:30:05.920 only be accomplished

00:30:06.680 by acting

00:30:07.620 on it

00:30:08.220 you know

00:30:08.600 and therefore

00:30:09.420 moving

00:30:10.620 right

00:30:10.880 so the way

00:30:11.700 to map the cup

00:30:12.580 you know

00:30:13.240 is to touch it

00:30:14.000 with your fingers

00:30:14.580 in the end

00:30:15.520 there is a

00:30:16.680 an analogous

00:30:17.540 kind of motion

00:30:18.420 in conceptual

00:30:20.100 space

00:30:20.760 and you know

00:30:21.500 even you know

00:30:22.260 abstract ideas

00:30:23.680 like I think

00:30:25.100 some of the examples

00:30:25.760 you give in the book

00:30:26.280 are like you know

00:30:26.700 democracy

00:30:27.260 right you know

00:30:28.180 or money

00:30:29.080 or how we

00:30:30.600 understand these things

00:30:31.420 so let's go back

00:30:32.800 to the first thing

00:30:33.440 you said there

00:30:33.880 the idea that

00:30:35.160 there's motor cortex

00:30:36.100 and sensory cortex

00:30:37.020 is sort of

00:30:38.180 no longer

00:30:38.920 considered right

00:30:39.620 as you mentioned

00:30:41.040 we

00:30:41.620 the neurons

00:30:42.900 that

00:30:43.540 in these cortical

00:30:44.440 columns

00:30:44.820 there are certain

00:30:45.360 neurons that are

00:30:46.140 the motor output neurons

00:30:47.400 these are in a

00:30:48.500 particular layer 5

00:30:50.000 as they're called

00:30:50.700 and so in the

00:30:52.220 motor cortex

00:30:52.780 they were really big

00:30:53.720 and they project

00:30:54.280 to the spinal cord

00:30:55.520 and say oh

00:30:56.420 that's how you

00:30:56.920 move your fingers

00:30:57.560 but if you look

00:30:58.780 at the neurons

00:30:59.500 the columns

00:31:00.640 in the visual cortex

00:31:01.760 the parts that get

00:31:02.500 input from the eyes

00:31:03.480 they have the same

00:31:04.880 layer 5 cells

00:31:05.700 and these cells

00:31:06.860 project to

00:31:07.800 a part of the brain

00:31:08.660 called the superior

00:31:09.200 colliculus

00:31:09.660 which is what

00:31:10.920 controls eye motion

00:31:11.860 so this goes

00:31:13.360 against the original

00:31:13.940 idea oh

00:31:14.500 there's sensory cortex

00:31:15.360 and motor cortex

00:31:16.060 no one believes

00:31:17.020 that well

00:31:17.380 I don't know

00:31:17.840 nobody

00:31:18.080 but very few

00:31:19.120 people believe

00:31:19.580 that anymore

00:31:20.020 it's as far

00:31:21.040 as we know

00:31:21.500 every part

00:31:22.160 of the cortex

00:31:22.620 has a motor output

00:31:23.520 and so every part

00:31:24.960 of the cortex

00:31:25.360 is getting some

00:31:26.040 sort of input

00:31:26.600 and it has

00:31:27.200 some motor output

00:31:28.100 and so the basic

00:31:29.780 algorithm of cortex

00:31:30.600 is a sensory

00:31:31.440 motor system

00:31:32.440 it's not divided

00:31:33.660 it's not like

00:31:34.320 we have sensory

00:31:34.940 areas and motor

00:31:36.000 areas

00:31:36.260 as far as we know

00:31:37.240 ever it's been seen

00:31:38.160 there's these

00:31:38.580 motor cells

00:31:39.520 everywhere

00:31:39.980 so we can put

00:31:41.680 that aside

00:31:42.300 now

00:31:43.340 I can

00:31:44.720 very clearly

00:31:46.340 walk you through

00:31:47.160 and in some sense

00:31:48.840 prove

00:31:49.340 from logic

00:31:50.400 that

00:31:51.200 when you're learning

00:31:52.480 what a coffee cup

00:31:53.480 feels like

00:31:54.020 and I could even

00:31:54.520 do this for vision

00:31:55.220 that you have to have

00:31:56.940 this idea of a reference

00:31:57.960 frame

00:31:58.260 that you have to know

00:31:59.780 where your finger is

00:32:00.500 relative to the cup

00:32:01.320 and that's how you

00:32:02.480 build a model of it

00:32:03.140 and so we can build out

00:32:04.600 this cortical column

00:32:05.540 that explains

00:32:06.060 how it does that

00:32:07.060 how do your

00:32:07.580 parts of your cortex

00:32:09.440 representing your fingers

00:32:10.520 are able to learn

00:32:11.360 the structure

00:32:11.760 of a coffee cup

00:32:12.320 now

00:32:13.420 Mountcastle

00:32:14.200 go back to him

00:32:14.900 he said

00:32:15.600 look

00:32:15.840 it's the same

00:32:16.300 algorithm everywhere

00:32:17.020 and he says

00:32:19.200 it looks the same

00:32:19.760 everywhere

00:32:20.060 so it's the same

00:32:20.620 algorithm everywhere

00:32:21.300 so that would sort of

00:32:22.540 say hmm

00:32:23.100 well if I'm thinking

00:32:23.940 about something

00:32:24.640 that doesn't seem

00:32:25.520 like a sensory

00:32:26.660 motor system

00:32:27.440 like I'm not

00:32:28.000 touching something

00:32:28.680 or looking

00:32:29.020 I'm just thinking

00:32:29.700 about something

00:32:30.300 that would

00:32:31.060 if Mountcastle

00:32:31.860 was right

00:32:32.340 then the same

00:32:33.400 basic algorithm

00:32:34.160 would be applying

00:32:34.800 there

00:32:35.060 so that was

00:32:35.540 one constraint

00:32:36.200 like well

00:32:36.700 that you know

00:32:38.000 and the evidence

00:32:38.780 is that Mountcastle

00:32:39.520 is right

00:32:39.800 I mean the physical

00:32:41.060 evidence suggests

00:32:41.740 he's right

00:32:42.140 it just becomes

00:32:43.080 a little bit odd

00:32:43.580 to think like

00:32:44.060 well how is language

00:32:45.140 like this

00:32:45.620 and how is

00:32:46.080 mathematics like

00:32:47.260 you know

00:32:47.640 touching a coffee cup

00:32:48.520 but then we realize

00:32:50.100 that it's just

00:32:51.060 reference frames

00:32:51.840 are a way of

00:32:52.220 storing everything

00:32:53.020 and in the way

00:32:55.260 we move

00:32:55.860 through a reference

00:32:56.460 frame

00:32:56.660 it's like

00:32:57.160 how do you move

00:32:57.720 from one location

00:32:58.460 how do the neurons

00:32:59.460 activate one location

00:33:01.280 after another location

00:33:02.160 after another location

00:33:03.000 we do that

00:33:04.160 to this idea

00:33:05.420 of movement

00:33:05.860 so I'm moving

00:33:06.920 if I want to

00:33:07.460 access the locations

00:33:08.680 on a coffee cup

00:33:09.300 I move my finger

00:33:10.080 but the same concept

00:33:11.600 could apply to mathematics

00:33:12.700 or to politics

00:33:14.460 but you're not

00:33:15.480 actually physically

00:33:16.300 moving something

00:33:17.000 but you're still

00:33:17.800 walking through

00:33:18.980 a structure

00:33:19.560 a good bridge example

00:33:21.800 is if I say to you

00:33:23.440 you know

00:33:23.680 imagine your house

00:33:24.540 and you know

00:33:25.560 I ask you to walk

00:33:26.440 you know

00:33:26.800 tell me about your house

00:33:27.640 what you'll do

00:33:28.240 is you'll mentally

00:33:29.560 imagine walking

00:33:30.600 through your house

00:33:31.340 it won't be random

00:33:32.600 you just won't have

00:33:33.180 random thoughts

00:33:33.800 come to your head

00:33:34.340 but you will mentally

00:33:35.880 imagine walking

00:33:36.640 through your house

00:33:37.180 and as you walk

00:33:37.880 through your house

00:33:38.440 you'll recall

00:33:39.360 what is supposed

00:33:40.120 to be seen

00:33:40.580 in different directions

00:33:41.320 you can say

00:33:41.740 oh I'll walk

00:33:42.180 in the front door

00:33:42.740 and I'll look

00:33:43.060 to the right

00:33:43.420 what do I see

00:33:44.000 I'll look to the left

00:33:44.580 what do I see

00:33:45.100 this is sort of

00:33:46.720 an example

00:33:47.640 you could relate

00:33:48.240 it to something

00:33:48.640 physically

00:33:49.020 you could move to

00:33:49.840 but that's pretty much

00:33:51.180 what's going on

00:33:51.780 when you're thinking

00:33:52.200 about anything

00:33:52.760 if you're thinking

00:33:53.880 about your podcast

00:33:54.600 and how you get

00:33:55.380 more subscribers

00:33:56.080 you have a model

00:33:57.200 of that in your head

00:33:57.980 and you're

00:33:58.360 you are

00:33:59.160 trying it out

00:34:00.960 thinking about

00:34:01.540 different aspects

00:34:02.180 by literally

00:34:02.860 invoking these

00:34:03.740 different locations

00:34:04.440 and reference frames

00:34:05.240 and so that's

00:34:06.800 sort of the core

00:34:07.540 of all knowledge

00:34:08.200 yeah it's interesting

00:34:09.380 I guess back to

00:34:10.720 Mountcastle for a second

00:34:11.640 one piece of evidence

00:34:13.360 in favor of this view

00:34:14.600 of a common

00:34:15.760 cortical algorithm

00:34:16.840 is the fact that

00:34:18.220 adjacent areas

00:34:19.180 of cortex

00:34:19.740 can be

00:34:20.460 appropriated

00:34:22.020 by

00:34:22.800 various functions

00:34:24.580 you know

00:34:25.160 if you

00:34:25.960 lose your vision

00:34:27.080 say

00:34:27.440 classical visual cortex

00:34:29.760 can be appropriated

00:34:30.720 by other senses

00:34:31.940 and there's this

00:34:32.700 plasticity

00:34:33.640 that can

00:34:35.000 ignore

00:34:35.620 some of the

00:34:36.240 previous boundaries

00:34:37.480 between

00:34:38.240 separate senses

00:34:39.560 in the cortex

00:34:40.560 yeah that's right

00:34:41.880 there's this

00:34:42.700 tremendous plasticity

00:34:43.740 and

00:34:44.360 and you can also

00:34:45.580 recover from

00:34:46.240 various sorts of

00:34:47.020 trauma and so on

00:34:47.880 I mean there's

00:34:48.500 some rewiring

00:34:49.160 has to occur

00:34:49.760 but it does show

00:34:51.060 that that

00:34:51.500 whatever's going

00:34:52.260 whatever the

00:34:52.800 circuitry in the

00:34:53.380 visual cortex

00:34:53.960 was you know

00:34:55.000 quote

00:34:55.480 if you were

00:34:56.880 a sighted person

00:34:57.660 what it would do

00:34:58.740 if you're not

00:34:59.560 a sighted person

00:35:00.200 well it'll just

00:35:00.840 do something else

00:35:01.620 and so

00:35:02.560 it's not

00:35:03.320 and so

00:35:04.080 that is a

00:35:04.680 very very strong

00:35:05.360 argument for that

00:35:06.580 there's a famous

00:35:07.220 scientist

00:35:07.680 Bacharita

00:35:08.880 who did

00:35:09.360 an experiment

00:35:10.600 where he

00:35:11.160 I'm trying to

00:35:12.080 remember the

00:35:12.400 animal he used

00:35:13.120 maybe you can

00:35:14.120 recall it

00:35:14.600 but anyway

00:35:14.980 it'll come to me

00:35:16.220 a ferret

00:35:16.980 I think it was a

00:35:17.480 ferret

00:35:17.680 we took the

00:35:18.640 they took a

00:35:19.500 before the

00:35:20.480 animal's born

00:35:21.100 he took the

00:35:21.560 optic nerve

00:35:22.220 and ran it

00:35:23.100 over to one

00:35:23.600 part of the

00:35:24.040 a different part

00:35:24.760 of the neocortex

00:35:25.280 and took the

00:35:25.820 auditory nerve

00:35:26.340 and ran it

00:35:26.680 to a different

00:35:26.980 part of the neocortex

00:35:27.940 you know

00:35:28.380 basically rewired

00:35:30.100 the animal

00:35:30.520 I'm not sure

00:35:30.880 we do these

00:35:31.360 experiments today

00:35:32.040 but and you

00:35:33.420 know and the

00:35:34.260 argument was that

00:35:34.920 the animals you

00:35:35.820 know still saw

00:35:36.520 and still heard

00:35:37.200 and so on

00:35:37.680 maybe not as

00:35:38.340 well as an

00:35:39.660 unaltered one

00:35:40.540 but the evidence

00:35:41.660 was that yeah

00:35:42.360 that really works

00:35:43.240 so what is

00:35:45.500 genetically determined

00:35:46.900 and what is

00:35:48.360 learned here

00:35:49.020 it seems that

00:35:50.040 the genetics

00:35:51.100 at minimum

00:35:51.880 are determining

00:35:52.900 what is hooked

00:35:55.020 up to what

00:35:55.820 initially

00:35:56.580 right you know

00:35:57.400 barring

00:35:57.660 yeah that was

00:35:58.220 roughly roughly

00:35:59.260 that's right

00:35:59.760 I think you

00:36:00.720 know like where

00:36:01.380 do the eyes

00:36:01.860 the optic nerve

00:36:02.740 from the eyes

00:36:03.180 where do they

00:36:03.620 project and

00:36:04.280 where do the

00:36:04.720 regions that get

00:36:05.440 the input from

00:36:06.580 the eyes where

00:36:07.040 do they project

00:36:07.640 and so this

00:36:08.880 rough sort of

00:36:10.080 overall architecture

00:36:11.240 is specified

00:36:12.620 and as we

00:36:13.800 just talked

00:36:14.240 through trauma

00:36:14.840 and other reasons

00:36:15.500 sometimes that

00:36:16.100 architecture can

00:36:16.820 get rewired

00:36:17.660 I think also

00:36:18.940 the basic

00:36:20.780 algorithm that

00:36:21.840 goes on in

00:36:22.500 each of these

00:36:22.900 cortical columns

00:36:23.760 the circuitry

00:36:24.780 inside the

00:36:26.080 neocortex is

00:36:26.780 pretty well

00:36:27.180 determined by

00:36:27.900 genetics and

00:36:29.940 in fact what

00:36:30.740 one of

00:36:31.040 myocast's

00:36:31.520 arguments was

00:36:32.300 that humans

00:36:33.120 the human

00:36:33.980 neocortex got

00:36:34.780 large and

00:36:35.940 we have a

00:36:36.280 very large

00:36:36.720 one relative

00:36:37.240 to our body

00:36:37.780 size just

00:36:38.680 because all

00:36:39.160 evolution had

00:36:40.420 to do is

00:36:40.800 discover just

00:36:41.360 make more

00:36:41.740 copies of

00:36:42.300 these columns

00:36:42.820 you don't

00:36:43.380 have to do

00:36:44.400 anything new

00:36:45.040 just make

00:36:45.480 more copies

00:36:46.000 and that's

00:36:46.300 something easy

00:36:46.860 for genes

00:36:47.300 to specify

00:36:47.880 and so

00:36:49.060 human brains

00:36:49.940 got large

00:36:50.680 quickly in

00:36:51.420 evolutionary time

00:36:52.300 by that just

00:36:53.320 replicate more

00:36:54.180 of it type

00:36:54.720 of thing

00:36:55.040 okay so

00:36:56.800 let's go

00:36:57.640 beyond the

00:36:58.700 human now

00:36:59.540 and talk

00:37:00.400 about artificial

00:37:01.940 intelligence

00:37:02.580 and before

00:37:04.900 we talk about

00:37:05.380 the risks

00:37:06.000 or the

00:37:06.620 imagined risks

00:37:07.860 tell me

00:37:09.380 what you

00:37:09.860 think the

00:37:11.040 path looks

00:37:11.840 like going

00:37:12.640 forward

00:37:13.040 what are

00:37:13.580 we doing

00:37:13.840 now and

00:37:14.460 what do

00:37:14.660 you think

00:37:14.900 we need

00:37:15.420 to do

00:37:15.900 to have

00:37:16.760 our dreams

00:37:17.320 of true

00:37:18.400 artificial

00:37:19.220 general

00:37:19.820 intelligence

00:37:20.380 realized

00:37:21.260 well

00:37:22.040 you know

00:37:23.680 today's

00:37:24.280 AI

00:37:24.680 as powerful

00:37:25.880 as it is

00:37:26.540 and successful

00:37:27.160 as it is

00:37:27.900 I think

00:37:29.260 most senior

00:37:30.700 AI practitioners

00:37:31.640 will admit

00:37:32.420 and many of

00:37:33.900 them have

00:37:34.340 that they

00:37:35.220 don't really

00:37:35.520 think they're

00:37:35.960 intelligent

00:37:36.380 you know

00:37:37.140 they're

00:37:37.420 they're

00:37:37.820 really

00:37:38.060 wonderful

00:37:38.460 pattern

00:37:38.860 classifiers

00:37:39.480 and they

00:37:39.920 can do

00:37:40.160 all kinds

00:37:40.500 of clever

00:37:40.800 things

00:37:41.260 but there

00:37:42.520 are very

00:37:42.840 few

00:37:43.080 practitioners

00:37:43.600 would say

00:37:44.220 hey this

00:37:44.720 AI system

00:37:45.420 that's

00:37:45.780 recognizing

00:37:46.180 faces is

00:37:46.860 really

00:37:47.080 intelligent

00:37:47.540 and there's

00:37:48.960 sort of a

00:37:49.420 lack of

00:37:49.800 understanding

00:37:50.360 what intelligence

00:37:51.300 is and how

00:37:52.080 to go forward

00:37:52.700 and how do

00:37:53.400 you make a

00:37:53.760 system that

00:37:54.200 could solve

00:37:55.260 general

00:37:55.720 problems

00:37:56.380 could do

00:37:57.220 more than

00:37:57.580 one thing

00:37:58.140 right

00:37:58.520 and so

00:38:00.200 in the second

00:38:00.920 part of my

00:38:01.260 book I lay

00:38:01.840 out what I

00:38:02.620 believe are the

00:38:03.360 requirements to

00:38:04.040 do that

00:38:04.460 and my

00:38:05.440 approach has

00:38:05.920 always been

00:38:06.580 for 40

00:38:07.660 years has

00:38:08.240 been like

00:38:08.660 well I

00:38:09.040 think we

00:38:09.320 need to

00:38:09.540 first figure

00:38:09.960 out what

00:38:10.260 brains do

00:38:10.940 and how

00:38:12.340 they do

00:38:12.660 them and

00:38:13.320 then we'll

00:38:13.840 know how

00:38:14.180 to build

00:38:14.480 intelligent

00:38:14.840 machines

00:38:15.200 because we

00:38:16.340 just don't

00:38:16.820 seem able

00:38:17.700 to intuit

00:38:18.380 what an

00:38:19.060 intelligent

00:38:19.320 machine is

00:38:20.060 so I

00:38:21.560 think what

00:38:22.580 I the way

00:38:23.100 I look at

00:38:23.500 this problem

00:38:24.040 if we want

00:38:24.660 to make you

00:38:25.400 know what's

00:38:25.680 the what's

00:38:26.100 the recipe

00:38:26.560 for making

00:38:27.020 an intelligent

00:38:27.420 machine is

00:38:28.960 you have to

00:38:29.400 say what are

00:38:29.840 the principles

00:38:30.380 by which the

00:38:31.100 brain works

00:38:31.600 that we need

00:38:32.080 to replicate

00:38:32.540 and which

00:38:32.940 principles don't

00:38:33.620 we need to

00:38:34.000 replicate and

00:38:35.380 so I made

00:38:36.420 a list of

00:38:36.820 these in

00:38:37.100 the book

00:38:37.360 but if you

00:38:38.980 can think of

00:38:39.300 a very high

00:38:39.820 level they

00:38:40.460 have to have

00:38:40.920 some sort of

00:38:41.520 embodiment they

00:38:42.260 have to have

00:38:42.580 the ability to

00:38:43.160 move their

00:38:43.600 sensors somehow

00:38:44.700 in the world

00:38:45.380 you know you

00:38:46.360 can't really

00:38:47.480 learn how to

00:38:48.760 use tools and

00:38:49.580 how to you

00:38:50.460 know run

00:38:51.220 factories and

00:38:52.000 and how to

00:38:53.120 you know do

00:38:53.520 things unless you

00:38:54.280 can move in the

00:38:54.840 world and

00:38:56.160 it requires these

00:38:57.540 reference frames I

00:38:58.380 was talking about

00:38:58.920 because movement

00:38:59.640 requires reference

00:39:00.680 frames but that's

00:39:01.540 not a controversial

00:39:02.200 statement it's

00:39:03.040 this it's just a

00:39:04.040 fact you're going

00:39:05.060 to have to have

00:39:05.500 know where things

00:39:06.480 are in the world

00:39:07.320 and and then the

00:39:08.900 final there's a

00:39:10.340 set of things but

00:39:11.320 one of the other

00:39:12.500 big ones which we

00:39:13.500 haven't talked about

00:39:14.180 yet and which is

00:39:14.820 where the title of

00:39:15.500 the book comes

00:39:15.940 from a thousand

00:39:17.400 brains is that the

00:39:18.840 way to think about

00:39:19.500 our near cortex it

00:39:21.180 has 150,000 of

00:39:22.480 these columns we

00:39:23.700 have essentially

00:39:24.260 150,000 separate

00:39:26.120 modeling systems going

00:39:27.240 on in our brain and

00:39:28.660 they work together

00:39:29.580 by voting and so

00:39:31.960 that concept of a

00:39:33.220 distributed

00:39:33.840 intelligence system

00:39:35.600 is important we're

00:39:37.200 not just one thing

00:39:38.160 we it feels like

00:39:39.140 we're one thing but

00:39:40.120 we're really 150,000

00:39:41.240 of these things and

00:39:42.640 we're only conscious

00:39:43.460 of being one thing

00:39:44.460 but that's not

00:39:45.060 really what's

00:39:45.440 happening under the

00:39:46.020 covers so those

00:39:47.960 are some of the

00:39:48.400 key ideas I would

00:39:49.440 just stick to very

00:39:50.160 very high ideas it

00:39:50.960 has to have an

00:39:51.540 embodiment has to

00:39:52.340 be able to move

00:39:52.960 its sensors has to

00:39:54.520 be able to organize

00:39:55.260 information and

00:39:56.020 reference frames and

00:39:57.640 it has to be

00:39:58.320 distributed and

00:39:59.500 that's how we

00:40:00.000 can do multiple

00:40:01.320 sensors and

00:40:02.160 sensory integration

00:40:03.000 things like that

00:40:03.700 I guess I question

00:40:06.840 the criteria of

00:40:09.140 embodiment and

00:40:10.720 movement right I

00:40:12.680 mean I understand

00:40:13.580 that practically

00:40:15.060 speaking that's how

00:40:16.660 a useful intelligence

00:40:17.660 can get trained up

00:40:18.920 in our world to

00:40:20.640 do things you know

00:40:22.040 physically in our

00:40:23.120 world but it

00:40:24.480 seems like you

00:40:24.920 could have a

00:40:25.580 perfectly intelligent

00:40:27.060 system you know

00:40:28.220 i.e. a mind

00:40:29.660 that is turned

00:40:32.340 loose on you

00:40:34.160 know simulated

00:40:34.800 worlds and are

00:40:36.180 capable of solving

00:40:37.280 problems that don't

00:40:38.560 require effectors of

00:40:41.280 any kind I mean

00:40:41.960 you know chess is

00:40:43.220 obviously a very low

00:40:44.680 level analogy but

00:40:45.760 just imagine a

00:40:46.940 thousand things like

00:40:48.040 chess that represent

00:40:49.680 real you know

00:40:51.560 theory building or

00:40:52.700 cognition you know

00:40:53.960 in a box yeah I

00:40:56.040 I think you're

00:40:56.740 right and and so

00:40:57.620 when I use the

00:40:58.820 word movement or

00:40:59.580 embodiment and I

00:41:00.880 I'm careful to

00:41:01.620 define it in the

00:41:02.200 book because it

00:41:03.660 doesn't have to be

00:41:04.480 physical it you

00:41:06.980 know I example I

00:41:08.240 gave you can imagine

00:41:09.200 intelligent agent that

00:41:10.440 lives in the

00:41:11.020 internet and right

00:41:12.700 in movement is

00:41:13.640 following links right

00:41:15.040 it's not a physical

00:41:15.840 thing but there's

00:41:17.420 still this conceptual

00:41:18.980 mathematical idea of

00:41:20.140 what it means to

00:41:20.640 move yeah and so

00:41:22.040 and so you're

00:41:22.880 changing the location

00:41:23.960 of of some

00:41:25.420 representation and

00:41:26.940 that could be

00:41:27.400 virtual it could be

00:41:28.460 you know it doesn't

00:41:29.360 have to have a

00:41:29.800 physical embodiment

00:41:31.120 but but in the end

00:41:32.620 you can't you can't

00:41:34.240 learn about the world

00:41:34.960 just by looking at a

00:41:35.660 set of pictures

00:41:36.300 right that's not

00:41:38.040 not going to happen

00:41:38.760 you can learn to

00:41:39.440 classify pictures but

00:41:40.620 so so some AI

00:41:43.100 systems will have to

00:41:44.100 be physically embodied

00:41:45.340 like a like like a

00:41:46.800 robot if I guess if

00:41:47.760 you want many will

00:41:48.840 not be many will be

00:41:49.840 virtual but they

00:41:51.040 all have this

00:41:51.780 internal process

00:41:52.820 which we could I

00:41:53.900 could point to the

00:41:54.640 thing that says here's

00:41:56.040 where the reference

00:41:56.620 frame is here's where

00:41:57.400 your current location

00:41:58.160 is here's how it's

00:41:58.980 moving to a new

00:41:59.600 location based on some

00:42:00.700 movement vector you

00:42:02.040 know like a verb and a

00:42:03.160 word can you can think

00:42:04.140 of that as like an

00:42:04.680 action and so you can

00:42:06.720 have an action that's

00:42:07.820 not physical but it's

00:42:08.880 still an action and it

00:42:09.760 moves to a new location

00:42:10.720 in this internal

00:42:11.640 representation right

00:42:12.860 right okay well

00:42:15.020 let's talk about risk

00:42:15.940 because this is the

00:42:16.660 place where I think you

00:42:17.560 and I have very

00:42:18.380 different intuitions

00:42:20.120 you you are as far

00:42:22.540 as I can tell from

00:42:23.300 your book you seem

00:42:24.980 very sanguine about

00:42:26.940 AI risk and really

00:42:30.700 you seem to think

00:42:32.580 that the only real

00:42:33.620 risk and the serious

00:42:34.800 risk of things going

00:42:35.940 very badly for us is

00:42:37.440 that bad people will

00:42:39.740 do bad things with

00:42:41.060 much more powerful

00:42:41.960 tools so the

00:42:44.100 heuristic here would

00:42:44.940 be you know don't

00:42:45.560 give your super

00:42:46.200 intelligent AI to the

00:42:47.440 next Hitler because

00:42:49.020 that would be bad

00:42:49.780 but other than that

00:42:51.420 the generic problem

00:42:53.080 of self-replication

00:42:54.380 which you talk about

00:42:55.820 briefly and we you

00:42:57.380 point out we face that

00:42:58.460 on other fronts like

00:42:59.440 with you know with the

00:43:00.440 pandemic where we've

00:43:01.880 been dealing with

00:43:02.640 I mean so natural

00:43:03.260 viruses and bacteria

00:43:04.440 or computer viruses

00:43:06.220 I mean there's there's

00:43:06.980 anything that can

00:43:07.700 self-replicate can be

00:43:09.300 dangerous but that

00:43:11.480 aside you seem quite

00:43:13.700 confident that AI will

00:43:15.120 not get away from us

00:43:16.640 there won't be an

00:43:17.840 intelligence explosion

00:43:18.880 and we don't have to

00:43:21.340 worry too much about

00:43:22.260 the so-called

00:43:22.960 alignment problem

00:43:23.940 and at one point you

00:43:25.620 even question whether

00:43:26.920 it makes sense to

00:43:28.080 expect that we'll

00:43:29.060 produce something that

00:43:31.000 can be appropriately

00:43:31.940 called superhuman

00:43:33.140 intelligence so perhaps

00:43:35.140 you can explain the

00:43:36.760 basis for your optimism

00:43:38.500 here so I think what

00:43:41.400 most people and perhaps

00:43:42.500 yourself have fears

00:43:44.160 about is is they they

00:43:46.760 they use humans as an

00:43:48.680 example of how things

00:43:50.240 can go wrong and so we

00:43:51.960 think about the

00:43:52.420 alignment problem or

00:43:53.440 you think about you

00:43:54.700 know motivations of an

00:43:55.820 AI system well okay

00:43:57.780 does the AI system have

00:43:59.020 motivations or not does

00:44:00.500 it have a desire to do

00:44:01.660 anything now as a human

00:44:03.600 an animal we all have

00:44:04.600 desires right but if you

00:44:06.940 if you take apart what

00:44:08.880 parts of the human brain

00:44:10.300 are are doing different

00:44:12.340 parts there's some parts

00:44:14.160 that are just building

00:44:14.860 this model of the world

00:44:15.860 and this is the core of

00:44:17.400 our intelligence this is

00:44:18.320 what it means to be

00:44:19.620 intelligent that part

00:44:21.220 itself is is is benign

00:44:23.200 it has no motivations on

00:44:24.680 its own it doesn't desire

00:44:25.940 to do anything I use an

00:44:27.620 example of a map you know

00:44:28.900 a map is a model of the

00:44:30.040 world and you can a map

00:44:33.540 can be very powerful tool

00:44:36.960 for some to do good or to

00:44:38.180 do bad but on its own the

00:44:39.940 map doesn't do anything

00:44:40.920 so if you think about

00:44:42.220 the neocortex on its own

00:44:43.920 it it sits on top of the

00:44:45.540 rest of your brain and

00:44:47.060 the rest of your brain is

00:44:48.120 really what makes us

00:44:49.360 motivated it gets us you

00:44:50.960 know we have our our our

00:44:52.680 good sides and our bad

00:44:53.860 sides you know our desire

00:44:55.140 to maintain our life and

00:44:57.000 have sex and aggression

00:44:58.680 and all this stuff the

00:44:59.960 neocortex is just sitting

00:45:00.900 there it's like a map it

00:45:01.740 says you know I

00:45:02.400 understand the world and

00:45:03.500 you can use me as how as

00:45:04.500 you want so when we did

00:45:05.880 when we build intelligent

00:45:06.780 machines we have the

00:45:08.920 option and and I think

00:45:10.680 almost the imperative not

00:45:12.200 to build the old parts of

00:45:13.240 the brain to you know why

00:45:14.360 do that we just just have

00:45:16.080 this thing which is

00:45:17.200 inherently smart but on

00:45:19.180 its own doesn't really want

00:45:19.940 to do anything and and

00:45:21.760 so there's some of the

00:45:22.700 some of the risks that

00:45:23.700 come about from the

00:45:25.340 people's fears about the

00:45:27.340 alignment problem

00:45:28.220 specifically is that the

00:45:31.280 the intelligent agent will

00:45:32.900 decide on its own or

00:45:34.000 decide for some reason to

00:45:35.840 do things that are in

00:45:36.920 its best interest not in

00:45:38.060 our best interest or

00:45:38.840 maybe it'll listen to us

00:45:40.440 but then not listen to

00:45:41.320 us or something like

00:45:42.100 this I just don't see

00:45:43.740 how that can physically

00:45:44.760 happen and and for

00:45:46.680 people most people don't

00:45:47.800 understand the separation

00:45:49.020 they just assume that

00:45:50.040 this intelligence is

00:45:50.900 wrapped up in these all

00:45:52.020 these all the things that

00:45:52.960 make us human the

00:45:54.420 intelligence explosion

00:45:55.300 problem is a separate

00:45:56.220 issue I'm not sure

00:45:58.340 which one of those you're

00:45:59.040 more worried about

00:45:59.940 yeah well let's let's

00:46:01.380 deal with the alignment

00:46:03.360 issue first I mean I do

00:46:05.240 think that's more

00:46:06.040 critical but let's see

00:46:08.700 if I can capture what

00:46:09.640 troubles me about this

00:46:10.840 picture you painted

00:46:12.500 here it seems that

00:46:13.700 you're to my mind

00:46:15.140 you're you're being

00:46:16.300 strangely anthropomorphic

00:46:20.100 on one side but not

00:46:22.260 anthropomorphic enough

00:46:23.880 on the other I mean so

00:46:25.400 like you know you think

00:46:27.180 that to understand

00:46:28.680 intelligence and and

00:46:29.940 actually truly implement

00:46:31.080 it in machines we

00:46:32.720 really have to be focused

00:46:35.460 on ourselves first and

00:46:37.080 we have to understand

00:46:38.000 how the human brain

00:46:39.160 works and then emulate

00:46:40.460 those principles pretty

00:46:42.660 directly in machines

00:46:44.440 that strikes me as

00:46:46.120 possibly true but

00:46:47.240 possibly not true and

00:46:48.560 if if I had to bet I

00:46:50.320 think I would probably

00:46:51.780 bet against it although

00:46:53.400 even here you seem to

00:46:55.080 be not taking full

00:46:58.060 account of what the

00:46:59.660 human brain is doing I

00:47:00.620 mean like we you know we

00:47:01.500 can't partition reason

00:47:03.580 and emotion as clearly

00:47:05.600 as we thought we could

00:47:07.360 hundreds of years ago and

00:47:08.440 in fact you know certain

00:47:09.680 emotions you know certain

00:47:10.860 drives are built into our

00:47:12.800 being able to reason

00:47:14.560 effectively I think that's

00:47:16.580 you know I I'll take an

00:47:17.780 exception to that I know I

00:47:19.320 know this is a opinion

00:47:20.500 that you had Lisa Barrett on

00:47:23.380 your program recently

00:47:24.100 yeah like Antonio Damasio is

00:47:25.860 the person who's banged on

00:47:27.160 about this the most yeah I

00:47:28.260 know and I just disagree I

00:47:30.180 just it's you know you

00:47:31.940 can separate these two and

00:47:34.080 and I can say this because

00:47:36.440 I understand actually what's

00:47:38.180 going on in the neocortex

00:47:39.620 and I can see what I have a

00:47:42.080 very good sense of what

00:47:43.100 these actual neurons are

00:47:44.060 actually doing when it's

00:47:45.240 modeling the world and so

00:47:46.300 on and in you do not it

00:47:48.800 does not require this

00:47:50.000 emotional component a human

00:47:51.880 does now you say you know

00:47:53.800 on one hand I don't argue

00:47:55.440 we should replicate the

00:47:56.340 brain I say we should

00:47:57.200 replicate the structures

00:47:58.240 of the neocortex right

00:47:59.580 which is not replicating

00:48:00.600 the brain it's just one

00:48:02.600 part of the brain and so

00:48:04.120 I'm specifically saying

00:48:05.320 you know I don't really

00:48:06.860 care too much about how

00:48:07.740 the spinal cord works or

00:48:09.000 how you know the brain

00:48:10.380 stem does this or that

00:48:11.560 it's interesting maybe I

00:48:12.960 know a little bit about

00:48:13.580 it but that's not

00:48:14.700 important the cortex sits

00:48:15.840 on top of another

00:48:16.760 structure and the cortex

00:48:17.820 does its own thing and

00:48:19.100 they interact of course

00:48:20.200 they interact and our

00:48:21.980 emotions affect what we

00:48:23.300 learn and what we don't

00:48:24.040 learn but it doesn't

00:48:25.080 have to be that way in a

00:48:27.180 system another system

00:48:28.580 that we build that's the

00:48:29.980 way humans are

00:48:30.580 structured yeah okay so

00:48:31.400 I would I would agree

00:48:32.380 with that except the

00:48:33.600 boundary between what is

00:48:35.880 an emotion or a drive or

00:48:37.700 a motivation or a goal and

00:48:40.060 what is a value neutral

00:48:42.560 mapping of reality you know

00:48:45.220 I think that boundary is

00:48:46.840 perhaps harder to specify

00:48:48.900 than than than you think it

00:48:51.340 is and that certain of

00:48:53.640 these things are connected

00:48:54.920 right which is to I mean

00:48:56.560 here's an example this is

00:48:57.660 probably not a perfect

00:48:59.200 analogy but this gets at

00:49:00.980 some of the surprising

00:49:01.940 features of cognition that

00:49:03.340 may await us so we think

00:49:05.420 intuitively that

00:49:07.040 understanding a proposition

00:49:09.260 is cognitively quite

00:49:12.100 distinct from believing it

00:49:14.160 right so like I can give you

00:49:15.600 a statement that you can

00:49:17.320 believe or disbelieve or be

00:49:19.140 uncertain about and I can

00:49:20.340 say you know there's two

00:49:21.600 plus two equals four two

00:49:22.760 plus two equals five and

00:49:24.640 that can give you some

00:49:25.480 gigantic number and say

00:49:26.720 this number is prime and

00:49:28.380 presumably in the first

00:49:30.060 condition you'll say yes I

00:49:31.660 believe that in the second

00:49:32.700 you'll say no that's false

00:49:33.960 and in the third you won't

00:49:35.740 know whether or not it's

00:49:37.520 prime or not so those are

00:49:39.240 distinct states that we can

00:49:40.720 intuitively differentiate but

00:49:42.400 there's also evidence to

00:49:44.440 suggest that merely

00:49:46.060 comprehending a statement if I

00:49:47.840 give you a statement and you

00:49:49.040 parse it successfully the

00:49:51.520 parsing itself contains an

00:49:54.500 actual default acceptance of

00:49:56.840 it as true and rejecting it

00:50:00.080 as false is a separate

00:50:01.920 operation added to that I

00:50:03.700 mean and there's there's not

00:50:04.720 a ton of evidence for this

00:50:05.740 but there's certainly some

00:50:07.280 behavioral evidence so if I

00:50:08.580 put you in a paradigm where

00:50:10.340 we gave you statements that

00:50:12.360 were true and false and all

00:50:13.560 you had to do was judge them

00:50:14.800 true and false and they were

00:50:16.260 all matched for complexity so

00:50:18.760 you know two plus two equals

00:50:20.020 four is no more or less

00:50:22.040 complex than two plus two

00:50:23.220 equals five but it'll take

00:50:25.040 you longer systematically

00:50:26.640 longer to judge very simple

00:50:28.920 statements to be false than

00:50:30.260 to judge them to be true

00:50:31.540 suggesting that you're doing a

00:50:33.340 further operation now we can

00:50:35.440 remain agnostic as to whether

00:50:37.220 or not that's actually true but

00:50:38.500 if true it's counterintuitive

00:50:41.160 that merely understanding

00:50:43.060 something entails you know some

00:50:45.240 credence giving epistemic

00:50:47.400 credence given to it by default

00:50:49.100 and that to reject it as false

00:50:51.400 represents a subsequent act but

00:50:54.220 like that's the kind of thing

00:50:55.200 that you know already we're on

00:50:56.720 territory that is not coldly

00:50:59.540 rational some of the all too

00:51:02.120 apish appetites have kind of

00:51:03.500 crept into cognition here in

00:51:05.260 ways that we didn't really budget

00:51:08.340 for and so the question is just

00:51:11.040 how much of that is avoidable in

00:51:13.400 building a new type of mind

00:51:15.240 well I you know I'm not familiar

00:51:17.500 with that specific research and

00:51:19.540 so I haven't heard of that but I

00:51:22.300 you know to me none of these

00:51:23.660 things are surprising in any way

00:51:26.200 you just if you start thinking

00:51:28.580 about the brain is basically

00:51:30.380 trying to build models it's

00:51:31.900 constantly trying to build models

00:51:33.140 in fact you're as you walk around

00:51:35.440 the your life day-to-day moment to

00:51:37.800 moment and you see things you're

00:51:39.600 building the model the model is

00:51:40.680 being constructed even like where are

00:51:42.100 things in the refrigerator right

00:51:43.260 now your brain will update you

00:51:44.920 open the fridge oh the milk's on

00:51:46.060 the left today whatever and so

00:51:47.620 if someone gives you a

00:51:48.440 proposition like two plus two

00:51:49.840 equals five you know I don't

00:51:51.560 know what the evidence that you

00:51:52.540 believe it and then falsify it

00:51:54.340 but I certainly imagine you can

00:51:55.980 imagine it trying to see if it's

00:51:57.520 right it'd be like me saying to

00:51:58.760 you hey you know Sam the milk was

00:52:00.520 on the right in your

00:52:01.140 refrigerator and you'd have to

00:52:02.780 think about it for a second you

00:52:03.780 say well let me think okay I

00:52:05.040 opened it no last time I saw it

00:52:06.240 was on the left you know no

00:52:07.300 that's wrong but you would walk

00:52:08.960 through the process of trying to

00:52:10.420 imagine it and trying to see does

00:52:12.600 that fit my model and yes or no

00:52:14.580 and I can I don't it's not

00:52:16.380 surprising to me that would that

00:52:18.760 you would have to process it the

00:52:20.520 way as if it was true it's just

00:52:22.120 matter saying can you imagine this

00:52:23.640 go imagine it do you think it's

00:52:24.780 right it's not like I believe that

00:52:27.040 now I falsified it it's it's more

00:52:29.140 like well actually I'll just give you

00:52:30.600 one other datum here so because

00:52:32.500 it's just intellectually interesting

00:52:33.880 and socially all too

00:52:35.120 consequential this effect goes by

00:52:38.240 several names I think but you know

00:52:40.240 one is the illusory truth effect

00:52:41.980 which is even in the act of

00:52:43.760 disconfirming something you know to

00:52:46.040 be false you know a some specious

00:52:47.920 rumor or conspiracy theory merely

00:52:50.640 having to invoke it I mean have

00:52:52.680 people entertain the concept again

00:52:54.720 even in the context of debunking it

00:52:56.640 ramifies a belief in it in many many

00:53:01.500 people it's just oh yeah it becomes

00:53:03.060 harder to discredit things because

00:53:05.040 you have to talk about them in the

00:53:06.180 first place yeah I mean so look

00:53:08.320 we're talking about language here

00:53:10.100 right yeah and in language so much

00:53:12.800 of what we humans know is via

00:53:14.320 language and we have no idea if it's

00:53:16.420 true when someone says something to

00:53:17.860 you right how do you know and so you

00:53:20.520 it's you have to so I mean I gave an

00:53:23.660 example like I've never been to the

00:53:25.100 city of Havana well I believe it's

00:53:27.280 there I believe it's true but I don't

00:53:28.940 know I've never been there I've never

00:53:30.240 actually touched it or smelled it or

00:53:31.720 saw it so maybe it's false so I just

00:53:35.280 I mean this is one of the issues we

00:53:37.440 have I have a whole chapter on false

00:53:38.740 beliefs because so much of our

00:53:40.940 knowledge of the world is built up on

00:53:42.440 language and the default assumption

00:53:45.460 under language that if someone says

00:53:47.860 something it's true it's like it's a

00:53:50.100 pattern in the world you're going to

00:53:51.340 accept it if I touch a coffee cup I

00:53:53.240 accept that that's what it feels right

00:53:54.980 and if I look at something I accept

00:53:57.520 that's what it looks like well someone

00:53:58.960 says something my initial acceptance is

00:54:01.020 okay that's what it is so you know and

00:54:03.860 then instead of fact well someone says

00:54:05.240 something that's false of course well

00:54:07.460 that's a problem because just just by

00:54:10.040 the fact that I've experienced it it's

00:54:11.880 now part of my world model and I and

00:54:13.680 if that's what you're referring to I can

00:54:15.200 see this is really a problem of

00:54:16.580 language we face and this is this is the

00:54:19.200 root cause of almost all of our false

00:54:20.760 beliefs is that someone just says

00:54:22.740 something enough times and that's good

00:54:25.560 enough and you have to seek out

00:54:28.600 contrary evidence evidence for it yeah

00:54:30.960 sometimes it's good enough even when

00:54:32.200 you're the one saying it you just

00:54:33.440 overhear you're you're the voice of

00:54:35.840 your own mind saying it and no I know

00:54:37.800 that's that's that's it's been proven

00:54:41.060 that everyone is susceptible to that

00:54:43.080 kind of distortion of our beliefs or

00:54:45.340 especially our memories just remembering

00:54:47.420 something over and over again changes it

00:54:48.840 you know yeah okay so let's get back to

00:54:51.240 AI risk here because here's where I think

00:54:54.600 you and I have very different intuitions I

00:54:57.120 mean the intuition that many of us have

00:54:59.720 the people who have informed my views

00:55:02.620 here people like Stuart Russell who you

00:55:05.400 probably know at Berkeley yeah and Nick

00:55:07.800 Bostrom and Eliezer Yudkowsky and just

00:55:10.760 lots of people in this spot worrying about

00:55:14.460 the same thing to one another degree the

00:55:17.540 intuition is that you don't get a second

00:55:20.660 chance to create a truly autonomous

00:55:24.580 super intelligence right like it seems that

00:55:27.720 in principle this is the kind of thing you

00:55:30.040 have to get right on the first try right

00:55:32.840 and having to get anything right on the

00:55:34.140 first try just seems extraordinarily

00:55:36.540 dangerous because we we rarely if ever do

00:55:40.080 that when doing something complicated

00:55:42.040 and another way of putting this is that

00:55:44.460 it seems like in the space of all possible

00:55:47.620 super intelligent minds there are more ways

00:55:51.580 to build one that isn't perfectly aligned

00:55:54.780 with our long-term well-being then there are

00:55:59.460 ways to build one that is perfectly aligned

00:56:02.020 with our long-term well-being and you know

00:56:05.860 from my point of view what what you know

00:56:08.240 your optimism and the optimism of many other

00:56:11.020 people you know who take your side of this

00:56:12.920 debate is based on is a it's not really

00:56:16.300 taking the prospect of intelligence seriously

00:56:20.940 enough and the autonomy that is intrinsic to

00:56:24.180 it I mean if we actually built a true

00:56:27.600 general intelligence what that means is that

00:56:31.740 we would suddenly find ourselves in

00:56:33.620 relationship to something that we actually

00:56:37.860 can't perfectly understand it's like it will

00:56:41.740 be analogous to a strange person walking into

00:56:45.180 the room you know you're in relationship and if

00:56:48.400 this person can think a thousand times or a

00:56:52.120 million times faster than you can and has

00:56:56.160 goals that are less than perfectly aligned with

00:56:59.520 your own that's going to be a problem

00:57:02.160 eventually we can't find ourselves in a state

00:57:05.780 of perpetual negotiation with systems that

00:57:09.560 are more competent and powerful and intelligent

00:57:12.220 I think I think there's I think there's two

00:57:14.220 mistakes in your argument the first one is

00:57:16.760 you say my intuition and your intuition I think

00:57:19.520 most of the people who have this fear have an

00:57:22.080 intuition about what

00:57:23.520 if you'd like to continue listening to this

00:57:32.320 conversation you'll need to subscribe at

00:57:34.520 samharris.org once you do you'll get access to

00:57:37.460 all full-length episodes of the making sense

00:57:39.200 podcast along with other subscriber only content

00:57:41.980 including bonus episodes and amas and the

00:57:45.460 conversations I've been having on the waking up app

00:57:47.400 the making sense podcast is ad-free and relies

00:57:50.760 entirely on listener support and you can

00:57:53.120 subscribe now at samharris.org

00:57:55.980 AWF

00:58:02.140 for example

00:58:03.160 in a

00:58:03.620 SMITH

00:58:04.740 from the

00:58:05.020 moment

00:58:05.280 honest

00:58:06.340 in time

00:58:08.380 go

00:58:09.300 to the

00:58:10.000 報告

00:58:10.200 another

00:58:12.040 balloon

Making Sense - Sam Harris - July 09, 2021

#255 — The Future of Intelligence

Episode Stats

Length

Words per Minute

Word Count

Sentence Count

Misogynist Sentences

Hate Speech Sentences

Summary

Transcript