Jeff Hawkins: Thousand Brains Theory of Intelligence #25

Transcript

00:00:00 The following is a conversation with Jeff Hawkins.

00:00:02 He’s the founder of the Redwood Center

00:00:04 for Theoretical Neuroscience in 2002, and NuMenta in 2005.

00:00:08 In his 2004 book, titled On Intelligence,

00:00:11 and in the research before and after,

00:00:13 he and his team have worked to reverse engineer

00:00:16 the neural cortex, and propose artificial intelligence

00:00:19 architectures, approaches, and ideas

00:00:21 that are inspired by the human brain.

00:00:23 These ideas include Hierarchical Tupperware Memory,

00:00:25 HTM, from 2004, and new work,

00:00:28 the Thousand Brains Theory of Intelligence

00:00:30 from 2017, 18, and 19.

00:00:33 Jeff’s ideas have been an inspiration

00:00:36 to many who have looked for progress

00:00:38 beyond the current machine learning approaches,

00:00:40 but they have also received criticism

00:00:42 for lacking a body of empirical evidence

00:00:44 supporting the models.

00:00:46 This is always a challenge when seeking more

00:00:48 than small incremental steps forward in AI.

00:00:51 Jeff is a brilliant mind, and many of the ideas

00:00:54 he has developed and aggregated from neuroscience

00:00:56 are worth understanding and thinking about.

00:00:59 There are limits to deep learning,

00:01:00 as it is currently defined.

00:01:02 Forward progress in AI is shrouded in mystery.

00:01:05 My hope is that conversations like this

00:01:07 can help provide an inspiring spark for new ideas.

00:01:11 This is the Artificial Intelligence Podcast.

00:01:14 If you enjoy it, subscribe on YouTube, iTunes,

00:01:16 or simply connect with me on Twitter

00:01:18 at Lex Friedman, spelled F R I D.

00:01:21 And now, here’s my conversation with Jeff Hawkins.

00:01:26 Are you more interested in understanding the human brain

00:01:29 or in creating artificial systems

00:01:32 that have many of the same qualities

00:01:34 but don’t necessarily require that you actually understand

00:01:38 the underpinning workings of our mind?

00:01:41 So there’s a clear answer to that question.

00:01:44 My primary interest is understanding the human brain.

00:01:46 No question about it.

00:01:47 But I also firmly believe that we will not be able

00:01:53 to create fully intelligent machines

00:01:55 until we understand how the human brain works.

00:01:57 So I don’t see those as separate problems.

00:02:00 I think there’s limits to what can be done

00:02:01 with machine intelligence if you don’t understand

00:02:03 the principles by which the brain works.

00:02:05 And so I actually believe that studying the brain

00:02:07 is actually the fastest way to get to machine intelligence.

00:02:11 And within that, let me ask the impossible question,

00:02:14 how do you, not define, but at least think

00:02:17 about what it means to be intelligent?

00:02:19 So I didn’t try to answer that question first.

00:02:22 We said, let’s just talk about how the brain works

00:02:24 and let’s figure out how certain parts of the brain,

00:02:26 mostly the neocortex, but some other parts too.

00:02:29 The parts of the brain most associated with intelligence.

00:02:32 And let’s discover the principles by how they work.

00:02:35 Because intelligence isn’t just like some mechanism

00:02:39 and it’s not just some capabilities.

00:02:40 It’s like, okay, we don’t even know

00:02:42 where to begin on this stuff.

00:02:44 And so now that we’ve made a lot of progress on this,

00:02:49 after we’ve made a lot of progress

00:02:50 on how the neocortex works, and we can talk about that,

00:02:53 I now have a very good idea what’s gonna be required

00:02:55 to make intelligent machines.

00:02:57 I can tell you today, some of the things

00:02:59 are gonna be necessary, I believe,

00:03:02 to create intelligent machines.

00:03:03 Well, so we’ll get there.

00:03:04 We’ll get to the neocortex and some of the theories

00:03:07 of how the whole thing works.

00:03:09 And you’re saying, as we understand more and more

00:03:11 about the neocortex, about our own human mind,

00:03:14 we’ll be able to start to more specifically define

00:03:17 what it means to be intelligent.

00:03:18 It’s not useful to really talk about that until.

00:03:21 I don’t know if it’s not useful.

00:03:23 Look, there’s a long history of AI, as you know.

00:03:26 And there’s been different approaches taken to it.

00:03:28 And who knows, maybe they’re all useful.

00:03:32 So the good old fashioned AI, the expert systems,

00:03:37 the current convolutional neural networks,

00:03:38 they all have their utility.

00:03:40 They all have a value in the world.

00:03:43 But I would think almost everyone agree

00:03:45 that none of them are really intelligent

00:03:46 in a sort of a deep way that humans are.

00:03:49 And so it’s just the question of how do you get

00:03:53 from where those systems were or are today

00:03:56 to where a lot of people think we’re gonna go.

00:03:59 And there’s a big, big gap there, a huge gap.

00:04:02 And I think the quickest way of bridging that gap

00:04:06 is to figure out how the brain does that.

00:04:08 And then we can sit back and look and say,

00:04:10 oh, which of these principles that the brain works on

00:04:12 are necessary and which ones are not?

00:04:15 Clearly, we don’t have to build this in,

00:04:16 and intelligent machines aren’t gonna be built

00:04:18 out of organic living cells.

00:04:22 But there’s a lot of stuff that goes on the brain

00:04:24 that’s gonna be necessary.

00:04:25 So let me ask maybe, before we get into the fun details,

00:04:30 let me ask maybe a depressing or a difficult question.

00:04:33 Do you think it’s possible that we will never

00:04:36 be able to understand how our brain works,

00:04:38 that maybe there’s aspects to the human mind,

00:04:41 like we ourselves cannot introspectively get to the core,

00:04:46 that there’s a wall you eventually hit?

00:04:48 Yeah, I don’t believe that’s the case.

00:04:52 I have never believed that’s the case.

00:04:53 There’s not been a single thing humans have ever put

00:04:56 their minds to that we’ve said, oh, we reached the wall,

00:04:58 we can’t go any further.

00:04:59 It’s just, people keep saying that.

00:05:01 People used to believe that about life.

00:05:03 Alain Vital, right, there’s like,

00:05:05 what’s the difference between living matter

00:05:06 and nonliving matter, something special

00:05:07 that we never understand.

00:05:09 We no longer think that.

00:05:10 So there’s no historical evidence that suggests this

00:05:14 is the case, and I just never even consider

00:05:16 that’s a possibility.

00:05:17 I would also say, today, we understand so much

00:05:21 about the neocortex.

00:05:22 We’ve made tremendous progress in the last few years

00:05:25 that I no longer think of it as an open question.

00:05:30 The answers are very clear to me.

00:05:32 The pieces we don’t know are clear to me,

00:05:34 but the framework is all there, and it’s like,

00:05:36 oh, okay, we’re gonna be able to do this.

00:05:38 This is not a problem anymore, just takes time and effort,

00:05:41 but there’s no mystery, a big mystery anymore.

00:05:44 So then let’s get into it for people like myself

00:05:47 who are not very well versed in the human brain,

00:05:52 except my own.

00:05:54 Can you describe to me, at the highest level,

00:05:57 what are the different parts of the human brain,

00:05:59 and then zooming in on the neocortex,

00:06:02 the parts of the neocortex, and so on,

00:06:04 a quick overview.

00:06:05 Yeah, sure.

00:06:06 The human brain, we can divide it roughly into two parts.

00:06:10 There’s the old parts, lots of pieces,

00:06:14 and then there’s the new part.

00:06:15 The new part is the neocortex.

00:06:18 It’s new because it didn’t exist before mammals.

00:06:20 The only mammals have a neocortex,

00:06:22 and in humans, in primates, it’s very large.

00:06:24 In the human brain, the neocortex occupies

00:06:26 about 70 to 75% of the volume of the brain.

00:06:30 It’s huge.

00:06:32 And the old parts of the brain are,

00:06:34 there’s lots of pieces there.

00:06:36 There’s the spinal cord, and there’s the brain stem,

00:06:38 and the cerebellum, and the different parts

00:06:40 of the basal ganglia, and so on.

00:06:42 In the old parts of the brain,

00:06:42 you have the autonomic regulation,

00:06:44 like breathing and heart rate.

00:06:46 You have basic behaviors, so like walking and running

00:06:49 are controlled by the old parts of the brain.

00:06:51 All the emotional centers of the brain

00:06:53 are in the old part of the brain,

00:06:53 so when you feel anger or hungry, lust,

00:06:55 or things like that, those are all

00:06:56 in the old parts of the brain.

00:06:57 And we associate with the neocortex

00:07:02 all the things we think about as sort of

00:07:04 high level perception and cognitive functions,

00:07:08 anything from seeing and hearing and touching things

00:07:12 to language to mathematics and engineering

00:07:15 and science and so on.

00:07:16 Those are all associated with the neocortex,

00:07:19 and they’re certainly correlated.

00:07:21 Our abilities in those regards are correlated

00:07:23 with the relative size of our neocortex

00:07:25 compared to other mammals.

00:07:27 So that’s like the rough division,

00:07:30 and you obviously can’t understand

00:07:32 the neocortex completely isolated,

00:07:35 but you can understand a lot of it

00:07:37 with just a few interfaces to the old parts of the brain,

00:07:40 and so it gives you a system to study.

00:07:44 The other remarkable thing about the neocortex,

00:07:48 compared to the old parts of the brain,

00:07:49 is the neocortex is extremely uniform.

00:07:52 It’s not visibly or anatomically,

00:07:57 it’s very, I always like to say

00:07:59 it’s like the size of a dinner napkin,

00:08:01 about two and a half millimeters thick,

00:08:03 and it looks remarkably the same everywhere.

00:08:05 Everywhere you look in that two and a half millimeters

00:08:07 is this detailed architecture,

00:08:10 and it looks remarkably the same everywhere,

00:08:11 and that’s across species.

00:08:12 A mouse versus a cat and a dog and a human.

00:08:15 Where if you look at the old parts of the brain,

00:08:17 there’s lots of little pieces do specific things.

00:08:19 So it’s like the old parts of our brain evolved,

00:08:22 like this is the part that controls heart rate,

00:08:23 and this is the part that controls this,

00:08:24 and this is this kind of thing,

00:08:25 and that’s this kind of thing,

00:08:27 and these evolved for eons a long, long time,

00:08:30 and they have their specific functions,

00:08:31 and all of a sudden mammals come along,

00:08:33 and they got this thing called the neocortex,

00:08:35 and it got large by just replicating the same thing

00:08:38 over and over and over again.

00:08:39 This is like, wow, this is incredible.

00:08:42 So all the evidence we have,

00:08:46 and this is an idea that was first articulated

00:08:50 in a very cogent and beautiful argument

00:08:52 by a guy named Vernon Malcastle in 1978, I think it was,

00:08:56 that the neocortex all works on the same principle.

00:09:01 So language, hearing, touch, vision, engineering,

00:09:05 all these things are basically underlying,

00:09:06 are all built on the same computational substrate.

00:09:10 They’re really all the same problem.

00:09:11 So the low level of the building blocks all look similar.

00:09:14 Yeah, and they’re not even that low level.

00:09:16 We’re not talking about like neurons.

00:09:17 We’re talking about this very complex circuit

00:09:19 that exists throughout the neocortex.

00:09:21 It’s remarkably similar.

00:09:23 It’s like, yes, you see variations of it here and there,

00:09:26 more of the cell, less and less, and so on.

00:09:29 But what Malcastle argued was, he says,

00:09:32 you know, if you take a section of neocortex,

00:09:35 why is one a visual area and one is a auditory area?

00:09:38 Or why is, and his answer was,

00:09:41 it’s because one is connected to eyes

00:09:43 and one is connected to ears.

00:09:45 Literally, you mean just it’s most closest

00:09:47 in terms of number of connections

00:09:49 to the sensor. Literally, literally,

00:09:50 if you took the optic nerve and attached it

00:09:53 to a different part of the neocortex,

00:09:55 that part would become a visual region.

00:09:57 This actually, this experiment was actually done

00:10:00 by Merkankasur in developing, I think it was lemurs,

00:10:04 I can’t remember what it was, some animal.

00:10:06 And there’s a lot of evidence to this.

00:10:08 You know, if you take a blind person,

00:10:09 a person who’s born blind at birth,

00:10:12 they’re born with a visual neocortex.

00:10:15 It doesn’t, may not get any input from the eyes

00:10:18 because of some congenital defect or something.

00:10:21 And that region becomes, does something else.

00:10:24 It picks up another task.

00:10:27 So, and it’s, so it’s this very complex thing.

00:10:32 It’s not like, oh, they’re all built on neurons.

00:10:33 No, they’re all built in this very complex circuit

00:10:36 and somehow that circuit underlies everything.

00:10:40 And so this is the, it’s called

00:10:43 the common cortical algorithm, if you will.

00:10:45 Some scientists just find it hard to believe

00:10:47 and they just, I can’t believe that’s true,

00:10:50 but the evidence is overwhelming in this case.

00:10:52 And so a large part of what it means

00:10:54 to figure out how the brain creates intelligence

00:10:56 and what is intelligence in the brain

00:10:59 is to understand what that circuit does.

00:11:02 If you can figure out what that circuit does,

00:11:05 as amazing as it is, then you can,

00:11:06 then you understand what all these

00:11:08 other cognitive functions are.

00:11:10 So if you were to sort of put neocortex

00:11:13 outside of your book on intelligence,

00:11:15 you look, if you wrote a giant tome, a textbook

00:11:18 on the neocortex, and you look maybe

00:11:21 a couple of centuries from now,

00:11:23 how much of what we know now would still be accurate

00:11:26 two centuries from now?

00:11:27 So how close are we in terms of understanding?

00:11:30 I have to speak from my own particular experience here.

00:11:32 So I run a small research lab here.

00:11:35 It’s like any other research lab.

00:11:38 I’m sort of the principal investigator.

00:11:39 There’s actually two of us

00:11:40 and there’s a bunch of other people.

00:11:42 And this is what we do.

00:11:43 We study the neocortex and we publish our results

00:11:46 and so on.

00:11:46 So about three years ago,

00:11:49 we had a real breakthrough in this field.

00:11:52 Just tremendous breakthrough.

00:11:53 We’ve now published, I think, three papers on it.

00:11:56 And so I have a pretty good understanding

00:12:00 of all the pieces and what we’re missing.

00:12:02 I would say that almost all the empirical data

00:12:06 we’ve collected about the brain, which is enormous.

00:12:08 If you don’t know the neuroscience literature,

00:12:10 it’s just incredibly big.

00:12:13 And it’s, for the most part, all correct.

00:12:16 It’s facts and experimental results and measurements

00:12:21 and all kinds of stuff.

00:12:22 But none of that has been really assimilated

00:12:25 into a theoretical framework.

00:12:27 It’s data without, in the language of Thomas Kuhn,

00:12:32 the historian, would be a sort of a pre paradigm science.

00:12:35 Lots of data, but no way to fit it together.

00:12:38 I think almost all of that’s correct.

00:12:39 There’s just gonna be some mistakes in there.

00:12:42 And for the most part,

00:12:43 there aren’t really good cogent theories about it,

00:12:45 how to put it together.

00:12:47 It’s not like we have two or three competing good theories,

00:12:50 which ones are right and which ones are wrong.

00:12:51 It’s like, nah, people are just scratching their heads.

00:12:54 Some people have given up

00:12:55 on trying to figure out what the whole thing does.

00:12:57 In fact, there’s very, very few labs that we do

00:13:01 that focus really on theory

00:13:03 and all this unassimilated data and trying to explain it.

00:13:06 So it’s not like we’ve got it wrong.

00:13:08 It’s just that we haven’t got it at all.

00:13:11 So it’s really, I would say, pretty early days

00:13:15 in terms of understanding the fundamental theory’s forces

00:13:19 of the way our mind works.

00:13:20 I don’t think so.

00:13:21 I would have said that’s true five years ago.

00:13:25 So as I said,

00:13:26 we had some really big breakthroughs on this recently

00:13:29 and we started publishing papers on this.

00:13:30 So we’ll get to that.

00:13:34 But so I don’t think it’s,

00:13:35 I’m an optimist and from where I sit today,

00:13:38 most people would disagree with this,

00:13:39 but from where I sit today, from what I know,

00:13:43 it’s not super early days anymore.

00:13:44 We are, the way these things go

00:13:46 is it’s not a linear path, right?

00:13:48 You don’t just start accumulating

00:13:49 and get better and better and better.

00:13:50 No, all this stuff you’ve collected,

00:13:52 none of it makes sense.

00:13:53 All these different things are just sort of around.

00:13:55 And then you’re gonna have some breaking points

00:13:57 where all of a sudden, oh my God, now we got it right.

00:13:59 That’s how it goes in science.

00:14:01 And I personally feel like we passed that little thing

00:14:04 about a couple of years ago,

00:14:06 all that big thing a couple of years ago.

00:14:07 So we can talk about that.

00:14:09 Time will tell if I’m right,

00:14:11 but I feel very confident about it.

00:14:12 That’s why I’m willing to say it on tape like this.

00:14:15 At least very optimistic.

00:14:18 So let’s, before those few years ago,

00:14:20 let’s take a step back to HTM,

00:14:23 the hierarchical temporal memory theory,

00:14:26 which you first proposed on intelligence

00:14:27 and went through a few different generations.

00:14:29 Can you describe what it is,

00:14:31 how it evolved through the three generations

00:14:33 since you first put it on paper?

00:14:35 Yeah, so one of the things that neuroscientists

00:14:39 just sort of missed for many, many years,

00:14:42 and especially people who were thinking about theory,

00:14:45 was the nature of time in the brain.

00:14:49 Brains process information through time.

00:14:51 The information coming into the brain

00:14:52 is constantly changing.

00:14:55 The patterns from my speech right now,

00:14:57 if you were listening to it at normal speed,

00:15:00 would be changing on your ears

00:15:01 about every 10 milliseconds or so, you’d have a change.

00:15:04 This constant flow, when you look at the world,

00:15:06 your eyes are moving constantly,

00:15:08 three to five times a second,

00:15:09 and the input’s completely changing.

00:15:11 If I were to touch something like a coffee cup,

00:15:13 as I move my fingers, the input changes.

00:15:15 So this idea that the brain works on time changing patterns

00:15:19 is almost completely, or was almost completely missing

00:15:22 from a lot of the basic theories,

00:15:23 like fears of vision and so on.

00:15:25 It’s like, oh no, we’re gonna put this image

00:15:26 in front of you and flash it and say, what is it?

00:15:29 Convolutional neural networks work that way today, right?

00:15:32 Classify this picture.

00:15:34 But that’s not what vision is like.

00:15:35 Vision is this sort of crazy time based pattern

00:15:38 that’s going all over the place,

00:15:40 and so is touch and so is hearing.

00:15:41 So the first part of hierarchical temporal memory

00:15:43 was the temporal part.

00:15:45 It’s to say, you won’t understand the brain,

00:15:48 nor will you understand intelligent machines

00:15:50 unless you’re dealing with time based patterns.

00:15:52 The second thing was, the memory component of it was,

00:15:55 is to say that we aren’t just processing input,

00:16:00 we learn a model of the world.

00:16:02 And the memory stands for that model.

00:16:05 The point of the brain, the part of the neocortex,

00:16:07 it learns a model of the world.

00:16:08 We have to store things, our experiences,

00:16:11 in a form that leads to a model of the world.

00:16:14 So we can move around the world,

00:16:15 we can pick things up and do things and navigate

00:16:17 and know how it’s going on.

00:16:18 So that’s what the memory referred to.

00:16:19 And many people just, they were thinking about

00:16:22 like certain processes without memory at all.

00:16:25 They’re just like processing things.

00:16:26 And then finally, the hierarchical component

00:16:29 was a reflection to that the neocortex,

00:16:32 although it’s this uniform sheet of cells,

00:16:35 different parts of it project to other parts,

00:16:37 which project to other parts.

00:16:39 And there is a sort of rough hierarchy in terms of that.

00:16:43 So the hierarchical temporal memory is just saying,

00:16:45 look, we should be thinking about the brain

00:16:47 as time based, model memory based,

00:16:52 and hierarchical processing.

00:16:54 And that was a placeholder for a bunch of components

00:16:58 that we would then plug into that.

00:17:00 We still believe all those things I just said,

00:17:02 but we now know so much more that I’m stopping to use

00:17:06 the word hierarchical temporal memory yet

00:17:08 because it’s insufficient to capture the stuff we know.

00:17:11 So again, it’s not incorrect, but it’s,

00:17:13 I now know more and I would rather describe it

00:17:15 more accurately.

00:17:16 Yeah, so you’re basically, we could think of HTM

00:17:20 as emphasizing that there’s three aspects of intelligence

00:17:24 that are important to think about

00:17:25 whatever the eventual theory it converges to.

00:17:28 So in terms of time, how do you think of nature of time

00:17:32 across different time scales?

00:17:33 So you mentioned things changing,

00:17:36 sensory inputs changing every 10, 20 minutes.

00:17:39 What about every few minutes, every few months and years?

00:17:42 Well, if you think about a neuroscience problem,

00:17:44 the brain problem, neurons themselves can stay active

00:17:49 for certain periods of time, parts of the brain

00:17:52 where they stay active for minutes.

00:17:54 You could hold a certain perception or an activity

00:17:59 for a certain period of time,

00:18:01 but most of them don’t last that long.

00:18:04 And so if you think about your thoughts

00:18:07 are the activity of neurons,

00:18:09 if you’re gonna wanna involve something

00:18:10 that happened a long time ago,

00:18:11 even just this morning, for example,

00:18:14 the neurons haven’t been active throughout that time.

00:18:16 So you have to store that.

00:18:17 So if I ask you, what did you have for breakfast today?

00:18:20 That is memory, that is you’ve built into your model

00:18:23 the world now, you remember that.

00:18:24 And that memory is in the synapses,

00:18:27 is basically in the formation of synapses.

00:18:29 And so you’re sliding into what,

00:18:34 you know, it’s the different timescales.

00:18:36 There’s timescales of which we are like understanding

00:18:39 my language and moving about and seeing things rapidly

00:18:41 and over time, that’s the timescales

00:18:42 of activities of neurons.

00:18:44 But if you wanna get in longer timescales,

00:18:46 then it’s more memory.

00:18:47 And we have to invoke those memories to say,

00:18:49 oh yes, well now I can remember what I had for breakfast

00:18:51 because I stored that someplace.

00:18:54 I may forget it tomorrow, but I’d store it for now.

00:18:58 So does memory also need to have,

00:19:02 so the hierarchical aspect of reality

00:19:06 is not just about concepts, it’s also about time?

00:19:08 Do you think of it that way?

00:19:10 Yeah, time is infused in everything.

00:19:12 It’s like you really can’t separate it out.

00:19:15 If I ask you, what is your, you know,

00:19:18 how’s the brain learn a model of this coffee cup here?

00:19:21 I have a coffee cup and I’m at the coffee cup.

00:19:23 I say, well, time is not an inherent property

00:19:25 of the model I have of this cup,

00:19:28 whether it’s a visual model or a tactile model.

00:19:31 I can sense it through time,

00:19:32 but the model itself doesn’t really have much time.

00:19:34 If I asked you, if I said,

00:19:36 well, what is the model of my cell phone?

00:19:38 My brain has learned a model of the cell phone.

00:19:40 So if you have a smartphone like this,

00:19:43 and I said, well, this has time aspects to it.

00:19:45 I have expectations when I turn it on,

00:19:48 what’s gonna happen, what or how long it’s gonna take

00:19:50 to do certain things, if I bring up an app,

00:19:52 what sequences, and so I have,

00:19:54 and it’s like melodies in the world, you know?

00:19:57 Melody has a sense of time.

00:19:58 So many things in the world move and act,

00:20:01 and there’s a sense of time related to them.

00:20:03 Some don’t, but most things do actually.

00:20:08 So it’s sort of infused throughout the models of the world.

00:20:12 You build a model of the world,

00:20:13 you’re learning the structure of the objects in the world,

00:20:16 and you’re also learning how those things change

00:20:18 through time.

00:20:20 Okay, so it really is just a fourth dimension

00:20:23 that’s infused deeply, and you have to make sure

00:20:26 that your models of intelligence incorporate it.

00:20:30 So, like you mentioned, the state of neuroscience

00:20:34 is deeply empirical, a lot of data collection.

00:20:37 It’s, you know, that’s where it is.

00:20:41 You mentioned Thomas Kuhn, right?

00:20:43 Yeah.

00:20:44 And then you’re proposing a theory of intelligence,

00:20:48 and which is really the next step,

00:20:50 the really important step to take,

00:20:52 but why is HTM, or what we’ll talk about soon,

00:21:00 the right theory?

00:21:03 So is it more in the, is it backed by intuition?

00:21:07 Is it backed by evidence?

00:21:09 Is it backed by a mixture of both?

00:21:11 Is it kind of closer to where string theory is in physics,

00:21:15 where there’s mathematical components

00:21:18 which show that, you know what,

00:21:21 it seems that this, it fits together too well

00:21:24 for it not to be true, which is where string theory is.

00:21:28 Is that where you’re kind of seeing?

00:21:29 It’s a mixture of all those things,

00:21:30 although definitely where we are right now

00:21:32 is definitely much more on the empirical side

00:21:34 than, let’s say, string theory.

00:21:37 The way this goes about, we’re theorists, right?

00:21:39 So we look at all this data, and we’re trying to come up

00:21:41 with some sort of model that explains it, basically,

00:21:44 and there’s, unlike string theory,

00:21:46 there’s vast more amounts of empirical data here

00:21:50 that I think than most physicists deal with.

00:21:54 And so our challenge is to sort through that

00:21:57 and figure out what kind of constructs would explain this.

00:22:02 And when we have an idea,

00:22:04 you come up with a theory of some sort,

00:22:06 you have lots of ways of testing it.

00:22:08 First of all, there are 100 years of assimilated,

00:22:13 assimilated, unassimilated empirical data from neuroscience.

00:22:16 So we go back and read papers,

00:22:18 and we say, oh, did someone find this already?

00:22:20 We can predict X, Y, and Z,

00:22:23 and maybe no one’s even talked about it

00:22:25 since 1972 or something, but we go back and find that,

00:22:28 and we say, oh, either it can support the theory

00:22:31 or it can invalidate the theory.

00:22:33 And we say, okay, we have to start over again.

00:22:34 Oh, no, it’s supportive, let’s keep going with that one.

00:22:38 So the way I kind of view it, when we do our work,

00:22:42 we look at all this empirical data,

00:22:45 and what I call it is a set of constraints.

00:22:47 We’re not interested in something

00:22:48 that’s biologically inspired.

00:22:49 We’re trying to figure out how the actual brain works.

00:22:52 So every piece of empirical data

00:22:53 is a constraint on a theory.

00:22:55 In theory, if you have the correct theory,

00:22:57 it needs to explain every pin, right?

00:22:59 So we have this huge number of constraints on the problem,

00:23:03 which initially makes it very, very difficult.

00:23:05 If you don’t have many constraints,

00:23:07 you can make up stuff all the day.

00:23:08 You can say, oh, here’s an answer on how you can do this,

00:23:10 you can do that, you can do this.

00:23:11 But if you consider all biology as a set of constraints,

00:23:13 all neuroscience as a set of constraints,

00:23:15 and even if you’re working in one little part

00:23:17 of the neocortex, for example,

00:23:18 there are hundreds and hundreds of constraints.

00:23:20 These are empirical constraints

00:23:22 that it’s very, very difficult initially

00:23:24 to come up with a theoretical framework for that.

00:23:27 But when you do, and it solves all those constraints

00:23:30 at once, you have a high confidence

00:23:32 that you got something close to correct.

00:23:35 It’s just mathematically almost impossible not to be.

00:23:39 So that’s the curse and the advantage of what we have.

00:23:43 The curse is we have to solve,

00:23:45 we have to meet all these constraints, which is really hard.

00:23:48 But when you do meet them,

00:23:50 then you have a great confidence

00:23:53 that you’ve discovered something.

00:23:54 In addition, then we work with scientific labs.

00:23:58 So we’ll say, oh, there’s something we can’t find,

00:24:00 we can predict something,

00:24:01 but we can’t find it anywhere in the literature.

00:24:04 So we will then, we have people we’ve collaborated with,

00:24:06 we’ll say, sometimes they’ll say, you know what?

00:24:09 I have some collected data, which I didn’t publish,

00:24:11 but we can go back and look at it

00:24:13 and see if we can find that,

00:24:14 which is much easier than designing a new experiment.

00:24:17 You know, neuroscience experiments take a long time, years.

00:24:20 So, although some people are doing that now too.

00:24:23 So, but between all of these things,

00:24:27 I think it’s a reasonable,

00:24:30 actually a very, very good approach.

00:24:31 We are blessed with the fact that we can test our theories

00:24:35 out the yin yang here because there’s so much

00:24:37 unassimilar data and we can also falsify our theories

00:24:39 very easily, which we do often.

00:24:41 So it’s kind of reminiscent to whenever that was

00:24:44 with Copernicus, you know, when you figure out

00:24:47 that the sun’s at the center of the solar system

00:24:51 as opposed to earth, the pieces just fall into place.

00:24:54 Yeah, I think that’s the general nature of aha moments

00:24:59 is, and it’s Copernicus, it could be,

00:25:02 you could say the same thing about Darwin,

00:25:05 you could say the same thing about, you know,

00:25:07 about the double helix,

00:25:09 that people have been working on a problem for so long

00:25:12 and have all this data and they can’t make sense of it,

00:25:14 they can’t make sense of it.

00:25:15 But when the answer comes to you

00:25:17 and everything falls into place,

00:25:19 it’s like, oh my gosh, that’s it.

00:25:21 That’s got to be right.

00:25:23 I asked both Jim Watson and Francis Crick about this.

00:25:29 I asked them, you know, when you were working on

00:25:31 trying to discover the structure of the double helix,

00:25:35 and when you came up with the sort of the structure

00:25:39 that ended up being correct, but it was sort of a guess,

00:25:44 you know, it wasn’t really verified yet.

00:25:45 I said, did you know that it was right?

00:25:48 And they both said, absolutely.

00:25:50 So we absolutely knew it was right.

00:25:51 And it doesn’t matter if other people didn’t believe it

00:25:54 or not, we knew it was right.

00:25:55 They’d get around to thinking it

00:25:56 and agree with it eventually anyway.

00:25:59 And that’s the kind of thing you hear a lot with scientists

00:26:01 who really are studying a difficult problem.

00:26:04 And I feel that way too about our work.

00:26:07 Have you talked to Crick or Watson about the problem

00:26:10 you’re trying to solve, the, of finding the DNA of the brain?

00:26:15 Yeah, in fact, Francis Crick was very interested in this

00:26:19 in the latter part of his life.

00:26:21 And in fact, I got interested in brains

00:26:23 by reading an essay he wrote in 1979

00:26:26 called Thinking About the Brain.

00:26:28 And that was when I decided I’m gonna leave my profession

00:26:32 of computers and engineering and become a neuroscientist.

00:26:35 Just reading that one essay from Francis Crick.

00:26:37 I got to meet him later in life.

00:26:41 I spoke at the Salk Institute and he was in the audience.

00:26:44 And then I had a tea with him afterwards.

00:26:48 He was interested in a different problem.

00:26:50 He was focused on consciousness.

00:26:53 The easy problem, right?

00:26:54 Well, I think it’s the red herring.

00:26:58 And so we weren’t really overlapping a lot there.

00:27:02 Jim Watson, who’s still alive,

00:27:05 is also interested in this problem.

00:27:07 And he was, when he was director

00:27:09 of the Cold Spring Harbor Laboratories,

00:27:12 he was really sort of behind moving in the direction

00:27:15 of neuroscience there.

00:27:16 And so he had a personal interest in this field.

00:27:20 And I have met with him numerous times.

00:27:23 And in fact, the last time was a little bit over a year ago,

00:27:27 I gave a talk at Cold Spring Harbor Labs

00:27:30 about the progress we were making in our work.

00:27:34 And it was a lot of fun because he said,

00:27:39 well, you wouldn’t be coming here

00:27:41 unless you had something important to say.

00:27:42 So I’m gonna go attend your talk.

00:27:44 So he sat in the very front row.

00:27:46 Next to him was the director of the lab, Bruce Stillman.

00:27:50 So these guys are in the front row of this auditorium.

00:27:52 Nobody else in the auditorium wants to sit in the front row

00:27:54 because there’s Jim Watson and there’s the director.

00:27:56 And I gave a talk and then I had dinner with him afterwards.

00:28:03 But there’s a great picture of my colleague Subitai Amantak

00:28:07 where I’m up there sort of like screaming the basics

00:28:09 of this new framework we have.

00:28:11 And Jim Watson’s on the edge of his chair.

00:28:13 He’s literally on the edge of his chair,

00:28:15 like intently staring up at the screen.

00:28:17 And when he discovered the structure of DNA,

00:28:21 the first public talk he gave

00:28:23 was at Cold Spring Harbor Labs.

00:28:25 And there’s a picture, there’s a famous picture

00:28:27 of Jim Watson standing at the whiteboard

00:28:29 with an overrated thing pointing at something,

00:28:31 pointing at the double helix with his pointer.

00:28:33 And it actually looks a lot like the picture of me.

00:28:34 So there was a sort of funny,

00:28:36 there’s Arian talking about the brain

00:28:37 and there’s Jim Watson staring intently at it.

00:28:39 And of course there with, whatever, 60 years earlier,

00:28:41 he was standing pointing at the double helix.

00:28:44 That’s one of the great discoveries in all of,

00:28:47 whatever, biology, science, all science and DNA.

00:28:49 So it’s funny that there’s echoes of that in your presentation.

00:28:54 Do you think, in terms of evolutionary timeline and history,

00:28:58 the development of the neocortex was a big leap?

00:29:01 Or is it just a small step?

00:29:07 So like, if we ran the whole thing over again,

00:29:09 from the birth of life on Earth,

00:29:12 how likely would we develop the mechanism of the neocortex?

00:29:15 Okay, well those are two separate questions.

00:29:17 One is, was it a big leap?

00:29:18 And one was how likely it is, okay?

00:29:21 They’re not necessarily related.

00:29:22 Maybe correlated.

00:29:23 Maybe correlated, maybe not.

00:29:25 And we don’t really have enough data

00:29:26 to make a judgment about that.

00:29:28 I would say definitely it was a big leap.

00:29:29 And I can tell you why.

00:29:30 I don’t think it was just another incremental step.

00:29:34 I don’t get that at the moment.

00:29:35 I don’t really have any idea how likely it is.

00:29:38 If we look at evolution,

00:29:39 we have one data point, which is Earth, right?

00:29:42 Life formed on Earth billions of years ago,

00:29:45 whether it was introduced here or it created it here,

00:29:48 or someone introduced it, we don’t really know,

00:29:49 but it was here early.

00:29:51 It took a long, long time to get to multicellular life.

00:29:55 And then for multicellular life,

00:29:58 it took a long, long time to get the neocortex.

00:30:02 And we’ve only had the neocortex for a few 100,000 years.

00:30:05 So that’s like nothing, okay?

00:30:08 So is it likely?

00:30:09 Well, it certainly isn’t something

00:30:10 that happened right away on Earth.

00:30:13 And there were multiple steps to get there.

00:30:15 So I would say it’s probably not gonna be something

00:30:17 that would happen instantaneously

00:30:18 on other planets that might have life.

00:30:20 It might take several billion years on average.

00:30:23 Is it likely?

00:30:24 I don’t know, but you’d have to survive

00:30:25 for several billion years to find out.

00:30:27 Probably.

00:30:29 Is it a big leap?

00:30:30 Yeah, I think it is a qualitative difference

00:30:35 in all other evolutionary steps.

00:30:37 I can try to describe that if you’d like.

00:30:39 Sure, in which way?

00:30:41 Yeah, I can tell you how.

00:30:43 Pretty much, let’s start with a little preface.

00:30:47 Many of the things that humans are able to do

00:30:50 do not have obvious survival advantages precedent.

00:30:58 We could create music, is that,

00:31:00 is there a really survival advantage to that?

00:31:02 Maybe, maybe not.

00:31:03 What about mathematics?

00:31:04 Is there a real survival advantage to mathematics?

00:31:07 Well, you could stretch it.

00:31:09 You can try to figure these things out, right?

00:31:13 But most of evolutionary history,

00:31:14 everything had immediate survival advantages to it.

00:31:18 So, I’ll tell you a story, which I like,

00:31:22 may or may not be true, but the story goes as follows.

00:31:29 Organisms have been evolving for,

00:31:30 since the beginning of life here on Earth,

00:31:33 and adding this sort of complexity onto that,

00:31:35 and this sort of complexity onto that,

00:31:36 and the brain itself is evolved this way.

00:31:39 In fact, there’s old parts, and older parts,

00:31:42 and older, older parts of the brain

00:31:43 that kind of just keeps calming on new things,

00:31:45 and we keep adding capabilities.

00:31:47 When we got to the neocortex,

00:31:48 initially it had a very clear survival advantage

00:31:52 in that it produced better vision,

00:31:54 and better hearing, and better touch,

00:31:55 and maybe, and so on.

00:31:57 But what I think happens is that evolution discovered,

00:32:01 it took a mechanism, and this is in our recent theories,

00:32:05 but it took a mechanism evolved a long time ago

00:32:08 for navigating in the world, for knowing where you are.

00:32:10 These are the so called grid cells and place cells

00:32:13 of an old part of the brain.

00:32:15 And it took that mechanism for building maps of the world,

00:32:20 and knowing where you are on those maps,

00:32:22 and how to navigate those maps,

00:32:24 and turns it into a sort of a slimmed down,

00:32:27 idealized version of it.

00:32:29 And that idealized version could now apply

00:32:31 to building maps of other things.

00:32:32 Maps of coffee cups, and maps of phones,

00:32:35 maps of mathematics.

00:32:36 Concepts almost.

00:32:37 Concepts, yes, and not just almost, exactly.

00:32:40 And so, and it just started replicating this stuff, right?

00:32:44 You just think more, and more, and more.

00:32:45 So we went from being sort of dedicated purpose

00:32:48 neural hardware to solve certain problems

00:32:51 that are important to survival,

00:32:53 to a general purpose neural hardware

00:32:55 that could be applied to all problems.

00:32:58 And now it’s escaped the orbit of survival.

00:33:02 We are now able to apply it to things

00:33:04 which we find enjoyment,

00:33:08 but aren’t really clearly survival characteristics.

00:33:13 And that it seems to only have happened in humans,

00:33:16 to the large extent.

00:33:19 And so that’s what’s going on,

00:33:20 where we sort of have,

00:33:22 we’ve sort of escaped the gravity of evolutionary pressure,

00:33:26 in some sense, in the neocortex.

00:33:28 And it now does things which are not,

00:33:31 that are really interesting,

00:33:32 discovering models of the universe,

00:33:34 which may not really help us.

00:33:36 Does it matter?

00:33:37 How does it help us surviving,

00:33:38 knowing that there might be multiverses,

00:33:40 or that there might be the age of the universe,

00:33:42 or how do various stellar things occur?

00:33:46 It doesn’t really help us survive at all.

00:33:47 But we enjoy it, and that’s what happened.

00:33:50 Or at least not in the obvious way, perhaps.

00:33:53 It is required,

00:33:56 if you look at the entire universe in an evolutionary way,

00:33:58 it’s required for us to do interplanetary travel,

00:34:00 and therefore survive past our own sun.

00:34:03 But you know, let’s not get too.

00:34:04 Yeah, but evolution works at one time frame,

00:34:07 it’s survival, if you think of survival of the phenotype,

00:34:11 survival of the individual.

00:34:13 What you’re talking about there is spans well beyond that.

00:34:16 So there’s no genetic,

00:34:18 I’m not transferring any genetic traits to my children

00:34:23 that are gonna help them survive better on Mars.

00:34:26 Totally different mechanism, that’s right.

00:34:28 So let’s get into the new, as you’ve mentioned,

00:34:31 this idea of the, I don’t know if you have a nice name,

00:34:34 thousand.

00:34:35 We call it the thousand brain theory of intelligence.

00:34:37 I like it.

00:34:38 Can you talk about this idea of a spatial view of concepts

00:34:43 and so on?

00:34:44 Yeah, so can I just describe sort of the,

00:34:46 there’s an underlying core discovery,

00:34:49 which then everything comes from that.

00:34:51 That’s a very simple, this is really what happened.

00:34:55 We were deep into problems about understanding

00:34:58 how we build models of stuff in the world

00:35:00 and how we make predictions about things.

00:35:03 And I was holding a coffee cup just like this in my hand.

00:35:07 And my finger was touching the side, my index finger.

00:35:10 And then I moved it to the top

00:35:12 and I was gonna feel the rim at the top of the cup.

00:35:15 And I asked myself a very simple question.

00:35:18 I said, well, first of all, I say,

00:35:20 I know that my brain predicts what it’s gonna feel

00:35:22 before it touches it.

00:35:23 You can just think about it and imagine it.

00:35:26 And so we know that the brain’s making predictions

00:35:27 all the time.

00:35:28 So the question is, what does it take to predict that?

00:35:31 And there’s a very interesting answer.

00:35:33 First of all, it says the brain has to know

00:35:35 it’s touching a coffee cup.

00:35:36 It has to have a model of a coffee cup.

00:35:38 It needs to know where the finger currently is

00:35:41 on the cup relative to the cup.

00:35:43 Because when I make a movement,

00:35:44 it needs to know where it’s going to be on the cup

00:35:46 after the movement is completed relative to the cup.

00:35:50 And then it can make a prediction

00:35:51 about what it’s gonna sense.

00:35:53 So this told me that the neocortex,

00:35:54 which is making this prediction,

00:35:56 needs to know that it’s sensing it’s touching a cup.

00:35:59 And it needs to know the location of my finger

00:36:01 relative to that cup in a reference frame of the cup.

00:36:04 It doesn’t matter where the cup is relative to my body.

00:36:06 It doesn’t matter its orientation.

00:36:08 None of that matters.

00:36:09 It’s where my finger is relative to the cup,

00:36:10 which tells me then that the neocortex

00:36:13 has a reference frame that’s anchored to the cup.

00:36:17 Because otherwise I wouldn’t be able to say the location

00:36:19 and I wouldn’t be able to predict my new location.

00:36:21 And then we quickly, very instantly can say,

00:36:24 well, every part of my skin could touch this cup.

00:36:26 And therefore every part of my skin is making predictions

00:36:28 and every part of my skin must have a reference frame

00:36:30 that it’s using to make predictions.

00:36:33 So the big idea is that throughout the neocortex,

00:36:39 there are, everything is being stored

00:36:44 and referenced in reference frames.

00:36:46 You can think of them like XYZ reference frames,

00:36:48 but they’re not like that.

00:36:50 We know a lot about the neural mechanisms for this,

00:36:52 but the brain thinks in reference frames.

00:36:54 And as an engineer, if you’re an engineer,

00:36:56 this is not surprising.

00:36:57 You’d say, if I wanted to build a CAD model

00:37:00 of the coffee cup, well, I would bring it up

00:37:02 and some CAD software, and I would assign

00:37:04 some reference frame and say this features

00:37:05 at this locations and so on.

00:37:06 But the fact that this, the idea that this is occurring

00:37:09 throughout the neocortex everywhere, it was a novel idea.

00:37:14 And then a zillion things fell into place after that,

00:37:19 a zillion.

00:37:19 So now we think about the neocortex

00:37:21 as processing information quite differently

00:37:23 than we used to do it.

00:37:24 We used to think about the neocortex

00:37:25 as processing sensory data and extracting features

00:37:28 from that sensory data and then extracting features

00:37:30 from the features, very much like a deep learning network

00:37:33 does today.

00:37:34 But that’s not how the brain works at all.

00:37:36 The brain works by assigning everything,

00:37:39 every input, everything to reference frames.

00:37:41 And there are thousands, hundreds of thousands

00:37:44 of them active at once in your neocortex.

00:37:47 It’s a surprising thing to think about,

00:37:49 but once you sort of internalize this,

00:37:51 you understand that it explains almost every,

00:37:54 almost all the mysteries we’ve had about this structure.

00:37:57 So one of the consequences of that

00:38:00 is that every small part of the neocortex,

00:38:02 say a millimeter square, and there’s 150,000 of those.

00:38:06 So it’s about 150,000 square millimeters.

00:38:08 If you take every little square millimeter of the cortex,

00:38:11 it’s got some input coming into it

00:38:13 and it’s gonna have reference frames

00:38:14 where it’s assigned that input to.

00:38:16 And each square millimeter can learn

00:38:19 complete models of objects.

00:38:20 So what do I mean by that?

00:38:22 If I’m touching the coffee cup,

00:38:23 well, if I just touch it in one place,

00:38:25 I can’t learn what this coffee cup is

00:38:27 because I’m just feeling one part.

00:38:28 But if I move it around the cup

00:38:31 and touch it at different areas,

00:38:32 I can build up a complete model of the cup

00:38:34 because I’m now filling in that three dimensional map,

00:38:36 which is the coffee cup.

00:38:37 I can say, oh, what am I feeling

00:38:38 at all these different locations?

00:38:39 That’s the basic idea, it’s more complicated than that.

00:38:43 But so through time, and we talked about time earlier,

00:38:46 through time, even a single column,

00:38:48 which is only looking at, or a single part of the cortex,

00:38:50 which is only looking at a small part of the world,

00:38:52 can build up a complete model of an object.

00:38:55 And so if you think about the part of the brain,

00:38:57 which is getting input from all my fingers,

00:38:59 so they’re spread across the top of your head here.

00:39:01 This is the somatosensory cortex.

00:39:04 There’s columns associated

00:39:05 with all the different areas of my skin.

00:39:07 And what we believe is happening

00:39:10 is that all of them are building models of this cup,

00:39:12 every one of them, or things.

00:39:15 They’re not all building,

00:39:16 not every column or every part of the cortex

00:39:18 builds models of everything,

00:39:19 but they’re all building models of something.

00:39:21 And so you have, so when I touch this cup with my hand,

00:39:26 there are multiple models of the cup being invoked.

00:39:28 If I look at it with my eyes,

00:39:30 there are, again, many models of the cup being invoked,

00:39:32 because each part of the visual system,

00:39:34 the brain doesn’t process an image.

00:39:35 That’s a misleading idea.

00:39:38 It’s just like your fingers touching the cup,

00:39:40 so different parts of my retina

00:39:41 are looking at different parts of the cup.

00:39:42 And thousands and thousands of models of the cup

00:39:45 are being invoked at once.

00:39:47 And they’re all voting with each other,

00:39:48 trying to figure out what’s going on.

00:39:50 So that’s why we call it the thousand brains theory

00:39:51 of intelligence, because there isn’t one model of a cup.

00:39:54 There are thousands of models of this cup.

00:39:56 There are thousands of models of your cellphone

00:39:57 and about cameras and microphones and so on.

00:40:00 It’s a distributed modeling system,

00:40:02 which is very different

00:40:03 than the way people have thought about it.

00:40:04 And so that’s a really compelling and interesting idea.

00:40:07 I have two first questions.

00:40:08 So one, on the ensemble part of everything coming together,

00:40:12 you have these thousand brains.

00:40:14 How do you know which one has done the best job

00:40:17 of forming the…

00:40:18 Great question.

00:40:19 Let me try to explain it.

00:40:20 There’s a problem that’s known in neuroscience

00:40:23 called the sensor fusion problem.

00:40:25 Yes.

00:40:26 And so the idea is there’s something like,

00:40:27 oh, the image comes from the eye.

00:40:29 There’s a picture on the retina

00:40:30 and then it gets projected to the neocortex.

00:40:32 Oh, by now it’s all spread out all over the place

00:40:35 and it’s kind of squirrely and distorted

00:40:37 and pieces are all over the…

00:40:39 It doesn’t look like a picture anymore.

00:40:40 When does it all come back together again?

00:40:43 Or you might say, well, yes,

00:40:45 but I also have sounds or touches associated with the cup.

00:40:48 So I’m seeing the cup and touching the cup.

00:40:50 How do they get combined together again?

00:40:52 So it’s called the sensor fusion problem.

00:40:54 As if all these disparate parts

00:40:55 have to be brought together into one model someplace.

00:40:59 That’s the wrong idea.

00:41:01 The right idea is that you’ve got all these guys voting.

00:41:03 There’s auditory models of the cup.

00:41:05 There’s visual models of the cup.

00:41:06 There’s tactile models of the cup.

00:41:09 In the vision system,

00:41:10 there might be ones that are more focused on black and white

00:41:12 and ones focusing on color.

00:41:13 It doesn’t really matter.

00:41:14 There’s just thousands and thousands of models of this cup.

00:41:17 And they vote.

00:41:17 They don’t actually come together in one spot.

00:41:20 Just literally think of it this way.

00:41:21 Imagine you have these columns

00:41:24 that are like about the size of a little piece of spaghetti.

00:41:26 Like a two and a half millimeters tall

00:41:28 and about a millimeter in wide.

00:41:30 They’re not physical, but you could think of them that way.

00:41:33 And each one’s trying to guess what this thing is

00:41:35 or touching.

00:41:36 Now, they can do a pretty good job

00:41:38 if they’re allowed to move over time.

00:41:40 So I can reach my hand into a black box

00:41:41 and move my finger around an object.

00:41:43 And if I touch enough spaces, I go, okay,

00:41:45 now I know what it is.

00:41:46 But often we don’t do that.

00:41:48 Often I can just reach and grab something with my hand

00:41:49 all at once and I get it.

00:41:51 Or if I had to look through the world through a straw,

00:41:53 so I’m only invoking one little column,

00:41:55 I can only see part of something

00:41:56 because I have to move the straw around.

00:41:58 But if I open my eyes, I see the whole thing at once.

00:42:00 So what we think is going on

00:42:01 is all these little pieces of spaghetti,

00:42:03 if you will, all these little columns in the cortex,

00:42:05 are all trying to guess what it is that they’re sensing.

00:42:08 They’ll do a better guess if they have time

00:42:10 and can move over time.

00:42:11 So if I move my eyes, I move my fingers.

00:42:13 But if they don’t, they have a poor guess.

00:42:16 It’s a probabilistic guess of what they might be touching.

00:42:20 Now, imagine they can post their probability

00:42:22 at the top of a little piece of spaghetti.

00:42:24 Each one of them says,

00:42:25 I think, and it’s not really a probability distribution.

00:42:27 It’s more like a set of possibilities.

00:42:29 In the brain, it doesn’t work as a probability distribution.

00:42:31 It works as more like what we call a union.

00:42:34 So you could say, and one column says,

00:42:35 I think it could be a coffee cup,

00:42:37 a soda can, or a water bottle.

00:42:39 And another column says,

00:42:40 I think it could be a coffee cup

00:42:42 or a telephone or a camera or whatever, right?

00:42:46 And all these guys are saying what they think it might be.

00:42:49 And there’s these long range connections

00:42:51 in certain layers in the cortex.

00:42:53 So there’s in some layers in some cells types

00:42:56 in each column, send the projections across the brain.

00:43:00 And that’s the voting occurs.

00:43:01 And so there’s a simple associative memory mechanism.

00:43:04 We’ve described this in a recent paper

00:43:06 and we’ve modeled this that says,

00:43:09 they can all quickly settle on the only

00:43:11 or the one best answer for all of them.

00:43:14 If there is a single best answer,

00:43:16 they all vote and say, yep, it’s gotta be the coffee cup.

00:43:18 And at that point, they all know it’s a coffee cup.

00:43:21 And at that point, everyone acts as if it’s a coffee cup.

00:43:23 They’re like, yep, we know it’s a coffee,

00:43:24 even though I’ve only seen one little piece of this world,

00:43:26 I know it’s a coffee cup I’m touching

00:43:27 or I’m seeing or whatever.

00:43:28 And so you can think of all these columns

00:43:30 are looking at different parts in different places,

00:43:33 different sensory input, different locations,

00:43:35 they’re all different.

00:43:36 But this layer that’s doing the voting, it solidifies.

00:43:40 It’s just like it crystallizes and says,

00:43:42 oh, we all know what we’re doing.

00:43:44 And so you don’t bring these models together in one model,

00:43:46 you just vote and there’s a crystallization of the vote.

00:43:49 Great, that’s at least a compelling way

00:43:51 to think about the way you form a model of the world.

00:43:58 Now, you talk about a coffee cup.

00:44:00 Do you see this, as far as I understand,

00:44:03 you are proposing this as well,

00:44:04 that this extends to much more than coffee cups?

00:44:06 Yeah.

00:44:07 It does.

00:44:09 Or at least the physical world,

00:44:10 it expands to the world of concepts.

00:44:14 Yeah, it does.

00:44:15 And well, first, the primary thing is evidence for that

00:44:18 is that the regions of the neocortex

00:44:20 that are associated with language

00:44:22 or high level thought or mathematics

00:44:23 or things like that,

00:44:24 they look like the regions of the neocortex

00:44:26 that process vision, hearing, and touch.

00:44:28 They don’t look any different.

00:44:29 Or they look only marginally different.

00:44:32 And so one would say, well, if Vernon Mountcastle,

00:44:36 who proposed that all the parts of the neocortex

00:44:38 do the same thing, if he’s right,

00:44:41 then the parts that are doing language

00:44:42 or mathematics or physics

00:44:44 are working on the same principle.

00:44:45 They must be working on the principle of reference frames.

00:44:48 So that’s a little odd thought.

00:44:51 But of course, we had no prior idea

00:44:53 how these things happen.

00:44:55 So let’s go with that.

00:44:57 And we, in our recent paper,

00:44:59 we talked a little bit about that.

00:45:01 I’ve been working on it more since.

00:45:02 I have better ideas about it now.

00:45:05 I’m sitting here very confident

00:45:06 that that’s what’s happening.

00:45:08 And I can give you some examples

00:45:09 that help you think about that.

00:45:11 It’s not we understand it completely,

00:45:12 but I understand it better than I’ve described it

00:45:14 in any paper so far.

00:45:15 So, but we did put that idea out there.

00:45:17 It says, okay, this is,

00:45:18 it’s a good place to start, you know?

00:45:22 And the evidence would suggest it’s how it’s happening.

00:45:24 And then we can start tackling that problem

00:45:26 one piece at a time.

00:45:27 Like, what does it mean to do high level thought?

00:45:29 What does it mean to do language?

00:45:30 How would that fit into a reference frame framework?

00:45:34 Yeah, so there’s a,

00:45:35 I don’t know if you could tell me if there’s a connection,

00:45:37 but there’s an app called Anki

00:45:40 that helps you remember different concepts.

00:45:42 And they talk about like a memory palace

00:45:45 that helps you remember completely random concepts

00:45:47 by trying to put them in a physical space in your mind

00:45:51 and putting them next to each other.

00:45:52 It’s called the method of loci.

00:45:53 Loci, yeah.

00:45:54 For some reason, that seems to work really well.

00:45:57 Now, that’s a very narrow kind of application

00:45:59 of just remembering some facts.

00:46:00 But that’s a very, very telling one.

00:46:03 Yes, exactly.

00:46:04 So this seems like you’re describing a mechanism

00:46:06 why this seems to work.

00:46:09 So basically the way what we think is going on

00:46:11 is all things you know, all concepts, all ideas,

00:46:15 words, everything you know are stored in reference frames.

00:46:20 And so if you want to remember something,

00:46:24 you have to basically navigate through a reference frame

00:46:26 the same way a rat navigates through a maze

00:46:28 and the same way my finger navigates to this coffee cup.

00:46:31 You are moving through some space.

00:46:33 And so if you have a random list of things

00:46:35 you were asked to remember,

00:46:37 by assigning them to a reference frame,

00:46:39 you’ve already know very well to see your house, right?

00:46:42 And the idea of the method of loci is you can say,

00:46:43 okay, in my lobby, I’m going to put this thing.

00:46:45 And then the bedroom, I put this one.

00:46:47 I go down the hall, I put this thing.

00:46:48 And then you want to recall those facts

00:46:50 or recall those things.

00:46:51 You just walk mentally, you walk through your house.

00:46:54 You’re mentally moving through a reference frame

00:46:56 that you already had.

00:46:57 And that tells you,

00:46:59 there’s two things that are really important about that.

00:47:00 It tells us the brain prefers to store things

00:47:02 in reference frames.

00:47:03 And that the method of recalling things

00:47:06 or thinking, if you will,

00:47:08 is to move mentally through those reference frames.

00:47:11 You could move physically through some reference frames,

00:47:13 like I could physically move through the reference frame

00:47:15 of this coffee cup.

00:47:16 I can also mentally move through the reference frame

00:47:17 of the coffee cup, imagining me touching it.

00:47:19 But I can also mentally move my house.

00:47:22 And so now we can ask yourself,

00:47:24 or are all concepts stored this way?

00:47:26 There was some recent research using human subjects

00:47:31 in fMRI, and I’m going to apologize for not knowing

00:47:33 the name of the scientists who did this.

00:47:36 But what they did is they put humans in this fMRI machine,

00:47:41 which is one of these imaging machines.

00:47:42 And they gave the humans tasks to think about birds.

00:47:46 So they had different types of birds,

00:47:47 and birds that look big and small,

00:47:49 and long necks and long legs, things like that.

00:47:52 And what they could tell from the fMRI

00:47:55 was a very clever experiment.

00:47:57 You get to tell when humans were thinking about the birds,

00:48:00 that the birds, the knowledge of birds

00:48:03 was arranged in a reference frame,

00:48:05 similar to the ones that are used

00:48:07 when you navigate in a room.

00:48:08 That these are called grid cells,

00:48:10 and there are grid cell like patterns of activity

00:48:12 in the neocortex when they do this.

00:48:15 So it’s a very clever experiment.

00:48:18 And what it basically says,

00:48:20 that even when you’re thinking about something abstract,

00:48:22 and you’re not really thinking about it as a reference frame,

00:48:24 it tells us the brain is actually using a reference frame.

00:48:26 And it’s using the same neural mechanisms.

00:48:28 These grid cells are the basic same neural mechanism

00:48:30 that we propose that grid cells,

00:48:32 which exist in the old part of the brain,

00:48:34 the entorhinal cortex, that that mechanism

00:48:37 is now similar mechanism is used throughout the neocortex.

00:48:40 It’s the same nature to preserve this interesting way

00:48:43 of creating reference frames.

00:48:44 And so now they have empirical evidence

00:48:46 that when you think about concepts like birds,

00:48:49 that you’re using reference frames

00:48:51 that are built on grid cells.

00:48:53 So that’s similar to the method of loci,

00:48:55 but in this case, the birds are related.

00:48:56 So they create their own reference frame,

00:48:58 which is consistent with bird space.

00:49:01 And when you think about something, you go through that.

00:49:03 You can make the same example,

00:49:04 let’s take mathematics.

00:49:06 Let’s say you wanna prove a conjecture.

00:49:09 What is a conjecture?

00:49:10 A conjecture is a statement you believe to be true,

00:49:13 but you haven’t proven it.

00:49:15 And so it might be an equation.

00:49:16 I wanna show that this is equal to that.

00:49:19 And you have some places you start with.

00:49:21 You say, well, I know this is true,

00:49:22 and I know this is true.

00:49:23 And I think that maybe to get to the final proof,

00:49:25 I need to go through some intermediate results.

00:49:28 What I believe is happening is literally these equations

00:49:33 or these points are assigned to a reference frame,

00:49:36 a mathematical reference frame.

00:49:37 And when you do mathematical operations,

00:49:39 a simple one might be multiply or divide,

00:49:41 but you might be a little plus transform or something else.

00:49:44 That is like a movement in the reference frame of the math.

00:49:47 And so you’re literally trying to discover a path

00:49:50 from one location to another location

00:49:52 in a space of mathematics.

00:49:56 And if you can get to these intermediate results,

00:49:58 then you know your map is pretty good,

00:50:00 and you know you’re using the right operations.

00:50:02 Much of what we think about is solving hard problems

00:50:05 is designing the correct reference frame for that problem,

00:50:08 figuring out how to organize the information

00:50:11 and what behaviors I wanna use in that space

00:50:14 to get me there.

00:50:16 Yeah, so if you dig in an idea of this reference frame,

00:50:19 whether it’s the math, you start a set of axioms

00:50:21 to try to get to proving the conjecture.

00:50:25 Can you try to describe, maybe take a step back,

00:50:28 how you think of the reference frame in that context?

00:50:30 Is it the reference frame that the axioms are happy in?

00:50:36 Is it the reference frame that might contain everything?

00:50:38 Is it a changing thing as you?

00:50:41 You have many, many reference frames.

00:50:43 I mean, in fact, the way the theory,

00:50:44 the thousand brain theory of intelligence says

00:50:46 that every single thing in the world

00:50:47 has its own reference frame.

00:50:48 So every word has its own reference frames.

00:50:50 And we can talk about this.

00:50:52 The mathematics work out,

00:50:54 this is no problem for neurons to do this.

00:50:55 But how many reference frames does a coffee cup have?

00:50:58 Well, it’s on a table.

00:51:00 Let’s say you ask how many reference frames

00:51:03 could a column in my finger

00:51:06 that’s touching the coffee cup have?

00:51:07 Because there are many, many copy,

00:51:09 there are many, many models of the coffee cup.

00:51:10 So the coffee, there is no one model of a coffee cup.

00:51:13 There are many models of a coffee cup.

00:51:14 And you could say, well,

00:51:15 how many different things can my finger learn?

00:51:17 Is this the question you want to ask?

00:51:19 Imagine I say every concept, every idea,

00:51:21 everything you’ve ever know about that you can say,

00:51:23 I know that thing has a reference frame

00:51:27 associated with it.

00:51:28 And what we do when we build composite objects,

00:51:30 we assign reference frames

00:51:32 to point another reference frame.

00:51:33 So my coffee cup has multiple components to it.

00:51:37 It’s got a limb, it’s got a cylinder, it’s got a handle.

00:51:40 And those things have their own reference frames

00:51:42 and they’re assigned to a master reference frame,

00:51:45 which is called this cup.

00:51:46 And now I have this Numenta logo on it.

00:51:48 Well, that’s something that exists elsewhere in the world.

00:51:50 It’s its own thing.

00:51:51 So it has its own reference frame.

00:51:52 So we now have to say,

00:51:53 how can I assign the Numenta logo reference frame

00:51:56 onto the cylinder or onto the coffee cup?

00:51:59 So it’s all, we talked about this in the paper

00:52:01 that came out in December of this last year.

00:52:06 The idea of how you can assign reference frames

00:52:08 to reference frames, how neurons could do this.

00:52:10 So, well, my question is,

00:52:12 even though you mentioned reference frames a lot,

00:52:14 I almost feel it’s really useful to dig into

00:52:16 how you think of what a reference frame is.

00:52:20 I mean, it was already helpful for me to understand

00:52:22 that you think of reference frames

00:52:23 as something there is a lot of.

00:52:26 Okay, so let’s just say that we’re gonna have

00:52:28 some neurons in the brain, not many, actually,

00:52:31 10,000, 20,000 are gonna create

00:52:32 a whole bunch of reference frames.

00:52:34 What does it mean?

00:52:35 What is a reference frame?

00:52:37 First of all, these reference frames are different

00:52:40 than the ones you might be used to.

00:52:42 We know lots of reference frames.

00:52:43 For example, we know the Cartesian coordinates, X, Y, Z,

00:52:46 that’s a type of reference frame.

00:52:47 We know longitude and latitude,

00:52:50 that’s a different type of reference frame.

00:52:52 If I look at a printed map,

00:52:54 you might have columns A through M,

00:52:58 and rows one through 20,

00:53:00 that’s a different type of reference frame.

00:53:01 It’s kind of a Cartesian coordinate reference frame.

00:53:04 The interesting thing about the reference frames

00:53:06 in the brain, and we know this because these

00:53:08 have been established through neuroscience

00:53:10 studying the entorhinal cortex.

00:53:12 So I’m not speculating here, okay?

00:53:13 This is known neuroscience in an old part of the brain.

00:53:16 The way these cells create reference frames,

00:53:18 they have no origin.

00:53:20 So what it’s more like, you have a point,

00:53:24 a point in some space, and you,

00:53:27 given a particular movement,

00:53:29 you can then tell what the next point should be.

00:53:32 And you can then tell what the next point would be,

00:53:34 and so on.

00:53:35 You can use this to calculate

00:53:38 how to get from one point to another.

00:53:40 So how do I get from my house to my home,

00:53:43 or how do I get my finger from the side of my cup

00:53:44 to the top of the cup?

00:53:46 How do I get from the axioms to the conjecture?

00:53:50 So it’s a different type of reference frame,

00:53:54 and I can, if you want, I can describe in more detail,

00:53:57 I can paint a picture of how you might want

00:53:59 to think about that.

00:53:59 It’s really helpful to think it’s something

00:54:00 you can move through, but is there,

00:54:03 is it helpful to think of it as spatial in some sense,

00:54:08 or is there something that’s more?

00:54:09 No, it’s definitely spatial.

00:54:11 It’s spatial in a mathematical sense.

00:54:13 How many dimensions?

00:54:14 Can it be a crazy number of dimensions?

00:54:16 Well, that’s an interesting question.

00:54:17 In the old part of the brain, the entorhinal cortex,

00:54:20 they studied rats, and initially it looks like,

00:54:22 oh, this is just two dimensional.

00:54:24 It’s like the rat is in some box in the maze or whatever,

00:54:27 and they know where the rat is using

00:54:28 these two dimensional reference frames

00:54:30 to know where it is in the maze.

00:54:32 We said, well, okay, but what about bats?

00:54:35 That’s a mammal, and they fly in three dimensional space.

00:54:38 How do they do that?

00:54:39 They seem to know where they are, right?

00:54:41 So this is a current area of active research,

00:54:44 and it seems like somehow the neurons

00:54:46 in the entorhinal cortex can learn three dimensional space.

00:54:50 We just, two members of our team,

00:54:52 along with Elif Fett from MIT,

00:54:55 just released a paper this literally last week.

00:54:59 It’s on bioRxiv, where they show that you can,

00:55:03 if you, the way these things work,

00:55:05 and I won’t get, unless you want to,

00:55:06 I won’t get into the detail,

00:55:08 but grid cells can represent any n dimensional space.

00:55:12 It’s not inherently limited.

00:55:15 You can think of it this way.

00:55:16 If you had two dimensional, the way it works

00:55:18 is you had a bunch of two dimensional slices.

00:55:20 That’s the way these things work.

00:55:21 There’s a whole bunch of two dimensional models,

00:55:24 and you can just, you can slice up

00:55:26 any n dimensional space with two dimensional projections.

00:55:29 So, and you could have one dimensional models.

00:55:31 So there’s nothing inherent about the mathematics

00:55:34 about the way the neurons do this,

00:55:35 which constrain the dimensionality of the space,

00:55:39 which I think was important.

00:55:41 So obviously I have a three dimensional map of this cup.

00:55:44 Maybe it’s even more than that, I don’t know.

00:55:46 But it’s clearly a three dimensional map of the cup.

00:55:48 I don’t just have a projection of the cup.

00:55:50 But when I think about birds,

00:55:52 or when I think about mathematics,

00:55:53 perhaps it’s more than three dimensions.

00:55:55 Who knows?

00:55:56 So in terms of each individual column

00:56:00 building up more and more information over time,

00:56:04 do you think that mechanism is well understood?

00:56:06 In your mind, you’ve proposed a lot of architectures there.

00:56:09 Is that a key piece, or is it,

00:56:11 is the big piece, the thousand brain theory of intelligence,

00:56:16 the ensemble of it all?

00:56:17 Well, I think they’re both big.

00:56:18 I mean, clearly the concept, as a theorist,

00:56:20 the concept is most exciting, right?

00:56:23 The high level concept.

00:56:23 The high level concept.

00:56:24 This is a totally new way of thinking

00:56:26 about how the neocortex works.

00:56:27 So that is appealing.

00:56:28 It has all these ramifications.

00:56:30 And with that, as a framework for how the brain works,

00:56:33 you can make all kinds of predictions

00:56:34 and solve all kinds of problems.

00:56:36 Now we’re trying to work through

00:56:37 many of these details right now.

00:56:38 Okay, how do the neurons actually do this?

00:56:40 Well, it turns out, if you think about grid cells

00:56:42 and place cells in the old parts of the brain,

00:56:44 there’s a lot that’s known about them,

00:56:45 but there’s still some mysteries.

00:56:47 There’s a lot of debate about exactly the details,

00:56:49 how these work and what are the signs.

00:56:50 And we have that still, that same level of detail,

00:56:52 that same level of concern.

00:56:54 What we spend here most of our time doing

00:56:56 is trying to make a very good list

00:57:00 of the things we don’t understand yet.

00:57:02 That’s the key part here.

00:57:04 What are the constraints?

00:57:05 It’s not like, oh, this thing seems to work, we’re done.

00:57:07 No, it’s like, okay, it kind of works,

00:57:08 but these are other things we know it has to do

00:57:10 and it’s not doing those yet.

00:57:12 I would say we’re well on the way here.

00:57:15 We’re not done yet.

00:57:17 There’s a lot of trickiness to this system,

00:57:20 but the basic principles about how different layers

00:57:23 in the neocortex are doing much of this, we understand.

00:57:27 But there’s some fundamental parts

00:57:28 that we don’t understand as well.

00:57:30 So what would you say is one of the harder open problems

00:57:34 or one of the ones that have been bothering you,

00:57:37 keeping you up at night the most?

00:57:38 Oh, well, right now, this is a detailed thing

00:57:40 that wouldn’t apply to most people, okay?

00:57:42 Sure.

00:57:43 But you want me to answer that question?

00:57:44 Yeah, please.

00:57:46 We’ve talked about as if, oh,

00:57:48 to predict what you’re going to sense on this coffee cup,

00:57:50 I need to know where my finger is gonna be

00:57:52 on the coffee cup.

00:57:53 That is true, but it’s insufficient.

00:57:56 Think about my finger touches the edge of the coffee cup.

00:57:58 My finger can touch it at different orientations.

00:58:01 I can rotate my finger around here and that doesn’t change.

00:58:06 I can make that prediction and somehow,

00:58:08 so it’s not just the location.

00:58:10 There’s an orientation component of this as well.

00:58:13 This is known in the old parts of the brain too.

00:58:15 There’s things called head direction cells,

00:58:16 which way the rat is facing.

00:58:18 It’s the same kind of basic idea.

00:58:20 So if my finger were a rat, you know, in three dimensions,

00:58:23 I have a three dimensional orientation

00:58:25 and I have a three dimensional location.

00:58:27 If I was a rat, I would have a,

00:58:28 you might think of it as a two dimensional location,

00:58:30 a two dimensional orientation,

00:58:31 a one dimensional orientation,

00:58:32 like just which way is it facing?

00:58:35 So how the two components work together,

00:58:38 how it is that I combine orientation,

00:58:41 the orientation of my sensor,

00:58:43 as well as the location is a tricky problem.

00:58:49 And I think I’ve made progress on it.

00:58:52 So at a bigger version of that,

00:58:55 so perspective is super interesting, but super specific.

00:58:58 Yeah, I warned you.

00:59:00 No, no, no, that’s really good,

00:59:01 but there’s a more general version of that.

00:59:03 Do you think context matters,

00:59:06 the fact that we’re in a building in North America,

00:59:10 that we, in the day and age where we have mugs?

00:59:15 I mean, there’s all this extra information

00:59:19 that you bring to the table about everything else

00:59:22 in the room that’s outside of just the coffee cup.

00:59:24 How does it get connected, do you think?

00:59:27 Yeah, and that is another really interesting question.

00:59:30 I’m gonna throw that under the rubric

00:59:32 or the name of attentional problems.

00:59:35 First of all, we have this model,

00:59:36 I have many, many models.

00:59:38 And also the question, does it matter?

00:59:40 Well, it matters for certain things, of course it does.

00:59:42 Maybe what we think of that as a coffee cup

00:59:44 in another part of the world

00:59:45 is viewed as something completely different.

00:59:47 Or maybe our logo, which is very benign

00:59:50 in this part of the world,

00:59:51 it means something very different

00:59:52 in another part of the world.

00:59:53 So those things do matter.

00:59:57 I think the way to think about it is the following,

01:00:00 one way to think about it,

01:00:01 is we have all these models of the world, okay?

01:00:04 And we model everything.

01:00:06 And as I said earlier, I kind of snuck it in there,

01:00:08 our models are actually, we build composite structure.

01:00:12 So every object is composed of other objects,

01:00:15 which are composed of other objects,

01:00:16 and they become members of other objects.

01:00:18 So this room has chairs and a table and a room

01:00:20 and walls and so on.

01:00:21 Now we can just arrange these things in a certain way

01:00:24 and go, oh, that’s the nomenclature conference room.

01:00:26 So, and what we do is when we go around the world

01:00:31 and we experience the world,

01:00:33 by walking into a room, for example,

01:00:35 the first thing I do is I can say,

01:00:36 oh, I’m in this room, do I recognize the room?

01:00:38 Then I can say, oh, look, there’s a table here.

01:00:41 And by attending to the table,

01:00:43 I’m then assigning this table in the context of the room.

01:00:45 Then I can say, oh, on the table, there’s a coffee cup.

01:00:48 Oh, and on the table, there’s a logo.

01:00:49 And in the logo, there’s the word Nementa.

01:00:51 Oh, and look in the logo, there’s the letter E.

01:00:53 Oh, and look, it has an unusual serif.

01:00:55 And it doesn’t actually, but I pretended to serif.

01:00:59 So the point is your attention is kind of drilling

01:01:03 deep in and out of these nested structures.

01:01:07 And I can pop back up and I can pop back down.

01:01:09 I can pop back up and I can pop back down.

01:01:10 So when I attend to the coffee cup,

01:01:13 I haven’t lost the context of everything else,

01:01:15 but it’s sort of, there’s this sort of nested structure.

01:01:18 So the attention filters the reference frame information

01:01:22 for that particular period of time?

01:01:24 Yes, it basically, moment to moment,

01:01:26 you attend the sub components,

01:01:28 and then you can attend the sub components

01:01:29 to sub components.

01:01:30 And you can move up and down.

01:01:31 You can move up and down.

01:01:32 We do that all the time.

01:01:33 You’re not even, now that I’m aware of it,

01:01:35 I’m very conscious of it.

01:01:36 But until, but most people don’t even think about this.

01:01:39 You just walk in a room and you don’t say,

01:01:41 oh, I looked at the chair and I looked at the board

01:01:43 and looked at that word on the board

01:01:44 and I looked over here, what’s going on, right?

01:01:47 So what percent of your day are you deeply aware of this?

01:01:50 And what part can you actually relax and just be Jeff?

01:01:52 Me personally, like my personal day?

01:01:54 Yeah.

01:01:55 Unfortunately, I’m afflicted with too much of the former.

01:02:01 Well, unfortunately or unfortunately.

01:02:02 Yeah.

01:02:03 You don’t think it’s useful?

01:02:04 Oh, it is useful, totally useful.

01:02:06 I think about this stuff almost all the time.

01:02:09 And one of my primary ways of thinking

01:02:12 is when I’m in sleep at night,

01:02:13 I always wake up in the middle of the night.

01:02:15 And then I stay awake for at least an hour

01:02:17 with my eyes shut in sort of a half sleep state

01:02:20 thinking about these things.

01:02:21 I come up with answers to problems very often

01:02:23 in that sort of half sleeping state.

01:02:25 I think about it on my bike ride, I think about it on walks.

01:02:27 I’m just constantly thinking about this.

01:02:28 I have to almost schedule time

01:02:32 to not think about this stuff

01:02:34 because it’s very, it’s mentally taxing.

01:02:37 Are you, when you’re thinking about this stuff,

01:02:39 are you thinking introspectively,

01:02:41 like almost taking a step outside of yourself

01:02:43 and trying to figure out what is your mind doing right now?

01:02:45 I do that all the time, but that’s not all I do.

01:02:49 I’m constantly observing myself.

01:02:50 So as soon as I started thinking about grid cells,

01:02:53 for example, and getting into that,

01:02:55 I started saying, oh, well, grid cells

01:02:56 can have my place of sense in the world.

01:02:58 That’s where you know where you are.

01:02:59 And it’s interesting, we always have a sense

01:03:01 of where we are unless we’re lost.

01:03:03 And so I started at night when I got up

01:03:04 to go to the bathroom, I would start trying to do it

01:03:06 completely with my eyes closed all the time.

01:03:08 And I would test my sense of grid cells.

01:03:10 I would walk five feet and say, okay, I think I’m here.

01:03:13 Am I really there?

01:03:14 What’s my error?

01:03:15 And then I would calculate my error again

01:03:16 and see how the errors could accumulate.

01:03:17 So even something as simple as getting up

01:03:19 in the middle of the night to go to the bathroom,

01:03:20 I’m testing these theories out.

01:03:22 It’s kind of fun.

01:03:23 I mean, the coffee cup is an example of that too.

01:03:25 So I find that these sort of everyday introspections

01:03:30 are actually quite helpful.

01:03:32 It doesn’t mean you can ignore the science.

01:03:34 I mean, I spend hours every day

01:03:37 reading ridiculously complex papers.

01:03:40 That’s not nearly as much fun,

01:03:41 but you have to sort of build up those constraints

01:03:44 and the knowledge about the field and who’s doing what

01:03:46 and what exactly they think is happening here.

01:03:48 And then you can sit back and say,

01:03:50 okay, let’s try to piece this all together.

01:03:53 Let’s come up with some, I’m very,

01:03:56 in this group here, people, they know they do,

01:03:58 I do this all the time.

01:03:59 I come in with these introspective ideas and say,

01:04:01 well, have you ever thought about this?

01:04:02 Now watch, well, let’s all do this together.

01:04:04 And it’s helpful.

01:04:05 It’s not, as long as you don’t,

01:04:09 all you did was that, then you’re just making up stuff.

01:04:12 But if you’re constraining it by the reality

01:04:14 of the neuroscience, then it’s really helpful.

01:04:17 So let’s talk a little bit about deep learning

01:04:20 and the successes in the applied space of neural networks,

01:04:26 ideas of training model on data

01:04:29 and these simple computational units,

01:04:31 artificial neurons that with backpropagation,

01:04:36 statistical ways of being able to generalize

01:04:40 from the training set onto data

01:04:42 that’s similar to that training set.

01:04:44 So where do you think are the limitations

01:04:47 of those approaches?

01:04:48 What do you think are its strengths

01:04:50 relative to your major efforts

01:04:52 of constructing a theory of human intelligence?

01:04:56 Well, I’m not an expert in this field.

01:04:57 I’m somewhat knowledgeable.

01:04:59 So, but I’m not.

01:04:59 Some of it is in just your intuition.

01:05:01 What are your?

01:05:02 Well, I have a little bit more than intuition,

01:05:03 but I just want to say like,

01:05:05 you know, one of the things that you asked me,

01:05:07 do I spend all my time thinking about neuroscience?

01:05:09 I do.

01:05:10 That’s to the exclusion of thinking about things

01:05:11 like convolutional neural networks.

01:05:13 But I try to stay current.

01:05:15 So look, I think it’s great, the progress they’ve made.

01:05:17 It’s fantastic.

01:05:18 And as I mentioned earlier,

01:05:19 it’s very highly useful for many things.

01:05:22 The models that we have today are actually derived

01:05:26 from a lot of neuroscience principles.

01:05:28 There are distributed processing systems

01:05:30 and distributed memory systems,

01:05:31 and that’s how the brain works.

01:05:33 They use things that we might call them neurons,

01:05:35 but they’re really not neurons at all.

01:05:37 So we can just, they’re not really neurons.

01:05:39 So they’re distributed processing systems.

01:05:41 And that nature of hierarchy,

01:05:44 that came also from neuroscience.

01:05:47 And so there’s a lot of things,

01:05:48 the learning rules, basically,

01:05:49 not back prop, but other, you know,

01:05:51 sort of heavy on top of that.

01:05:52 I’d be curious to say they’re not neurons at all.

01:05:55 Can you describe in which way?

01:05:56 I mean, some of it is obvious,

01:05:57 but I’d be curious if you have specific ways

01:06:00 in which you think are the biggest differences.

01:06:02 Yeah, we had a paper in 2016 called

01:06:04 Why Neurons Have Thousands of Synapses.

01:06:06 And if you read that paper,

01:06:09 you’ll know what I’m talking about here.

01:06:11 A real neuron in the brain is a complex thing.

01:06:14 And let’s just start with the synapses on it,

01:06:17 which is a connection between neurons.

01:06:19 Real neurons can have everywhere

01:06:20 from five to 30,000 synapses on them.

01:06:25 The ones near the cell body,

01:06:27 the ones that are close to the soma of the cell body,

01:06:30 those are like the ones that people model

01:06:32 in artificial neurons.

01:06:33 There is a few hundred of those.

01:06:35 Maybe they can affect the cell.

01:06:37 They can make the cell become active.

01:06:39 95% of the synapses can’t do that.

01:06:43 They’re too far away.

01:06:44 So if you activate one of those synapses,

01:06:45 it just doesn’t affect the cell body enough

01:06:47 to make any difference.

01:06:48 Any one of them individually.

01:06:50 Any one of them individually,

01:06:50 or even if you do a mass of them.

01:06:54 What real neurons do is the following.

01:06:57 If you activate or you get 10 to 20 of them

01:07:03 active at the same time,

01:07:04 meaning they’re all receiving an input at the same time,

01:07:06 and those 10 to 20 synapses or 40 synapses

01:07:09 within a very short distance on the dendrite,

01:07:11 like 40 microns, a very small area.

01:07:13 So if you activate a bunch of these

01:07:14 right next to each other at some distant place,

01:07:17 what happens is it creates

01:07:19 what’s called the dendritic spike.

01:07:21 And the dendritic spike travels through the dendrites

01:07:24 and can reach the soma or the cell body.

01:07:27 Now, when it gets there, it changes the voltage,

01:07:31 which is sort of like gonna make the cell fire,

01:07:33 but never enough to make the cell fire.

01:07:36 It’s sort of what we call, it says we depolarize the cell,

01:07:38 you raise the voltage a little bit,

01:07:39 but not enough to do anything.

01:07:41 It’s like, well, what good is that?

01:07:42 And then it goes back down again.

01:07:44 So we propose a theory,

01:07:47 which I’m very confident in basics are,

01:07:50 is that what’s happening there is

01:07:52 those 95% of the synapses are recognizing

01:07:55 dozens to hundreds of unique patterns.

01:07:58 They can write about 10, 20 synapses at a time,

01:08:02 and they’re acting like predictions.

01:08:04 So the neuron actually is a predictive engine on its own.

01:08:07 It can fire when it gets enough,

01:08:09 what they call proximal input

01:08:10 from those ones near the cell fire,

01:08:11 but it can get ready to fire from dozens to hundreds

01:08:15 of patterns that it recognizes from the other guys.

01:08:18 And the advantage of this to the neuron

01:08:21 is that when it actually does produce a spike

01:08:23 in action potential,

01:08:24 it does so slightly sooner than it would have otherwise.

01:08:27 And so what could is slightly sooner?

01:08:29 Well, the slightly sooner part is it,

01:08:31 all the excitatory neurons in the brain

01:08:34 are surrounded by these inhibitory neurons,

01:08:36 and they’re very fast, the inhibitory neurons,

01:08:38 these basket cells.

01:08:40 And if I get my spike out

01:08:42 a little bit sooner than someone else,

01:08:44 I inhibit all my neighbors around me, right?

01:08:47 And what you end up with is a different representation.

01:08:49 You end up with a reputation that matches your prediction.

01:08:52 It’s a sparser representation,

01:08:53 meaning fewer neurons are active,

01:08:55 but it’s much more specific.

01:08:57 And so we showed how networks of these neurons

01:09:00 can do very sophisticated temporal prediction, basically.

01:09:04 So this, summarize this,

01:09:07 real neurons in the brain are time based prediction engines,

01:09:10 and there’s no concept of this at all

01:09:14 in artificial, what we call point neurons.

01:09:18 I don’t think you can build a brain without them.

01:09:20 I don’t think you can build intelligence without them,

01:09:21 because it’s where a large part of the time comes from.

01:09:26 These are predictive models, and the time is,

01:09:29 there’s a prior and a prediction and an action,

01:09:32 and it’s inherent through every neuron in the neocortex.

01:09:34 So I would say that point neurons sort of model

01:09:37 a piece of that, and not very well at that either.

01:09:40 But like for example, synapses are very unreliable,

01:09:46 and you cannot assign any precision to them.

01:09:49 So even one digit of precision is not possible.

01:09:52 So the way real neurons work is they don’t add these,

01:09:55 they don’t change these weights accurately

01:09:57 like artificial neural networks do.

01:09:59 They basically form new synapses,

01:10:01 and so what you’re trying to always do is

01:10:03 detect the presence of some 10 to 20

01:10:06 active synapses at the same time,

01:10:08 as opposed, and they’re almost binary.

01:10:11 It’s like, because you can’t really represent

01:10:12 anything much finer than that.

01:10:14 So these are the kind of,

01:10:16 and I think that’s actually another essential component,

01:10:18 because the brain works on sparse patterns,

01:10:20 and all that mechanism is based on sparse patterns,

01:10:24 and I don’t actually think you could build real brains

01:10:26 or machine intelligence without

01:10:29 incorporating some of those ideas.

01:10:30 It’s hard to even think about the complexity

01:10:32 that emerges from the fact that

01:10:34 the timing of the firing matters in the brain,

01:10:37 the fact that you form new synapses,

01:10:40 and I mean, everything you just mentioned

01:10:44 in the past couple minutes.

01:10:44 Trust me, if you spend time on it,

01:10:46 you can get your mind around it.

01:10:47 It’s not like, it’s no longer a mystery to me.

01:10:49 No, but sorry, as a function, in a mathematical way,

01:10:53 can you start getting an intuition about

01:10:56 what gets it excited, what not,

01:10:58 and what kind of representation?

01:10:59 Yeah, it’s not as easy as,

01:11:02 there’s many other types of neural networks

01:11:04 that are more amenable to pure analysis,

01:11:09 especially very simple networks.

01:11:10 Oh, I have four neurons, and they’re doing this.

01:11:12 Can we describe to them mathematically

01:11:14 what they’re doing type of thing?

01:11:16 Even the complexity of convolutional neural networks today,

01:11:19 it’s sort of a mystery.

01:11:20 They can’t really describe the whole system.

01:11:22 And so it’s different.

01:11:24 My colleague Subitai Ahmad, he did a nice paper on this.

01:11:31 You can get all this stuff on our website

01:11:32 if you’re interested,

01:11:34 talking about sort of the mathematical properties

01:11:36 of sparse representations.

01:11:37 And so what we can do is we can show mathematically,

01:11:40 for example, why 10 to 20 synapses to recognize a pattern

01:11:44 is the correct number, is the right number you’d wanna use.

01:11:47 And by the way, that matches biology.

01:11:49 We can show mathematically some of these concepts

01:11:53 about the show why the brain is so robust

01:11:58 to noise and error and fallout and so on.

01:12:01 We can show that mathematically

01:12:02 as well as empirically in simulations.

01:12:05 But the system can’t be analyzed completely.

01:12:07 Any complex system can’t, and so that’s out of the realm.

01:12:11 But there is mathematical benefits and intuitions

01:12:17 that can be derived from mathematics.

01:12:19 And we try to do that as well.

01:12:20 Most of our papers have a section about that.

01:12:23 So I think it’s refreshing and useful for me

01:12:25 to be talking to you about deep neural networks,

01:12:29 because your intuition basically says

01:12:30 that we can’t achieve anything like intelligence

01:12:34 with artificial neural networks.

01:12:35 Well, not in the current form.

01:12:36 Not in the current form.

01:12:37 I’m sure we can do it in the ultimate form, sure.

01:12:40 So let me dig into it

01:12:41 and see what your thoughts are there a little bit.

01:12:43 So I’m not sure if you read this little blog post

01:12:45 called Bitter Lesson by Rich Sutton recently.

01:12:49 He’s a reinforcement learning pioneer.

01:12:51 I’m not sure if you’re familiar with him.

01:12:53 His basic idea is that all the stuff we’ve done in AI

01:12:56 in the past 70 years, he’s one of the old school guys.

01:13:02 The biggest lesson learned is that all the tricky things

01:13:06 we’ve done, they benefit in the short term,

01:13:10 but in the long term, what wins out

01:13:12 is a simple general method that just relies on Moore’s law,

01:13:16 on computation getting faster and faster.

01:13:19 This is what he’s saying.

01:13:21 This is what has worked up to now.

01:13:23 This is what has worked up to now.

01:13:25 If you’re trying to build a system,

01:13:29 if we’re talking about,

01:13:30 he’s not concerned about intelligence.

01:13:31 He’s concerned about a system that works

01:13:34 in terms of making predictions

01:13:36 on applied narrow AI problems, right?

01:13:38 That’s what this discussion is about.

01:13:40 That you just try to go as general as possible

01:13:44 and wait years or decades for the computation

01:13:48 to make it actually.

01:13:50 Is he saying that as a criticism

01:13:51 or is he saying this is a prescription

01:13:53 of what we ought to be doing?

01:13:54 Well, it’s very difficult.

01:13:55 He’s saying this is what has worked

01:13:57 and yes, a prescription, but it’s a difficult prescription

01:14:00 because it says all the fun things

01:14:02 you guys are trying to do, we are trying to do.

01:14:05 He’s part of the community.

01:14:07 He’s saying it’s only going to be short term gains.

01:14:10 So this all leads up to a question, I guess,

01:14:13 on artificial neural networks

01:14:15 and maybe our own biological neural networks

01:14:19 is do you think if we just scale things up significantly,

01:14:23 so take these dumb artificial neurons,

01:14:27 the point neurons, I like that term.

01:14:30 If we just have a lot more of them,

01:14:33 do you think some of the elements

01:14:34 that we see in the brain may start emerging?

01:14:38 No, I don’t think so.

01:14:39 We can do bigger problems of the same type.

01:14:43 I mean, it’s been pointed out by many people

01:14:45 that today’s convolutional neural networks

01:14:46 aren’t really much different

01:14:47 than the ones we had quite a while ago.

01:14:50 They’re bigger and train more

01:14:51 and we have more labeled data and so on.

01:14:56 But I don’t think you can get to the kind of things

01:14:58 I know the brain can do and that we think about

01:15:01 as intelligence by just scaling it up.

01:15:03 So that may be, it’s a good description

01:15:06 of what’s happened in the past,

01:15:07 what’s happened recently with the reemergence

01:15:09 of artificial neural networks.

01:15:12 It may be a good prescription

01:15:14 for what’s gonna happen in the short term.

01:15:17 But I don’t think that’s the path.

01:15:19 I’ve said that earlier.

01:15:20 There’s an alternate path.

01:15:21 I should mention to you, by the way,

01:15:22 that we’ve made sufficient progress

01:15:25 on the whole cortical theory in the last few years

01:15:28 that last year we decided to start actively pursuing

01:15:35 how do we get these ideas embedded into machine learning?

01:15:40 Well, that’s, again, being led by my colleague,

01:15:41 Subed Tariman, and he’s more of a machine learning guy.

01:15:45 I’m more of a neuroscience guy.

01:15:46 So this is now, I wouldn’t say our focus,

01:15:51 but it is now an equal focus here

01:15:54 because we need to proselytize what we’ve learned

01:15:58 and we need to show how it’s beneficial

01:16:01 to the machine learning layer.

01:16:03 So we’re putting, we have a plan in place right now.

01:16:05 In fact, we just did our first paper on this.

01:16:07 I can tell you about that.

01:16:09 But one of the reasons I wanna talk to you

01:16:11 is because I’m trying to get more people

01:16:14 in the machine learning community to say,

01:16:15 I need to learn about this stuff.

01:16:17 And maybe we should just think about this a bit more

01:16:19 about what we’ve learned about the brain

01:16:20 and what are those team at Nimenta, what have they done?

01:16:23 Is that useful for us?

01:16:25 Yeah, so is there elements of all the cortical theory

01:16:28 that things we’ve been talking about

01:16:29 that may be useful in the short term?

01:16:31 Yes, in the short term, yes.

01:16:33 This is the, sorry to interrupt,

01:16:34 but the open question is,

01:16:37 it certainly feels from my perspective

01:16:39 that in the long term,

01:16:41 some of the ideas we’ve been talking about

01:16:42 will be extremely useful.

01:16:44 The question is whether in the short term.

01:16:46 Well, this is always what I would call

01:16:48 the entrepreneur’s dilemma.

01:16:50 So you have this long term vision,

01:16:53 oh, we’re gonna all be driving electric cars

01:16:55 or we’re all gonna have computers

01:16:56 or we’re all gonna, whatever.

01:16:59 And you’re at some point in time and you say,

01:17:01 I can see that long term vision,

01:17:02 I’m sure it’s gonna happen.

01:17:03 How do I get there without killing myself?

01:17:05 Without going out of business, right?

01:17:07 That’s the challenge.

01:17:08 That’s the dilemma.

01:17:09 That’s the really difficult thing to do.

01:17:11 So we’re facing that right now.

01:17:13 So ideally what you’d wanna do

01:17:14 is find some steps along the way

01:17:16 that you can get there incrementally.

01:17:17 You don’t have to like throw it all out

01:17:19 and start over again.

01:17:20 The first thing that we’ve done

01:17:22 is we focus on the sparse representations.

01:17:25 So just in case you don’t know what that means

01:17:28 or some of the listeners don’t know what that means,

01:17:31 in the brain, if I have like 10,000 neurons,

01:17:34 what you would see is maybe 2% of them active at a time.

01:17:36 You don’t see 50%, you don’t see 30%,

01:17:39 you might see 2%.

01:17:41 And it’s always like that.

01:17:42 For any set of sensory inputs?

01:17:44 It doesn’t matter if anything,

01:17:45 doesn’t matter any part of the brain.

01:17:47 But which neurons differs?

01:17:51 Which neurons are active?

01:17:52 Yeah, so let’s say I take 10,000 neurons

01:17:55 that are representing something.

01:17:56 They’re sitting there in a little block together.

01:17:57 It’s a teeny little block of neurons, 10,000 neurons.

01:18:00 And they’re representing a location,

01:18:01 they’re representing a cup,

01:18:02 they’re representing the input from my sensors.

01:18:04 I don’t know, it doesn’t matter.

01:18:05 It’s representing something.

01:18:07 The way the representations occur,

01:18:09 it’s always a sparse representation.

01:18:10 Meaning it’s a population code.

01:18:11 So which 200 cells are active tells me what’s going on.

01:18:14 It’s not, individual cells aren’t that important at all.

01:18:18 It’s the population code that matters.

01:18:20 And when you have sparse population codes,

01:18:23 then all kinds of beautiful properties come out of them.

01:18:26 So the brain uses sparse population codes.

01:18:28 We’ve written and described these benefits

01:18:30 in some of our papers.

01:18:32 So they give this tremendous robustness to the systems.

01:18:37 Brains are incredibly robust.

01:18:39 Neurons are dying all the time and spasming

01:18:41 and synapses are falling apart all the time.

01:18:43 And it keeps working.

01:18:45 So what Sibutai and Louise, one of our other engineers here

01:18:51 have done, have shown they’re introducing sparseness

01:18:55 into convolutional neural networks.

01:18:56 Now other people are thinking along these lines,

01:18:58 but we’re going about it in a more principled way, I think.

01:19:00 And we’re showing that if you enforce sparseness

01:19:04 throughout these convolutional neural networks

01:19:07 in both the act, which sort of,

01:19:09 which neurons are active and the connections between them,

01:19:12 that you get some very desirable properties.

01:19:15 So one of the current hot topics in deep learning right now

01:19:18 are these adversarial examples.

01:19:20 So, you know, you give me any deep learning network

01:19:23 and I can give you a picture that looks perfect

01:19:26 and you’re going to call it, you know,

01:19:27 you’re going to say the monkey is, you know, an airplane.

01:19:30 So that’s a problem.

01:19:32 And DARPA just announced some big thing.

01:19:34 They’re trying to, you know, have some contest for this.

01:19:36 But if you enforce sparse representations here,

01:19:40 many of these problems go away.

01:19:41 They’re much more robust and they’re not easy to fool.

01:19:44 So we’ve already shown some of those results,

01:19:48 just literally in January or February,

01:19:51 just like last month we did that.

01:19:53 And you can, I think it’s on bioRxiv right now,

01:19:57 or on iRxiv, you can read about it.

01:19:59 But, so that’s like a baby step, okay?

01:20:03 That’s taking something from the brain.

01:20:04 We know about sparseness.

01:20:05 We know why it’s important.

01:20:06 We know what it gives the brain.

01:20:08 So let’s try to enforce that onto this.

01:20:09 What’s your intuition why sparsity leads to robustness?

01:20:12 Because it feels like it would be less robust.

01:20:15 Why would you feel the rest robust to you?

01:20:17 So it just feels like if the fewer neurons are involved,

01:20:24 the more fragile the representation.

01:20:26 But I didn’t say there was lots of few neurons.

01:20:28 I said, let’s say 200.

01:20:29 That’s a lot.

01:20:31 There’s still a lot, it’s just.

01:20:32 So here’s an intuition for it.

01:20:35 This is a bit technical, so for engineers,

01:20:39 machine learning people, this will be easy,

01:20:41 but all the listeners, maybe not.

01:20:44 If you’re trying to classify something,

01:20:45 you’re trying to divide some very high dimensional space

01:20:48 into different pieces, A and B.

01:20:50 And you’re trying to create some point where you say,

01:20:52 all these points in this high dimensional space are A,

01:20:54 and all these points in this high dimensional space are B.

01:20:57 And if you have points that are close to that line,

01:21:01 it’s not very robust.

01:21:02 It works for all the points you know about,

01:21:04 but it’s not very robust,

01:21:07 because you can just move a little bit

01:21:08 and you’ve crossed over the line.

01:21:10 When you have sparse representations,

01:21:12 imagine I pick, I’m gonna pick 200 cells active

01:21:16 out of 10,000, okay?

01:21:19 So I have 200 cells active.

01:21:20 Now let’s say I pick randomly another,

01:21:22 a different representation, 200.

01:21:24 The overlap between those is gonna be very small,

01:21:26 just a few.

01:21:28 I can pick millions of samples randomly of 200 neurons,

01:21:32 and not one of them will overlap more than just a few.

01:21:36 So one way to think about it is,

01:21:39 if I wanna fool one of these representations

01:21:41 to look like one of those other representations,

01:21:43 I can’t move just one cell, or two cells,

01:21:45 or three cells, or four cells.

01:21:46 I have to move 100 cells.

01:21:49 And that makes them robust.

01:21:52 In terms of further, so you mentioned sparsity.

01:21:56 What would be the next thing?

01:21:57 Yeah.

01:21:58 Okay, so we have, we picked one.

01:22:00 We don’t know if it’s gonna work well yet.

01:22:02 So again, we’re trying to come up with incremental ways

01:22:04 to moving from brain theory to add pieces

01:22:07 to machine learning, current machine learning world,

01:22:10 and one step at a time.

01:22:12 So the next thing we’re gonna try to do

01:22:13 is sort of incorporate some of the ideas

01:22:15 of the thousand brains theory,

01:22:19 that you have many, many models that are voting.

01:22:22 Now that idea is not new.

01:22:23 There’s a mixture of models that’s been around

01:22:25 for a long time.

01:22:27 But the way the brain does it is a little different.

01:22:29 And the way it votes is different.

01:22:33 And the kind of way it represents uncertainty

01:22:36 is different.

01:22:37 So we’re just starting this work,

01:22:39 but we’re gonna try to see if we can sort of incorporate

01:22:42 some of the principles of voting,

01:22:43 or principles of the thousand brain theory.

01:22:45 Like lots of simple models that talk to each other

01:22:49 in a certain way.

01:22:53 And can we build more machines, systems that learn faster

01:22:57 and also, well mostly are multimodal

01:23:03 and robust to multimodal type of issues.

01:23:07 So one of the challenges there

01:23:09 is the machine learning computer vision community

01:23:13 has certain sets of benchmarks,

01:23:15 sets of tests based on which they compete.

01:23:18 And I would argue, especially from your perspective,

01:23:22 that those benchmarks aren’t that useful

01:23:24 for testing the aspects that the brain is good at,

01:23:28 or intelligence.

01:23:29 They’re not really testing intelligence.

01:23:31 They’re very fine.

01:23:32 And it’s been extremely useful

01:23:34 for developing specific mathematical models,

01:23:37 but it’s not useful in the long term

01:23:40 for creating intelligence.

01:23:41 So you think you also have a role in proposing

01:23:44 better tests?

01:23:47 Yeah, this is a very,

01:23:48 you’ve identified a very serious problem.

01:23:51 First of all, the tests that they have

01:23:53 are the tests that they want.

01:23:54 Not the tests of the other things

01:23:55 that we’re trying to do, right?

01:23:58 You know, what are the, so on.

01:24:01 The second thing is sometimes these,

01:24:04 to be competitive in these tests,

01:24:06 you have to have huge data sets and huge computing power.

01:24:10 And so, you know, and we don’t have that here.

01:24:13 We don’t have it as well as other big teams

01:24:15 that big companies do.

01:24:18 So there’s numerous issues there.

01:24:20 You know, we come out, you know,

01:24:22 where our approach to this is all based on,

01:24:24 in some sense, you might argue, elegance.

01:24:26 We’re coming at it from like a theoretical base

01:24:27 that we think, oh my God, this is so clearly elegant.

01:24:29 This is how brains work.

01:24:30 This is what intelligence is.

01:24:31 But the machine learning world has gotten in this phase

01:24:33 where they think it doesn’t matter.

01:24:35 Doesn’t matter what you think,

01:24:36 as long as you do, you know, 0.1% better on this benchmark,

01:24:39 that’s what, that’s all that matters.

01:24:40 And that’s a problem.

01:24:43 You know, we have to figure out how to get around that.

01:24:46 That’s a challenge for us.

01:24:47 That’s one of the challenges that we have to deal with.

01:24:50 So I agree, you’ve identified a big issue.

01:24:52 It’s difficult for those reasons.

01:24:55 But you know, part of the reasons I’m talking to you here

01:24:59 today is I hope I’m gonna get some machine learning people

01:25:01 to say, I’m gonna read those papers.

01:25:03 Those might be some interesting ideas.

01:25:04 I’m tired of doing this 0.1% improvement stuff, you know?

01:25:08 Well, that’s why I’m here as well,

01:25:10 because I think machine learning now as a community

01:25:13 is at a place where the next step needs to be orthogonal

01:25:18 to what has received success in the past.

01:25:21 Well, you see other leaders saying this,

01:25:23 machine learning leaders, you know,

01:25:25 Jeff Hinton with his capsules idea.

01:25:27 Many people have gotten up to say, you know,

01:25:29 we’re gonna hit road map, maybe we should look at the brain,

01:25:32 you know, things like that.

01:25:33 So hopefully that thinking will occur organically.

01:25:38 And then we’re in a nice position for people to come

01:25:40 and look at our work and say,

01:25:41 well, what can we learn from these guys?

01:25:43 Yeah, MIT is launching a billion dollar computing college

01:25:47 that’s centered around this idea, so.

01:25:49 Is it on this idea of what?

01:25:50 Well, the idea that, you know,

01:25:52 the humanities, psychology, and neuroscience

01:25:54 have to work all together to get to build the S.

01:25:58 Yeah, I mean, Stanford just did

01:26:00 this Human Centered AI Center.

01:26:02 I’m a little disappointed in these initiatives

01:26:04 because, you know, they’re focusing

01:26:08 on sort of the human side of it,

01:26:09 and it could very easily slip into

01:26:12 how humans interact with intelligent machines,

01:26:16 which is nothing wrong with that,

01:26:17 but that’s not, that is orthogonal

01:26:19 to what we’re trying to do.

01:26:20 We’re trying to say, like,

01:26:21 what is the essence of intelligence?

01:26:22 I don’t care.

01:26:23 In fact, I wanna build intelligent machines

01:26:25 that aren’t emotional, that don’t smile at you,

01:26:28 that, you know, that aren’t trying to tuck you in at night.

01:26:31 Yeah, there is that pattern that you,

01:26:34 when you talk about understanding humans

01:26:36 is important for understanding intelligence,

01:26:38 that you start slipping into topics of ethics

01:26:41 or, yeah, like you said,

01:26:43 the interactive elements as opposed to,

01:26:45 no, no, no, we have to zoom in on the brain,

01:26:47 study what the human brain, the baby, the…

01:26:51 Let’s study what a brain does.

01:26:52 Does.

01:26:53 And then we can decide which parts of that

01:26:54 we wanna recreate in some system,

01:26:57 but until you have that theory about what the brain does,

01:26:59 what’s the point, you know, it’s just,

01:27:01 you’re gonna be wasting time, I think.

01:27:02 Right, just to break it down

01:27:04 on the artificial neural network side,

01:27:05 maybe you could speak to this

01:27:06 on the biological neural network side,

01:27:09 the process of learning versus the process of inference.

01:27:13 Maybe you can explain to me,

01:27:15 is there a difference between,

01:27:18 you know, in artificial neural networks,

01:27:19 there’s a difference between the learning stage

01:27:21 and the inference stage.

01:27:22 Do you see the brain as something different?

01:27:24 One of the big distinctions that people often say,

01:27:29 I don’t know how correct it is,

01:27:30 is artificial neural networks need a lot of data.

01:27:32 They’re very inefficient learning.

01:27:34 Do you see that as a correct distinction

01:27:37 from the biology of the human brain,

01:27:40 that the human brain is very efficient,

01:27:41 or is that just something we deceive ourselves?

01:27:44 No, it is efficient, obviously.

01:27:45 We can learn new things almost instantly.

01:27:47 And so what elements do you think are useful?

01:27:50 Yeah, I can talk about that.

01:27:50 You brought up two issues there.

01:27:52 So remember I talked early about the constraints

01:27:54 we always feel, well, one of those constraints

01:27:57 is the fact that brains are continually learning.

01:28:00 That’s not something we said, oh, we can add that later.

01:28:03 That’s something that was upfront,

01:28:05 had to be there from the start,

01:28:08 made our problems harder.

01:28:11 But we showed, going back to the 2016 paper

01:28:14 on sequence memory, we showed how that happens,

01:28:16 how the brains infer and learn at the same time.

01:28:19 And our models do that.

01:28:21 And they’re not two separate phases,

01:28:24 or two separate sets of time.

01:28:26 I think that’s a big, big problem in AI,

01:28:29 at least for many applications, not for all.

01:28:33 So I can talk about that.

01:28:34 There are some, it gets detailed,

01:28:37 there are some parts of the neocortex in the brain

01:28:39 where actually what’s going on,

01:28:41 there’s these cycles of activity in the brain.

01:28:46 And there’s very strong evidence

01:28:49 that you’re doing more of inference

01:28:51 on one part of the phase,

01:28:52 and more of learning on the other part of the phase.

01:28:54 So the brain can actually sort of separate

01:28:55 different populations of cells

01:28:56 or going back and forth like this.

01:28:58 But in general, I would say that’s an important problem.

01:29:01 We have all of our networks that we’ve come up with do both.

01:29:05 And they’re continuous learning networks.

01:29:08 And you mentioned benchmarks earlier.

01:29:10 Well, there are no benchmarks about that.

01:29:12 So we have to, we get in our little soapbox,

01:29:17 and hey, by the way, this is important,

01:29:19 and here’s a mechanism for doing that.

01:29:20 But until you can prove it to someone

01:29:23 in some commercial system or something, it’s a little harder.

01:29:26 So yeah, one of the things I had to linger on that

01:29:28 is in some ways to learn the concept of a coffee cup,

01:29:33 you only need this one coffee cup

01:29:35 and maybe some time alone in a room with it.

01:29:37 Well, the first thing is,

01:29:39 imagine I reach my hand into a black box

01:29:41 and I’m reaching, I’m trying to touch something.

01:29:43 I don’t know upfront if it’s something I already know

01:29:46 or if it’s a new thing.

01:29:47 And I have to, I’m doing both at the same time.

01:29:50 I don’t say, oh, let’s see if it’s a new thing.

01:29:53 Oh, let’s see if it’s an old thing.

01:29:54 I don’t do that.

01:29:55 As I go, my brain says, oh, it’s new or it’s not new.

01:29:59 And if it’s new, I start learning what it is.

01:30:02 And by the way, it starts learning from the get go,

01:30:04 even if it’s gonna recognize it.

01:30:06 So they’re not separate problems.

01:30:08 And so that’s the thing there.

01:30:10 The other thing you mentioned was the fast learning.

01:30:13 So I was just talking about continuous learning,

01:30:15 but there’s also fast learning.

01:30:16 Literally, I can show you this coffee cup

01:30:18 and I say, here’s a new coffee cup.

01:30:20 It’s got the logo on it.

01:30:21 Take a look at it, done, you’re done.

01:30:23 You can predict what it’s gonna look like,

01:30:25 you know, in different positions.

01:30:27 So I can talk about that too.

01:30:29 In the brain, the way learning occurs,

01:30:34 I mentioned this earlier, but I’ll mention it again.

01:30:35 The way learning occurs,

01:30:36 imagine I am a section of a dendrite of a neuron,

01:30:40 and I’m gonna learn something new.

01:30:43 Doesn’t matter what it is.

01:30:44 I’m just gonna learn something new.

01:30:46 I need to recognize a new pattern.

01:30:48 So what I’m gonna do is I’m gonna form new synapses.

01:30:52 New synapses, we’re gonna rewire the brain

01:30:55 onto that section of the dendrite.

01:30:57 Once I’ve done that, everything else that neuron has learned

01:31:01 is not affected by it.

01:31:02 That’s because it’s isolated

01:31:04 to that small section of the dendrite.

01:31:06 They’re not all being added together, like a point neuron.

01:31:09 So if I learn something new on this segment here,

01:31:11 it doesn’t change any of the learning

01:31:13 that occur anywhere else in that neuron.

01:31:14 So I can add something without affecting previous learning.

01:31:18 And I can do it quickly.

01:31:20 Now let’s talk, we can talk about the quickness,

01:31:22 how it’s done in real neurons.

01:31:24 You might say, well, doesn’t it take time to form synapses?

01:31:26 Yes, it can take maybe an hour to form a new synapse.

01:31:30 We can form memories quicker than that,

01:31:32 and I can explain that how it happens too, if you want.

01:31:35 But it’s getting a bit neurosciencey.

01:31:39 That’s great, but is there an understanding

01:31:41 of these mechanisms at every level?

01:31:43 Yeah.

01:31:43 So from the short term memories and the forming.

01:31:48 So this idea of synaptogenesis, the growth of new synapses,

01:31:51 that’s well described, it’s well understood.

01:31:54 And that’s an essential part of learning.

01:31:55 That is learning.

01:31:56 That is learning.

01:31:58 Okay.

01:32:01 Going back many, many years,

01:32:03 people, you know, it was, what’s his name,

01:32:06 the psychologist who proposed, Hebb, Donald Hebb.

01:32:09 He proposed that learning was the modification

01:32:12 of the strength of a connection between two neurons.

01:32:15 People interpreted that as the modification

01:32:18 of the strength of a synapse.

01:32:19 He didn’t say that.

01:32:20 He just said there’s a modification

01:32:22 between the effect of one neuron and another.

01:32:24 So synaptogenesis is totally consistent

01:32:26 with what Donald Hebb said.

01:32:28 But anyway, there’s these mechanisms,

01:32:29 the growth of new synapses.

01:32:30 You can go online, you can watch a video

01:32:32 of a synapse growing in real time.

01:32:33 It’s literally, you can see this little thing going boop.

01:32:37 It’s pretty impressive.

01:32:38 So those mechanisms are known.

01:32:39 Now there’s another thing that we’ve speculated

01:32:42 and we’ve written about,

01:32:43 which is consistent with known neuroscience,

01:32:45 but it’s less proven.

01:32:48 And this is the idea, how do I form a memory

01:32:50 really, really quickly?

01:32:51 Like instantaneous.

01:32:52 If it takes an hour to grow a synapse,

01:32:54 like that’s not instantaneous.

01:32:56 So there are types of synapses called silent synapses.

01:33:01 They look like a synapse, but they don’t do anything.

01:33:04 They’re just sitting there.

01:33:04 It’s like if an action potential comes in,

01:33:07 it doesn’t release any neurotransmitter.

01:33:10 Some parts of the brain have more of these than others.

01:33:12 For example, the hippocampus has a lot of them,

01:33:14 which is where we associate most short term memory with.

01:33:18 So what we speculated, again, in that 2016 paper,

01:33:22 we proposed that the way we form very quick memories,

01:33:26 very short term memories, or quick memories,

01:33:28 is that we convert silent synapses into active synapses.

01:33:33 It’s like saying a synapse has a zero weight

01:33:36 and a one weight,

01:33:37 but the longterm memory has to be formed by synaptogenesis.

01:33:41 So you can remember something really quickly

01:33:43 by just flipping a bunch of these guys from silent to active.

01:33:46 It’s not from 0.1 to 0.15.

01:33:49 It’s like, it doesn’t do anything

01:33:50 till it releases transmitter.

01:33:52 And if I do that over a bunch of these,

01:33:53 I’ve got a very quick short term memory.

01:33:56 So I guess the lesson behind this

01:33:58 is that most neural networks today are fully connected.

01:34:01 Every neuron connects every other neuron

01:34:03 from layer to layer.

01:34:04 That’s not correct in the brain.

01:34:06 We don’t want that.

01:34:06 We actually don’t want that.

01:34:08 It’s bad.

01:34:09 You want a very sparse connectivity

01:34:10 so that any neuron connects to some subset of the neurons

01:34:14 in the other layer.

01:34:15 And it does so on a dendrite by dendrite segment basis.

01:34:18 So it’s a very some parcelated out type of thing.

01:34:21 And that then learning is not adjusting all these weights,

01:34:25 but learning is just saying,

01:34:26 okay, connect to these 10 cells here right now.

01:34:30 In that process, you know, with artificial neural networks,

01:34:32 it’s a very simple process of backpropagation

01:34:36 that adjusts the weights.

01:34:37 The process of synaptogenesis.

01:34:40 Synaptogenesis.

01:34:40 Synaptogenesis.

01:34:42 It’s even easier.

01:34:43 It’s even easier.

01:34:43 It’s even easier.

01:34:44 Backpropagation requires something

01:34:47 that really can’t happen in brains.

01:34:48 This backpropagation of this error signal,

01:34:51 that really can’t happen.

01:34:52 People are trying to make it happen in brains,

01:34:53 but it doesn’t happen in brains.

01:34:54 This is pure Hebbian learning.

01:34:56 Well, synaptogenesis is pure Hebbian learning.

01:34:58 It’s basically saying,

01:35:00 there’s a population of cells over here

01:35:01 that are active right now.

01:35:03 And there’s a population of cells over here

01:35:04 active right now.

01:35:05 How do I form connections between those active cells?

01:35:07 And it’s literally saying this guy became active.

01:35:11 These 100 neurons here became active

01:35:13 before this neuron became active.

01:35:15 So form connections to those ones.

01:35:17 That’s it.

01:35:17 There’s no propagation of error, nothing.

01:35:19 All the networks we do,

01:35:20 all the models we have work on almost completely on

01:35:25 Hebbian learning,

01:35:26 but on dendritic segments

01:35:30 and multiple synapses at the same time.

01:35:33 So now let’s sort of turn the question

01:35:34 that you already answered,

01:35:35 and maybe you can answer it again.

01:35:38 If you look at the history of artificial intelligence,

01:35:41 where do you think we stand?

01:35:43 How far are we from solving intelligence?

01:35:45 You said you were very optimistic.

01:35:47 Can you elaborate on that?

01:35:48 Yeah, it’s always the crazy question to ask

01:35:53 because no one can predict the future.

01:35:55 Absolutely.

01:35:55 So I’ll tell you a story.

01:35:58 I used to run a different neuroscience institute

01:36:01 called the Redwood Neuroscience Institute,

01:36:02 and we would hold these symposiums

01:36:04 and we’d get like 35 scientists

01:36:06 from around the world to come together.

01:36:08 And I used to ask them all the same question.

01:36:10 I would say, well, how long do you think it’ll be

01:36:11 before we understand how the neocortex works?

01:36:14 And everyone went around the room

01:36:15 and they had introduced the name

01:36:16 and they have to answer that question.

01:36:18 So I got, the typical answer was 50 to 100 years.

01:36:22 Some people would say 500 years.

01:36:24 Some people said never.

01:36:25 I said, why are you a neuroscientist?

01:36:27 It’s never gonna, it’s a good pay.

01:36:32 It’s interesting.

01:36:34 So, you know, but it doesn’t work like that.

01:36:36 As I mentioned earlier, these are not,

01:36:38 these are step functions.

01:36:39 Things happen and then bingo, they happen.

01:36:41 You can’t predict that.

01:36:43 I feel I’ve already passed a step function.

01:36:45 So if I can do my job correctly over the next five years,

01:36:50 then, meaning I can proselytize these ideas.

01:36:53 I can convince other people they’re right.

01:36:56 We can show that other people,

01:36:58 machine learning people should pay attention

01:37:00 to these ideas.

01:37:01 Then we’re definitely in an under 20 year timeframe.

01:37:04 If I can do those things, if I’m not successful in that,

01:37:07 and this is the last time anyone talks to me

01:37:09 and no one reads our papers and you know,

01:37:12 and I’m wrong or something like that,

01:37:13 then I don’t know.

01:37:15 But it’s not 50 years.

01:37:21 Think about electric cars.

01:37:22 How quickly are they gonna populate the world?

01:37:24 It probably takes about a 20 year span.

01:37:27 It’ll be something like that.

01:37:28 But I think if I can do what I said, we’re starting it.

01:37:31 And of course there could be other,

01:37:34 you said step functions.

01:37:35 It could be everybody gives up on your ideas for 20 years

01:37:40 and then all of a sudden somebody picks it up again.

01:37:42 Wait, that guy was onto something.

01:37:43 Yeah, so that would be a failure on my part, right?

01:37:47 Think about Charles Babbage.

01:37:49 Charles Babbage, he’s the guy who invented the computer

01:37:52 back in the 18 something, 1800s.

01:37:55 And everyone forgot about it until 100 years later.

01:37:59 And say, hey, this guy figured this stuff out

01:38:00 a long time ago.

01:38:02 But he was ahead of his time.

01:38:03 I don’t think, as I said,

01:38:06 I recognize this is part of any entrepreneur’s challenge.

01:38:09 I use entrepreneur broadly in this case.

01:38:11 I’m not meaning like I’m building a business

01:38:12 or trying to sell something.

01:38:13 I mean, I’m trying to sell ideas.

01:38:15 And this is the challenge as to how you get people

01:38:19 to pay attention to you, how do you get them

01:38:21 to give you positive or negative feedback,

01:38:24 how do you get the people to act differently

01:38:25 based on your ideas.

01:38:27 So we’ll see how well we do on that.

01:38:30 So you know that there’s a lot of hype

01:38:32 behind artificial intelligence currently.

01:38:34 Do you, as you look to spread the ideas

01:38:39 that are of neocortical theory, the things you’re working on,

01:38:43 do you think there’s some possibility

01:38:45 we’ll hit an AI winter once again?

01:38:47 Yeah, it’s certainly a possibility.

01:38:48 No question about it.

01:38:49 Is that something you worry about?

01:38:50 Yeah, well, I guess, do I worry about it?

01:38:54 I haven’t decided yet if that’s good or bad for my mission.

01:38:57 That’s true, that’s very true.

01:38:59 Because it’s almost like you need the winter

01:39:02 to refresh the palette.

01:39:04 Yeah, it’s like, I want, here’s what you wanna have it is.

01:39:07 You want, like to the extent that everyone is so thrilled

01:39:12 about the current state of machine learning and AI

01:39:15 and they don’t imagine they need anything else,

01:39:18 it makes my job harder.

01:39:19 If everything crashed completely

01:39:22 and every student left the field

01:39:24 and there was no money for anybody to do anything

01:39:26 and it became an embarrassment

01:39:27 to talk about machine intelligence and AI,

01:39:29 that wouldn’t be good for us either.

01:39:30 You want sort of the soft landing approach, right?

01:39:33 You want enough people, the senior people in AI

01:39:36 and machine learning to say, you know,

01:39:37 we need other approaches.

01:39:38 We really need other approaches.

01:39:40 Damn, we need other approaches.

01:39:42 Maybe we should look to the brain.

01:39:43 Okay, let’s look to the brain.

01:39:44 Who’s got some brain ideas?

01:39:45 Okay, let’s start a little project on the side here

01:39:47 trying to do brain idea related stuff.

01:39:49 That’s the ideal outcome we would want.

01:39:51 So I don’t want a total winter

01:39:53 and yet I don’t want it to be sunny all the time either.

01:39:57 So what do you think it takes to build a system

01:40:00 with human level intelligence

01:40:03 where once demonstrated you would be very impressed?

01:40:06 So does it have to have a body?

01:40:08 Does it have to have the C word we used before,

01:40:12 consciousness as an entirety in a holistic sense?

01:40:19 First of all, I don’t think the goal

01:40:20 is to create a machine that is human level intelligence.

01:40:23 I think it’s a false goal.

01:40:24 Back to Turing, I think it was a false statement.

01:40:27 We want to understand what intelligence is

01:40:29 and then we can build intelligent machines

01:40:30 of all different scales, all different capabilities.

01:40:34 A dog is intelligent.

01:40:35 I don’t need, that’d be pretty good to have a dog.

01:40:38 But what about something that doesn’t look

01:40:39 like an animal at all, in different spaces?

01:40:41 So my thinking about this is that

01:40:44 we want to define what intelligence is,

01:40:46 agree upon what makes an intelligent system.

01:40:48 We can then say, okay, we’re now gonna build systems

01:40:51 that work on those principles or some subset of them

01:40:54 and we can apply them to all different types of problems.

01:40:57 And the kind, the idea, it’s not computing.

01:41:00 We don’t ask, if I take a little one chip computer,

01:41:05 I don’t say, well, that’s not a computer

01:41:06 because it’s not as powerful as this big server over here.

01:41:09 No, no, because we know that what the principles

01:41:11 of computing are and I can apply those principles

01:41:12 to a small problem or into a big problem.

01:41:14 And same, intelligence needs to get there.

01:41:16 We have to say, these are the principles.

01:41:17 I can make a small one, a big one.

01:41:19 I can make them distributed.

01:41:19 I can put them on different sensors.

01:41:21 They don’t have to be human like at all.

01:41:23 Now, you did bring up a very interesting question

01:41:24 about embodiment.

01:41:25 Does it have to have a body?

01:41:27 It has to have some concept of movement.

01:41:30 It has to be able to move through these reference frames

01:41:33 I talked about earlier.

01:41:34 Whether it’s physically moving,

01:41:35 like I need, if I’m gonna have an AI

01:41:37 that understands coffee cups,

01:41:38 it’s gonna have to pick up the coffee cup

01:41:40 and touch it and look at it with its eyes and hands

01:41:43 or something equivalent to that.

01:41:45 If I have a mathematical AI,

01:41:48 maybe it needs to move through mathematical spaces.

01:41:51 I could have a virtual AI that lives in the internet

01:41:55 and its movements are traversing links

01:41:58 and digging into files,

01:42:00 but it’s got a location that it’s traveling

01:42:03 through some space.

01:42:04 You can’t have an AI that just take some flash thing input.

01:42:09 We call it flash inference.

01:42:10 Here’s a pattern, done.

01:42:12 No, it’s movement pattern, movement pattern,

01:42:15 movement pattern, attention, digging, building structure,

01:42:19 figuring out the model of the world.

01:42:20 So some sort of embodiment,

01:42:22 whether it’s physical or not, has to be part of it.

01:42:25 So self awareness and the way to be able to answer

01:42:28 where am I?

01:42:28 Well, you’re bringing up self,

01:42:29 that’s a different topic, self awareness.

01:42:31 No, the very narrow definition of self,

01:42:33 meaning knowing a sense of self enough to know

01:42:37 where am I in the space where it’s actually.

01:42:39 Yeah, basically the system needs to know its location

01:42:43 or each component of the system needs to know

01:42:46 where it is in the world at that point in time.

01:42:48 So self awareness and consciousness.

01:42:51 Do you think one, from the perspective of neuroscience

01:42:55 and neurocortex, these are interesting topics,

01:42:58 solvable topics.

01:42:59 Do you have any ideas of why the heck it is

01:43:02 that we have a subjective experience at all?

01:43:04 Yeah, I have a lot of thoughts on that.

01:43:05 And is it useful or is it just a side effect of us?

01:43:08 It’s interesting to think about.

01:43:10 I don’t think it’s useful as a means to figure out

01:43:13 how to build intelligent machines.

01:43:16 It’s something that systems do

01:43:20 and we can talk about what it is that are like,

01:43:22 well, if I build a system like this,

01:43:23 then it would be self aware.

01:43:25 Or if I build it like this, it wouldn’t be self aware.

01:43:28 So that’s a choice I can have.

01:43:30 It’s not like, oh my God, it’s self aware.

01:43:32 I can’t turn, I heard an interview recently

01:43:35 with this philosopher from Yale,

01:43:37 I can’t remember his name, I apologize for that.

01:43:39 But he was talking about,

01:43:39 well, if these computers are self aware,

01:43:41 then it would be a crime to unplug them.

01:43:42 And I’m like, oh, come on, that’s not,

01:43:45 I unplug myself every night, I go to sleep.

01:43:47 Is that a crime?

01:43:48 I plug myself in again in the morning and there I am.

01:43:51 So people get kind of bent out of shape about this.

01:43:56 I have very definite, very detailed understanding

01:43:59 or opinions about what it means to be conscious

01:44:02 and what it means to be self aware.

01:44:04 I don’t think it’s that interesting a problem.

01:44:06 You’ve talked to Christoph Koch.

01:44:08 He thinks that’s the only problem.

01:44:10 I didn’t actually listen to your interview with him,

01:44:12 but I know him and I know that’s the thing he cares about.

01:44:15 He also thinks intelligence and consciousness are disjoint.

01:44:18 So I mean, it’s not, you don’t have to have one or the other.

01:44:21 So he is.

01:44:21 I disagree with that.

01:44:22 I just totally disagree with that.

01:44:24 So where’s your thoughts and consciousness,

01:44:26 where does it emerge from?

01:44:27 Because it is.

01:44:28 So then we have to break it down to the two parts, okay?

01:44:30 Because consciousness isn’t one thing.

01:44:32 That’s part of the problem with that term

01:44:33 is it means different things to different people

01:44:35 and there’s different components of it.

01:44:37 There is a concept of self awareness, okay?

01:44:40 That can be very easily explained.

01:44:43 You have a model of your own body.

01:44:46 The neocortex models things in the world

01:44:48 and it also models your own body.

01:44:50 And then it has a memory.

01:44:53 It can remember what you’ve done, okay?

01:44:55 So it can remember what you did this morning,

01:44:57 can remember what you had for breakfast and so on.

01:44:59 And so I can say to you, okay, Lex,

01:45:03 were you conscious this morning when you had your bagel?

01:45:06 And you’d say, yes, I was conscious.

01:45:08 Now what if I could take your brain

01:45:10 and revert all the synapses back

01:45:12 to the state they were this morning?

01:45:14 And then I said to you, Lex,

01:45:15 were you conscious when you ate the bagel?

01:45:17 And you said, no, I wasn’t conscious.

01:45:18 I said, here’s a video of eating the bagel.

01:45:19 And you said, I wasn’t there.

01:45:22 That’s not possible

01:45:23 because I must’ve been unconscious at that time.

01:45:25 So we can just make this one to one correlation

01:45:27 between memory of your body’s trajectory through the world

01:45:31 over some period of time,

01:45:32 a memory and the ability to recall that memory

01:45:34 is what you would call conscious.

01:45:35 I was conscious of that, it’s a self awareness.

01:45:38 And any system that can recall,

01:45:41 memorize what it’s done recently

01:45:43 and bring that back and invoke it again

01:45:46 would say, yeah, I’m aware.

01:45:48 I remember what I did.

01:45:49 All right, I got it.

01:45:51 That’s an easy one.

01:45:52 Although some people think that’s a hard one.

01:45:54 The more challenging part of consciousness

01:45:57 is this one that’s sometimes used

01:45:59 going by the word of qualia,

01:46:01 which is, why does an object seem red?

01:46:04 Or what is pain?

01:46:06 And why does pain feel like something?

01:46:08 Why do I feel redness?

01:46:10 Or why do I feel painness?

01:46:12 And then I could say, well,

01:46:13 why does sight seems different than hearing?

01:46:15 It’s the same problem.

01:46:16 It’s really, these are all just neurons.

01:46:18 And so how is it that,

01:46:20 why does looking at you feel different than hearing you?

01:46:24 It feels different, but there’s just neurons in my head.

01:46:26 They’re all doing the same thing.

01:46:27 So that’s an interesting question.

01:46:29 The best treatise I’ve read about this

01:46:31 is by a guy named Oregon.

01:46:33 He wrote a book called,

01:46:35 Why Red Doesn’t Sound Like a Bell.

01:46:37 It’s a little, it’s not a trade book, easy to read,

01:46:42 but it, and it’s an interesting question.

01:46:46 Take something like color.

01:46:47 Color really doesn’t exist in the world.

01:46:49 It’s not a property of the world.

01:46:51 Property of the world that exists is light frequency.

01:46:54 And that gets turned into,

01:46:55 we have certain cells in the retina

01:46:57 that respond to different frequencies

01:46:59 different than others.

01:47:00 And so when they enter the brain,

01:47:01 you just have a bunch of axons

01:47:02 that are firing at different rates.

01:47:04 And from that, we perceive color.

01:47:06 But there is no color in the brain.

01:47:07 I mean, there’s no color coming in on those synapses.

01:47:10 It’s just a correlation between some axons

01:47:14 and some property of frequency.

01:47:17 And that isn’t even color itself.

01:47:18 Frequency doesn’t have a color.

01:47:20 It’s just what it is.

01:47:22 So then the question is,

01:47:24 well, why does it even appear to have a color at all?

01:47:27 Just as you’re describing it,

01:47:29 there seems to be a connection to those ideas

01:47:31 of reference frames.

01:47:32 I mean, it just feels like consciousness

01:47:37 having the subject,

01:47:38 assigning the feeling of red to the actual color

01:47:42 or to the wavelength is useful for intelligence.

01:47:47 Yeah, I think that’s a good way of putting it.

01:47:49 It’s useful as a predictive mechanism

01:47:51 or useful as a generalization idea.

01:47:53 It’s a way of grouping things together to say,

01:47:55 it’s useful to have a model like this.

01:47:57 So think about the well known syndrome

01:48:02 that people who’ve lost a limb experience

01:48:04 called phantom limbs.

01:48:06 And what they claim is they can have their arm is removed,

01:48:12 but they feel their arm.

01:48:13 That not only feel it, they know it’s there.

01:48:15 It’s there, I know it’s there.

01:48:17 They’ll swear to you that it’s there.

01:48:19 And then they can feel pain in their arm

01:48:20 and they’ll feel pain in their finger.

01:48:21 And if they move their non existent arm behind their back,

01:48:25 then they feel the pain behind their back.

01:48:27 So this whole idea that your arm exists

01:48:30 is a model of your brain.

01:48:31 It may or may not really exist.

01:48:34 And just like, but it’s useful to have a model of something

01:48:38 that sort of correlates to things in the world.

01:48:40 So you can make predictions about what would happen

01:48:41 when those things occur.

01:48:43 It’s a little bit of a fuzzy,

01:48:44 but I think you’re getting quite towards the answer there.

01:48:46 It’s useful for the model to express things certain ways

01:48:51 that we can then map them into these reference frames

01:48:53 and make predictions about them.

01:48:55 I need to spend more time on this topic.

01:48:57 It doesn’t bother me.

01:48:58 Do you really need to spend more time?

01:49:00 Yeah, I know.

01:49:01 It does feel special that we have subjective experience,

01:49:04 but I’m yet to know why.

01:49:07 I’m just personally curious.

01:49:09 It’s not necessary for the work we’re doing here.

01:49:11 I don’t think I need to solve that problem

01:49:13 to build intelligent machines at all, not at all.

01:49:15 But there is sort of the silly notion

01:49:17 that you described briefly

01:49:20 that doesn’t seem so silly to us humans is,

01:49:23 if you’re successful building intelligent machines,

01:49:27 it feels wrong to then turn them off.

01:49:30 Because if you’re able to build a lot of them,

01:49:33 it feels wrong to then be able to turn off the…

01:49:38 Well, why?

01:49:39 Let’s break that down a bit.

01:49:41 As humans, why do we fear death?

01:49:43 There’s two reasons we fear death.

01:49:47 Well, first of all, I’ll say,

01:49:47 when you’re dead, it doesn’t matter at all.

01:49:48 Who cares?

01:49:49 You’re dead.

01:49:50 So why do we fear death?

01:49:51 We fear death for two reasons.

01:49:53 One is because we are programmed genetically to fear death.

01:49:57 That’s a survival and pop beginning of the genes thing.

01:50:02 And we also are programmed to feel sad

01:50:05 when people we know die.

01:50:06 We don’t feel sad for someone we don’t know dies.

01:50:08 There’s people dying right now,

01:50:09 they’re only just gonna say,

01:50:10 I don’t feel bad about them,

01:50:11 because I don’t know them.

01:50:12 But if I knew them, I’d feel really bad.

01:50:13 So again, these are old brain,

01:50:16 genetically embedded things that we fear death.

01:50:19 It’s outside of those uncomfortable feelings.

01:50:24 There’s nothing else to worry about.

01:50:25 Well, wait, hold on a second.

01:50:27 Do you know the denial of death by Becker?

01:50:30 No.

01:50:31 There’s a thought that death is,

01:50:36 our whole conception of our world model

01:50:41 kind of assumes immortality.

01:50:43 And then death is this terror that underlies it all.

01:50:47 So like…

01:50:47 Some people’s world model, not mine.

01:50:50 But, okay, so what Becker would say

01:50:52 is that you’re just living in an illusion.

01:50:54 You’ve constructed an illusion for yourself

01:50:56 because it’s such a terrible terror,

01:50:59 the fact that this…

01:51:00 What’s the illusion?

01:51:01 The illusion that death doesn’t matter.

01:51:02 You’re still not coming to grips with…

01:51:04 The illusion of what?

01:51:05 That death is…

01:51:07 Going to happen.

01:51:08 Oh, like it’s not gonna happen?

01:51:10 You’re actually operating.

01:51:11 You haven’t, even though you said you’ve accepted it,

01:51:14 you haven’t really accepted the notion that you’re gonna die

01:51:16 is what you say.

01:51:16 So it sounds like you disagree with that notion.

01:51:21 Yeah, yeah, totally.

01:51:22 I literally, every night I go to bed, it’s like dying.

01:51:28 Like little deaths.

01:51:28 It’s little deaths.

01:51:29 And if I didn’t wake up, it wouldn’t matter to me.

01:51:32 Only if I knew that was gonna happen would it be bothersome.

01:51:35 If I didn’t know it was gonna happen, how would I know?

01:51:37 Then I would worry about my wife.

01:51:39 So imagine I was a loner and I lived in Alaska

01:51:43 and I lived out there and there was no animals.

01:51:45 Nobody knew I existed.

01:51:46 I was just eating these roots all the time.

01:51:48 And nobody knew I was there.

01:51:51 And one day I didn’t wake up.

01:51:54 What pain in the world would there exist?

01:51:57 Well, so most people that think about this problem

01:51:59 would say that you’re just deeply enlightened

01:52:01 or are completely delusional.

01:52:04 One of the two.

01:52:05 But I would say that’s a very enlightened way

01:52:10 to see the world.

01:52:13 That’s the rational one as well.

01:52:14 It’s rational, that’s right.

01:52:15 But the fact is we don’t,

01:52:19 I mean, we really don’t have an understanding

01:52:22 of why the heck it is we’re born and why we die

01:52:24 and what happens after we die.

01:52:25 Well, maybe there isn’t a reason, maybe there is.

01:52:27 So I’m interested in those big problems too, right?

01:52:30 You interviewed Max Tegmark,

01:52:32 and there’s people like that, right?

01:52:33 I’m interested in those big problems as well.

01:52:35 And in fact, when I was young,

01:52:38 I made a list of the biggest problems I could think of.

01:52:41 First, why does anything exist?

01:52:43 Second, why do we have the laws of physics that we have?

01:52:46 Third, is life inevitable?

01:52:49 And why is it here?

01:52:50 Fourth, is intelligence inevitable?

01:52:52 And why is it here?

01:52:53 I stopped there because I figured

01:52:55 if you can make a truly intelligent system,

01:52:57 that will be the quickest way

01:52:59 to answer the first three questions.

01:53:01 I’m serious.

01:53:04 And so I said, my mission, you asked me earlier,

01:53:07 my first mission is to understand the brain,

01:53:09 but I felt that is the shortest way

01:53:10 to get to true machine intelligence.

01:53:12 And I wanna get to true machine intelligence

01:53:13 because even if it doesn’t occur in my lifetime,

01:53:15 other people will benefit from it

01:53:17 because I think it’ll occur in my lifetime,

01:53:19 but 20 years, you never know.

01:53:23 But that will be the quickest way for us to,

01:53:26 we can make super mathematicians,

01:53:27 we can make super space explorers,

01:53:29 we can make super physicist brains that do these things

01:53:34 and that can run experiments that we can’t run.

01:53:37 We don’t have the abilities to manipulate things and so on,

01:53:40 but we can build intelligent machines that do all those things

01:53:42 with the ultimate goal of finding out the answers

01:53:46 to the other questions.

01:53:48 Let me ask you another depressing and difficult question,

01:53:51 which is once we achieve that goal of creating,

01:53:57 no, of understanding intelligence,

01:54:01 do you think we would be happier,

01:54:02 more fulfilled as a species?

01:54:04 The understanding intelligence

01:54:05 or understanding the answers to the big questions?

01:54:07 Understanding intelligence.

01:54:08 Oh, totally, totally.

01:54:11 It would be far more fun place to live.

01:54:13 You think so?

01:54:14 Oh yeah, why not?

01:54:15 I mean, just put aside this terminator nonsense

01:54:19 and just think about, you can think about,

01:54:24 we can talk about the risks of AI if you want.

01:54:26 I’d love to, so let’s talk about.

01:54:28 But I think the world would be far better knowing things.

01:54:30 We’re always better than know things.

01:54:32 Do you think it’s better, is it a better place to live in

01:54:35 that I know that our planet is one of many

01:54:37 in the solar system and the solar system’s one of many

01:54:39 in the galaxy?

01:54:40 I think it’s a more, I dread, I sometimes think like,

01:54:43 God, what would it be like to live 300 years ago?

01:54:45 I’d be looking up at the sky, I can’t understand anything.

01:54:47 Oh my God, I’d be like going to bed every night going,

01:54:49 what’s going on here?

01:54:50 Well, I mean, in some sense I agree with you,

01:54:52 but I’m not exactly sure.

01:54:54 So I’m also a scientist, so I share your views,

01:54:57 but I’m not, we’re like rolling down the hill together.

01:55:02 What’s down the hill?

01:55:03 I feel like we’re climbing a hill.

01:55:05 Whatever.

01:55:06 We’re getting closer to enlightenment

01:55:07 and you’re going down the hill.

01:55:10 We’re climbing, we’re getting pulled up a hill

01:55:12 by our curiosity.

01:55:13 Our curiosity is, we’re pulling ourselves up the hill

01:55:16 by our curiosity.

01:55:16 Yeah, Sisyphus was doing the same thing with the rock.

01:55:19 Yeah, yeah, yeah, yeah.

01:55:20 But okay, our happiness aside, do you have concerns

01:55:24 about, you talk about Sam Harris, Elon Musk,

01:55:29 of existential threats of intelligent systems?

01:55:31 No, I’m not worried about existential threats at all.

01:55:33 There are some things we really do need to worry about.

01:55:36 Even today’s AI, we have things we have to worry about.

01:55:38 We have to worry about privacy

01:55:39 and about how it impacts false beliefs in the world.

01:55:42 And we have real problems and things to worry about

01:55:47 with today’s AI.

01:55:48 And that will continue as we create more intelligent systems.

01:55:51 There’s no question, the whole issue

01:55:53 about making intelligent armaments and weapons

01:55:57 is something that really we have to think about carefully.

01:55:59 I don’t think of those as existential threats.

01:56:01 I think those are the kind of threats we always face

01:56:04 and we’ll have to face them here

01:56:05 and we’ll have to deal with them.

01:56:10 We could talk about what people think

01:56:12 are the existential threats,

01:56:13 but when I hear people talking about them,

01:56:16 they all sound hollow to me.

01:56:17 They’re based on ideas, they’re based on people

01:56:20 who really have no idea what intelligence is.

01:56:22 And if they knew what intelligence was,

01:56:24 they wouldn’t say those things.

01:56:26 So those are not experts in the field.

01:56:28 Yeah, so there’s two, right?

01:56:32 So one is like super intelligence.

01:56:33 So a system that becomes far, far superior

01:56:37 in reasoning ability than us humans.

01:56:43 How is that an existential threat?

01:56:46 Then, so there’s a lot of ways in which it could be.

01:56:49 One way is us humans are actually irrational, inefficient

01:56:54 and get in the way of, not happiness,

01:57:00 but whatever the objective function is

01:57:02 of maximizing that objective function.

01:57:04 Super intelligent.

01:57:05 The paperclip problem and things like that.

01:57:06 So the paperclip problem but with the super intelligent.

01:57:09 Yeah, yeah, yeah, yeah.

01:57:10 So we already face this threat in some sense.

01:57:15 They’re called bacteria.

01:57:17 These are organisms in the world

01:57:18 that would like to turn everything into bacteria.

01:57:21 And they’re constantly morphing,

01:57:23 they’re constantly changing to evade our protections.

01:57:26 And in the past, they have killed huge swaths

01:57:30 of populations of humans on this planet.

01:57:33 So if you wanna worry about something

01:57:34 that’s gonna multiply endlessly, we have it.

01:57:38 And I’m far more worried in that regard.

01:57:40 I’m far more worried that some scientists in the laboratory

01:57:43 will create a super virus or a super bacteria

01:57:45 that we cannot control.

01:57:47 That is a more of an existential threat.

01:57:49 Putting an intelligence thing on top of it

01:57:52 actually seems to make it less existential to me.

01:57:54 It’s like, it limits its power.

01:57:56 It limits where it can go.

01:57:57 It limits the number of things it can do in many ways.

01:57:59 A bacteria is something you can’t even see.

01:58:03 So that’s only one of those problems.

01:58:04 Yes, exactly.

01:58:05 So the other one, just in your intuition about intelligence,

01:58:09 when you think about intelligence of us humans,

01:58:12 do you think of that as something,

01:58:14 if you look at intelligence on a spectrum

01:58:16 from zero to us humans,

01:58:18 do you think you can scale that to something far,

01:58:22 far superior to all the mechanisms we’ve been talking about?

01:58:24 I wanna make another point here, Lex, before I get there.

01:58:28 Intelligence is the neocortex.

01:58:30 It is not the entire brain.

01:58:34 The goal is not to make a human.

01:58:36 The goal is not to make an emotional system.

01:58:38 The goal is not to make a system

01:58:39 that wants to have sex and reproduce.

01:58:41 Why would I build that?

01:58:42 If I wanna have a system that wants to reproduce

01:58:44 and have sex, make bacteria, make computer viruses.

01:58:47 Those are bad things, don’t do that.

01:58:49 Those are really bad, don’t do those things.

01:58:52 Regulate those.

01:58:53 But if I just say I want an intelligent system,

01:58:56 why does it have to have any of the human like emotions?

01:58:58 Why does it even care if it lives?

01:59:00 Why does it even care if it has food?

01:59:02 It doesn’t care about those things.

01:59:03 It’s just, you know, it’s just in a trance

01:59:06 thinking about mathematics or it’s out there

01:59:08 just trying to build the space for it on Mars.

01:59:14 That’s a choice we make.

01:59:15 Don’t make human like things,

01:59:17 don’t make replicating things,

01:59:18 don’t make things that have emotions,

01:59:19 just stick to the neocortex.

01:59:21 So that’s a view actually that I share

01:59:23 but not everybody shares in the sense that

01:59:25 you have faith and optimism about us as engineers of systems,

01:59:29 humans as builders of systems to not put in stupid, not.

01:59:34 So this is why I mentioned the bacteria one.

01:59:37 Because you might say, well, some person’s gonna do that.

01:59:40 Well, some person today could create a bacteria

01:59:42 that’s resistant to all the known antibacterial agents.

01:59:46 So we already have that threat.

01:59:49 We already know this is going on.

01:59:51 It’s not a new threat.

01:59:52 So just accept that and then we have to deal with it, right?

01:59:56 Yeah, so my point is nothing to do with intelligence.

01:59:59 Intelligence is a separate component

02:00:01 that you might apply to a system

02:00:03 that wants to reproduce and do stupid things.

02:00:06 Let’s not do that.

02:00:07 Yeah, in fact, it is a mystery

02:00:08 why people haven’t done that yet.

02:00:10 My dad is a physicist, believes that the reason,

02:00:14 he says, for example, nuclear weapons haven’t proliferated

02:00:18 amongst evil people.

02:00:19 So one belief that I share is that

02:00:21 there’s not that many evil people in the world

02:00:25 that would use, whether it’s bacteria or nuclear weapons

02:00:31 or maybe the future AI systems to do bad.

02:00:35 So the fraction is small.

02:00:36 And the second is that it’s actually really hard,

02:00:38 technically, so the intersection between evil

02:00:42 and competent is small in terms of, and that’s the.

02:00:45 And by the way, to really annihilate humanity,

02:00:47 you’d have to have sort of the nuclear winter phenomenon,

02:00:50 which is not one person shooting or even 10 bombs.

02:00:54 You’d have to have some automated system

02:00:56 that detonates a million bombs

02:00:58 or whatever many thousands we have.

02:01:00 So extreme evil combined with extreme competence.

02:01:03 And to start with building some stupid system

02:01:05 that would automatically, Dr. Strangelove type of thing,

02:01:08 you know, I mean, look, we could have

02:01:11 some nuclear bomb go off in some major city in the world.

02:01:14 I think that’s actually quite likely, even in my lifetime.

02:01:17 I don’t think that’s an unlikely thing.

02:01:18 And it’d be a tragedy.

02:01:20 But it won’t be an existential threat.

02:01:23 And it’s the same as, you know, the virus of 1917,

02:01:26 whatever it was, you know, the influenza.

02:01:30 These bad things can happen and the plague and so on.

02:01:33 We can’t always prevent them.

02:01:35 We always try, but we can’t.

02:01:37 But they’re not existential threats

02:01:38 until we combine all those crazy things together.

02:01:41 So on the spectrum of intelligence from zero to human,

02:01:45 do you have a sense of whether it’s possible

02:01:47 to create several orders of magnitude

02:01:51 or at least double that of human intelligence?

02:01:54 Talking about neuro context.

02:01:55 I think it’s the wrong thing to say double the intelligence.

02:01:59 Break it down into different components.

02:02:01 Can I make something that’s a million times fast

02:02:03 than a human brain?

02:02:04 Yes, I can do that.

02:02:06 Could I make something that is,

02:02:09 has a lot more storage than the human brain?

02:02:10 Yes, I could do that.

02:02:11 More common, more copies of common.

02:02:13 Can I make something that attaches

02:02:14 to different sensors than human brain?

02:02:16 Yes, I can do that.

02:02:17 Could I make something that’s distributed?

02:02:19 So these people, yeah, we talked early

02:02:21 about the departure of the neocortex voting.

02:02:23 They don’t have to be co located.

02:02:24 Like, you know, they can be all around the place.

02:02:25 I could do that too.

02:02:29 Those are the levers I have, but is it more intelligent?

02:02:32 Well, it depends what I train it on.

02:02:33 What is it doing?

02:02:34 If it’s.

02:02:35 Well, so here’s the thing.

02:02:36 So let’s say larger neocortex

02:02:39 and or whatever size that allows for higher

02:02:44 and higher hierarchies to form,

02:02:47 we’re talking about reference frames and concepts.

02:02:50 Could I have something that’s a super physicist

02:02:51 or a super mathematician?

02:02:52 Yes.

02:02:53 And the question is, once you have a super physicist,

02:02:56 will they be able to understand something?

02:03:00 Do you have a sense that it will be orders of math,

02:03:02 like us compared to ants?

02:03:03 Could we ever understand it?

02:03:04 Yeah.

02:03:06 Most people cannot understand general relativity.

02:03:11 It’s a really hard thing to get.

02:03:13 I mean, yeah, you can paint it in a fuzzy picture,

02:03:15 stretchy space, you know?

02:03:17 But the field equations to do that

02:03:19 and the deep intuitions are really, really hard.

02:03:23 And I’ve tried, I’m unable to do it.

02:03:25 Like it’s easy to get special relativity,

02:03:28 but general relativity, man, that’s too much.

02:03:32 And so we already live with this to some extent.

02:03:34 The vast majority of people can’t understand actually

02:03:37 what the vast majority of other people actually know.

02:03:40 We’re just, either we don’t have the effort to,

02:03:41 or we can’t, or we don’t have time,

02:03:43 or just not smart enough, whatever.

02:03:46 But we have ways of communicating.

02:03:48 Einstein has spoken in a way that I can understand.

02:03:51 He’s given me analogies that are useful.

02:03:54 I can use those analogies from my own work

02:03:56 and think about concepts that are similar.

02:04:01 It’s not stupid.

02:04:02 It’s not like he’s existing some other plane

02:04:04 and there’s no connection with my plane in the world here.

02:04:06 So that will occur.

02:04:07 It already has occurred.

02:04:09 That’s what my point of this story is.

02:04:10 It already has occurred.

02:04:11 We live it every day.

02:04:14 One could argue that when we create machine intelligence

02:04:17 that think a million times faster than us

02:04:18 that it’ll be so far we can’t make the connections.

02:04:20 But you know, at the moment,

02:04:23 everything that seems really, really hard

02:04:25 to figure out in the world,

02:04:26 when you actually figure it out, it’s not that hard.

02:04:29 You know, almost everyone can understand the multiverses.

02:04:32 Almost everyone can understand quantum physics.

02:04:34 Almost everyone can understand these basic things,

02:04:36 even though hardly any people could figure those things out.

02:04:39 Yeah, but really understand.

02:04:41 But you don’t need to really.

02:04:42 Only a few people really understand.

02:04:43 You need to only understand the projections,

02:04:47 the sprinkles of the useful insights from that.

02:04:50 That was my example of Einstein, right?

02:04:51 His general theory of relativity is one thing

02:04:53 that very, very, very few people can get.

02:04:56 And what if we just said those other few people

02:04:58 are also artificial intelligences?

02:05:00 How bad is that?

02:05:01 In some sense they are, right?

02:05:02 Yeah, they say already.

02:05:04 I mean, Einstein wasn’t a really normal person.

02:05:06 He had a lot of weird quirks.

02:05:07 And so did the other people who worked with him.

02:05:09 So, you know, maybe they already were sort of

02:05:11 this astral plane of intelligence that,

02:05:14 we live with it already.

02:05:15 It’s not a problem.

02:05:17 It’s still useful and, you know.

02:05:20 So do you think we are the only intelligent life

02:05:22 out there in the universe?

02:05:24 I would say that intelligent life

02:05:27 has and will exist elsewhere in the universe.

02:05:29 I’ll say that.

02:05:31 There was a question about

02:05:32 contemporaneous intelligence life,

02:05:34 which is hard to even answer

02:05:35 when we think about relativity and the nature of space time.

02:05:39 Can’t say what exactly is this time

02:05:41 someplace else in the world.

02:05:43 But I think it’s, you know,

02:05:44 I do worry a lot about the filter idea,

02:05:48 which is that perhaps intelligent species

02:05:52 don’t last very long.

02:05:54 And so we haven’t been around very long.

02:05:55 And as a technological species,

02:05:57 we’ve been around for almost nothing, you know.

02:05:59 What, 200 years, something like that.

02:06:02 And we don’t have any data,

02:06:04 a good data point on whether it’s likely

02:06:06 that we’ll survive or not.

02:06:08 So do I think that there have been intelligent life

02:06:10 elsewhere in the universe?

02:06:11 Almost certainly, of course.

02:06:13 In the past, in the future, yes.

02:06:16 Does it survive for a long time?

02:06:17 I don’t know.

02:06:18 This is another reason I’m excited about our work,

02:06:21 is our work meaning the general world of AI.

02:06:24 I think we can build intelligent machines

02:06:28 that outlast us.

02:06:32 You know, they don’t have to be tied to Earth.

02:06:34 They don’t have to, you know,

02:06:35 I’m not saying they’re recreating, you know,

02:06:38 aliens, I’m just saying,

02:06:40 if I asked myself,

02:06:41 and this might be a good point to end on here.

02:06:44 If I asked myself, you know,

02:06:45 what’s special about our species?

02:06:47 We’re not particularly interesting physically.

02:06:49 We don’t fly, we’re not good swimmers,

02:06:51 we’re not very fast, we’re not very strong, you know.

02:06:54 It’s our brain, that’s the only thing.

02:06:55 And we are the only species on this planet

02:06:57 that’s built the model of the world

02:06:58 that extends beyond what we can actually sense.

02:07:01 We’re the only people who know about

02:07:03 the far side of the moon, and the other universes,

02:07:05 and I mean, other galaxies, and other stars,

02:07:07 and about what happens in the atom.

02:07:09 There’s no, that knowledge doesn’t exist anywhere else.

02:07:12 It’s only in our heads.

02:07:13 Cats don’t do it, dogs don’t do it,

02:07:15 monkeys don’t do it, it’s just on.

02:07:16 And that is what we’ve created that’s unique.

02:07:18 Not our genes, it’s knowledge.

02:07:20 And if I asked me, what is the legacy of humanity?

02:07:23 What should our legacy be?

02:07:25 It should be knowledge.

02:07:25 We should preserve our knowledge

02:07:27 in a way that it can exist beyond us.

02:07:30 And I think the best way of doing that,

02:07:32 in fact you have to do it,

02:07:33 is it has to go along with intelligent machines

02:07:34 that understand that knowledge.

02:07:37 It’s a very broad idea, but we should be thinking,

02:07:41 I call it a state planning for humanity.

02:07:43 We should be thinking about what we wanna leave behind

02:07:46 when as a species we’re no longer here.

02:07:49 And that’ll happen sometime.

02:07:51 Sooner or later it’s gonna happen.

02:07:52 And understanding intelligence and creating intelligence

02:07:56 gives us a better chance to prolong.

02:07:58 It does give us a better chance to prolong life, yes.

02:08:01 It gives us a chance to live on other planets.

02:08:03 But even beyond that, I mean our solar system

02:08:06 will disappear one day, just given enough time.

02:08:08 So I don’t know, I doubt we’ll ever be able to travel

02:08:12 to other things, but we could tell the stars,

02:08:15 but we could send intelligent machines to do that.

02:08:17 So you have an optimistic, a hopeful view of our knowledge

02:08:23 of the echoes of human civilization

02:08:26 living through the intelligent systems we create?

02:08:29 Oh, totally.

02:08:30 Well, I think the intelligent systems we create

02:08:31 are in some sense the vessel for bringing them beyond Earth

02:08:36 or making them last beyond humans themselves.

02:08:39 How do you feel about that?

02:08:41 That they won’t be human, quote unquote?

02:08:43 Who cares?

02:08:45 Human, what is human?

02:08:46 Our species are changing all the time.

02:08:48 Human today is not the same as human just 50 years ago.

02:08:52 What is human?

02:08:53 Do we care about our genetics?

02:08:54 Why is that important?

02:08:56 As I point out, our genetics are no more interesting

02:08:58 than a bacterium’s genetics.

02:08:59 It’s no more interesting than a monkey’s genetics.

02:09:01 What we have, what’s unique and what’s valuable

02:09:04 is our knowledge, what we’ve learned about the world.

02:09:07 And that is the rare thing.

02:09:09 That’s the thing we wanna preserve.

02:09:11 It’s, who cares about our genes?

02:09:13 That’s not.

02:09:14 It’s the knowledge.

02:09:16 It’s the knowledge.

02:09:17 That’s a really good place to end.

02:09:19 Thank you so much for talking to me.

02:09:20 No, it was fun.