Hugging Face's logo

Datasets:
shubhamg2208
/
lexicap

Tasks:
Text Classification
Text Generation
Sub-tasks:
sentiment-analysis
dialogue-modeling
language-modeling
Languages:
English
Size:
n<1K
Tags:
karpathy,whisper,openai
Dataset card Files Community
lexicap / vtt /episode_025_small.vtt
Shubham Gupta
Add readme and files
a3be5d0
raw
history blame
211 kB
WEBVTT
00:00.000 --> 00:02.360
The following is a conversation with Jeff Hawkins.
00:02.360 --> 00:04.140
He's the founder of the Redwood Center
00:04.140 --> 00:06.560
for Theoretical and Neuroscience in 2002
00:06.560 --> 00:08.980
and New Menta in 2005.
00:08.980 --> 00:11.920
In his 2004 book titled On Intelligence
00:11.920 --> 00:13.880
and in the research before and after,
00:13.880 --> 00:16.200
he and his team have worked to reverse engineer
00:16.200 --> 00:19.160
the New York Cortex and propose artificial intelligence
00:19.160 --> 00:21.360
architectures approaches and ideas
00:21.360 --> 00:23.640
that are inspired by the human brain.
00:23.640 --> 00:25.960
These ideas include hierarchical temporal memory,
00:25.960 --> 00:30.080
HTM from 2004, and New Work, The Thousand's Brain's Theory
00:30.080 --> 00:33.800
of Intelligence from 2017, 18, and 19.
00:33.800 --> 00:36.120
Jeff's ideas have been an inspiration
00:36.120 --> 00:38.960
to many who have looked for progress beyond the current
00:38.960 --> 00:41.760
machine learning approaches, but they have also received
00:41.760 --> 00:44.680
criticism for lacking a body of empirical evidence
00:44.680 --> 00:46.240
supporting the models.
00:46.240 --> 00:49.040
This is always a challenge when seeking more than small
00:49.040 --> 00:51.440
incremental steps forward in AI.
00:51.440 --> 00:54.120
Jeff is a brilliant mind and many of the ideas
00:54.120 --> 00:56.520
he has developed and aggregated from neuroscience
00:56.520 --> 00:59.160
are worth understanding and thinking about.
00:59.160 --> 01:00.960
There are limits to deep learning
01:00.960 --> 01:02.960
as it is currently defined.
01:02.960 --> 01:05.800
Forward progress in AI is shrouded in mystery.
01:05.800 --> 01:07.800
My hope is that conversations like this
01:07.800 --> 01:11.480
can help provide an inspiring spark for new ideas.
01:11.480 --> 01:14.060
This is the Artificial Intelligence Podcast.
01:14.060 --> 01:16.200
If you enjoy it, subscribe on YouTube,
01:16.200 --> 01:18.680
iTunes, or simply connect with me on Twitter
01:18.680 --> 01:21.560
at Lex Freedman spelled F R I D.
01:21.560 --> 01:25.360
And now here's my conversation with Jeff Hawkins.
01:26.800 --> 01:29.880
Are you more interested in understanding the human brain
01:29.880 --> 01:32.040
or in creating artificial systems
01:32.040 --> 01:34.680
that have many of the same qualities,
01:34.680 --> 01:38.600
but don't necessarily require that you actually understand
01:38.600 --> 01:41.520
the underpinning workings of our mind?
01:41.520 --> 01:44.020
So there's a clear answer to that question.
01:44.020 --> 01:46.800
My primary interest is understanding the human brain.
01:46.800 --> 01:51.520
No question about it, but I also firmly believe
01:51.520 --> 01:53.880
that we will not be able to create
01:53.880 --> 01:55.040
fully intelligent machines
01:55.040 --> 01:57.280
until we understand how the human brain works.
01:57.280 --> 02:00.120
So I don't see those as separate problems.
02:00.120 --> 02:01.720
I think there's limits to what can be done
02:01.720 --> 02:02.640
with machine intelligence
02:02.640 --> 02:04.040
if you don't understand the principles
02:04.040 --> 02:05.680
by which the brain works.
02:05.680 --> 02:07.880
And so I actually believe that studying the brain
02:07.880 --> 02:11.960
is actually the fastest way to get to machine intelligence.
02:11.960 --> 02:14.640
And within that, let me ask the impossible question.
02:14.640 --> 02:17.160
How do you not define, but at least think
02:17.160 --> 02:19.400
about what it means to be intelligent?
02:19.400 --> 02:22.240
So I didn't try to answer that question first.
02:22.240 --> 02:24.480
We said, let's just talk about how the brain works.
02:24.480 --> 02:26.680
And let's figure out how certain parts of the brain,
02:26.680 --> 02:29.880
mostly the neocortex, but some other parts too,
02:29.880 --> 02:32.320
the parts of the brain most associated with intelligence.
02:32.320 --> 02:35.800
And let's discover the principles by how they work.
02:35.800 --> 02:39.320
Because intelligence isn't just like some mechanism
02:39.320 --> 02:40.640
and it's not just some capabilities.
02:40.640 --> 02:43.000
It's like, okay, we don't even know where to begin
02:43.000 --> 02:44.000
on this stuff.
02:44.000 --> 02:49.000
And so now that we've made a lot of progress on this,
02:49.000 --> 02:50.480
after we've made a lot of progress
02:50.480 --> 02:53.200
on how the neocortex works, and we can talk about that,
02:53.200 --> 02:54.560
I now have a very good idea
02:54.560 --> 02:57.200
what's gonna be required to make intelligent machines.
02:57.200 --> 02:59.600
I can tell you today, some of the things
02:59.600 --> 03:02.120
are gonna be necessary, I believe,
03:02.120 --> 03:03.480
to create intelligent machines.
03:03.480 --> 03:04.600
Well, so we'll get there.
03:04.600 --> 03:07.440
We'll get to the neocortex and some of the theories
03:07.440 --> 03:09.200
of how the whole thing works.
03:09.200 --> 03:11.720
And you're saying, as we understand more and more
03:12.720 --> 03:14.800
about the neocortex, about our own human mind,
03:14.800 --> 03:17.720
we'll be able to start to more specifically define
03:17.720 --> 03:18.680
what it means to be intelligent.
03:18.680 --> 03:21.880
It's not useful to really talk about that until...
03:21.880 --> 03:23.560
I don't know if it's not useful.
03:23.560 --> 03:26.200
Look, there's a long history of AI, as you know.
03:26.200 --> 03:28.920
And there's been different approaches taken to it.
03:28.920 --> 03:30.160
And who knows?
03:30.160 --> 03:32.280
Maybe they're all useful.
03:32.280 --> 03:37.280
So the good old fashioned AI, the expert systems,
03:37.360 --> 03:38.960
the current convolutional neural networks,
03:38.960 --> 03:40.440
they all have their utility.
03:41.280 --> 03:43.800
They all have a value in the world.
03:43.800 --> 03:45.280
But I would think almost everyone agreed
03:45.280 --> 03:46.640
that none of them are really intelligent
03:46.640 --> 03:49.880
in a sort of a deep way that humans are.
03:49.880 --> 03:53.600
And so it's just the question of how do you get
03:53.600 --> 03:56.440
from where those systems were or are today
03:56.440 --> 03:59.240
to where a lot of people think we're gonna go.
03:59.240 --> 04:02.320
And there's a big, big gap there, a huge gap.
04:02.320 --> 04:06.240
And I think the quickest way of bridging that gap
04:06.240 --> 04:08.840
is to figure out how the brain does that.
04:08.840 --> 04:10.160
And then we can sit back and look and say,
04:10.160 --> 04:13.000
oh, what are these principles that the brain works on
04:13.000 --> 04:15.160
are necessary and which ones are not?
04:15.160 --> 04:16.640
Clearly, we don't have to build this in,
04:16.640 --> 04:18.520
and tellage machines aren't gonna be built
04:18.520 --> 04:22.760
out of organic living cells.
04:22.760 --> 04:24.720
But there's a lot of stuff that goes on the brain
04:24.720 --> 04:25.920
that's gonna be necessary.
04:25.920 --> 04:30.320
So let me ask maybe, before we get into the fun details,
04:30.320 --> 04:33.080
let me ask maybe a depressing or a difficult question.
04:33.080 --> 04:36.240
Do you think it's possible that we will never
04:36.240 --> 04:38.120
be able to understand how our brain works,
04:38.120 --> 04:41.840
that maybe there's aspects to the human mind
04:41.840 --> 04:46.160
like we ourselves cannot introspectively get to the core,
04:46.160 --> 04:48.120
that there's a wall you eventually hit?
04:48.120 --> 04:50.200
Yeah, I don't believe that's the case.
04:52.040 --> 04:53.240
I have never believed that's the case.
04:53.240 --> 04:55.760
There's not been a single thing we've ever,
04:55.760 --> 04:57.760
humans have ever put their minds to that we've said,
04:57.760 --> 05:00.320
oh, we reached the wall, we can't go any further.
05:00.320 --> 05:01.680
People keep saying that.
05:01.680 --> 05:03.400
People used to believe that about life, you know,
05:03.400 --> 05:04.480
Alain Vitao, right?
05:04.480 --> 05:06.360
There's like, what's the difference in living matter
05:06.360 --> 05:07.280
and nonliving matter?
05:07.280 --> 05:09.120
Something special you never understand.
05:09.120 --> 05:10.640
We no longer think that.
05:10.640 --> 05:14.720
So there's no historical evidence that suggests this is the case
05:14.720 --> 05:17.600
and I just never even consider that's a possibility.
05:17.600 --> 05:21.840
I would also say today, we understand so much
05:21.840 --> 05:22.800
about the neocortex.
05:22.800 --> 05:25.480
We've made tremendous progress in the last few years
05:25.480 --> 05:29.160
that I no longer think of as an open question.
05:30.000 --> 05:32.560
The answers are very clear to me and the pieces
05:32.560 --> 05:34.800
that we don't know are clear to me,
05:34.800 --> 05:37.440
but the framework is all there and it's like, oh, okay,
05:37.440 --> 05:38.600
we're gonna be able to do this.
05:38.600 --> 05:39.960
This is not a problem anymore.
05:39.960 --> 05:42.680
It just takes time and effort, but there's no mystery,
05:42.680 --> 05:44.040
a big mystery anymore.
05:44.040 --> 05:47.800
So then let's get into it for people like myself
05:47.800 --> 05:52.800
who are not very well versed in the human brain,
05:52.960 --> 05:53.840
except my own.
05:54.800 --> 05:57.320
Can you describe to me at the highest level,
05:57.320 --> 05:59.120
what are the different parts of the human brain
05:59.120 --> 06:02.080
and then zooming in on the neocortex,
06:02.080 --> 06:05.480
the parts of the neocortex and so on, a quick overview.
06:05.480 --> 06:06.640
Yeah, sure.
06:06.640 --> 06:10.800
The human brain, we can divide it roughly into two parts.
06:10.800 --> 06:14.200
There's the old parts, lots of pieces,
06:14.200 --> 06:15.680
and then there's the new part.
06:15.680 --> 06:18.040
The new part is the neocortex.
06:18.040 --> 06:20.440
It's new because it didn't exist before mammals.
06:20.440 --> 06:23.000
The only mammals have a neocortex and in humans
06:23.000 --> 06:24.760
and primates is very large.
06:24.760 --> 06:29.400
In the human brain, the neocortex occupies about 70 to 75%
06:29.400 --> 06:30.640
of the volume of the brain.
06:30.640 --> 06:32.080
It's huge.
06:32.080 --> 06:34.840
And the old parts of the brain are,
06:34.840 --> 06:36.000
there's lots of pieces there.
06:36.000 --> 06:38.760
There's a spinal cord and there's the brainstem
06:38.760 --> 06:40.240
and the cerebellum and the different parts
06:40.240 --> 06:42.040
of the basal ganglion and so on.
06:42.040 --> 06:42.960
In the old parts of the brain,
06:42.960 --> 06:44.800
you have the autonomic regulation,
06:44.800 --> 06:46.280
like breathing and heart rate.
06:46.280 --> 06:48.240
You have basic behaviors.
06:48.240 --> 06:49.960
So like walking and running are controlled
06:49.960 --> 06:51.400
by the old parts of the brain.
06:51.400 --> 06:53.080
All the emotional centers of the brain
06:53.080 --> 06:53.920
are in the old part of the brain.
06:53.920 --> 06:55.080
So when you feel anger or hungry,
06:55.080 --> 06:56.080
lust or things like that,
06:56.080 --> 06:57.880
those are all in the old parts of the brain.
06:59.080 --> 07:02.160
And we associate with the neocortex
07:02.160 --> 07:03.320
all the things we think about
07:03.320 --> 07:05.760
as sort of high level perception.
07:05.760 --> 07:10.760
And cognitive functions, anything from seeing and hearing
07:10.920 --> 07:14.560
and touching things to language, to mathematics
07:14.560 --> 07:16.920
and engineering and science and so on.
07:16.920 --> 07:19.760
Those are all associated with the neocortex.
07:19.760 --> 07:21.760
And they're certainly correlated.
07:21.760 --> 07:24.000
Our abilities in those regards are correlated
07:24.000 --> 07:25.800
with the relative size of our neocortex
07:25.800 --> 07:27.960
compared to other mammals.
07:27.960 --> 07:30.520
So that's like the rough division.
07:30.520 --> 07:32.760
And you obviously can't understand
07:32.760 --> 07:35.160
the neocortex completely isolated,
07:35.160 --> 07:37.040
but you can understand a lot of it
07:37.040 --> 07:38.720
with just a few interfaces
07:38.720 --> 07:40.320
to the old parts of the brain.
07:40.320 --> 07:44.960
And so it gives you a system to study.
07:44.960 --> 07:48.040
The other remarkable thing about the neocortex
07:48.040 --> 07:49.880
compared to the old parts of the brain
07:49.880 --> 07:52.880
is the neocortex is extremely uniform.
07:52.880 --> 07:55.720
It's not visually or anatomically,
07:55.720 --> 07:58.800
or it's very, it's like a,
07:58.800 --> 08:00.080
I always like to say it's like the size
08:00.080 --> 08:03.720
of a dinner napkin, about two and a half millimeters thick.
08:03.720 --> 08:06.000
And it looks remarkably the same everywhere.
08:06.000 --> 08:07.920
Everywhere you look in that two and a half millimeters
08:07.920 --> 08:10.080
is this detailed architecture.
08:10.080 --> 08:11.560
And it looks remarkably the same everywhere.
08:11.560 --> 08:12.600
And that's a cross species,
08:12.600 --> 08:15.360
a mouse versus a cat and a dog and a human.
08:15.360 --> 08:17.080
Where if you look at the old parts of the brain,
08:17.080 --> 08:19.640
there's lots of little pieces do specific things.
08:19.640 --> 08:22.040
So it's like the old parts of a brain evolved,
08:22.040 --> 08:23.640
like this is the part that controls heart rate
08:23.640 --> 08:24.840
and this is the part that controls this
08:24.840 --> 08:25.800
and this is this kind of thing.
08:25.800 --> 08:27.200
And that's this kind of thing.
08:27.200 --> 08:30.080
And these evolve for eons of a long, long time
08:30.080 --> 08:31.600
and they have those specific functions.
08:31.600 --> 08:33.240
And all of a sudden mammals come along
08:33.240 --> 08:35.200
and they got this thing called the neocortex
08:35.200 --> 08:38.200
and it got large by just replicating the same thing
08:38.200 --> 08:39.440
over and over and over again.
08:39.440 --> 08:42.680
This is like, wow, this is incredible.
08:42.680 --> 08:46.240
So all the evidence we have,
08:46.240 --> 08:50.040
and this is an idea that was first articulated
08:50.040 --> 08:52.080
in a very cogent and beautiful argument
08:52.080 --> 08:55.720
by a guy named Vernon Malkassel in 1978, I think it was,
08:56.880 --> 09:01.640
that the neocortex all works on the same principle.
09:01.640 --> 09:05.320
So language, hearing, touch, vision, engineering,
09:05.320 --> 09:07.040
all these things are basically underlying
09:07.040 --> 09:10.400
or all built in the same computational substrate.
09:10.400 --> 09:11.880
They're really all the same problem.
09:11.880 --> 09:14.880
So the low level of the building blocks all look similar.
09:14.880 --> 09:16.320
Yeah, and they're not even that low level.
09:16.320 --> 09:17.920
We're not talking about like neurons.
09:17.920 --> 09:19.960
We're talking about this very complex circuit
09:19.960 --> 09:23.560
that exists throughout the neocortex is remarkably similar.
09:23.560 --> 09:26.400
It is, it's like, yes, you see variations of it here
09:26.400 --> 09:29.680
and they're more of the cell, that's not old and so on.
09:29.680 --> 09:31.840
But what Malkassel argued was,
09:31.840 --> 09:35.640
it says, you know, if you take a section on neocortex,
09:35.640 --> 09:38.640
why is one a visual area and one is a auditory area?
09:38.640 --> 09:41.240
Or why is, and his answer was,
09:41.240 --> 09:43.240
it's because one is connected to eyes
09:43.240 --> 09:45.440
and one is connected to ears.
09:45.440 --> 09:47.840
Literally, you mean just as most closest
09:47.840 --> 09:50.440
in terms of the number of connections to the sensor?
09:50.440 --> 09:52.920
Literally, if you took the optic nerve
09:52.920 --> 09:55.320
and attached it to a different part of the neocortex,
09:55.320 --> 09:58.000
that part would become a visual region.
09:58.000 --> 10:00.400
This actually, this experiment was actually done
10:00.400 --> 10:05.000
by Murgankasur in developing, I think it was lemurs,
10:05.000 --> 10:06.720
I can't remember what it was, it's some animal.
10:06.720 --> 10:08.560
And there's a lot of evidence to this.
10:08.560 --> 10:09.960
You know, if you take a blind person,
10:09.960 --> 10:12.240
a person is born blind at birth,
10:12.240 --> 10:15.480
they're born with a visual neocortex.
10:15.480 --> 10:18.320
It doesn't, may not get any input from the eyes
10:18.320 --> 10:21.280
because of some congenital defect or something.
10:21.280 --> 10:24.720
And that region becomes, does something else.
10:24.720 --> 10:27.000
It picks up another task.
10:27.000 --> 10:32.000
So, and it's, so it's this very complex thing.
10:32.280 --> 10:33.760
It's not like, oh, they're all built on neurons.
10:33.760 --> 10:36.480
No, they're all built in this very complex circuit.
10:36.480 --> 10:40.280
And somehow that circuit underlies everything.
10:40.280 --> 10:44.760
And so this is, it's called the common cortical algorithm,
10:44.760 --> 10:47.960
if you will, some scientists just find it hard to believe.
10:47.960 --> 10:50.040
And they just say, I can't believe that's true.
10:50.040 --> 10:52.080
But the evidence is overwhelming in this case.
10:52.080 --> 10:54.320
And so a large part of what it means
10:54.320 --> 10:56.440
to figure out how the brain creates intelligence
10:56.440 --> 10:59.840
and what is intelligence in the brain
10:59.840 --> 11:02.040
is to understand what that circuit does.
11:02.040 --> 11:05.040
If you can figure out what that circuit does,
11:05.040 --> 11:06.920
as amazing as it is, then you can,
11:06.920 --> 11:10.480
then you understand what all these other cognitive functions are.
11:10.480 --> 11:13.280
So if you were to sort of put neocortex
11:13.280 --> 11:15.160
outside of your book on intelligence,
11:15.160 --> 11:17.480
you look, if you wrote a giant tome,
11:17.480 --> 11:19.800
a textbook on the neocortex,
11:19.800 --> 11:23.760
and you look maybe a couple of centuries from now,
11:23.760 --> 11:26.520
how much of what we know now would still be accurate
11:26.520 --> 11:27.680
two centuries from now.
11:27.680 --> 11:30.840
So how close are we in terms of understanding?
11:30.840 --> 11:33.000
I have to speak from my own particular experience here.
11:33.000 --> 11:36.440
So I run a small research lab here.
11:36.440 --> 11:38.040
It's like any other research lab.
11:38.040 --> 11:39.440
I'm sort of the principal investigator.
11:39.440 --> 11:40.280
There's actually two of us,
11:40.280 --> 11:42.560
and there's a bunch of other people.
11:42.560 --> 11:43.840
And this is what we do.
11:43.840 --> 11:44.960
We started the neocortex,
11:44.960 --> 11:46.960
and we publish our results and so on.
11:46.960 --> 11:48.520
So about three years ago,
11:49.840 --> 11:52.480
we had a real breakthrough in this field.
11:52.480 --> 11:53.320
Just tremendous breakthrough.
11:53.320 --> 11:56.520
We started, we now publish, I think, three papers on it.
11:56.520 --> 12:00.200
And so I have a pretty good understanding
12:00.200 --> 12:02.320
of all the pieces and what we're missing.
12:02.320 --> 12:06.280
I would say that almost all the empirical data
12:06.280 --> 12:08.520
we've collected about the brain, which is enormous.
12:08.520 --> 12:10.320
If you don't know the neuroscience literature,
12:10.320 --> 12:13.960
it's just incredibly big.
12:13.960 --> 12:16.840
And it's, for the most part, all correct.
12:16.840 --> 12:20.240
It's facts and experimental results
12:20.240 --> 12:22.960
and measurements and all kinds of stuff.
12:22.960 --> 12:25.800
But none of that has been really assimilated
12:25.800 --> 12:27.840
into a theoretical framework.
12:27.840 --> 12:32.240
It's data without, in the language of Thomas Kuhn,
12:32.240 --> 12:35.280
the historian, it would be sort of a preparadigm science.
12:35.280 --> 12:38.160
Lots of data, but no way to fit it in together.
12:38.160 --> 12:39.520
I think almost all of that's correct.
12:39.520 --> 12:42.120
There's gonna be some mistakes in there.
12:42.120 --> 12:43.240
And for the most part,
12:43.240 --> 12:45.480
there aren't really good cogent theories
12:45.480 --> 12:47.240
about how to put it together.
12:47.240 --> 12:50.040
It's not like we have two or three competing good theories,
12:50.040 --> 12:51.520
which ones are right and which ones are wrong.
12:51.520 --> 12:53.720
It's like, yeah, people just scratching their heads
12:53.720 --> 12:55.560
throwing things, you know, some people giving up
12:55.560 --> 12:57.560
on trying to figure out what the whole thing does.
12:57.560 --> 13:00.960
In fact, there's very, very few labs that we do
13:00.960 --> 13:03.280
that focus really on theory
13:03.280 --> 13:06.760
and all this unassimilated data and trying to explain it.
13:06.760 --> 13:08.880
So it's not like we've got it wrong.
13:08.880 --> 13:11.120
It's just that we haven't got it at all.
13:11.120 --> 13:15.040
So it's really, I would say, pretty early days
13:15.040 --> 13:18.360
in terms of understanding the fundamental theories,
13:18.360 --> 13:20.240
forces of the way our mind works.
13:20.240 --> 13:21.080
I don't think so.
13:21.080 --> 13:23.760
I would have said that's true five years ago.
13:25.360 --> 13:28.600
So as I said, we had some really big breakthroughs
13:28.600 --> 13:30.800
on this recently and we started publishing papers on this.
13:30.800 --> 13:34.240
So you can get to that.
13:34.240 --> 13:36.760
But so I don't think it's, you know, I'm an optimist
13:36.760 --> 13:38.280
and from where I sit today,
13:38.280 --> 13:39.440
most people would disagree with this,
13:39.440 --> 13:41.640
but from where I sit today, from what I know,
13:43.240 --> 13:44.920
it's not super early days anymore.
13:44.920 --> 13:46.840
We are, you know, the way these things go
13:46.840 --> 13:48.200
is it's not a linear path, right?
13:48.200 --> 13:49.840
You don't just start accumulating
13:49.840 --> 13:50.800
and get better and better and better.
13:50.800 --> 13:52.920
No, you got all the stuff you've collected.
13:52.920 --> 13:53.760
None of it makes sense.
13:53.760 --> 13:55.640
All these different things are just sort of around.
13:55.640 --> 13:57.120
And then you're going to have some breaking points
13:57.120 --> 13:59.400
all of a sudden, oh my God, now we got it right.
13:59.400 --> 14:01.120
That's how it goes in science.
14:01.120 --> 14:04.480
And I personally feel like we passed that little thing
14:04.480 --> 14:06.320
about a couple of years ago.
14:06.320 --> 14:07.560
All that big thing a couple of years ago.
14:07.560 --> 14:09.600
So we can talk about that.
14:09.600 --> 14:11.000
Time will tell if I'm right,
14:11.000 --> 14:12.640
but I feel very confident about it.
14:12.640 --> 14:15.120
That's when we'll just say it on tape like this.
14:15.120 --> 14:18.040
At least very optimistic.
14:18.040 --> 14:20.160
So let's, before those few years ago,
14:20.160 --> 14:23.200
let's take a step back to HTM,
14:23.200 --> 14:25.960
the hierarchical temporal memory theory,
14:25.960 --> 14:27.480
which you first proposed on intelligence
14:27.480 --> 14:29.280
and went through a few different generations.
14:29.280 --> 14:31.200
Can you describe what it is,
14:31.200 --> 14:33.560
how it would evolve through the three generations
14:33.560 --> 14:35.360
since you first put it on paper?
14:35.360 --> 14:39.240
Yeah, so one of the things that neuroscientists
14:39.240 --> 14:42.920
just sort of missed for many, many years.
14:42.920 --> 14:45.720
And especially people were thinking about theory
14:45.720 --> 14:47.720
was the nature of time in the brain.
14:47.720 --> 14:50.440
Brain's process, information through time,
14:50.440 --> 14:53.280
the information coming into the brain is constantly changing.
14:53.280 --> 14:56.160
The patterns from my speech right now,
14:56.160 --> 14:58.520
if you're listening to it at normal speed,
14:58.520 --> 15:00.080
would be changing on your ears
15:00.080 --> 15:02.680
about every 10 milliseconds or so, you'd have a change.
15:02.680 --> 15:05.320
This constant flow, when you look at the world,
15:05.320 --> 15:06.800
your eyes are moving constantly,
15:06.800 --> 15:08.240
three to five times a second,
15:08.240 --> 15:09.920
and the input's completely, completely.
15:09.920 --> 15:11.800
If I were to touch something like a coffee cup
15:11.800 --> 15:13.880
as I move my fingers, the input changes.
15:13.880 --> 15:16.840
So this idea that the brain works on time
15:16.840 --> 15:19.640
changing patterns is almost completely,
15:19.640 --> 15:21.080
or was almost completely missing
15:21.080 --> 15:23.520
from a lot of the basic theories like fears of vision
15:23.520 --> 15:24.360
and so on.
15:24.360 --> 15:26.280
It's like, oh no, we're gonna put this image in front of you
15:26.280 --> 15:28.360
and flash it and say, what is it?
15:28.360 --> 15:31.120
A convolutional neural network's worked that way today, right?
15:31.120 --> 15:33.280
Classified this picture.
15:33.280 --> 15:35.120
But that's not what vision is like.
15:35.120 --> 15:37.760
Vision is this sort of crazy time based pattern
15:37.760 --> 15:39.080
that's going all over the place,
15:39.080 --> 15:40.920
and so is touch and so is hearing.
15:40.920 --> 15:42.880
So the first part of a hierarchical temporal memory
15:42.880 --> 15:44.280
was the temporal part.
15:44.280 --> 15:47.680
It's to say, you won't understand the brain,
15:47.680 --> 15:49.360
nor will you understand intelligent machines
15:49.360 --> 15:51.720
unless you're dealing with time based patterns.
15:51.720 --> 15:54.760
The second thing was, the memory component of it was,
15:54.760 --> 15:59.760
is to say that we aren't just processing input,
15:59.760 --> 16:02.000
we learn a model of the world.
16:02.000 --> 16:04.000
And the memory stands for that model.
16:04.000 --> 16:06.640
The point of the brain, part of the neocortex,
16:06.640 --> 16:07.840
it learns a model of the world.
16:07.840 --> 16:10.840
We have to store things that are experiences
16:10.840 --> 16:13.520
in a form that leads to a model of the world.
16:13.520 --> 16:15.080
So we can move around the world,
16:15.080 --> 16:16.240
we can pick things up and do things
16:16.240 --> 16:17.520
and navigate and know how it's going on.
16:17.520 --> 16:19.320
So that's what the memory referred to.
16:19.320 --> 16:22.320
And many people just, they were thinking about like,
16:22.320 --> 16:24.480
certain processes without memory at all.
16:24.480 --> 16:26.120
They're just like processing things.
16:26.120 --> 16:28.320
And then finally, the hierarchical component
16:28.320 --> 16:31.640
was a reflection to that the neocortex,
16:31.640 --> 16:33.920
although it's just a uniform sheet of cells,
16:33.920 --> 16:36.920
different parts of it project to other parts,
16:36.920 --> 16:38.680
which project to other parts.
16:38.680 --> 16:42.400
And there is a sort of rough hierarchy in terms of that.
16:42.400 --> 16:46.000
So the hierarchical temporal memory is just saying,
16:46.000 --> 16:47.720
look, we should be thinking about the brain
16:47.720 --> 16:52.720
as time based, model memory based and hierarchical processing.
16:54.760 --> 16:58.160
And that was a placeholder for a bunch of components
16:58.160 --> 17:00.720
that we would then plug into that.
17:00.720 --> 17:02.600
We still believe all those things I just said,
17:02.600 --> 17:06.960
but we now know so much more that I'm stopping to use
17:06.960 --> 17:08.200
the word hierarchical temporal memory yet
17:08.200 --> 17:11.320
because it's insufficient to capture the stuff we know.
17:11.320 --> 17:12.960
So again, it's not incorrect,
17:12.960 --> 17:15.800
but I now know more and I would rather describe it
17:15.800 --> 17:16.800
more accurately.
17:16.800 --> 17:20.360
Yeah, so you're basically, we could think of HTM
17:20.360 --> 17:24.800
as emphasizing that there's three aspects of intelligence
17:24.800 --> 17:25.920
that are important to think about
17:25.920 --> 17:28.880
whatever the eventual theory converges to.
17:28.880 --> 17:32.480
So in terms of time, how do you think of nature of time
17:32.480 --> 17:33.880
across different time scales?
17:33.880 --> 17:36.800
So you mentioned things changing,
17:36.800 --> 17:39.160
sensory inputs changing every 10, 20 minutes.
17:39.160 --> 17:40.520
What about every few minutes?
17:40.520 --> 17:42.120
Every few months and years?
17:42.120 --> 17:44.840
Well, if you think about a neuroscience problem,
17:44.840 --> 17:49.640
the brain problem, neurons themselves can stay active
17:49.640 --> 17:51.560
for certain periods of time.
17:51.560 --> 17:53.280
They're parts of the brain where they stay active
17:53.280 --> 17:56.680
for minutes, so you could hold a certain perception
17:56.680 --> 18:01.320
or an activity for a certain period of time,
18:01.320 --> 18:04.480
but not most of them don't last that long.
18:04.480 --> 18:07.160
And so if you think about your thoughts
18:07.160 --> 18:09.080
or the activity neurons,
18:09.080 --> 18:10.680
if you're gonna wanna involve something
18:10.680 --> 18:11.920
that happened a long time ago,
18:11.920 --> 18:14.400
even just this morning, for example,
18:14.400 --> 18:16.360
the neurons haven't been active throughout that time.
18:16.360 --> 18:17.800
So you have to store that.
18:17.800 --> 18:20.720
So by I ask you, what did you have for breakfast today?
18:20.720 --> 18:22.000
That is memory.
18:22.000 --> 18:24.160
That is, you've built it into your model of the world now.
18:24.160 --> 18:27.880
You remember that and that memory is in the synapses,
18:27.880 --> 18:30.080
it's basically in the formation of synapses.
18:30.080 --> 18:35.080
And so you're sliding into what used to different time scales.
18:36.760 --> 18:38.280
There's time scales of which we are
18:38.280 --> 18:40.440
like understanding my language and moving about
18:40.440 --> 18:41.840
and seeing things rapidly and over time.
18:41.840 --> 18:44.280
That's the time scales of activities of neurons.
18:44.280 --> 18:46.200
But if you wanna get in longer time scales,
18:46.200 --> 18:48.840
then it's more memory and we have to invoke those memories
18:48.840 --> 18:50.960
to say, oh, yes, well, now I can remember
18:50.960 --> 18:54.160
what I had for breakfast because I stored that someplace.
18:54.160 --> 18:58.200
I may forget it tomorrow, but I'd store it for now.
18:58.200 --> 19:01.600
So does memory also need to have,
19:02.880 --> 19:06.240
so the hierarchical aspect of reality
19:06.240 --> 19:07.720
is not just about concepts,
19:07.720 --> 19:08.800
it's also about time.
19:08.800 --> 19:10.280
Do you think of it that way?
19:10.280 --> 19:12.840
Yeah, time is infused in everything.
19:12.840 --> 19:15.560
It's like, you really can't separate it out.
19:15.560 --> 19:19.560
If I ask you, what is your, how's the brain
19:19.560 --> 19:21.360
learn a model of this coffee cup here?
19:21.360 --> 19:23.200
I have a coffee cup, then I met the coffee cup.
19:23.200 --> 19:26.000
I said, well, time is not an inherent property
19:26.000 --> 19:28.520
of the model I have of this cup,
19:28.520 --> 19:31.440
whether it's a visual model or tactile model.
19:31.440 --> 19:32.600
I can sense it through time,
19:32.600 --> 19:34.880
but the model itself doesn't really have much time.
19:34.880 --> 19:36.560
If I asked you, if I say, well,
19:36.560 --> 19:39.000
what is the model of my cell phone?
19:39.000 --> 19:41.480
My brain has learned a model of the cell phones.
19:41.480 --> 19:43.360
If you have a smartphone like this,
19:43.360 --> 19:45.680
and I said, well, this has time aspects to it.
19:45.680 --> 19:48.040
I have expectations when I turn it on,
19:48.040 --> 19:49.480
what's gonna happen, what water,
19:49.480 --> 19:51.960
how long it's gonna take to do certain things,
19:51.960 --> 19:54.040
if I bring up an app, what sequences,
19:54.040 --> 19:56.520
and so I have instant, it's like melodies in the world,
19:56.520 --> 19:58.560
you know, melody has a sense of time.
19:58.560 --> 20:01.200
So many things in the world move and act,
20:01.200 --> 20:03.720
and there's a sense of time related to them.
20:03.720 --> 20:08.280
Some don't, but most things do actually.
20:08.280 --> 20:12.120
So it's sort of infused throughout the models of the world.
20:12.120 --> 20:13.720
You build a model of the world,
20:13.720 --> 20:16.400
you're learning the structure of the objects in the world,
20:16.400 --> 20:17.840
and you're also learning
20:17.840 --> 20:19.760
how those things change through time.
20:20.760 --> 20:23.920
Okay, so it really is just a fourth dimension
20:23.920 --> 20:25.280
that's infused deeply,
20:25.280 --> 20:26.760
and you have to make sure
20:26.760 --> 20:30.960
that your models of intelligence incorporate it.
20:30.960 --> 20:34.840
So, like you mentioned, the state of neuroscience
20:34.840 --> 20:36.000
is deeply empirical.
20:36.000 --> 20:40.120
A lot of data collection, it's, you know,
20:40.120 --> 20:43.120
that's where it is, you mentioned Thomas Kuhn, right?
20:43.120 --> 20:44.560
Yeah.
20:44.560 --> 20:48.040
And then you're proposing a theory of intelligence,
20:48.040 --> 20:50.520
and which is really the next step,
20:50.520 --> 20:52.920
the really important step to take,
20:52.920 --> 20:57.920
but why is HTM, or what we'll talk about soon,
20:57.920 --> 21:01.160
the right theory?
21:01.160 --> 21:05.160
So is it more in this, is it backed by intuition,
21:05.160 --> 21:09.160
is it backed by evidence, is it backed by a mixture of both?
21:09.160 --> 21:12.800
Is it kind of closer to where string theory is in physics,
21:12.800 --> 21:15.800
where there's mathematical components
21:15.800 --> 21:18.160
which show that, you know what,
21:18.160 --> 21:20.160
it seems that this,
21:20.160 --> 21:23.560
it fits together too well for it not to be true,
21:23.560 --> 21:25.360
which is where string theory is.
21:25.360 --> 21:28.080
Is that where you're kind of thinking?
21:28.080 --> 21:30.080
It's a mixture of all those things,
21:30.080 --> 21:32.080
although definitely where we are right now,
21:32.080 --> 21:34.080
it's definitely much more on the empirical side
21:34.080 --> 21:36.080
than, let's say, string theory.
21:36.080 --> 21:39.080
The way this goes about, we're theorists, right?
21:39.080 --> 21:41.080
So we look at all this data,
21:41.080 --> 21:43.080
and we're trying to come up with some sort of model
21:43.080 --> 21:45.080
that explains it, basically,
21:45.080 --> 21:47.080
and there's, unlike string theory,
21:47.080 --> 21:50.080
there's vast more amounts of empirical data here
21:50.080 --> 21:54.080
than I think that most physicists deal with.
21:54.080 --> 21:57.080
And so our challenge is to sort through that
21:57.080 --> 22:01.080
and figure out what kind of constructs would explain this.
22:01.080 --> 22:04.080
And when we have an idea,
22:04.080 --> 22:06.080
you come up with a theory of some sort,
22:06.080 --> 22:08.080
you have lots of ways of testing it.
22:08.080 --> 22:10.080
First of all, I am, you know,
22:10.080 --> 22:14.080
there are 100 years of assimilated,
22:14.080 --> 22:16.080
und assimilated empirical data from neuroscience.
22:16.080 --> 22:18.080
So we go back and repapers, and we say,
22:18.080 --> 22:20.080
oh, did someone find this already?
22:20.080 --> 22:23.080
We can predict X, Y, and Z,
22:23.080 --> 22:25.080
and maybe no one's even talked about it
22:25.080 --> 22:27.080
since 1972 or something,
22:27.080 --> 22:29.080
but we go back and find that, and we say,
22:29.080 --> 22:31.080
oh, either it can support the theory
22:31.080 --> 22:33.080
or it can invalidate the theory.
22:33.080 --> 22:35.080
And then we say, okay, we have to start over again.
22:35.080 --> 22:37.080
Oh, no, it's support. Let's keep going with that one.
22:37.080 --> 22:40.080
So the way I kind of view it,
22:40.080 --> 22:43.080
when we do our work, we come up,
22:43.080 --> 22:45.080
we look at all this empirical data,
22:45.080 --> 22:47.080
and it's what I call it is a set of constraints.
22:47.080 --> 22:49.080
We're not interested in something that's biologically inspired.
22:49.080 --> 22:52.080
We're trying to figure out how the actual brain works.
22:52.080 --> 22:55.080
So every piece of empirical data is a constraint on a theory.
22:55.080 --> 22:57.080
If you have the correct theory,
22:57.080 --> 22:59.080
it needs to explain every pin, right?
22:59.080 --> 23:02.080
So we have this huge number of constraints on the problem,
23:02.080 --> 23:05.080
which initially makes it very, very difficult.
23:05.080 --> 23:07.080
If you don't have many constraints,
23:07.080 --> 23:09.080
you can make up stuff all the day.
23:09.080 --> 23:11.080
You can say, oh, here's an answer to how you can do this,
23:11.080 --> 23:13.080
you can do that, you can do this.
23:13.080 --> 23:15.080
But if you consider all biology as a set of constraints,
23:15.080 --> 23:17.080
all neuroscience as a set of constraints,
23:17.080 --> 23:19.080
and even if you're working in one little part of the Neocortex,
23:19.080 --> 23:21.080
for example, there are hundreds and hundreds of constraints.
23:21.080 --> 23:23.080
There are a lot of empirical constraints
23:23.080 --> 23:25.080
that it's very, very difficult initially
23:25.080 --> 23:27.080
to come up with a theoretical framework for that.
23:27.080 --> 23:31.080
But when you do, and it solves all those constraints at once,
23:31.080 --> 23:33.080
you have a high confidence
23:33.080 --> 23:36.080
that you got something close to correct.
23:36.080 --> 23:39.080
It's just mathematically almost impossible not to be.
23:39.080 --> 23:43.080
So that's the curse and the advantage of what we have.
23:43.080 --> 23:47.080
The curse is we have to meet all these constraints,
23:47.080 --> 23:49.080
which is really hard.
23:49.080 --> 23:51.080
But when you do meet them,
23:51.080 --> 23:53.080
then you have a great confidence
23:53.080 --> 23:55.080
that you've discovered something.
23:55.080 --> 23:58.080
In addition, then we work with scientific labs.
23:58.080 --> 24:00.080
So we'll say, oh, there's something we can't find,
24:00.080 --> 24:02.080
we can predict something,
24:02.080 --> 24:04.080
but we can't find it anywhere in the literature.
24:04.080 --> 24:07.080
So we will then, we have people we collaborated with,
24:07.080 --> 24:09.080
we'll say, sometimes they'll say, you know what,
24:09.080 --> 24:11.080
I have some collected data, which I didn't publish,
24:11.080 --> 24:13.080
but we can go back and look at it
24:13.080 --> 24:15.080
and see if we can find that,
24:15.080 --> 24:17.080
which is much easier than designing a new experiment.
24:17.080 --> 24:20.080
You know, neuroscience experiments take a long time, years.
24:20.080 --> 24:23.080
So although some people are doing that now too.
24:23.080 --> 24:27.080
So, but between all of these things,
24:27.080 --> 24:29.080
I think it's a reasonable,
24:29.080 --> 24:32.080
it's actually a very, very good approach.
24:32.080 --> 24:35.080
We are blessed with the fact that we can test our theories
24:35.080 --> 24:37.080
out to yin and yang here,
24:37.080 --> 24:39.080
because there's so much on a similar data,
24:39.080 --> 24:41.080
and we can also falsify our theories very easily,
24:41.080 --> 24:43.080
which we do often.
24:43.080 --> 24:46.080
So it's kind of reminiscent to whenever that was with Copernicus,
24:46.080 --> 24:49.080
you know, when you figure out that the sun is at the center,
24:49.080 --> 24:53.080
the solar system as opposed to Earth,
24:53.080 --> 24:55.080
the pieces just fall into place.
24:55.080 --> 24:59.080
Yeah, I think that's the general nature of the Ha moments,
24:59.080 --> 25:02.080
is in Copernicus, it could be,
25:02.080 --> 25:05.080
you could say the same thing about Darwin,
25:05.080 --> 25:07.080
you could say the same thing about, you know,
25:07.080 --> 25:09.080
about the double helix,
25:09.080 --> 25:13.080
that people have been working on a problem for so long,
25:13.080 --> 25:14.080
and have all this data,
25:14.080 --> 25:15.080
and they can't make sense of it, they can't make sense of it.
25:15.080 --> 25:17.080
But when the answer comes to you,
25:17.080 --> 25:19.080
and everything falls into place,
25:19.080 --> 25:21.080
it's like, oh my gosh, that's it.
25:21.080 --> 25:23.080
That's got to be right.
25:23.080 --> 25:28.080
I asked both Jim Watson and Francis Crick about this.
25:28.080 --> 25:30.080
I asked them, you know,
25:30.080 --> 25:33.080
when you were working on trying to discover the structure
25:33.080 --> 25:35.080
of the double helix,
25:35.080 --> 25:38.080
and when you came up with the sort of,
25:38.080 --> 25:42.080
the structure that ended up being correct,
25:42.080 --> 25:44.080
but it was sort of a guess, you know,
25:44.080 --> 25:46.080
it wasn't really verified yet.
25:46.080 --> 25:48.080
I said, did you know that it was right?
25:48.080 --> 25:50.080
And they both said, absolutely.
25:50.080 --> 25:52.080
We absolutely knew it was right.
25:52.080 --> 25:55.080
And it doesn't matter if other people didn't believe it or not,
25:55.080 --> 25:57.080
we knew it was right, they'd get around to thinking it
25:57.080 --> 25:59.080
and agree with it eventually anyway.
25:59.080 --> 26:01.080
And that's the kind of thing you hear a lot with scientists
26:01.080 --> 26:04.080
who really are studying a difficult problem,
26:04.080 --> 26:07.080
and I feel that way too, about our work.
26:07.080 --> 26:10.080
Have you talked to Crick or Watson about the problem
26:10.080 --> 26:15.080
you're trying to solve, the, of finding the DNA of the brain?
26:15.080 --> 26:16.080
Yeah.
26:16.080 --> 26:19.080
In fact, Francis Crick was very interested in this,
26:19.080 --> 26:21.080
in the latter part of his life.
26:21.080 --> 26:23.080
And in fact, I got interested in brains
26:23.080 --> 26:26.080
by reading an essay he wrote in 1979
26:26.080 --> 26:28.080
called Thinking About the Brain.
26:28.080 --> 26:30.080
And that was when I decided
26:30.080 --> 26:33.080
I'm going to leave my profession of computers and engineering
26:33.080 --> 26:35.080
and become a neuroscientist.
26:35.080 --> 26:37.080
Just reading that one essay from Francis Crick.
26:37.080 --> 26:39.080
I got to meet him later in life.
26:39.080 --> 26:43.080
I got to, I spoke at the Salk Institute
26:43.080 --> 26:44.080
and he was in the audience
26:44.080 --> 26:47.080
and then I had a tea with him afterwards.
26:47.080 --> 26:50.080
You know, he was interested in a different problem.
26:50.080 --> 26:52.080
He was focused on consciousness.
26:52.080 --> 26:54.080
The easy problem, right?
26:54.080 --> 26:58.080
Well, I think it's the red herring
26:58.080 --> 27:01.080
and so we weren't really overlapping a lot there.
27:01.080 --> 27:05.080
Jim Watson, who's still alive,
27:05.080 --> 27:07.080
is also interested in this problem
27:07.080 --> 27:11.080
and when he was director of the Coltsman Harbor Laboratories,
27:11.080 --> 27:13.080
he was really sort of behind
27:13.080 --> 27:16.080
moving in the direction of neuroscience there.
27:16.080 --> 27:19.080
And so he had a personal interest in this field
27:19.080 --> 27:23.080
and I have met with him numerous times.
27:23.080 --> 27:25.080
And in fact, the last time,
27:25.080 --> 27:27.080
a little bit over a year ago,
27:27.080 --> 27:30.080
I gave a talk at Coltsman Harbor Labs
27:30.080 --> 27:34.080
about the progress we were making in our work.
27:34.080 --> 27:39.080
And it was a lot of fun because he said,
27:39.080 --> 27:41.080
well, you wouldn't be coming here
27:41.080 --> 27:42.080
unless you had something important to say,
27:42.080 --> 27:44.080
so I'm going to go attend your talk.
27:44.080 --> 27:46.080
So he sat in the very front row.
27:46.080 --> 27:50.080
Next to him was the director of the lab, Bruce Stillman.
27:50.080 --> 27:52.080
So these guys were in the front row of this auditorium, right?
27:52.080 --> 27:54.080
So nobody else in the auditorium wants to sit in the front row
27:54.080 --> 27:57.080
because there's Jim Watson there as the director.
27:57.080 --> 28:03.080
And I gave a talk and then I had dinner with Jim afterwards.
28:03.080 --> 28:06.080
But there's a great picture of my colleague,
28:06.080 --> 28:08.080
Subitai Amantik, where I'm up there
28:08.080 --> 28:11.080
sort of expiring the basics of this new framework we have.
28:11.080 --> 28:13.080
And Jim Watson's on the edge of his chair.
28:13.080 --> 28:15.080
He's literally on the edge of his chair,
28:15.080 --> 28:17.080
like, internally staring up at the screen.
28:17.080 --> 28:21.080
And when he discovered the structure of DNA,
28:21.080 --> 28:25.080
the first public talk he gave was at Coltsman Harbor Labs.
28:25.080 --> 28:27.080
And there's a picture, there's a famous picture
28:27.080 --> 28:29.080
of Jim Watson standing at the whiteboard
28:29.080 --> 28:31.080
with an overhead thing pointing at something,
28:31.080 --> 28:33.080
pointing at the double helix at this pointer.
28:33.080 --> 28:35.080
And it actually looks a lot like the picture of me.
28:35.080 --> 28:37.080
So there was a sort of funny, there's an area talking about the brain
28:37.080 --> 28:39.080
and there's Jim Watson staring up at the tent.
28:39.080 --> 28:41.080
And of course, there was, you know, whatever,
28:41.080 --> 28:44.080
60 years earlier he was standing pointing at the double helix.
28:44.080 --> 28:47.080
It's one of the great discoveries in all of, you know,
28:47.080 --> 28:50.080
whatever, by all the science, all science and DNA.
28:50.080 --> 28:54.080
So it's the funny that there's echoes of that in your presentation.
28:54.080 --> 28:58.080
Do you think in terms of evolutionary timeline and history,
28:58.080 --> 29:01.080
the development of the neocortex was a big leap?
29:01.080 --> 29:06.080
Or is it just a small step?
29:06.080 --> 29:09.080
So, like, if we ran the whole thing over again,
29:09.080 --> 29:12.080
from the birth of life on Earth,
29:12.080 --> 29:15.080
how likely would we develop the mechanism of the neocortex?
29:15.080 --> 29:17.080
Okay, well, those are two separate questions.
29:17.080 --> 29:19.080
One, was it a big leap?
29:19.080 --> 29:21.080
And one was how likely it is, okay?
29:21.080 --> 29:23.080
They're not necessarily related.
29:23.080 --> 29:25.080
Maybe correlated.
29:25.080 --> 29:28.080
And we don't really have enough data to make a judgment about that.
29:28.080 --> 29:30.080
I would say definitely it was a big leap.
29:30.080 --> 29:31.080
And I can tell you why.
29:31.080 --> 29:34.080
I don't think it was just another incremental step.
29:34.080 --> 29:36.080
I'll get that in a moment.
29:36.080 --> 29:38.080
I don't really have any idea how likely it is.
29:38.080 --> 29:41.080
If we look at evolution, we have one data point,
29:41.080 --> 29:43.080
which is Earth, right?
29:43.080 --> 29:45.080
Life formed on Earth billions of years ago,
29:45.080 --> 29:48.080
whether it was introduced here or it created it here
29:48.080 --> 29:50.080
or someone introduced it we don't really know,
29:50.080 --> 29:51.080
but it was here early.
29:51.080 --> 29:55.080
It took a long, long time to get to multicellular life.
29:55.080 --> 29:58.080
And then from multicellular life,
29:58.080 --> 30:02.080
it took a long, long time to get the neocortex.
30:02.080 --> 30:05.080
And we've only had the neocortex for a few hundred thousand years.
30:05.080 --> 30:07.080
So that's like nothing.
30:07.080 --> 30:09.080
Okay, so is it likely?
30:09.080 --> 30:13.080
Well, certainly it isn't something that happened right away on Earth.
30:13.080 --> 30:15.080
And there were multiple steps to get there.
30:15.080 --> 30:17.080
So I would say it's probably not going to something that would happen
30:17.080 --> 30:20.080
instantaneously on other planets that might have life.
30:20.080 --> 30:23.080
It might take several billion years on average.
30:23.080 --> 30:24.080
Is it likely?
30:24.080 --> 30:25.080
I don't know.
30:25.080 --> 30:28.080
But you'd have to survive for several billion years to find out.
30:28.080 --> 30:29.080
Probably.
30:29.080 --> 30:30.080
Is it a big leap?
30:30.080 --> 30:35.080
Yeah, I think it is a qualitative difference
30:35.080 --> 30:38.080
in all other evolutionary steps.
30:38.080 --> 30:40.080
I can try to describe that if you'd like.
30:40.080 --> 30:42.080
Sure, in which way?
30:42.080 --> 30:44.080
Yeah, I can tell you how.
30:44.080 --> 30:48.080
Pretty much, let's start with a little preface.
30:48.080 --> 30:54.080
Maybe the things that humans are able to do do not have obvious
30:54.080 --> 30:59.080
survival advantages precedent.
30:59.080 --> 31:00.080
We create music.
31:00.080 --> 31:03.080
Is there a really survival advantage to that?
31:03.080 --> 31:04.080
Maybe, maybe not.
31:04.080 --> 31:05.080
What about mathematics?
31:05.080 --> 31:09.080
Is there a real survival advantage to mathematics?
31:09.080 --> 31:10.080
You can stretch it.
31:10.080 --> 31:13.080
You can try to figure these things out, right?
31:13.080 --> 31:18.080
But most of evolutionary history, everything had immediate survival
31:18.080 --> 31:19.080
advantages to it.
31:19.080 --> 31:22.080
I'll tell you a story, which I like.
31:22.080 --> 31:25.080
It may not be true.
31:25.080 --> 31:29.080
But the story goes as follows.
31:29.080 --> 31:34.080
Organisms have been evolving since the beginning of life here on Earth.
31:34.080 --> 31:37.080
Adding this sort of complexity onto that and this sort of complexity onto that.
31:37.080 --> 31:40.080
And the brain itself is evolved this way.
31:40.080 --> 31:44.080
There's an old part, an older part, an older, older part to the brain that kind of just
31:44.080 --> 31:47.080
keeps calming on new things and we keep adding capabilities.
31:47.080 --> 31:52.080
When we got to the neocortex, initially it had a very clear survival advantage
31:52.080 --> 31:56.080
in that it produced better vision and better hearing and better touch and maybe
31:56.080 --> 31:58.080
a new place and so on.
31:58.080 --> 32:04.080
But what I think happens is that evolution took a mechanism, and this is in our
32:04.080 --> 32:08.080
recent theory, but it took a mechanism that evolved a long time ago for
32:08.080 --> 32:10.080
navigating in the world, for knowing where you are.
32:10.080 --> 32:14.080
These are the so called grid cells and place cells of that old part of the brain.
32:14.080 --> 32:21.080
And it took that mechanism for building maps of the world and knowing where you are
32:21.080 --> 32:26.080
on those maps and how to navigate those maps and turns it into a sort of a slim
32:26.080 --> 32:29.080
down idealized version of it.
32:29.080 --> 32:32.080
And that idealized version could now apply to building maps of other things,
32:32.080 --> 32:36.080
maps of coffee cups and maps of phones, maps of mathematics.
32:36.080 --> 32:40.080
Concepts, yes, and not just almost, exactly.
32:40.080 --> 32:44.080
And it just started replicating this stuff.
32:44.080 --> 32:46.080
You just think more and more and more.
32:46.080 --> 32:51.080
So we went from being sort of dedicated purpose neural hardware to solve certain
32:51.080 --> 32:56.080
problems that are important to survival to a general purpose neural hardware
32:56.080 --> 33:02.080
that could be applied to all problems and now it's escaped the orbit of survival.
33:02.080 --> 33:08.080
It's, we are now able to apply it to things which we find enjoyment, you know,
33:08.080 --> 33:13.080
but aren't really clearly survival characteristics.
33:13.080 --> 33:19.080
And that it seems to only have happened in humans to the large extent.
33:19.080 --> 33:24.080
And so that's what's going on where we sort of have, we've sort of escaped the
33:24.080 --> 33:28.080
gravity of evolutionary pressure in some sense in the near cortex.
33:28.080 --> 33:32.080
And it now does things which are not, that are really interesting,
33:32.080 --> 33:36.080
discovering models of the universe, which may not really help us.
33:36.080 --> 33:37.080
It doesn't matter.
33:37.080 --> 33:41.080
How does it help us surviving knowing that there might be multiple verses or that
33:41.080 --> 33:44.080
there might be, you know, the age of the universe or how do, you know,
33:44.080 --> 33:46.080
various stellar things occur?
33:46.080 --> 33:47.080
It doesn't really help us survive at all.
33:47.080 --> 33:50.080
But we enjoy it and that's what happened.
33:50.080 --> 33:53.080
Or at least not in the obvious way, perhaps.
33:53.080 --> 33:58.080
It is required, if you look at the entire universe in an evolutionary way,
33:58.080 --> 34:03.080
it's required for us to do interplanetary travel and therefore survive past our own fun.
34:03.080 --> 34:05.080
But you know, let's not get too quick.
34:05.080 --> 34:07.080
Yeah, but, you know, evolution works at one time frame.
34:07.080 --> 34:11.080
It's survival, if you think of survival of the phenotype,
34:11.080 --> 34:13.080
survival of the individual.
34:13.080 --> 34:16.080
What you're talking about there is spans well beyond that.
34:16.080 --> 34:22.080
So there's no genetic, I'm not transferring any genetic traits to my children.
34:22.080 --> 34:25.080
That are going to help them survive better on Mars.
34:25.080 --> 34:27.080
Totally different mechanism.
34:27.080 --> 34:32.080
So let's get into the new, as you've mentioned, this idea,
34:32.080 --> 34:35.080
I don't know if you have a nice name, thousand.
34:35.080 --> 34:37.080
We call it the thousand brain theory of intelligence.
34:37.080 --> 34:38.080
I like it.
34:38.080 --> 34:44.080
So can you talk about this idea of spatial view of concepts and so on?
34:44.080 --> 34:45.080
Yeah.
34:45.080 --> 34:49.080
So can I just describe sort of the, there's an underlying core discovery,
34:49.080 --> 34:51.080
which then everything comes from that.
34:51.080 --> 34:55.080
That's a very simple, this is really what happened.
34:55.080 --> 35:00.080
We were deep into problems about understanding how we build models of stuff in the world
35:00.080 --> 35:03.080
and how we make predictions about things.
35:03.080 --> 35:07.080
And I was holding a coffee cup just like this in my hand.
35:07.080 --> 35:10.080
And I had my finger was touching the side, my index finger.
35:10.080 --> 35:15.080
And then I moved it to the top and I was going to feel the rim at the top of the cup.
35:15.080 --> 35:18.080
And I asked myself a very simple question.
35:18.080 --> 35:22.080
I said, well, first of all, let's say I know that my brain predicts what it's going to feel
35:22.080 --> 35:23.080
before it touches it.
35:23.080 --> 35:25.080
You can just think about it and imagine it.
35:25.080 --> 35:28.080
And so we know that the brain's making predictions all the time.
35:28.080 --> 35:31.080
So the question is, what does it take to predict that?
35:31.080 --> 35:33.080
And there's a very interesting answer.
35:33.080 --> 35:36.080
First of all, it says the brain has to know it's touching a coffee cup.
35:36.080 --> 35:38.080
It has to have a model of a coffee cup.
35:38.080 --> 35:43.080
It needs to know where the finger currently is on the cup, relative to the cup.
35:43.080 --> 35:46.080
Because when I make a movement, it needs to know where it's going to be on the cup
35:46.080 --> 35:50.080
after the movement is completed, relative to the cup.
35:50.080 --> 35:53.080
And then it can make a prediction about what it's going to sense.
35:53.080 --> 35:56.080
So this told me that the neocortex, which is making this prediction,
35:56.080 --> 35:59.080
needs to know that it's sensing it's touching a cup.
35:59.080 --> 36:02.080
And it needs to know the location of my finger relative to that cup
36:02.080 --> 36:04.080
in a reference frame of the cup.
36:04.080 --> 36:06.080
It doesn't matter where the cup is relative to my body.
36:06.080 --> 36:08.080
It doesn't matter its orientation.
36:08.080 --> 36:09.080
None of that matters.
36:09.080 --> 36:13.080
It's where my finger is relative to the cup, which tells me then that the neocortex
36:13.080 --> 36:17.080
has a reference frame that's anchored to the cup.
36:17.080 --> 36:19.080
Because otherwise, I wouldn't be able to say the location
36:19.080 --> 36:21.080
and I wouldn't be able to predict my new location.
36:21.080 --> 36:24.080
And then we quickly, very instantly, you can say,
36:24.080 --> 36:26.080
well, every part of my skin could touch this cup
36:26.080 --> 36:28.080
and therefore every part of my skin is making predictions
36:28.080 --> 36:30.080
and every part of my skin must have a reference frame
36:30.080 --> 36:33.080
that it's using to make predictions.
36:33.080 --> 36:39.080
So the big idea is that throughout the neocortex,
36:39.080 --> 36:47.080
there are, everything is being stored and referenced in reference frames.
36:47.080 --> 36:49.080
You can think of them like XYZ reference frames,
36:49.080 --> 36:50.080
but they're not like that.
36:50.080 --> 36:52.080
We know a lot about the neural mechanisms for this.
36:52.080 --> 36:55.080
But the brain thinks in reference frames.
36:55.080 --> 36:58.080
And as an engineer, if you're an engineer, this is not surprising.
36:58.080 --> 37:01.080
You'd say, if I were to build a CAD model of the coffee cup,
37:01.080 --> 37:03.080
well, I would bring it up in some CAD software
37:03.080 --> 37:05.080
and I would assign some reference frame and say,
37:05.080 --> 37:07.080
this features at this location and so on.
37:07.080 --> 37:10.080
But the fact that this, the idea that this is occurring
37:10.080 --> 37:14.080
throughout the neocortex everywhere, it was a novel idea.
37:14.080 --> 37:20.080
And then a zillion things fell into place after that, a zillion.
37:20.080 --> 37:23.080
So now we think about the neocortex as processing information
37:23.080 --> 37:25.080
quite differently than we used to do it.
37:25.080 --> 37:28.080
We used to think about the neocortex as processing sensory data
37:28.080 --> 37:30.080
and extracting features from that sensory data
37:30.080 --> 37:32.080
and then extracting features from the features
37:32.080 --> 37:35.080
very much like a deep learning network does today.
37:35.080 --> 37:36.080
But that's not how the brain works at all.
37:36.080 --> 37:39.080
The brain works by assigning everything,
37:39.080 --> 37:41.080
every input, everything to reference frames,
37:41.080 --> 37:44.080
and there are thousands, hundreds of thousands of them
37:44.080 --> 37:47.080
active at once in your neocortex.
37:47.080 --> 37:49.080
It's a surprising thing to think about,
37:49.080 --> 37:51.080
but once you sort of internalize this,
37:51.080 --> 37:54.080
you understand that it explains almost every,
37:54.080 --> 37:57.080
almost all the mysteries we've had about this structure.
37:57.080 --> 38:00.080
So one of the consequences of that is that
38:00.080 --> 38:04.080
every small part of the neocortex, say a millimeter square,
38:04.080 --> 38:06.080
and there's 150,000 of those.
38:06.080 --> 38:08.080
So it's about 150,000 square millimeters.
38:08.080 --> 38:11.080
If you take every little square millimeter of the cortex,
38:11.080 --> 38:13.080
it's got some input coming into it,
38:13.080 --> 38:15.080
and it's going to have reference frames
38:15.080 --> 38:17.080
where it's assigning that input to.
38:17.080 --> 38:21.080
And each square millimeter can learn complete models of objects.
38:21.080 --> 38:22.080
So what do I mean by that?
38:22.080 --> 38:23.080
If I'm touching the coffee cup,
38:23.080 --> 38:25.080
well, if I just touch it in one place,
38:25.080 --> 38:27.080
I can't learn what this coffee cup is
38:27.080 --> 38:29.080
because I'm just feeling one part.
38:29.080 --> 38:32.080
But if I move it around the cup and touch it in different areas,
38:32.080 --> 38:34.080
I can build up a complete model of the cup
38:34.080 --> 38:36.080
because I'm now filling in that three dimensional map,
38:36.080 --> 38:37.080
which is the coffee cup.
38:37.080 --> 38:39.080
I can say, oh, what am I feeling in all these different locations?
38:39.080 --> 38:40.080
That's the basic idea.
38:40.080 --> 38:42.080
It's more complicated than that.
38:42.080 --> 38:46.080
But so through time, and we talked about time earlier,
38:46.080 --> 38:48.080
through time, even a single column,
38:48.080 --> 38:50.080
which is only looking at, or a single part of the cortex,
38:50.080 --> 38:52.080
which is only looking at a small part of the world,
38:52.080 --> 38:54.080
can build up a complete model of an object.
38:54.080 --> 38:57.080
And so if you think about the part of the brain,
38:57.080 --> 38:59.080
which is getting input from all my fingers,
38:59.080 --> 39:01.080
so they're spread across the top of your head here.
39:01.080 --> 39:03.080
This is the somatosensory cortex.
39:03.080 --> 39:07.080
There's columns associated with all the different areas of my skin.
39:07.080 --> 39:10.080
And what we believe is happening is that
39:10.080 --> 39:12.080
all of them are building models of this cup,
39:12.080 --> 39:15.080
every one of them, or things.
39:15.080 --> 39:18.080
Not every column or every part of the cortex
39:18.080 --> 39:19.080
builds models of everything,
39:19.080 --> 39:21.080
but they're all building models of something.
39:21.080 --> 39:26.080
And so when I touch this cup with my hand,
39:26.080 --> 39:29.080
there are multiple models of the cup being invoked.
39:29.080 --> 39:30.080
If I look at it with my eyes,
39:30.080 --> 39:32.080
there are again many models of the cup being invoked,
39:32.080 --> 39:34.080
because each part of the visual system,
39:34.080 --> 39:36.080
the brain doesn't process an image.
39:36.080 --> 39:38.080
That's a misleading idea.
39:38.080 --> 39:40.080
It's just like your fingers touching the cup,
39:40.080 --> 39:43.080
so different parts of my retina are looking at different parts of the cup.
39:43.080 --> 39:45.080
And thousands and thousands of models of the cup
39:45.080 --> 39:47.080
are being invoked at once.
39:47.080 --> 39:49.080
And they're all voting with each other,
39:49.080 --> 39:50.080
trying to figure out what's going on.
39:50.080 --> 39:52.080
So that's why we call it the thousand brains theory of intelligence,
39:52.080 --> 39:54.080
because there isn't one model of a cup.
39:54.080 --> 39:56.080
There are thousands of models of this cup.
39:56.080 --> 39:58.080
There are thousands of models of your cell phone,
39:58.080 --> 40:01.080
and about cameras and microphones and so on.
40:01.080 --> 40:03.080
It's a distributed modeling system,
40:03.080 --> 40:05.080
which is very different than what people have thought about it.
40:05.080 --> 40:07.080
So that's a really compelling and interesting idea.
40:07.080 --> 40:09.080
I have two first questions.
40:09.080 --> 40:12.080
So one, on the ensemble part of everything coming together,
40:12.080 --> 40:14.080
you have these thousand brains.
40:14.080 --> 40:19.080
How do you know which one has done the best job of forming the cup?
40:19.080 --> 40:20.080
Great question. Let me try to explain.
40:20.080 --> 40:23.080
There's a problem that's known in neuroscience
40:23.080 --> 40:25.080
called the sensor fusion problem.
40:25.080 --> 40:26.080
Yes.
40:26.080 --> 40:28.080
And so the idea is something like,
40:28.080 --> 40:29.080
oh, the image comes from the eye.
40:29.080 --> 40:30.080
There's a picture on the retina.
40:30.080 --> 40:32.080
And it gets projected to the neocortex.
40:32.080 --> 40:35.080
Oh, by now it's all sped out all over the place,
40:35.080 --> 40:37.080
and it's kind of squirrely and distorted,
40:37.080 --> 40:39.080
and pieces are all over the, you know,
40:39.080 --> 40:41.080
it doesn't look like a picture anymore.
40:41.080 --> 40:43.080
When does it all come back together again?
40:43.080 --> 40:44.080
Right?
40:44.080 --> 40:46.080
Or you might say, well, yes, but I also,
40:46.080 --> 40:48.080
I also have sounds or touches associated with the cup.
40:48.080 --> 40:50.080
So I'm seeing the cup and touching the cup.
40:50.080 --> 40:52.080
How do they get combined together again?
40:52.080 --> 40:54.080
So this is called the sensor fusion problem.
40:54.080 --> 40:57.080
As if all these disparate parts have to be brought together
40:57.080 --> 40:59.080
into one model someplace.
40:59.080 --> 41:01.080
That's the wrong idea.
41:01.080 --> 41:03.080
The right idea is that you get all these guys voting.
41:03.080 --> 41:05.080
There's auditory models of the cup,
41:05.080 --> 41:07.080
there's visual models of the cup,
41:07.080 --> 41:09.080
there's tactile models of the cup.
41:09.080 --> 41:11.080
In the vision system, there might be ones
41:11.080 --> 41:13.080
that are more focused on black and white,
41:13.080 --> 41:14.080
ones versioned on color.
41:14.080 --> 41:15.080
It doesn't really matter.
41:15.080 --> 41:17.080
There's just thousands and thousands of models of this cup.
41:17.080 --> 41:18.080
And they vote.
41:18.080 --> 41:20.080
They don't actually come together in one spot.
41:20.080 --> 41:22.080
Just literally think of it this way.
41:22.080 --> 41:25.080
Imagine you have, each columns are like about the size
41:25.080 --> 41:26.080
of a little piece of spaghetti.
41:26.080 --> 41:27.080
Okay?
41:27.080 --> 41:28.080
Like a two and a half millimeters tall
41:28.080 --> 41:30.080
and about a millimeter in white.
41:30.080 --> 41:33.080
They're not physical like, but you can think of them that way.
41:33.080 --> 41:36.080
And each one's trying to guess what this thing is or touching.
41:36.080 --> 41:38.080
Now they can, they can do a pretty good job
41:38.080 --> 41:40.080
if they're allowed to move over time.
41:40.080 --> 41:42.080
So I can reach my hand into a black box and move my finger
41:42.080 --> 41:44.080
around an object and if I touch enough space,
41:44.080 --> 41:46.080
it's like, okay, I know what it is.
41:46.080 --> 41:48.080
But often we don't do that.
41:48.080 --> 41:50.080
Often I can just reach and grab something with my hand
41:50.080 --> 41:51.080
all at once and I get it.
41:51.080 --> 41:53.080
Or if I had to look through the world through a straw,
41:53.080 --> 41:55.080
so I'm only invoking one little column,
41:55.080 --> 41:57.080
I can only see part of something because I have to move
41:57.080 --> 41:58.080
the straw around.
41:58.080 --> 42:00.080
But if I open my eyes to see the whole thing at once.
42:00.080 --> 42:02.080
So what we think is going on is all these little pieces
42:02.080 --> 42:05.080
of spaghetti, all these little columns in the cortex
42:05.080 --> 42:08.080
are all trying to guess what it is that they're sensing.
42:08.080 --> 42:10.080
They'll do a better guess if they have time
42:10.080 --> 42:11.080
and can move over time.
42:11.080 --> 42:13.080
So if I move my eyes and move my fingers.
42:13.080 --> 42:16.080
But if they don't, they have a, they have a poor guess.
42:16.080 --> 42:19.080
It's a, it's a probabilistic guess of what they might be touching.
42:19.080 --> 42:22.080
Now imagine they can post their probability
42:22.080 --> 42:24.080
at the top of little piece of spaghetti.
42:24.080 --> 42:25.080
Each one of them says, I think,
42:25.080 --> 42:27.080
and it's not really a probability distribution.
42:27.080 --> 42:29.080
It's more like a set of possibilities in the brain.
42:29.080 --> 42:31.080
It doesn't work as a probability distribution.
42:31.080 --> 42:33.080
It works as more like what we call a union.
42:33.080 --> 42:35.080
You could say, and one column says,
42:35.080 --> 42:39.080
I think it could be a coffee cup, a soda can or a water bottle.
42:39.080 --> 42:42.080
And another column says, I think it could be a coffee cup
42:42.080 --> 42:45.080
or a, you know, telephone or camera or whatever.
42:45.080 --> 42:46.080
Right.
42:46.080 --> 42:49.080
And all these guys are saying what they think it might be.
42:49.080 --> 42:51.080
And there's these long range connections
42:51.080 --> 42:53.080
in certain layers in the cortex.
42:53.080 --> 42:57.080
So there's some layers in some cell types in each column
42:57.080 --> 42:59.080
send the projections across the brain.
42:59.080 --> 43:01.080
And that's the voting occurs.
43:01.080 --> 43:04.080
And so there's a simple associative memory mechanism.
43:04.080 --> 43:07.080
We've described this in a recent paper and we've modeled this
43:07.080 --> 43:11.080
that says they can all quickly settle on the only
43:11.080 --> 43:14.080
or the one best answer for all of them.
43:14.080 --> 43:17.080
If there is a single best answer, they all vote and say,
43:17.080 --> 43:19.080
yep, it's got to be the coffee cup.
43:19.080 --> 43:21.080
And at that point, they all know it's a coffee cup.
43:21.080 --> 43:23.080
And at that point, everyone acts as if it's a coffee cup.
43:23.080 --> 43:24.080
Yeah, we know it's a coffee.
43:24.080 --> 43:26.080
Even though I've only seen one little piece of this world,
43:26.080 --> 43:28.080
I know it's a coffee cup I'm touching or I'm seeing or whatever.
43:28.080 --> 43:31.080
And so you can think of all these columns are looking
43:31.080 --> 43:33.080
at different parts and different places,
43:33.080 --> 43:35.080
different sensory input, different locations.
43:35.080 --> 43:36.080
They're all different.
43:36.080 --> 43:40.080
But this layer that's doing the voting, it solidifies.
43:40.080 --> 43:43.080
It crystallizes and says, oh, we all know what we're doing.
43:43.080 --> 43:46.080
And so you don't bring these models together in one model,
43:46.080 --> 43:49.080
you just vote and there's a crystallization of the vote.
43:49.080 --> 43:50.080
Great.
43:50.080 --> 43:56.080
That's at least a compelling way to think about the way you
43:56.080 --> 43:58.080
form a model of the world.
43:58.080 --> 44:00.080
Now, you talk about a coffee cup.
44:00.080 --> 44:04.080
Do you see this as far as I understand that you were proposing
44:04.080 --> 44:07.080
this as well, that this extends to much more than coffee cups?
44:07.080 --> 44:09.080
Yeah, it does.
44:09.080 --> 44:11.080
Or at least the physical world.
44:11.080 --> 44:14.080
It expands to the world of concepts.
44:14.080 --> 44:15.080
Yeah, it does.
44:15.080 --> 44:18.080
And well, the first, the primary phase of evidence for that
44:18.080 --> 44:21.080
is that the regions of the neocortex that are associated
44:21.080 --> 44:24.080
with language or high level thought or mathematics or things
44:24.080 --> 44:26.080
like that, they look like the regions of the neocortex
44:26.080 --> 44:28.080
that process vision and hearing and touch.
44:28.080 --> 44:31.080
They don't look any different or they look only marginally
44:31.080 --> 44:32.080
different.
44:32.080 --> 44:36.080
And so one would say, well, if Vernon Mountcastle,
44:36.080 --> 44:39.080
who proposed that all the parts of the neocortex
44:39.080 --> 44:42.080
are the same thing, if he's right, then the parts
44:42.080 --> 44:44.080
that are doing language or mathematics or physics
44:44.080 --> 44:46.080
are working on the same principle.
44:46.080 --> 44:48.080
They must be working on the principle of reference frames.
44:48.080 --> 44:51.080
So that's a little odd thought.
44:51.080 --> 44:55.080
But of course, we had no prior idea how these things happen.
44:55.080 --> 44:57.080
So let's go with that.
44:57.080 --> 45:01.080
And in our recent paper, we talked a little bit about that.
45:01.080 --> 45:03.080
I've been working on it more since.
45:03.080 --> 45:05.080
I have better ideas about it now.
45:05.080 --> 45:08.080
I'm sitting here very confident that that's what's happening.
45:08.080 --> 45:11.080
And I can give you some examples to help you think about that.
45:11.080 --> 45:13.080
It's not that we understand it completely,
45:13.080 --> 45:15.080
but I understand it better than I've described it in any paper
45:15.080 --> 45:16.080
so far.
45:16.080 --> 45:18.080
But we did put that idea out there.
45:18.080 --> 45:22.080
It's a good place to start.
45:22.080 --> 45:25.080
And the evidence would suggest it's how it's happening.
45:25.080 --> 45:27.080
And then we can start tackling that problem one piece at a time.
45:27.080 --> 45:29.080
What does it mean to do high level thought?
45:29.080 --> 45:30.080
What does it mean to do language?
45:30.080 --> 45:34.080
How would that fit into a reference framework?
45:34.080 --> 45:38.080
I don't know if you could tell me if there's a connection,
45:38.080 --> 45:42.080
but there's an app called Anki that helps you remember different concepts.
45:42.080 --> 45:46.080
And they talk about like a memory palace that helps you remember
45:46.080 --> 45:50.080
completely random concepts by trying to put them in a physical space
45:50.080 --> 45:52.080
in your mind and putting them next to each other.
45:52.080 --> 45:54.080
It's called the method of loci.
45:54.080 --> 45:57.080
For some reason, that seems to work really well.
45:57.080 --> 46:00.080
Now that's a very narrow kind of application of just remembering some facts.
46:00.080 --> 46:03.080
But that's a very, very telling one.
46:03.080 --> 46:04.080
Yes, exactly.
46:04.080 --> 46:09.080
So this seems like you're describing a mechanism why this seems to work.
46:09.080 --> 46:13.080
So basically the way what we think is going on is all things you know,
46:13.080 --> 46:17.080
all concepts, all ideas, words, everything, you know,
46:17.080 --> 46:20.080
are stored in reference frames.
46:20.080 --> 46:24.080
And so if you want to remember something,
46:24.080 --> 46:27.080
you have to basically navigate through a reference frame the same way
46:27.080 --> 46:28.080
a rat navigates to a man.
46:28.080 --> 46:31.080
Even the same way my finger rat navigates to this coffee cup.
46:31.080 --> 46:33.080
You are moving through some space.
46:33.080 --> 46:37.080
And so if you have a random list of things you would ask to remember
46:37.080 --> 46:39.080
by assigning them to a reference frame,
46:39.080 --> 46:42.080
you've already know very well to see your house, right?
46:42.080 --> 46:44.080
And the idea of the method of loci is you can say,
46:44.080 --> 46:46.080
okay, in my lobby, I'm going to put this thing.
46:46.080 --> 46:48.080
And then the bedroom, I put this one.
46:48.080 --> 46:49.080
I go down the hall, I put this thing.
46:49.080 --> 46:51.080
And then you want to recall those facts.
46:51.080 --> 46:52.080
So recall those things.
46:52.080 --> 46:53.080
You just walk mentally.
46:53.080 --> 46:54.080
You walk through your house.
46:54.080 --> 46:57.080
You're mentally moving through a reference frame that you already had.
46:57.080 --> 47:00.080
And that tells you there's two things that are really important about that.
47:00.080 --> 47:03.080
It tells us the brain prefers to store things in reference frames.
47:03.080 --> 47:08.080
And the method of recalling things or thinking, if you will,
47:08.080 --> 47:11.080
is to move mentally through those reference frames.
47:11.080 --> 47:13.080
You could move physically through some reference frames,
47:13.080 --> 47:16.080
like I could physically move through the reference frame of this coffee cup.
47:16.080 --> 47:18.080
I can also mentally move through the reference frame of the coffee cup,
47:18.080 --> 47:19.080
imagining me touching it.
47:19.080 --> 47:22.080
But I can also mentally move my house.
47:22.080 --> 47:26.080
And so now we can ask ourselves, are all concepts stored this way?
47:26.080 --> 47:32.080
There was some recent research using human subjects in fMRI.
47:32.080 --> 47:36.080
And I'm going to apologize for not knowing the name of the scientists who did this.
47:36.080 --> 47:41.080
But what they did is they put humans in this fMRI machine,
47:41.080 --> 47:42.080
which was one of these imaging machines.
47:42.080 --> 47:46.080
And they gave the humans tasks to think about birds.
47:46.080 --> 47:49.080
So they had different types of birds, and birds that looked big and small
47:49.080 --> 47:51.080
and long necks and long legs, things like that.
47:51.080 --> 47:56.080
And what they could tell from the fMRI was a very clever experiment.
47:56.080 --> 48:00.080
You get to tell when humans were thinking about the birds,
48:00.080 --> 48:05.080
that the birds, the knowledge of birds was arranged in a reference frame
48:05.080 --> 48:08.080
similar to the ones that are used when you navigate in a room.
48:08.080 --> 48:10.080
These are called grid cells.
48:10.080 --> 48:14.080
And there are grid cell like patterns of activity in the neocortex when they do this.
48:14.080 --> 48:18.080
So that, it's a very clever experiment.
48:18.080 --> 48:22.080
And what it basically says is that even when you're thinking about something abstract
48:22.080 --> 48:24.080
and you're not really thinking about it as a reference frame,
48:24.080 --> 48:27.080
it tells us the brain is actually using a reference frame.
48:27.080 --> 48:29.080
And it's using the same neural mechanisms.
48:29.080 --> 48:32.080
These grid cells are the basic same neural mechanisms that we propose
48:32.080 --> 48:36.080
that grid cells, which exist in the old part of the brain, the entomonic cortex,
48:36.080 --> 48:40.080
that that mechanism is now similar mechanism, is used throughout the neocortex.
48:40.080 --> 48:44.080
It's the same nature to preserve this interesting way of creating reference frames.
48:44.080 --> 48:49.080
And so now they have empirical evidence that when you think about concepts like birds
48:49.080 --> 48:53.080
that you're using reference frames that are built on grid cells.
48:53.080 --> 48:55.080
So that's similar to the method of loci.
48:55.080 --> 48:57.080
But in this case, the birds are related so that makes,
48:57.080 --> 49:01.080
they create their own reference frame, which is consistent with bird space.
49:01.080 --> 49:03.080
And when you think about something, you go through that.
49:03.080 --> 49:04.080
You can make the same example.
49:04.080 --> 49:06.080
Let's take a math mathematics.
49:06.080 --> 49:08.080
Let's say you want to prove a conjecture.
49:08.080 --> 49:09.080
Okay.
49:09.080 --> 49:10.080
What is a conjecture?
49:10.080 --> 49:13.080
A conjecture is a statement you believe to be true,
49:13.080 --> 49:15.080
but you haven't proven it.
49:15.080 --> 49:17.080
And so it might be an equation.
49:17.080 --> 49:19.080
I want to show that this is equal to that.
49:19.080 --> 49:21.080
And you have some places you start with.
49:21.080 --> 49:23.080
You say, well, I know this is true and I know this is true.
49:23.080 --> 49:26.080
And I think that maybe to get to the final proof,
49:26.080 --> 49:28.080
I need to go through some intermediate results.
49:28.080 --> 49:33.080
What I believe is happening is literally these equations
49:33.080 --> 49:36.080
or these points are assigned to a reference frame,
49:36.080 --> 49:38.080
a mathematical reference frame.
49:38.080 --> 49:40.080
And when you do mathematical operations,
49:40.080 --> 49:42.080
a simple one might be multiply or divide,
49:42.080 --> 49:44.080
maybe a little plus transform or something else.
49:44.080 --> 49:47.080
That is like a movement in the reference frame of the math.
49:47.080 --> 49:50.080
And so you're literally trying to discover a path
49:50.080 --> 49:56.080
from one location to another location in a space of mathematics.
49:56.080 --> 49:58.080
And if you can get to these intermediate results,
49:58.080 --> 50:00.080
then you know your map is pretty good
50:00.080 --> 50:03.080
and you know you're using the right operations.
50:03.080 --> 50:06.080
Much of what we think about is solving hard problems
50:06.080 --> 50:09.080
is designing the correct reference frame for that problem,
50:09.080 --> 50:12.080
how to organize the information, and what behaviors
50:12.080 --> 50:15.080
I want to use in that space to get me there.
50:15.080 --> 50:19.080
Yeah, so if you dig in on an idea of this reference frame,
50:19.080 --> 50:21.080
whether it's the math, you start a set of axioms
50:21.080 --> 50:24.080
to try to get to proving the conjecture.
50:24.080 --> 50:27.080
Can you try to describe, maybe take a step back,
50:27.080 --> 50:30.080
how you think of the reference frame in that context?
50:30.080 --> 50:35.080
Is it the reference frame that the axioms are happy in?
50:35.080 --> 50:38.080
Is it the reference frame that might contain everything?
50:38.080 --> 50:41.080
Is it a changing thing as you...
50:41.080 --> 50:43.080
You have many, many reference frames.
50:43.080 --> 50:45.080
In fact, the way the thousand brain theories of intelligence
50:45.080 --> 50:48.080
says that every single thing in the world has its own reference frame.
50:48.080 --> 50:50.080
So every word has its own reference frames.
50:50.080 --> 50:52.080
And we can talk about this.
50:52.080 --> 50:55.080
The mathematics work out this is no problem for neurons to do this.
50:55.080 --> 50:58.080
But how many reference frames does the coffee cup have?
50:58.080 --> 51:03.080
Well, let's say you ask how many reference frames
51:03.080 --> 51:07.080
could the column in my finger that's touching the coffee cup have
51:07.080 --> 51:10.080
because there are many, many models of the coffee cup.
51:10.080 --> 51:12.080
So there is no model of the coffee cup.
51:12.080 --> 51:14.080
There are many models of the coffee cup.
51:14.080 --> 51:17.080
And you can say, well, how many different things can my finger learn?
51:17.080 --> 51:19.080
Is this the question you want to ask?
51:19.080 --> 51:21.080
Imagine I say every concept, every idea,
51:21.080 --> 51:23.080
everything you've ever know about that you can say,
51:23.080 --> 51:28.080
I know that thing has a reference frame associated with it.
51:28.080 --> 51:30.080
And what we do when we build composite objects,
51:30.080 --> 51:34.080
we assign reference frames to point another reference frame.
51:34.080 --> 51:37.080
So my coffee cup has multiple components to it.
51:37.080 --> 51:38.080
It's got a limb.
51:38.080 --> 51:39.080
It's got a cylinder.
51:39.080 --> 51:40.080
It's got a handle.
51:40.080 --> 51:43.080
And those things have their own reference frames.
51:43.080 --> 51:45.080
And they're assigned to a master reference frame,
51:45.080 --> 51:46.080
which is called this cup.
51:46.080 --> 51:48.080
And now I have this mental logo on it.
51:48.080 --> 51:50.080
Well, that's something that exists elsewhere in the world.
51:50.080 --> 51:51.080
It's its own thing.
51:51.080 --> 51:52.080
So it has its own reference frame.
51:52.080 --> 51:56.080
So we now have to say, how can I assign the mental logo reference frame
51:56.080 --> 51:59.080
onto the cylinder or onto the coffee cup?
51:59.080 --> 52:04.080
So we talked about this in the paper that came out in December
52:04.080 --> 52:06.080
of this last year.
52:06.080 --> 52:09.080
The idea of how you can assign reference frames to reference frames,
52:09.080 --> 52:10.080
how neurons could do this.
52:10.080 --> 52:14.080
So my question is, even though you mentioned reference frames a lot,
52:14.080 --> 52:18.080
I almost feel it's really useful to dig into how you think
52:18.080 --> 52:20.080
of what a reference frame is.
52:20.080 --> 52:22.080
It was already helpful for me to understand that you think
52:22.080 --> 52:26.080
of reference frames as something there is a lot of.
52:26.080 --> 52:29.080
OK, so let's just say that we're going to have some neurons
52:29.080 --> 52:32.080
in the brain, not many actually, 10,000, 20,000,
52:32.080 --> 52:34.080
are going to create a whole bunch of reference frames.
52:34.080 --> 52:35.080
What does it mean?
52:35.080 --> 52:37.080
What is a reference frame?
52:37.080 --> 52:40.080
First of all, these reference frames are different than the ones
52:40.080 --> 52:42.080
you might be used to.
52:42.080 --> 52:43.080
We know lots of reference frames.
52:43.080 --> 52:45.080
For example, we know the Cartesian coordinates,
52:45.080 --> 52:47.080
XYZ, that's a type of reference frame.
52:47.080 --> 52:50.080
We know longitude and latitude.
52:50.080 --> 52:52.080
That's a different type of reference frame.
52:52.080 --> 52:55.080
If I look at a printed map, it might have columns,
52:55.080 --> 52:59.080
A through M and rows, 1 through 20,
52:59.080 --> 53:01.080
that's a different type of reference frame.
53:01.080 --> 53:04.080
It's kind of a Cartesian reference frame.
53:04.080 --> 53:07.080
The interesting thing about the reference frames in the brain,
53:07.080 --> 53:09.080
and we know this because these have been established
53:09.080 --> 53:12.080
through neuroscience studying the entorhinal cortex.
53:12.080 --> 53:13.080
So I'm not speculating here.
53:13.080 --> 53:16.080
This is known neuroscience in an old part of the brain.
53:16.080 --> 53:18.080
The way these cells create reference frames,
53:18.080 --> 53:20.080
they have no origin.
53:20.080 --> 53:24.080
So what it's more like, you have a point,
53:24.080 --> 53:26.080
a point in some space,
53:26.080 --> 53:29.080
and you, given a particular movement,
53:29.080 --> 53:32.080
you can then tell what the next point should be.
53:32.080 --> 53:34.080
And you can then tell what the next point would be.
53:34.080 --> 53:35.080
And so on.
53:35.080 --> 53:40.080
You can use this to calculate how to get from one point to another.
53:40.080 --> 53:43.080
So how do I get from my house to my home,
53:43.080 --> 53:45.080
or how do I get my finger from the side of my cup
53:45.080 --> 53:46.080
to the top of the cup?
53:46.080 --> 53:52.080
How do I get from the axioms to the conjecture?
53:52.080 --> 53:54.080
So it's a different type of reference frame.
53:54.080 --> 53:57.080
And I can, if you want, I can describe in more detail.
53:57.080 --> 53:59.080
I can paint a picture how you might want to think about that.
53:59.080 --> 54:00.080
It's really helpful to think.
54:00.080 --> 54:02.080
It's something you can move through.
54:02.080 --> 54:03.080
Yeah.
54:03.080 --> 54:08.080
But is it helpful to think of it as spatial in some sense,
54:08.080 --> 54:09.080
or is there something?
54:09.080 --> 54:11.080
No, it's definitely spatial.
54:11.080 --> 54:13.080
It's spatial in a mathematical sense.
54:13.080 --> 54:14.080
How many dimensions?
54:14.080 --> 54:16.080
Can it be a crazy number of dimensions?
54:16.080 --> 54:17.080
Well, that's an interesting question.
54:17.080 --> 54:20.080
In the old part of the brain, the entorhinal cortex,
54:20.080 --> 54:22.080
they studied rats.
54:22.080 --> 54:24.080
And initially, it looks like, oh, this is just two dimensional.
54:24.080 --> 54:27.080
It's like the rat is in some box in a maze or whatever,
54:27.080 --> 54:29.080
and they know whether the rat is using these two dimensional
54:29.080 --> 54:32.080
reference frames and know where it is in the maze.
54:32.080 --> 54:35.080
We say, OK, well, what about bats?
54:35.080 --> 54:38.080
That's a mammal, and they fly in three dimensional space.
54:38.080 --> 54:39.080
How do they do that?
54:39.080 --> 54:41.080
They seem to know where they are, right?
54:41.080 --> 54:44.080
So this is a current area of active research,
54:44.080 --> 54:47.080
and it seems like somehow the neurons in the entorhinal cortex
54:47.080 --> 54:50.080
can learn three dimensional space.
54:50.080 --> 54:55.080
We just, two members of our team, along with Ilefet from MIT,
54:55.080 --> 54:59.080
just released a paper this literally last week,
54:59.080 --> 55:03.080
it's on bioarchive, where they show that you can,
55:03.080 --> 55:06.080
the way these things work, and unless you want to,
55:06.080 --> 55:10.080
I won't get into the detail, but grid cells
55:10.080 --> 55:12.080
can represent any n dimensional space.
55:12.080 --> 55:15.080
It's not inherently limited.
55:15.080 --> 55:18.080
You can think of it this way, if you had two dimensional,
55:18.080 --> 55:21.080
the way it works is you had a bunch of two dimensional slices.
55:21.080 --> 55:22.080
That's the way these things work.
55:22.080 --> 55:24.080
There's a whole bunch of two dimensional models,
55:24.080 --> 55:27.080
and you can slice up any n dimensional space
55:27.080 --> 55:29.080
with two dimensional projections.
55:29.080 --> 55:31.080
And you could have one dimensional models.
55:31.080 --> 55:34.080
So there's nothing inherent about the mathematics
55:34.080 --> 55:36.080
about the way the neurons do this,
55:36.080 --> 55:39.080
which constrained the dimensionality of the space,
55:39.080 --> 55:41.080
which I think was important.
55:41.080 --> 55:44.080
So obviously, I have a three dimensional map of this cup.
55:44.080 --> 55:46.080
Maybe it's even more than that, I don't know.
55:46.080 --> 55:48.080
But it's a clearly three dimensional map of the cup.
55:48.080 --> 55:50.080
I don't just have a projection of the cup.
55:50.080 --> 55:52.080
But when I think about birds,
55:52.080 --> 55:53.080
or when I think about mathematics,
55:53.080 --> 55:55.080
perhaps it's more than three dimensions.
55:55.080 --> 55:56.080
Who knows?
55:56.080 --> 56:00.080
So in terms of each individual column
56:00.080 --> 56:04.080
building up more and more information over time,
56:04.080 --> 56:06.080
do you think that mechanism is well understood?
56:06.080 --> 56:10.080
In your mind, you've proposed a lot of architectures there.
56:10.080 --> 56:14.080
Is that a key piece, or is it, is the big piece,
56:14.080 --> 56:16.080
the thousand brain theory of intelligence,
56:16.080 --> 56:18.080
the ensemble of it all?
56:18.080 --> 56:19.080
Well, I think they're both big.
56:19.080 --> 56:21.080
I mean, clearly the concept, as a theorist,
56:21.080 --> 56:23.080
the concept is most exciting, right?
56:23.080 --> 56:24.080
A high level concept.
56:24.080 --> 56:25.080
A high level concept.
56:25.080 --> 56:26.080
This is a totally new way of thinking about
56:26.080 --> 56:27.080
how the near characteristics work.
56:27.080 --> 56:29.080
So that is appealing.
56:29.080 --> 56:31.080
It has all these ramifications.
56:31.080 --> 56:34.080
And with that, as a framework for how the brain works,
56:34.080 --> 56:35.080
you can make all kinds of predictions
56:35.080 --> 56:36.080
and solve all kinds of problems.
56:36.080 --> 56:38.080
Now we're trying to work through many of these details right now.
56:38.080 --> 56:40.080
Okay, how do the neurons actually do this?
56:40.080 --> 56:42.080
Well, it turns out, if you think about grid cells
56:42.080 --> 56:44.080
and place cells in the old parts of the brain,
56:44.080 --> 56:46.080
there's a lot that's known about them,
56:46.080 --> 56:47.080
but there's still some mysteries.
56:47.080 --> 56:49.080
There's a lot of debate about exactly the details,
56:49.080 --> 56:50.080
how these work, and what are the signs.
56:50.080 --> 56:52.080
And we have that same level of detail,
56:52.080 --> 56:54.080
that same level of concern.
56:54.080 --> 56:56.080
What we spend here, most of our time doing,
56:56.080 --> 56:59.080
is trying to make a very good list
56:59.080 --> 57:02.080
of the things we don't understand yet.
57:02.080 --> 57:04.080
That's the key part here.
57:04.080 --> 57:05.080
What are the constraints?
57:05.080 --> 57:07.080
It's not like, oh, this seems to work, we're done.
57:07.080 --> 57:09.080
It's like, okay, it kind of works,
57:09.080 --> 57:11.080
but these are other things we know it has to do,
57:11.080 --> 57:13.080
and it's not doing those yet.
57:13.080 --> 57:15.080
I would say we're well on the way here.
57:15.080 --> 57:17.080
We're not done yet.
57:17.080 --> 57:20.080
There's a lot of trickiness to this system,
57:20.080 --> 57:23.080
but the basic principles about how different layers
57:23.080 --> 57:27.080
in the neocortex are doing much of this, we understand.
57:27.080 --> 57:29.080
But there's some fundamental parts
57:29.080 --> 57:30.080
that we don't understand as well.
57:30.080 --> 57:34.080
So what would you say is one of the harder open problems,
57:34.080 --> 57:37.080
or one of the ones that have been bothering you,
57:37.080 --> 57:39.080
keeping you up at night the most?
57:39.080 --> 57:41.080
Well, right now, this is a detailed thing
57:41.080 --> 57:43.080
that wouldn't apply to most people, okay?
57:43.080 --> 57:44.080
Sure.
57:44.080 --> 57:45.080
But you want me to answer that question?
57:45.080 --> 57:46.080
Yeah, please.
57:46.080 --> 57:49.080
We've talked about, as if, oh, to predict
57:49.080 --> 57:51.080
what you're going to sense on this coffee cup,
57:51.080 --> 57:54.080
I need to know where my finger's going to be on the coffee cup.
57:54.080 --> 57:56.080
That is true, but it's insufficient.
57:56.080 --> 57:59.080
Think about my finger touching the edge of the coffee cup.
57:59.080 --> 58:02.080
My finger can touch it at different orientations.
58:02.080 --> 58:05.080
I can touch it at my finger around here.
58:05.080 --> 58:06.080
And that doesn't change.
58:06.080 --> 58:09.080
I can make that prediction, and somehow,
58:09.080 --> 58:10.080
so it's not just the location.
58:10.080 --> 58:13.080
There's an orientation component of this as well.
58:13.080 --> 58:15.080
This is known in the old part of the brain, too.
58:15.080 --> 58:17.080
There's things called head direction cells,
58:17.080 --> 58:18.080
which way the rat is facing.
58:18.080 --> 58:20.080
It's the same kind of basic idea.
58:20.080 --> 58:23.080
So if my finger were a rat, you know, in three dimensions,
58:23.080 --> 58:25.080
I have a three dimensional orientation,
58:25.080 --> 58:27.080
and I have a three dimensional location.
58:27.080 --> 58:29.080
If I was a rat, I would have a,
58:29.080 --> 58:31.080
I think it was a two dimensional location,
58:31.080 --> 58:33.080
or one dimensional orientation, like this,
58:33.080 --> 58:35.080
which way is it facing?
58:35.080 --> 58:38.080
So how the two components work together,
58:38.080 --> 58:41.080
how does it, I combine orientation,
58:41.080 --> 58:43.080
the orientation of my sensor,
58:43.080 --> 58:47.080
as well as the location,
58:47.080 --> 58:49.080
is a tricky problem.
58:49.080 --> 58:52.080
And I think I've made progress on it.
58:52.080 --> 58:55.080
So at a bigger version of that,
58:55.080 --> 58:57.080
the perspective is super interesting,
58:57.080 --> 58:58.080
but super specific.
58:58.080 --> 58:59.080
Yeah, I warned you.
58:59.080 --> 59:01.080
No, no, no, it's really good,
59:01.080 --> 59:04.080
but there's a more general version of that.
59:04.080 --> 59:06.080
Do you think context matters?
59:06.080 --> 59:10.080
The fact that we are in a building in North America,
59:10.080 --> 59:15.080
that we, in the day and age where we have mugs,
59:15.080 --> 59:18.080
I mean, there's all this extra information
59:18.080 --> 59:22.080
that you bring to the table about everything else in the room
59:22.080 --> 59:24.080
that's outside of just the coffee cup.
59:24.080 --> 59:25.080
Of course it is.
59:25.080 --> 59:27.080
How does it get connected, do you think?
59:27.080 --> 59:30.080
Yeah, and that is another really interesting question.
59:30.080 --> 59:32.080
I'm going to throw that under the rubric
59:32.080 --> 59:34.080
or the name of attentional problems.
59:34.080 --> 59:36.080
First of all, we have this model.
59:36.080 --> 59:37.080
I have many, many models.
59:37.080 --> 59:39.080
And also the question, does it matter?
59:39.080 --> 59:41.080
Well, it matters for certain things.
59:41.080 --> 59:42.080
Of course it does.
59:42.080 --> 59:44.080
Maybe what we think about as a coffee cup
59:44.080 --> 59:47.080
in another part of the world is viewed as something completely different.
59:47.080 --> 59:51.080
Or maybe our logo, which is very benign in this part of the world,
59:51.080 --> 59:53.080
it means something very different in another part of the world.
59:53.080 --> 59:56.080
So those things do matter.
59:56.080 --> 1:00:00.080
I think the way to think about it as the following,
1:00:00.080 --> 1:00:01.080
one way to think about it,
1:00:01.080 --> 1:00:03.080
is we have all these models of the world.
1:00:03.080 --> 1:00:06.080
And we model everything.
1:00:06.080 --> 1:00:08.080
And as I said earlier, I kind of snuck it in there.
1:00:08.080 --> 1:00:12.080
Our models are actually, we build composite structures.
1:00:12.080 --> 1:00:15.080
So every object is composed of other objects,
1:00:15.080 --> 1:00:16.080
which are composed of other objects,
1:00:16.080 --> 1:00:18.080
and they become members of other objects.
1:00:18.080 --> 1:00:21.080
So this room is chairs and a table and a room and walls and so on.
1:00:21.080 --> 1:00:24.080
Now we can just arrange these things in a certain way.
1:00:24.080 --> 1:00:27.080
And you go, oh, that's in the Nementa conference room.
1:00:27.080 --> 1:00:32.080
So, and what we do is when we go around the world,
1:00:32.080 --> 1:00:34.080
when we experience the world,
1:00:34.080 --> 1:00:36.080
by walking to a room, for example,
1:00:36.080 --> 1:00:38.080
the first thing I do is like, oh, I'm in this room.
1:00:38.080 --> 1:00:39.080
Do I recognize the room?
1:00:39.080 --> 1:00:42.080
Then I can say, oh, look, there's a table here.
1:00:42.080 --> 1:00:44.080
And by attending to the table,
1:00:44.080 --> 1:00:46.080
I'm then assigning this table in a context of the room.
1:00:46.080 --> 1:00:48.080
Then I say, oh, on the table, there's a coffee cup.
1:00:48.080 --> 1:00:50.080
Oh, and on the table, there's a logo.
1:00:50.080 --> 1:00:52.080
And in the logo, there's the word Nementa.
1:00:52.080 --> 1:00:54.080
So if you look in the logo, there's the letter E.
1:00:54.080 --> 1:00:56.080
And look, it has an unusual surf.
1:00:56.080 --> 1:00:59.080
It doesn't actually, but I pretend it does.
1:00:59.080 --> 1:01:05.080
So the point is your attention is kind of drilling deep in and out
1:01:05.080 --> 1:01:07.080
of these nested structures.
1:01:07.080 --> 1:01:09.080
And I can pop back up and I can pop back down.
1:01:09.080 --> 1:01:11.080
I can pop back up and I can pop back down.
1:01:11.080 --> 1:01:13.080
So when I attend to the coffee cup,
1:01:13.080 --> 1:01:15.080
I haven't lost the context of everything else,
1:01:15.080 --> 1:01:19.080
but it's sort of, there's this sort of nested structure.
1:01:19.080 --> 1:01:22.080
The attention filters the reference frame formation
1:01:22.080 --> 1:01:24.080
for that particular period of time.
1:01:24.080 --> 1:01:25.080
Yes.
1:01:25.080 --> 1:01:28.080
It basically, a moment to moment, you attend the subcomponents
1:01:28.080 --> 1:01:30.080
and then you can attend the subcomponents to subcomponents.
1:01:30.080 --> 1:01:31.080
You can move up and down.
1:01:31.080 --> 1:01:32.080
You can move up and down.
1:01:32.080 --> 1:01:33.080
We do that all the time.
1:01:33.080 --> 1:01:35.080
You're not even, now that I'm aware of it,
1:01:35.080 --> 1:01:37.080
I'm very conscious of it.
1:01:37.080 --> 1:01:40.080
But most people don't even think about this.
1:01:40.080 --> 1:01:42.080
You know, you just walk in a room and you don't say,
1:01:42.080 --> 1:01:43.080
oh, I looked at the chair and I looked at the board
1:01:43.080 --> 1:01:44.080
and looked at that word on the board
1:01:44.080 --> 1:01:45.080
and I looked over here.
1:01:45.080 --> 1:01:46.080
What's going on?
1:01:46.080 --> 1:01:47.080
Right.
1:01:47.080 --> 1:01:50.080
So what percentage of your day are you deeply aware of this?
1:01:50.080 --> 1:01:53.080
In what part can you actually relax and just be Jeff?
1:01:53.080 --> 1:01:55.080
Me personally, like my personal day.
1:01:55.080 --> 1:01:56.080
Yeah.
1:01:56.080 --> 1:02:01.080
Unfortunately, I'm afflicted with too much of the former.
1:02:01.080 --> 1:02:03.080
Well, unfortunately or unfortunately.
1:02:03.080 --> 1:02:04.080
Yeah.
1:02:04.080 --> 1:02:05.080
You don't think it's useful?
1:02:05.080 --> 1:02:06.080
Oh, it is useful.
1:02:06.080 --> 1:02:07.080
Totally useful.
1:02:07.080 --> 1:02:09.080
I think about this stuff almost all the time.
1:02:09.080 --> 1:02:13.080
And one of my primary ways of thinking is
1:02:13.080 --> 1:02:14.080
when I'm asleep at night,
1:02:14.080 --> 1:02:16.080
I always wake up in the middle of the night
1:02:16.080 --> 1:02:19.080
and I stay awake for at least an hour with my eyes shut
1:02:19.080 --> 1:02:21.080
in sort of a half sleep state thinking about these things.
1:02:21.080 --> 1:02:23.080
I come up with answers to problems very often
1:02:23.080 --> 1:02:25.080
in that sort of half sleeping state.
1:02:25.080 --> 1:02:27.080
I think about on my bike ride, I think about on walks.
1:02:27.080 --> 1:02:29.080
I'm just constantly thinking about this.
1:02:29.080 --> 1:02:34.080
I have to almost schedule time to not think about this stuff
1:02:34.080 --> 1:02:37.080
because it's very, it's mentally taxing.
1:02:37.080 --> 1:02:39.080
Are you, when you're thinking about this stuff,
1:02:39.080 --> 1:02:41.080
are you thinking introspectively,
1:02:41.080 --> 1:02:43.080
like almost taking a step outside of yourself
1:02:43.080 --> 1:02:45.080
and trying to figure out what is your mind doing right now?
1:02:45.080 --> 1:02:48.080
I do that all the time, but that's not all I do.
1:02:48.080 --> 1:02:50.080
I'm constantly observing myself.
1:02:50.080 --> 1:02:52.080
So as soon as I started thinking about grid cells,
1:02:52.080 --> 1:02:54.080
for example, and getting into that,
1:02:54.080 --> 1:02:57.080
I started saying, oh, well, grid cells can have my place of sense
1:02:57.080 --> 1:02:58.080
in the world.
1:02:58.080 --> 1:02:59.080
That's where you know where you are.
1:02:59.080 --> 1:03:01.080
And it's interesting, we always have a sense of where we are
1:03:01.080 --> 1:03:02.080
unless we're lost.
1:03:02.080 --> 1:03:05.080
And so I started at night when I got up to go to the bathroom,
1:03:05.080 --> 1:03:07.080
I would start trying to do it completely with my eyes closed
1:03:07.080 --> 1:03:09.080
all the time and I would test my sense of grid cells.
1:03:09.080 --> 1:03:13.080
I would walk five feet and say, okay, I think I'm here.
1:03:13.080 --> 1:03:14.080
Am I really there?
1:03:14.080 --> 1:03:15.080
What's my error?
1:03:15.080 --> 1:03:17.080
And then I would calculate my error again and see how the errors
1:03:17.080 --> 1:03:18.080
accumulate.
1:03:18.080 --> 1:03:20.080
So even something as simple as getting up in the middle of the
1:03:20.080 --> 1:03:22.080
night to go to the bathroom, I'm testing these theories out.
1:03:22.080 --> 1:03:23.080
It's kind of fun.
1:03:23.080 --> 1:03:25.080
I mean, the coffee cup is an example of that too.
1:03:25.080 --> 1:03:30.080
So I think I find that these sort of everyday introspections
1:03:30.080 --> 1:03:32.080
are actually quite helpful.
1:03:32.080 --> 1:03:34.080
It doesn't mean you can ignore the science.
1:03:34.080 --> 1:03:38.080
I mean, I spend hours every day reading ridiculously complex
1:03:38.080 --> 1:03:39.080
papers.
1:03:39.080 --> 1:03:41.080
That's not nearly as much fun,
1:03:41.080 --> 1:03:44.080
but you have to sort of build up those constraints and the knowledge
1:03:44.080 --> 1:03:47.080
about the field and who's doing what and what exactly they think
1:03:47.080 --> 1:03:48.080
is happening here.
1:03:48.080 --> 1:03:51.080
And then you can sit back and say, okay, let's try to have pieces
1:03:51.080 --> 1:03:52.080
all together.
1:03:52.080 --> 1:03:56.080
Let's come up with some, you know, I'm very in this group here
1:03:56.080 --> 1:03:58.080
and people, they know they do this.
1:03:58.080 --> 1:03:59.080
I do this all the time.
1:03:59.080 --> 1:04:01.080
I come in with these introspective ideas and say, well,
1:04:01.080 --> 1:04:02.080
there we ever thought about this.
1:04:02.080 --> 1:04:04.080
Now watch, well, let's all do this together.
1:04:04.080 --> 1:04:06.080
And it's helpful.
1:04:06.080 --> 1:04:10.080
It's not, as long as you don't, if all you did was that,
1:04:10.080 --> 1:04:12.080
then you're just making up stuff, right?
1:04:12.080 --> 1:04:15.080
But if you're constraining it by the reality of the neuroscience,
1:04:15.080 --> 1:04:17.080
then it's really helpful.
1:04:17.080 --> 1:04:22.080
So let's talk a little bit about deep learning and the successes
1:04:22.080 --> 1:04:28.080
in the applied space of neural networks, ideas of training model
1:04:28.080 --> 1:04:31.080
on data and these simple computational units,
1:04:31.080 --> 1:04:37.080
artificial neurons that with back propagation have statistical
1:04:37.080 --> 1:04:42.080
ways of being able to generalize from the training set on to
1:04:42.080 --> 1:04:44.080
data that similar to that training set.
1:04:44.080 --> 1:04:48.080
So where do you think are the limitations of those approaches?
1:04:48.080 --> 1:04:52.080
What do you think are strengths relative to your major efforts
1:04:52.080 --> 1:04:55.080
of constructing a theory of human intelligence?
1:04:55.080 --> 1:04:56.080
Yeah.
1:04:56.080 --> 1:04:58.080
Well, I'm not an expert in this field.
1:04:58.080 --> 1:04:59.080
I'm somewhat knowledgeable.
1:04:59.080 --> 1:05:00.080
So, but I'm not.
1:05:00.080 --> 1:05:02.080
A little bit in just your intuition.
1:05:02.080 --> 1:05:04.080
Well, I have a little bit more than intuition,
1:05:04.080 --> 1:05:07.080
but I just want to say like, you know, one of the things that you asked me,
1:05:07.080 --> 1:05:09.080
do I spend all my time thinking about neuroscience?
1:05:09.080 --> 1:05:10.080
I do.
1:05:10.080 --> 1:05:12.080
That's to the exclusion of thinking about things like convolutional neural
1:05:12.080 --> 1:05:13.080
networks.
1:05:13.080 --> 1:05:15.080
But I try to stay current.
1:05:15.080 --> 1:05:18.080
So look, I think it's great the progress they've made.
1:05:18.080 --> 1:05:19.080
It's fantastic.
1:05:19.080 --> 1:05:23.080
And as I mentioned earlier, it's very highly useful for many things.
1:05:23.080 --> 1:05:27.080
The models that we have today are actually derived from a lot of
1:05:27.080 --> 1:05:28.080
neuroscience principles.
1:05:28.080 --> 1:05:31.080
They are distributed processing systems and distributed memory systems,
1:05:31.080 --> 1:05:33.080
and that's how the brain works.
1:05:33.080 --> 1:05:36.080
And they use things that we might call them neurons,
1:05:36.080 --> 1:05:37.080
but they're really not neurons at all.
1:05:37.080 --> 1:05:39.080
So we can just, they're not really neurons.
1:05:39.080 --> 1:05:42.080
So they're distributed processing systems.
1:05:42.080 --> 1:05:47.080
And nature of hierarchy that came also from neuroscience.
1:05:47.080 --> 1:05:50.080
And so there's a lot of things, the learning rules, basically,
1:05:50.080 --> 1:05:52.080
not backprop, but other, you know, sort of heavy entire learning.
1:05:52.080 --> 1:05:55.080
I'll be curious to say they're not neurons at all.
1:05:55.080 --> 1:05:56.080
Can you describe in which way?
1:05:56.080 --> 1:06:00.080
I mean, some of it is obvious, but I'd be curious if you have specific
1:06:00.080 --> 1:06:02.080
ways in which you think are the biggest differences.
1:06:02.080 --> 1:06:06.080
Yeah, we had a paper in 2016 called Why Neurons of Thousands of Synapses.
1:06:06.080 --> 1:06:11.080
And if you read that paper, you'll know what I'm talking about here.
1:06:11.080 --> 1:06:14.080
A real neuron in the brain is a complex thing.
1:06:14.080 --> 1:06:18.080
Let's just start with the synapses on it, which is a connection between neurons.
1:06:18.080 --> 1:06:24.080
Real neurons can everywhere from five to 30,000 synapses on them.
1:06:24.080 --> 1:06:30.080
The ones near the cell body, the ones that are close to the soma, the cell body,
1:06:30.080 --> 1:06:33.080
those are like the ones that people model in artificial neurons.
1:06:33.080 --> 1:06:35.080
There's a few hundred of those.
1:06:35.080 --> 1:06:37.080
Maybe they can affect the cell.
1:06:37.080 --> 1:06:39.080
They can make the cell become active.
1:06:39.080 --> 1:06:43.080
95% of the synapses can't do that.
1:06:43.080 --> 1:06:44.080
They're too far away.
1:06:44.080 --> 1:06:47.080
So if you activate one of those synapses, it just doesn't affect the cell body
1:06:47.080 --> 1:06:49.080
enough to make any difference.
1:06:49.080 --> 1:06:50.080
Any one of them individually.
1:06:50.080 --> 1:06:53.080
Any one of them individually, or even if you do a mass of them.
1:06:53.080 --> 1:06:57.080
What real neurons do is the following.
1:06:57.080 --> 1:07:04.080
If you activate, or you get 10 to 20 of them active at the same time,
1:07:04.080 --> 1:07:06.080
meaning they're all receiving an input at the same time,
1:07:06.080 --> 1:07:10.080
and those 10 to 20 synapses or 40 synapses are within a very short distance
1:07:10.080 --> 1:07:13.080
on the dendrite, like 40 microns, a very small area.
1:07:13.080 --> 1:07:17.080
So if you activate a bunch of these right next to each other at some distant place,
1:07:17.080 --> 1:07:21.080
what happens is it creates what's called the dendritic spike.
1:07:21.080 --> 1:07:24.080
And dendritic spike travels through the dendrites
1:07:24.080 --> 1:07:27.080
and can reach the soma or the cell body.
1:07:27.080 --> 1:07:31.080
Now, when it gets there, it changes the voltage,
1:07:31.080 --> 1:07:33.080
which is sort of like going to make the cell fire,
1:07:33.080 --> 1:07:35.080
but never enough to make the cell fire.
1:07:35.080 --> 1:07:38.080
It's sort of what we call, it says we depolarize the cell.
1:07:38.080 --> 1:07:41.080
You raise the voltage a little bit, but not enough to do anything.
1:07:41.080 --> 1:07:42.080
It's like, well, what good is that?
1:07:42.080 --> 1:07:44.080
And then it goes back down again.
1:07:44.080 --> 1:07:50.080
So we proposed a theory, which I'm very confident in basics are,
1:07:50.080 --> 1:07:54.080
is that what's happening there is those 95% of the synapses
1:07:54.080 --> 1:07:58.080
are recognizing dozens to hundreds of unique patterns.
1:07:58.080 --> 1:08:01.080
They can write, you know, about 10, 20 synapses at a time,
1:08:01.080 --> 1:08:04.080
and they're acting like predictions.
1:08:04.080 --> 1:08:07.080
So the neuron actually is a predictive engine on its own.
1:08:07.080 --> 1:08:11.080
It can fire when it gets enough, what they call proximal input from those ones
1:08:11.080 --> 1:08:15.080
near the cell fire, but it can get ready to fire from dozens to hundreds
1:08:15.080 --> 1:08:17.080
of patterns that it recognizes from the other guys.
1:08:17.080 --> 1:08:22.080
And the advantage of this to the neuron is that when it actually does produce
1:08:22.080 --> 1:08:27.080
a spike in action potential, it does so slightly sooner than it would have otherwise.
1:08:27.080 --> 1:08:29.080
And so what could just slightly sooner?
1:08:29.080 --> 1:08:33.080
Well, the slightly sooner part is it, there's all the neurons in the,
1:08:33.080 --> 1:08:36.080
the excited throwing neurons in the brain are surrounded by these inhibitory neurons,
1:08:36.080 --> 1:08:40.080
and they're very fast, the inhibitory neurons, these baskets all.
1:08:40.080 --> 1:08:44.080
And if I get my spike out a little bit sooner than someone else,
1:08:44.080 --> 1:08:46.080
I inhibit all my neighbors around me, right?
1:08:46.080 --> 1:08:49.080
And what you end up with is a different representation.
1:08:49.080 --> 1:08:52.080
You end up with a representation that matches your prediction.
1:08:52.080 --> 1:08:55.080
It's a sparser representation, meaning fewer neurons are active,
1:08:55.080 --> 1:08:57.080
but it's much more specific.
1:08:57.080 --> 1:09:04.080
And so we showed how networks of these neurons can do very sophisticated temporal prediction, basically.
1:09:04.080 --> 1:09:10.080
So this summarizes real neurons in the brain are time based prediction engines,
1:09:10.080 --> 1:09:17.080
and there's no concept of this at all in artificial, what we call point neurons.
1:09:17.080 --> 1:09:19.080
I don't think you can mail the brain without them.
1:09:19.080 --> 1:09:25.080
I don't think you can build intelligence without them because it's where a large part of the time comes from.
1:09:25.080 --> 1:09:31.080
These are predictive models and the time is, there's a prior prediction and an action,
1:09:31.080 --> 1:09:34.080
and it's inherent through every neuron in the neocortex.
1:09:34.080 --> 1:09:38.080
So I would say that point neurons sort of model a piece of that,
1:09:38.080 --> 1:09:45.080
and not very well at that either, but, you know, like, for example, synapses are very unreliable,
1:09:45.080 --> 1:09:49.080
and you cannot assign any precision to them.
1:09:49.080 --> 1:09:52.080
So even one digit of precision is not possible.
1:09:52.080 --> 1:09:57.080
So the way real neurons work is they don't add these, they don't change these weights accurately,
1:09:57.080 --> 1:09:59.080
like artificial neural networks do.
1:09:59.080 --> 1:10:03.080
They basically form new synapses, and so what you're trying to always do is
1:10:03.080 --> 1:10:09.080
detect the presence of some 10 to 20 active synapses at the same time as opposed,
1:10:09.080 --> 1:10:11.080
and they're almost binary.
1:10:11.080 --> 1:10:14.080
It's like, because you can't really represent anything much finer than that.
1:10:14.080 --> 1:10:18.080
So these are the kind of, and I think that's actually another essential component
1:10:18.080 --> 1:10:24.080
because the brain works on sparse patterns, and all that mechanism is based on sparse patterns,
1:10:24.080 --> 1:10:28.080
and I don't actually think you could build real brains or machine intelligence
1:10:28.080 --> 1:10:30.080
without incorporating some of those ideas.
1:10:30.080 --> 1:10:34.080
It's hard to even think about the complexity that emerges from the fact that
1:10:34.080 --> 1:10:40.080
the timing of the firing matters in the brain, the fact that you form new synapses,
1:10:40.080 --> 1:10:44.080
and everything you just mentioned in the past couple minutes.
1:10:44.080 --> 1:10:47.080
Trust me, if you spend time on it, you can get your mind around it.
1:10:47.080 --> 1:10:49.080
It's not like it's no longer a mystery to me.
1:10:49.080 --> 1:10:53.080
No, but sorry, as a function in a mathematical way,
1:10:53.080 --> 1:10:58.080
can you start getting an intuition about what gets it excited, what not,
1:10:58.080 --> 1:11:00.080
and what kind of representation?
1:11:00.080 --> 1:11:04.080
Yeah, it's not as easy as there are many other types of neural networks
1:11:04.080 --> 1:11:10.080
that are more amenable to pure analysis, especially very simple networks.
1:11:10.080 --> 1:11:12.080
You know, oh, I have four neurons, and they're doing this.
1:11:12.080 --> 1:11:16.080
Can we describe them mathematically what they're doing type of thing?
1:11:16.080 --> 1:11:19.080
Even the complexity of convolutional neural networks today,
1:11:19.080 --> 1:11:23.080
it's sort of a mystery. They can't really describe the whole system.
1:11:23.080 --> 1:11:25.080
And so it's different.
1:11:25.080 --> 1:11:31.080
My colleague, Subitain Ahmad, he did a nice paper on this.
1:11:31.080 --> 1:11:34.080
You can get all the stuff on our website if you're interested.
1:11:34.080 --> 1:11:38.080
Talking about sort of mathematical properties of sparse representations,
1:11:38.080 --> 1:11:42.080
and so what we can do is we can show mathematically, for example,
1:11:42.080 --> 1:11:46.080
why 10 to 20 synapses to recognize a pattern is the correct number,
1:11:46.080 --> 1:11:48.080
is the right number you'd want to use.
1:11:48.080 --> 1:11:50.080
And by the way, that matches biology.
1:11:50.080 --> 1:11:55.080
We can show mathematically some of these concepts about the show
1:11:55.080 --> 1:12:01.080
why the brain is so robust to noise and error and fallout and so on.
1:12:01.080 --> 1:12:05.080
We can show that mathematically as well as empirically in simulations.
1:12:05.080 --> 1:12:08.080
But the system can't be analyzed completely.
1:12:08.080 --> 1:12:12.080
Any complex system can, and so that's out of the realm.
1:12:12.080 --> 1:12:19.080
But there is mathematical benefits and intuitions that can be derived from mathematics.
1:12:19.080 --> 1:12:21.080
And we try to do that as well.
1:12:21.080 --> 1:12:23.080
Most of our papers have a section about that.
1:12:23.080 --> 1:12:28.080
So I think it's refreshing and useful for me to be talking to you about deep neural networks,
1:12:28.080 --> 1:12:36.080
because your intuition basically says that we can't achieve anything like intelligence with artificial neural networks.
1:12:36.080 --> 1:12:37.080
Well, not in the current form.
1:12:37.080 --> 1:12:38.080
Not in the current form.
1:12:38.080 --> 1:12:40.080
I'm sure we can do it in the ultimate form, sure.
1:12:40.080 --> 1:12:43.080
So let me dig into it and see what your thoughts are there a little bit.
1:12:43.080 --> 1:12:49.080
So I'm not sure if you read this little blog post called Bitter Lesson by Rich Sutton recently.
1:12:49.080 --> 1:12:51.080
He's a reinforcement learning pioneer.
1:12:51.080 --> 1:12:53.080
I'm not sure if you're familiar with him.
1:12:53.080 --> 1:13:02.080
His basic idea is that all the stuff we've done in AI in the past 70 years, he's one of the old school guys.
1:13:02.080 --> 1:13:10.080
The biggest lesson learned is that all the tricky things we've done don't, you know, they benefit in the short term.
1:13:10.080 --> 1:13:20.080
But in the long term, what wins out is a simple general method that just relies on Moore's law on computation getting faster and faster.
1:13:20.080 --> 1:13:21.080
This is what he's saying.
1:13:21.080 --> 1:13:23.080
This is what has worked up to now.
1:13:23.080 --> 1:13:25.080
This is what has worked up to now.
1:13:25.080 --> 1:13:31.080
If you're trying to build a system, if we're talking about, he's not concerned about intelligence.
1:13:31.080 --> 1:13:38.080
He's concerned about a system that works in terms of making predictions on applied, narrow AI problems.
1:13:38.080 --> 1:13:41.080
That's what the discussion is about.
1:13:41.080 --> 1:13:50.080
That you just try to go as general as possible and wait years or decades for the computation to make it actually possible.
1:13:50.080 --> 1:13:54.080
Is he saying that as a criticism or is he saying this is a prescription of what we ought to be doing?
1:13:54.080 --> 1:13:55.080
Well, it's very difficult.
1:13:55.080 --> 1:13:57.080
He's saying this is what has worked.
1:13:57.080 --> 1:14:03.080
And yes, a prescription, but it's a difficult prescription because it says all the fun things you guys are trying to do.
1:14:03.080 --> 1:14:05.080
We are trying to do.
1:14:05.080 --> 1:14:07.080
He's part of the community.
1:14:07.080 --> 1:14:11.080
He's saying it's only going to be short term gains.
1:14:11.080 --> 1:14:19.080
This all leads up to a question, I guess, on artificial neural networks and maybe our own biological neural networks.
1:14:19.080 --> 1:14:24.080
Do you think if we just scale things up significantly?
1:14:24.080 --> 1:14:28.080
Take these dumb artificial neurons, the point neurons.
1:14:28.080 --> 1:14:30.080
I like that term.
1:14:30.080 --> 1:14:36.080
If we just have a lot more of them, do you think some of the elements that we see in the brain
1:14:36.080 --> 1:14:38.080
may start emerging?
1:14:38.080 --> 1:14:39.080
No, I don't think so.
1:14:39.080 --> 1:14:43.080
We can do bigger problems of the same type.
1:14:43.080 --> 1:14:50.080
I mean, it's been pointed out by many people that today's convolutional neural networks aren't really much different than the ones we had quite a while ago.
1:14:50.080 --> 1:14:56.080
We just, they're bigger and train more and we have more labeled data and so on.
1:14:56.080 --> 1:15:03.080
But I don't think you can get to the kind of things I know the brain can do and that we think about as intelligence by just scaling it up.
1:15:03.080 --> 1:15:12.080
So that may be, it's a good description of what's happened in the past, what's happened recently with the reemergence of artificial neural networks.
1:15:12.080 --> 1:15:17.080
It may be a good prescription for what's going to happen in the short term.
1:15:17.080 --> 1:15:19.080
But I don't think that's the path.
1:15:19.080 --> 1:15:20.080
I've said that earlier.
1:15:20.080 --> 1:15:21.080
There's an alternate path.
1:15:21.080 --> 1:15:29.080
I should mention to you, by the way, that we've made sufficient progress on our, the whole cortical theory in the last few years.
1:15:29.080 --> 1:15:40.080
But last year, we decided to start actively pursuing how we get these ideas embedded into machine learning.
1:15:40.080 --> 1:15:45.080
That's, again, being led by my colleague, and he's more of a machine learning guy.
1:15:45.080 --> 1:15:47.080
I'm more of an neuroscience guy.
1:15:47.080 --> 1:15:58.080
So this is now our, I wouldn't say our focus, but it is now an equal focus here because we need to proselytize what we've learned.
1:15:58.080 --> 1:16:03.080
And we need to show how it's beneficial to the machine learning.
1:16:03.080 --> 1:16:05.080
So we're putting, we have a plan in place right now.
1:16:05.080 --> 1:16:07.080
In fact, we just did our first paper on this.
1:16:07.080 --> 1:16:09.080
I can tell you about that.
1:16:09.080 --> 1:16:15.080
But, you know, one of the reasons I want to talk to you is because I'm trying to get more people in the machine learning community to say,
1:16:15.080 --> 1:16:17.080
I need to learn about this stuff.
1:16:17.080 --> 1:16:21.080
And maybe we should just think about this a bit more about what we've learned about the brain.
1:16:21.080 --> 1:16:23.080
And what are those team, what have they done?
1:16:23.080 --> 1:16:25.080
Is that useful for us?
1:16:25.080 --> 1:16:32.080
Yeah, so is there elements of all the, the cortical theory that things we've been talking about that may be useful in the short term?
1:16:32.080 --> 1:16:34.080
Yes, in the short term, yes.
1:16:34.080 --> 1:16:41.080
This is the, sorry to interrupt, but the, the open question is it, it certainly feels from my perspective that in the long term,
1:16:41.080 --> 1:16:44.080
some of the ideas we've been talking about will be extremely useful.
1:16:44.080 --> 1:16:46.080
The question is whether in the short term.
1:16:46.080 --> 1:16:51.080
Well, this is a, always what we, I would call the entrepreneur's dilemma.
1:16:51.080 --> 1:16:59.080
You have this long term vision, oh, we're going to all be driving electric cars or we're all going to have computers or we're all going to whatever.
1:16:59.080 --> 1:17:03.080
And, and you're at some point in time and you say, I can see that long term vision.
1:17:03.080 --> 1:17:04.080
I'm sure it's going to happen.
1:17:04.080 --> 1:17:07.080
How do I get there without killing myself, you know, without going out of business?
1:17:07.080 --> 1:17:09.080
That's the challenge.
1:17:09.080 --> 1:17:10.080
That's the dilemma.
1:17:10.080 --> 1:17:11.080
That's the really difficult thing to do.
1:17:11.080 --> 1:17:13.080
So we're facing that right now.
1:17:13.080 --> 1:17:17.080
So ideally what you'd want to do is find some steps along the way that you can get there incrementally.
1:17:17.080 --> 1:17:20.080
You don't have to like throw it all out and start over again.
1:17:20.080 --> 1:17:25.080
The first thing that we've done is we focus on these sparse representations.
1:17:25.080 --> 1:17:30.080
So just in case you don't know what that means or some of the listeners don't know what that means.
1:17:30.080 --> 1:17:37.080
In the brain, if I have like 10,000 neurons, what you would see is maybe 2% of them active at a time.
1:17:37.080 --> 1:17:41.080
You don't see 50%, you don't see 30%, you might see 2%.
1:17:41.080 --> 1:17:42.080
And it's always like that.
1:17:42.080 --> 1:17:44.080
For any set of sensory inputs.
1:17:44.080 --> 1:17:45.080
It doesn't matter anything.
1:17:45.080 --> 1:17:47.080
It doesn't matter any part of the brain.
1:17:47.080 --> 1:17:51.080
But which neurons differs?
1:17:51.080 --> 1:17:52.080
Which neurons are active?
1:17:52.080 --> 1:17:53.080
Yeah.
1:17:53.080 --> 1:17:54.080
So let me put this.
1:17:54.080 --> 1:17:56.080
Let's say I take 10,000 neurons that are representing something.
1:17:56.080 --> 1:17:58.080
They're sitting there in a little block together.
1:17:58.080 --> 1:18:00.080
It's a teeny little block of neurons, 10,000 neurons.
1:18:00.080 --> 1:18:01.080
And they're representing a location.
1:18:01.080 --> 1:18:02.080
They're representing a cop.
1:18:02.080 --> 1:18:04.080
They're representing the input from my sensors.
1:18:04.080 --> 1:18:05.080
I don't know.
1:18:05.080 --> 1:18:06.080
It doesn't matter.
1:18:06.080 --> 1:18:07.080
It's representing something.
1:18:07.080 --> 1:18:10.080
The way the representations occur, it's always a sparse representation.
1:18:10.080 --> 1:18:12.080
Meaning it's a population code.
1:18:12.080 --> 1:18:15.080
So which 200 cells are active tells me what's going on.
1:18:15.080 --> 1:18:18.080
It's not individual cells aren't that important at all.
1:18:18.080 --> 1:18:20.080
It's the population code that matters.
1:18:20.080 --> 1:18:23.080
And when you have sparse population codes,
1:18:23.080 --> 1:18:26.080
then all kinds of beautiful properties come out of them.
1:18:26.080 --> 1:18:29.080
So the brain uses sparse population codes that we've written
1:18:29.080 --> 1:18:32.080
and described these benefits in some of our papers.
1:18:32.080 --> 1:18:37.080
So they give this tremendous robustness to the systems.
1:18:37.080 --> 1:18:39.080
You know, brains are incredibly robust.
1:18:39.080 --> 1:18:42.080
Neurons are dying all the time and spasming and synapses falling apart.
1:18:42.080 --> 1:18:45.080
And, you know, all the time and it keeps working.
1:18:45.080 --> 1:18:52.080
So what Subitai and Louise, one of our other engineers here have done,
1:18:52.080 --> 1:18:56.080
have shown that they're introducing sparseness into convolutional neural networks.
1:18:56.080 --> 1:18:58.080
Now other people are thinking along these lines,
1:18:58.080 --> 1:19:00.080
but we're going about it in a more principled way, I think.
1:19:00.080 --> 1:19:06.080
And we're showing that if you enforce sparseness throughout these convolutional neural networks,
1:19:06.080 --> 1:19:13.080
in both the sort of which neurons are active and the connections between them,
1:19:13.080 --> 1:19:15.080
that you get some very desirable properties.
1:19:15.080 --> 1:19:20.080
So one of the current hot topics in deep learning right now are these adversarial examples.
1:19:20.080 --> 1:19:23.080
So, you know, you give me any deep learning network
1:19:23.080 --> 1:19:27.080
and I can give you a picture that looks perfect and you're going to call it, you know,
1:19:27.080 --> 1:19:30.080
you're going to say the monkey is, you know, an airplane.
1:19:30.080 --> 1:19:32.080
So that's a problem.
1:19:32.080 --> 1:19:36.080
And DARPA just announced some big thing and we're trying to, you know, have some contests for this.
1:19:36.080 --> 1:19:40.080
But if you enforce sparse representations here,
1:19:40.080 --> 1:19:41.080
many of these problems go away.
1:19:41.080 --> 1:19:45.080
They're much more robust and they're not easy to fool.
1:19:45.080 --> 1:19:48.080
So we've already shown some of those results,
1:19:48.080 --> 1:19:53.080
just literally in January or February, just like last month we did that.
1:19:53.080 --> 1:19:59.080
And you can, I think it's on bioarchive right now or on iCry, you can read about it.
1:19:59.080 --> 1:20:02.080
But so that's like a baby step.
1:20:02.080 --> 1:20:04.080
Okay. That's a take something from the brain.
1:20:04.080 --> 1:20:05.080
We know, we know about sparseness.
1:20:05.080 --> 1:20:06.080
We know why it's important.
1:20:06.080 --> 1:20:08.080
We know what it gives the brain.
1:20:08.080 --> 1:20:09.080
So let's try to enforce that onto this.
1:20:09.080 --> 1:20:12.080
What's your intuition why sparsity leads to robustness?
1:20:12.080 --> 1:20:14.080
Because it feels like it would be less robust.
1:20:14.080 --> 1:20:17.080
Why would you feel the rest robust to you?
1:20:17.080 --> 1:20:24.080
So it, it just feels like if the fewer neurons are involved,
1:20:24.080 --> 1:20:26.080
the more fragile the representation.
1:20:26.080 --> 1:20:28.080
Yeah, but I didn't say there was lots of few.
1:20:28.080 --> 1:20:30.080
I said, let's say 200.
1:20:30.080 --> 1:20:31.080
That's a lot.
1:20:31.080 --> 1:20:32.080
There's still a lot.
1:20:32.080 --> 1:20:33.080
Yeah.
1:20:33.080 --> 1:20:35.080
So here's an intuition for it.
1:20:35.080 --> 1:20:37.080
This is a bit technical.
1:20:37.080 --> 1:20:41.080
So for, you know, for engineers, machine learning people this be easy,
1:20:41.080 --> 1:20:44.080
but God's listeners, maybe not.
1:20:44.080 --> 1:20:46.080
If you're trying to classify something,
1:20:46.080 --> 1:20:50.080
you're trying to divide some very high dimensional space into different pieces, A and B.
1:20:50.080 --> 1:20:55.080
And you're trying to create some point where you say all these points in this high dimensional space are A
1:20:55.080 --> 1:20:57.080
and all these points in this high dimensional space are B.
1:20:57.080 --> 1:21:03.080
And if you have points that are close to that line, it's not very robust.
1:21:03.080 --> 1:21:07.080
It works for all the points you know about, but it's, it's not very robust
1:21:07.080 --> 1:21:10.080
because you can just move a little bit and you've crossed over the line.
1:21:10.080 --> 1:21:14.080
When you have sparse representations, imagine I pick, I have,
1:21:14.080 --> 1:21:18.080
I'm going to pick 200 cells active out of, out of 10,000.
1:21:18.080 --> 1:21:19.080
Okay.
1:21:19.080 --> 1:21:20.080
So I have 200 cells active.
1:21:20.080 --> 1:21:24.080
Now let's say I pick randomly another, a different representation, 200.
1:21:24.080 --> 1:21:27.080
The overlap between those is going to be very small, just a few.
1:21:27.080 --> 1:21:36.080
I can pick millions of samples randomly of 200 neurons and not one of them will overlap more than just a few.
1:21:36.080 --> 1:21:43.080
So one way to think about it is if I want to fool one of these representations to look like one of those other representations,
1:21:43.080 --> 1:21:46.080
I can't move just one cell or two cells or three cells or four cells.
1:21:46.080 --> 1:21:48.080
I have to move 100 cells.
1:21:48.080 --> 1:21:52.080
And that makes them robust.
1:21:52.080 --> 1:21:56.080
In terms of further, so you mentioned sparsity.
1:21:56.080 --> 1:21:57.080
Will we be the next thing?
1:21:57.080 --> 1:21:58.080
Yeah.
1:21:58.080 --> 1:21:59.080
Okay.
1:21:59.080 --> 1:22:00.080
So we have, we picked one.
1:22:00.080 --> 1:22:02.080
We don't know if it's going to work well yet.
1:22:02.080 --> 1:22:08.080
So again, we're trying to come up incremental ways of moving from brain theory to add pieces to machine learning,
1:22:08.080 --> 1:22:12.080
current machine learning world in one step at a time.
1:22:12.080 --> 1:22:20.080
So the next thing we're going to try to do is, is sort of incorporate some of the ideas of the, the thousand brains theory that you have many,
1:22:20.080 --> 1:22:22.080
many models and that are voting.
1:22:22.080 --> 1:22:23.080
Now that idea is not new.
1:22:23.080 --> 1:22:26.080
There's a mixture of models that's been around for a long time.
1:22:26.080 --> 1:22:29.080
But the way the brain does it is a little different.
1:22:29.080 --> 1:22:36.080
And, and the way it votes is different and the kind of way it represents uncertainty is different.
1:22:36.080 --> 1:22:43.080
So we're just starting this work, but we're going to try to see if we can sort of incorporate some of the principles of voting
1:22:43.080 --> 1:22:53.080
or principles of a thousand brain theory, like lots of simple models that talk to each other in a, in a very certain way.
1:22:53.080 --> 1:23:07.080
And can we build more machines and systems that learn faster and, and also, well, mostly are multimodal and robust to multimodal type of issues.
1:23:07.080 --> 1:23:15.080
So one of the challenges there is, you know, the machine learning computer vision community has certain sets of benchmarks.
1:23:15.080 --> 1:23:18.080
So it's a test based on which they compete.
1:23:18.080 --> 1:23:29.080
And I would argue, especially from your perspective, that those benchmarks aren't that useful for testing the aspects that the brain is good at or intelligent.
1:23:29.080 --> 1:23:31.080
They're not really testing intelligence.
1:23:31.080 --> 1:23:41.080
They're very fine and has been extremely useful for developing specific mathematical models, but it's not useful in the long term for creating intelligence.
1:23:41.080 --> 1:23:46.080
So do you think you also have a role in proposing better tests?
1:23:46.080 --> 1:23:50.080
Yeah, this is a very, you've identified a very serious problem.
1:23:50.080 --> 1:23:57.080
First of all, the tests that they have are the tests that they want, not the tests of the other things that we're trying to do.
1:23:57.080 --> 1:24:01.080
Right. You know, what are the, so on.
1:24:01.080 --> 1:24:10.080
The second thing is sometimes these to be competitive in these tests, you have to have huge data sets and huge computing power.
1:24:10.080 --> 1:24:13.080
And so, you know, and we don't have that here.
1:24:13.080 --> 1:24:18.080
We don't have it as well as other big teams that big companies do.
1:24:18.080 --> 1:24:20.080
So there's numerous issues there.
1:24:20.080 --> 1:24:26.080
You know, we come at it, you know, we're our approach to this is all based on, in some sense, you might argue elegance.
1:24:26.080 --> 1:24:30.080
You know, we're coming at it from like a theoretical base that we think, oh my God, this is so clearly elegant.
1:24:30.080 --> 1:24:31.080
This is how brains work.
1:24:31.080 --> 1:24:32.080
This is what intelligence is.
1:24:32.080 --> 1:24:35.080
But the machine learning world has gotten in this phase where they think it doesn't matter.
1:24:35.080 --> 1:24:39.080
Doesn't matter what you think, as long as you do, you know, 0.1% better on this benchmark.
1:24:39.080 --> 1:24:41.080
That's what that's all that matters.
1:24:41.080 --> 1:24:43.080
And that's a problem.
1:24:43.080 --> 1:24:46.080
You know, we have to figure out how to get around that.
1:24:46.080 --> 1:24:47.080
That's a challenge for us.
1:24:47.080 --> 1:24:50.080
That's one of the challenges we have to deal with.
1:24:50.080 --> 1:24:53.080
So I agree you've identified a big issue.
1:24:53.080 --> 1:24:55.080
It's difficult for those reasons.
1:24:55.080 --> 1:25:02.080
But, you know, part of the reasons I'm talking to you here today is I hope I'm going to get some machine learning people to say,
1:25:02.080 --> 1:25:03.080
I'm going to read those papers.
1:25:03.080 --> 1:25:04.080
Those might be some interesting ideas.
1:25:04.080 --> 1:25:08.080
I'm tired of doing this 0.1% improvement stuff, you know.
1:25:08.080 --> 1:25:21.080
Well, that's why I'm here as well, because I think machine learning now as a community is at a place where the next step is needs to be orthogonal to what has received success in the past.
1:25:21.080 --> 1:25:27.080
You see other leaders saying this, machine learning leaders, you know, Jeff Hinton with his capsules idea.
1:25:27.080 --> 1:25:33.080
Many people have gotten up saying, you know, we're going to hit road, maybe we should look at the brain, you know, things like that.
1:25:33.080 --> 1:25:37.080
So hopefully that thinking will occur organically.
1:25:37.080 --> 1:25:43.080
And then we're in a nice position for people to come and look at our work and say, well, what can we learn from these guys?
1:25:43.080 --> 1:25:49.080
Yeah, MIT is just launching a billion dollar computing college that's centered around this idea.
1:25:49.080 --> 1:25:51.080
On this idea of what?
1:25:51.080 --> 1:25:59.080
Well, the idea that, you know, the humanities, psychology, neuroscience have to work all together to get to build the S.
1:25:59.080 --> 1:26:02.080
Yeah, I mean, Stanford just did this human center today, I think.
1:26:02.080 --> 1:26:10.080
I'm a little disappointed in these initiatives because, you know, they're focusing on sort of the human side of it,
1:26:10.080 --> 1:26:17.080
and it can very easily slip into how humans interact with intelligent machines, which is nothing wrong with that.
1:26:17.080 --> 1:26:20.080
But that's not, that is orthogonal to what we're trying to do.
1:26:20.080 --> 1:26:22.080
We're trying to say, like, what is the essence of intelligence?
1:26:22.080 --> 1:26:23.080
I don't care.
1:26:23.080 --> 1:26:31.080
In fact, I want to build intelligent machines that aren't emotional, that don't smile at you, that, you know, that aren't trying to tuck you in at night.
1:26:31.080 --> 1:26:38.080
Yeah, there is that pattern that you, when you talk about understanding humans is important for understanding intelligence.
1:26:38.080 --> 1:26:47.080
You start slipping into topics of ethics or, yeah, like you said, the interactive elements as opposed to, no, no, no, let's zoom in on the brain,
1:26:47.080 --> 1:26:51.080
study what the human brain, the baby, the...
1:26:51.080 --> 1:26:53.080
Let's study what a brain does.
1:26:53.080 --> 1:26:57.080
And then we can decide which parts of that we want to recreate in some system.
1:26:57.080 --> 1:27:00.080
But until you have that theory about what the brain does, what's the point?
1:27:00.080 --> 1:27:03.080
You know, it's just, you're going to be wasting time, I think.
1:27:03.080 --> 1:27:09.080
Just to break it down on the artificial neural network side, maybe you can speak to this on the, on the biologic neural network side,
1:27:09.080 --> 1:27:13.080
the process of learning versus the process of inference.
1:27:13.080 --> 1:27:22.080
Maybe you can explain to me, what, is there a difference between, you know, in artificial neural networks, there's a difference between the learning stage and the inference stage?
1:27:22.080 --> 1:27:23.080
Yeah.
1:27:23.080 --> 1:27:25.080
Do you see the brain as something different?
1:27:25.080 --> 1:27:33.080
One of the big distinctions that people often say, I don't know how correct it is, is artificial neural networks need a lot of data.
1:27:33.080 --> 1:27:34.080
They're very inefficient learning.
1:27:34.080 --> 1:27:35.080
Yeah.
1:27:35.080 --> 1:27:42.080
Do you see that as a correct distinction from the biology of the human brain, that the human brain is very efficient?
1:27:42.080 --> 1:27:44.080
Or is that just something we deceive ourselves with?
1:27:44.080 --> 1:27:45.080
No, it is efficient, obviously.
1:27:45.080 --> 1:27:47.080
We can learn new things almost instantly.
1:27:47.080 --> 1:27:50.080
And so what elements do you think...
1:27:50.080 --> 1:27:51.080
Yeah, I can talk about that.
1:27:51.080 --> 1:27:52.080
You brought up two issues there.
1:27:52.080 --> 1:28:00.080
So remember I talked early about the constraints, we always feel, well, one of those constraints is the fact that brains are continually learning.
1:28:00.080 --> 1:28:03.080
That's not something we said, oh, we can add that later.
1:28:03.080 --> 1:28:11.080
That's something that was upfront, had to be there from the start, made our problems harder.
1:28:11.080 --> 1:28:19.080
But we showed, going back to the 2016 paper on sequence memory, we showed how that happens, how the brains infer and learn at the same time.
1:28:19.080 --> 1:28:22.080
And our models do that.
1:28:22.080 --> 1:28:26.080
They're not two separate phases or two separate sets of time.
1:28:26.080 --> 1:28:33.080
I think that's a big, big problem in AI, at least for many applications, not for all.
1:28:33.080 --> 1:28:34.080
So I can talk about that.
1:28:34.080 --> 1:28:37.080
It gets detailed.
1:28:37.080 --> 1:28:46.080
There are some parts of the neocortex in the brain where actually what's going on, there's these cycles of activity in the brain.
1:28:46.080 --> 1:28:54.080
And there's very strong evidence that you're doing more of inference on one part of the phase and more of learning on the other part of the phase.
1:28:54.080 --> 1:28:58.080
So the brain can actually sort of separate different populations of cells that are going back and forth like this.
1:28:58.080 --> 1:29:01.080
But in general, I would say that's an important problem.
1:29:01.080 --> 1:29:05.080
We have all of our networks that we've come up with do both.
1:29:05.080 --> 1:29:08.080
They're continuous learning networks.
1:29:08.080 --> 1:29:10.080
And you mentioned benchmarks earlier.
1:29:10.080 --> 1:29:12.080
Well, there are no benchmarks about that.
1:29:12.080 --> 1:29:13.080
Exactly.
1:29:13.080 --> 1:29:19.080
So we have to like, we get in our little soapbox and hey, by the way, this is important.
1:29:19.080 --> 1:29:21.080
And here's the mechanism for doing that.
1:29:21.080 --> 1:29:26.080
But until you can prove it to someone in some commercial system or something, it's a little harder.
1:29:26.080 --> 1:29:33.080
So one of the things I had to linger on that is in some ways to learn the concept of a coffee cup.
1:29:33.080 --> 1:29:38.080
You only need this one coffee cup and maybe some time alone in a room with it.
1:29:38.080 --> 1:29:43.080
So the first thing is I imagine I reach my hand into a black box and I'm reaching, I'm trying to touch something.
1:29:43.080 --> 1:29:47.080
I don't know up front if it's something I already know or if it's a new thing.
1:29:47.080 --> 1:29:50.080
And I have to, I'm doing both at the same time.
1:29:50.080 --> 1:29:53.080
I don't say, oh, let's see if it's a new thing.
1:29:53.080 --> 1:29:55.080
Oh, let's see if it's an old thing.
1:29:55.080 --> 1:29:56.080
I don't do that.
1:29:56.080 --> 1:29:59.080
As I go, my brain says, oh, it's new or it's not new.
1:29:59.080 --> 1:30:02.080
And if it's new, I start learning what it is.
1:30:02.080 --> 1:30:06.080
And by the way, it starts learning from the get go, even if it's going to recognize it.
1:30:06.080 --> 1:30:09.080
So they're not separate problems.
1:30:09.080 --> 1:30:10.080
And so that's the thing there.
1:30:10.080 --> 1:30:13.080
The other thing you mentioned was the fast learning.
1:30:13.080 --> 1:30:17.080
So I was just talking about continuous learning, but there's also fast learning.
1:30:17.080 --> 1:30:20.080
Literally, I can show you this coffee cup and I say, here's a new coffee cup.
1:30:20.080 --> 1:30:21.080
It's got the logo on it.
1:30:21.080 --> 1:30:22.080
Take a look at it.
1:30:22.080 --> 1:30:23.080
Done.
1:30:23.080 --> 1:30:24.080
You're done.
1:30:24.080 --> 1:30:27.080
You can predict what it's going to look like, you know, in different positions.
1:30:27.080 --> 1:30:29.080
So I can talk about that too.
1:30:29.080 --> 1:30:30.080
Yes.
1:30:30.080 --> 1:30:34.080
In the brain, the way learning occurs.
1:30:34.080 --> 1:30:36.080
I mentioned this earlier, but I mentioned it again.
1:30:36.080 --> 1:30:40.080
The way learning occurs, I imagine I have a section of a dendrite of a neuron.
1:30:40.080 --> 1:30:43.080
And I want to learn, I'm going to learn something new.
1:30:43.080 --> 1:30:44.080
It just doesn't matter what it is.
1:30:44.080 --> 1:30:46.080
I'm just going to learn something new.
1:30:46.080 --> 1:30:48.080
I need to recognize a new pattern.
1:30:48.080 --> 1:30:52.080
So what I'm going to do is I'm going to form new synapses.
1:30:52.080 --> 1:30:57.080
New synapses, we're going to rewire the brain onto that section of the dendrite.
1:30:57.080 --> 1:31:02.080
Once I've done that, everything else that neuron has learned is not affected by it.
1:31:02.080 --> 1:31:06.080
Now, it's because it's isolated to that small section of the dendrite.
1:31:06.080 --> 1:31:09.080
They're not all being added together, like a point neuron.
1:31:09.080 --> 1:31:13.080
So if I learn something new on this segment here, it doesn't change any of the learning
1:31:13.080 --> 1:31:15.080
that occur anywhere else in that neuron.
1:31:15.080 --> 1:31:18.080
So I can add something without affecting previous learning.
1:31:18.080 --> 1:31:20.080
And I can do it quickly.
1:31:20.080 --> 1:31:24.080
Now, let's talk, we can talk about the quickness, how it's done in real neurons.
1:31:24.080 --> 1:31:27.080
You might say, well, doesn't it take time to form synapses?
1:31:27.080 --> 1:31:30.080
Yes, it can take maybe an hour to form a new synapse.
1:31:30.080 --> 1:31:32.080
We can form memories quicker than that.
1:31:32.080 --> 1:31:35.080
And I can explain that happens too, if you want.
1:31:35.080 --> 1:31:38.080
But it's getting a bit neurosciencey.
1:31:38.080 --> 1:31:40.080
That's great.
1:31:40.080 --> 1:31:43.080
But is there an understanding of these mechanisms at every level?
1:31:43.080 --> 1:31:48.080
So from the short term memories and the forming new connections.
1:31:48.080 --> 1:31:51.080
So this idea of synaptogenesis, the growth of new synapses,
1:31:51.080 --> 1:31:54.080
that's well described, as well understood.
1:31:54.080 --> 1:31:56.080
And that's an essential part of learning.
1:31:56.080 --> 1:31:57.080
That is learning.
1:31:57.080 --> 1:31:58.080
That is learning.
1:31:58.080 --> 1:32:00.080
Okay.
1:32:00.080 --> 1:32:04.080
You know, back, you know, going back many, many years, people, you know,
1:32:04.080 --> 1:32:08.080
was, what's his name, the psychologist proposed,
1:32:08.080 --> 1:32:10.080
Hebb, Donald Hebb.
1:32:10.080 --> 1:32:13.080
He proposed that learning was the modification of the strength
1:32:13.080 --> 1:32:15.080
of a connection between two neurons.
1:32:15.080 --> 1:32:19.080
People interpreted that as the modification of the strength of a synapse.
1:32:19.080 --> 1:32:21.080
He didn't say that.
1:32:21.080 --> 1:32:24.080
He just said there's a modification between the effect of one neuron and another.
1:32:24.080 --> 1:32:28.080
So synaptogenesis is totally consistent with Donald Hebb said.
1:32:28.080 --> 1:32:31.080
But anyway, there's these mechanisms, the growth of new synapse.
1:32:31.080 --> 1:32:34.080
You can go online, you can watch a video of a synapse growing in real time.
1:32:34.080 --> 1:32:36.080
It's literally, you can see this little thing going.
1:32:36.080 --> 1:32:38.080
It's pretty impressive.
1:32:38.080 --> 1:32:40.080
So that those mechanisms are known.
1:32:40.080 --> 1:32:43.080
Now, there's another thing that we've speculated and we've written about,
1:32:43.080 --> 1:32:48.080
which is consistent with no neuroscience, but it's less proven.
1:32:48.080 --> 1:32:49.080
And this is the idea.
1:32:49.080 --> 1:32:51.080
How do I form a memory really, really quickly?
1:32:51.080 --> 1:32:53.080
Like instantaneous.
1:32:53.080 --> 1:32:56.080
If it takes an hour to grow a synapse, like that's not instantaneous.
1:32:56.080 --> 1:33:01.080
So there are types of synapses called silent synapses.
1:33:01.080 --> 1:33:04.080
They look like a synapse, but they don't do anything.
1:33:04.080 --> 1:33:05.080
They're just sitting there.
1:33:05.080 --> 1:33:10.080
It's like if an action potential comes in, it doesn't release any neurotransmitter.
1:33:10.080 --> 1:33:12.080
Some parts of the brain have more of these than others.
1:33:12.080 --> 1:33:14.080
For example, the hippocampus has a lot of them,
1:33:14.080 --> 1:33:18.080
which is where we associate most short term memory with.
1:33:18.080 --> 1:33:22.080
So what we speculated, again, in that 2016 paper,
1:33:22.080 --> 1:33:26.080
we proposed that the way we form very quick memories,
1:33:26.080 --> 1:33:29.080
very short term memories, or quick memories,
1:33:29.080 --> 1:33:34.080
is that we convert silent synapses into active synapses.
1:33:34.080 --> 1:33:38.080
It's like saying a synapse has a zero weight and a one weight.
1:33:38.080 --> 1:33:41.080
But the long term memory has to be formed by synaptogenesis.
1:33:41.080 --> 1:33:43.080
So you can remember something really quickly
1:33:43.080 --> 1:33:46.080
by just flipping a bunch of these guys from silent to active.
1:33:46.080 --> 1:33:49.080
It's not from 0.1 to 0.15.
1:33:49.080 --> 1:33:52.080
It doesn't do anything until it releases transmitter.
1:33:52.080 --> 1:33:56.080
If I do that over a bunch of these, I've got a very quick short term memory.
1:33:56.080 --> 1:34:01.080
So I guess the lesson behind this is that most neural networks today are fully connected.
1:34:01.080 --> 1:34:04.080
Every neuron connects every other neuron from layer to layer.
1:34:04.080 --> 1:34:06.080
That's not correct in the brain.
1:34:06.080 --> 1:34:07.080
We don't want that.
1:34:07.080 --> 1:34:08.080
We actually don't want that.
1:34:08.080 --> 1:34:09.080
It's bad.
1:34:09.080 --> 1:34:13.080
You want a very sparse connectivity so that any neuron connects
1:34:13.080 --> 1:34:15.080
to some subset of the neurons in the other layer,
1:34:15.080 --> 1:34:18.080
and it does so on a dendrite by dendrite segment basis.
1:34:18.080 --> 1:34:21.080
So it's a very parcelated out type of thing.
1:34:21.080 --> 1:34:25.080
And that then learning is not adjusting all these weights,
1:34:25.080 --> 1:34:29.080
but learning is just saying, OK, connect to these 10 cells here right now.
1:34:29.080 --> 1:34:32.080
In that process, you know, with artificial neural networks,
1:34:32.080 --> 1:34:37.080
it's a very simple process of back propagation that adjusts the weights.
1:34:37.080 --> 1:34:39.080
The process of synaptogenesis.
1:34:39.080 --> 1:34:40.080
Synaptogenesis.
1:34:40.080 --> 1:34:41.080
Synaptogenesis.
1:34:41.080 --> 1:34:42.080
It's even easier.
1:34:42.080 --> 1:34:43.080
It's even easier.
1:34:43.080 --> 1:34:44.080
It's even easier.
1:34:44.080 --> 1:34:49.080
Back propagation requires something that really can't happen in brains.
1:34:49.080 --> 1:34:51.080
This back propagation of this error signal.
1:34:51.080 --> 1:34:52.080
They really can't happen.
1:34:52.080 --> 1:34:55.080
People are trying to make it happen in brains, but it doesn't happen in brain.
1:34:55.080 --> 1:34:57.080
This is pure Hebbian learning.
1:34:57.080 --> 1:34:59.080
Well, synaptogenesis is pure Hebbian learning.
1:34:59.080 --> 1:35:03.080
It's basically saying there's a population of cells over here that are active right now,
1:35:03.080 --> 1:35:05.080
and there's a population of cells over here active right now.
1:35:05.080 --> 1:35:08.080
How do I form connections between those active cells?
1:35:08.080 --> 1:35:11.080
And it's literally saying this guy became active.
1:35:11.080 --> 1:35:15.080
These 100 neurons here became active before this neuron became active.
1:35:15.080 --> 1:35:17.080
So form connections to those ones.
1:35:17.080 --> 1:35:18.080
That's it.
1:35:18.080 --> 1:35:20.080
There's no propagation of error, nothing.
1:35:20.080 --> 1:35:26.080
All the networks we do, all models we have work on almost completely on Hebbian learning,
1:35:26.080 --> 1:35:33.080
but on dendritic segments and multiple synaptoses at the same time.
1:35:33.080 --> 1:35:36.080
So now let's turn the question that you already answered,
1:35:36.080 --> 1:35:38.080
and maybe you can answer it again.
1:35:38.080 --> 1:35:43.080
If you look at the history of artificial intelligence, where do you think we stand?
1:35:43.080 --> 1:35:45.080
How far are we from solving intelligence?
1:35:45.080 --> 1:35:47.080
You said you were very optimistic.
1:35:47.080 --> 1:35:48.080
Yeah.
1:35:48.080 --> 1:35:49.080
Can you elaborate on that?
1:35:49.080 --> 1:35:55.080
Yeah, it's always the crazy question to ask, because no one can predict the future.
1:35:55.080 --> 1:35:56.080
Absolutely.
1:35:56.080 --> 1:35:58.080
So I'll tell you a story.
1:35:58.080 --> 1:36:02.080
I used to run a different neuroscience institute called the Redwood Neuroscience Institute,
1:36:02.080 --> 1:36:07.080
and we would hold these symposiums, and we'd get like 35 scientists from around the world to come together.
1:36:07.080 --> 1:36:09.080
And I used to ask them all the same question.
1:36:09.080 --> 1:36:13.080
I would say, well, how long do you think it'll be before we understand how the New York Cortex works?
1:36:13.080 --> 1:36:17.080
And everyone went around the room, and they had introduced the name, and they had to answer that question.
1:36:17.080 --> 1:36:22.080
So I got, the typical answer was 50 to 100 years.
1:36:22.080 --> 1:36:24.080
Some people would say 500 years.
1:36:24.080 --> 1:36:25.080
Some people said never.
1:36:25.080 --> 1:36:27.080
I said, why are you a neuroscience institute?
1:36:27.080 --> 1:36:28.080
Never.
1:36:28.080 --> 1:36:30.080
It's good pay.
1:36:30.080 --> 1:36:33.080
It's interesting.
1:36:33.080 --> 1:36:37.080
But it doesn't work like that.
1:36:37.080 --> 1:36:39.080
As I mentioned earlier, these are step functions.
1:36:39.080 --> 1:36:41.080
Things happen, and then bingo, they happen.
1:36:41.080 --> 1:36:43.080
You can't predict that.
1:36:43.080 --> 1:36:45.080
I feel I've already passed a step function.
1:36:45.080 --> 1:36:53.080
So if I can do my job correctly over the next five years, then meaning I can proselytize these ideas.
1:36:53.080 --> 1:36:55.080
I can convince other people they're right.
1:36:55.080 --> 1:37:01.080
We can show that machine learning people should pay attention to these ideas.
1:37:01.080 --> 1:37:04.080
Then we're definitely in an under 20 year time frame.
1:37:04.080 --> 1:37:09.080
If I can do those things, if I'm not successful in that, and this is the last time anyone talks to me,
1:37:09.080 --> 1:37:15.080
and no one reads our papers, and I'm wrong or something like that, then I don't know.
1:37:15.080 --> 1:37:17.080
But it's not 50 years.
1:37:17.080 --> 1:37:22.080
It's the same thing about electric cars.
1:37:22.080 --> 1:37:24.080
How quickly are they going to populate the world?
1:37:24.080 --> 1:37:27.080
It probably takes about a 20 year span.
1:37:27.080 --> 1:37:28.080
It'll be something like that.
1:37:28.080 --> 1:37:31.080
But I think if I can do what I said, we're starting it.
1:37:31.080 --> 1:37:35.080
Of course, there could be other use of step functions.
1:37:35.080 --> 1:37:42.080
It could be everybody gives up on your ideas for 20 years, and then all of a sudden somebody picks it up again.
1:37:42.080 --> 1:37:44.080
Wait, that guy was onto something.
1:37:44.080 --> 1:37:46.080
That would be a failure on my part.
1:37:46.080 --> 1:37:49.080
Think about Charles Babbage.
1:37:49.080 --> 1:37:55.080
Charles Babbage used to invented the computer back in the 1800s.
1:37:55.080 --> 1:37:59.080
Everyone forgot about it until 100 years later.
1:37:59.080 --> 1:38:03.080
This guy figured this stuff out a long time ago, but he was ahead of his time.
1:38:03.080 --> 1:38:09.080
As I said, I recognize this is part of any entrepreneur's challenge.
1:38:09.080 --> 1:38:11.080
I use entrepreneur broadly in this case.
1:38:11.080 --> 1:38:13.080
I'm not meaning like I'm building a business trying to sell something.
1:38:13.080 --> 1:38:15.080
I'm trying to sell ideas.
1:38:15.080 --> 1:38:20.080
This is the challenge as to how you get people to pay attention to you.
1:38:20.080 --> 1:38:24.080
How do you get them to give you positive or negative feedback?
1:38:24.080 --> 1:38:27.080
How do you get the people to act differently based on your ideas?
1:38:27.080 --> 1:38:30.080
We'll see how well we do on that.
1:38:30.080 --> 1:38:34.080
There's a lot of hype behind artificial intelligence currently.
1:38:34.080 --> 1:38:43.080
Do you, as you look to spread the ideas that are in your cortical theory, the things you're working on,
1:38:43.080 --> 1:38:47.080
do you think there's some possibility we'll hit an AI winter once again?
1:38:47.080 --> 1:38:49.080
It's certainly a possibility.
1:38:49.080 --> 1:38:51.080
That's something you worry about?
1:38:51.080 --> 1:38:54.080
I guess, do I worry about it?
1:38:54.080 --> 1:38:58.080
I haven't decided yet if that's good or bad for my mission.
1:38:58.080 --> 1:38:59.080
That's true.
1:38:59.080 --> 1:39:04.080
That's very true because it's almost like you need the winter to refresh the pallet.
1:39:04.080 --> 1:39:08.080
Here's what you want to have it.
1:39:08.080 --> 1:39:15.080
To the extent that everyone is so thrilled about the current state of machine learning and AI,
1:39:15.080 --> 1:39:20.080
and they don't imagine they need anything else, it makes my job harder.
1:39:20.080 --> 1:39:24.080
If everything crashed completely and every student left the field,
1:39:24.080 --> 1:39:26.080
and there was no money for anybody to do anything,
1:39:26.080 --> 1:39:29.080
and it became an embarrassment to talk about machine intelligence and AI,
1:39:29.080 --> 1:39:31.080
that wouldn't be good for us either.
1:39:31.080 --> 1:39:33.080
You want the soft landing approach, right?
1:39:33.080 --> 1:39:37.080
You want enough people, the senior people in AI and machine learning to say,
1:39:37.080 --> 1:39:39.080
you know, we need other approaches.
1:39:39.080 --> 1:39:40.080
We really need other approaches.
1:39:40.080 --> 1:39:42.080
Damn, we need other approaches.
1:39:42.080 --> 1:39:43.080
Maybe we should look to the brain.
1:39:43.080 --> 1:39:44.080
Okay, let's look to the brain.
1:39:44.080 --> 1:39:45.080
Who's got some brain ideas?
1:39:45.080 --> 1:39:49.080
Okay, let's start a little project on the side here trying to do brain idea related stuff.
1:39:49.080 --> 1:39:51.080
That's the ideal outcome we would want.
1:39:51.080 --> 1:39:53.080
So I don't want a total winter,
1:39:53.080 --> 1:39:57.080
and yet I don't want it to be sunny all the time either.
1:39:57.080 --> 1:40:02.080
So what do you think it takes to build a system with human level intelligence
1:40:02.080 --> 1:40:06.080
where once demonstrated, you would be very impressed?
1:40:06.080 --> 1:40:08.080
So does it have to have a body?
1:40:08.080 --> 1:40:18.080
Does it have to have the C word we used before consciousness as an entirety in a holistic sense?
1:40:18.080 --> 1:40:23.080
First of all, I don't think the goal is to create a machine that is human level intelligence.
1:40:23.080 --> 1:40:24.080
I think it's a false goal.
1:40:24.080 --> 1:40:26.080
Back to Turing, I think it was a false statement.
1:40:26.080 --> 1:40:28.080
We want to understand what intelligence is,
1:40:28.080 --> 1:40:31.080
and then we can build intelligent machines of all different scales,
1:40:31.080 --> 1:40:33.080
all different capabilities.
1:40:33.080 --> 1:40:35.080
You know, a dog is intelligent.
1:40:35.080 --> 1:40:38.080
You know, that would be pretty good to have a dog, you know,
1:40:38.080 --> 1:40:41.080
but what about something that doesn't look like an animal at all in different spaces?
1:40:41.080 --> 1:40:45.080
So my thinking about this is that we want to define what intelligence is,
1:40:45.080 --> 1:40:48.080
agree upon what makes an intelligent system.
1:40:48.080 --> 1:40:52.080
We can then say, okay, we're now going to build systems that work on those principles,
1:40:52.080 --> 1:40:57.080
or some subset of them, and we can apply them to all different types of problems.
1:40:57.080 --> 1:41:00.080
And the kind, the idea, it's not computing.
1:41:00.080 --> 1:41:05.080
We don't ask, if I take a little, you know, little one chip computer,
1:41:05.080 --> 1:41:08.080
I don't say, well, that's not a computer because it's not as powerful as this, you know,
1:41:08.080 --> 1:41:09.080
big server over here.
1:41:09.080 --> 1:41:11.080
No, no, because we know that what the principles are computing on,
1:41:11.080 --> 1:41:14.080
and I can apply those principles to a small problem or into a big problem.
1:41:14.080 --> 1:41:16.080
And same, intelligence needs to get there.
1:41:16.080 --> 1:41:17.080
We have to say, these are the principles.
1:41:17.080 --> 1:41:19.080
I can make a small one, a big one.
1:41:19.080 --> 1:41:20.080
I can make them distributed.
1:41:20.080 --> 1:41:21.080
I can put them on different sensors.
1:41:21.080 --> 1:41:23.080
They don't have to be human like at all.
1:41:23.080 --> 1:41:25.080
Now, you did bring up a very interesting question about embodiment.
1:41:25.080 --> 1:41:27.080
Does it have to have a body?
1:41:27.080 --> 1:41:30.080
It has to have some concept of movement.
1:41:30.080 --> 1:41:33.080
It has to be able to move through these reference frames.
1:41:33.080 --> 1:41:35.080
I talked about earlier, whether it's physically moving,
1:41:35.080 --> 1:41:38.080
like I need, if I'm going to have an AI that understands coffee cups,
1:41:38.080 --> 1:41:42.080
it's going to have to pick up the coffee cup and touch it and look at it with its eyes
1:41:42.080 --> 1:41:45.080
and hands or something equivalent to that.
1:41:45.080 --> 1:41:51.080
If I have a mathematical AI, maybe it needs to move through mathematical spaces.
1:41:51.080 --> 1:41:55.080
I could have a virtual AI that lives in the Internet
1:41:55.080 --> 1:42:00.080
and its movements are traversing links and digging into files,
1:42:00.080 --> 1:42:04.080
but it's got a location that it's traveling through some space.
1:42:04.080 --> 1:42:08.080
You can't have an AI that just takes some flash thing input,
1:42:08.080 --> 1:42:10.080
you know, we call it flash inference.
1:42:10.080 --> 1:42:13.080
Here's a pattern done.
1:42:13.080 --> 1:42:16.080
No, it's movement pattern, movement pattern, movement pattern.
1:42:16.080 --> 1:42:18.080
Attention, digging, building, building structure,
1:42:18.080 --> 1:42:20.080
just trying to figure out the model of the world.
1:42:20.080 --> 1:42:25.080
So some sort of embodiment, whether it's physical or not, has to be part of it.
1:42:25.080 --> 1:42:28.080
So self awareness in the way to be able to answer where am I?
1:42:28.080 --> 1:42:31.080
You bring up self awareness, it's a different topic, self awareness.
1:42:31.080 --> 1:42:37.080
No, the very narrow definition, meaning knowing a sense of self enough to know
1:42:37.080 --> 1:42:40.080
where am I in the space where essentially.
1:42:40.080 --> 1:42:43.080
The system needs to know its location,
1:42:43.080 --> 1:42:48.080
where each component of the system needs to know where it is in the world at that point in time.
1:42:48.080 --> 1:42:51.080
So self awareness and consciousness.
1:42:51.080 --> 1:42:56.080
Do you think, one, from the perspective of neuroscience and neurocortex,
1:42:56.080 --> 1:42:59.080
these are interesting topics, solvable topics,
1:42:59.080 --> 1:43:04.080
do you have any ideas of why the heck it is that we have a subjective experience at all?
1:43:04.080 --> 1:43:05.080
Yeah, I have a lot of questions.
1:43:05.080 --> 1:43:08.080
And is it useful, or is it just a side effect of us?
1:43:08.080 --> 1:43:10.080
It's interesting to think about.
1:43:10.080 --> 1:43:16.080
I don't think it's useful as a means to figure out how to build intelligent machines.
1:43:16.080 --> 1:43:21.080
It's something that systems do, and we can talk about what it is,
1:43:21.080 --> 1:43:25.080
that are like, well, if I build a system like this, then it would be self aware.
1:43:25.080 --> 1:43:28.080
Or if I build it like this, it wouldn't be self aware.
1:43:28.080 --> 1:43:30.080
So that's a choice I can have.
1:43:30.080 --> 1:43:32.080
It's not like, oh my God, it's self aware.
1:43:32.080 --> 1:43:37.080
I heard an interview recently with this philosopher from Yale.
1:43:37.080 --> 1:43:38.080
I can't remember his name.
1:43:38.080 --> 1:43:39.080
I apologize for that.
1:43:39.080 --> 1:43:41.080
But he was talking about, well, if these computers were self aware,
1:43:41.080 --> 1:43:43.080
then it would be a crime to unplug them.
1:43:43.080 --> 1:43:45.080
And I'm like, oh, come on.
1:43:45.080 --> 1:43:46.080
I unplug myself every night.
1:43:46.080 --> 1:43:47.080
I go to sleep.
1:43:47.080 --> 1:43:49.080
Is that a crime?
1:43:49.080 --> 1:43:51.080
I plug myself in again in the morning.
1:43:51.080 --> 1:43:53.080
There I am.
1:43:53.080 --> 1:43:56.080
People get kind of bent out of shape about this.
1:43:56.080 --> 1:44:02.080
I have very detailed understanding or opinions about what it means to be conscious
1:44:02.080 --> 1:44:04.080
and what it means to be self aware.
1:44:04.080 --> 1:44:07.080
I don't think it's that interesting a problem.
1:44:07.080 --> 1:44:08.080
You talked about Christoph Koch.
1:44:08.080 --> 1:44:10.080
He thinks that's the only problem.
1:44:10.080 --> 1:44:12.080
I didn't actually listen to your interview with him.
1:44:12.080 --> 1:44:15.080
But I know him, and I know that's the thing.
1:44:15.080 --> 1:44:18.080
He also thinks intelligence and consciousness are disjoint.
1:44:18.080 --> 1:44:21.080
So I mean, it's not, you don't have to have one or the other.
1:44:21.080 --> 1:44:23.080
I just agree with that.
1:44:23.080 --> 1:44:24.080
I just totally disagree with that.
1:44:24.080 --> 1:44:26.080
So where's your thoughts and consciousness?
1:44:26.080 --> 1:44:28.080
Where does it emerge from?
1:44:28.080 --> 1:44:30.080
Then we have to break it down to the two parts.
1:44:30.080 --> 1:44:32.080
Because consciousness isn't one thing.
1:44:32.080 --> 1:44:34.080
That's part of the problem with that term.
1:44:34.080 --> 1:44:36.080
It means different things to different people.
1:44:36.080 --> 1:44:38.080
And there's different components of it.
1:44:38.080 --> 1:44:40.080
There is a concept of self awareness.
1:44:40.080 --> 1:44:44.080
That can be very easily explained.
1:44:44.080 --> 1:44:46.080
You have a model of your own body.
1:44:46.080 --> 1:44:48.080
The neocortex models things in the world.
1:44:48.080 --> 1:44:50.080
And it also models your own body.
1:44:50.080 --> 1:44:53.080
And then it has a memory.
1:44:53.080 --> 1:44:55.080
It can remember what you've done.
1:44:55.080 --> 1:44:57.080
So it can remember what you did this morning.
1:44:57.080 --> 1:44:59.080
It can remember what you had for breakfast and so on.
1:44:59.080 --> 1:45:02.080
And so I can say to you, okay, Lex,
1:45:02.080 --> 1:45:06.080
were you conscious this morning when you had your bagel?
1:45:06.080 --> 1:45:08.080
And you'd say, yes, I was conscious.
1:45:08.080 --> 1:45:11.080
Now, what if I could take your brain and revert all the synapses
1:45:11.080 --> 1:45:13.080
back to the state they were this morning?
1:45:13.080 --> 1:45:15.080
And then I said to you, Lex,
1:45:15.080 --> 1:45:17.080
were you conscious when you ate the bagel?
1:45:17.080 --> 1:45:18.080
And he said, no, I wasn't conscious.
1:45:18.080 --> 1:45:20.080
I said, here's a video of eating the bagel.
1:45:20.080 --> 1:45:22.080
And he said, I wasn't there.
1:45:22.080 --> 1:45:25.080
That's not possible because I must have been unconscious at that time.
1:45:25.080 --> 1:45:27.080
So we can just make this one to one correlation
1:45:27.080 --> 1:45:30.080
between memory of your body's trajectory through the world
1:45:30.080 --> 1:45:32.080
over some period of time.
1:45:32.080 --> 1:45:34.080
And the ability to recall that memory
1:45:34.080 --> 1:45:36.080
is what you would call conscious.
1:45:36.080 --> 1:45:38.080
I was conscious of that. It's self awareness.
1:45:38.080 --> 1:45:41.080
And any system that can recall,
1:45:41.080 --> 1:45:43.080
memorize what it's done recently
1:45:43.080 --> 1:45:46.080
and bring that back and invoke it again
1:45:46.080 --> 1:45:48.080
would say, yeah, I'm aware.
1:45:48.080 --> 1:45:51.080
I remember what I did. All right, I got it.
1:45:51.080 --> 1:45:54.080
That's an easy one. Although some people think that's a hard one.
1:45:54.080 --> 1:45:57.080
The more challenging part of consciousness
1:45:57.080 --> 1:45:59.080
is this is one that's sometimes used
1:45:59.080 --> 1:46:01.080
by the word Aqualia,
1:46:01.080 --> 1:46:04.080
which is, you know, why does an object seem red?
1:46:04.080 --> 1:46:06.080
Or what is pain?
1:46:06.080 --> 1:46:08.080
And why does pain feel like something?
1:46:08.080 --> 1:46:10.080
Why do I feel redness?
1:46:10.080 --> 1:46:12.080
So why do I feel a little painless in a way?
1:46:12.080 --> 1:46:14.080
And then I could say, well, why does sight
1:46:14.080 --> 1:46:16.080
seems different than hearing? You know, it's the same problem.
1:46:16.080 --> 1:46:18.080
It's really, you know, these are all just neurons.
1:46:18.080 --> 1:46:21.080
And so how is it that why does looking at you
1:46:21.080 --> 1:46:24.080
feel different than, you know, hearing you?
1:46:24.080 --> 1:46:26.080
It feels different, but there's just neurons in my head.
1:46:26.080 --> 1:46:28.080
They're all doing the same thing.
1:46:28.080 --> 1:46:30.080
So that's an interesting question.
1:46:30.080 --> 1:46:32.080
The best treatise I've read about this
1:46:32.080 --> 1:46:34.080
is by a guy named Oregon.
1:46:34.080 --> 1:46:38.080
He wrote a book called Why Red Doesn't Sound Like a Bell.
1:46:38.080 --> 1:46:42.080
It's a little, it's not a trade book, easy to read,
1:46:42.080 --> 1:46:45.080
but it, and it's an interesting question.
1:46:45.080 --> 1:46:47.080
Take something like color.
1:46:47.080 --> 1:46:49.080
Color really doesn't exist in the world.
1:46:49.080 --> 1:46:51.080
It's not a property of the world.
1:46:51.080 --> 1:46:54.080
Property of the world that exists is light frequency,
1:46:54.080 --> 1:46:57.080
and that gets turned into we have certain cells
1:46:57.080 --> 1:46:59.080
in the retina that respond to different frequencies
1:46:59.080 --> 1:47:00.080
different than others.
1:47:00.080 --> 1:47:02.080
And so when they enter the brain, you just have a bunch
1:47:02.080 --> 1:47:04.080
of axons that are firing at different rates,
1:47:04.080 --> 1:47:06.080
and from that we perceive color.
1:47:06.080 --> 1:47:08.080
But there is no color in the brain.
1:47:08.080 --> 1:47:11.080
I mean, there's no color coming in on those synapses.
1:47:11.080 --> 1:47:14.080
It's just a correlation between some axons
1:47:14.080 --> 1:47:17.080
and some property of frequency.
1:47:17.080 --> 1:47:19.080
And that isn't even color itself.
1:47:19.080 --> 1:47:21.080
Frequency doesn't have a color.
1:47:21.080 --> 1:47:23.080
It's just what it is.
1:47:23.080 --> 1:47:25.080
So then the question is, well, why does it even
1:47:25.080 --> 1:47:27.080
appear to have a color at all?
1:47:27.080 --> 1:47:30.080
Just as you're describing it, there seems to be a connection
1:47:30.080 --> 1:47:32.080
to those ideas of reference frames.
1:47:32.080 --> 1:47:38.080
I mean, it just feels like consciousness having the subject,
1:47:38.080 --> 1:47:42.080
assigning the feeling of red to the actual color
1:47:42.080 --> 1:47:47.080
or to the wavelength is useful for intelligence.
1:47:47.080 --> 1:47:49.080
Yeah, I think that's a good way of putting it.
1:47:49.080 --> 1:47:51.080
It's useful as a predictive mechanism,
1:47:51.080 --> 1:47:53.080
or useful as a generalization idea.
1:47:53.080 --> 1:47:55.080
It's a way of grouping things together to say
1:47:55.080 --> 1:47:58.080
it's useful to have a model like this.
1:47:58.080 --> 1:48:02.080
Think about the well known syndrome that people
1:48:02.080 --> 1:48:06.080
who've lost a limb experience called phantom limbs.
1:48:06.080 --> 1:48:11.080
And what they claim is they can have their arms removed
1:48:11.080 --> 1:48:13.080
but they feel the arm.
1:48:13.080 --> 1:48:15.080
Not only feel it, they know it's there.
1:48:15.080 --> 1:48:16.080
It's there.
1:48:16.080 --> 1:48:17.080
I know it's there.
1:48:17.080 --> 1:48:19.080
They'll swear to you that it's there.
1:48:19.080 --> 1:48:20.080
And then they can feel pain in their arm.
1:48:20.080 --> 1:48:22.080
And they'll feel pain in their finger.
1:48:22.080 --> 1:48:25.080
And if they move their non existent arm behind their back,
1:48:25.080 --> 1:48:27.080
then they feel the pain behind their back.
1:48:27.080 --> 1:48:30.080
So this whole idea that your arm exists
1:48:30.080 --> 1:48:31.080
is a model of your brain.
1:48:31.080 --> 1:48:34.080
It may or may not really exist.
1:48:34.080 --> 1:48:38.080
And just like, but it's useful to have a model of something
1:48:38.080 --> 1:48:40.080
that sort of correlates to things in the world
1:48:40.080 --> 1:48:42.080
so you can make predictions about what would happen
1:48:42.080 --> 1:48:43.080
when those things occur.
1:48:43.080 --> 1:48:44.080
It's a little bit of a fuzzy,
1:48:44.080 --> 1:48:46.080
but I think you're getting quite towards the answer there.
1:48:46.080 --> 1:48:51.080
It's useful for the model to express things certain ways
1:48:51.080 --> 1:48:53.080
that we can then map them into these reference frames
1:48:53.080 --> 1:48:55.080
and make predictions about them.
1:48:55.080 --> 1:48:57.080
I need to spend more time on this topic.
1:48:57.080 --> 1:48:58.080
It doesn't bother me.
1:48:58.080 --> 1:49:00.080
Do you really need to spend more time on this?
1:49:00.080 --> 1:49:01.080
Yeah.
1:49:01.080 --> 1:49:04.080
It does feel special that we have subjective experience,
1:49:04.080 --> 1:49:07.080
but I'm yet to know why.
1:49:07.080 --> 1:49:08.080
I'm just personally curious.
1:49:08.080 --> 1:49:10.080
It's not necessary for the work we're doing here.
1:49:10.080 --> 1:49:12.080
I don't think I need to solve that problem
1:49:12.080 --> 1:49:14.080
to build intelligent machines at all.
1:49:14.080 --> 1:49:15.080
Not at all.
1:49:15.080 --> 1:49:19.080
But there is sort of the silly notion that you described briefly
1:49:19.080 --> 1:49:22.080
that doesn't seem so silly to us humans is,
1:49:22.080 --> 1:49:26.080
you know, if you're successful building intelligent machines,
1:49:26.080 --> 1:49:29.080
it feels wrong to then turn them off.
1:49:29.080 --> 1:49:32.080
Because if you're able to build a lot of them,
1:49:32.080 --> 1:49:36.080
it feels wrong to then be able to, you know,
1:49:36.080 --> 1:49:38.080
to turn off the...
1:49:38.080 --> 1:49:39.080
Well, why?
1:49:39.080 --> 1:49:41.080
Let's break that down a bit.
1:49:41.080 --> 1:49:43.080
As humans, why do we fear death?
1:49:43.080 --> 1:49:46.080
There's two reasons we fear death.
1:49:46.080 --> 1:49:48.080
Well, first of all, I'll say when you're dead, it doesn't matter.
1:49:48.080 --> 1:49:49.080
Okay.
1:49:49.080 --> 1:49:50.080
You're dead.
1:49:50.080 --> 1:49:51.080
So why do we fear death?
1:49:51.080 --> 1:49:53.080
We fear death for two reasons.
1:49:53.080 --> 1:49:57.080
One is because we are programmed genetically to fear death.
1:49:57.080 --> 1:50:02.080
That's a survival and propaganda in the genes thing.
1:50:02.080 --> 1:50:06.080
And we also are programmed to feel sad when people we know die.
1:50:06.080 --> 1:50:08.080
We don't feel sad for someone we don't know dies.
1:50:08.080 --> 1:50:09.080
There's people dying right now.
1:50:09.080 --> 1:50:10.080
They're only scared to say,
1:50:10.080 --> 1:50:11.080
I don't feel bad about them because I don't know them.
1:50:11.080 --> 1:50:13.080
But I knew they might feel really bad.
1:50:13.080 --> 1:50:18.080
So again, these are old brain genetically embedded things
1:50:18.080 --> 1:50:20.080
that we fear death.
1:50:20.080 --> 1:50:23.080
Outside of those uncomfortable feelings,
1:50:23.080 --> 1:50:25.080
there's nothing else to worry about.
1:50:25.080 --> 1:50:27.080
Well, wait, hold on a second.
1:50:27.080 --> 1:50:30.080
Do you know the denial of death by Beckard?
1:50:30.080 --> 1:50:36.080
You know, there's a thought that death is, you know,
1:50:36.080 --> 1:50:43.080
our whole conception of our world model kind of assumes immortality.
1:50:43.080 --> 1:50:47.080
And then death is this terror that underlies it all.
1:50:47.080 --> 1:50:50.080
So like, well, some people's world model, not mine.
1:50:50.080 --> 1:50:51.080
But okay.
1:50:51.080 --> 1:50:54.080
So what, what Becker would say is that you're just living in an illusion.
1:50:54.080 --> 1:50:58.080
You've constructed illusion for yourself because it's such a terrible terror.
1:50:58.080 --> 1:51:02.080
The fact that this illusion, the illusion that death doesn't matter.
1:51:02.080 --> 1:51:05.080
You're still not coming to grips with the illusion of what?
1:51:05.080 --> 1:51:08.080
That death is going to happen.
1:51:08.080 --> 1:51:10.080
Oh, it's not going to happen.
1:51:10.080 --> 1:51:11.080
You're, you're actually operating.
1:51:11.080 --> 1:51:13.080
You haven't, even though you said you've accepted it,
1:51:13.080 --> 1:51:16.080
you haven't really accepted the notion of death is what you say.
1:51:16.080 --> 1:51:21.080
So it sounds like it sounds like you disagree with that notion.
1:51:21.080 --> 1:51:22.080
I mean, totally.
1:51:22.080 --> 1:51:27.080
Like, I literally every night, every night I go to bed, it's like dying.
1:51:27.080 --> 1:51:28.080
Little deaths.
1:51:28.080 --> 1:51:29.080
Little deaths.
1:51:29.080 --> 1:51:32.080
And if I didn't wake up, it wouldn't matter to me.
1:51:32.080 --> 1:51:35.080
Only if I knew that was going to happen would it be bothersome.
1:51:35.080 --> 1:51:37.080
If I didn't know it was going to happen, how would I know?
1:51:37.080 --> 1:51:39.080
Then I would worry about my wife.
1:51:39.080 --> 1:51:40.080
Yeah.
1:51:40.080 --> 1:51:44.080
So imagine, imagine I was a loner and I lived in Alaska and I lived them out there and there
1:51:44.080 --> 1:51:45.080
was no animals.
1:51:45.080 --> 1:51:46.080
Nobody knew I existed.
1:51:46.080 --> 1:51:48.080
I was just eating these roots all the time.
1:51:48.080 --> 1:51:50.080
And nobody knew I was there.
1:51:50.080 --> 1:51:53.080
And one day I didn't wake up.
1:51:53.080 --> 1:51:56.080
Where, what, what pain in the world would there exist?
1:51:56.080 --> 1:52:01.080
Well, so most people that think about this problem would say that you're just deeply enlightened
1:52:01.080 --> 1:52:04.080
or are completely delusional.
1:52:04.080 --> 1:52:05.080
Wow.
1:52:05.080 --> 1:52:14.080
But I would say, I would say that's a very enlightened way to see the world is that that's the rational one.
1:52:14.080 --> 1:52:15.080
Well, I think it's rational.
1:52:15.080 --> 1:52:16.080
That's right.
1:52:16.080 --> 1:52:22.080
But the fact is we don't, I mean, we really don't have an understanding of why the heck
1:52:22.080 --> 1:52:26.080
it is we're born and why we die and what happens after we die.
1:52:26.080 --> 1:52:27.080
Well, maybe there isn't a reason.
1:52:27.080 --> 1:52:28.080
Maybe there is.
1:52:28.080 --> 1:52:30.080
So I'm interested in those big problems too, right?
1:52:30.080 --> 1:52:33.080
You know, you, you interviewed Max Tagmark, you know, and there's people like that, right?
1:52:33.080 --> 1:52:35.080
I'm interested in those big problems as well.
1:52:35.080 --> 1:52:41.080
And in fact, when I was young, I made a list of the biggest problems I could think of.
1:52:41.080 --> 1:52:43.080
First, why does anything exist?
1:52:43.080 --> 1:52:46.080
Second, why did we have the laws of physics that we have?
1:52:46.080 --> 1:52:49.080
Third, is life inevitable?
1:52:49.080 --> 1:52:50.080
And why is it here?
1:52:50.080 --> 1:52:52.080
Fourth, is intelligence inevitable?
1:52:52.080 --> 1:52:53.080
And why is it here?
1:52:53.080 --> 1:52:58.080
I stopped there because I figured if you can make a truly intelligent system, we'll be,
1:52:58.080 --> 1:53:03.080
that'll be the quickest way to answer the first three questions.
1:53:03.080 --> 1:53:04.080
I'm serious.
1:53:04.080 --> 1:53:05.080
Yeah.
1:53:05.080 --> 1:53:09.080
And so I said, my mission, you know, you asked me earlier, my first mission is to understand
1:53:09.080 --> 1:53:12.080
the brain, but I felt that is the shortest way to get to true machine intelligence.
1:53:12.080 --> 1:53:16.080
And I want to get to true machine intelligence because even if it doesn't occur in my lifetime,
1:53:16.080 --> 1:53:19.080
other people will benefit from it because I think it'll occur in my lifetime.
1:53:19.080 --> 1:53:21.080
But, you know, 20 years, you never know.
1:53:21.080 --> 1:53:27.080
And, but that will be the quickest way for us to, you know, we can make super mathematicians.
1:53:27.080 --> 1:53:29.080
We can make super space explorers.
1:53:29.080 --> 1:53:36.080
We can make super physicists brains that do these things and that can run experiments
1:53:36.080 --> 1:53:37.080
that we can't run.
1:53:37.080 --> 1:53:40.080
We don't have the abilities to manipulate things and so on.
1:53:40.080 --> 1:53:42.080
But we can build and tell the machines to do all those things.
1:53:42.080 --> 1:53:48.080
And with the ultimate goal of finding out the answers to the other questions.
1:53:48.080 --> 1:53:56.080
Let me ask, you know, the depressing and difficult question, which is once we achieve that goal,
1:53:56.080 --> 1:54:03.080
do you, of creating, no, of understanding intelligence, do you think we would be happier,
1:54:03.080 --> 1:54:05.080
more fulfilled as a species?
1:54:05.080 --> 1:54:08.080
The understanding intelligence or understanding the answers to the big questions?
1:54:08.080 --> 1:54:09.080
Understanding intelligence.
1:54:09.080 --> 1:54:11.080
Oh, totally.
1:54:11.080 --> 1:54:12.080
Totally.
1:54:12.080 --> 1:54:14.080
It would be far more fun place to live.
1:54:14.080 --> 1:54:15.080
You think so?
1:54:15.080 --> 1:54:16.080
Oh, yeah.
1:54:16.080 --> 1:54:17.080
Why not?
1:54:17.080 --> 1:54:22.080
I just put aside this, you know, terminator nonsense and, and, and, and just think about,
1:54:22.080 --> 1:54:26.080
you can think about the, we can talk about the risk of AI if you want.
1:54:26.080 --> 1:54:27.080
I'd love to.
1:54:27.080 --> 1:54:28.080
So let's talk about.
1:54:28.080 --> 1:54:30.080
But I think the world is far better knowing things.
1:54:30.080 --> 1:54:32.080
We're always better than no things.
1:54:32.080 --> 1:54:33.080
Do you think it's better?
1:54:33.080 --> 1:54:38.080
Is it a better place to live in that I know that our planet is one of many in the solar system
1:54:38.080 --> 1:54:40.080
and the solar system is one of many in the galaxy?
1:54:40.080 --> 1:54:44.080
I think it's a more, I dread, I used to, I sometimes think like, God, what would be like
1:54:44.080 --> 1:54:47.080
the 300 years ago, I'd be looking up the sky, I can't understand anything.
1:54:47.080 --> 1:54:48.080
Oh my God.
1:54:48.080 --> 1:54:50.080
I'd be like going to bed every night going, what's going on here?
1:54:50.080 --> 1:54:54.080
Well, I mean, in some sense, I agree with you, but I'm not exactly sure.
1:54:54.080 --> 1:54:55.080
So I'm also a scientist.
1:54:55.080 --> 1:55:01.080
So I have, I share your views, but I'm not, we're, we're like rolling down the hill together.
1:55:01.080 --> 1:55:03.080
What's down the hill?
1:55:03.080 --> 1:55:05.080
I feel for climbing a hill.
1:55:05.080 --> 1:55:07.080
Whatever we're getting, we're getting closer to enlightenment.
1:55:07.080 --> 1:55:08.080
Whatever.
1:55:08.080 --> 1:55:12.080
We're climbing, we're getting pulled up a hill.
1:55:12.080 --> 1:55:14.080
Pulled up by our curiosity.
1:55:14.080 --> 1:55:17.080
We're pulling ourselves up the hill by our curiosity.
1:55:17.080 --> 1:55:19.080
Yeah, sycophers are doing the same thing with the rock.
1:55:19.080 --> 1:55:21.080
Yeah, yeah, yeah.
1:55:21.080 --> 1:55:29.080
But okay, our happiness aside, do you have concerns about, you know, you talk about Sam Harris, Elon Musk,
1:55:29.080 --> 1:55:32.080
of existential threats of intelligence systems?
1:55:32.080 --> 1:55:34.080
No, I'm not worried about existential threats at all.
1:55:34.080 --> 1:55:36.080
There are some things we really do need to worry about.
1:55:36.080 --> 1:55:38.080
Even today's AI, we have things we have to worry about.
1:55:38.080 --> 1:55:43.080
We have to worry about privacy and about how it impacts false beliefs in the world.
1:55:43.080 --> 1:55:48.080
And we have real problems that, and things to worry about with today's AI.
1:55:48.080 --> 1:55:51.080
And that will continue as we create more intelligent systems.
1:55:51.080 --> 1:55:59.080
There's no question, you know, the whole issue about, you know, making intelligent armament and weapons is something that really we have to think about carefully.
1:55:59.080 --> 1:56:01.080
I don't think of those as existential threats.
1:56:01.080 --> 1:56:09.080
I think those are the kind of threats we always face and we'll have to face them here and we'll have to deal with them.
1:56:09.080 --> 1:56:17.080
We can talk about what people think are the existential threats, but when I hear people talking about them, they all sound hollow to me.
1:56:17.080 --> 1:56:21.080
They're based on ideas, they're based on people who really have no idea what intelligence is.
1:56:21.080 --> 1:56:26.080
And if they knew what intelligence was, they wouldn't say those things.
1:56:26.080 --> 1:56:29.080
So those are not experts in the field, you know.
1:56:29.080 --> 1:56:31.080
So there's two, right?
1:56:31.080 --> 1:56:33.080
So one is like superintelligence.
1:56:33.080 --> 1:56:42.080
So a system that becomes far, far superior in reasoning ability than us humans.
1:56:42.080 --> 1:56:45.080
And how is that an existential threat?
1:56:45.080 --> 1:56:49.080
So there's a lot of ways in which it could be.
1:56:49.080 --> 1:57:00.080
One way is us humans are actually irrational, inefficient and get in the way of not happiness,
1:57:00.080 --> 1:57:05.080
but whatever the objective function is of maximizing that objective function and superintelligence.
1:57:05.080 --> 1:57:07.080
The paperclip problem and things like that.
1:57:07.080 --> 1:57:09.080
So the paperclip problem, but with the superintelligence.
1:57:09.080 --> 1:57:10.080
Yeah, yeah, yeah.
1:57:10.080 --> 1:57:15.080
So we already faced this threat in some sense.
1:57:15.080 --> 1:57:17.080
They're called bacteria.
1:57:17.080 --> 1:57:21.080
These are organisms in the world that would like to turn everything into bacteria.
1:57:21.080 --> 1:57:23.080
And they're constantly morphing.
1:57:23.080 --> 1:57:26.080
They're constantly changing to evade our protections.
1:57:26.080 --> 1:57:33.080
And in the past, they have killed huge swaths of populations of humans on this planet.
1:57:33.080 --> 1:57:38.080
So if you want to worry about something that's going to multiply endlessly, we have it.
1:57:38.080 --> 1:57:43.080
And I'm far more worried in that regard, I'm far more worried that some scientists in the laboratory
1:57:43.080 --> 1:57:47.080
will create a super virus or a super bacteria that we cannot control.
1:57:47.080 --> 1:57:49.080
That is a more existential threat.
1:57:49.080 --> 1:57:54.080
Putting an intelligence thing on top of it actually seems to make it less existential to me.
1:57:54.080 --> 1:57:56.080
It's like, it limits its power.
1:57:56.080 --> 1:57:57.080
It limits where it can go.
1:57:57.080 --> 1:57:59.080
It limits the number of things it can do in many ways.
1:57:59.080 --> 1:58:02.080
A bacteria is something you can't even see.
1:58:02.080 --> 1:58:04.080
So that's only one of those problems.
1:58:04.080 --> 1:58:05.080
Yes, exactly.
1:58:05.080 --> 1:58:09.080
So the other one, just in your intuition about intelligence,
1:58:09.080 --> 1:58:12.080
when you think about intelligence of us humans,
1:58:12.080 --> 1:58:14.080
do you think of that as something,
1:58:14.080 --> 1:58:18.080
if you look at intelligence on a spectrum from zero to us humans,
1:58:18.080 --> 1:58:24.080
do you think you can scale that to something far superior to all the mechanisms we've been talking about?
1:58:24.080 --> 1:58:27.080
I want to make another point here, Alex, before I get there.
1:58:27.080 --> 1:58:30.080
Intelligence is the neocortex.
1:58:30.080 --> 1:58:32.080
It is not the entire brain.
1:58:32.080 --> 1:58:36.080
The goal is not to make a human.
1:58:36.080 --> 1:58:38.080
The goal is not to make an emotional system.
1:58:38.080 --> 1:58:41.080
The goal is not to make a system that wants to have sex and reproduce.
1:58:41.080 --> 1:58:42.080
Why would I build that?
1:58:42.080 --> 1:58:44.080
If I want to have a system that wants to reproduce and have sex,
1:58:44.080 --> 1:58:47.080
make bacteria, make computer viruses.
1:58:47.080 --> 1:58:48.080
Those are bad things.
1:58:48.080 --> 1:58:49.080
Don't do that.
1:58:49.080 --> 1:58:50.080
Those are really bad.
1:58:50.080 --> 1:58:51.080
Don't do those things.
1:58:51.080 --> 1:58:53.080
Regulate those.
1:58:53.080 --> 1:58:56.080
But if I just say, I want an intelligent system,
1:58:56.080 --> 1:58:58.080
why doesn't it have to have any human like emotions?
1:58:58.080 --> 1:59:00.080
Why does it even care if it lives?
1:59:00.080 --> 1:59:02.080
Why does it even care if it has food?
1:59:02.080 --> 1:59:04.080
It doesn't care about those things.
1:59:04.080 --> 1:59:07.080
It's just in a trance thinking about mathematics,
1:59:07.080 --> 1:59:12.080
or it's out there just trying to build the space for it on Mars.
1:59:12.080 --> 1:59:15.080
That's a choice we make.
1:59:15.080 --> 1:59:17.080
Don't make human like things.
1:59:17.080 --> 1:59:18.080
Don't make replicating things.
1:59:18.080 --> 1:59:19.080
Don't make things that have emotions.
1:59:19.080 --> 1:59:21.080
Just stick to the neocortex.
1:59:21.080 --> 1:59:24.080
That's a view, actually, that I share, but not everybody shares,
1:59:24.080 --> 1:59:28.080
in the sense that you have faith and optimism about us as engineers
1:59:28.080 --> 1:59:31.080
and systems, humans as builders of systems,
1:59:31.080 --> 1:59:35.080
to not put in different stupid things.
1:59:35.080 --> 1:59:37.080
This is why I mentioned the bacteria one,
1:59:37.080 --> 1:59:40.080
because you might say, well, some person's going to do that.
1:59:40.080 --> 1:59:42.080
Well, some person today could create a bacteria
1:59:42.080 --> 1:59:46.080
that's resistant to all the known antibacterial agents.
1:59:46.080 --> 1:59:49.080
So we already have that threat.
1:59:49.080 --> 1:59:51.080
We already know this is going on.
1:59:51.080 --> 1:59:52.080
It's not a new threat.
1:59:52.080 --> 1:59:56.080
So just accept that, and then we have to deal with it, right?
1:59:56.080 --> 1:59:59.080
Yeah, so my point is nothing to do with intelligence.
1:59:59.080 --> 2:00:02.080
Intelligence is a separate component that you might apply
2:00:02.080 --> 2:00:05.080
to a system that wants to reproduce and do stupid things.
2:00:05.080 --> 2:00:07.080
Let's not do that.
2:00:07.080 --> 2:00:10.080
Yeah, in fact, it is a mystery why people haven't done that yet.
2:00:10.080 --> 2:00:14.080
My dad as a physicist believes that the reason,
2:00:14.080 --> 2:00:19.080
for example, nuclear weapons haven't proliferated amongst evil people.
2:00:19.080 --> 2:00:25.080
So one belief that I share is that there's not that many evil people in the world
2:00:25.080 --> 2:00:32.080
that would use whether it's bacteria or nuclear weapons,
2:00:32.080 --> 2:00:35.080
or maybe the future AI systems to do bad.
2:00:35.080 --> 2:00:37.080
So the fraction is small.
2:00:37.080 --> 2:00:40.080
And the second is that it's actually really hard, technically.
2:00:40.080 --> 2:00:45.080
So the intersection between evil and competent is small.
2:00:45.080 --> 2:00:47.080
And by the way, to really annihilate humanity,
2:00:47.080 --> 2:00:51.080
you'd have to have sort of the nuclear winter phenomenon,
2:00:51.080 --> 2:00:54.080
which is not one person shooting or even 10 bombs.
2:00:54.080 --> 2:00:58.080
You'd have to have some automated system that detonates a million bombs,
2:00:58.080 --> 2:01:00.080
or whatever many thousands we have.
2:01:00.080 --> 2:01:03.080
So it's extreme evil combined with extreme competence.
2:01:03.080 --> 2:01:06.080
And despite building some stupid system that would automatically,
2:01:06.080 --> 2:01:10.080
you know, Dr. Strangelup type of thing, you know,
2:01:10.080 --> 2:01:14.080
I mean, look, we could have some nuclear bomb go off in some major city in the world.
2:01:14.080 --> 2:01:17.080
I think that's actually quite likely, even in my lifetime.
2:01:17.080 --> 2:01:20.080
I don't think that's an unlikely thing, and it would be a tragedy.
2:01:20.080 --> 2:01:23.080
But it won't be an existential threat.
2:01:23.080 --> 2:01:27.080
And it's the same as, you know, the virus of 1917 or whatever it was,
2:01:27.080 --> 2:01:29.080
you know, the influenza.
2:01:29.080 --> 2:01:33.080
These bad things can happen and the plague and so on.
2:01:33.080 --> 2:01:35.080
We can't always prevent it.
2:01:35.080 --> 2:01:37.080
We always try, but we can't.
2:01:37.080 --> 2:01:41.080
But they're not existential threats until we combine all those crazy things together.
2:01:41.080 --> 2:01:45.080
So on the spectrum of intelligence from zero to human,
2:01:45.080 --> 2:01:51.080
do you have a sense of whether it's possible to create several orders of magnitude
2:01:51.080 --> 2:01:54.080
or at least double that of human intelligence,
2:01:54.080 --> 2:01:56.080
to talk about neural cortex?
2:01:56.080 --> 2:01:58.080
I think it's the wrong thing to say, double the intelligence.
2:01:58.080 --> 2:02:01.080
Break it down into different components.
2:02:01.080 --> 2:02:04.080
Can I make something that's a million times faster than a human brain?
2:02:04.080 --> 2:02:06.080
Yes, I can do that.
2:02:06.080 --> 2:02:10.080
Could I make something that is, has a lot more storage than a human brain?
2:02:10.080 --> 2:02:13.080
Yes, I can do that. More copies come.
2:02:13.080 --> 2:02:16.080
Can I make something that attaches to different sensors than a human brain?
2:02:16.080 --> 2:02:17.080
Yes, I can do that.
2:02:17.080 --> 2:02:19.080
Could I make something that's distributed?
2:02:19.080 --> 2:02:23.080
We talked earlier about the departure of neural cortex voting.
2:02:23.080 --> 2:02:25.080
They don't have to be co located.
2:02:25.080 --> 2:02:29.080
They can be all around the places. I could do that too.
2:02:29.080 --> 2:02:32.080
Those are the levers I have, but is it more intelligent?
2:02:32.080 --> 2:02:35.080
What depends what I train in on? What is it doing?
2:02:35.080 --> 2:02:37.080
So here's the thing.
2:02:37.080 --> 2:02:46.080
Let's say larger neural cortex and or whatever size that allows for higher and higher hierarchies
2:02:46.080 --> 2:02:49.080
to form, we're talking about reference frames and concepts.
2:02:49.080 --> 2:02:53.080
So I could, could I have something that's a super physicist or a super mathematician? Yes.
2:02:53.080 --> 2:02:59.080
And the question is, once you have a super physicist, will they be able to understand something?
2:02:59.080 --> 2:03:03.080
Do you have a sense that it will be orders, like us compared to ants?
2:03:03.080 --> 2:03:04.080
Could we ever understand it?
2:03:04.080 --> 2:03:05.080
Yeah.
2:03:05.080 --> 2:03:11.080
Most people cannot understand general relativity.
2:03:11.080 --> 2:03:13.080
It's a really hard thing to get.
2:03:13.080 --> 2:03:17.080
I mean, you can paint it in a fuzzy picture, stretchy space, you know?
2:03:17.080 --> 2:03:18.080
Yeah.
2:03:18.080 --> 2:03:23.080
But the field equations to do that and the deep intuitions are really, really hard.
2:03:23.080 --> 2:03:26.080
And I've tried, I'm unable to do it.
2:03:26.080 --> 2:03:32.080
It's easy to get special relative, but general relative, man, that's too much.
2:03:32.080 --> 2:03:35.080
And so we already live with this to some extent.
2:03:35.080 --> 2:03:40.080
The vast majority of people can't understand actually what the vast majority of other people actually know.
2:03:40.080 --> 2:03:45.080
We're just either we don't have the effort to or we can't or we don't have time or just not smart enough, whatever.
2:03:45.080 --> 2:03:48.080
So, but we have ways of communicating.
2:03:48.080 --> 2:03:51.080
Einstein has spoken in a way that I can understand.
2:03:51.080 --> 2:03:54.080
He's given me analogies that are useful.
2:03:54.080 --> 2:04:00.080
I can use those analogies for my own work and think about, you know, concepts that are similar.
2:04:00.080 --> 2:04:02.080
It's not stupid.
2:04:02.080 --> 2:04:04.080
It's not like he's existed in some other plane.
2:04:04.080 --> 2:04:06.080
There's no connection to my plane in the world here.
2:04:06.080 --> 2:04:07.080
So that will occur.
2:04:07.080 --> 2:04:09.080
It already has occurred.
2:04:09.080 --> 2:04:12.080
That's when my point at this story is it already has occurred.
2:04:12.080 --> 2:04:14.080
We live it every day.
2:04:14.080 --> 2:04:21.080
One could argue that with we create machine intelligence that think a million times faster than us that it'll be so far we can't make the connections.
2:04:21.080 --> 2:04:29.080
But, you know, at the moment, everything that seems really, really hard to figure out in the world when you actually figure it out is not that hard.
2:04:29.080 --> 2:04:32.080
You know, almost everyone can understand the multiverses.
2:04:32.080 --> 2:04:34.080
Almost everyone can understand quantum physics.
2:04:34.080 --> 2:04:38.080
Almost everyone can understand these basic things, even though hardly any people could figure those things out.
2:04:38.080 --> 2:04:40.080
Yeah, but really understand.
2:04:40.080 --> 2:04:43.080
But you don't need to really, only a few people really understand.
2:04:43.080 --> 2:04:49.080
You need to only understand the projections, the sprinkles of the useful insights from that.
2:04:49.080 --> 2:04:51.080
That was my example of Einstein, right?
2:04:51.080 --> 2:04:55.080
His general theory of relativity is one thing that very, very, very few people can get.
2:04:55.080 --> 2:04:59.080
And what if we just said those other few people are also artificial intelligences?
2:04:59.080 --> 2:05:01.080
How bad is that?
2:05:01.080 --> 2:05:02.080
In some sense they are, right?
2:05:02.080 --> 2:05:03.080
Yeah, they say already.
2:05:03.080 --> 2:05:05.080
I mean, Einstein wasn't a really normal person.
2:05:05.080 --> 2:05:07.080
He had a lot of weird quirks.
2:05:07.080 --> 2:05:09.080
And so the other people who work with him.
2:05:09.080 --> 2:05:15.080
So, you know, maybe they already were sort of this actual plane of intelligence that we live with it already.
2:05:15.080 --> 2:05:17.080
It's not a problem.
2:05:17.080 --> 2:05:20.080
It's still useful and, you know.
2:05:20.080 --> 2:05:24.080
So do you think we are the only intelligent life out there in the universe?
2:05:24.080 --> 2:05:29.080
I would say that intelligent life has and will exist elsewhere in the universe.
2:05:29.080 --> 2:05:31.080
I'll say that.
2:05:31.080 --> 2:05:39.080
There is a question about contemporaneous intelligence life, which is hard to even answer when we think about relativity and the nature of space time.
2:05:39.080 --> 2:05:43.080
We can't say what exactly is this time someplace else in the world.
2:05:43.080 --> 2:05:54.080
But I think it's, you know, I do worry a lot about the filter idea, which is that perhaps intelligent species don't last very long.
2:05:54.080 --> 2:05:56.080
And so we haven't been around very long.
2:05:56.080 --> 2:06:02.080
As a technological species, we've been around for almost nothing, you know, what, 200 years or something like that.
2:06:02.080 --> 2:06:08.080
And we don't have any data, a good data point on whether it's likely that we'll survive or not.
2:06:08.080 --> 2:06:12.080
So do I think that there have been intelligent life elsewhere in the universe?
2:06:12.080 --> 2:06:14.080
Almost certainly, of course.
2:06:14.080 --> 2:06:16.080
In the past, in the future, yes.
2:06:16.080 --> 2:06:18.080
Does it survive for a long time?
2:06:18.080 --> 2:06:19.080
I don't know.
2:06:19.080 --> 2:06:25.080
This is another reason I'm excited about our work, is our work meaning the general world of AI.
2:06:25.080 --> 2:06:31.080
I think we can build intelligent machines that outlast us.
2:06:31.080 --> 2:06:34.080
You know, they don't have to be tied to Earth.
2:06:34.080 --> 2:06:39.080
They don't have to, you know, I'm not saying they're recreating, you know, you know, aliens.
2:06:39.080 --> 2:06:44.080
I'm just saying, if I asked myself, and this might be a good point to end on here.
2:06:44.080 --> 2:06:47.080
If I asked myself, you know, what's special about our species?
2:06:47.080 --> 2:06:49.080
We're not particularly interesting physically.
2:06:49.080 --> 2:06:51.080
We're not, we don't fly.
2:06:51.080 --> 2:06:52.080
We're not good swimmers.
2:06:52.080 --> 2:06:53.080
We're not very fast.
2:06:53.080 --> 2:06:54.080
We're not very strong, you know.
2:06:54.080 --> 2:06:55.080
It's our brain.
2:06:55.080 --> 2:06:56.080
That's the only thing.
2:06:56.080 --> 2:07:01.080
And we are the only species on this planet that's built the model of the world that extends beyond what we can actually sense.
2:07:01.080 --> 2:07:09.080
We're the only people who know about the far side of the moon and other universes and other galaxies and other stars and what happens in the atom.
2:07:09.080 --> 2:07:12.080
That knowledge doesn't exist anywhere else.
2:07:12.080 --> 2:07:13.080
It's only in our heads.
2:07:13.080 --> 2:07:14.080
Cats don't do it.
2:07:14.080 --> 2:07:15.080
Dogs don't do it.
2:07:15.080 --> 2:07:16.080
Monkeys don't do it.
2:07:16.080 --> 2:07:18.080
That is what we've created that's unique.
2:07:18.080 --> 2:07:19.080
Not our genes.
2:07:19.080 --> 2:07:20.080
It's knowledge.
2:07:20.080 --> 2:07:24.080
And if I ask me, what is the legacy of humanity?
2:07:24.080 --> 2:07:25.080
What should our legacy be?
2:07:25.080 --> 2:07:26.080
It should be knowledge.
2:07:26.080 --> 2:07:30.080
We should preserve our knowledge in a way that it can exist beyond us.
2:07:30.080 --> 2:07:38.080
And I think the best way of doing that, in fact, you have to do it, is to have to go along with intelligent machines to understand that knowledge.
2:07:38.080 --> 2:07:44.080
It's a very broad idea, but we should be thinking, I call it a state planning for humanity.
2:07:44.080 --> 2:07:49.080
We should be thinking about what we want to leave behind when as a species we're no longer here.
2:07:49.080 --> 2:07:51.080
And that will happen sometime.
2:07:51.080 --> 2:07:52.080
Sooner or later, it's going to happen.
2:07:52.080 --> 2:07:58.080
And understanding intelligence and creating intelligence gives us a better chance to prolong.
2:07:58.080 --> 2:08:00.080
It does give us a better chance to prolong life.
2:08:00.080 --> 2:08:01.080
Yes.
2:08:01.080 --> 2:08:03.080
It gives us a chance to live on other planets.
2:08:03.080 --> 2:08:07.080
But even beyond that, I mean, our solar system will disappear one day.
2:08:07.080 --> 2:08:09.080
It's given enough time.
2:08:09.080 --> 2:08:10.080
So I don't know.
2:08:10.080 --> 2:08:14.080
I doubt we will ever be able to travel to other things.
2:08:14.080 --> 2:08:18.080
But we could tell the stars, but we could send intelligent machines to do that.
2:08:18.080 --> 2:08:29.080
Do you have an optimistic, a hopeful view of our knowledge of the echoes of human civilization living through the intelligent systems we create?
2:08:29.080 --> 2:08:30.080
Oh, totally.
2:08:30.080 --> 2:08:32.080
Well, I think the intelligent systems are greater.
2:08:32.080 --> 2:08:39.080
In some sense, the vessel for bringing them beyond Earth or making them last beyond humans themselves.
2:08:39.080 --> 2:08:41.080
So how do you feel about that?
2:08:41.080 --> 2:08:44.080
That they won't be human, quote unquote.
2:08:44.080 --> 2:08:48.080
Human, what is human? Our species are changing all the time.
2:08:48.080 --> 2:08:52.080
Human today is not the same as human just 50 years ago.
2:08:52.080 --> 2:08:54.080
What is human? Do we care about our genetics?
2:08:54.080 --> 2:08:56.080
Why is that important?
2:08:56.080 --> 2:08:59.080
As I point out, our genetics are no more interesting than a bacterium's genetics.
2:08:59.080 --> 2:09:01.080
It's no more interesting than a monkey's genetics.
2:09:01.080 --> 2:09:07.080
What we have, what's unique and what's valuable is our knowledge, what we've learned about the world.
2:09:07.080 --> 2:09:09.080
And that is the rare thing.
2:09:09.080 --> 2:09:11.080
That's the thing we want to preserve.
2:09:11.080 --> 2:09:15.080
Who cares about our genes?
2:09:15.080 --> 2:09:17.080
It's the knowledge.
2:09:17.080 --> 2:09:19.080
That's a really good place to end.
2:09:19.080 --> 2:09:42.080
Thank you so much for talking to me.