François Chollet: Measures of Intelligence
AI 与机器学习心理与人性生物与进化音乐与艺术技术与编程
📋 章节目录
暂无章节信息
🔑 关键词
intelligencehumantestdatalearninginstancedonhumansablelanguagegoingpossiblearctaskpriorsdeeppapermachineinterestingmeasure
💬 精彩语录
暂无语录
🎙️ 完整对话（3098 条）
Lex Fridman (00:00.000)
The following is a conversation with Francois Chollet,
以下是与弗朗索瓦·乔莱的对话，
Lex Fridman (00:03.220)
his second time on the podcast.
他第二次参加播客。
Lex Fridman (00:05.260)
He's both a world class engineer and a philosopher
他既是世界级的工程师又是哲学家
Lex Fridman (00:09.580)
in the realm of deep learning and artificial intelligence.
在深度学习和人工智能领域。
Lex Fridman (00:13.180)
This time, we talk a lot about his paper titled
这次，我们详细讨论了他的论文，题为
François Chollet (00:16.200)
on the measure of intelligence that discusses
关于讨论的智力测量
Lex Fridman (00:19.040)
how we might define and measure general intelligence
我们如何定义和衡量一般智力
François Chollet (00:22.440)
in our computing machinery.
在我们的计算机器中。
Lex Fridman (00:24.640)
Quick summary of the sponsors,
赞助商的快速摘要，
François Chollet (00:26.420)
Babbel, Masterclass, and Cash App.
Babbel、Masterclass 和 Cash App。
Lex Fridman (00:29.460)
Click the sponsor links in the description
单击说明中的赞助商链接
François Chollet (00:31.240)
to get a discount and to support this podcast.
获得折扣并支持此播客。
Lex Fridman (00:34.500)
As a side note, let me say that the serious,
作为旁注，让我严肃地说，
François Chollet (00:36.880)
rigorous scientific study
严谨的科学研究
Lex Fridman (00:38.720)
of artificial general intelligence is a rare thing.
人工智能的发展是一件罕见的事情。
François Chollet (00:42.220)
The mainstream machine learning community works
主流机器学习社区的工作
Lex Fridman (00:44.080)
on very narrow AI with very narrow benchmarks.
在非常狭窄的人工智能和非常狭窄的基准上。
François Chollet (00:47.740)
This is very good for incremental
这对于增量非常有好处
Lex Fridman (00:49.920)
and sometimes big incremental progress.
有时还有巨大的渐进式进展。
François Chollet (00:53.200)
On the other hand, the outside the mainstream,
另一方面，在主流之外，
Lex Fridman (00:56.060)
renegade, you could say, AGI community works
François Chollet (01:00.020)
on approaches that verge on the philosophical
Lex Fridman (01:03.000)
and even the literary without big public benchmarks.
François Chollet (01:07.300)
Walking the line between the two worlds is a rare breed,
Lex Fridman (01:10.640)
but it doesn't have to be.
François Chollet (01:12.360)
I ran the AGI series at MIT as an attempt
Lex Fridman (01:15.320)
to inspire more people to walk this line.
François Chollet (01:17.700)
Deep mind and open AI for a time
Lex Fridman (01:20.020)
and still on occasion walk this line.
François Chollet (01:23.180)
Francois Chollet does as well.
Lex Fridman (01:25.860)
I hope to also.
François Chollet (01:27.620)
It's a beautiful dream to work towards
Lex Fridman (01:29.880)
and to make real one day.
François Chollet (01:32.480)
If you enjoy this thing, subscribe on YouTube,
Lex Fridman (01:34.580)
review it with five stars on Apple Podcast,
François Chollet (01:36.760)
follow on Spotify, support on Patreon,
Lex Fridman (01:39.020)
or connect with me on Twitter at Lex Friedman.
François Chollet (01:42.020)
As usual, I'll do a few minutes of ads now
Lex Fridman (01:44.240)
and no ads in the middle.
François Chollet (01:45.780)
I try to make these interesting,
Lex Fridman (01:47.440)
but I give you timestamps so you can skip.
Lex Fridman (01:50.620)
But still, please do check out the sponsors
Lex Fridman (01:52.660)
by clicking the links in the description.
François Chollet (01:54.580)
It's the best way to support this podcast.
Lex Fridman (01:57.900)
This show is sponsored by Babbel,
François Chollet (02:00.100)
an app and website that gets you speaking
Lex Fridman (02:02.460)
in a new language within weeks.
François Chollet (02:04.360)
Go to babbel.com and use code Lex to get three months free.
Lex Fridman (02:08.200)
They offer 14 languages, including Spanish, French,
François Chollet (02:11.460)
Italian, German, and yes, Russian.
Lex Fridman (02:15.220)
Daily lessons are 10 to 15 minutes,
François Chollet (02:17.340)
super easy, effective,
Lex Fridman (02:19.060)
designed by over 100 language experts.
François Chollet (02:22.240)
Let me read a few lines from the Russian poem
Lex Fridman (02:24.700)
Noch, ulitsa, fanar, apteka, by Alexander Bloch,
François Chollet (02:29.020)
that you'll start to understand if you sign up to Babbel.
Lex Fridman (02:32.580)
Noch, ulitsa, fanar, apteka,
François Chollet (02:35.220)
Bessmysliny, ituskly, svet,
Lex Fridman (02:38.100)
Zhevi esho, khod chetvert veka,
François Chollet (02:41.140)
Vse budet tak, ishoda, net.
Lex Fridman (02:44.700)
Now, I say that you'll start to understand this poem
François Chollet (02:48.500)
because Russian starts with a language
Lex Fridman (02:51.420)
and ends with vodka.
François Chollet (02:54.020)
Now, the latter part is definitely not endorsed
Lex Fridman (02:56.600)
or provided by Babbel.
François Chollet (02:58.020)
It will probably lose me this sponsorship,
Lex Fridman (03:00.340)
although it hasn't yet.
Lex Fridman (03:02.460)
But once you graduate with Babbel,
Lex Fridman (03:04.460)
you can enroll in my advanced course
François Chollet (03:06.120)
of late night Russian conversation over vodka.
Lex Fridman (03:09.200)
No app for that yet.
Lex Fridman (03:11.260)
So get started by visiting babbel.com
Lex Fridman (03:13.740)
and use code Lex to get three months free.
François Chollet (03:18.180)
This show is also sponsored by Masterclass.
Lex Fridman (03:20.980)
Sign up at masterclass.com slash Lex
François Chollet (03:23.380)
to get a discount and to support this podcast.
Lex Fridman (03:26.580)
When I first heard about Masterclass,
François Chollet (03:28.060)
I thought it was too good to be true.
Lex Fridman (03:29.980)
I still think it's too good to be true.
François Chollet (03:32.340)
For $180 a year, you get an all access pass
Lex Fridman (03:35.420)
to watch courses from, to list some of my favorites.
François Chollet (03:38.740)
Chris Hatfield on space exploration,
Lex Fridman (03:41.340)
hope to have him in this podcast one day.
François Chollet (03:43.500)
Neil Dugras Tyson on scientific thinking and communication,
Lex Fridman (03:46.660)
Neil too.
François Chollet (03:47.900)
Will Wright, creator of SimCity and Sims
Lex Fridman (03:50.140)
on game design, Carlos Santana on guitar,
François Chollet (03:52.780)
Kary Kasparov on chess, Daniel Nagrano on poker,
Lex Fridman (03:55.980)
and many more.
François Chollet (03:57.240)
Chris Hatfield explaining how rockets work
Lex Fridman (03:59.700)
and the experience of being watched at the space
François Chollet (04:01.740)
alone is worth the money.
Lex Fridman (04:03.300)
By the way, you can watch it on basically any device.
François Chollet (04:06.540)
Once again, sign up at masterclass.com slash Lex
Lex Fridman (04:09.380)
to get a discount and to support this podcast.
François Chollet (04:13.340)
This show finally is presented by Cash App,
Lex Fridman (04:16.460)
the number one finance app in the App Store.
François Chollet (04:18.720)
When you get it, use code LexPodcast.
Lex Fridman (04:21.220)
Cash App lets you send money to friends,
François Chollet (04:23.300)
buy Bitcoin, and invest in the stock market
Lex Fridman (04:25.460)
with as little as $1.
François Chollet (04:27.260)
Since Cash App allows you to send
Lex Fridman (04:28.980)
and receive money digitally,
François Chollet (04:30.540)
let me mention a surprising fact related to physical money.
Lex Fridman (04:33.860)
Of all the currency in the world,
François Chollet (04:35.700)
roughly 8% of it is actually physical money.
Lex Fridman (04:39.300)
The other 92% of the money only exists digitally,
Lex Fridman (04:42.820)
and that's only going to increase.
Lex Fridman (04:45.280)
So again, if you get Cash App from the App Store
François Chollet (04:47.400)
through Google Play and use code LexPodcast,
Lex Fridman (04:50.660)
you get 10 bucks,
Lex Fridman (04:51.740)
and Cash App will also donate $10 to FIRST,
Lex Fridman (04:54.420)
an organization that is helping to advance robotics
Lex Fridman (04:57.000)
and STEM education for young people around the world.
Lex Fridman (05:00.500)
And now here's my conversation with Francois Chalet.
Lex Fridman (05:05.060)
What philosophers, thinkers, or ideas
Lex Fridman (05:07.360)
had a big impact on you growing up and today?
Lex Fridman (05:10.700)
So one author that had a big impact on me
Lex Fridman (05:14.860)
when I read his books as a teenager was Jean Piaget,
François Chollet (05:18.820)
who is a Swiss psychologist,
Lex Fridman (05:21.380)
is considered to be the father of developmental psychology.
Lex Fridman (05:25.540)
And he has a large body of work about
Lex Fridman (05:28.700)
basically how intelligence develops in children.
Lex Fridman (05:33.380)
And so it's very old work,
Lex Fridman (05:35.500)
like most of it is from the 1930s, 1940s.
Lex Fridman (05:39.140)
So it's not quite up to date.
Lex Fridman (05:40.900)
It's actually superseded by many newer developments
François Chollet (05:43.820)
in developmental psychology.
Lex Fridman (05:45.660)
But to me, it was very interesting, very striking,
Lex Fridman (05:49.600)
and actually shaped the early ways
Lex Fridman (05:51.340)
in which I started thinking about the mind
Lex Fridman (05:53.820)
and the development of intelligence as a teenager.
Lex Fridman (05:56.220)
His actual ideas or the way he thought about it
François Chollet (05:58.460)
or just the fact that you could think
Lex Fridman (05:59.840)
about the developing mind at all?
François Chollet (06:01.600)
I guess both.
Lex Fridman (06:02.500)
Jean Piaget is the author that really introduced me
François Chollet (06:04.940)
to the notion that intelligence and the mind
Lex Fridman (06:07.980)
is something that you construct throughout your life
Lex Fridman (06:11.120)
and that children construct it in stages.
Lex Fridman (06:15.780)
And I thought that was a very interesting idea,
François Chollet (06:17.460)
which is, of course, very relevant to AI,
Lex Fridman (06:20.460)
to building artificial minds.
François Chollet (06:23.180)
Another book that I read around the same time
Lex Fridman (06:25.860)
that had a big impact on me,
Lex Fridman (06:28.900)
and there was actually a little bit of overlap
Lex Fridman (06:32.100)
with Jean Piaget as well,
Lex Fridman (06:32.980)
and I read it around the same time,
Lex Fridman (06:35.340)
is Geoff Hawking's On Intelligence, which is a classic.
Lex Fridman (06:39.860)
And he has this vision of the mind
Lex Fridman (06:42.500)
as a multi scale hierarchy of temporal prediction modules.
Lex Fridman (06:47.820)
And these ideas really resonated with me,
Lex Fridman (06:50.020)
like the notion of a modular hierarchy
François Chollet (06:55.440)
of potentially compression functions
Lex Fridman (07:00.100)
or prediction functions.
François Chollet (07:01.700)
I thought it was really, really interesting,
Lex Fridman (07:03.980)
and it shaped the way I started thinking
François Chollet (07:07.100)
about how to build minds.
Lex Fridman (07:09.760)
The hierarchical nature, which aspect?
François Chollet (07:13.740)
Also, he's a neuroscientist, so he was thinking actual,
Lex Fridman (07:17.520)
he was basically talking about how our mind works.
François Chollet (07:20.580)
Yeah, the notion that cognition is prediction
Lex Fridman (07:23.260)
was an idea that was kind of new to me at the time
Lex Fridman (07:25.460)
and that I really loved at the time.
Lex Fridman (07:27.840)
And yeah, and the notion that there are multiple scales
François Chollet (07:31.900)
of processing in the brain.
Lex Fridman (07:35.320)
The hierarchy.
François Chollet (07:36.260)
Yes.
Lex Fridman (07:37.100)
This was before deep learning.
François Chollet (07:38.600)
These ideas of hierarchies in AI
Lex Fridman (07:41.140)
have been around for a long time,
François Chollet (07:43.180)
even before on intelligence.
Lex Fridman (07:45.020)
They've been around since the 1980s.
Lex Fridman (07:48.980)
And yeah, that was before deep learning.
Lex Fridman (07:50.500)
But of course, I think these ideas really found
François Chollet (07:53.500)
their practical implementation in deep learning.
Lex Fridman (07:58.100)
What about the memory side of things?
François Chollet (07:59.740)
I think he was talking about knowledge representation.
Lex Fridman (08:02.860)
Do you think about memory a lot?
François Chollet (08:04.420)
One way you can think of neural networks
Lex Fridman (08:06.340)
as a kind of memory, you're memorizing things,
Lex Fridman (08:10.780)
but it doesn't seem to be the kind of memory
Lex Fridman (08:14.260)
that's in our brains,
François Chollet (08:16.880)
or it doesn't have the same rich complexity,
Lex Fridman (08:18.660)
long term nature that's in our brains.
François Chollet (08:20.660)
Yes, the brain is more of a sparse access memory
Lex Fridman (08:23.980)
so that you can actually retrieve very precisely
François Chollet (08:27.740)
like bits of your experience.
Lex Fridman (08:30.100)
The retrieval aspect, you can like introspect,
François Chollet (08:33.500)
you can ask yourself questions.
Lex Fridman (08:35.300)
I guess you can program your own memory
Lex Fridman (08:38.260)
and language is actually the tool you use to do that.
Lex Fridman (08:41.700)
I think language is a kind of operating system for the mind
Lex Fridman (08:46.360)
and use language.
Lex Fridman (08:47.820)
Well, one of the uses of language is as a query
François Chollet (08:51.800)
that you run over your own memory,
Lex Fridman (08:53.860)
use words as keys to retrieve specific experiences
François Chollet (08:57.940)
or specific concepts, specific thoughts.
Lex Fridman (09:00.140)
Like language is a way you store thoughts,
François Chollet (09:02.380)
not just in writing, in the physical world,
Lex Fridman (09:04.740)
but also in your own mind.
Lex Fridman (09:06.100)
And it's also how you retrieve them.
Lex Fridman (09:07.580)
Like, imagine if you didn't have language,
François Chollet (09:10.000)
then you would have to,
Lex Fridman (09:11.740)
you would not really have a self,
François Chollet (09:14.340)
internally triggered way of retrieving past thoughts.
Lex Fridman (09:18.620)
You would have to rely on external experiences.
François Chollet (09:21.300)
For instance, you see a specific site,
Lex Fridman (09:24.020)
you smell a specific smell and that brings up memories,
Lex Fridman (09:26.780)
but you would not really have a way
Lex Fridman (09:28.700)
to deliberately access these memories without language.
François Chollet (09:32.740)
Well, the interesting thing you mentioned
Lex Fridman (09:33.980)
is you can also program the memory.
François Chollet (09:37.420)
You can change it probably with language.
Lex Fridman (09:39.980)
Yeah, using language, yes.
François Chollet (09:41.500)
Well, let me ask you a Chomsky question,
Lex Fridman (09:44.100)
which is like, first of all,
Lex Fridman (09:45.980)
do you think language is like fundamental,
Lex Fridman (09:49.100)
like there's turtles, what's at the bottom of the turtles?
François Chollet (09:54.460)
They don't go, it can't be turtles all the way down.
Lex Fridman (09:57.260)
Is language at the bottom of cognition of everything?
François Chollet (10:00.260)
Is like language, the fundamental aspect
Lex Fridman (10:05.300)
of like what it means to be a thinking thing?
François Chollet (10:10.700)
No, I don't think so.
Lex Fridman (10:12.100)
I think language is.
Lex Fridman (10:12.940)
You disagree with Norm Chomsky?
Lex Fridman (10:14.620)
Yes, I think language is a layer on top of cognition.
Lex Fridman (10:17.900)
So it is fundamental to cognition in the sense that
Lex Fridman (10:21.740)
to use a computing metaphor,
François Chollet (10:23.380)
I see language as the operating system of the brain,
Lex Fridman (10:28.060)
of the human mind.
Lex Fridman (10:29.500)
And the operating system is a layer on top of the computer.
Lex Fridman (10:33.180)
The computer exists before the operating system,
Lex Fridman (10:36.140)
but the operating system is how you make it truly useful.
Lex Fridman (10:39.500)
And the operating system is most likely Windows, not Linux,
François Chollet (10:43.940)
because language is messy.
Lex Fridman (10:45.860)
Yeah, it's messy and it's pretty difficult
François Chollet (10:49.460)
to inspect it, introspect it.
Lex Fridman (10:53.140)
How do you think about language?
François Chollet (10:55.100)
Like we use actually sort of human interpretable language,
Lex Fridman (11:00.060)
but is there something like a deeper,
Lex Fridman (11:03.100)
that's closer to like logical type of statements?
Lex Fridman (11:08.860)
Like, yeah, what is the nature of language, do you think?
François Chollet (11:16.140)
Like is there something deeper than like the syntactic rules
Lex Fridman (11:18.540)
we construct?
François Chollet (11:19.380)
Is there something that doesn't require utterances
Lex Fridman (11:22.860)
or writing or so on?
François Chollet (11:25.580)
Are you asking about the possibility
Lex Fridman (11:27.460)
that there could exist languages for thinking
Lex Fridman (11:30.900)
that are not made of words?
Lex Fridman (11:32.820)
Yeah.
François Chollet (11:33.660)
Yeah, I think so.
Lex Fridman (11:34.500)
I think, so the mind is layers, right?
Lex Fridman (11:38.580)
And language is almost like the outermost,
Lex Fridman (11:41.780)
the uppermost layer.
Lex Fridman (11:44.620)
But before we think in words,
Lex Fridman (11:46.780)
I think we think in terms of emotion in space
Lex Fridman (11:51.100)
and we think in terms of physical actions.
Lex Fridman (11:54.180)
And I think babies in particular,
François Chollet (11:56.860)
probably expresses thoughts in terms of the actions
Lex Fridman (12:01.380)
that they've seen or that they can perform
Lex Fridman (12:03.700)
and in terms of motions of objects in their environment
Lex Fridman (12:08.020)
before they start thinking in terms of words.
François Chollet (12:10.860)
It's amazing to think about that
Lex Fridman (12:13.900)
as the building blocks of language.
Lex Fridman (12:16.780)
So like the kind of actions and ways the babies see the world
Lex Fridman (12:21.820)
as like more fundamental
François Chollet (12:23.260)
than the beautiful Shakespearean language
Lex Fridman (12:26.220)
you construct on top of it.
Lex Fridman (12:28.620)
And we probably don't have any idea
Lex Fridman (12:30.500)
what that looks like, right?
François Chollet (12:31.700)
Like what, because it's important
Lex Fridman (12:34.020)
for them trying to engineer it into AI systems.
François Chollet (12:38.460)
I think visual analogies and motion
Lex Fridman (12:42.060)
is a fundamental building block of the mind.
Lex Fridman (12:45.380)
And you actually see it reflected in language.
Lex Fridman (12:48.540)
Like language is full of special metaphors.
Lex Fridman (12:51.820)
And when you think about things,
Lex Fridman (12:53.820)
I consider myself very much as a visual thinker.
François Chollet (12:57.380)
You often express these thoughts
Lex Fridman (13:01.140)
by using things like visualizing concepts
François Chollet (13:06.500)
in 2D space or like you solve problems
Lex Fridman (13:09.940)
by imagining yourself navigating a concept space.
Lex Fridman (13:14.940)
So I don't know if you have this sort of experience.
Lex Fridman (13:17.940)
You said visualizing concept space.
Lex Fridman (13:19.860)
So like, so I certainly think about,
Lex Fridman (13:24.820)
I certainly visualize mathematical concepts,
Lex Fridman (13:27.980)
but you mean like in concept space,
Lex Fridman (13:32.340)
visually you're embedding ideas
François Chollet (13:34.860)
into a three dimensional space
Lex Fridman (13:36.940)
you can explore with your mind essentially?
François Chollet (13:38.820)
You should be more like 2D, but yeah.
Lex Fridman (13:40.340)
2D?
François Chollet (13:41.180)
Yeah.
Lex Fridman (13:42.100)
You're a flatlander.
François Chollet (13:43.180)
You're, okay.
Lex Fridman (13:45.700)
No, I do not.
François Chollet (13:49.660)
I always have to, before I jump from concept to concept,
Lex Fridman (13:52.780)
I have to put it back down on paper.
François Chollet (13:57.100)
It has to be on paper.
Lex Fridman (13:58.060)
I can only travel on 2D paper, not inside my mind.
François Chollet (14:03.340)
You're able to move inside your mind.
Lex Fridman (14:05.340)
But even if you're writing like a paper, for instance,
Lex Fridman (14:07.900)
don't you have like a spatial representation of your paper?
Lex Fridman (14:11.020)
Like you visualize where ideas lie topologically
François Chollet (14:16.660)
in relationship to other ideas,
Lex Fridman (14:18.980)
kind of like a subway map of the ideas in your paper.
François Chollet (14:22.500)
Yeah, that's true.
Lex Fridman (14:23.380)
I mean, there is, in papers, I don't know about you,
Lex Fridman (14:27.900)
but it feels like there's a destination.
Lex Fridman (14:32.540)
There's a key idea that you want to arrive at.
Lex Fridman (14:36.220)
And a lot of it is in the fog
Lex Fridman (14:39.340)
and you're trying to kind of,
François Chollet (14:40.820)
it's almost like, what's that called
Lex Fridman (14:46.180)
when you do a path planning search from both directions,
François Chollet (14:49.900)
from the start and from the end.
Lex Fridman (14:52.700)
And then you find, you do like shortest path,
Lex Fridman (14:54.740)
but like, you know, in game playing,
Lex Fridman (14:57.380)
you do this with like A star from both sides.
Lex Fridman (15:01.020)
And you see where we're on the join.
Lex Fridman (15:03.420)
Yeah, so you kind of do, at least for me,
François Chollet (15:05.740)
I think like, first of all,
Lex Fridman (15:07.100)
just exploring from the start from like first principles,
Lex Fridman (15:10.800)
what do I know, what can I start proving from that, right?
Lex Fridman (15:15.620)
And then from the destination,
François Chollet (15:18.060)
if you start backtracking,
Lex Fridman (15:20.460)
like if I want to show some kind of sets of ideas,
Lex Fridman (15:25.400)
what would it take to show them and you kind of backtrack,
Lex Fridman (15:28.300)
but like, yeah,
François Chollet (15:29.140)
I don't think I'm doing all that in my mind though.
Lex Fridman (15:31.260)
Like I'm putting it down on paper.
Lex Fridman (15:33.180)
Do you use mind maps to organize your ideas?
Lex Fridman (15:35.500)
Yeah, I like mind maps.
François Chollet (15:37.740)
Let's get into this,
Lex Fridman (15:38.580)
because I've been so jealous of people.
François Chollet (15:41.180)
I haven't really tried it.
Lex Fridman (15:42.120)
I've been jealous of people that seem to like,
François Chollet (15:45.500)
they get like this fire of passion in their eyes
Lex Fridman (15:48.140)
because everything starts making sense.
François Chollet (15:50.020)
It's like Tom Cruise in the movie
Lex Fridman (15:51.940)
was like moving stuff around.
François Chollet (15:53.820)
Some of the most brilliant people I know use mind maps.
Lex Fridman (15:55.900)
I haven't tried really.
Lex Fridman (15:57.660)
Can you explain what the hell a mind map is?
Lex Fridman (16:01.240)
I guess mind map is a way to make
François Chollet (16:03.700)
kind of like the mess inside your mind
Lex Fridman (16:05.940)
to just put it on paper so that you gain more control over it.
François Chollet (16:10.020)
It's a way to organize things on paper
Lex Fridman (16:13.020)
and as kind of like a consequence
François Chollet (16:16.420)
of organizing things on paper,
Lex Fridman (16:17.940)
they start being more organized inside your own mind.
Lex Fridman (16:20.300)
So what does that look like?
Lex Fridman (16:21.540)
You put, like, do you have an example?
Lex Fridman (16:23.980)
Like what's the first thing you write on paper?
Lex Fridman (16:27.360)
What's the second thing you write?
François Chollet (16:28.980)
I mean, typically you draw a mind map
Lex Fridman (16:31.660)
to organize the way you think about a topic.
Lex Fridman (16:34.860)
So you would start by writing down
Lex Fridman (16:37.340)
like the key concept about that topic.
François Chollet (16:39.580)
Like you would write intelligence or something,
Lex Fridman (16:42.220)
and then you would start adding associative connections.
François Chollet (16:45.660)
Like what do you think about
Lex Fridman (16:46.860)
when you think about intelligence?
Lex Fridman (16:48.100)
What do you think are the key elements of intelligence?
Lex Fridman (16:50.460)
So maybe you would have language, for instance,
Lex Fridman (16:52.340)
and you'd have motion.
Lex Fridman (16:53.420)
And so you would start drawing notes with these things.
Lex Fridman (16:55.460)
And then you would see what do you think about
Lex Fridman (16:57.220)
when you think about motion and so on.
Lex Fridman (16:59.140)
And you would go like that, like a tree.
Lex Fridman (17:00.620)
Is it a tree mostly or is it a graph too, like a tree?
François Chollet (17:05.660)
Oh, it's more of a graph than a tree.
Lex Fridman (17:07.980)
And it's not limited to just writing down words.
François Chollet (17:13.260)
You can also draw things.
Lex Fridman (17:15.940)
And it's not supposed to be purely hierarchical, right?
François Chollet (17:21.660)
The point is that once you start writing it down,
Lex Fridman (17:24.540)
you can start reorganizing it so that it makes more sense,
Lex Fridman (17:27.500)
so that it's connected in a more effective way.
Lex Fridman (17:29.940)
See, but I'm so OCD that you just mentioned
François Chollet (17:34.460)
intelligence and language and motion.
Lex Fridman (17:37.060)
I would start becoming paranoid
François Chollet (17:39.100)
that the categorization isn't perfect.
Lex Fridman (17:41.980)
Like that I would become paralyzed with the mind map
François Chollet (17:47.860)
that like this may not be.
Lex Fridman (17:49.660)
So like the, even though you're just doing
François Chollet (17:52.660)
associative kind of connections,
Lex Fridman (17:55.380)
there's an implied hierarchy that's emerging.
Lex Fridman (17:58.460)
And I would start becoming paranoid
Lex Fridman (17:59.900)
that it's not the proper hierarchy.
Lex Fridman (18:02.340)
So you're not just, one way to see mind maps
Lex Fridman (18:04.940)
is you're putting thoughts on paper.
François Chollet (18:07.060)
It's like a stream of consciousness,
Lex Fridman (18:10.580)
but then you can also start getting paranoid.
Lex Fridman (18:12.220)
Well, is this the right hierarchy?
Lex Fridman (18:15.140)
Sure, which it's mind maps, your mind map.
François Chollet (18:17.780)
You're free to draw anything you want.
Lex Fridman (18:19.420)
You're free to draw any connection you want.
Lex Fridman (18:20.860)
And you can just make a different mind map
Lex Fridman (18:23.420)
if you think the central node is not the right node.
François Chollet (18:26.260)
Yeah, I suppose there's a fear of being wrong.
Lex Fridman (18:29.700)
If you want to organize your ideas
François Chollet (18:32.660)
by writing down what you think,
Lex Fridman (18:35.540)
which I think is very effective.
François Chollet (18:37.380)
Like how do you know what you think about something
Lex Fridman (18:40.140)
if you don't write it down, right?
François Chollet (18:42.940)
If you do that, the thing is that it imposes
Lex Fridman (18:46.180)
much more syntactic structure over your ideas,
François Chollet (18:49.980)
which is not required with mind maps.
Lex Fridman (18:51.540)
So mind map is kind of like a lower level,
François Chollet (18:54.180)
more freehand way of organizing your thoughts.
Lex Fridman (18:57.900)
And once you've drawn it,
François Chollet (18:59.580)
then you can start actually voicing your thoughts
Lex Fridman (19:03.620)
in terms of, you know, paragraphs.
Lex Fridman (19:05.380)
It's a two dimensional aspect of layout too, right?
Lex Fridman (19:08.780)
Yeah.
François Chollet (19:09.620)
It's a kind of flower, I guess, you start.
Lex Fridman (19:12.860)
There's usually, you want to start with a central concept?
François Chollet (19:15.820)
Yes.
Lex Fridman (19:16.660)
Then you move out.
François Chollet (19:17.500)
Typically it ends up more like a subway map.
Lex Fridman (19:19.140)
So it ends up more like a graph,
François Chollet (19:20.660)
a topological graph without a root node.
Lex Fridman (19:23.500)
Yeah, so like in a subway map,
François Chollet (19:25.020)
there are some nodes that are more connected than others.
Lex Fridman (19:27.300)
And there are some nodes that are more important than others.
Lex Fridman (19:30.940)
So there are destinations,
Lex Fridman (19:32.380)
but it's not going to be purely like a tree, for instance.
François Chollet (19:36.420)
Yeah, it's fascinating to think that
Lex Fridman (19:38.540)
if there's something to that about the way our mind thinks.
François Chollet (19:42.420)
By the way, I just kind of remembered obvious thing
Lex Fridman (19:45.820)
that I have probably thousands of documents
François Chollet (19:49.020)
in Google Doc at this point, that are bullet point lists,
Lex Fridman (19:53.620)
which is, you can probably map a mind map
François Chollet (19:57.860)
to a bullet point list.
Lex Fridman (1:00:02.180)
How much intelligence, how do you
Lex Fridman (1:00:05.740)
build an intelligent system?
Lex Fridman (1:00:07.140)
And the coupled problem, how hard is this problem?
Lex Fridman (1:00:11.420)
How much intelligence does this problem actually require?
Lex Fridman (1:00:14.380)
So we get to cheat because we get
François Chollet (1:00:18.460)
to look at the problem.
Lex Fridman (1:00:20.700)
It's not like you get to close our eyes
Lex Fridman (1:00:22.860)
and completely new to driving.
Lex Fridman (1:00:24.740)
We get to do what we do as human beings, which
François Chollet (1:00:27.020)
is for the majority of our life before we ever
Lex Fridman (1:00:31.100)
learn, quote unquote, to drive.
François Chollet (1:00:32.460)
We get to watch other cars and other people drive.
Lex Fridman (1:00:35.460)
We get to be in cars.
François Chollet (1:00:36.540)
We get to watch.
Lex Fridman (1:00:37.540)
We get to see movies about cars.
François Chollet (1:00:39.500)
We get to observe all this stuff.
Lex Fridman (1:00:42.700)
And that's similar to what neural networks are doing.
François Chollet (1:00:45.060)
It's getting a lot of data, and the question
Lex Fridman (1:00:50.340)
is, yeah, how many leaps of reasoning genius
Lex Fridman (1:00:55.740)
is required to be able to actually effectively drive?
Lex Fridman (1:00:59.420)
I think it's a good example of driving.
François Chollet (1:01:01.260)
I mean, sure, you've seen a lot of cars in your life
Lex Fridman (1:01:06.260)
before you learned to drive.
Lex Fridman (1:01:07.700)
But let's say you've learned to drive in Silicon Valley,
Lex Fridman (1:01:10.620)
and now you rent a car in Tokyo.
François Chollet (1:01:14.100)
Well, now everyone is driving on the other side of the road,
Lex Fridman (1:01:16.820)
and the signs are different, and the roads
François Chollet (1:01:19.220)
are more narrow and so on.
Lex Fridman (1:01:20.500)
So it's a very, very different environment.
Lex Fridman (1:01:22.660)
And a smart human, even an average human,
Lex Fridman (1:01:26.780)
should be able to just zero shot it,
François Chollet (1:01:29.300)
to just be operational in this very different environment
Lex Fridman (1:01:34.260)
right away, despite having had no contact with the novel
François Chollet (1:01:40.500)
complexity that is contained in this environment.
Lex Fridman (1:01:44.140)
And that novel complexity is not just an interpolation
François Chollet (1:01:49.780)
over the situations that you've encountered previously,
Lex Fridman (1:01:52.420)
like learning to drive in the US.
François Chollet (1:01:55.060)
I would say the reason I ask is one
Lex Fridman (1:01:57.300)
of the most interesting tests of intelligence
François Chollet (1:01:59.940)
we have today actively, which is driving,
Lex Fridman (1:02:04.460)
in terms of having an impact on the world.
Lex Fridman (1:02:06.740)
When do you think we'll pass that test of intelligence?
Lex Fridman (1:02:09.900)
So I don't think driving is that much of a test of intelligence,
François Chollet (1:02:13.380)
because again, there is no task for which skill at that task
Lex Fridman (1:02:18.500)
demonstrates intelligence, unless it's
François Chollet (1:02:21.980)
a kind of meta task that involves acquiring new skills.
Lex Fridman (1:02:26.540)
So I don't think, I think you can actually
François Chollet (1:02:28.260)
solve driving without having any real amount of intelligence.
Lex Fridman (1:02:35.060)
For instance, if you did have infinite trained data,
François Chollet (1:02:39.540)
you could just literally train an end to end deep learning
Lex Fridman (1:02:42.660)
model that does driving, provided infinite trained data.
François Chollet (1:02:45.700)
The only problem with the whole idea
Lex Fridman (1:02:48.940)
is collecting a data set that's sufficiently comprehensive,
François Chollet (1:02:53.500)
that covers the very long tail of possible situations
Lex Fridman (1:02:56.380)
you might encounter.
Lex Fridman (1:02:57.260)
And it's really just a scale problem.
Lex Fridman (1:02:59.380)
So I think there's nothing fundamentally wrong
François Chollet (1:03:04.500)
with this plan, with this idea.
Lex Fridman (1:03:06.500)
It's just that it strikes me as a fairly inefficient thing
François Chollet (1:03:11.260)
to do, because you run into this scaling issue with diminishing
Lex Fridman (1:03:17.340)
returns.
François Chollet (1:03:17.860)
Whereas if instead you took a more manual engineering
Lex Fridman (1:03:21.980)
approach, where you use deep learning modules in combination
François Chollet (1:03:29.020)
with engineering an explicit model of the surrounding
Lex Fridman (1:03:33.220)
of the cars, and you bridge the two in a clever way,
François Chollet (1:03:36.100)
your model will actually start generalizing
Lex Fridman (1:03:38.900)
much earlier and more effectively
François Chollet (1:03:40.900)
than the end to end deep learning model.
Lex Fridman (1:03:42.540)
So why would you not go with the more manual engineering
Lex Fridman (1:03:46.500)
oriented approach?
Lex Fridman (1:03:47.900)
Even if you created that system, either the end
François Chollet (1:03:50.620)
to end deep learning model system that's
Lex Fridman (1:03:52.500)
running infinite data, or the slightly more human system,
François Chollet (1:03:58.500)
I don't think achieving L5 would demonstrate
Lex Fridman (1:04:02.740)
general intelligence or intelligence
François Chollet (1:04:04.540)
of any generality at all.
Lex Fridman (1:04:05.740)
Again, the only possible test of generality in AI
François Chollet (1:04:10.580)
would be a test that looks at skill acquisition
Lex Fridman (1:04:12.740)
over unknown tasks.
François Chollet (1:04:14.500)
For instance, you could take your L5 driver
Lex Fridman (1:04:17.380)
and ask it to learn to pilot a commercial airplane,
François Chollet (1:04:21.540)
for instance.
Lex Fridman (1:04:22.420)
And then you would look at how much human involvement is
François Chollet (1:04:25.180)
required and how much strength data
Lex Fridman (1:04:26.740)
is required for the system to learn to pilot an airplane.
Lex Fridman (1:04:29.860)
And that gives you a measure of how intelligent
Lex Fridman (1:04:35.020)
that system really is.
François Chollet (1:04:35.860)
Yeah, well, I mean, that's a big leap.
Lex Fridman (1:04:37.540)
I get you.
Lex Fridman (1:04:38.060)
But I'm more interested, as a problem, I would see,
Lex Fridman (1:04:42.820)
to me, driving is a black box that
François Chollet (1:04:47.380)
can generate novel situations at some rate,
Lex Fridman (1:04:51.180)
what people call edge cases.
Lex Fridman (1:04:53.500)
So it does have newness that keeps being like,
Lex Fridman (1:04:56.380)
we're confronted, let's say, once a month.
François Chollet (1:04:59.460)
It is a very long tail, yes.
Lex Fridman (1:05:00.660)
It's a long tail.
François Chollet (1:05:01.460)
That doesn't mean you cannot solve it just
Lex Fridman (1:05:05.620)
by training a statistical model and a lot of data.
François Chollet (1:05:08.740)
Huge amount of data.
Lex Fridman (1:05:09.820)
It's really a matter of scale.
Lex Fridman (1:05:11.900)
But I guess what I'm saying is if you have a vehicle that
Lex Fridman (1:05:16.020)
achieves level 5, it is going to be able to deal
François Chollet (1:05:21.580)
with new situations.
Lex Fridman (1:05:23.980)
Or, I mean, the data is so large that the rate of new situations
François Chollet (1:05:30.860)
is very low.
Lex Fridman (1:05:32.100)
Yes.
François Chollet (1:05:33.140)
That's not intelligent.
Lex Fridman (1:05:34.220)
So if we go back to your kind of definition of intelligence,
François Chollet (1:05:37.780)
it's the efficiency.
Lex Fridman (1:05:39.460)
With which you can adapt to new situations,
François Chollet (1:05:42.380)
to truly new situations, not situations you've seen before.
Lex Fridman (1:05:45.700)
Not situations that could be anticipated by your creators,
François Chollet (1:05:48.460)
by the creators of the system, but truly new situations.
Lex Fridman (1:05:51.740)
The efficiency with which you acquire new skills.
François Chollet (1:05:54.940)
If you require, if in order to pick up a new skill,
Lex Fridman (1:05:58.260)
you require a very extensive training
François Chollet (1:06:03.180)
data set of most possible situations
Lex Fridman (1:06:05.900)
that can occur in the practice of that skill,
François Chollet (1:06:08.940)
then the system is not intelligent.
Lex Fridman (1:06:10.620)
It is mostly just a lookup table.
François Chollet (1:06:15.060)
Yeah.
Lex Fridman (1:06:16.140)
Well, likewise, if in order to acquire a skill,
François Chollet (1:06:20.100)
you need a human engineer to write down
Lex Fridman (1:06:23.300)
a bunch of rules that cover most or every possible situation.
François Chollet (1:06:26.940)
Likewise, the system is not intelligent.
Lex Fridman (1:06:29.620)
The system is merely the output artifact
François Chollet (1:06:33.100)
of a process that happens in the minds of the engineers that
Lex Fridman (1:06:39.300)
are creating it.
François Chollet (1:06:40.820)
It is encoding an abstraction that's
Lex Fridman (1:06:44.700)
produced by the human mind.
Lex Fridman (1:06:46.420)
And intelligence would actually be
Lex Fridman (1:06:51.500)
the process of autonomously producing this abstraction.
François Chollet (1:06:56.260)
Yeah.
Lex Fridman (1:06:57.180)
Not like if you take an abstraction
Lex Fridman (1:06:59.260)
and you encode it on a piece of paper or in a computer program,
Lex Fridman (1:07:02.900)
the abstraction itself is not intelligent.
François Chollet (1:07:05.940)
What's intelligent is the agent that's
Lex Fridman (1:07:09.220)
capable of producing these abstractions.
François Chollet (1:07:11.780)
Yeah, it feels like there's a little bit of a gray area.
Lex Fridman (1:07:16.500)
Because you're basically saying that deep learning forms
François Chollet (1:07:18.860)
abstractions, too.
Lex Fridman (1:07:21.500)
But those abstractions do not seem
François Chollet (1:07:24.660)
to be effective for generalizing far outside of the things
Lex Fridman (1:07:29.140)
that it's already seen.
Lex Fridman (1:07:30.100)
But generalize a little bit.
Lex Fridman (1:07:31.620)
Yeah, absolutely.
François Chollet (1:07:32.620)
No, deep learning does generalize a little bit.
Lex Fridman (1:07:34.820)
Generalization is not binary.
François Chollet (1:07:36.980)
It's more like a spectrum.
Lex Fridman (1:07:38.140)
Yeah.
Lex Fridman (1:07:38.740)
And there's a certain point, it's a gray area,
Lex Fridman (1:07:40.860)
but there's a certain point where
François Chollet (1:07:42.500)
there's an impressive degree of generalization that happens.
Lex Fridman (1:07:47.340)
No, I guess exactly what you were saying
François Chollet (1:07:50.420)
is intelligence is how efficiently you're
Lex Fridman (1:07:56.420)
able to generalize far outside of the distribution of things
François Chollet (1:08:02.300)
you've seen already.
Lex Fridman (1:08:03.260)
Yes.
Lex Fridman (1:08:03.780)
So it's both the distance of how far you can,
Lex Fridman (1:08:07.180)
how new, how radically new something is,
Lex Fridman (1:08:10.180)
and how efficiently you're able to deal with that.
Lex Fridman (1:08:12.740)
So you can think of intelligence as a measure of an information
François Chollet (1:08:17.420)
conversion ratio.
Lex Fridman (1:08:19.140)
Imagine a space of possible situations.
Lex Fridman (1:08:23.420)
And you've covered some of them.
Lex Fridman (1:08:27.860)
So you have some amount of information
François Chollet (1:08:30.180)
about your space of possible situations
Lex Fridman (1:08:32.020)
that's provided by the situations you already know.
Lex Fridman (1:08:34.420)
And that's, on the other hand, also provided
Lex Fridman (1:08:36.540)
by the prior knowledge that the system brings
François Chollet (1:08:40.420)
to the table, the prior knowledge embedded
Lex Fridman (1:08:42.340)
in the system.
Lex Fridman (1:08:43.660)
So the system starts with some information
Lex Fridman (1:08:46.420)
about the problem, about the task.
Lex Fridman (1:08:48.860)
And it's about going from that information
Lex Fridman (1:08:52.500)
to a program, what we would call a skill program,
François Chollet (1:08:55.340)
a behavioral program, that can cover a large area
Lex Fridman (1:08:58.860)
of possible situation space.
Lex Fridman (1:09:01.660)
And essentially, the ratio between that area
Lex Fridman (1:09:04.100)
and the amount of information you start with is intelligence.
Lex Fridman (1:09:09.740)
So a very smart agent can make efficient use
Lex Fridman (1:09:14.180)
of very little information about a new problem
Lex Fridman (1:09:17.580)
and very little prior knowledge as well
Lex Fridman (1:09:19.580)
to cover a very large area of potential situations
François Chollet (1:09:23.380)
in that problem without knowing what these future new situations
Lex Fridman (1:09:28.500)
are going to be.
Lex Fridman (1:09:31.140)
So one of the other big things you talk about in the paper,
Lex Fridman (1:09:34.540)
we've talked about a little bit already,
Lex Fridman (1:09:36.300)
but let's talk about it some more,
Lex Fridman (1:09:37.860)
is the actual tests of intelligence.
Lex Fridman (1:09:41.020)
So if we look at human and machine intelligence,
Lex Fridman (1:09:45.980)
do you think tests of intelligence
François Chollet (1:09:48.100)
should be different for humans and machines,
Lex Fridman (1:09:50.340)
or how we think about testing of intelligence?
François Chollet (1:09:54.420)
Are these fundamentally the same kind of intelligences
Lex Fridman (1:09:59.740)
that we're after, and therefore, the tests should be similar?
Lex Fridman (1:10:03.780)
So if your goal is to create AIs that are more humanlike,
Lex Fridman (1:10:10.540)
then it would be super valuable, obviously,
François Chollet (1:10:12.540)
to have a test that's universal, that applies to both AIs
Lex Fridman (1:10:18.500)
and humans, so that you could establish
François Chollet (1:10:20.820)
a comparison between the two, that you
Lex Fridman (1:10:23.260)
could tell exactly how intelligent,
François Chollet (1:10:27.340)
in terms of human intelligence, a given system is.
Lex Fridman (1:10:30.420)
So that said, the constraints that
François Chollet (1:10:34.260)
apply to artificial intelligence and to human intelligence
Lex Fridman (1:10:37.620)
are very different.
Lex Fridman (1:10:39.340)
And your test should account for this difference.
Lex Fridman (1:10:44.860)
Because if you look at artificial systems,
François Chollet (1:10:47.140)
it's always possible for an experimenter
Lex Fridman (1:10:50.420)
to buy arbitrary levels of skill at arbitrary tasks,
François Chollet (1:10:55.580)
either by injecting hardcoded prior knowledge
Lex Fridman (1:11:01.100)
into the system via rules and so on that
François Chollet (1:11:05.660)
come from the human mind, from the minds of the programmers,
Lex Fridman (1:11:08.660)
and also buying higher levels of skill
François Chollet (1:11:12.980)
just by training on more data.
Lex Fridman (1:11:15.620)
For instance, you could generate an infinity
François Chollet (1:11:17.860)
of different Go games, and you could train a Go playing
Lex Fridman (1:11:21.660)
system that way, but you could not directly compare it
François Chollet (1:11:26.820)
to human Go playing skills.
Lex Fridman (1:11:28.620)
Because a human that plays Go had
François Chollet (1:11:31.100)
to develop that skill in a very constrained environment.
Lex Fridman (1:11:34.660)
They had a limited amount of time.
François Chollet (1:11:36.580)
They had a limited amount of energy.
Lex Fridman (1:11:38.940)
And of course, this started from a different set of priors.
François Chollet (1:11:42.620)
This started from innate human priors.
Lex Fridman (1:11:48.540)
So I think if you want to compare
François Chollet (1:11:49.860)
the intelligence of two systems, like the intelligence of an AI
Lex Fridman (1:11:53.260)
and the intelligence of a human, you have to control for priors.
François Chollet (1:11:59.780)
You have to start from the same set of knowledge priors
Lex Fridman (1:12:04.500)
about the task, and you have to control
François Chollet (1:12:06.940)
for experience, that is to say, for training data.
Lex Fridman (1:12:11.140)
So what's priors?
Lex Fridman (1:12:14.980)
So prior is whatever information you
Lex Fridman (1:12:18.340)
have about a given task before you
François Chollet (1:12:21.020)
start learning about this task.
Lex Fridman (1:12:23.100)
And how's that different from experience?
François Chollet (1:12:25.780)
Well, experience is acquired.
Lex Fridman (1:12:28.020)
So for instance, if you're trying to play Go,
François Chollet (1:12:31.100)
your experience with Go is all the Go games
Lex Fridman (1:12:33.900)
you've played, or you've seen, or you've simulated
François Chollet (1:12:37.060)
in your mind, let's say.
Lex Fridman (1:12:38.500)
And your priors are things like, well,
François Chollet (1:12:42.740)
Go is a game on the 2D grid.
Lex Fridman (1:12:45.860)
And we have lots of hardcoded priors
François Chollet (1:12:48.780)
about the organization of 2D space.
Lex Fridman (1:12:53.180)
And the rules of how the dynamics of the physics
Lex Fridman (1:12:58.340)
of this game in this 2D space?
Lex Fridman (1:12:59.980)
Yes.
Lex Fridman (1:13:00.580)
And the idea that you have what winning is.
Lex Fridman (1:13:04.300)
Yes, exactly.
Lex Fridman (1:13:05.580)
And other board games can also share some similarities with Go.
Lex Fridman (1:13:09.660)
And if you've played these board games, then,
François Chollet (1:13:12.060)
with respect to the game of Go, that
Lex Fridman (1:13:13.860)
would be part of your priors about the game.
François Chollet (1:13:16.300)
Well, it's interesting to think about the game of Go
Lex Fridman (1:13:18.500)
is how many priors are actually brought to the table.
François Chollet (1:13:22.620)
When you look at self play, reinforcement learning based
Lex Fridman (1:13:27.500)
mechanisms that do learning, it seems
François Chollet (1:13:29.300)
like the number of priors is pretty low.
Lex Fridman (1:13:31.020)
Yes.
Lex Fridman (1:13:31.380)
But you're saying you should be expec...
Lex Fridman (1:13:32.980)
There is a 2D special priors in the carbonate.
François Chollet (1:13:35.700)
Right.
Lex Fridman (1:13:36.460)
But you should be clear at making
François Chollet (1:13:39.020)
those priors explicit.
Lex Fridman (1:13:40.460)
Yes.
Lex Fridman (1:13:41.820)
So in particular, I think if your goal
Lex Fridman (1:13:44.060)
is to measure a humanlike form of intelligence,
François Chollet (1:13:47.700)
then you should clearly establish
Lex Fridman (1:13:49.700)
that you want the AI you're testing
François Chollet (1:13:52.820)
to start from the same set of priors that humans start with.
Lex Fridman (1:13:57.500)
Right.
Lex Fridman (1:13:58.820)
So I mean, to me personally, but I think to a lot of people,
Lex Fridman (1:14:02.740)
the human side of things is very interesting.
Lex Fridman (1:14:05.300)
So testing intelligence for humans.
Lex Fridman (1:14:08.020)
What do you think is a good test of human intelligence?
François Chollet (1:14:14.420)
Well, that's the question that psychometrics is interested in.
Lex Fridman (1:14:19.820)
There's an entire subfield of psychology
François Chollet (1:14:22.420)
that deals with this question.
Lex Fridman (1:14:23.860)
So what's psychometrics?
François Chollet (1:14:25.180)
The psychometrics is the subfield of psychology
Lex Fridman (1:14:27.980)
that tries to measure, quantify aspects of the human mind.
Lex Fridman (1:14:33.940)
So in particular, our cognitive abilities, intelligence,
Lex Fridman (1:14:36.940)
and personality traits as well.
Lex Fridman (1:14:39.660)
So what are, it might be a weird question,
Lex Fridman (1:14:43.620)
but what are the first principles of psychometrics
Lex Fridman (1:14:49.700)
this operates on?
Lex Fridman (1:14:52.100)
What are the priors it brings to the table?
Lex Fridman (1:14:55.340)
So it's a field with a fairly long history.
Lex Fridman (1:15:01.940)
So psychology sometimes gets a bad reputation
François Chollet (1:15:05.500)
for not having very reproducible results.
Lex Fridman (1:15:09.020)
And psychometrics has actually some fairly solidly
François Chollet (1:15:12.420)
reproducible results.
Lex Fridman (1:15:14.180)
So the ideal goals of the field is a test
François Chollet (1:15:17.980)
should be reliable, which is a notion tied to reproducibility.
Lex Fridman (1:15:23.060)
It should be valid, meaning that it should actually
François Chollet (1:15:26.540)
measure what you say it measures.
Lex Fridman (1:15:30.860)
So for instance, if you're saying
François Chollet (1:15:32.780)
that you're measuring intelligence,
Lex Fridman (1:15:34.140)
then your test results should be correlated
François Chollet (1:15:36.620)
with things that you expect to be correlated
Lex Fridman (1:15:39.140)
with intelligence like success in school
François Chollet (1:15:41.500)
or success in the workplace and so on.
Lex Fridman (1:15:43.580)
Should be standardized, meaning that you
François Chollet (1:15:46.540)
can administer your tests to many different people
Lex Fridman (1:15:48.980)
in some conditions.
Lex Fridman (1:15:50.780)
And it should be free from bias.
Lex Fridman (1:15:52.860)
Meaning that, for instance, if your test involves
François Chollet (1:15:57.140)
the English language, then you have
Lex Fridman (1:15:59.100)
to be aware that this creates a bias against people
François Chollet (1:16:02.500)
who have English as their second language
Lex Fridman (1:16:04.340)
or people who can't speak English at all.
Lex Fridman (1:16:07.300)
So of course, these principles for creating
Lex Fridman (1:16:10.100)
psychometric tests are very much an ideal.
François Chollet (1:16:13.420)
I don't think every psychometric test is really either
Lex Fridman (1:16:17.540)
reliable, valid, or free from bias.
Lex Fridman (1:16:22.060)
But at least the field is aware of these weaknesses
Lex Fridman (1:16:25.740)
and is trying to address them.
Lex Fridman (1:16:27.380)
So it's kind of interesting.
Lex Fridman (1:16:30.100)
Ultimately, you're only able to measure,
François Chollet (1:16:31.820)
like you said previously, the skill.
Lex Fridman (1:16:34.420)
But you're trying to do a bunch of measures
François Chollet (1:16:36.420)
of different skills that correlate,
Lex Fridman (1:16:38.820)
as you mentioned, strongly with some general concept
François Chollet (1:16:41.780)
of cognitive ability.
Lex Fridman (1:16:43.340)
Yes, yes.
Lex Fridman (1:16:44.060)
So what's the G factor?
Lex Fridman (1:16:46.620)
So right, there are many different kinds
François Chollet (1:16:48.140)
of tests of intelligence.
Lex Fridman (1:16:50.620)
And each of them is interesting in different aspects
François Chollet (1:16:55.340)
of intelligence.
Lex Fridman (1:16:56.060)
Some of them will deal with language.
François Chollet (1:16:57.580)
Some of them will deal with spatial vision,
Lex Fridman (1:17:00.940)
maybe mental rotations, numbers, and so on.
François Chollet (1:17:04.420)
When you run these very different tests at scale,
Lex Fridman (1:17:08.580)
what you start seeing is that there
François Chollet (1:17:10.940)
are clusters of correlations among test results.
Lex Fridman (1:17:14.220)
So for instance, if you look at homework at school,
François Chollet (1:17:19.300)
you will see that people who do well at math
Lex Fridman (1:17:21.780)
are also likely statistically to do well in physics.
Lex Fridman (1:17:25.500)
And what's more, people who do well at math and physics
Lex Fridman (1:17:30.060)
are also statistically likely to do well
François Chollet (1:17:32.620)
in things that sound completely unrelated,
Lex Fridman (1:17:35.580)
like writing an English essay, for instance.
Lex Fridman (1:17:38.420)
And so when you see clusters of correlations
Lex Fridman (1:17:42.700)
in statistical terms, you would explain them
François Chollet (1:17:46.140)
with the latent variable.
Lex Fridman (1:17:47.540)
And the latent variable that would, for instance, explain
François Chollet (1:17:51.100)
the relationship between being good at math
Lex Fridman (1:17:53.020)
and being good at physics would be cognitive ability.
Lex Fridman (1:17:57.020)
And the G factor is the latent variable
Lex Fridman (1:18:00.780)
that explains the fact that every test of intelligence
François Chollet (1:18:05.540)
that you can come up with results on this test
Lex Fridman (1:18:09.340)
end up being correlated.
Lex Fridman (1:18:10.540)
So there is some single unique variable
Lex Fridman (1:18:16.180)
that explains these correlations.
François Chollet (1:18:17.820)
That's the G factor.
Lex Fridman (1:18:18.820)
So it's a statistical construct.
François Chollet (1:18:20.380)
It's not really something you can directly measure,
Lex Fridman (1:18:23.060)
for instance, in a person.
Lex Fridman (1:18:25.540)
But it's there.
Lex Fridman (1:18:26.540)
But it's there.
François Chollet (1:18:27.220)
It's there.
Lex Fridman (1:18:27.740)
It's there at scale.
Lex Fridman (1:18:28.740)
And that's also one thing I want to mention about psychometrics.
Lex Fridman (1:18:33.460)
Like when you talk about measuring intelligence
François Chollet (1:18:36.620)
in humans, for instance, some people
Lex Fridman (1:18:38.660)
get a little bit worried.
François Chollet (1:18:40.060)
They will say, that sounds dangerous.
Lex Fridman (1:18:41.940)
Maybe that sounds potentially discriminatory, and so on.
Lex Fridman (1:18:44.340)
And they're not wrong.
Lex Fridman (1:18:46.460)
And the thing is, personally, I'm
François Chollet (1:18:48.220)
not interested in psychometrics as a way
Lex Fridman (1:18:51.100)
to characterize one individual person.
François Chollet (1:18:54.740)
Like if I get your psychometric personality
Lex Fridman (1:18:59.180)
assessments or your IQ, I don't think that actually
François Chollet (1:19:01.780)
tells me much about you as a person.
Lex Fridman (1:19:05.020)
I think psychometrics is most useful as a statistical tool.
Lex Fridman (1:19:10.300)
So it's most useful at scale.
Lex Fridman (1:19:12.500)
It's most useful when you start getting test results
François Chollet (1:19:15.420)
for a large number of people.
Lex Fridman (1:19:17.420)
And you start cross correlating these test results.
François Chollet (1:19:20.580)
Because that gives you information
Lex Fridman (1:19:23.620)
about the structure of the human mind,
François Chollet (1:19:26.420)
in particular about the structure
Lex Fridman (1:19:28.300)
of human cognitive abilities.
Lex Fridman (1:19:29.780)
So at scale, psychometrics paints a certain picture
Lex Fridman (1:19:34.860)
of the human mind.
Lex Fridman (1:19:35.620)
And that's interesting.
Lex Fridman (1:19:37.220)
And that's what's relevant to AI, the structure
François Chollet (1:19:39.540)
of human cognitive abilities.
Lex Fridman (1:19:41.060)
Yeah, it gives you an insight into it.
François Chollet (1:19:42.860)
I mean, to me, I remember when I learned about G factor,
Lex Fridman (1:19:45.820)
it seemed like it would be impossible for it
François Chollet (1:19:52.820)
to be real, even as a statistical variable.
Lex Fridman (1:19:55.500)
Like it felt kind of like astrology.
François Chollet (1:19:59.020)
Like it's like wishful thinking among psychologists.
Lex Fridman (1:20:01.980)
But the more I learned, I realized that there's some.
François Chollet (1:20:05.420)
I mean, I'm not sure what to make about human beings,
Lex Fridman (1:20:07.620)
the fact that the G factor is a thing.
François Chollet (1:20:10.580)
There's a commonality across all of human species,
Lex Fridman (1:20:13.260)
that there does seem to be a strong correlation
François Chollet (1:20:15.340)
between cognitive abilities.
Lex Fridman (1:20:17.140)
That's kind of fascinating, actually.
Lex Fridman (1:20:19.140)
So human cognitive abilities have a structure.
Lex Fridman (1:20:22.780)
Like the most mainstream theory of the structure
François Chollet (1:20:25.380)
of cognitive abilities is called CHC theory.
Lex Fridman (1:20:28.780)
It's Cattell, Horn, Carroll.
François Chollet (1:20:30.660)
It's named after the three psychologists who
Lex Fridman (1:20:33.180)
contributed key pieces of it.
Lex Fridman (1:20:35.340)
And it describes cognitive abilities
Lex Fridman (1:20:38.620)
as a hierarchy with three levels.
Lex Fridman (1:20:41.060)
And at the top, you have the G factor.
Lex Fridman (1:20:43.140)
Then you have broad cognitive abilities,
François Chollet (1:20:46.140)
for instance fluid intelligence, that
Lex Fridman (1:20:49.340)
encompass a broad set of possible kinds of tasks
François Chollet (1:20:54.940)
that are all related.
Lex Fridman (1:20:57.100)
And then you have narrow cognitive abilities
François Chollet (1:20:59.900)
at the last level, which is closer to task specific skill.
Lex Fridman (1:21:04.340)
And there are actually different theories of the structure
François Chollet (1:21:09.100)
of cognitive abilities that just emerge
Lex Fridman (1:21:10.700)
from different statistical analysis of IQ test results.
Lex Fridman (1:21:14.500)
But they all describe a hierarchy with a kind of G
Lex Fridman (1:21:18.500)
factor at the top.
Lex Fridman (1:21:21.140)
And you're right that the G factor,
Lex Fridman (1:21:23.740)
it's not quite real in the sense that it's not something
François Chollet (1:21:27.620)
you can observe and measure, like your height,
Lex Fridman (1:21:29.660)
for instance.
Lex Fridman (1:21:30.340)
But it's real in the sense that you
Lex Fridman (1:21:32.940)
see it in a statistical analysis of the data.
François Chollet (1:21:37.780)
One thing I want to mention is that the fact
Lex Fridman (1:21:39.700)
that there is a G factor does not really
François Chollet (1:21:41.540)
mean that human intelligence is general in a strong sense.
Lex Fridman (1:21:45.740)
It does not mean human intelligence
François Chollet (1:21:47.220)
can be applied to any problem at all,
Lex Fridman (1:21:50.340)
and that someone who has a high IQ
François Chollet (1:21:52.140)
is going to be able to solve any problem at all.
Lex Fridman (1:21:54.100)
That's not quite what it means.
François Chollet (1:21:55.260)
I think one popular analogy to understand it
Lex Fridman (1:22:00.420)
is the sports analogy.
François Chollet (1:22:03.340)
If you consider the concept of physical fitness,
Lex Fridman (1:22:06.660)
it's a concept that's very similar to intelligence
François Chollet (1:22:09.220)
because it's a useful concept.
Lex Fridman (1:22:11.340)
It's something you can intuitively understand.
François Chollet (1:22:14.460)
Some people are fit, maybe like you.
Lex Fridman (1:22:17.620)
Some people are not as fit, maybe like me.
Lex Fridman (1:22:20.540)
But none of us can fly.
Lex Fridman (1:22:22.980)
Absolutely.
François Chollet (1:22:23.700)
It's constrained to a specific set of skills.
Lex Fridman (1:22:25.460)
Even if you're very fit, that doesn't
François Chollet (1:22:27.060)
mean you can do anything at all in any environment.
Lex Fridman (1:22:31.020)
You obviously cannot fly.
François Chollet (1:22:32.420)
You cannot survive at the bottom of the ocean and so on.
Lex Fridman (1:22:36.020)
And if you were a scientist and you
François Chollet (1:22:38.540)
wanted to precisely define and measure physical fitness
Lex Fridman (1:22:42.820)
in humans, then you would come up with a battery of tests.
François Chollet (1:22:47.500)
You would have running 100 meter, playing soccer,
Lex Fridman (1:22:51.580)
playing table tennis, swimming, and so on.
Lex Fridman (1:22:54.260)
And if you ran these tests over many different people,
Lex Fridman (1:22:58.420)
you would start seeing correlations in test results.
François Chollet (1:23:01.220)
For instance, people who are good at soccer
Lex Fridman (1:23:03.020)
are also good at sprinting.
Lex Fridman (1:23:05.620)
And you would explain these correlations
Lex Fridman (1:23:08.580)
with physical abilities that are strictly
François Chollet (1:23:11.660)
analogous to cognitive abilities.
Lex Fridman (1:23:14.020)
And then you would start also observing correlations
François Chollet (1:23:17.060)
between biological characteristics,
Lex Fridman (1:23:21.220)
like maybe lung volume is correlated with being
François Chollet (1:23:24.900)
a fast runner, for instance, in the same way
Lex Fridman (1:23:27.820)
that there are neurophysical correlates of cognitive
François Chollet (1:23:32.500)
abilities.
Lex Fridman (1:23:33.940)
And at the top of the hierarchy of physical abilities
François Chollet (1:23:38.620)
that you would be able to observe,
Lex Fridman (1:23:39.980)
you would have a G factor, a physical G factor, which
François Chollet (1:23:43.340)
would map to physical fitness.
Lex Fridman (1:23:45.740)
And as you just said, that doesn't
François Chollet (1:23:47.980)
mean that people with high physical fitness can't fly.
Lex Fridman (1:23:51.340)
It doesn't mean human morphology and human physiology
François Chollet (1:23:54.500)
is universal.
Lex Fridman (1:23:55.660)
It's actually super specialized.
François Chollet (1:23:57.860)
We can only do the things that we were evolved to do.
Lex Fridman (1:24:04.100)
We are not appropriate to, you could not
François Chollet (1:24:08.340)
exist on Venus or Mars or in the void of space
Lex Fridman (1:24:11.100)
or the bottom of the ocean.
Lex Fridman (1:24:12.460)
So that said, one thing that's really striking and remarkable
Lex Fridman (1:24:17.740)
is that our morphology generalizes
François Chollet (1:24:23.060)
far beyond the environments that we evolved for.
Lex Fridman (1:24:27.260)
Like in a way, you could say we evolved to run after prey
Lex Fridman (1:24:31.180)
in the savanna, right?
Lex Fridman (1:24:32.900)
That's very much where our human morphology comes from.
Lex Fridman (1:24:36.820)
And that said, we can do a lot of things
Lex Fridman (1:24:40.220)
that are completely unrelated to that.
François Chollet (1:24:42.980)
We can climb mountains.
Lex Fridman (1:24:44.260)
We can swim across lakes.
François Chollet (1:24:47.260)
We can play table tennis.
Lex Fridman (1:24:48.980)
I mean, table tennis is very different from what
Lex Fridman (1:24:51.060)
we were evolved to do, right?
Lex Fridman (1:24:53.100)
So our morphology, our bodies, our sense and motor
François Chollet (1:24:56.300)
affordances have a degree of generality
Lex Fridman (1:24:59.500)
that is absolutely remarkable, right?
Lex Fridman (1:25:02.180)
And I think cognition is very similar to that.
Lex Fridman (1:25:05.300)
Our cognitive abilities have a degree of generality
François Chollet (1:25:08.260)
that goes far beyond what the mind was initially
Lex Fridman (1:25:11.180)
supposed to do, which is why we can play music and write
François Chollet (1:25:14.540)
novels and go to Mars and do all kinds of crazy things.
Lex Fridman (1:25:18.580)
But it's not universal in the same way
François Chollet (1:25:20.780)
that human morphology and our body
Lex Fridman (1:25:23.420)
is not appropriate for actually most of the universe by volume.
François Chollet (1:25:27.500)
In the same way, you could say that the human mind is not
Lex Fridman (1:25:29.940)
really appropriate for most of problem space,
François Chollet (1:25:32.620)
potential problem space by volume.
Lex Fridman (1:25:35.460)
So we have very strong cognitive biases, actually,
François Chollet (1:25:39.660)
that mean that there are certain types of problems
Lex Fridman (1:25:42.620)
that we handle very well and certain types of problems
François Chollet (1:25:45.380)
that we are completely in adapted for.
Lex Fridman (1:25:48.260)
So that's really how we'd interpret the G factor.
François Chollet (1:25:52.420)
It's not a sign of strong generality.
Lex Fridman (1:25:56.340)
It's really just the broadest cognitive ability.
Lex Fridman (1:26:01.020)
But our abilities, whether we are
Lex Fridman (1:26:03.020)
talking about sensory motor abilities or cognitive
François Chollet (1:26:05.820)
abilities, they still remain very specialized
Lex Fridman (1:26:09.460)
in the human condition, right?
François Chollet (1:26:12.420)
Within the constraints of the human cognition,
Lex Fridman (1:26:16.300)
they're general.
François Chollet (1:26:18.300)
Yes, absolutely.
Lex Fridman (1:26:19.500)
But the constraints, as you're saying, are very limited.
François Chollet (1:26:22.140)
I think what's limiting.
Lex Fridman (1:26:23.860)
So we evolved our cognition and our body
François Chollet (1:26:26.980)
evolved in very specific environments.
Lex Fridman (1:26:29.420)
Because our environment was so variable, fast changing,
Lex Fridman (1:26:32.740)
and so unpredictable, part of the constraints
Lex Fridman (1:26:35.740)
that drove our evolution is generality itself.
Lex Fridman (1:26:39.540)
So we were, in a way, evolved to be able to improvise
Lex Fridman (1:26:42.780)
in all kinds of physical or cognitive environments.
Lex Fridman (1:26:47.540)
And for this reason, it turns out
Lex Fridman (1:26:49.900)
that the minds and bodies that we ended up with
François Chollet (1:26:55.060)
can be applied to much, much broader scope
Lex Fridman (1:26:58.020)
than what they were evolved for.
Lex Fridman (1:27:00.060)
And that's truly remarkable.
Lex Fridman (1:27:01.740)
And that's a degree of generalization
François Chollet (1:27:03.940)
that is far beyond anything you can see in artificial systems
Lex Fridman (1:27:07.540)
today.
François Chollet (1:27:10.300)
That said, it does not mean that human intelligence
Lex Fridman (1:27:14.500)
is anywhere universal.
François Chollet (1:27:16.380)
Yeah, it's not general.
Lex Fridman (1:27:18.900)
It's a kind of exciting topic for people,
François Chollet (1:27:21.140)
even outside of artificial intelligence, is IQ tests.
Lex Fridman (1:27:27.580)
I think it's Mensa, whatever.
François Chollet (1:27:29.220)
There's different degrees of difficulty for questions.
Lex Fridman (1:27:32.420)
We talked about this offline a little bit, too,
François Chollet (1:27:34.700)
about difficult questions.
Lex Fridman (1:27:37.500)
What makes a question on an IQ test more difficult or less
Lex Fridman (1:27:42.300)
difficult, do you think?
Lex Fridman (1:27:43.700)
So the thing to keep in mind is that there's
François Chollet (1:27:46.500)
no such thing as a question that's intrinsically difficult.
Lex Fridman (1:27:51.540)
It has to be difficult to suspect to the things you
Lex Fridman (1:27:54.580)
already know and the things you can already do, right?
Lex Fridman (1:27:58.540)
So in terms of an IQ test question,
François Chollet (1:28:02.740)
typically it would be structured, for instance,
Lex Fridman (1:28:05.980)
as a set of demonstration input and output pairs, right?
Lex Fridman (1:28:11.860)
And then you would be given a test input, a prompt,
Lex Fridman (1:28:15.420)
and you would need to recognize or produce
François Chollet (1:28:18.700)
the corresponding output.
Lex Fridman (1:28:20.300)
And in that narrow context, you could say a difficult question
François Chollet (1:28:26.060)
is a question where the input prompt is
Lex Fridman (1:28:31.580)
very surprising and unexpected, given the training examples.
François Chollet (1:28:36.540)
Just even the nature of the patterns
Lex Fridman (1:28:38.340)
that you're observing in the input prompt.
François Chollet (1:28:40.180)
For instance, let's say you have a rotation problem.
Lex Fridman (1:28:43.260)
You must relate the shape by 90 degrees.
François Chollet (1:28:46.660)
If I give you two examples and then I give you one prompt,
Lex Fridman (1:28:50.500)
which is actually one of the two training examples,
François Chollet (1:28:53.020)
then there is zero generalization difficulty
Lex Fridman (1:28:55.700)
for the task.
François Chollet (1:28:56.380)
It's actually a trivial task.
Lex Fridman (1:28:57.500)
You just recognize that it's one of the training examples,
Lex Fridman (1:29:00.780)
and you produce the same answer.
Lex Fridman (1:29:02.300)
Now, if it's a more complex shape,
François Chollet (1:29:05.580)
there is a little bit more generalization,
Lex Fridman (1:29:07.700)
but it remains that you are still
François Chollet (1:29:09.860)
doing the same thing at this time,
Lex Fridman (1:29:12.060)
as you were being demonstrated at training time.
François Chollet (1:29:15.060)
A difficult task starts to require some amount of test
Lex Fridman (1:29:20.300)
time adaptation, some amount of improvisation, right?
Lex Fridman (1:29:25.100)
So consider, I don't know, you're
Lex Fridman (1:29:29.260)
teaching a class on quantum physics or something.
François Chollet (1:29:34.020)
If you wanted to test the understanding that students
Lex Fridman (1:29:40.460)
have of the material, you would come up
François Chollet (1:29:42.220)
with an exam that's very different from anything
Lex Fridman (1:29:47.740)
they've seen on the internet when they were cramming.
François Chollet (1:29:51.940)
On the other hand, if you wanted to make it easy,
Lex Fridman (1:29:54.780)
you would just give them something
François Chollet (1:29:56.340)
that's very similar to the mock exams
Lex Fridman (1:30:00.420)
that they've taken, something that's
François Chollet (1:30:03.060)
just a simple interpolation of questions
Lex Fridman (1:30:05.220)
that they've already seen.
Lex Fridman (1:30:07.260)
And so that would be an easy exam.
Lex Fridman (1:30:09.220)
It's very similar to what you've been trained on.
Lex Fridman (1:30:11.940)
And a difficult exam is one that really probes your understanding
Lex Fridman (1:30:15.460)
because it forces you to improvise.
François Chollet (1:30:18.980)
It forces you to do things that are
Lex Fridman (1:30:22.180)
different from what you were exposed to before.
Lex Fridman (1:30:24.780)
So that said, it doesn't mean that the exam that
Lex Fridman (1:30:29.100)
requires improvisation is intrinsically hard, right?
François Chollet (1:30:32.700)
Because maybe you're a quantum physics expert.
Lex Fridman (1:30:35.820)
So when you take the exam, this is actually
François Chollet (1:30:37.780)
stuff that, despite being new to the students,
Lex Fridman (1:30:40.300)
it's not new to you, right?
Lex Fridman (1:30:42.900)
So it can only be difficult with respect
Lex Fridman (1:30:46.020)
to what the test taker already knows
Lex Fridman (1:30:49.380)
and with respect to the information
Lex Fridman (1:30:51.780)
that the test taker has about the task.
Lex Fridman (1:30:54.700)
So that's what I mean by controlling for priors
Lex Fridman (1:30:57.860)
what the information you bring to the table.
Lex Fridman (1:30:59.900)
And the experience.
Lex Fridman (1:31:00.660)
And the experience, which is to train data.
Lex Fridman (1:31:02.660)
So in the case of the quantum physics exam,
Lex Fridman (1:31:05.580)
that would be all the course material itself
Lex Fridman (1:31:09.740)
and all the mock exams that students
Lex Fridman (1:31:11.500)
might have taken online.
François Chollet (1:31:12.820)
Yeah, it's interesting because I've also sent you an email.
Lex Fridman (1:31:17.700)
I asked you, I've been in just this curious question
François Chollet (1:31:21.820)
of what's a really hard IQ test question.
Lex Fridman (1:31:27.300)
And I've been talking to also people
François Chollet (1:31:30.580)
who have designed IQ tests.
Lex Fridman (1:31:32.540)
There's a few folks on the internet, it's like a thing.
François Chollet (1:31:34.420)
People are really curious about it.
Lex Fridman (1:31:36.180)
First of all, most of the IQ tests they designed,
François Chollet (1:31:39.460)
they like religiously protect against the correct answers.
Lex Fridman (1:31:45.620)
Like you can't find the correct answers anywhere.
François Chollet (1:31:48.380)
In fact, the question is ruined once you know,
Lex Fridman (1:31:50.620)
even like the approach you're supposed to take.
Lex Fridman (1:31:53.700)
So they're very...
Lex Fridman (1:31:54.540)
That said, the approach is implicit in the training examples.
Lex Fridman (1:31:58.420)
So if you release the training examples, it's over.
Lex Fridman (1:32:02.740)
Which is why in Arc, for instance,
François Chollet (1:32:04.980)
there is a test set that is private and no one has seen it.
Lex Fridman (1:32:09.140)
No, for really tough IQ questions, it's not obvious.
François Chollet (1:32:13.580)
It's not because the ambiguity.
Lex Fridman (1:32:17.100)
Like it's, I mean, we'll have to look through them,
Lex Fridman (1:32:20.780)
but like some number sequences and so on,
Lex Fridman (1:32:22.860)
it's not completely clear.
Lex Fridman (1:32:25.060)
So like you can get a sense, but there's like some,
Lex Fridman (1:32:30.540)
you know, when you look at a number sequence, I don't know,
François Chollet (1:32:36.140)
like your Fibonacci number sequence,
Lex Fridman (1:32:37.620)
if you look at the first few numbers,
François Chollet (1:32:39.580)
that sequence could be completed in a lot of different ways.
Lex Fridman (1:32:42.980)
And you know, some are, if you think deeply,
François Chollet (1:32:45.620)
are more correct than others.
Lex Fridman (1:32:46.900)
Like there's a kind of intuitive simplicity
Lex Fridman (1:32:51.300)
and elegance to the correct solution.
Lex Fridman (1:32:53.020)
Yes.
François Chollet (1:32:53.860)
I am personally not a fan of ambiguity
Lex Fridman (1:32:56.420)
in test questions actually,
Lex Fridman (1:32:58.660)
but I think you can have difficulty
Lex Fridman (1:33:01.140)
without requiring ambiguity simply by making the test
François Chollet (1:33:05.620)
require a lot of extrapolation over the training examples.
Lex Fridman (1:33:09.500)
But the beautiful question is difficult,
Lex Fridman (1:33:13.340)
but gives away everything
Lex Fridman (1:33:14.500)
when you give the training example.
François Chollet (1:33:17.180)
Basically, yes.
Lex Fridman (1:33:18.460)
Meaning that, so the tests I'm interested in creating
François Chollet (1:33:24.020)
are not necessarily difficult for humans
Lex Fridman (1:33:27.740)
because human intelligence is the benchmark.
François Chollet (1:33:31.580)
They're supposed to be difficult for machines
Lex Fridman (1:33:34.380)
in ways that are easy for humans.
François Chollet (1:33:36.300)
Like I think an ideal test of human and machine intelligence
Lex Fridman (1:33:40.820)
is a test that is actionable,
François Chollet (1:33:44.380)
that highlights the need for progress,
Lex Fridman (1:33:48.260)
and that highlights the direction
François Chollet (1:33:50.060)
in which you should be making progress.
Lex Fridman (1:33:51.500)
I think we'll talk about the ARC challenge
Lex Fridman (1:33:54.340)
and the test you've constructed
Lex Fridman (1:33:55.580)
and you have these elegant examples.
François Chollet (1:33:58.100)
I think that highlight,
Lex Fridman (1:33:59.180)
like this is really easy for us humans,
Lex Fridman (1:34:01.820)
but it's really hard for machines.
Lex Fridman (1:34:04.580)
But on the, you know, the designing an IQ test
François Chollet (1:34:09.220)
for IQs of like higher than 160 and so on,
Lex Fridman (1:34:13.380)
you have to say, you have to take that
Lex Fridman (1:34:15.220)
and put it on steroids, right?
Lex Fridman (1:34:16.500)
You have to think like, what is hard for humans?
Lex Fridman (1:34:19.540)
And that's a fascinating exercise in itself, I think.
Lex Fridman (1:34:25.940)
And it was an interesting question
François Chollet (1:34:27.740)
of what it takes to create a really hard question for humans
Lex Fridman (1:34:32.300)
because you again have to do the same process
François Chollet (1:34:36.340)
as you mentioned, which is, you know,
Lex Fridman (1:34:39.900)
something basically where the experience
François Chollet (1:34:45.100)
that you have likely to have encountered
Lex Fridman (1:34:46.900)
throughout your whole life,
François Chollet (1:34:48.740)
even if you've prepared for IQ tests,
Lex Fridman (1:34:51.780)
which is a big challenge,
François Chollet (1:34:53.380)
that this will still be novel for you.
Lex Fridman (1:34:55.820)
Yeah, I mean, novelty is a requirement.
François Chollet (1:34:58.900)
You should not be able to practice for the questions
Lex Fridman (1:35:02.100)
that you're gonna be tested on.
François Chollet (1:35:03.780)
That's important because otherwise what you're doing
Lex Fridman (1:35:06.700)
is not exhibiting intelligence.
Lex Fridman (1:35:08.180)
What you're doing is just retrieving
Lex Fridman (1:35:10.900)
what you've been exposed before.
François Chollet (1:35:12.380)
It's the same thing as deep learning model.
Lex Fridman (1:35:14.500)
If you train a deep learning model
François Chollet (1:35:15.900)
on all the possible answers, then it will ace your test
Lex Fridman (1:35:20.100)
in the same way that, you know,
François Chollet (1:35:24.420)
a stupid student can still ace the test
Lex Fridman (1:35:28.100)
if they cram for it.
François Chollet (1:35:30.140)
They memorize, you know,
Lex Fridman (1:35:32.500)
a hundred different possible mock exams.
Lex Fridman (1:35:34.980)
And then they hope that the actual exam
Lex Fridman (1:35:37.180)
will be a very simple interpolation of the mock exams.
Lex Fridman (1:35:41.180)
And that student could just be a deep learning model
Lex Fridman (1:35:43.180)
at that point.
Lex Fridman (1:35:44.020)
But you can actually do that
Lex Fridman (1:35:45.900)
without any understanding of the material.
Lex Fridman (1:35:48.180)
And in fact, many students pass their exams
Lex Fridman (1:35:50.540)
in exactly this way.
Lex Fridman (1:35:51.940)
And if you want to avoid that,
Lex Fridman (1:35:53.140)
you need an exam that's unlike anything they've seen
François Chollet (1:35:56.660)
that really probes their understanding.
Lex Fridman (1:36:00.020)
So how do we design an IQ test for machines,
Lex Fridman (1:36:05.020)
an intelligent test for machines?
Lex Fridman (1:36:07.860)
All right, so in the paper I outline
François Chollet (1:36:10.300)
a number of requirements that you expect of such a test.
Lex Fridman (1:36:14.780)
And in particular, we should start by acknowledging
François Chollet (1:36:19.620)
the priors that we expect to be required
Lex Fridman (1:36:23.300)
in order to perform the test.
Lex Fridman (1:36:25.260)
So we should be explicit about the priors, right?
Lex Fridman (1:36:28.100)
And if the goal is to compare machine intelligence
Lex Fridman (1:36:31.780)
and human intelligence,
Lex Fridman (1:36:32.740)
then we should assume human cognitive priors, right?
Lex Fridman (1:36:36.980)
And secondly, we should make sure that we are testing
Lex Fridman (1:36:42.020)
for skill acquisition ability,
François Chollet (1:36:44.820)
skill acquisition efficiency in particular,
Lex Fridman (1:36:46.740)
and not for skill itself.
François Chollet (1:36:48.580)
Meaning that every task featured in your test
Lex Fridman (1:36:51.860)
should be novel and should not be something
François Chollet (1:36:54.420)
that you can anticipate.
Lex Fridman (1:36:55.980)
So for instance, it should not be possible
Lex Fridman (1:36:57.980)
to brute force the space of possible questions, right?
Lex Fridman (1:37:02.860)
To pre generate every possible question and answer.
Lex Fridman (1:37:06.940)
So it should be tasks that cannot be anticipated,
Lex Fridman (1:37:10.660)
not just by the system itself,
Lex Fridman (1:37:12.460)
but by the creators of the system, right?
Lex Fridman (1:37:15.940)
Yeah, you know what's fascinating?
François Chollet (1:37:17.660)
I mean, one of my favorite aspects of the paper
Lex Fridman (1:37:20.820)
and the work you do with the ARC challenge
François Chollet (1:37:22.860)
is the process of making priors explicit.
Lex Fridman (1:37:28.940)
Just even that act alone is a really powerful one
François Chollet (1:37:33.420)
of like, what are, it's a really powerful question
Lex Fridman (1:37:39.260)
asked of us humans.
Lex Fridman (1:37:40.500)
What are the priors that we bring to the table?
Lex Fridman (1:37:44.180)
So the next step is like, once you have those priors,
Lex Fridman (1:37:46.900)
how do you use them to solve a novel task?
Lex Fridman (1:37:50.060)
But like, just even making the priors explicit
François Chollet (1:37:52.940)
is a really difficult and really powerful step.
Lex Fridman (1:37:56.140)
And that's like visually beautiful
Lex Fridman (1:37:58.940)
and conceptually philosophically beautiful part
Lex Fridman (1:38:01.340)
of the work you did with, and I guess continue to do
François Chollet (1:38:06.020)
probably with the paper and the ARC challenge.
Lex Fridman (1:38:08.460)
Can you talk about some of the priors
Lex Fridman (1:38:10.740)
that we're talking about here?
Lex Fridman (1:38:12.380)
Yes, so a researcher has done a lot of work
François Chollet (1:38:15.380)
on what exactly are the knowledge priors
Lex Fridman (1:38:19.460)
that are innate to humans is Elizabeth Spelke from Harvard.
Lex Fridman (1:38:26.500)
So she developed the core knowledge theory,
Lex Fridman (1:38:30.580)
which outlines four different core knowledge systems.
Lex Fridman (1:38:36.500)
So systems of knowledge that we are basically
Lex Fridman (1:38:39.180)
either born with or that we are hardwired
François Chollet (1:38:43.660)
to acquire very early on in our development.
Lex Fridman (1:38:47.180)
And there's no strong distinction between the two.
François Chollet (1:38:52.060)
Like if you are primed to acquire
Lex Fridman (1:38:57.060)
a certain type of knowledge in just a few weeks,
François Chollet (1:39:01.220)
you might as well just be born with it.
Lex Fridman (1:39:03.500)
It's just part of who you are.
Lex Fridman (1:39:06.460)
And so there are four different core knowledge systems.
Lex Fridman (1:39:09.500)
Like the first one is the notion of objectness
Lex Fridman (1:39:13.460)
and basic physics.
Lex Fridman (1:39:16.340)
Like you recognize that something that moves
François Chollet (1:39:20.700)
coherently, for instance, is an object.
Lex Fridman (1:39:23.220)
So we intuitively, naturally, innately divide the world
François Chollet (1:39:28.260)
into objects based on this notion of coherence,
Lex Fridman (1:39:31.260)
physical coherence.
Lex Fridman (1:39:32.740)
And in terms of elementary physics,
Lex Fridman (1:39:34.700)
there's the fact that objects can bump against each other
Lex Fridman (1:39:41.620)
and the fact that they can occlude each other.
Lex Fridman (1:39:44.460)
So these are things that we are essentially born with
François Chollet (1:39:48.300)
or at least that we are going to be acquiring extremely early
Lex Fridman (1:39:52.500)
because we're really hardwired to acquire them.
Lex Fridman (1:39:55.620)
So a bunch of points, pixels that move together
Lex Fridman (1:39:59.940)
on objects are partly the same object.
François Chollet (1:40:02.820)
Yes.
Lex Fridman (1:40:07.660)
I don't smoke weed, but if I did,
François Chollet (1:40:11.260)
that's something I could sit all night
Lex Fridman (1:40:13.100)
and just think about, remember what I wrote in your paper,
François Chollet (1:40:15.700)
just objectness, I wasn't self aware, I guess,
Lex Fridman (1:40:19.700)
of that particular prior.
François Chollet (1:40:23.180)
That's such a fascinating prior that like...
Lex Fridman (1:40:28.500)
That's the most basic one, but actually...
François Chollet (1:40:30.940)
Objectness, just identity, just objectness.
Lex Fridman (1:40:34.420)
It's very basic, I suppose, but it's so fundamental.
François Chollet (1:40:39.060)
It is fundamental to human cognition.
Lex Fridman (1:40:41.380)
Yeah.
François Chollet (1:40:42.220)
The second prior that's also fundamental is agentness,
Lex Fridman (1:40:46.660)
which is not a real world, a real world, so agentness.
François Chollet (1:40:50.740)
The fact that some of these objects
Lex Fridman (1:40:53.340)
that you segment your environment into,
François Chollet (1:40:56.540)
some of these objects are agents.
Lex Fridman (1:40:58.940)
So what's an agent?
François Chollet (1:41:00.300)
It's basically, it's an object that has goals.
Lex Fridman (1:41:05.380)
That has what?
François Chollet (1:41:06.340)
That has goals, that is capable of pursuing goals.
Lex Fridman (1:41:09.420)
So for instance, if you see two dots
François Chollet (1:41:12.580)
moving in roughly synchronized fashion,
Lex Fridman (1:41:16.300)
you will intuitively infer that one of the dots
François Chollet (1:41:19.820)
is pursuing the other.
Lex Fridman (1:41:21.620)
So that one of the dots is...
Lex Fridman (1:41:24.980)
And one of the dots is an agent
Lex Fridman (1:41:27.380)
and its goal is to avoid the other dot.
Lex Fridman (1:41:29.460)
And one of the dots, the other dot is also an agent
Lex Fridman (1:41:32.740)
and its goal is to catch the first dot.
François Chollet (1:41:35.860)
Belke has shown that babies as young as three months
Lex Fridman (1:41:40.540)
identify agentness and goal directedness
François Chollet (1:41:45.220)
in their environment.
Lex Fridman (1:41:46.420)
Another prior is basic geometry and topology,
François Chollet (1:41:52.140)
like the notion of distance,
Lex Fridman (1:41:53.660)
the ability to navigate in your environment and so on.
François Chollet (1:41:57.620)
This is something that is fundamentally hardwired
Lex Fridman (1:42:01.380)
into our brain.
François Chollet (1:42:02.700)
It's in fact backed by very specific neural mechanisms,
Lex Fridman (1:42:07.100)
like for instance, grid cells and place cells.
Lex Fridman (1:42:10.820)
So it's something that's literally hard coded
Lex Fridman (1:42:15.260)
at the neural level in our hippocampus.
Lex Fridman (1:42:19.940)
And the last prior would be the notion of numbers.
Lex Fridman (1:42:23.580)
Like numbers are not actually a cultural construct.
François Chollet (1:42:26.460)
We are intuitively, innately able to do some basic counting
Lex Fridman (1:42:31.460)
and to compare quantities.
Lex Fridman (1:42:34.100)
So it doesn't mean we can do arbitrary arithmetic.
Lex Fridman (1:42:37.660)
Counting, the actual counting.
François Chollet (1:42:39.020)
Counting, like counting one, two, three ish,
Lex Fridman (1:42:41.500)
then maybe more than three.
François Chollet (1:42:43.700)
You can also compare quantities.
Lex Fridman (1:42:45.140)
If I give you three dots and five dots,
François Chollet (1:42:48.580)
you can tell the side with five dots has more dots.
Lex Fridman (1:42:52.500)
So this is actually an innate prior.
Lex Fridman (1:42:56.580)
So that said, the list may not be exhaustive.
Lex Fridman (1:43:00.020)
So SpellKey is still, you know,
François Chollet (1:43:02.580)
passing the potential existence of new knowledge systems.
Lex Fridman (1:43:08.500)
For instance, knowledge systems that we deal
François Chollet (1:43:12.100)
with social relationships.
Lex Fridman (1:43:15.940)
Yeah, I mean, and there could be...
François Chollet (1:43:17.700)
Which is much less relevant to something like ARC
Lex Fridman (1:43:22.060)
or IQ test and so on.
François Chollet (1:43:22.900)
Right.
Lex Fridman (1:43:23.740)
There could be stuff that's like you said,
François Chollet (1:43:26.740)
rotation, symmetry, is there like...
Lex Fridman (1:43:29.020)
Symmetry is really interesting.
François Chollet (1:43:31.060)
It's very likely that there is, speaking about rotation,
Lex Fridman (1:43:34.380)
that there is in the brain, a hard coded system
François Chollet (1:43:38.900)
that is capable of performing rotations.
Lex Fridman (1:43:42.060)
One famous experiment that people did in the...
François Chollet (1:43:45.660)
I don't remember which was exactly,
Lex Fridman (1:43:48.180)
but in the 70s was that people found that
François Chollet (1:43:53.180)
if you asked people, if you give them two different shapes
Lex Fridman (1:43:57.580)
and one of the shapes is a rotated version
François Chollet (1:44:01.420)
of the first shape, and you ask them,
Lex Fridman (1:44:03.340)
is that shape a rotated version of the first shape or not?
Lex Fridman (1:44:07.060)
What you see is that the time it takes people to answer
Lex Fridman (1:44:11.140)
is linearly proportional, right, to the angle of rotation.
Lex Fridman (1:44:16.140)
So it's almost like you have somewhere in your brain
Lex Fridman (1:44:19.660)
like a turntable with a fixed speed.
Lex Fridman (1:44:24.020)
And if you want to know if two objects are a rotated version
Lex Fridman (1:44:28.620)
of each other, you put the object on the turntable,
François Chollet (1:44:31.700)
you let it move around a little bit,
Lex Fridman (1:44:34.740)
and then you stop when you have a match.
Lex Fridman (1:44:37.580)
And that's really interesting.
Lex Fridman (1:44:40.140)
So what's the ARC challenge?
Lex Fridman (1:44:42.740)
So in the paper, I outline all these principles
Lex Fridman (1:44:47.380)
that a good test of machine intelligence
Lex Fridman (1:44:50.140)
and human intelligence should follow.
Lex Fridman (1:44:51.940)
And the ARC challenge is one attempt
François Chollet (1:44:55.300)
to embody as many of these principles as possible.
Lex Fridman (1:44:58.540)
So I don't think it's anywhere near a perfect attempt, right?
François Chollet (1:45:03.780)
It does not actually follow every principle,
Lex Fridman (1:45:06.060)
but it is what I was able to do given the constraints.
Lex Fridman (1:45:10.700)
So the format of ARC is very similar to classic IQ tests,
Lex Fridman (1:45:15.540)
in particular Raven's Progressive Metrices.
Lex Fridman (1:45:18.020)
Raven's?
Lex Fridman (1:45:18.980)
Yeah, Raven's Progressive Metrices.
François Chollet (1:45:20.580)
I mean, if you've done IQ tests in the past,
Lex Fridman (1:45:22.820)
you know what that is, probably.
François Chollet (1:45:24.220)
Or at least you've seen it, even if you
Lex Fridman (1:45:25.620)
don't know what it's called.
Lex Fridman (1:45:26.980)
And so you have a set of tasks, that's what they're called.
Lex Fridman (1:45:32.300)
And for each task, you have training data,
François Chollet (1:45:37.180)
which is a set of input and output pairs.
Lex Fridman (1:45:40.260)
So an input or output pair is a grid of colors, basically.
François Chollet (1:45:45.540)
The grid, the size of the grid is variables.
Lex Fridman (1:45:48.500)
The size of the grid is variable.
Lex Fridman (1:45:51.380)
And you're given an input, and you must transform it
Lex Fridman (1:45:56.100)
into the proper output.
Lex Fridman (1:45:59.020)
And so you're shown a few demonstrations
Lex Fridman (1:46:02.060)
of a task in the form of existing input output pairs,
Lex Fridman (1:46:05.100)
and then you're given a new input.
Lex Fridman (1:46:06.860)
And you must provide, you must produce the correct output.
Lex Fridman (1:46:12.620)
And the assumptions in Arc is that every task should only
Lex Fridman (1:46:22.860)
require core knowledge priors, should not
François Chollet (1:46:27.660)
require any outside knowledge.
Lex Fridman (1:46:30.460)
So for instance, no language, no English, nothing like this.
François Chollet (1:46:36.900)
No concepts taken from our human experience,
Lex Fridman (1:46:41.540)
like trees, dogs, cats, and so on.
Lex Fridman (1:46:44.340)
So only reasoning tasks that are built on top
Lex Fridman (1:46:49.700)
of core knowledge priors.
Lex Fridman (1:46:52.060)
And some of the tasks are actually explicitly
Lex Fridman (1:46:56.260)
trying to probe specific forms of abstraction.
François Chollet (1:47:02.220)
Part of the reason why I wanted to create Arc
Lex Fridman (1:47:05.500)
is I'm a big believer in when you're
François Chollet (1:47:11.740)
faced with a problem as murky as understanding
Lex Fridman (1:47:18.340)
how to autonomously generate abstraction in a machine,
François Chollet (1:47:22.380)
you have to coevolve the solution and the problem.
Lex Fridman (1:47:27.180)
And so part of the reason why I designed Arc
François Chollet (1:47:29.380)
was to clarify my ideas about the nature of abstraction.
Lex Fridman (1:47:34.660)
And some of the tasks are actually
François Chollet (1:47:36.220)
designed to probe bits of that theory.
Lex Fridman (1:47:39.900)
And there are things that turn out
François Chollet (1:47:42.340)
to be very easy for humans to perform, including young kids,
Lex Fridman (1:47:46.740)
but turn out to be near impossible for machines.
Lex Fridman (1:47:50.500)
So what have you learned from the nature of abstraction
Lex Fridman (1:47:53.780)
from designing that?
Lex Fridman (1:47:58.380)
Can you clarify what you mean?
Lex Fridman (1:47:59.620)
One of the things you wanted to try to understand
François Chollet (1:48:02.300)
was this idea of abstraction.
Lex Fridman (1:48:06.020)
Yes, so clarifying my own ideas about abstraction
François Chollet (1:48:10.380)
by forcing myself to produce tasks that
Lex Fridman (1:48:13.700)
would require the ability to produce
François Chollet (1:48:17.020)
that form of abstraction in order to solve them.
Lex Fridman (1:48:19.900)
Got it.
François Chollet (1:48:20.860)
OK, so and by the way, just the people should check out.
Lex Fridman (1:48:24.060)
I'll probably overlay if you're watching the video part.
Lex Fridman (1:48:26.380)
But the grid input output with the different colors
Lex Fridman (1:48:32.180)
on the grid, that's it.
François Chollet (1:48:34.340)
I mean, it's a very simple world,
Lex Fridman (1:48:36.300)
but it's kind of beautiful.
François Chollet (1:48:37.460)
It's very similar to classic IQ tests.
Lex Fridman (1:48:39.740)
It's not very original in that sense.
François Chollet (1:48:41.620)
The main difference with IQ tests
Lex Fridman (1:48:43.260)
is that we make the priors explicit, which is not
François Chollet (1:48:46.860)
usually the case in IQ tests.
Lex Fridman (1:48:48.580)
So you make it explicit that everything should only
François Chollet (1:48:50.820)
be built on top of core knowledge priors.
Lex Fridman (1:48:53.860)
I also think it's generally more diverse than IQ tests
François Chollet (1:48:58.620)
in general.
Lex Fridman (1:49:00.300)
And it perhaps requires a bit more manual work
François Chollet (1:49:03.820)
to produce solutions, because you
Lex Fridman (1:49:05.460)
have to click around on a grid for a while.
François Chollet (1:49:08.500)
Sometimes the grids can be as large as 30 by 30 cells.
Lex Fridman (1:49:12.020)
So how did you come up, if you can reveal, with the questions?
Lex Fridman (1:49:18.020)
What's the process of the questions?
Lex Fridman (1:49:19.580)
Was it mostly you that came up with the questions?
Lex Fridman (1:49:23.380)
How difficult is it to come up with a question?
Lex Fridman (1:49:25.780)
Is this scalable to a much larger number?
François Chollet (1:49:30.700)
If we think, with IQ tests, you might not necessarily
Lex Fridman (1:49:33.740)
want it to or need it to be scalable.
François Chollet (1:49:36.460)
With machines, it's possible, you
Lex Fridman (1:49:39.580)
could argue, that it needs to be scalable.
Lex Fridman (1:49:41.620)
So there are 1,000 questions, 1,000 tasks,
Lex Fridman (1:49:46.500)
including the test set, the prior test set.
François Chollet (1:49:49.140)
I think it's fairly difficult in the sense
Lex Fridman (1:49:51.060)
that a big requirement is that every task should
François Chollet (1:49:54.500)
be novel and unique and unpredictable.
Lex Fridman (1:50:00.140)
You don't want to create your own little world that
François Chollet (1:50:04.460)
is simple enough that it would be possible for a human
Lex Fridman (1:50:08.860)
to reverse and generate and write down
François Chollet (1:50:12.580)
an algorithm that could generate every possible arc
Lex Fridman (1:50:15.940)
task and their solution.
Lex Fridman (1:50:17.060)
So that would completely invalidate the test.
Lex Fridman (1:50:19.340)
So you're constantly coming up with new stuff.
François Chollet (1:50:21.700)
Yeah, you need a source of novelty,
Lex Fridman (1:50:24.820)
of unfakeable novelty.
Lex Fridman (1:50:27.860)
And one thing I found is that, as a human,
Lex Fridman (1:50:32.020)
you are not a very good source of unfakeable novelty.
Lex Fridman (1:50:36.460)
And so you have to base the creation of these tasks
Lex Fridman (1:50:40.580)
quite a bit.
François Chollet (1:50:41.100)
There are only so many unique tasks
Lex Fridman (1:50:42.980)
that you can do in a given day.
Lex Fridman (1:50:45.580)
So that means coming up with truly original new ideas.
Lex Fridman (1:50:49.860)
Did psychedelics help you at all?
François Chollet (1:50:52.380)
No, I'm just kidding.
Lex Fridman (1:50:53.780)
But I mean, that's fascinating to think about.
Lex Fridman (1:50:55.820)
So you would be walking or something like that.
Lex Fridman (1:50:58.780)
Are you constantly thinking of something totally new?
François Chollet (1:51:02.860)
Yes.
Lex Fridman (1:51:06.020)
This is hard.
François Chollet (1:51:06.980)
This is hard.
Lex Fridman (1:51:07.620)
Yeah, I mean, I'm not saying you've done anywhere
François Chollet (1:51:10.980)
near a perfect job at it.
Lex Fridman (1:51:12.380)
There is some amount of redundancy,
Lex Fridman (1:51:14.540)
and there are many imperfections in ARC.
Lex Fridman (1:51:16.740)
So that said, you should consider
François Chollet (1:51:18.540)
ARC as a work in progress.
Lex Fridman (1:51:19.820)
It is not the definitive state.
François Chollet (1:51:25.180)
The ARC tasks today are not the definitive state of the test.
Lex Fridman (1:51:29.300)
I want to keep refining it in the future.
François Chollet (1:51:32.780)
I also think it should be possible to open up
Lex Fridman (1:51:36.180)
the creation of tasks to a broad audience
François Chollet (1:51:38.660)
to do crowdsourcing.
Lex Fridman (1:51:40.860)
That would involve several levels of filtering,
François Chollet (1:51:43.180)
obviously.
Lex Fridman (1:51:44.140)
But I think it's possible to apply crowdsourcing
François Chollet (1:51:46.260)
to develop a much bigger and much more diverse ARC data set.
Lex Fridman (1:51:51.140)
That would also be free of potentially some
François Chollet (1:51:54.020)
of my own personal biases.
Lex Fridman (1:51:56.700)
Is there always need to be a part of ARC
Lex Fridman (1:51:59.220)
that the test is hidden?
Lex Fridman (1:52:02.900)
Yes, absolutely.
François Chollet (1:52:04.140)
It is imperative that the tests that you're
Lex Fridman (1:52:08.900)
using to actually benchmark algorithms
François Chollet (1:52:11.900)
is not accessible to the people developing these algorithms.
Lex Fridman (1:52:15.220)
Because otherwise, what's going to happen
François Chollet (1:52:16.860)
is that the human engineers are just
Lex Fridman (1:52:19.100)
going to solve the tasks themselves
Lex Fridman (1:52:21.820)
and encode their solution in program form.
Lex Fridman (1:52:24.820)
But that, again, what you're seeing here
François Chollet (1:52:27.420)
is the process of intelligence happening
Lex Fridman (1:52:30.100)
in the mind of the human.
Lex Fridman (1:52:31.180)
And then you're just capturing its crystallized output.
Lex Fridman (1:52:35.460)
But that crystallized output is not the same thing
François Chollet (1:52:38.260)
as the process it generated.
Lex Fridman (1:52:40.020)
It's not intelligent in itself.
Lex Fridman (1:52:41.340)
So what, by the way, the idea of crowdsourcing it
Lex Fridman (1:52:43.980)
is fascinating.
François Chollet (1:52:45.860)
I think the creation of questions
Lex Fridman (1:52:49.860)
is really exciting for people.
François Chollet (1:52:51.460)
I think there's a lot of really brilliant people
Lex Fridman (1:52:53.980)
out there that love to create these kinds of stuff.
François Chollet (1:52:56.220)
Yeah, one thing that kind of surprised me
Lex Fridman (1:52:59.060)
that I wasn't expecting is that lots of people
François Chollet (1:53:01.620)
seem to actually enjoy ARC as a kind of game.
Lex Fridman (1:53:05.980)
And I was releasing it as a test,
François Chollet (1:53:08.820)
as a benchmark of fluid general intelligence.
Lex Fridman (1:53:14.100)
And lots of people just, including kids,
François Chollet (1:53:17.100)
just started enjoying it as a game.
Lex Fridman (1:53:18.900)
So I think that's encouraging.
François Chollet (1:53:20.980)
Yeah, I'm fascinated by it.
Lex Fridman (1:53:22.300)
There's a world of people who create IQ questions.
François Chollet (1:53:25.940)
I think that's a cool activity for machines and for humans.
Lex Fridman (1:53:32.660)
And humans are themselves fascinated
François Chollet (1:53:35.420)
by taking the questions, like measuring
Lex Fridman (1:53:40.220)
their own intelligence.
François Chollet (1:53:42.300)
I mean, that's just really compelling.
Lex Fridman (1:53:44.420)
It's really interesting to me, too.
François Chollet (1:53:47.020)
One of the cool things about ARC, you said,
Lex Fridman (1:53:48.740)
is kind of inspired by IQ tests or whatever
François Chollet (1:53:51.620)
follows a similar process.
Lex Fridman (1:53:53.460)
But because of its nature, because of the context
François Chollet (1:53:56.060)
in which it lives, it immediately
Lex Fridman (1:53:59.020)
forces you to think about the nature of intelligence
François Chollet (1:54:01.660)
as opposed to just the test of your own.
Lex Fridman (1:54:04.220)
It forces you to really think.
François Chollet (1:54:06.020)
I don't know if it's within the question,
Lex Fridman (1:54:09.900)
inherent in the question, or just the fact
François Chollet (1:54:11.860)
that it lives in the test that's supposed
Lex Fridman (1:54:13.780)
to be a test of machine intelligence.
François Chollet (1:54:15.340)
Absolutely.
Lex Fridman (1:54:15.900)
As you solve ARC tasks as a human,
François Chollet (1:54:20.660)
you will be forced to basically introspect
Lex Fridman (1:54:24.700)
how you come up with solutions.
Lex Fridman (1:54:27.060)
And that forces you to reflect on the human problem solving
Lex Fridman (1:54:32.660)
process.
Lex Fridman (1:54:33.820)
And the way your own mind generates
Lex Fridman (1:54:38.780)
abstract representations of the problems it's exposed to.
François Chollet (1:54:44.780)
I think it's due to the fact that the set of core knowledge
Lex Fridman (1:54:48.860)
priors that ARC is built upon is so small.
François Chollet (1:54:52.460)
It's all a recombination of a very, very small set
Lex Fridman (1:54:58.660)
of assumptions.
Lex Fridman (1:55:00.460)
OK, so what's the future of ARC?
Lex Fridman (1:55:02.900)
So you held ARC as a challenge, as part
François Chollet (1:55:05.420)
of like a Kaggle competition.
Lex Fridman (1:55:06.700)
Yes.
François Chollet (1:55:07.180)
Kaggle competition.
Lex Fridman (1:55:08.420)
And what do you think?
Lex Fridman (1:55:11.860)
Do you think that's something that
Lex Fridman (1:55:13.060)
continues for five years, 10 years,
Lex Fridman (1:55:16.060)
like just continues growing?
Lex Fridman (1:55:17.820)
Yes, absolutely.
Lex Fridman (1:55:18.940)
So ARC itself will keep evolving.
Lex Fridman (1:55:21.340)
So I've talked about crowdsourcing.
François Chollet (1:55:22.780)
I think that's a good avenue.
Lex Fridman (1:55:26.180)
Another thing I'm starting is I'll
François Chollet (1:55:29.340)
be collaborating with folks from the psychology department
Lex Fridman (1:55:32.700)
at NYU to do human testing on ARC.
Lex Fridman (1:55:36.660)
And I think there are lots of interesting questions
Lex Fridman (1:55:38.940)
you can start asking, especially as you start correlating
François Chollet (1:55:43.940)
machine solutions to ARC tasks and the human characteristics
Lex Fridman (1:55:49.420)
of solutions.
François Chollet (1:55:50.060)
Like for instance, you can try to see
Lex Fridman (1:55:52.020)
if there's a relationship between the human perceived
François Chollet (1:55:55.660)
difficulty of a task and the machine perceived.
Lex Fridman (1:55:59.420)
Yes, and exactly some measure of machine
François Chollet (1:56:01.940)
perceived difficulty.
Lex Fridman (1:56:02.780)
Yeah, it's a nice playground in which
François Chollet (1:56:04.900)
to explore this very difference.
Lex Fridman (1:56:06.340)
It's the same thing as we talked about the autonomous vehicles.
François Chollet (1:56:09.260)
The things that could be difficult for humans
Lex Fridman (1:56:10.900)
might be very different than the things that are difficult.
Lex Fridman (1:56:13.100)
And formalizing or making explicit that difference
Lex Fridman (1:56:17.300)
in difficulty may teach us something fundamental
François Chollet (1:56:21.020)
about intelligence.
Lex Fridman (1:56:22.340)
So one thing I think we did well with ARC
François Chollet (1:56:26.420)
is that it's proving to be a very actionable test in the sense
Lex Fridman (1:56:33.060)
that machine performance on ARC started at very much zero
François Chollet (1:56:37.700)
initially, while humans found actually the task very easy.
Lex Fridman (1:56:43.340)
And that alone was like a big red flashing light saying
François Chollet (1:56:48.180)
that something is going on and that we are missing something.
Lex Fridman (1:56:52.380)
And at the same time, machine performance
François Chollet (1:56:55.420)
did not stay at zero for very long.
Lex Fridman (1:56:57.660)
Actually, within two weeks of the Kaggle competition,
François Chollet (1:57:00.260)
we started having a nonzero number.
Lex Fridman (1:57:03.220)
And now the state of the art is around 20%
François Chollet (1:57:06.460)
of the test set solved.
Lex Fridman (1:57:10.260)
And so ARC is actually a challenge
François Chollet (1:57:12.500)
where our capabilities start at zero, which indicates
Lex Fridman (1:57:16.860)
the need for progress.
Lex Fridman (1:57:18.180)
But it's also not an impossible challenge.
Lex Fridman (1:57:20.580)
It's not accessible.
François Chollet (1:57:21.500)
You can start making progress basically right away.
Lex Fridman (1:57:25.260)
At the same time, we are still very far
François Chollet (1:57:28.380)
from having solved it.
Lex Fridman (1:57:29.420)
And that's actually a very positive outcome
François Chollet (1:57:32.820)
of the competition is that the competition has proven
Lex Fridman (1:57:35.900)
that there was no obvious shortcut to solve these tasks.
François Chollet (1:57:41.740)
Yeah, so the test held up.
Lex Fridman (1:57:43.180)
Yeah, exactly.
François Chollet (1:57:44.340)
That was the primary reason to use the Kaggle competition
Lex Fridman (1:57:46.900)
is to check if some clever person was
François Chollet (1:57:51.540)
going to hack the benchmark that did not happen.
Lex Fridman (1:57:56.380)
People who are solving the task are essentially doing it.
François Chollet (1:58:01.060)
Well, in a way, they're actually exploring some flaws of ARC
Lex Fridman (1:58:05.580)
that we will need to address in the future,
François Chollet (1:58:07.380)
especially they're essentially anticipating
Lex Fridman (1:58:09.900)
what sort of tasks may be contained in the test set.
François Chollet (1:58:13.780)
Right, which is kind of, yeah, that's the kind of hacking.
Lex Fridman (1:58:18.460)
It's human hacking of the test.
François Chollet (1:58:20.180)
Yes, that said, with the state of the art,
Lex Fridman (1:58:23.380)
it's like 20% we're still very, very far from human level,
François Chollet (1:58:28.220)
which is closer to 100%.
Lex Fridman (1:58:30.940)
And I do believe that it will take a while
François Chollet (1:58:35.540)
until we reach human parity on ARC.
Lex Fridman (1:58:40.500)
And that by the time we have human parity,
François Chollet (1:58:43.540)
we will have AI systems that are probably
Lex Fridman (1:58:47.020)
pretty close to human level in terms of general fluid
François Chollet (1:58:50.740)
intelligence, which is, I mean, they are not
Lex Fridman (1:58:53.260)
going to be necessarily human like.
François Chollet (1:58:54.940)
They're not necessarily, you would not necessarily
Lex Fridman (1:58:58.780)
recognize them as being an AGI.
Lex Fridman (1:59:01.860)
But they would be capable of a degree of generalization
Lex Fridman (1:59:06.860)
that matches the generalization performed
François Chollet (1:59:09.820)
by human fluid intelligence.
Lex Fridman (1:59:11.300)
Sure.
François Chollet (1:59:11.860)
I mean, this is a good point in terms
Lex Fridman (1:59:13.380)
of general fluid intelligence to mention in your paper.
François Chollet (1:59:17.700)
You describe different kinds of generalizations,
Lex Fridman (1:59:21.060)
local, broad, extreme.
Lex Fridman (1:59:23.460)
And there's a kind of a hierarchy that you form.
Lex Fridman (1:59:25.660)
So when we say generalizations, what are we talking about?
Lex Fridman (1:59:31.820)
What kinds are there?
Lex Fridman (1:59:33.180)
Right, so generalization is a very old idea.
François Chollet (1:59:37.020)
I mean, it's even older than machine learning.
Lex Fridman (1:59:39.420)
In the context of machine learning,
François Chollet (1:59:40.980)
you say a system generalizes if it can make sense of an input
Lex Fridman (1:59:47.140)
it has not yet seen.
Lex Fridman (1:59:49.580)
And that's what I would call system centric generalization,
Lex Fridman (1:59:54.940)
generalization with respect to novelty
François Chollet (20:01.460)
It's the same, it's a, no, it's not, it's a tree.
Lex Fridman (20:05.060)
It's a tree, yeah.
Lex Fridman (20:06.220)
So I create trees,
Lex Fridman (20:07.900)
but also they don't have the visual element.
François Chollet (20:10.740)
Like, I guess I'm comfortable with the structure.
Lex Fridman (20:13.460)
It feels like the narrowness,
François Chollet (20:15.740)
the constraints feel more comforting.
Lex Fridman (20:18.260)
If you have thousands of documents
François Chollet (20:20.300)
with your own thoughts in Google Docs,
Lex Fridman (20:23.100)
why don't you write some kind of search engine,
François Chollet (20:26.580)
like maybe a mind map, a piece of software,
Lex Fridman (20:30.900)
mind mapping software, where you write down a concept
Lex Fridman (20:33.980)
and then it gives you sentences or paragraphs
Lex Fridman (20:37.500)
from your thousand Google Docs document
François Chollet (20:39.700)
that match this concept.
Lex Fridman (20:41.220)
The problem is it's so deeply, unlike mind maps,
François Chollet (20:45.300)
it's so deeply rooted in natural language.
Lex Fridman (20:48.460)
So it's not, it's not semantically searchable,
François Chollet (20:54.420)
I would say, because the categories are very,
Lex Fridman (20:57.220)
you kind of mentioned intelligence, language, and motion.
François Chollet (21:00.700)
They're very strong, semantic.
Lex Fridman (21:02.580)
Like, it feels like the mind map forces you
François Chollet (21:05.020)
to be semantically clear and specific.
Lex Fridman (21:09.780)
The bullet points list I have are sparse,
François Chollet (21:13.860)
desperate thoughts that poetically represent
Lex Fridman (21:20.340)
a category like motion, as opposed to saying motion.
Lex Fridman (21:25.260)
So unfortunately, that's the same problem with the internet.
Lex Fridman (21:28.980)
That's why the idea of semantic web is difficult to get.
François Chollet (21:32.340)
It's, most language on the internet is a giant mess
Lex Fridman (21:37.980)
of natural language that's hard to interpret, which,
Lex Fridman (21:42.500)
so do you think there's something to mind maps as,
Lex Fridman (21:46.180)
you actually originally brought it up
François Chollet (21:48.100)
as we were talking about kind of cognition and language.
Lex Fridman (21:53.580)
Do you think there's something to mind maps
François Chollet (21:55.300)
about how our brain actually deals,
Lex Fridman (21:58.100)
like think reasons about things?
François Chollet (22:01.740)
It's possible.
Lex Fridman (22:02.580)
I think it's reasonable to assume that there is
François Chollet (22:07.100)
some level of topological processing in the brain,
Lex Fridman (22:10.620)
that the brain is very associative in nature.
Lex Fridman (22:15.140)
And I also believe that a topological space
Lex Fridman (22:20.660)
is a better medium to encode thoughts
François Chollet (22:25.420)
than a geometric space.
Lex Fridman (22:27.540)
So I think...
François Chollet (22:28.380)
What's the difference in a topological
Lex Fridman (22:29.740)
and a geometric space?
François Chollet (22:31.060)
Well, if you're talking about topologies,
Lex Fridman (22:34.100)
then points are either connected or not.
Lex Fridman (22:36.220)
So a topology is more like a subway map.
Lex Fridman (22:38.660)
And geometry is when you're interested
François Chollet (22:41.660)
in the distance between things.
Lex Fridman (22:43.900)
And in a subway map,
François Chollet (22:44.740)
you don't really have the concept of distance.
Lex Fridman (22:46.340)
You only have the concept of whether there is a train
François Chollet (22:48.420)
going from station A to station B.
Lex Fridman (22:52.820)
And what we do in deep learning is that we're actually
François Chollet (22:55.620)
dealing with geometric spaces.
Lex Fridman (22:57.740)
We are dealing with concept vectors, word vectors,
François Chollet (23:01.540)
that have a distance between them
Lex Fridman (23:03.300)
to express in terms of that product.
Lex Fridman (23:05.340)
So we are not really building topological models usually.
Lex Fridman (23:10.780)
I think you're absolutely right.
François Chollet (23:11.820)
Like distance is a fundamental importance in deep learning.
Lex Fridman (23:16.540)
I mean, it's the continuous aspect of it.
François Chollet (23:19.300)
Yes, because everything is a vector
Lex Fridman (23:21.180)
and everything has to be a vector
François Chollet (23:22.500)
because everything has to be differentiable.
Lex Fridman (23:24.500)
If your space is discrete, it's no longer differentiable.
François Chollet (23:26.860)
You cannot do deep learning in it anymore.
Lex Fridman (23:29.660)
Well, you could, but you can only do it by embedding it
François Chollet (23:32.420)
in a bigger continuous space.
Lex Fridman (23:35.620)
So if you do topology in the context of deep learning,
François Chollet (23:39.380)
you have to do it by embedding your topology
Lex Fridman (23:41.100)
in the geometry.
François Chollet (23:42.820)
Well, let me zoom out for a second.
Lex Fridman (23:46.220)
Let's get into your paper on the measure of intelligence
François Chollet (23:50.180)
that you put out in 2019.
Lex Fridman (23:52.860)
Yes.
François Chollet (23:53.700)
Okay.
Lex Fridman (23:54.540)
November.
François Chollet (23:55.380)
November.
Lex Fridman (23:57.700)
Yeah, remember 2019?
François Chollet (23:59.420)
That was a different time.
Lex Fridman (24:01.100)
Yeah, I remember.
François Chollet (24:02.780)
I still remember.
Lex Fridman (24:06.500)
It feels like a different world.
François Chollet (24:09.620)
You could travel, you could actually go outside
Lex Fridman (24:12.620)
and see friends.
François Chollet (24:15.100)
Yeah.
Lex Fridman (24:16.260)
Let me ask the most absurd question.
François Chollet (24:18.940)
I think there's some nonzero probability
Lex Fridman (24:21.740)
there'll be a textbook one day, like 200 years from now
François Chollet (24:25.220)
on artificial intelligence,
Lex Fridman (24:27.740)
or it'll be called like just intelligence
François Chollet (24:30.660)
cause humans will already be gone.
Lex Fridman (24:32.460)
It'll be your picture with a quote.
François Chollet (24:35.220)
This is, you know, one of the early biological systems
Lex Fridman (24:39.020)
would consider the nature of intelligence
Lex Fridman (24:41.580)
and there'll be like a definition
Lex Fridman (24:43.180)
of how they thought about intelligence.
François Chollet (24:45.180)
Which is one of the things you do in your paper
Lex Fridman (24:46.860)
on measure intelligence is to ask like,
François Chollet (24:51.060)
well, what is intelligence
Lex Fridman (24:52.620)
and how to test for intelligence and so on.
Lex Fridman (24:55.540)
So is there a spiffy quote about what is intelligence?
Lex Fridman (25:01.860)
What is the definition of intelligence
Lex Fridman (25:03.900)
according to Francois Chollet?
Lex Fridman (25:06.740)
Yeah, so do you think the super intelligent AIs
François Chollet (25:10.740)
of the future will want to remember us
Lex Fridman (25:13.900)
the way we remember humans from the past?
Lex Fridman (25:16.060)
And do you think they will be, you know,
Lex Fridman (25:18.500)
they won't be ashamed of having a biological origin?
François Chollet (25:22.340)
No, I think it would be a niche topic.
Lex Fridman (25:24.660)
It won't be that interesting,
Lex Fridman (25:25.820)
but it'll be like the people that study
Lex Fridman (25:29.420)
in certain contexts like historical civilization
François Chollet (25:33.100)
that no longer exists, the Aztecs and so on.
Lex Fridman (25:36.340)
That's how it'll be seen.
Lex Fridman (25:38.260)
And it'll be study in also the context on social media.
Lex Fridman (25:42.340)
There'll be hashtags about the atrocity
François Chollet (25:46.700)
committed to human beings
Lex Fridman (25:49.340)
when the robots finally got rid of them.
François Chollet (25:52.500)
Like it was a mistake.
Lex Fridman (25:55.180)
You'll be seen as a giant mistake,
Lex Fridman (25:57.020)
but ultimately in the name of progress
Lex Fridman (26:00.060)
and it created a better world
François Chollet (26:01.540)
because humans were over consuming the resources
Lex Fridman (26:05.220)
and they were not very rational
Lex Fridman (26:07.260)
and were destructive in the end in terms of productivity
Lex Fridman (26:11.060)
and putting more love in the world.
Lex Fridman (26:13.820)
And so within that context,
Lex Fridman (26:15.300)
there'll be a chapter about these biological systems.
François Chollet (26:17.420)
It seems to have a very detailed vision of that hit here.
Lex Fridman (26:20.380)
You should write a sci fi novel about it.
François Chollet (26:22.340)
I'm working on a sci fi novel currently, yes.
Lex Fridman (26:28.100)
Self published, yeah.
François Chollet (26:29.460)
The definition of intelligence.
Lex Fridman (26:30.740)
So intelligence is the efficiency
François Chollet (26:34.660)
with which you acquire new skills at tasks
Lex Fridman (26:39.380)
that you did not previously know about,
Lex Fridman (26:41.940)
that you did not prepare for, right?
Lex Fridman (26:44.700)
So intelligence is not skill itself.
François Chollet (26:47.780)
It's not what you know, it's not what you can do.
Lex Fridman (26:50.740)
It's how well and how efficiently
François Chollet (26:52.900)
you can learn new things.
Lex Fridman (26:54.580)
New things.
François Chollet (26:55.580)
Yes.
Lex Fridman (26:56.420)
The idea of newness there
François Chollet (26:58.100)
seems to be fundamentally important.
Lex Fridman (27:01.180)
Yes.
Lex Fridman (27:02.020)
So you would see intelligence on display, for instance.
Lex Fridman (27:05.780)
Whenever you see a human being or an AI creature
François Chollet (27:09.980)
adapt to a new environment that it does not see before,
Lex Fridman (27:13.900)
that its creators did not anticipate.
François Chollet (27:16.620)
When you see adaptation, when you see improvisation,
Lex Fridman (27:19.340)
when you see generalization, that's intelligence.
François Chollet (27:22.500)
In reverse, if you have a system
Lex Fridman (27:24.460)
that when you put it in a slightly new environment,
François Chollet (27:27.100)
it cannot adapt, it cannot improvise,
Lex Fridman (27:30.060)
it cannot deviate from what it's hard coded to do
François Chollet (27:33.380)
or what it has been trained to do,
Lex Fridman (27:38.700)
that is a system that is not intelligent.
François Chollet (27:41.060)
There's actually a quote from Einstein
Lex Fridman (27:43.580)
that captures this idea, which is,
François Chollet (27:46.780)
the measure of intelligence is the ability to change.
Lex Fridman (27:50.740)
I like that quote.
François Chollet (27:51.740)
I think it captures at least part of this idea.
Lex Fridman (27:54.940)
You know, there might be something interesting
François Chollet (27:56.460)
about the difference between your definition and Einstein's.
Lex Fridman (27:59.500)
I mean, he's just being Einstein and clever,
Lex Fridman (28:04.740)
but acquisition of new ability to deal with new things
Lex Fridman (28:09.740)
versus ability to just change.
Lex Fridman (28:14.100)
What's the difference between those two things?
Lex Fridman (28:16.820)
So just change in itself.
Lex Fridman (28:19.260)
Do you think there's something to that?
Lex Fridman (28:21.300)
Just being able to change.
François Chollet (28:23.780)
Yes, being able to adapt.
Lex Fridman (28:25.540)
So not change, but certainly change its direction.
François Chollet (28:30.060)
Being able to adapt yourself to your environment.
Lex Fridman (28:34.420)
Whatever the environment is.
François Chollet (28:35.660)
That's a big part of intelligence.
Lex Fridman (28:37.460)
And intelligence is more precisely, you know,
Lex Fridman (28:40.020)
how efficiently you're able to adapt,
Lex Fridman (28:42.460)
how efficiently you're able to basically master your environment,
Lex Fridman (28:45.740)
how efficiently you can acquire new skills.
Lex Fridman (28:49.140)
And I think there's a big distinction to be drawn
François Chollet (28:52.300)
between intelligence, which is a process,
Lex Fridman (28:56.220)
and the output of that process, which is skill.
Lex Fridman (29:01.420)
So for instance, if you have a very smart human brain,
Lex Fridman (29:04.900)
so for instance, if you have a very smart human programmer
François Chollet (29:08.980)
that considers the game of chess,
Lex Fridman (29:10.780)
and that writes down a static program that can play chess,
François Chollet (29:16.180)
then the intelligence is the process
Lex Fridman (29:19.140)
of developing that program.
Lex Fridman (29:20.660)
But the program itself is just encoding
Lex Fridman (29:25.660)
the output artifact of that process.
François Chollet (29:28.100)
The program itself is not intelligent.
Lex Fridman (29:30.020)
And the way you tell it's not intelligent
François Chollet (29:31.860)
is that if you put it in a different context,
Lex Fridman (29:34.020)
you ask it to play Go or something,
François Chollet (29:36.060)
it's not going to be able to perform well
Lex Fridman (29:37.780)
without human involvement,
François Chollet (29:38.900)
because the source of intelligence,
Lex Fridman (29:41.100)
the entity that is capable of that process
François Chollet (29:43.140)
is the human programmer.
Lex Fridman (29:44.380)
So we should be able to tell the difference
François Chollet (29:47.940)
between the process and its output.
Lex Fridman (29:50.100)
We should not confuse the output and the process.
François Chollet (29:53.260)
It's the same as, you know,
Lex Fridman (29:54.860)
do not confuse a road building company
Lex Fridman (29:58.780)
and one specific road,
Lex Fridman (2:00:00.380)
for the specific system you're considering.
Lex Fridman (2:00:02.980)
So I think a good test of intelligence
Lex Fridman (2:00:05.060)
should actually deal with developer aware generalization,
François Chollet (2:00:09.900)
which is slightly stronger than system centric generalization.
Lex Fridman (2:00:13.500)
So developer aware generalization
François Chollet (2:00:16.020)
would be the ability to generalize
Lex Fridman (2:00:19.860)
to novelty or uncertainty that not only the system itself has
François Chollet (2:00:24.220)
not access to, but the developer of the system
Lex Fridman (2:00:26.660)
could not have access to either.
François Chollet (2:00:29.380)
That's a fascinating meta definition.
Lex Fridman (2:00:32.380)
So the system is basically the edge case thing
François Chollet (2:00:37.700)
we're talking about with autonomous vehicles.
Lex Fridman (2:00:39.780)
Neither the developer nor the system
François Chollet (2:00:41.620)
know about the edge cases in my encounter.
Lex Fridman (2:00:44.420)
So it's up to the system should be
François Chollet (2:00:47.020)
able to generalize the thing that nobody expected,
Lex Fridman (2:00:51.660)
neither the designer of the training data,
François Chollet (2:00:54.860)
nor obviously the contents of the training data.
Lex Fridman (2:00:59.060)
That's a fascinating definition.
Lex Fridman (2:01:00.580)
So you can see degrees of generalization as a spectrum.
Lex Fridman (2:01:04.540)
And the lowest level is what machine learning
François Chollet (2:01:08.060)
is trying to do is the assumption
Lex Fridman (2:01:10.780)
that any new situation is going to be sampled
François Chollet (2:01:15.220)
from a static distribution of possible situations
Lex Fridman (2:01:18.340)
and that you already have a representative sample
François Chollet (2:01:21.500)
of the distribution.
Lex Fridman (2:01:22.420)
That's your training data.
Lex Fridman (2:01:23.860)
And so in machine learning, you generalize to a new sample
Lex Fridman (2:01:26.700)
from a known distribution.
Lex Fridman (2:01:28.780)
And the ways in which your new sample will be new or different
Lex Fridman (2:01:34.020)
are ways that are already understood by the developers
François Chollet (2:01:38.140)
of the system.
Lex Fridman (2:01:39.420)
So you are generalizing to known unknowns
François Chollet (2:01:43.020)
for one specific task.
Lex Fridman (2:01:45.100)
That's what you would call robustness.
François Chollet (2:01:47.500)
You are robust to things like noise, small variations,
Lex Fridman (2:01:50.180)
and so on for one fixed known distribution
François Chollet (2:01:56.620)
that you know through your training data.
Lex Fridman (2:01:59.300)
And the higher degree would be flexibility
François Chollet (2:02:05.060)
in machine intelligence.
Lex Fridman (2:02:06.380)
So flexibility would be something
François Chollet (2:02:08.620)
like an L5 cell driving car or maybe a robot that
Lex Fridman (2:02:12.500)
can pass the coffee cup test, which
François Chollet (2:02:16.820)
is the notion that you'd be given a random kitchen
Lex Fridman (2:02:21.460)
somewhere in the country.
Lex Fridman (2:02:22.460)
And you would have to go make a cup of coffee in that kitchen.
Lex Fridman (2:02:28.460)
So flexibility would be the ability
François Chollet (2:02:30.820)
to deal with unknown unknowns, so things that could not,
Lex Fridman (2:02:35.300)
dimensions of viability that could not
François Chollet (2:02:37.180)
have been possibly foreseen by the creators of the system
Lex Fridman (2:02:41.100)
within one specific task.
Lex Fridman (2:02:42.860)
So generalizing to the long tail of situations in self driving,
Lex Fridman (2:02:47.020)
for instance, would be flexibility.
Lex Fridman (2:02:48.540)
So you have robustness, flexibility, and finally,
Lex Fridman (2:02:51.700)
you would have extreme generalization,
François Chollet (2:02:53.700)
which is basically flexibility, but instead
Lex Fridman (2:02:57.740)
of just considering one specific domain,
François Chollet (2:03:01.180)
like driving or domestic robotics,
Lex Fridman (2:03:03.340)
you're considering an open ended range of possible domains.
Lex Fridman (2:03:07.740)
So a robot would be capable of extreme generalization
Lex Fridman (2:03:12.620)
if, let's say, it's designed and trained for cooking,
François Chollet (2:03:18.060)
for instance.
Lex Fridman (2:03:19.820)
And if I buy the robot and if it's
François Chollet (2:03:24.580)
able to teach itself gardening in a couple of weeks,
Lex Fridman (2:03:28.780)
it would be capable of extreme generalization, for instance.
Lex Fridman (2:03:32.300)
So the ultimate goal is extreme generalization.
Lex Fridman (2:03:34.300)
Yes.
Lex Fridman (2:03:34.820)
So creating a system that is so general that it could
Lex Fridman (2:03:40.020)
essentially achieve human skill parity over arbitrary tasks
Lex Fridman (2:03:46.140)
and arbitrary domains with the same level of improvisation
Lex Fridman (2:03:50.820)
and adaptation power as humans when
François Chollet (2:03:53.740)
it encounters new situations.
Lex Fridman (2:03:55.380)
And it would do so over basically the same range
François Chollet (2:03:59.780)
of possible domains and tasks as humans
Lex Fridman (2:04:02.780)
and using essentially the same amount of training
François Chollet (2:04:05.500)
experience of practice as humans would require.
Lex Fridman (2:04:07.860)
That would be human level extreme generalization.
Lex Fridman (2:04:10.900)
So I don't actually think humans are anywhere
Lex Fridman (2:04:14.620)
near the optimal intelligence bounds
François Chollet (2:04:19.580)
if there is such a thing.
Lex Fridman (2:04:21.300)
So I think for humans or in general?
François Chollet (2:04:23.820)
In general.
Lex Fridman (2:04:25.140)
I think it's quite likely that there
François Chollet (2:04:26.780)
is a hard limit to how intelligent any system can be.
Lex Fridman (2:04:33.860)
But at the same time, I don't think humans are anywhere
François Chollet (2:04:35.980)
near that limit.
Lex Fridman (2:04:39.180)
Yeah, last time I think we talked,
François Chollet (2:04:40.780)
I think you had this idea that we're only
Lex Fridman (2:04:43.820)
as intelligent as the problems we face.
François Chollet (2:04:46.580)
Sort of we are bounded by the problems.
Lex Fridman (2:04:51.300)
In a way, yes.
François Chollet (2:04:51.940)
We are bounded by our environments,
Lex Fridman (2:04:55.100)
and we are bounded by the problems we try to solve.
François Chollet (2:04:58.100)
Yeah.
Lex Fridman (2:04:59.220)
Yeah.
Lex Fridman (2:04:59.700)
What do you make of Neuralink and outsourcing
Lex Fridman (2:05:03.820)
some of the brain power, like brain computer interfaces?
Lex Fridman (2:05:07.140)
Do you think we can expand or augment our intelligence?
Lex Fridman (2:05:13.460)
I am fairly skeptical of neural interfaces
François Chollet (2:05:18.340)
because they are trying to fix one specific bottleneck
Lex Fridman (2:05:23.780)
in human machine cognition, which
François Chollet (2:05:26.380)
is the bandwidth bottleneck, input and output
Lex Fridman (2:05:29.700)
of information in the brain.
Lex Fridman (2:05:31.820)
And my perception of the problem is that bandwidth is not
Lex Fridman (2:05:37.820)
at this time a bottleneck at all.
François Chollet (2:05:41.140)
Meaning that we already have sensors
Lex Fridman (2:05:43.580)
that enable us to take in far more information than what
François Chollet (2:05:48.300)
we can actually process.
Lex Fridman (2:05:50.420)
Well, to push back on that a little bit,
François Chollet (2:05:53.260)
to sort of play devil's advocate a little bit,
Lex Fridman (2:05:55.420)
is if you look at the internet, Wikipedia, let's say Wikipedia,
François Chollet (2:05:58.980)
I would say that humans, after the advent of Wikipedia,
Lex Fridman (2:06:03.300)
are much more intelligent.
François Chollet (2:06:05.860)
Yes, I think that's a good one.
Lex Fridman (2:06:07.820)
But that's also not about, that's about externalizing
François Chollet (2:06:14.180)
our intelligence via information processing systems,
Lex Fridman (2:06:18.140)
external information processing systems,
François Chollet (2:06:19.740)
which is very different from brain computer interfaces.
Lex Fridman (2:06:23.780)
Right, but the question is whether if we have direct
François Chollet (2:06:27.980)
access, if our brain has direct access to Wikipedia without
Lex Fridman (2:06:31.940)
Your brain already has direct access to Wikipedia.
François Chollet (2:06:34.540)
It's on your phone.
Lex Fridman (2:06:35.900)
And you have your hands and your eyes and your ears
Lex Fridman (2:06:39.380)
and so on to access that information.
Lex Fridman (2:06:42.140)
And the speed at which you can access it
François Chollet (2:06:44.340)
Is bottlenecked by the cognition.
Lex Fridman (2:06:45.700)
I think it's already close, fairly close to optimal,
François Chollet (2:06:49.620)
which is why speed reading, for instance, does not work.
Lex Fridman (2:06:53.340)
The faster you read, the less you understand.
Lex Fridman (2:06:55.980)
But maybe it's because it uses the eyes.
Lex Fridman (2:06:58.420)
So maybe.
Lex Fridman (2:07:00.540)
So I don't believe so.
Lex Fridman (2:07:01.460)
I think the brain is very slow.
François Chollet (2:07:04.620)
It typically operates, you know, the fastest things
Lex Fridman (2:07:07.860)
that happen in the brain are at the level of 50 milliseconds.
François Chollet (2:07:11.420)
Forming a conscious thought can potentially
Lex Fridman (2:07:14.580)
take entire seconds, right?
Lex Fridman (2:07:16.740)
And you can already read pretty fast.
Lex Fridman (2:07:19.220)
So I think the speed at which you can take information in
Lex Fridman (2:07:23.460)
and even the speed at which you can output information
Lex Fridman (2:07:26.460)
can only be very incrementally improved.
François Chollet (2:07:29.900)
Maybe there's a question.
Lex Fridman (2:07:31.100)
If you're a very fast typer, if you're a very trained typer,
François Chollet (2:07:34.380)
the speed at which you can express your thoughts
Lex Fridman (2:07:36.660)
is already the speed at which you can form your thoughts.
François Chollet (2:07:40.500)
Right, so that's kind of an idea that there are
Lex Fridman (2:07:44.540)
fundamental bottlenecks to the human mind.
Lex Fridman (2:07:47.020)
But it's possible that everything we have
Lex Fridman (2:07:50.260)
in the human mind is just to be able to survive
François Chollet (2:07:53.140)
in the environment.
Lex Fridman (2:07:54.420)
And there's a lot more to expand.
François Chollet (2:07:58.300)
Maybe, you know, you said the speed of the thought.
Lex Fridman (2:08:02.420)
So I think augmenting human intelligence
Lex Fridman (2:08:06.780)
is a very valid and very powerful avenue, right?
Lex Fridman (2:08:09.900)
And that's what computers are about.
François Chollet (2:08:12.260)
In fact, that's what all of culture and civilization
Lex Fridman (2:08:15.900)
is about.
François Chollet (2:08:16.740)
Our culture is externalized cognition
Lex Fridman (2:08:20.620)
and we rely on culture to think constantly.
François Chollet (2:08:23.740)
Yeah, I mean, that's another, yeah.
Lex Fridman (2:08:26.620)
Not just computers, not just phones and the internet.
François Chollet (2:08:29.140)
I mean, all of culture, like language, for instance,
Lex Fridman (2:08:32.460)
is a form of externalized cognition.
François Chollet (2:08:34.020)
Books are obviously externalized cognition.
Lex Fridman (2:08:37.460)
Yeah, that's a good point.
Lex Fridman (2:08:38.580)
And you can scale that externalized cognition
Lex Fridman (2:08:42.060)
far beyond the capability of the human brain.
Lex Fridman (2:08:45.180)
And you could see civilization itself
Lex Fridman (2:08:48.900)
is it has capabilities that are far beyond any individual brain
Lex Fridman (2:08:54.260)
and will keep scaling it because it's not
Lex Fridman (2:08:55.940)
rebound by individual brains.
François Chollet (2:08:59.140)
It's a different kind of system.
Lex Fridman (2:09:01.340)
Yeah, and that system includes nonhuman, nonhumans.
François Chollet (2:09:06.260)
First of all, it includes all the other biological systems,
Lex Fridman (2:09:08.700)
which are probably contributing to the overall intelligence
François Chollet (2:09:11.660)
of the organism.
Lex Fridman (2:09:12.900)
And then computers are part of it.
François Chollet (2:09:14.460)
Nonhuman systems are probably not contributing much,
Lex Fridman (2:09:16.860)
but AIs are definitely contributing to that.
François Chollet (2:09:19.700)
Like Google search, for instance, is a big part of it.
Lex Fridman (2:09:24.260)
Yeah, yeah, a huge part, a part that we can't probably
François Chollet (2:09:29.660)
introspect.
Lex Fridman (2:09:31.060)
Like how the world has changed in the past 20 years,
François Chollet (2:09:33.780)
it's probably very difficult for us
Lex Fridman (2:09:35.220)
to be able to understand until, of course,
François Chollet (2:09:38.620)
whoever created the simulation we're in is probably
Lex Fridman (2:09:41.740)
doing metrics, measuring the progress.
François Chollet (2:09:44.940)
There was probably a big spike in performance.
Lex Fridman (2:09:48.340)
They're enjoying this.
Lex Fridman (2:09:51.580)
So what are your thoughts on the Turing test
Lex Fridman (2:09:56.020)
and the Lobner Prize, which is one
François Chollet (2:10:00.340)
of the most famous attempts at the test of artificial
Lex Fridman (2:10:05.700)
intelligence by doing a natural language open dialogue test
Lex Fridman (2:10:11.740)
that's judged by humans as far as how well the machine did?
Lex Fridman (2:10:18.860)
So I'm not a fan of the Turing test.
François Chollet (2:10:21.460)
Itself or any of its variants for two reasons.
Lex Fridman (2:10:25.940)
So first of all, it's really coping out
François Chollet (2:10:34.140)
of trying to define and measure intelligence
Lex Fridman (2:10:37.660)
because it's entirely outsourcing that
François Chollet (2:10:40.620)
to a panel of human judges.
Lex Fridman (2:10:43.380)
And these human judges, they may not themselves
François Chollet (2:10:47.420)
have any proper methodology.
Lex Fridman (2:10:49.700)
They may not themselves have any proper definition
François Chollet (2:10:52.660)
of intelligence.
Lex Fridman (2:10:53.620)
They may not be reliable.
Lex Fridman (2:10:54.780)
So the Turing test is already failing
Lex Fridman (2:10:57.260)
one of the core psychometrics principles, which
François Chollet (2:10:59.620)
is reliability because you have biased human judges.
Lex Fridman (2:11:04.620)
It's also violating the standardization requirement
Lex Fridman (2:11:07.900)
and the freedom from bias requirement.
Lex Fridman (2:11:10.140)
And so it's really a cope out because you are outsourcing
François Chollet (2:11:13.900)
everything that matters, which is precisely describing
Lex Fridman (2:11:17.380)
intelligence and finding a standalone test to measure it.
François Chollet (2:11:22.180)
You're outsourcing everything to people.
Lex Fridman (2:11:25.260)
So it's really a cope out.
Lex Fridman (2:11:26.340)
And by the way, we should keep in mind
Lex Fridman (2:11:28.860)
that when Turing proposed the imitation game,
François Chollet (2:11:33.940)
it was not meaning for the imitation game
Lex Fridman (2:11:36.780)
to be an actual goal for the field of AI
Lex Fridman (2:11:40.700)
and actual test of intelligence.
Lex Fridman (2:11:42.460)
It was using the imitation game as a thought experiment
François Chollet (2:11:48.780)
in a philosophical discussion in his 1950 paper.
Lex Fridman (2:11:53.580)
He was trying to argue that theoretically, it
François Chollet (2:11:58.820)
should be possible for something very much like the human mind,
Lex Fridman (2:12:04.220)
indistinguishable from the human mind,
François Chollet (2:12:06.100)
to be encoded in a Turing machine.
Lex Fridman (2:12:08.060)
And at the time, that was a very daring idea.
François Chollet (2:12:14.540)
It was stretching credulity.
Lex Fridman (2:12:16.580)
But nowadays, I think it's fairly well accepted
François Chollet (2:12:20.140)
that the mind is an information processing system
Lex Fridman (2:12:22.660)
and that you could probably encode it into a computer.
Lex Fridman (2:12:25.420)
So another reason why I'm not a fan of this type of test
Lex Fridman (2:12:29.380)
is that the incentives that it creates
François Chollet (2:12:34.220)
are incentives that are not conducive to proper scientific
Lex Fridman (2:12:39.740)
research.
François Chollet (2:12:40.780)
If your goal is to trick, to convince a panel of human
Lex Fridman (2:12:45.700)
judges that they are talking to a human,
François Chollet (2:12:48.460)
then you have an incentive to rely on tricks
Lex Fridman (2:12:53.420)
and prestidigitation.
François Chollet (2:12:56.500)
In the same way that, let's say, you're doing physics
Lex Fridman (2:12:59.180)
and you want to solve teleportation.
Lex Fridman (2:13:01.500)
And what if the test that you set out to pass
Lex Fridman (2:13:04.660)
is you need to convince a panel of judges
Lex Fridman (2:13:07.460)
that teleportation took place?
Lex Fridman (2:13:09.500)
And they're just sitting there and watching what you're doing.
Lex Fridman (2:13:12.580)
And that is something that you can achieve with David
Lex Fridman (2:13:17.540)
Copperfield could achieve it in his show at Vegas.
Lex Fridman (2:13:22.780)
And what he's doing is very elaborate.
Lex Fridman (2:13:25.260)
But it's not physics.
François Chollet (2:13:29.180)
It's not making any progress in our understanding
Lex Fridman (2:13:31.740)
of the universe.
François Chollet (2:13:32.620)
To push back on that is possible.
Lex Fridman (2:13:34.780)
That's the hope with these kinds of subjective evaluations
François Chollet (2:13:39.020)
is that it's easier to solve it generally
Lex Fridman (2:13:41.940)
than it is to come up with tricks that convince
François Chollet (2:13:45.420)
a large number of judges.
Lex Fridman (2:13:46.620)
That's the hope.
François Chollet (2:13:47.340)
In practice, it turns out that it's
Lex Fridman (2:13:49.300)
very easy to deceive people in the same way
François Chollet (2:13:51.500)
that you can do magic in Vegas.
Lex Fridman (2:13:54.380)
You can actually very easily convince people
François Chollet (2:13:57.300)
that they're talking to a human when they're actually
Lex Fridman (2:13:59.500)
talking to an algorithm.
François Chollet (2:14:00.740)
I just disagree.
Lex Fridman (2:14:01.740)
I disagree with that.
François Chollet (2:14:02.660)
I think it's easy.
Lex Fridman (2:14:03.620)
I would push.
François Chollet (2:14:05.100)
No, it's not easy.
Lex Fridman (2:14:07.340)
It's doable.
François Chollet (2:14:08.300)
It's very easy because we are biased.
Lex Fridman (2:14:12.260)
We have theory of mind.
François Chollet (2:14:13.860)
We are constantly projecting emotions, intentions, agentness.
Lex Fridman (2:14:21.020)
Agentness is one of our core innate priors.
François Chollet (2:14:24.260)
We are projecting these things on everything around us.
Lex Fridman (2:14:26.820)
Like if you paint a smiley on a rock,
François Chollet (2:14:31.260)
the rock becomes happy in our eyes.
Lex Fridman (2:14:33.420)
And because we have this extreme bias that
François Chollet (2:14:36.540)
permits everything we see around us,
Lex Fridman (2:14:39.740)
it's actually pretty easy to trick people.
François Chollet (2:14:41.780)
I just disagree with that.
Lex Fridman (2:14:44.300)
I so totally disagree with that.
François Chollet (2:14:45.820)
You brilliantly put as a huge, the anthropomorphization
Lex Fridman (2:14:50.500)
that we naturally do, the agentness of that word.
Lex Fridman (2:14:53.140)
Is that a real word?
Lex Fridman (2:14:53.980)
No, it's not a real word.
François Chollet (2:14:55.500)
I like it.
Lex Fridman (2:14:56.020)
But it's a useful word.
François Chollet (2:14:57.780)
It's a useful word.
Lex Fridman (2:14:58.620)
Let's make it real.
François Chollet (2:14:59.660)
It's a huge help.
Lex Fridman (2:15:01.020)
But I still think it's really difficult to convince.
François Chollet (2:15:04.900)
If you do like the Alexa Prize formulation,
Lex Fridman (2:15:07.940)
where you talk for an hour, there's
François Chollet (2:15:10.420)
formulations of the test you can create,
Lex Fridman (2:15:12.460)
where it's very difficult.
Lex Fridman (2:15:13.780)
So I like the Alexa Prize better because it's more pragmatic.
Lex Fridman (2:15:18.100)
It's more practical.
François Chollet (2:15:19.540)
It's actually incentivizing developers
Lex Fridman (2:15:22.100)
to create something that's useful as a human machine
François Chollet (2:15:27.860)
interface.
Lex Fridman (2:15:29.300)
So that's slightly better than just the imitation.
Lex Fridman (2:15:31.780)
So I like it.
Lex Fridman (2:15:34.100)
Your idea is like a test which hopefully
François Chollet (2:15:36.980)
help us in creating intelligent systems as a result.
Lex Fridman (2:15:39.620)
Like if you create a system that passes it,
François Chollet (2:15:41.700)
it'll be useful for creating further intelligent systems.
Lex Fridman (2:15:44.740)
Yes, at least.
François Chollet (2:15:46.100)
Yeah.
Lex Fridman (2:15:47.620)
Just to kind of comment, I'm a little bit surprised
Lex Fridman (2:15:51.740)
how little inspiration people draw from the Turing test
Lex Fridman (2:15:55.660)
today.
François Chollet (2:15:57.180)
The media and the popular press might write about it
Lex Fridman (2:15:59.420)
every once in a while.
François Chollet (2:16:00.900)
The philosophers might talk about it.
Lex Fridman (2:16:03.500)
But most engineers are not really inspired by it.
Lex Fridman (2:16:07.020)
And I know you don't like the Turing test,
Lex Fridman (2:16:11.340)
but we'll have this argument another time.
François Chollet (2:16:15.060)
There's something inspiring about it, I think.
Lex Fridman (2:16:18.620)
As a philosophical device in a physical discussion,
François Chollet (2:16:21.740)
I think there is something very interesting about it.
Lex Fridman (2:16:23.780)
I don't think it is in practical terms.
François Chollet (2:16:26.220)
I don't think it's conducive to progress.
Lex Fridman (2:16:29.060)
And one of the reasons why is that I
François Chollet (2:16:32.540)
think being very human like, being
Lex Fridman (2:16:35.300)
indistinguishable from a human is actually
François Chollet (2:16:37.540)
the very last step in the creation of machine
Lex Fridman (2:16:40.460)
intelligence.
François Chollet (2:16:41.020)
That the first ARs that will show strong generalization
Lex Fridman (2:16:46.820)
that will actually implement human like broad cognitive
François Chollet (2:16:52.500)
abilities, they will not actually behave or look
Lex Fridman (2:16:54.980)
anything like humans.
François Chollet (2:16:58.500)
Human likeness is the very last step in that process.
Lex Fridman (2:17:01.700)
And so a good test is a test that
François Chollet (2:17:03.780)
points you towards the first step on the ladder,
Lex Fridman (2:17:07.060)
not towards the top of the ladder.
Lex Fridman (2:17:08.900)
So to push back on that, I usually
Lex Fridman (2:17:11.980)
agree with you on most things.
François Chollet (2:17:13.460)
I remember you, I think at some point,
Lex Fridman (2:17:15.060)
tweeting something about the Turing test
François Chollet (2:17:17.100)
not being being counterproductive
Lex Fridman (2:17:19.020)
or something like that.
Lex Fridman (2:17:20.340)
And I think a lot of very smart people agree with that.
Lex Fridman (2:17:23.220)
I, a computation speaking, not very smart person,
François Chollet (2:17:31.460)
disagree with that.
Lex Fridman (2:17:32.300)
Because I think there's some magic
François Chollet (2:17:33.820)
to the interactivity with other humans.
Lex Fridman (2:17:36.900)
So to play devil's advocate on your statement,
François Chollet (2:17:39.620)
it's possible that in order to demonstrate
Lex Fridman (2:17:42.780)
the generalization abilities of a system,
François Chollet (2:17:45.540)
you have to show your ability, in conversation,
Lex Fridman (2:17:49.940)
show your ability to adjust, adapt to the conversation
François Chollet (2:17:55.380)
through not just like as a standalone system,
Lex Fridman (2:17:58.380)
but through the process of like the interaction,
François Chollet (2:18:01.380)
the game theoretic, where you really
Lex Fridman (2:18:05.700)
are changing the environment by your actions.
Lex Fridman (2:18:09.180)
So in the ARC challenge, for example,
Lex Fridman (2:18:11.660)
you're an observer.
François Chollet (2:18:12.820)
You can't scare the test into changing.
Lex Fridman (2:18:17.460)
You can't talk to the test.
François Chollet (2:18:19.380)
You can't play with it.
Lex Fridman (2:18:21.260)
So there's some aspect of that interactivity
François Chollet (2:18:24.300)
that becomes highly subjective, but it
Lex Fridman (2:18:26.140)
feels like it could be conducive to generalizability.
François Chollet (2:18:29.620)
I think you make a great point.
Lex Fridman (2:18:31.060)
The interactivity is a very good setting
François Chollet (2:18:33.580)
to force a system to show adaptation,
Lex Fridman (2:18:36.060)
to show generalization.
François Chollet (2:18:39.300)
That said, at the same time, it's
Lex Fridman (2:18:42.620)
not something very scalable, because you
François Chollet (2:18:44.860)
rely on human judges.
Lex Fridman (2:18:46.100)
It's not something reliable, because the human judges may
François Chollet (2:18:48.700)
not, may not.
Lex Fridman (2:18:49.420)
So you don't like human judges.
François Chollet (2:18:50.940)
Basically, yes.
Lex Fridman (2:18:51.860)
And I think so.
François Chollet (2:18:52.540)
I love the idea of interactivity.
Lex Fridman (2:18:56.140)
I initially wanted an ARC test that
François Chollet (2:18:59.620)
had some amount of interactivity where your score on a task
Lex Fridman (2:19:02.820)
would not be 1 or 0, if you can solve it or not,
Lex Fridman (2:19:05.380)
but would be the number of attempts
Lex Fridman (2:19:11.580)
that you can make before you hit the right solution, which
François Chollet (2:19:14.740)
means that now you can start applying
Lex Fridman (2:19:16.900)
the scientific method as you solve ARC tasks,
François Chollet (2:19:19.860)
that you can start formulating hypotheses and probing
Lex Fridman (2:19:23.780)
the system to see whether the observation will
François Chollet (2:19:27.300)
match the hypothesis or not.
Lex Fridman (2:19:28.660)
It would be amazing if you could also,
François Chollet (2:19:30.700)
even higher level than that, measure the quality of your attempts,
Lex Fridman (2:19:35.500)
which, of course, is impossible.
Lex Fridman (2:19:36.780)
But again, that gets subjective.
Lex Fridman (2:19:38.540)
How good was your thinking?
Lex Fridman (2:19:41.620)
How efficient was?
Lex Fridman (2:19:43.900)
So one thing that's interesting about this notion of scoring you
François Chollet (2:19:48.380)
as how many attempts you need is that you
Lex Fridman (2:19:50.500)
can start producing tasks that are way more ambiguous, right?
François Chollet (2:19:55.220)
Right.
Lex Fridman (2:19:56.500)
Because with the different attempts,
Lex Fridman (2:19:59.700)
you can actually probe that ambiguity, right?
Lex Fridman (2:20:03.300)
Right.
Lex Fridman (2:20:04.140)
So that's, in a sense, which is how good can
Lex Fridman (2:20:08.220)
you adapt to the uncertainty and reduce the uncertainty?
François Chollet (2:20:15.700)
Yes, it's half fast.
Lex Fridman (2:20:18.260)
It's the efficiency with which you reduce uncertainty
François Chollet (2:20:21.180)
in program space, exactly.
Lex Fridman (2:20:22.940)
Very difficult to come up with that kind of test, though.
François Chollet (2:20:24.940)
Yeah, so I would love to be able to create something like this.
Lex Fridman (2:20:28.340)
In practice, it would be very, very difficult, but yes.
François Chollet (2:20:33.140)
I mean, what you're doing, what you've done with the ARC challenge
Lex Fridman (2:20:36.140)
is brilliant.
François Chollet (2:20:37.620)
I'm also not surprised that it's not more popular,
Lex Fridman (2:20:40.940)
but I think it's picking up.
François Chollet (2:20:42.140)
It does its niche.
Lex Fridman (2:20:42.860)
It does its niche, yeah.
François Chollet (2:20:44.100)
Yeah.
Lex Fridman (2:20:44.900)
What are your thoughts about another test?
François Chollet (2:20:47.100)
I talked with Marcus Hutter.
Lex Fridman (2:20:48.940)
He has the Hutter Prize for compression of human knowledge.
Lex Fridman (2:20:51.660)
And the idea is really sort of quantify and reduce
Lex Fridman (2:20:55.620)
the test of intelligence purely to just the ability
François Chollet (2:20:58.260)
to compress.
Lex Fridman (2:20:59.580)
What's your thoughts about this intelligence as compression?
François Chollet (2:21:04.660)
I mean, it's a very fun test because it's
Lex Fridman (2:21:07.980)
such a simple idea, like you're given Wikipedia,
François Chollet (2:21:12.220)
basic English Wikipedia, and you must compress it.
Lex Fridman (2:21:15.500)
And so it stems from the idea that cognition is compression,
François Chollet (2:21:21.140)
that the brain is basically a compression algorithm.
Lex Fridman (2:21:24.020)
This is a very old idea.
François Chollet (2:21:25.620)
It's a very, I think, striking and beautiful idea.
Lex Fridman (2:21:30.540)
I used to believe it.
François Chollet (2:21:32.740)
I eventually had to realize that it was very much
Lex Fridman (2:21:36.140)
a flawed idea.
Lex Fridman (2:21:36.900)
So I no longer believe that cognition is compression.
Lex Fridman (2:21:41.420)
But I can tell you what's the difference.
Lex Fridman (2:21:44.620)
So it's very easy to believe that cognition and compression
Lex Fridman (2:21:48.820)
are the same thing.
Lex Fridman (2:21:51.660)
So Jeff Hawkins, for instance, says
Lex Fridman (2:21:53.220)
that cognition is prediction.
Lex Fridman (2:21:54.780)
And of course, prediction is basically the same thing
Lex Fridman (2:21:57.740)
as compression.
François Chollet (2:21:58.700)
It's just including the temporal axis.
Lex Fridman (2:22:03.580)
And it's very easy to believe this
François Chollet (2:22:05.060)
because compression is something that we
Lex Fridman (2:22:06.900)
do all the time very naturally.
François Chollet (2:22:09.020)
We are constantly compressing information.
Lex Fridman (2:22:12.020)
We are constantly trying.
François Chollet (2:22:15.660)
We have this bias towards simplicity.
Lex Fridman (2:22:17.940)
We are constantly trying to organize things in our mind
Lex Fridman (2:22:21.060)
and around us to be more regular.
Lex Fridman (2:22:24.460)
So it's a beautiful idea.
François Chollet (2:22:26.860)
It's very easy to believe.
Lex Fridman (2:22:28.620)
There is a big difference between what
François Chollet (2:22:31.580)
we do with our brains and compression.
Lex Fridman (2:22:33.980)
So compression is actually kind of a tool
François Chollet (2:22:38.220)
in the human cognitive toolkit that is used in many ways.
Lex Fridman (2:22:42.060)
But it's just a tool.
François Chollet (2:22:44.540)
It is a tool for cognition.
Lex Fridman (2:22:45.940)
It is not cognition itself.
Lex Fridman (2:22:47.620)
And the big fundamental difference
Lex Fridman (2:22:50.020)
is that cognition is about being able to operate
François Chollet (2:22:55.340)
in future situations that include fundamental uncertainty
Lex Fridman (2:23:00.740)
and novelty.
Lex Fridman (2:23:02.140)
So for instance, consider a child at age 10.
Lex Fridman (2:23:06.860)
And so they have 10 years of life experience.
François Chollet (2:23:10.100)
They've gotten pain, pleasure, rewards, and punishment
Lex Fridman (2:23:14.260)
in a period of time.
François Chollet (2:23:16.500)
If you were to generate the shortest behavioral program
Lex Fridman (2:23:21.980)
that would have basically run that child over these 10 years
François Chollet (2:23:26.740)
in an optimal way, the shortest optimal behavioral program
Lex Fridman (2:23:32.220)
given the experience of that child so far,
François Chollet (2:23:34.820)
well, that program, that compressed program,
Lex Fridman (2:23:37.540)
this is what you would get if the mind of the child
François Chollet (2:23:39.940)
was a compression algorithm essentially,
Lex Fridman (2:23:42.740)
would be utterly unable, inappropriate,
François Chollet (2:23:48.100)
to process the next 70 years in the life of that child.
Lex Fridman (2:23:54.380)
So in the models we build of the world,
François Chollet (2:23:59.020)
we are not trying to make them actually optimally compressed.
Lex Fridman (2:24:03.220)
We are using compression as a tool
François Chollet (2:24:06.660)
to promote simplicity and efficiency in our models.
Lex Fridman (2:24:10.060)
But they are not perfectly compressed
François Chollet (2:24:12.060)
because they need to include things
Lex Fridman (2:24:15.300)
that are seemingly useless today, that have seemingly
François Chollet (2:24:18.540)
been useless so far.
Lex Fridman (2:24:20.140)
But that may turn out to be useful in the future
François Chollet (2:24:24.140)
because you just don't know the future.
Lex Fridman (2:24:25.900)
And that's the fundamental principle
François Chollet (2:24:28.740)
that cognition, that intelligence arises from
Lex Fridman (2:24:31.260)
is that you need to be able to run
François Chollet (2:24:33.780)
appropriate behavioral programs except you have absolutely
Lex Fridman (2:24:36.660)
no idea what sort of context, environment, situation
François Chollet (2:24:40.940)
they are going to be running in.
Lex Fridman (2:24:42.260)
And you have to deal with that uncertainty,
François Chollet (2:24:45.020)
with that future anomaly.
Lex Fridman (2:24:46.580)
So an analogy that you can make is with investing,
François Chollet (2:24:52.500)
for instance.
Lex Fridman (2:24:54.460)
If I look at the past 20 years of stock market data,
Lex Fridman (2:24:59.540)
and I use a compression algorithm
Lex Fridman (2:25:01.860)
to figure out the best trading strategy,
François Chollet (2:25:04.420)
it's going to be you buy Apple stock, then
Lex Fridman (2:25:06.660)
maybe the past few years you buy Tesla stock or something.
Lex Fridman (2:25:10.420)
But is that strategy still going to be
Lex Fridman (2:25:13.300)
true for the next 20 years?
François Chollet (2:25:14.660)
Well, actually, probably not, which
Lex Fridman (2:25:17.980)
is why if you're a smart investor,
François Chollet (2:25:21.060)
you're not just going to be following the strategy that
Lex Fridman (2:25:26.340)
corresponds to compression of the past.
François Chollet (2:25:28.980)
You're going to be following, you're
Lex Fridman (2:25:31.660)
going to have a balanced portfolio, right?
François Chollet (2:25:34.860)
Because you just don't know what's going to happen.
Lex Fridman (2:25:38.180)
I mean, I guess in that same sense,
François Chollet (2:25:40.460)
the compression is analogous to what
Lex Fridman (2:25:42.540)
you talked about, which is local or robust generalization
François Chollet (2:25:45.900)
versus extreme generalization.
Lex Fridman (2:25:47.820)
It's much closer to that side of being able to generalize
François Chollet (2:25:52.420)
in the local sense.
Lex Fridman (2:25:53.420)
That's why as humans, when we are children, in our education,
Lex Fridman (2:25:59.980)
so a lot of it is driven by play, driven by curiosity.
Lex Fridman (2:26:04.180)
We are not efficiently compressing things.
François Chollet (2:26:07.900)
We're actually exploring.
Lex Fridman (2:26:09.620)
We are retaining all kinds of things
François Chollet (2:26:16.620)
from our environment that seem to be completely useless.
Lex Fridman (2:26:19.660)
Because they might turn out to be eventually useful, right?
Lex Fridman (2:26:24.380)
And that's what cognition is really about.
Lex Fridman (2:26:26.940)
And what makes it antagonistic to compression
François Chollet (2:26:29.300)
is that it is about hedging for future uncertainty.
Lex Fridman (2:26:33.980)
And that's antagonistic to compression.
François Chollet (2:26:35.860)
Yes.
Lex Fridman (2:26:36.580)
Officially hedging.
François Chollet (2:26:38.500)
Cognition leverages compression as a tool
Lex Fridman (2:26:41.660)
to promote efficiency and simplicity in our models.
François Chollet (2:26:47.420)
It's like Einstein said, make it simpler, but not,
Lex Fridman (2:26:52.260)
however that quote goes, but not too simple.
Lex Fridman (2:26:54.940)
So compression simplifies things,
Lex Fridman (2:26:57.700)
but you don't want to make it too simple.
François Chollet (2:27:00.100)
Yes.
Lex Fridman (2:27:00.740)
So a good model of the world is going
François Chollet (2:27:03.100)
to include all kinds of things that are completely useless,
Lex Fridman (2:27:06.100)
actually, just in case.
François Chollet (2:27:08.500)
Because you need diversity in the same way
Lex Fridman (2:27:10.020)
that in your portfolio.
François Chollet (2:27:11.140)
You need all kinds of stocks that may not
Lex Fridman (2:27:13.340)
have performed well so far, but you need diversity.
Lex Fridman (2:27:15.580)
And the reason you need diversity
Lex Fridman (2:27:16.980)
is because fundamentally you don't know what you're doing.
Lex Fridman (2:27:19.660)
And the same is true of the human mind,
Lex Fridman (2:27:22.020)
is that it needs to behave appropriately in the future.
Lex Fridman (2:27:26.860)
And it has no idea what the future is going to be like.
Lex Fridman (2:27:29.860)
But it's not going to be like the past.
Lex Fridman (2:27:31.460)
So compressing the past is not appropriate,
Lex Fridman (2:27:33.620)
because the past is not, it's not predictive of the future.
François Chollet (2:27:40.500)
Yeah, history repeats itself, but not perfectly.
Lex Fridman (2:27:44.740)
I don't think I asked you last time the most inappropriately
François Chollet (2:27:48.980)
absurd question.
Lex Fridman (2:27:51.180)
We've talked a lot about intelligence,
Lex Fridman (2:27:54.420)
but the bigger question from intelligence is of meaning.
Lex Fridman (2:28:00.860)
Intelligence systems are kind of goal oriented.
François Chollet (2:28:02.980)
They're always optimizing for a goal.
Lex Fridman (2:28:05.380)
If you look at the Hutter Prize, actually,
François Chollet (2:28:07.620)
I mean, there's always a clean formulation of a goal.
Lex Fridman (2:28:10.860)
But the natural question for us humans,
François Chollet (2:28:14.220)
since we don't know our objective function,
Lex Fridman (2:28:16.020)
is what is the meaning of it all?
Lex Fridman (2:28:18.460)
So the absurd question is, what, Francois,
Lex Fridman (2:28:22.980)
do you think is the meaning of life?
Lex Fridman (2:28:25.660)
What's the meaning of life?
Lex Fridman (2:28:26.820)
Yeah, that's a big question.
Lex Fridman (2:28:28.180)
And I think I can give you my answer, at least one
Lex Fridman (2:28:33.220)
of my answers.
Lex Fridman (2:28:34.540)
And so one thing that's very important in understanding who
Lex Fridman (2:28:42.220)
we are is that everything that makes up ourselves,
François Chollet (2:28:48.380)
that makes up who we are, even your most personal thoughts,
Lex Fridman (2:28:53.740)
is not actually your own.
François Chollet (2:28:55.700)
Even your most personal thoughts are expressed in words
Lex Fridman (2:29:00.060)
that you did not invent and are built on concepts and images
François Chollet (2:29:04.940)
that you did not invent.
Lex Fridman (2:29:06.900)
We are very much cultural beings.
François Chollet (2:29:10.940)
We are made of culture.
Lex Fridman (2:29:12.860)
What makes us different from animals, for instance?
Lex Fridman (2:29:16.660)
So everything about ourselves is an echo of the past.
Lex Fridman (2:29:22.860)
Is an echo of the past, an echo of people who lived before us.
François Chollet (2:29:29.900)
That's who we are.
Lex Fridman (2:29:31.420)
And in the same way, if we manage
François Chollet (2:29:35.300)
to contribute something to the collective edifice of culture,
Lex Fridman (2:29:41.780)
a new idea, maybe a beautiful piece of music,
François Chollet (2:29:44.580)
a work of art, a grand theory, a new world, maybe,
Lex Fridman (2:29:51.260)
that something is going to become
François Chollet (2:29:54.380)
a part of the minds of future humans, essentially, forever.
Lex Fridman (2:30:00.300)
So everything we do creates ripples
François Chollet (2:30:03.980)
that propagate into the future.
Lex Fridman (2:30:06.020)
And in a way, this is our path to immortality,
François Chollet (2:30:11.900)
is that as we contribute things to culture,
Lex Fridman (2:30:17.580)
culture in turn becomes future humans.
Lex Fridman (2:30:21.420)
And we keep influencing people thousands of years from now.
Lex Fridman (2:30:27.660)
So our actions today create ripples.
Lex Fridman (2:30:30.740)
And these ripples, I think, basically
Lex Fridman (2:30:35.140)
sum up the meaning of life.
François Chollet (2:30:37.620)
In the same way that we are the sum
Lex Fridman (2:30:42.540)
of the interactions between many different ripples that
François Chollet (2:30:45.500)
came from our past, we are ourselves
Lex Fridman (2:30:48.100)
creating ripples that will propagate into the future.
Lex Fridman (2:30:50.700)
And that's why we should be, this
Lex Fridman (2:30:53.460)
seems like perhaps an eighth thing to say,
Lex Fridman (2:30:56.060)
but we should be kind to others during our time on Earth
Lex Fridman (2:31:02.060)
because every act of kindness creates ripples.
Lex Fridman (2:31:05.660)
And in reverse, every act of violence also creates ripples.
Lex Fridman (2:31:09.380)
And you want to carefully choose which kind of ripples
François Chollet (2:31:13.260)
you want to create, and you want to propagate into the future.
Lex Fridman (2:31:16.460)
And in your case, first of all, beautifully put,
Lex Fridman (2:31:19.020)
but in your case, creating ripples
Lex Fridman (2:31:21.380)
into the future human and future AGI systems.
François Chollet (2:31:27.780)
Yes.
Lex Fridman (2:31:28.500)
It's fascinating.
François Chollet (2:31:29.500)
Our successors.
Lex Fridman (2:31:30.420)
I don't think there's a better way to end it,
François Chollet (2:31:34.500)
Francois, as always, for a second time.
Lex Fridman (2:31:37.180)
And I'm sure many times in the future,
François Chollet (2:31:39.340)
it's been a huge honor.
Lex Fridman (2:31:40.820)
You're one of the most brilliant people
François Chollet (2:31:43.380)
in the machine learning, computer science world.
Lex Fridman (2:31:47.500)
Again, it's a huge honor.
François Chollet (2:31:48.700)
Thanks for talking to me.
Lex Fridman (2:31:49.460)
It's been a pleasure.
François Chollet (2:31:50.540)
Thanks a lot for having me.
Lex Fridman (2:31:51.980)
We appreciate it.
François Chollet (2:31:53.900)
Thanks for listening to this conversation with Francois
Lex Fridman (2:31:56.220)
Chollet, and thank you to our sponsors, Babbel, Masterclass,
Lex Fridman (2:32:00.340)
and Cash App.
Lex Fridman (2:32:01.660)
Click the sponsor links in the description
François Chollet (2:32:03.900)
to get a discount and to support this podcast.
Lex Fridman (2:32:06.820)
If you enjoy this thing, subscribe on YouTube,
François Chollet (2:32:09.060)
review it with five stars on Apple Podcast,
Lex Fridman (2:32:11.340)
follow on Spotify, support on Patreon,
François Chollet (2:32:14.060)
or connect with me on Twitter at Lex Friedman.
Lex Fridman (2:32:17.780)
And now let me leave you with some words
François Chollet (2:32:19.420)
from René Descartes in 1668, an excerpt of which Francois
Lex Fridman (2:32:24.380)
includes and is on the measure of intelligence paper.
François Chollet (2:32:27.780)
If there were machines which bore a resemblance
Lex Fridman (2:32:30.300)
to our bodies and imitated our actions as closely as possible
François Chollet (2:32:34.420)
for all practical purposes, we should still
Lex Fridman (2:32:36.980)
have two very certain means of recognizing
François Chollet (2:32:40.020)
that they were not real men.
Lex Fridman (2:32:42.220)
The first is that they could never use words or put together
François Chollet (2:32:45.300)
signs, as we do in order to declare our thoughts to others.
Lex Fridman (2:32:49.820)
For we can certainly conceive of a machine so constructed
François Chollet (2:32:53.420)
that it utters words and even utters
Lex Fridman (2:32:55.540)
words that correspond to bodily actions causing
François Chollet (2:32:57.940)
a change in its organs.
Lex Fridman (2:32:59.580)
But it is not conceivable that such a machine should produce
François Chollet (2:33:03.380)
different arrangements of words so as
Lex Fridman (2:33:05.420)
to give an appropriately meaningful answer to whatever
François Chollet (2:33:08.620)
is said in its presence as the dullest of men can do.
Lex Fridman (2:33:12.780)
Here, Descartes is anticipating the Turing test,
Lex Fridman (2:33:15.460)
and the argument still continues to this day.
Lex Fridman (2:33:18.780)
Secondly, he continues, even though some machines might
François Chollet (2:33:22.140)
do some things as well as we do them, or perhaps even better,
Lex Fridman (2:33:26.580)
they would inevitably fail in others,
François Chollet (2:33:29.100)
which would reveal that they are acting not from understanding
Lex Fridman (2:33:32.420)
but only from the disposition of their organs.
François Chollet (2:33:36.780)
This is an incredible quote.
Lex Fridman (2:33:39.860)
Whereas reason is a universal instrument
François Chollet (2:33:43.220)
which can be used in all kinds of situations,
Lex Fridman (2:33:46.580)
these organs need some particular action.
François Chollet (2:33:49.060)
Hence, it is for all practical purposes
Lex Fridman (2:33:51.220)
impossible for a machine to have enough different organs
François Chollet (2:33:54.300)
to make it act in all the contingencies of life
Lex Fridman (2:33:57.780)
and the way in which our reason makes us act.
François Chollet (2:34:01.340)
That's the debate between mimicry and memorization
Lex Fridman (2:34:05.060)
versus understanding.
Lex Fridman (2:34:07.220)
So thank you for listening and hope to see you next time.
Lex Fridman (30:00.180)
because one specific road takes you from point A to point B,
Lex Fridman (30:03.460)
but a road building company can take you from,
Lex Fridman (30:06.180)
can make a path from anywhere to anywhere else.
François Chollet (30:08.980)
Yeah, that's beautifully put,
Lex Fridman (30:10.140)
but it's also to play devil's advocate a little bit.
François Chollet (30:15.460)
You know, it's possible that there's something
Lex Fridman (30:18.740)
more fundamental than us humans.
Lex Fridman (30:21.260)
So you kind of said the programmer creates
Lex Fridman (30:25.860)
the difference between the choir,
François Chollet (30:28.940)
the skill and the skill itself.
Lex Fridman (30:31.340)
There could be something like,
François Chollet (30:32.780)
you could argue the universe is more intelligent.
Lex Fridman (30:36.420)
Like the base intelligence that we should be trying
François Chollet (30:43.020)
to measure is something that created humans.
Lex Fridman (30:46.500)
We should be measuring God or the source of the universe
François Chollet (30:51.540)
as opposed to, like there could be a deeper intelligence.
Lex Fridman (30:55.140)
Sure.
François Chollet (30:55.980)
There's always deeper intelligence, I guess.
Lex Fridman (30:57.140)
You can argue that,
Lex Fridman (30:58.020)
but that does not take anything away
Lex Fridman (31:00.100)
from the fact that humans are intelligent.
Lex Fridman (31:01.900)
And you can tell that
Lex Fridman (31:03.260)
because they are capable of adaptation and generality.
François Chollet (31:07.020)
Got it.
Lex Fridman (31:07.860)
And you see that in particular in the fact
François Chollet (31:09.700)
that humans are capable of handling situations and tasks
Lex Fridman (31:16.780)
that are quite different from anything
François Chollet (31:19.780)
that any of our evolutionary ancestors
Lex Fridman (31:22.940)
has ever encountered.
Lex Fridman (31:24.540)
So we are capable of generalizing very much
Lex Fridman (31:27.140)
out of distribution,
François Chollet (31:28.100)
if you consider our evolutionary history
Lex Fridman (31:30.260)
as being in a way our training data.
François Chollet (31:33.260)
Of course, evolutionary biologists would argue
Lex Fridman (31:35.060)
that we're not going too far out of the distribution.
François Chollet (31:37.660)
We're like mapping the skills we've learned previously,
Lex Fridman (31:41.380)
desperately trying to like jam them
François Chollet (31:43.540)
into like these new situations.
Lex Fridman (31:47.060)
I mean, there's definitely a little bit of that,
Lex Fridman (31:49.460)
but it's pretty clear to me that we're able to,
Lex Fridman (31:53.660)
most of the things we do any given day
François Chollet (31:56.580)
in our modern civilization
Lex Fridman (31:58.060)
are things that are very, very different
François Chollet (32:00.860)
from what our ancestors a million years ago
Lex Fridman (32:03.900)
would have been doing in a given day.
Lex Fridman (32:05.900)
And your environment is very different.
Lex Fridman (32:07.540)
So I agree that everything we do,
François Chollet (32:12.180)
we do it with cognitive building blocks
Lex Fridman (32:14.220)
that we acquired over the course of evolution, right?
Lex Fridman (32:17.820)
And that anchors our cognition to a certain context,
Lex Fridman (32:22.180)
which is the human condition very much.
Lex Fridman (32:25.260)
But still our mind is capable of a pretty remarkable degree
Lex Fridman (32:29.500)
of generality far beyond anything we can create
François Chollet (32:32.700)
in artificial systems today.
Lex Fridman (32:34.100)
Like the degree in which the mind can generalize
François Chollet (32:37.740)
from its evolutionary history,
Lex Fridman (32:41.620)
can generalize away from its evolutionary history
François Chollet (32:43.940)
is much greater than the degree
Lex Fridman (32:46.500)
to which a deep learning system today
François Chollet (32:48.860)
can generalize away from its training data.
Lex Fridman (32:51.020)
And like the key point you're making,
François Chollet (32:52.380)
which I think is quite beautiful is like,
Lex Fridman (32:54.220)
we shouldn't measure, if we're talking about measurement,
François Chollet (32:58.660)
we shouldn't measure the skill.
Lex Fridman (33:01.620)
We should measure like the creation of the new skill,
François Chollet (33:04.340)
the ability to create that new skill.
Lex Fridman (33:06.780)
But it's tempting, like it's weird
François Chollet (33:10.940)
because the skill is a little bit of a small window
Lex Fridman (33:13.620)
into the system.
Lex Fridman (33:16.380)
So whenever you have a lot of skills,
Lex Fridman (33:19.420)
it's tempting to measure the skills.
François Chollet (33:21.900)
I mean, the skill is the only thing you can objectively
Lex Fridman (33:25.820)
measure, but yeah.
Lex Fridman (33:27.540)
So the thing to keep in mind is that
Lex Fridman (33:30.780)
when you see skill in the human,
François Chollet (33:35.060)
it gives you a strong signal that that human is intelligent
Lex Fridman (33:39.220)
because you know they weren't born with that skill typically.
François Chollet (33:42.740)
Like you see a very strong chess player,
Lex Fridman (33:45.220)
maybe you're a very strong chess player yourself.
François Chollet (33:47.540)
I think you're saying that because I'm Russian
Lex Fridman (33:51.020)
and now you're prejudiced, you assume.
François Chollet (33:53.860)
All Russians are good at chess.
Lex Fridman (33:54.700)
I'm biased, exactly.
François Chollet (33:55.540)
I'm biased, yeah.
Lex Fridman (33:56.900)
Well, you're definitely biased.
Lex Fridman (34:00.020)
So if you see a very strong chess player,
Lex Fridman (34:01.900)
you know they weren't born knowing how to play chess.
Lex Fridman (34:05.460)
So they had to acquire that skill
Lex Fridman (34:07.780)
with their limited resources, with their limited lifetime.
Lex Fridman (34:10.940)
And they did that because they are generally intelligent.
Lex Fridman (34:15.420)
And so they may as well have acquired any other skill.
François Chollet (34:18.980)
You know they have this potential.
Lex Fridman (34:21.180)
And on the other hand, if you see a computer playing chess,
François Chollet (34:25.700)
you cannot make the same assumptions
Lex Fridman (34:27.860)
because you cannot just assume
François Chollet (34:29.380)
the computer is generally intelligent.
Lex Fridman (34:30.860)
The computer may be born knowing how to play chess
François Chollet (34:35.300)
in the sense that it may have been programmed by a human
Lex Fridman (34:38.220)
that has understood chess for the computer
Lex Fridman (34:40.900)
and that has just encoded the output
Lex Fridman (34:44.180)
of that understanding in a static program.
Lex Fridman (34:46.020)
And that program is not intelligent.
Lex Fridman (34:49.420)
So let's zoom out just for a second and say like,
Lex Fridman (34:52.380)
what is the goal on the measure of intelligence paper?
Lex Fridman (34:57.460)
Like what do you hope to achieve with it?
Lex Fridman (34:59.020)
So the goal of the paper is to clear up
Lex Fridman (35:01.700)
some longstanding misunderstandings
François Chollet (35:04.580)
about the way we've been conceptualizing intelligence
Lex Fridman (35:08.380)
in the AI community and in the way we've been
François Chollet (35:12.500)
evaluating progress in AI.
Lex Fridman (35:16.780)
There's been a lot of progress recently in machine learning
Lex Fridman (35:19.060)
and people are extrapolating from that progress
Lex Fridman (35:22.140)
that we are about to solve general intelligence.
Lex Fridman (35:26.380)
And if you want to be able to evaluate these statements,
Lex Fridman (35:30.500)
you need to precisely define what you're talking about
François Chollet (35:33.820)
when you're talking about general intelligence.
Lex Fridman (35:35.580)
And you need a formal way, a reliable way to measure
Lex Fridman (35:40.580)
how much intelligence,
Lex Fridman (35:42.380)
how much general intelligence a system processes.
Lex Fridman (35:45.900)
And ideally this measure of intelligence
Lex Fridman (35:48.420)
should be actionable.
Lex Fridman (35:50.260)
So it should not just describe what intelligence is.
Lex Fridman (35:54.620)
It should not just be a binary indicator
François Chollet (35:56.860)
that tells you the system is intelligent or it isn't.
Lex Fridman (36:01.620)
It should be actionable.
Lex Fridman (36:03.060)
It should have explanatory power, right?
Lex Fridman (36:05.740)
So you could use it as a feedback signal.
François Chollet (36:08.580)
It would show you the way
Lex Fridman (36:10.980)
towards building more intelligent systems.
Lex Fridman (36:13.100)
So at the first level, you draw a distinction
Lex Fridman (36:16.500)
between two divergent views of intelligence.
François Chollet (36:21.780)
As we just talked about,
Lex Fridman (36:22.860)
intelligence is a collection of task specific skills
Lex Fridman (36:26.820)
and a general learning ability.
Lex Fridman (36:29.900)
So what's the difference between
François Chollet (36:32.300)
kind of this memorization of skills
Lex Fridman (36:35.580)
and a general learning ability?
François Chollet (36:37.820)
We've talked about it a little bit,
Lex Fridman (36:39.580)
but can you try to linger on this topic for a bit?
François Chollet (36:43.060)
Yeah, so the first part of the paper
Lex Fridman (36:45.460)
is an assessment of the different ways
François Chollet (36:49.100)
we've been thinking about intelligence
Lex Fridman (36:50.500)
and the different ways we've been evaluating progress in AI.
Lex Fridman (36:54.540)
And this tree of cognitive sciences
Lex Fridman (36:57.700)
has been shaped by two views of the human mind.
Lex Fridman (37:01.220)
And one view is the evolutionary psychology view
Lex Fridman (37:04.740)
in which the mind is a collection of fairly static
François Chollet (37:10.660)
special purpose ad hoc mechanisms
Lex Fridman (37:14.220)
that have been hard coded by evolution
François Chollet (37:17.620)
over our history as a species for a very long time.
Lex Fridman (37:22.500)
And early AI researchers,
François Chollet (37:27.940)
people like Marvin Minsky, for instance,
Lex Fridman (37:30.340)
they clearly subscribed to this view.
Lex Fridman (37:33.300)
And they saw the mind as a kind of
Lex Fridman (37:36.860)
collection of static programs
François Chollet (37:39.820)
similar to the programs they would run
Lex Fridman (37:42.140)
on like mainframe computers.
Lex Fridman (37:43.580)
And in fact, I think they very much understood the mind
Lex Fridman (37:48.060)
through the metaphor of the mainframe computer
Lex Fridman (37:50.540)
because that was the tool they were working with, right?
Lex Fridman (37:53.580)
And so you had these static programs,
François Chollet (37:55.100)
this collection of very different static programs
Lex Fridman (37:57.180)
operating over a database like memory.
Lex Fridman (38:00.060)
And in this picture, learning was not very important.
Lex Fridman (38:03.580)
Learning was considered to be just memorization.
Lex Fridman (38:05.660)
And in fact, learning is basically not featured
Lex Fridman (38:10.380)
in AI textbooks until the 1980s
François Chollet (38:14.620)
with the rise of machine learning.
Lex Fridman (38:16.940)
It's kind of fun to think about
François Chollet (38:18.780)
that learning was the outcast.
Lex Fridman (38:21.500)
Like the weird people working on learning,
François Chollet (38:24.060)
like the mainstream AI world was,
Lex Fridman (38:28.100)
I mean, I don't know what the best term is,
Lex Fridman (38:31.780)
but it's non learning.
Lex Fridman (38:33.900)
It was seen as like reasoning would not be learning based.
François Chollet (38:37.940)
Yes, it was considered that the mind
Lex Fridman (38:40.620)
was a collection of programs
François Chollet (38:43.180)
that were primarily logical in nature.
Lex Fridman (38:46.620)
And that's all you needed to do to create a mind
François Chollet (38:49.140)
was to write down these programs
Lex Fridman (38:50.860)
and they would operate over knowledge,
François Chollet (38:52.860)
which would be stored in some kind of database.
Lex Fridman (38:55.100)
And as long as your database would encompass,
François Chollet (38:57.300)
you know, everything about the world
Lex Fridman (38:59.380)
and your logical rules were comprehensive,
François Chollet (39:03.340)
then you would have a mind.
Lex Fridman (39:04.940)
So the other view of the mind
Lex Fridman (39:06.420)
is the brain as a sort of blank slate, right?
Lex Fridman (39:11.940)
This is a very old idea.
François Chollet (39:13.180)
You find it in John Locke's writings.
Lex Fridman (39:16.140)
This is the tabula rasa.
Lex Fridman (39:19.220)
And this is this idea that the mind
Lex Fridman (39:21.140)
is some kind of like information sponge
François Chollet (39:23.340)
that starts empty, that starts blank.
Lex Fridman (39:27.340)
And that absorbs knowledge and skills from experience, right?
Lex Fridman (39:34.340)
So it's a sponge that reflects the complexity of the world,
Lex Fridman (39:38.700)
the complexity of your life experience, essentially.
François Chollet (39:41.780)
That everything you know and everything you can do
Lex Fridman (39:44.340)
is a reflection of something you found
François Chollet (39:47.740)
in the outside world, essentially.
Lex Fridman (39:49.580)
So this is an idea that's very old.
François Chollet (39:51.580)
That was not very popular, for instance, in the 1970s.
Lex Fridman (39:56.780)
But that gained a lot of vitality recently
François Chollet (39:58.820)
with the rise of connectionism, in particular deep learning.
Lex Fridman (40:02.300)
And so today, deep learning
François Chollet (40:03.780)
is the dominant paradigm in AI.
Lex Fridman (40:06.540)
And I feel like lots of AI researchers
François Chollet (40:10.420)
are conceptualizing the mind via a deep learning metaphor.
Lex Fridman (40:14.980)
Like they see the mind as a kind of
François Chollet (40:17.820)
randomly initialized neural network that starts blank
Lex Fridman (40:21.660)
when you're born.
Lex Fridman (40:22.500)
And then that gets trained via exposure to trained data
Lex Fridman (40:26.100)
that acquires knowledge and skills
François Chollet (40:27.740)
via exposure to trained data.
Lex Fridman (40:29.220)
By the way, it's a small tangent.
François Chollet (40:32.700)
I feel like people who are thinking about intelligence
Lex Fridman (40:36.700)
are not conceptualizing it that way.
François Chollet (40:39.700)
I actually haven't met too many people
Lex Fridman (40:41.820)
who believe that a neural network
François Chollet (40:44.700)
will be able to reason, who seriously think that rigorously.
Lex Fridman (40:51.660)
Because I think it's actually an interesting worldview.
Lex Fridman (40:54.260)
And we'll talk about it more,
Lex Fridman (40:56.420)
but it's been impressive what neural networks
François Chollet (41:00.420)
have been able to accomplish.
Lex Fridman (41:02.100)
And to me, I don't know, you might disagree,
Lex Fridman (41:04.540)
but it's an open question whether like scaling size
Lex Fridman (41:09.820)
eventually might lead to incredible results
François Chollet (41:13.660)
to us mere humans will appear as if it's general.
Lex Fridman (41:17.060)
I mean, if you ask people who are seriously thinking
François Chollet (41:19.860)
about intelligence, they will definitely not say
Lex Fridman (41:22.660)
that all you need to do is,
François Chollet (41:24.900)
like the mind is just a neural network.
Lex Fridman (41:27.420)
However, it's actually a view that's very popular,
François Chollet (41:30.420)
I think, in the deep learning community
Lex Fridman (41:31.780)
that many people are kind of conceptually
François Chollet (41:35.460)
intellectually lazy about it.
Lex Fridman (41:37.140)
Right, it's a, but I guess what I'm saying exactly right,
François Chollet (41:40.500)
it's, I mean, I haven't met many people
Lex Fridman (41:44.740)
and I think it would be interesting to meet a person
François Chollet (41:47.740)
who is not intellectually lazy about this particular topic
Lex Fridman (41:50.260)
and still believes that neural networks will go all the way.
François Chollet (41:54.460)
I think Yama is probably closest to that
Lex Fridman (41:56.820)
with self supervised.
François Chollet (41:57.660)
There are definitely people who argue
Lex Fridman (41:59.660)
that current deep learning techniques
François Chollet (42:03.100)
are already the way to general artificial intelligence.
Lex Fridman (42:06.860)
And that all you need to do is to scale it up
François Chollet (42:09.460)
to all the available trained data.
Lex Fridman (42:12.780)
And that's, if you look at the waves
François Chollet (42:16.300)
that OpenAI's GPT3 model has made,
Lex Fridman (42:19.500)
you see echoes of this idea.
Lex Fridman (42:22.700)
So on that topic, GPT3, similar to GPT2 actually,
Lex Fridman (42:28.980)
have captivated some part of the imagination of the public.
François Chollet (42:33.060)
There's just a bunch of hype of different kind.
Lex Fridman (42:35.580)
That's, I would say it's emergent.
François Chollet (42:37.940)
It's not artificially manufactured.
Lex Fridman (42:39.820)
It's just like people just get excited
François Chollet (42:42.580)
for some strange reason.
Lex Fridman (42:43.780)
And in the case of GPT3, which is funny,
François Chollet (42:46.500)
that there's, I believe, a couple months delay
Lex Fridman (42:49.100)
from release to hype.
François Chollet (42:51.580)
Maybe I'm not historically correct on that,
Lex Fridman (42:56.780)
but it feels like there was a little bit of a lack of hype
Lex Fridman (43:01.260)
and then there's a phase shift into hype.
Lex Fridman (43:04.780)
But nevertheless, there's a bunch of cool applications
François Chollet (43:07.460)
that seem to captivate the imagination of the public
Lex Fridman (43:10.380)
about what this language model
François Chollet (43:12.140)
that's trained in unsupervised way
Lex Fridman (43:15.180)
without any fine tuning is able to achieve.
Lex Fridman (43:19.500)
So what do you make of that?
Lex Fridman (43:20.900)
What are your thoughts about GPT3?
François Chollet (43:22.940)
Yeah, so I think what's interesting about GPT3
Lex Fridman (43:25.700)
is the idea that it may be able to learn new tasks
François Chollet (43:31.180)
after just being shown a few examples.
Lex Fridman (43:33.580)
So I think if it's actually capable of doing that,
François Chollet (43:35.620)
that's novel and that's very interesting
Lex Fridman (43:37.580)
and that's something we should investigate.
François Chollet (43:39.900)
That said, I must say, I'm not entirely convinced
Lex Fridman (43:43.140)
that we have shown it's capable of doing that.
François Chollet (43:47.300)
It's very likely, given the amount of data
Lex Fridman (43:50.980)
that the model is trained on,
François Chollet (43:52.220)
that what it's actually doing is pattern matching
Lex Fridman (43:55.700)
a new task you give it with a task
François Chollet (43:58.060)
that it's been exposed to in its trained data.
Lex Fridman (44:00.100)
It's just recognizing the task
Lex Fridman (44:01.620)
instead of just developing a model of the task, right?
Lex Fridman (44:05.540)
But there's, sorry to interrupt,
François Chollet (44:07.660)
there's a parallel as to what you said before,
Lex Fridman (44:10.020)
which is it's possible to see GPT3 as like the prompts
François Chollet (44:14.620)
it's given as a kind of SQL query
Lex Fridman (44:17.780)
into this thing that it's learned,
François Chollet (44:19.580)
similar to what you said before,
Lex Fridman (44:20.860)
which is language is used to query the memory.
François Chollet (44:23.340)
Yes.
Lex Fridman (44:24.180)
So is it possible that neural network
François Chollet (44:26.940)
is a giant memorization thing,
Lex Fridman (44:29.300)
but then if it gets sufficiently giant,
François Chollet (44:32.260)
it'll memorize sufficiently large amounts
Lex Fridman (44:35.100)
of things in the world or it becomes,
Lex Fridman (44:37.860)
or intelligence becomes a querying machine?
Lex Fridman (44:40.580)
I think it's possible that a significant chunk
François Chollet (44:44.180)
of intelligence is this giant associative memory.
Lex Fridman (44:48.740)
I definitely don't believe that intelligence
François Chollet (44:51.340)
is just a giant associative memory,
Lex Fridman (44:53.740)
but it may well be a big component.
Lex Fridman (44:57.660)
So do you think GPT3, 4, 5,
Lex Fridman (45:02.660)
GPT10 will eventually, like, what do you think,
Lex Fridman (45:07.140)
where's the ceiling?
Lex Fridman (45:08.340)
Do you think you'll be able to reason?
François Chollet (45:11.980)
No, that's a bad question.
Lex Fridman (45:14.620)
Like, what is the ceiling is the better question.
Lex Fridman (45:17.340)
How well is it gonna scale?
Lex Fridman (45:18.500)
How good is GPTN going to be?
François Chollet (45:21.180)
Yeah.
Lex Fridman (45:22.020)
So I believe GPTN is gonna.
François Chollet (45:25.420)
GPTN.
Lex Fridman (45:26.860)
Is gonna improve on the strength of GPT2 and 3,
François Chollet (45:30.940)
which is it will be able to generate, you know,
Lex Fridman (45:33.980)
ever more plausible text in context.
François Chollet (45:37.660)
Just monotonically increasing performance.
Lex Fridman (45:41.260)
Yes, if you train a bigger model on more data,
François Chollet (45:44.340)
then your text will be increasingly more context aware
Lex Fridman (45:49.340)
and increasingly more plausible
François Chollet (45:51.220)
in the same way that GPT3 is much better
Lex Fridman (45:54.700)
at generating plausible text compared to GPT2.
Lex Fridman (45:57.500)
But that said, I don't think just scaling up the model
Lex Fridman (46:01.940)
to more transformer layers and more trained data
François Chollet (46:04.180)
is gonna address the flaws of GPT3,
Lex Fridman (46:07.020)
which is that it can generate plausible text,
Lex Fridman (46:09.900)
but that text is not constrained by anything else
Lex Fridman (46:13.620)
other than plausibility.
Lex Fridman (46:15.180)
So in particular, it's not constrained by factualness
Lex Fridman (46:19.180)
or even consistency, which is why it's very easy
François Chollet (46:21.820)
to get GPT3 to generate statements
Lex Fridman (46:23.860)
that are factually untrue.
François Chollet (46:26.260)
Or to generate statements that are even self contradictory.
Lex Fridman (46:29.580)
Right?
François Chollet (46:30.420)
Because it's only goal is plausibility,
Lex Fridman (46:35.420)
and it has no other constraints.
François Chollet (46:37.620)
It's not constrained to be self consistent, for instance.
Lex Fridman (46:40.300)
Right?
Lex Fridman (46:41.140)
And so for this reason, one thing that I thought
Lex Fridman (46:43.540)
was very interesting with GPT3 is that you can
François Chollet (46:46.780)
predetermine the answer it will give you
Lex Fridman (46:49.780)
by asking the question in a specific way,
François Chollet (46:52.020)
because it's very responsive to the way you ask the question.
Lex Fridman (46:55.260)
Since it has no understanding of the content of the question.
François Chollet (47:00.260)
Right.
Lex Fridman (47:01.100)
And if you have the same question in two different ways
François Chollet (47:05.620)
that are basically adversarially engineered
Lex Fridman (47:09.020)
to produce certain answer,
François Chollet (47:10.260)
you will get two different answers,
Lex Fridman (47:12.740)
two contradictory answers.
François Chollet (47:14.180)
It's very susceptible to adversarial attacks, essentially.
Lex Fridman (47:16.660)
Potentially, yes.
Lex Fridman (47:17.780)
So in general, the problem with these models,
Lex Fridman (47:20.820)
these generative models, is that they are very good
François Chollet (47:24.180)
at generating plausible text,
Lex Fridman (47:27.220)
but that's just not enough.
Lex Fridman (47:29.660)
Right?
Lex Fridman (47:33.620)
I think one avenue that would be very interesting
François Chollet (47:36.500)
to make progress is to make it possible
Lex Fridman (47:40.780)
to write programs over the latent space
François Chollet (47:43.860)
that these models operate on.
Lex Fridman (47:45.620)
That you would rely on these self supervised models
François Chollet (47:49.460)
to generate a sort of like pool of knowledge and concepts
Lex Fridman (47:54.340)
and common sense.
Lex Fridman (47:55.260)
And then you would be able to write
Lex Fridman (47:57.180)
explicit reasoning programs over it.
François Chollet (48:01.460)
Because the current problem with GPT3 is that
Lex Fridman (48:03.660)
it can be quite difficult to get it to do what you want to do.
François Chollet (48:09.420)
If you want to turn GPT3 into products,
Lex Fridman (48:12.420)
you need to put constraints on it.
François Chollet (48:14.780)
You need to force it to obey certain rules.
Lex Fridman (48:19.500)
So you need a way to program it explicitly.
François Chollet (48:22.540)
Yeah, so if you look at its ability
Lex Fridman (48:24.220)
to do program synthesis,
François Chollet (48:26.140)
it generates, like you said, something that's plausible.
Lex Fridman (48:29.060)
Yeah, so if you try to make it generate programs,
François Chollet (48:32.580)
it will perform well for any program
Lex Fridman (48:35.940)
that it has seen in its training data.
Lex Fridman (48:38.700)
But because program space is not interpretive, right?
Lex Fridman (48:42.940)
It's not going to be able to generalize to problems
François Chollet (48:46.740)
it hasn't seen before.
Lex Fridman (48:48.700)
Now that's currently, do you think sort of an absurd,
Lex Fridman (48:54.980)
but I think useful, I guess, intuition builder is,
Lex Fridman (49:00.340)
you know, the GPT3 has 175 billion parameters.
François Chollet (49:07.340)
Human brain has 100, has about a thousand times that
Lex Fridman (49:11.740)
or more in terms of number of synapses.
Lex Fridman (49:16.380)
Do you think, obviously, very different kinds of things,
Lex Fridman (49:21.180)
but there is some degree of similarity.
Lex Fridman (49:26.380)
Do you think, what do you think GPT will look like
Lex Fridman (49:30.700)
when it has 100 trillion parameters?
Lex Fridman (49:34.180)
You think our conversation might be in nature different?
Lex Fridman (49:39.100)
Like, because you've criticized GPT3 very effectively now.
Lex Fridman (49:42.940)
Do you think?
Lex Fridman (49:45.420)
No, I don't think so.
Lex Fridman (49:46.940)
So to begin with, the bottleneck with scaling up GPT3,
Lex Fridman (49:51.020)
GPT models, generative pre trained transformer models,
François Chollet (49:54.860)
is not going to be the size of the model
Lex Fridman (49:57.620)
or how long it takes to train it.
François Chollet (49:59.580)
The bottleneck is going to be the trained data
Lex Fridman (50:01.860)
because OpenAI is already training GPT3
Lex Fridman (50:05.540)
on a core of basically the entire web, right?
Lex Fridman (50:08.860)
And that's a lot of data.
Lex Fridman (50:09.820)
So you could imagine training on more data than that,
Lex Fridman (50:12.140)
like Google could train on more data than that,
Lex Fridman (50:14.460)
but it would still be only incrementally more data.
Lex Fridman (50:17.500)
And I don't recall exactly how much more data GPT3
François Chollet (50:21.340)
was trained on compared to GPT2,
Lex Fridman (50:22.820)
but it's probably at least like a hundred,
François Chollet (50:25.100)
maybe even a thousand X.
Lex Fridman (50:26.620)
I don't have the exact number.
François Chollet (50:28.460)
You're not going to be able to train a model
Lex Fridman (50:30.140)
on a hundred more data than what you're already doing.
Lex Fridman (50:34.180)
So that's brilliant.
Lex Fridman (50:35.300)
So it's easier to think of compute as a bottleneck
Lex Fridman (50:38.940)
and then arguing that we can remove that bottleneck.
Lex Fridman (50:41.380)
But we can remove the compute bottleneck.
François Chollet (50:43.060)
I don't think it's a big problem.
Lex Fridman (50:44.580)
If you look at the pace at which we've improved
François Chollet (50:48.500)
the efficiency of deep learning models
Lex Fridman (50:51.340)
in the past few years,
François Chollet (50:54.060)
I'm not worried about train time bottlenecks
Lex Fridman (50:57.180)
or model size bottlenecks.
François Chollet (50:59.580)
The bottleneck in the case
Lex Fridman (51:01.140)
of these generative transformer models
François Chollet (51:03.420)
is absolutely the trained data.
Lex Fridman (51:05.540)
What about the quality of the data?
François Chollet (51:07.740)
So, yeah.
Lex Fridman (51:08.580)
So the quality of the data is an interesting point.
François Chollet (51:10.900)
The thing is,
Lex Fridman (51:11.900)
if you're going to want to use these models
François Chollet (51:14.460)
in real products,
Lex Fridman (51:16.900)
then you want to feed them data
François Chollet (51:20.060)
that's as high quality, as factual,
Lex Fridman (51:23.460)
I would say as unbiased as possible,
François Chollet (51:25.620)
that there's not really such a thing
Lex Fridman (51:27.340)
as unbiased data in the first place.
Lex Fridman (51:30.500)
But you probably don't want to train it on Reddit,
Lex Fridman (51:34.020)
for instance.
François Chollet (51:34.860)
It sounds like a bad plan.
Lex Fridman (51:37.060)
So from my personal experience,
François Chollet (51:38.620)
working with large scale deep learning models.
Lex Fridman (51:42.740)
So at some point I was working on a model at Google
François Chollet (51:46.580)
that's trained on 350 million labeled images.
Lex Fridman (51:52.340)
It's an image classification model.
François Chollet (51:53.660)
That's a lot of images.
Lex Fridman (51:54.660)
That's like probably most publicly available images
François Chollet (51:58.140)
on the web at the time.
Lex Fridman (52:00.980)
And it was a very noisy data set
François Chollet (52:03.900)
because the labels were not originally annotated by hand,
Lex Fridman (52:07.820)
by humans.
François Chollet (52:08.660)
They were automatically derived from like tags
Lex Fridman (52:12.420)
on social media,
François Chollet (52:14.300)
or just keywords in the same page
Lex Fridman (52:16.820)
as the image was found and so on.
Lex Fridman (52:18.220)
So it was very noisy.
Lex Fridman (52:19.140)
And it turned out that you could easily get a better model,
François Chollet (52:25.340)
not just by training,
Lex Fridman (52:26.500)
like if you train on more of the noisy data,
François Chollet (52:29.980)
you get an incrementally better model,
Lex Fridman (52:31.540)
but you very quickly hit diminishing returns.
François Chollet (52:35.500)
On the other hand,
Lex Fridman (52:36.660)
if you train on smaller data set
François Chollet (52:38.420)
with higher quality annotations,
Lex Fridman (52:40.020)
quality annotations that are actually made by humans,
François Chollet (52:45.380)
you get a better model.
Lex Fridman (52:47.340)
And it also takes less time to train it.
François Chollet (52:49.860)
Yeah, that's fascinating.
Lex Fridman (52:51.580)
It's the self supervised learning.
François Chollet (52:53.500)
There's a way to get better doing the automated labeling.
Lex Fridman (52:58.780)
Yeah, so you can enrich or refine your labels
François Chollet (53:04.620)
in an automated way.
Lex Fridman (53:05.860)
That's correct.
Lex Fridman (53:07.460)
Do you have a hope for,
Lex Fridman (53:08.700)
I don't know if you're familiar
François Chollet (53:09.580)
with the idea of a semantic web.
Lex Fridman (53:11.980)
Is a semantic web just for people who are not familiar
Lex Fridman (53:15.620)
and is the idea of being able to convert the internet
Lex Fridman (53:20.620)
or be able to attach like semantic meaning
François Chollet (53:25.700)
to the words on the internet,
Lex Fridman (53:27.940)
the sentences, the paragraphs,
François Chollet (53:29.780)
to be able to convert information on the internet
Lex Fridman (53:33.940)
or some fraction of the internet
François Chollet (53:35.660)
into something that's interpretable by machines.
Lex Fridman (53:39.140)
That was kind of a dream for,
François Chollet (53:44.260)
I think the semantic web papers in the nineties,
Lex Fridman (53:47.020)
it's kind of the dream that, you know,
François Chollet (53:49.740)
the internet is full of rich, exciting information.
Lex Fridman (53:52.340)
Even just looking at Wikipedia,
François Chollet (53:54.420)
we should be able to use that as data for machines.
Lex Fridman (53:57.780)
And so far it's not,
François Chollet (53:58.980)
it's not really in a format that's available to machines.
Lex Fridman (54:01.220)
So no, I don't think the semantic web will ever work
Lex Fridman (54:04.540)
simply because it would be a lot of work, right?
Lex Fridman (54:08.020)
To make, to provide that information in structured form.
Lex Fridman (54:12.020)
And there is not really any incentive
Lex Fridman (54:13.820)
for anyone to provide that work.
Lex Fridman (54:16.340)
So I think the way forward to make the knowledge
Lex Fridman (54:21.180)
on the web available to machines
François Chollet (54:22.820)
is actually something closer to unsupervised deep learning.
Lex Fridman (54:29.140)
So GPT3 is actually a bigger step in the direction
François Chollet (54:32.220)
of making the knowledge of the web available to machines
Lex Fridman (54:34.940)
than the semantic web was.
François Chollet (54:36.660)
Yeah, perhaps in a human centric sense,
Lex Fridman (54:40.140)
it feels like GPT3 hasn't learned anything
François Chollet (54:47.300)
that could be used to reason.
Lex Fridman (54:50.340)
But that might be just the early days.
François Chollet (54:52.820)
Yeah, I think that's correct.
Lex Fridman (54:54.300)
I think the forms of reasoning that you see it perform
François Chollet (54:57.340)
are basically just reproducing patterns
Lex Fridman (55:00.660)
that it has seen in string data.
Lex Fridman (55:02.380)
So of course, if you're trained on the entire web,
Lex Fridman (55:06.580)
then you can produce an illusion of reasoning
François Chollet (55:09.340)
in many different situations.
Lex Fridman (55:10.740)
But it will break down if it's presented
François Chollet (55:13.100)
with a novel situation.
Lex Fridman (55:15.260)
That's the open question between the illusion of reasoning
Lex Fridman (55:17.660)
and actual reasoning, yeah.
Lex Fridman (55:18.700)
Yes.
François Chollet (55:19.660)
The power to adapt to something that is genuinely new.
Lex Fridman (55:22.780)
Because the thing is, even imagine you had,
François Chollet (55:28.020)
you could train on every bit of data
Lex Fridman (55:31.100)
ever generated in the history of humanity.
François Chollet (55:35.500)
It remains, that model would be capable
Lex Fridman (55:38.540)
of anticipating many different possible situations.
Lex Fridman (55:43.220)
But it remains that the future is
Lex Fridman (55:45.660)
going to be something different.
François Chollet (55:48.940)
For instance, if you train a GPT3 model on data
Lex Fridman (55:52.940)
from the year 2002, for instance,
Lex Fridman (55:55.700)
and then use it today, it's going to be missing many things.
Lex Fridman (55:58.260)
It's going to be missing many common sense
François Chollet (56:00.740)
facts about the world.
Lex Fridman (56:02.620)
It's even going to be missing vocabulary and so on.
François Chollet (56:05.820)
Yeah, it's interesting that GPT3 even doesn't have,
Lex Fridman (56:09.580)
I think, any information about the coronavirus.
François Chollet (56:13.580)
Yes.
Lex Fridman (56:14.980)
Which is why a system that's, you
François Chollet (56:19.620)
tell that the system is intelligent
Lex Fridman (56:21.300)
when it's capable to adapt.
Lex Fridman (56:22.860)
So intelligence is going to require
Lex Fridman (56:25.580)
some amount of continuous learning.
François Chollet (56:28.140)
It's also going to require some amount of improvisation.
Lex Fridman (56:31.020)
It's not enough to assume that what you're
François Chollet (56:33.980)
going to be asked to do is something
Lex Fridman (56:36.780)
that you've seen before, or something
François Chollet (56:39.300)
that is a simple interpolation of things you've seen before.
Lex Fridman (56:42.700)
Yeah.
François Chollet (56:43.340)
In fact, that model breaks down for even very
Lex Fridman (56:49.060)
tasks that look relatively simple from a distance,
François Chollet (56:52.300)
like L5 self driving, for instance.
Lex Fridman (56:55.660)
Google had a paper a couple of years
François Chollet (56:58.420)
back showing that something like 30 million different road
Lex Fridman (57:04.540)
situations were actually completely insufficient
François Chollet (57:07.180)
to train a driving model.
Lex Fridman (57:09.780)
It wasn't even L2, right?
Lex Fridman (57:11.740)
And that's a lot of data.
Lex Fridman (57:12.820)
That's a lot more data than the 20 or 30 hours of driving
François Chollet (57:16.940)
that a human needs to learn to drive,
Lex Fridman (57:19.580)
given the knowledge they've already accumulated.
François Chollet (57:21.900)
Well, let me ask you on that topic.
Lex Fridman (57:25.540)
Elon Musk, Tesla Autopilot, one of the only companies,
François Chollet (57:31.100)
I believe, is really pushing for a learning based approach.
Lex Fridman (57:34.660)
Are you skeptical that that kind of network
Lex Fridman (57:37.020)
can achieve level 4?
Lex Fridman (57:39.460)
L4 is probably achievable.
François Chollet (57:42.660)
L5 probably not.
Lex Fridman (57:44.420)
What's the distinction there?
Lex Fridman (57:45.860)
Is L5 is completely you can just fall asleep?
Lex Fridman (57:49.340)
Yeah, L5 is basically human level.
François Chollet (57:51.060)
Well, with driving, we have to be careful saying human level,
Lex Fridman (57:53.740)
because that's the most of the drivers.
François Chollet (57:57.180)
Yeah, that's the clearest example of cars
Lex Fridman (58:00.620)
will most likely be much safer than humans in many situations
François Chollet (58:05.020)
where humans fail.
Lex Fridman (58:06.540)
It's the vice versa question.
François Chollet (58:09.860)
I'll tell you, the thing is the amount of trained data
Lex Fridman (58:13.820)
you would need to anticipate for pretty much every possible
François Chollet (58:17.020)
situation you learn content in the real world
Lex Fridman (58:20.460)
is such that it's not entirely unrealistic
François Chollet (58:23.500)
to think that at some point in the future,
Lex Fridman (58:25.540)
we'll develop a system that's trained on enough data,
François Chollet (58:27.700)
especially provided that we can simulate a lot of that data.
Lex Fridman (58:32.340)
We don't necessarily need actual cars
François Chollet (58:34.500)
on the road for everything.
Lex Fridman (58:37.620)
But it's a massive effort.
Lex Fridman (58:39.780)
And it turns out you can create a system that's
Lex Fridman (58:42.100)
much more adaptive, that can generalize much better
François Chollet (58:45.180)
if you just add explicit models of the surroundings
Lex Fridman (58:52.060)
of the car.
Lex Fridman (58:53.580)
And if you use deep learning for what
Lex Fridman (58:55.180)
it's good at, which is to provide
François Chollet (58:57.460)
perceptive information.
Lex Fridman (58:59.500)
So in general, deep learning is a way
François Chollet (59:02.460)
to encode perception and a way to encode intuition.
Lex Fridman (59:05.740)
But it is not a good medium for any sort of explicit reasoning.
Lex Fridman (59:11.100)
And in AI systems today, strong generalization
Lex Fridman (59:15.940)
tends to come from explicit models,
François Chollet (59:21.020)
tend to come from abstractions in the human mind that
Lex Fridman (59:24.540)
are encoded in program form by a human engineer.
François Chollet (59:29.540)
These are the abstractions you can actually generalize, not
Lex Fridman (59:31.580)
the sort of weak abstraction that
François Chollet (59:33.380)
is learned by a neural network.
Lex Fridman (59:34.860)
Yeah, and the question is how much reasoning,
Lex Fridman (59:38.540)
how much strong abstractions are required
Lex Fridman (59:41.940)
to solve particular tasks like driving.
François Chollet (59:44.620)
That's the question.
Lex Fridman (59:46.540)
Or human life existence.
Lex Fridman (59:48.860)
How much strong abstractions does existence require?
Lex Fridman (59:53.340)
But more specifically on driving,
François Chollet (59:58.100)
that seems to be a coupled question about intelligence.
🔗 相关节目