François Chollet: Keras, Deep Learning, and the Progress of AI
AI 与机器学习心理与人性生物与进化技术与编程音乐与艺术
📋 章节目录
暂无章节信息
🔑 关键词
intelligencehumanlearningdeepdonsciencegoingdatapriorsproblemstryingspaceknowledgeprogresstensorflowbrainexperiencealgorithmsinstancekeras
💬 精彩语录
"And then, you know, the higher level of information you're trying to write, the longer it's going to take."
然后，您知道，您尝试编写的信息级别越高，所需的时间就越长。
— François Chollet （1:45:57.440）
🎙️ 完整对话（2301 条）
Lex Fridman (00:00.000)
The following is a conversation with Francois Chollet.
以下是与弗朗索瓦·肖莱的对话。
Lex Fridman (00:03.720)
He's the creator of Keras,
他是 Keras 的创造者，
Lex Fridman (00:05.760)
which is an open source deep learning library
这是一个开源深度学习库
Lex Fridman (00:08.080)
that is designed to enable fast, user friendly experimentation
旨在实现快速、用户友好的实验
Lex Fridman (00:11.480)
with deep neural networks.
与深度神经网络。
François Chollet (00:13.600)
It serves as an interface to several deep learning libraries,
它充当多个深度学习库的接口，
Lex Fridman (00:16.680)
most popular of which is TensorFlow,
其中最受欢迎的是 TensorFlow，
Lex Fridman (00:19.040)
and it was integrated into the TensorFlow main code base
并已集成到 TensorFlow 主代码库中
Lex Fridman (00:22.600)
a while ago.
不久前。
François Chollet (00:24.080)
Meaning, if you want to create, train,
意思是，如果你想创造、训练、
Lex Fridman (00:27.000)
and use neural networks,
并使用神经网络，
François Chollet (00:28.640)
probably the easiest and most popular option
可能是最简单和最受欢迎的选择
Lex Fridman (00:31.040)
is to use Keras inside TensorFlow.
就是在TensorFlow里面使用Keras。
François Chollet (00:34.840)
Aside from creating an exceptionally useful
除了创建一个非常有用的
Lex Fridman (00:37.240)
and popular library,
和流行的图书馆，
François Chollet (00:38.680)
Francois is also a world class AI researcher
Francois 也是一位世界级的人工智能研究员
Lex Fridman (00:41.920)
and software engineer at Google.
和谷歌的软件工程师。
Lex Fridman (00:44.560)
And he's definitely an outspoken,
而且他绝对是一个直言不讳的人
Lex Fridman (00:46.960)
if not controversial personality in the AI world,
如果不是人工智能界有争议的个性，
François Chollet (00:50.560)
especially in the realm of ideas
尤其是在思想领域
Lex Fridman (00:52.920)
around the future of artificial intelligence.
François Chollet (00:55.920)
This is the Artificial Intelligence Podcast.
Lex Fridman (00:58.600)
If you enjoy it, subscribe on YouTube,
François Chollet (01:01.000)
give it five stars on iTunes,
Lex Fridman (01:02.760)
support it on Patreon,
François Chollet (01:04.160)
or simply connect with me on Twitter
Lex Fridman (01:06.120)
at Lex Friedman, spelled F R I D M A N.
Lex Fridman (01:09.960)
And now, here's my conversation with Francois Chollet.
Lex Fridman (01:14.880)
You're known for not sugarcoating your opinions
Lex Fridman (01:17.320)
and speaking your mind about ideas in AI,
Lex Fridman (01:19.160)
especially on Twitter.
François Chollet (01:21.160)
It's one of my favorite Twitter accounts.
Lex Fridman (01:22.760)
So what's one of the more controversial ideas
Lex Fridman (01:26.320)
you've expressed online and gotten some heat for?
Lex Fridman (01:30.440)
How do you pick?
Lex Fridman (01:33.080)
How do I pick?
Lex Fridman (01:33.920)
Yeah, no, I think if you go through the trouble
François Chollet (01:36.880)
of maintaining a Twitter account,
Lex Fridman (01:39.640)
you might as well speak your mind, you know?
Lex Fridman (01:41.840)
Otherwise, what's even the point of having a Twitter account?
Lex Fridman (01:44.600)
It's like having a nice car
Lex Fridman (01:45.480)
and just leaving it in the garage.
Lex Fridman (01:48.600)
Yeah, so what's one thing for which I got
Lex Fridman (01:50.840)
a lot of pushback?
Lex Fridman (01:53.600)
Perhaps, you know, that time I wrote something
François Chollet (01:56.640)
about the idea of intelligence explosion,
Lex Fridman (02:00.920)
and I was questioning the idea
Lex Fridman (02:04.520)
and the reasoning behind this idea.
Lex Fridman (02:06.840)
And I got a lot of pushback on that.
François Chollet (02:09.640)
I got a lot of flak for it.
Lex Fridman (02:11.840)
So yeah, so intelligence explosion,
François Chollet (02:13.600)
I'm sure you're familiar with the idea,
Lex Fridman (02:14.960)
but it's the idea that if you were to build
François Chollet (02:18.800)
general AI problem solving algorithms,
Lex Fridman (02:22.920)
well, the problem of building such an AI,
François Chollet (02:27.480)
that itself is a problem that could be solved by your AI,
Lex Fridman (02:30.520)
and maybe it could be solved better
François Chollet (02:31.880)
than what humans can do.
Lex Fridman (02:33.760)
So your AI could start tweaking its own algorithm,
François Chollet (02:36.840)
could start making a better version of itself,
Lex Fridman (02:39.520)
and so on iteratively in a recursive fashion.
Lex Fridman (02:43.240)
And so you would end up with an AI
Lex Fridman (02:47.320)
with exponentially increasing intelligence.
François Chollet (02:50.080)
That's right.
Lex Fridman (02:50.920)
And I was basically questioning this idea,
François Chollet (02:55.880)
first of all, because the notion of intelligence explosion
Lex Fridman (02:59.040)
uses an implicit definition of intelligence
François Chollet (03:02.200)
that doesn't sound quite right to me.
Lex Fridman (03:05.360)
It considers intelligence as a property of a brain
François Chollet (03:11.200)
that you can consider in isolation,
Lex Fridman (03:13.680)
like the height of a building, for instance.
Lex Fridman (03:16.640)
But that's not really what intelligence is.
Lex Fridman (03:19.040)
Intelligence emerges from the interaction
François Chollet (03:22.200)
between a brain, a body,
Lex Fridman (03:25.240)
like embodied intelligence, and an environment.
Lex Fridman (03:28.320)
And if you're missing one of these pieces,
Lex Fridman (03:30.720)
then you cannot really define intelligence anymore.
Lex Fridman (03:33.800)
So just tweaking a brain to make it smaller and smaller
Lex Fridman (03:36.800)
doesn't actually make any sense to me.
Lex Fridman (03:39.120)
So first of all,
Lex Fridman (03:39.960)
you're crushing the dreams of many people, right?
Lex Fridman (03:43.000)
So there's a, let's look at like Sam Harris.
Lex Fridman (03:46.000)
Actually, a lot of physicists, Max Tegmark,
François Chollet (03:48.680)
people who think the universe
Lex Fridman (03:52.120)
is an information processing system,
François Chollet (03:54.640)
our brain is kind of an information processing system.
Lex Fridman (03:57.680)
So what's the theoretical limit?
François Chollet (03:59.400)
Like, it doesn't make sense that there should be some,
Lex Fridman (04:04.800)
it seems naive to think that our own brain
François Chollet (04:07.520)
is somehow the limit of the capabilities
Lex Fridman (04:10.000)
of this information system.
François Chollet (04:11.600)
I'm playing devil's advocate here.
Lex Fridman (04:13.600)
This information processing system.
Lex Fridman (04:15.600)
And then if you just scale it,
Lex Fridman (04:17.760)
if you're able to build something
François Chollet (04:19.360)
that's on par with the brain,
Lex Fridman (04:20.920)
you just, the process that builds it just continues
Lex Fridman (04:24.040)
and it'll improve exponentially.
Lex Fridman (04:26.400)
So that's the logic that's used actually
François Chollet (04:30.160)
by almost everybody
Lex Fridman (04:32.560)
that is worried about super human intelligence.
Lex Fridman (04:36.920)
So you're trying to make,
Lex Fridman (04:39.120)
so most people who are skeptical of that
François Chollet (04:40.960)
are kind of like, this doesn't,
Lex Fridman (04:43.000)
their thought process, this doesn't feel right.
François Chollet (04:46.520)
Like that's for me as well.
Lex Fridman (04:47.680)
So I'm more like, it doesn't,
François Chollet (04:51.440)
the whole thing is shrouded in mystery
Lex Fridman (04:52.800)
where you can't really say anything concrete,
Lex Fridman (04:55.840)
but you could say this doesn't feel right.
Lex Fridman (04:57.880)
This doesn't feel like that's how the brain works.
Lex Fridman (05:00.640)
And you're trying to with your blog posts
Lex Fridman (05:02.400)
and now making it a little more explicit.
Lex Fridman (05:05.680)
So one idea is that the brain isn't exist alone.
Lex Fridman (05:10.680)
It exists within the environment.
Lex Fridman (05:13.200)
So you can't exponentially,
Lex Fridman (05:15.680)
you would have to somehow exponentially improve
François Chollet (05:18.000)
the environment and the brain together almost.
Lex Fridman (05:20.920)
Yeah, in order to create something that's much smarter
François Chollet (05:25.960)
in some kind of,
Lex Fridman (05:27.840)
of course we don't have a definition of intelligence.
François Chollet (05:29.960)
That's correct, that's correct.
Lex Fridman (05:31.280)
I don't think, you should look at very smart people today,
François Chollet (05:34.280)
even humans, not even talking about AIs.
Lex Fridman (05:37.280)
I don't think their brain
Lex Fridman (05:38.640)
and the performance of their brain is the bottleneck
Lex Fridman (05:41.960)
to their expressed intelligence, to their achievements.
François Chollet (05:46.600)
You cannot just tweak one part of this system,
Lex Fridman (05:49.960)
like of this brain, body, environment system
Lex Fridman (05:52.840)
and expect that capabilities like what emerges
Lex Fridman (05:55.960)
out of this system to just explode exponentially.
François Chollet (06:00.280)
Because anytime you improve one part of a system
Lex Fridman (06:04.200)
with many interdependencies like this,
Lex Fridman (06:06.760)
there's a new bottleneck that arises, right?
Lex Fridman (06:09.520)
And I don't think even today for very smart people,
François Chollet (06:12.280)
their brain is not the bottleneck
Lex Fridman (06:15.000)
to the sort of problems they can solve, right?
François Chollet (06:17.560)
In fact, many very smart people today,
Lex Fridman (06:20.760)
you know, they are not actually solving
François Chollet (06:22.520)
any big scientific problems, they're not Einstein.
Lex Fridman (06:24.800)
They're like Einstein, but you know, the patent clerk days.
François Chollet (06:29.800)
Like Einstein became Einstein
Lex Fridman (06:31.920)
because this was a meeting of a genius
Lex Fridman (06:36.080)
with a big problem at the right time, right?
Lex Fridman (06:39.480)
But maybe this meeting could have never happened
Lex Fridman (06:42.480)
and then Einstein would have just been a patent clerk, right?
Lex Fridman (06:44.960)
And in fact, many people today are probably like
François Chollet (06:49.760)
genius level smart, but you wouldn't know
Lex Fridman (06:52.240)
because they're not really expressing any of that.
François Chollet (06:54.800)
Wow, that's brilliant.
Lex Fridman (06:55.640)
So we can think of the world, Earth,
Lex Fridman (06:58.520)
but also the universe as just as a space of problems.
Lex Fridman (07:02.720)
So all these problems and tasks are roaming it
François Chollet (07:05.160)
of various difficulty.
Lex Fridman (07:06.880)
And there's agents, creatures like ourselves
Lex Fridman (07:10.120)
and animals and so on that are also roaming it.
Lex Fridman (07:13.360)
And then you get coupled with a problem
Lex Fridman (07:16.480)
and then you solve it.
Lex Fridman (07:17.640)
But without that coupling,
François Chollet (07:19.880)
you can't demonstrate your quote unquote intelligence.
Lex Fridman (07:22.560)
Exactly, intelligence is the meeting
François Chollet (07:24.480)
of great problem solving capabilities
Lex Fridman (07:27.480)
with a great problem.
Lex Fridman (07:28.760)
And if you don't have the problem,
Lex Fridman (07:30.560)
you don't really express any intelligence.
François Chollet (07:32.280)
All you're left with is potential intelligence,
Lex Fridman (07:34.760)
like the performance of your brain
François Chollet (07:36.240)
or how high your IQ is,
Lex Fridman (07:38.680)
which in itself is just a number, right?
Lex Fridman (07:42.080)
So you mentioned problem solving capacity.
Lex Fridman (07:46.520)
Yeah.
Lex Fridman (07:47.360)
What do you think of as problem solving capacity?
Lex Fridman (07:51.800)
Can you try to define intelligence?
Lex Fridman (07:56.640)
Like what does it mean to be more or less intelligent?
Lex Fridman (08:00.000)
Is it completely coupled to a particular problem
Lex Fridman (08:03.000)
or is there something a little bit more universal?
Lex Fridman (08:05.720)
Yeah, I do believe all intelligence
François Chollet (08:07.440)
is specialized intelligence.
Lex Fridman (08:09.080)
Even human intelligence has some degree of generality.
François Chollet (08:12.200)
Well, all intelligent systems have some degree of generality
Lex Fridman (08:15.320)
but they're always specialized in one category of problems.
Lex Fridman (08:19.400)
So the human intelligence is specialized
Lex Fridman (08:21.880)
in the human experience.
Lex Fridman (08:23.560)
And that shows at various levels,
Lex Fridman (08:25.560)
that shows in some prior knowledge that's innate
François Chollet (08:30.200)
that we have at birth.
Lex Fridman (08:32.040)
Knowledge about things like agents,
François Chollet (08:35.360)
goal driven behavior, visual priors
Lex Fridman (08:38.080)
about what makes an object, priors about time and so on.
François Chollet (08:43.520)
That shows also in the way we learn.
Lex Fridman (08:45.360)
For instance, it's very, very easy for us
François Chollet (08:47.160)
to pick up language.
Lex Fridman (08:49.560)
It's very, very easy for us to learn certain things
François Chollet (08:52.080)
because we are basically hard coded to learn them.
Lex Fridman (08:54.920)
And we are specialized in solving certain kinds of problem
Lex Fridman (08:58.280)
and we are quite useless
Lex Fridman (08:59.720)
when it comes to other kinds of problems.
François Chollet (09:01.440)
For instance, we are not really designed
Lex Fridman (09:06.160)
to handle very long term problems.
François Chollet (09:08.800)
We have no capability of seeing the very long term.
Lex Fridman (09:12.880)
We don't have very much working memory.
Lex Fridman (09:18.000)
So how do you think about long term?
Lex Fridman (09:20.080)
Do you think long term planning,
Lex Fridman (09:21.360)
are we talking about scale of years, millennia?
Lex Fridman (09:24.880)
What do you mean by long term?
François Chollet (09:26.400)
We're not very good.
Lex Fridman (09:28.120)
Well, human intelligence is specialized
François Chollet (09:29.760)
in the human experience.
Lex Fridman (09:30.720)
And human experience is very short.
François Chollet (09:32.800)
One lifetime is short.
Lex Fridman (09:34.240)
Even within one lifetime,
François Chollet (09:35.880)
we have a very hard time envisioning things
Lex Fridman (09:40.000)
on a scale of years.
François Chollet (09:41.360)
It's very difficult to project yourself
Lex Fridman (09:43.240)
at a scale of five years, at a scale of 10 years and so on.
François Chollet (09:46.960)
We can solve only fairly narrowly scoped problems.
Lex Fridman (09:50.000)
So when it comes to solving bigger problems,
François Chollet (09:52.320)
larger scale problems,
Lex Fridman (09:53.760)
we are not actually doing it on an individual level.
Lex Fridman (09:56.360)
So it's not actually our brain doing it.
Lex Fridman (09:59.280)
We have this thing called civilization, right?
François Chollet (10:03.040)
Which is itself a sort of problem solving system,
Lex Fridman (10:06.600)
a sort of artificially intelligent system, right?
Lex Fridman (10:10.000)
And it's not running on one brain,
Lex Fridman (10:12.120)
it's running on a network of brains.
François Chollet (10:14.080)
In fact, it's running on much more
Lex Fridman (10:15.640)
than a network of brains.
François Chollet (10:16.760)
It's running on a lot of infrastructure,
Lex Fridman (10:20.080)
like books and computers and the internet
Lex Fridman (10:23.040)
and human institutions and so on.
Lex Fridman (10:25.800)
And that is capable of handling problems
François Chollet (10:30.240)
on a much greater scale than any individual human.
Lex Fridman (10:33.760)
If you look at computer science, for instance,
François Chollet (10:37.600)
that's an institution that solves problems
Lex Fridman (10:39.840)
and it is superhuman, right?
François Chollet (10:42.560)
It operates on a greater scale.
Lex Fridman (10:44.200)
It can solve much bigger problems
François Chollet (10:46.880)
than an individual human could.
Lex Fridman (10:49.080)
And science itself, science as a system, as an institution,
François Chollet (10:52.160)
is a kind of artificially intelligent problem solving
Lex Fridman (10:57.120)
algorithm that is superhuman.
François Chollet (10:59.360)
Yeah, it's, at least computer science
Lex Fridman (11:02.800)
is like a theorem prover at a scale of thousands,
François Chollet (11:07.720)
maybe hundreds of thousands of human beings.
Lex Fridman (11:10.400)
At that scale, what do you think is an intelligent agent?
Lex Fridman (11:14.680)
So there's us humans at the individual level,
Lex Fridman (11:18.280)
there is millions, maybe billions of bacteria in our skin.
François Chollet (11:23.880)
There is, that's at the smaller scale.
Lex Fridman (11:26.400)
You can even go to the particle level
François Chollet (11:29.160)
as systems that behave,
Lex Fridman (11:31.840)
you can say intelligently in some ways.
Lex Fridman (11:35.440)
And then you can look at the earth as a single organism,
Lex Fridman (11:37.840)
you can look at our galaxy
Lex Fridman (11:39.200)
and even the universe as a single organism.
Lex Fridman (11:42.160)
Do you think, how do you think about scale
Lex Fridman (11:44.680)
in defining intelligent systems?
Lex Fridman (11:46.280)
And we're here at Google, there is millions of devices
François Chollet (11:50.440)
doing computation just in a distributed way.
Lex Fridman (11:53.360)
How do you think about intelligence versus scale?
François Chollet (11:55.880)
You can always characterize anything as a system.
Lex Fridman (12:00.640)
I think people who talk about things
François Chollet (12:03.600)
like intelligence explosion,
Lex Fridman (12:05.320)
tend to focus on one agent is basically one brain,
François Chollet (12:08.760)
like one brain considered in isolation,
Lex Fridman (12:10.960)
like a brain, a jaw that's controlling a body
François Chollet (12:13.200)
in a very like top to bottom kind of fashion.
Lex Fridman (12:16.280)
And that body is pursuing goals into an environment.
Lex Fridman (12:19.480)
So it's a very hierarchical view.
Lex Fridman (12:20.720)
You have the brain at the top of the pyramid,
François Chollet (12:22.880)
then you have the body just plainly receiving orders.
Lex Fridman (12:25.960)
And then the body is manipulating objects
François Chollet (12:27.640)
in the environment and so on.
Lex Fridman (12:28.920)
So everything is subordinate to this one thing,
François Chollet (12:32.920)
this epicenter, which is the brain.
Lex Fridman (12:34.720)
But in real life, intelligent agents
Lex Fridman (12:37.120)
don't really work like this, right?
Lex Fridman (12:39.240)
There is no strong delimitation
François Chollet (12:40.920)
between the brain and the body to start with.
Lex Fridman (12:43.400)
You have to look not just at the brain,
Lex Fridman (12:45.000)
but at the nervous system.
Lex Fridman (12:46.560)
But then the nervous system and the body
François Chollet (12:48.840)
are naturally two separate entities.
Lex Fridman (12:50.760)
So you have to look at an entire animal as one agent.
Lex Fridman (12:53.960)
But then you start realizing as you observe an animal
Lex Fridman (12:57.000)
over any length of time,
François Chollet (13:00.200)
that a lot of the intelligence of an animal
Lex Fridman (13:03.160)
is actually externalized.
François Chollet (13:04.600)
That's especially true for humans.
Lex Fridman (13:06.240)
A lot of our intelligence is externalized.
François Chollet (13:08.880)
When you write down some notes,
Lex Fridman (13:10.360)
that is externalized intelligence.
François Chollet (13:11.960)
When you write a computer program,
Lex Fridman (13:14.000)
you are externalizing cognition.
Lex Fridman (13:16.000)
So it's externalizing books, it's externalized in computers,
Lex Fridman (13:19.720)
the internet, in other humans.
François Chollet (13:23.080)
It's externalizing language and so on.
Lex Fridman (13:25.400)
So there is no hard delimitation
François Chollet (13:30.480)
of what makes an intelligent agent.
Lex Fridman (13:32.640)
It's all about context.
François Chollet (13:34.960)
Okay, but AlphaGo is better at Go
Lex Fridman (13:38.800)
than the best human player.
François Chollet (13:42.520)
There's levels of skill here.
Lex Fridman (13:45.000)
So do you think there's such a ability,
François Chollet (13:48.600)
such a concept as intelligence explosion
Lex Fridman (13:52.800)
in a specific task?
Lex Fridman (13:54.760)
And then, well, yeah.
Lex Fridman (13:57.360)
Do you think it's possible to have a category of tasks
François Chollet (14:00.120)
on which you do have something
Lex Fridman (14:02.080)
like an exponential growth of ability
Lex Fridman (14:05.040)
to solve that particular problem?
Lex Fridman (14:07.440)
I think if you consider a specific vertical,
François Chollet (14:10.320)
it's probably possible to some extent.
Lex Fridman (14:15.320)
I also don't think we have to speculate about it
François Chollet (14:18.320)
because we have real world examples
Lex Fridman (14:22.280)
of recursively self improving intelligent systems, right?
Lex Fridman (14:26.920)
So for instance, science is a problem solving system,
Lex Fridman (14:30.920)
a knowledge generation system,
François Chollet (14:32.600)
like a system that experiences the world in some sense
Lex Fridman (14:36.240)
and then gradually understands it and can act on it.
Lex Fridman (14:40.160)
And that system is superhuman
Lex Fridman (14:42.120)
and it is clearly recursively self improving
François Chollet (14:45.600)
because science feeds into technology.
Lex Fridman (14:47.560)
Technology can be used to build better tools,
François Chollet (14:50.200)
better computers, better instrumentation and so on,
Lex Fridman (14:52.880)
which in turn can make science faster, right?
Lex Fridman (14:56.720)
So science is probably the closest thing we have today
Lex Fridman (15:00.560)
to a recursively self improving superhuman AI.
Lex Fridman (15:04.760)
And you can just observe is science,
Lex Fridman (15:08.520)
is scientific progress to the exploding,
François Chollet (15:10.320)
which itself is an interesting question.
Lex Fridman (15:12.800)
You can use that as a basis to try to understand
Lex Fridman (15:15.560)
what will happen with a superhuman AI
Lex Fridman (15:17.920)
that has a science like behavior.
François Chollet (15:21.000)
Let me linger on it a little bit more.
Lex Fridman (15:23.320)
What is your intuition why an intelligence explosion
Lex Fridman (15:27.600)
is not possible?
Lex Fridman (15:28.560)
Like taking the scientific,
François Chollet (15:30.920)
all the semi scientific revolutions,
Lex Fridman (15:33.240)
why can't we slightly accelerate that process?
Lex Fridman (15:38.080)
So you can absolutely accelerate
Lex Fridman (15:41.200)
any problem solving process.
Lex Fridman (15:43.120)
So a recursively self improvement
Lex Fridman (15:46.720)
is absolutely a real thing.
Lex Fridman (15:48.640)
But what happens with a recursively self improving system
Lex Fridman (15:51.880)
is typically not explosion
François Chollet (15:53.680)
because no system exists in isolation.
Lex Fridman (15:56.520)
And so tweaking one part of the system
François Chollet (15:58.640)
means that suddenly another part of the system
Lex Fridman (16:00.880)
becomes a bottleneck.
Lex Fridman (16:02.200)
And if you look at science, for instance,
Lex Fridman (16:03.800)
which is clearly a recursively self improving,
François Chollet (16:06.800)
clearly a problem solving system,
Lex Fridman (16:09.040)
scientific progress is not actually exploding.
François Chollet (16:12.000)
If you look at science,
Lex Fridman (16:13.520)
what you see is the picture of a system
François Chollet (16:16.480)
that is consuming an exponentially increasing
Lex Fridman (16:19.240)
amount of resources,
Lex Fridman (16:20.520)
but it's having a linear output
Lex Fridman (16:23.960)
in terms of scientific progress.
Lex Fridman (16:26.000)
And maybe that will seem like a very strong claim.
Lex Fridman (16:28.960)
Many people are actually saying that,
François Chollet (16:31.160)
scientific progress is exponential,
Lex Fridman (16:34.560)
but when they're claiming this,
François Chollet (16:36.120)
they're actually looking at indicators
Lex Fridman (16:38.400)
of resource consumption by science.
François Chollet (16:43.080)
For instance, the number of papers being published,
Lex Fridman (16:47.560)
the number of patents being filed and so on,
François Chollet (16:49.960)
which are just completely correlated
Lex Fridman (16:53.600)
with how many people are working on science today.
Lex Fridman (16:58.480)
So it's actually an indicator of resource consumption,
Lex Fridman (17:00.640)
but what you should look at is the output,
François Chollet (17:03.200)
is progress in terms of the knowledge
Lex Fridman (17:06.680)
that science generates,
François Chollet (17:08.040)
in terms of the scope and significance
Lex Fridman (17:10.640)
of the problems that we solve.
Lex Fridman (17:12.520)
And some people have actually been trying to measure that.
Lex Fridman (17:16.720)
Like Michael Nielsen, for instance,
François Chollet (17:20.160)
he had a very nice paper,
Lex Fridman (17:21.920)
I think that was last year about it.
Lex Fridman (17:25.200)
So his approach to measure scientific progress
Lex Fridman (17:28.360)
was to look at the timeline of scientific discoveries
François Chollet (17:33.720)
over the past, you know, 100, 150 years.
Lex Fridman (17:37.160)
And for each major discovery,
François Chollet (17:41.360)
ask a panel of experts to rate
Lex Fridman (17:44.360)
the significance of the discovery.
Lex Fridman (17:46.760)
And if the output of science as an institution
Lex Fridman (17:49.600)
were exponential,
François Chollet (17:50.440)
you would expect the temporal density of significance
Lex Fridman (17:56.600)
to go up exponentially.
François Chollet (17:58.160)
Maybe because there's a faster rate of discoveries,
Lex Fridman (18:00.960)
maybe because the discoveries are, you know,
François Chollet (18:02.960)
increasingly more important.
Lex Fridman (18:04.920)
And what actually happens
François Chollet (18:06.800)
if you plot this temporal density of significance
Lex Fridman (18:10.040)
measured in this way,
François Chollet (18:11.320)
is that you see very much a flat graph.
Lex Fridman (18:14.520)
You see a flat graph across all disciplines,
François Chollet (18:16.600)
across physics, biology, medicine, and so on.
Lex Fridman (18:19.720)
And it actually makes a lot of sense
François Chollet (18:22.480)
if you think about it,
Lex Fridman (18:23.320)
because think about the progress of physics
Lex Fridman (18:26.000)
110 years ago, right?
Lex Fridman (18:28.000)
It was a time of crazy change.
François Chollet (18:30.080)
Think about the progress of technology,
Lex Fridman (18:31.960)
you know, 170 years ago,
François Chollet (18:34.360)
when we started having, you know,
Lex Fridman (18:35.400)
replacing horses with cars,
François Chollet (18:37.560)
when we started having electricity and so on.
Lex Fridman (18:40.000)
It was a time of incredible change.
Lex Fridman (18:41.520)
And today is also a time of very, very fast change,
Lex Fridman (18:44.600)
but it would be an unfair characterization
François Chollet (18:48.040)
to say that today technology and science
Lex Fridman (18:50.560)
are moving way faster than they did 50 years ago
François Chollet (18:52.920)
or 100 years ago.
Lex Fridman (18:54.360)
And if you do try to rigorously plot
François Chollet (18:59.520)
the temporal density of the significance,
Lex Fridman (19:04.880)
yeah, of significance, sorry,
François Chollet (19:07.320)
you do see very flat curves.
Lex Fridman (19:09.720)
And you can check out the paper
François Chollet (19:12.040)
that Michael Nielsen had about this idea.
Lex Fridman (19:16.000)
And so the way I interpret it is,
François Chollet (19:20.000)
as you make progress in a given field,
Lex Fridman (19:24.160)
or in a given subfield of science,
François Chollet (19:26.120)
it becomes exponentially more difficult
Lex Fridman (19:28.680)
to make further progress.
François Chollet (19:30.440)
Like the very first person to work on information theory.
Lex Fridman (19:35.000)
If you enter a new field,
Lex Fridman (19:36.440)
and it's still the very early years,
Lex Fridman (19:37.920)
there's a lot of low hanging fruit you can pick.
François Chollet (19:41.160)
That's right, yeah.
Lex Fridman (19:42.000)
But the next generation of researchers
François Chollet (19:43.960)
is gonna have to dig much harder, actually,
Lex Fridman (19:48.160)
to make smaller discoveries,
François Chollet (19:50.640)
probably larger number of smaller discoveries,
Lex Fridman (19:52.640)
and to achieve the same amount of impact,
François Chollet (19:54.640)
you're gonna need a much greater head count.
Lex Fridman (19:57.480)
And that's exactly the picture you're seeing with science,
François Chollet (1:00:00.120)
over rule based models is going to be
Lex Fridman (1:00:02.760)
a cornerstone of AI research in the next century.
Lex Fridman (1:00:06.640)
And that doesn't mean we are going to drop deep learning.
Lex Fridman (1:00:10.200)
Deep learning is immensely useful.
François Chollet (1:00:11.880)
Like, being able to learn is a very flexible, adaptable,
Lex Fridman (1:00:17.200)
parametric model.
Lex Fridman (1:00:18.120)
So it's got to understand that's actually immensely useful.
Lex Fridman (1:00:20.720)
All it's doing is pattern cognition.
Lex Fridman (1:00:23.040)
But being good at pattern cognition, given lots of data,
Lex Fridman (1:00:25.640)
is just extremely powerful.
Lex Fridman (1:00:27.920)
So we are still going to be working on deep learning.
Lex Fridman (1:00:30.320)
We are going to be working on program synthesis.
François Chollet (1:00:31.840)
We are going to be combining the two in increasingly automated
Lex Fridman (1:00:34.680)
ways.
Lex Fridman (1:00:36.400)
So let's talk a little bit about data.
Lex Fridman (1:00:38.520)
You've tweeted, about 10,000 deep learning papers
François Chollet (1:00:44.600)
have been written about hard coding priors
Lex Fridman (1:00:47.080)
about a specific task in a neural network architecture
François Chollet (1:00:49.600)
works better than a lack of a prior.
Lex Fridman (1:00:52.440)
Basically, summarizing all these efforts,
François Chollet (1:00:55.120)
they put a name to an architecture.
Lex Fridman (1:00:56.920)
But really, what they're doing is hard coding some priors
François Chollet (1:00:59.280)
that improve the performance of the system.
Lex Fridman (1:01:01.560)
But which gets straight to the point is probably true.
Lex Fridman (1:01:06.880)
So you say that you can always buy performance by,
Lex Fridman (1:01:09.800)
in quotes, performance by either training on more data,
François Chollet (1:01:12.920)
better data, or by injecting task information
Lex Fridman (1:01:15.480)
to the architecture of the preprocessing.
François Chollet (1:01:18.400)
However, this isn't informative about the generalization power
Lex Fridman (1:01:21.280)
the techniques use, the fundamental ability
François Chollet (1:01:23.080)
to generalize.
Lex Fridman (1:01:24.200)
Do you think we can go far by coming up
François Chollet (1:01:26.800)
with better methods for this kind of cheating,
Lex Fridman (1:01:29.920)
for better methods of large scale annotation of data?
Lex Fridman (1:01:33.520)
So building better priors.
Lex Fridman (1:01:34.960)
If you automate it, it's not cheating anymore.
François Chollet (1:01:37.280)
Right.
Lex Fridman (1:01:38.360)
I'm joking about the cheating, but large scale.
Lex Fridman (1:01:41.600)
So basically, I'm asking about something
Lex Fridman (1:01:46.560)
that hasn't, from my perspective,
François Chollet (1:01:48.280)
been researched too much is exponential improvement
Lex Fridman (1:01:53.360)
in annotation of data.
Lex Fridman (1:01:55.960)
Do you often think about?
Lex Fridman (1:01:58.120)
I think it's actually been researched quite a bit.
François Chollet (1:02:00.840)
You just don't see publications about it.
Lex Fridman (1:02:02.720)
Because people who publish papers
François Chollet (1:02:05.840)
are going to publish about known benchmarks.
Lex Fridman (1:02:07.920)
Sometimes they're going to read a new benchmark.
François Chollet (1:02:09.800)
People who actually have real world large scale
Lex Fridman (1:02:12.200)
depending on problems, they're going
François Chollet (1:02:13.880)
to spend a lot of resources into data annotation
Lex Fridman (1:02:16.960)
and good data annotation pipelines,
Lex Fridman (1:02:18.400)
but you don't see any papers about it.
Lex Fridman (1:02:19.640)
That's interesting.
Lex Fridman (1:02:20.400)
So do you think, certainly resources,
Lex Fridman (1:02:22.720)
but do you think there's innovation happening?
François Chollet (1:02:24.840)
Oh, yeah.
Lex Fridman (1:02:25.880)
To clarify the point in the tweet.
Lex Fridman (1:02:28.880)
So machine learning in general is
Lex Fridman (1:02:31.160)
the science of generalization.
François Chollet (1:02:33.840)
You want to generate knowledge that
Lex Fridman (1:02:37.800)
can be reused across different data sets,
François Chollet (1:02:40.440)
across different tasks.
Lex Fridman (1:02:42.000)
And if instead you're looking at one data set
Lex Fridman (1:02:45.280)
and then you are hard coding knowledge about this task
Lex Fridman (1:02:50.000)
into your architecture, this is no more useful
François Chollet (1:02:54.040)
than training a network and then saying, oh, I
Lex Fridman (1:02:56.760)
found these weight values perform well.
Lex Fridman (1:03:01.920)
So David Ha, I don't know if you know David,
Lex Fridman (1:03:05.680)
he had a paper the other day about weight
François Chollet (1:03:08.760)
agnostic neural networks.
Lex Fridman (1:03:10.400)
And this is a very interesting paper
François Chollet (1:03:12.120)
because it really illustrates the fact
Lex Fridman (1:03:14.400)
that an architecture, even without weights,
François Chollet (1:03:17.400)
an architecture is knowledge about a task.
Lex Fridman (1:03:21.360)
It encodes knowledge.
Lex Fridman (1:03:23.640)
And when it comes to architectures
Lex Fridman (1:03:25.840)
that are uncrafted by researchers, in some cases,
François Chollet (1:03:30.440)
it is very, very clear that all they are doing
Lex Fridman (1:03:34.160)
is artificially reencoding the template that
François Chollet (1:03:38.880)
corresponds to the proper way to solve the task encoding
Lex Fridman (1:03:44.400)
a given data set.
François Chollet (1:03:45.200)
For instance, I know if you looked
Lex Fridman (1:03:48.120)
at the baby data set, which is about natural language
François Chollet (1:03:52.280)
question answering, it is generated by an algorithm.
Lex Fridman (1:03:55.520)
So this is a question answer pairs
François Chollet (1:03:57.680)
that are generated by an algorithm.
Lex Fridman (1:03:59.280)
The algorithm is solving a certain template.
François Chollet (1:04:01.520)
Turns out, if you craft a network that
Lex Fridman (1:04:04.400)
literally encodes this template, you
François Chollet (1:04:06.360)
can solve this data set with nearly 100% accuracy.
Lex Fridman (1:04:09.640)
But that doesn't actually tell you
François Chollet (1:04:11.160)
anything about how to solve question answering
Lex Fridman (1:04:14.640)
in general, which is the point.
François Chollet (1:04:17.680)
The question is just to linger on it,
Lex Fridman (1:04:19.400)
whether it's from the data side or from the size
François Chollet (1:04:21.560)
of the network.
Lex Fridman (1:04:23.280)
I don't know if you've read the blog post by Rich Sutton,
François Chollet (1:04:25.920)
The Bitter Lesson, where he says,
Lex Fridman (1:04:28.400)
the biggest lesson that we can read from 70 years of AI
François Chollet (1:04:31.480)
research is that general methods that leverage computation
Lex Fridman (1:04:34.720)
are ultimately the most effective.
Lex Fridman (1:04:37.160)
So as opposed to figuring out methods
Lex Fridman (1:04:39.720)
that can generalize effectively, do you
François Chollet (1:04:41.840)
think we can get pretty far by just having something
Lex Fridman (1:04:47.720)
that leverages computation and the improvement of computation?
François Chollet (1:04:51.520)
Yeah, so I think Rich is making a very good point, which
Lex Fridman (1:04:54.960)
is that a lot of these papers, which are actually
François Chollet (1:04:57.560)
all about manually hardcoding prior knowledge about a task
Lex Fridman (1:05:02.800)
into some system, it doesn't have
François Chollet (1:05:04.720)
to be deep learning architecture, but into some system.
Lex Fridman (1:05:08.600)
These papers are not actually making any impact.
François Chollet (1:05:11.920)
Instead, what's making really long term impact
Lex Fridman (1:05:14.800)
is very simple, very general systems
François Chollet (1:05:18.520)
that are really agnostic to all these tricks.
Lex Fridman (1:05:21.280)
Because these tricks do not generalize.
Lex Fridman (1:05:23.320)
And of course, the one general and simple thing
Lex Fridman (1:05:27.480)
that you should focus on is that which leverages computation.
François Chollet (1:05:33.160)
Because computation, the availability
Lex Fridman (1:05:36.200)
of large scale computation has been increasing exponentially
François Chollet (1:05:39.400)
following Moore's law.
Lex Fridman (1:05:40.560)
So if your algorithm is all about exploiting this,
François Chollet (1:05:44.080)
then your algorithm is suddenly exponentially improving.
Lex Fridman (1:05:47.440)
So I think Rich is definitely right.
François Chollet (1:05:52.400)
However, he's right about the past 70 years.
Lex Fridman (1:05:57.120)
He's like assessing the past 70 years.
François Chollet (1:05:59.440)
I am not sure that this assessment will still
Lex Fridman (1:06:02.360)
hold true for the next 70 years.
François Chollet (1:06:04.880)
It might to some extent.
Lex Fridman (1:06:07.160)
I suspect it will not.
François Chollet (1:06:08.560)
Because the truth of his assessment
Lex Fridman (1:06:11.560)
is a function of the context in which this research took place.
Lex Fridman (1:06:16.800)
And the context is changing.
Lex Fridman (1:06:18.600)
Moore's law might not be applicable anymore,
François Chollet (1:06:21.440)
for instance, in the future.
Lex Fridman (1:06:23.760)
And I do believe that when you tweak one aspect of a system,
François Chollet (1:06:31.200)
when you exploit one aspect of a system,
Lex Fridman (1:06:32.920)
some other aspect starts becoming the bottleneck.
François Chollet (1:06:36.480)
Let's say you have unlimited computation.
Lex Fridman (1:06:38.800)
Well, then data is the bottleneck.
Lex Fridman (1:06:41.440)
And I think we are already starting
Lex Fridman (1:06:43.560)
to be in a regime where our systems are
Lex Fridman (1:06:45.720)
so large in scale and so data ingrained
Lex Fridman (1:06:48.120)
that data today and the quality of data
Lex Fridman (1:06:50.360)
and the scale of data is the bottleneck.
Lex Fridman (1:06:53.040)
And in this environment, the bitter lesson from Rich
François Chollet (1:06:58.160)
is not going to be true anymore.
Lex Fridman (1:07:00.800)
So I think we are going to move from a focus
François Chollet (1:07:03.960)
on a computation scale to focus on data efficiency.
Lex Fridman (1:07:09.840)
Data efficiency.
Lex Fridman (1:07:10.720)
So that's getting to the question of symbolic AI.
Lex Fridman (1:07:13.120)
But to linger on the deep learning approaches,
Lex Fridman (1:07:16.280)
do you have hope for either unsupervised learning
Lex Fridman (1:07:19.240)
or reinforcement learning, which are
François Chollet (1:07:23.280)
ways of being more data efficient in terms
Lex Fridman (1:07:28.120)
of the amount of data they need that required human annotation?
Lex Fridman (1:07:31.560)
So unsupervised learning and reinforcement learning
Lex Fridman (1:07:34.280)
are frameworks for learning, but they are not
François Chollet (1:07:36.640)
like any specific technique.
Lex Fridman (1:07:39.000)
So usually when people say reinforcement learning,
Lex Fridman (1:07:41.200)
what they really mean is deep reinforcement learning,
Lex Fridman (1:07:43.320)
which is like one approach which is actually very questionable.
François Chollet (1:07:47.440)
The question I was asking was unsupervised learning
Lex Fridman (1:07:50.920)
with deep neural networks and deep reinforcement learning.
François Chollet (1:07:54.680)
Well, these are not really data efficient
Lex Fridman (1:07:56.840)
because you're still leveraging these huge parametric models
François Chollet (1:08:00.520)
point by point with gradient descent.
Lex Fridman (1:08:03.720)
It is more efficient in terms of the number of annotations,
François Chollet (1:08:08.000)
the density of annotations you need.
Lex Fridman (1:08:09.520)
So the idea being to learn the latent space around which
François Chollet (1:08:13.840)
the data is organized and then map the sparse annotations
Lex Fridman (1:08:17.960)
into it.
Lex Fridman (1:08:18.760)
And sure, I mean, that's clearly a very good idea.
Lex Fridman (1:08:23.560)
It's not really a topic I would be working on,
Lex Fridman (1:08:26.080)
but it's clearly a good idea.
Lex Fridman (1:08:28.040)
So it would get us to solve some problems that?
François Chollet (1:08:31.760)
It will get us to incremental improvements
Lex Fridman (1:08:34.880)
in labeled data efficiency.
Lex Fridman (1:08:38.240)
Do you have concerns about short term or long term threats
Lex Fridman (1:08:43.520)
from AI, from artificial intelligence?
François Chollet (1:08:47.800)
Yes, definitely to some extent.
Lex Fridman (1:08:50.640)
And what's the shape of those concerns?
François Chollet (1:08:52.800)
This is actually something I've briefly written about.
Lex Fridman (1:08:56.880)
But the capabilities of deep learning technology
François Chollet (1:09:02.680)
can be used in many ways that are
Lex Fridman (1:09:05.200)
concerning from mass surveillance with things
François Chollet (1:09:09.760)
like facial recognition.
Lex Fridman (1:09:11.880)
In general, tracking lots of data about everyone
Lex Fridman (1:09:15.440)
and then being able to making sense of this data
Lex Fridman (1:09:18.920)
to do identification, to do prediction.
François Chollet (1:09:22.240)
That's concerning.
Lex Fridman (1:09:23.160)
That's something that's being very aggressively pursued
François Chollet (1:09:26.560)
by totalitarian states like China.
Lex Fridman (1:09:31.440)
One thing I am very much concerned about
François Chollet (1:09:34.000)
is that our lives are increasingly online,
Lex Fridman (1:09:40.640)
are increasingly digital, made of information,
François Chollet (1:09:43.280)
made of information consumption and information production,
Lex Fridman (1:09:48.080)
our digital footprint, I would say.
Lex Fridman (1:09:51.800)
And if you absorb all of this data
Lex Fridman (1:09:56.280)
and you are in control of where you consume information,
François Chollet (1:10:01.440)
social networks and so on, recommendation engines,
Lex Fridman (1:10:06.960)
then you can build a sort of reinforcement
François Chollet (1:10:10.200)
loop for human behavior.
Lex Fridman (1:10:13.760)
You can observe the state of your mind at time t.
François Chollet (1:10:18.360)
You can predict how you would react
Lex Fridman (1:10:21.080)
to different pieces of content, how
François Chollet (1:10:23.800)
to get you to move your mind in a certain direction.
Lex Fridman (1:10:27.000)
And then you can feed you the specific piece of content
François Chollet (1:10:33.160)
that would move you in a specific direction.
Lex Fridman (1:10:35.680)
And you can do this at scale in terms
François Chollet (1:10:41.800)
of doing it continuously in real time.
Lex Fridman (1:10:44.960)
You can also do it at scale in terms
François Chollet (1:10:46.440)
of scaling this to many, many people, to entire populations.
Lex Fridman (1:10:50.480)
So potentially, artificial intelligence,
François Chollet (1:10:53.840)
even in its current state, if you combine it
Lex Fridman (1:10:57.440)
with the internet, with the fact that all of our lives
François Chollet (1:11:01.760)
are moving to digital devices and digital information
Lex Fridman (1:11:05.120)
consumption and creation, what you get
François Chollet (1:11:08.720)
is the possibility to achieve mass manipulation of behavior
Lex Fridman (1:11:14.480)
and mass psychological control.
Lex Fridman (1:11:16.840)
And this is a very real possibility.
Lex Fridman (1:11:18.520)
Yeah, so you're talking about any kind of recommender system.
François Chollet (1:11:22.080)
Let's look at the YouTube algorithm, Facebook,
Lex Fridman (1:11:26.160)
anything that recommends content you should watch next.
Lex Fridman (1:11:29.720)
And it's fascinating to think that there's
Lex Fridman (1:11:32.960)
some aspects of human behavior that you can say a problem of,
Lex Fridman (1:11:41.120)
is this person hold Republican beliefs or Democratic beliefs?
Lex Fridman (1:11:45.400)
And this is a trivial, that's an objective function.
Lex Fridman (1:11:50.240)
And you can optimize, and you can measure,
Lex Fridman (1:11:52.600)
and you can turn everybody into a Republican
François Chollet (1:11:54.360)
or everybody into a Democrat.
Lex Fridman (1:11:56.080)
I do believe it's true.
Lex Fridman (1:11:57.840)
So the human mind is very, if you look at the human mind
Lex Fridman (1:12:03.680)
as a kind of computer program, it
François Chollet (1:12:05.320)
has a very large exploit surface.
Lex Fridman (1:12:07.560)
It has many, many vulnerabilities.
François Chollet (1:12:09.360)
Exploit surfaces, yeah.
Lex Fridman (1:12:10.840)
Ways you can control it.
François Chollet (1:12:13.520)
For instance, when it comes to your political beliefs,
Lex Fridman (1:12:16.680)
this is very much tied to your identity.
Lex Fridman (1:12:19.400)
So for instance, if I'm in control of your news feed
Lex Fridman (1:12:23.040)
on your favorite social media platforms,
François Chollet (1:12:26.000)
this is actually where you're getting your news from.
Lex Fridman (1:12:29.360)
And of course, I can choose to only show you
François Chollet (1:12:32.960)
news that will make you see the world in a specific way.
Lex Fridman (1:12:37.120)
But I can also create incentives for you
François Chollet (1:12:41.920)
to post about some political beliefs.
Lex Fridman (1:12:44.720)
And then when I get you to express a statement,
François Chollet (1:12:47.960)
if it's a statement that me as the controller,
Lex Fridman (1:12:51.840)
I want to reinforce.
François Chollet (1:12:53.800)
I can just show it to people who will agree,
Lex Fridman (1:12:55.560)
and they will like it.
Lex Fridman (1:12:56.880)
And that will reinforce the statement in your mind.
Lex Fridman (1:12:59.280)
If this is a statement I want you to,
François Chollet (1:13:02.760)
this is a belief I want you to abandon,
Lex Fridman (1:13:05.320)
I can, on the other hand, show it to opponents.
François Chollet (1:13:09.600)
We'll attack you.
Lex Fridman (1:13:10.640)
And because they attack you, at the very least,
François Chollet (1:13:12.840)
next time you will think twice about posting it.
Lex Fridman (1:13:16.840)
But maybe you will even start believing this
François Chollet (1:13:20.280)
because you got pushback.
Lex Fridman (1:13:22.840)
So there are many ways in which social media platforms
François Chollet (1:13:28.440)
can potentially control your opinions.
Lex Fridman (1:13:30.520)
And today, so all of these things
François Chollet (1:13:35.040)
are already being controlled by AI algorithms.
Lex Fridman (1:13:38.240)
These algorithms do not have any explicit political goal
François Chollet (1:13:41.880)
today.
Lex Fridman (1:13:42.880)
Well, potentially they could, like if some totalitarian
François Chollet (1:13:48.680)
government takes over social media platforms
Lex Fridman (1:13:52.720)
and decides that now we are going to use this not just
François Chollet (1:13:55.360)
for mass surveillance, but also for mass opinion control
Lex Fridman (1:13:58.040)
and behavior control.
François Chollet (1:13:59.360)
Very bad things could happen.
Lex Fridman (1:14:01.840)
But what's really fascinating and actually quite concerning
François Chollet (1:14:06.480)
is that even without an explicit intent to manipulate,
Lex Fridman (1:14:11.280)
you're already seeing very dangerous dynamics
François Chollet (1:14:14.760)
in terms of how these content recommendation
Lex Fridman (1:14:18.160)
algorithms behave.
François Chollet (1:14:19.800)
Because right now, the goal, the objective function
Lex Fridman (1:14:24.920)
of these algorithms is to maximize engagement,
François Chollet (1:14:28.640)
which seems fairly innocuous at first.
Lex Fridman (1:14:32.520)
However, it is not because content
François Chollet (1:14:36.480)
that will maximally engage people, get people to react
Lex Fridman (1:14:42.000)
in an emotional way, get people to click on something.
François Chollet (1:14:44.720)
It is very often content that is not
Lex Fridman (1:14:52.200)
healthy to the public discourse.
François Chollet (1:14:54.400)
For instance, fake news are far more
Lex Fridman (1:14:58.200)
likely to get you to click on them than real news
François Chollet (1:15:01.320)
simply because they are not constrained to reality.
Lex Fridman (1:15:06.960)
So they can be as outrageous, as surprising,
François Chollet (1:15:11.360)
as good stories as you want because they're artificial.
Lex Fridman (1:15:15.880)
To me, that's an exciting world because so much good
François Chollet (1:15:18.880)
can come.
Lex Fridman (1:15:19.560)
So there's an opportunity to educate people.
François Chollet (1:15:24.520)
You can balance people's worldview with other ideas.
Lex Fridman (1:15:31.200)
So there's so many objective functions.
François Chollet (1:15:33.800)
The space of objective functions that
Lex Fridman (1:15:35.840)
create better civilizations is large, arguably infinite.
Lex Fridman (1:15:40.720)
But there's also a large space that
Lex Fridman (1:15:43.720)
creates division and destruction, civil war,
François Chollet (1:15:51.480)
a lot of bad stuff.
Lex Fridman (1:15:53.160)
And the worry is, naturally, probably that space
François Chollet (1:15:56.920)
is bigger, first of all.
Lex Fridman (1:15:59.160)
And if we don't explicitly think about what kind of effects
François Chollet (1:16:04.920)
are going to be observed from different objective functions,
Lex Fridman (1:16:08.320)
then we're going to get into trouble.
Lex Fridman (1:16:10.160)
But the question is, how do we get into rooms
Lex Fridman (1:16:14.480)
and have discussions, so inside Google, inside Facebook,
François Chollet (1:16:18.560)
inside Twitter, and think about, OK,
Lex Fridman (1:16:21.840)
how can we drive up engagement and, at the same time,
Lex Fridman (1:16:24.840)
create a good society?
Lex Fridman (1:16:28.200)
Is it even possible to have that kind
Lex Fridman (1:16:29.560)
of philosophical discussion?
Lex Fridman (1:16:31.720)
I think you can definitely try.
Lex Fridman (1:16:33.080)
So from my perspective, I would feel rather uncomfortable
Lex Fridman (1:16:37.280)
with companies that are uncomfortable with these new
François Chollet (1:16:41.560)
student algorithms, with them making explicit decisions
Lex Fridman (1:16:47.120)
to manipulate people's opinions or behaviors,
François Chollet (1:16:50.440)
even if the intent is good, because that's
Lex Fridman (1:16:53.480)
a very totalitarian mindset.
Lex Fridman (1:16:55.200)
So instead, what I would like to see
Lex Fridman (1:16:57.440)
is probably never going to happen,
François Chollet (1:16:58.880)
because it's not super realistic,
Lex Fridman (1:17:00.360)
but that's actually something I really care about.
François Chollet (1:17:02.520)
I would like all these algorithms
Lex Fridman (1:17:06.280)
to present configuration settings to their users,
Lex Fridman (1:17:10.560)
so that the users can actually make the decision about how
Lex Fridman (1:17:14.600)
they want to be impacted by these information
François Chollet (1:17:19.000)
recommendation, content recommendation algorithms.
Lex Fridman (1:17:21.960)
For instance, as a user of something
François Chollet (1:17:24.240)
like YouTube or Twitter, maybe I want
Lex Fridman (1:17:26.520)
to maximize learning about a specific topic.
Lex Fridman (1:17:30.280)
So I want the algorithm to feed my curiosity,
Lex Fridman (1:17:36.800)
which is in itself a very interesting problem.
Lex Fridman (1:17:38.760)
So instead of maximizing my engagement,
Lex Fridman (1:17:41.200)
it will maximize how fast and how much I'm learning.
Lex Fridman (1:17:44.600)
And it will also take into account the accuracy,
Lex Fridman (1:17:47.360)
hopefully, of the information I'm learning.
Lex Fridman (1:17:50.680)
So yeah, the user should be able to determine exactly
Lex Fridman (1:17:55.680)
how these algorithms are affecting their lives.
François Chollet (1:17:58.560)
I don't want actually any entity making decisions
Lex Fridman (1:18:03.520)
about in which direction they're going to try to manipulate me.
François Chollet (1:18:09.480)
I want technology.
Lex Fridman (1:18:11.680)
So AI, these algorithms are increasingly
François Chollet (1:18:14.280)
going to be our interface to a world that is increasingly
Lex Fridman (1:18:18.560)
made of information.
Lex Fridman (1:18:19.960)
And I want everyone to be in control of this interface,
Lex Fridman (1:18:25.840)
to interface with the world on their own terms.
Lex Fridman (1:18:29.160)
So if someone wants these algorithms
Lex Fridman (1:18:32.840)
to serve their own personal growth goals,
François Chollet (1:18:37.640)
they should be able to configure these algorithms
Lex Fridman (1:18:40.640)
in such a way.
François Chollet (1:18:41.800)
Yeah, but so I know it's painful to have explicit decisions.
Lex Fridman (1:18:46.680)
But there is underlying explicit decisions,
François Chollet (1:18:51.080)
which is some of the most beautiful fundamental
Lex Fridman (1:18:53.360)
philosophy that we have before us,
François Chollet (1:18:57.400)
which is personal growth.
Lex Fridman (1:19:01.120)
If I want to watch videos from which I can learn,
Lex Fridman (1:19:05.680)
what does that mean?
Lex Fridman (1:19:08.080)
So if I have a checkbox that wants to emphasize learning,
François Chollet (1:19:11.800)
there's still an algorithm with explicit decisions in it
Lex Fridman (1:19:15.480)
that would promote learning.
Lex Fridman (1:19:17.800)
What does that mean for me?
Lex Fridman (1:19:19.200)
For example, I've watched a documentary on flat Earth
François Chollet (1:19:22.800)
theory, I guess.
Lex Fridman (1:19:27.280)
I learned a lot.
François Chollet (1:19:28.240)
I'm really glad I watched it.
Lex Fridman (1:19:29.800)
It was a friend recommended it to me.
François Chollet (1:19:32.560)
Because I don't have such an allergic reaction to crazy
Lex Fridman (1:19:35.800)
people, as my fellow colleagues do.
Lex Fridman (1:19:37.640)
But it was very eye opening.
Lex Fridman (1:19:40.360)
And for others, it might not be.
François Chollet (1:19:42.120)
From others, they might just get turned off from that, same
Lex Fridman (1:19:45.560)
with Republican and Democrat.
Lex Fridman (1:19:47.160)
And it's a non trivial problem.
Lex Fridman (1:19:50.200)
And first of all, if it's done well,
François Chollet (1:19:52.880)
I don't think it's something that wouldn't happen,
Lex Fridman (1:19:56.560)
that YouTube wouldn't be promoting,
François Chollet (1:19:59.280)
or Twitter wouldn't be.
Lex Fridman (1:20:00.200)
It's just a really difficult problem,
Lex Fridman (1:20:02.280)
how to give people control.
Lex Fridman (1:20:05.520)
Well, it's mostly an interface design problem.
François Chollet (1:20:08.960)
The way I see it, you want to create technology
Lex Fridman (1:20:11.080)
that's like a mentor, or a coach, or an assistant,
Lex Fridman (1:20:16.400)
so that it's not your boss.
Lex Fridman (1:20:20.520)
You are in control of it.
François Chollet (1:20:22.560)
You are telling it what to do for you.
Lex Fridman (1:20:25.760)
And if you feel like it's manipulating you,
François Chollet (1:20:27.840)
it's not actually doing what you want.
Lex Fridman (1:20:31.760)
You should be able to switch to a different algorithm.
Lex Fridman (1:20:34.920)
So that's fine tune control.
Lex Fridman (1:20:36.440)
You kind of learn that you're trusting
François Chollet (1:20:38.840)
the human collaboration.
Lex Fridman (1:20:40.080)
I mean, that's how I see autonomous vehicles too,
François Chollet (1:20:41.920)
is giving as much information as possible,
Lex Fridman (1:20:44.480)
and you learn that dance yourself.
François Chollet (1:20:47.240)
Yeah, Adobe, I don't know if you use Adobe product
Lex Fridman (1:20:50.280)
for like Photoshop.
François Chollet (1:20:52.280)
They're trying to see if they can inject YouTube
Lex Fridman (1:20:55.040)
into their interface, but basically allow you
François Chollet (1:20:57.120)
to show you all these videos,
Lex Fridman (1:20:59.840)
that everybody's confused about what to do with features.
Lex Fridman (1:21:03.320)
So basically teach people by linking to,
Lex Fridman (1:21:07.120)
in that way, it's an assistant that uses videos
François Chollet (1:21:10.280)
as a basic element of information.
Lex Fridman (1:21:13.440)
Okay, so what practically should people do
François Chollet (1:21:18.240)
to try to fight against abuses of these algorithms,
Lex Fridman (1:21:24.000)
or algorithms that manipulate us?
François Chollet (1:21:27.400)
Honestly, it's a very, very difficult problem,
Lex Fridman (1:21:29.280)
because to start with, there is very little public awareness
François Chollet (1:21:32.800)
of these issues.
Lex Fridman (1:21:35.040)
Very few people would think there's anything wrong
François Chollet (1:21:38.520)
with the unused algorithm,
Lex Fridman (1:21:39.720)
even though there is actually something wrong already,
François Chollet (1:21:42.040)
which is that it's trying to maximize engagement
Lex Fridman (1:21:44.480)
most of the time, which has very negative side effects.
Lex Fridman (1:21:49.880)
So ideally, so the very first thing is to stop
Lex Fridman (1:21:56.160)
trying to purely maximize engagement,
Lex Fridman (1:21:59.560)
try to propagate content based on popularity, right?
Lex Fridman (1:22:06.560)
Instead, take into account the goals
Lex Fridman (1:22:11.040)
and the profiles of each user.
Lex Fridman (1:22:13.560)
So you will be, one example is, for instance,
François Chollet (1:22:16.920)
when I look at topic recommendations on Twitter,
Lex Fridman (1:22:20.800)
it's like, you know, they have this news tab
François Chollet (1:22:24.480)
with switch recommendations.
Lex Fridman (1:22:25.480)
It's always the worst coverage,
François Chollet (1:22:28.480)
because it's content that appeals
Lex Fridman (1:22:30.360)
to the smallest common denominator
François Chollet (1:22:34.080)
to all Twitter users, because they're trying to optimize.
Lex Fridman (1:22:37.080)
They're purely trying to optimize popularity.
François Chollet (1:22:39.040)
They're purely trying to optimize engagement.
Lex Fridman (1:22:41.320)
But that's not what I want.
Lex Fridman (1:22:42.960)
So they should put me in control of some setting
Lex Fridman (1:22:46.080)
so that I define what's the objective function
François Chollet (1:22:50.360)
that Twitter is going to be following
Lex Fridman (1:22:52.200)
to show me this content.
Lex Fridman (1:22:54.120)
And honestly, so this is all about interface design.
Lex Fridman (1:22:57.360)
And we are not, it's not realistic
François Chollet (1:22:59.440)
to give users control of a bunch of knobs
Lex Fridman (1:23:01.760)
that define algorithm.
François Chollet (1:23:03.400)
Instead, we should purely put them in charge
Lex Fridman (1:23:06.760)
of defining the objective function.
François Chollet (1:23:09.400)
Like, let the user tell us what they want to achieve,
Lex Fridman (1:23:13.240)
how they want this algorithm to impact their lives.
Lex Fridman (1:23:15.280)
So do you think it is that,
Lex Fridman (1:23:16.680)
or do they provide individual article by article
François Chollet (1:23:19.360)
reward structure where you give a signal,
Lex Fridman (1:23:21.600)
I'm glad I saw this, or I'm glad I didn't?
Lex Fridman (1:23:24.720)
So like a Spotify type feedback mechanism,
Lex Fridman (1:23:28.480)
it works to some extent.
François Chollet (1:23:30.680)
I'm kind of skeptical about it
Lex Fridman (1:23:32.000)
because the only way the algorithm,
François Chollet (1:23:34.880)
the algorithm will attempt to relate your choices
Lex Fridman (1:23:39.120)
with the choices of everyone else,
François Chollet (1:23:41.040)
which might, you know, if you have an average profile
Lex Fridman (1:23:45.000)
that works fine, I'm sure Spotify accommodations work fine
François Chollet (1:23:47.880)
if you just like mainstream stuff.
Lex Fridman (1:23:49.560)
If you don't, it can be, it's not optimal at all actually.
François Chollet (1:23:53.960)
It'll be in an efficient search
Lex Fridman (1:23:56.040)
for the part of the Spotify world that represents you.
Lex Fridman (1:24:00.800)
So it's a tough problem,
Lex Fridman (1:24:02.960)
but do note that even a feedback system
François Chollet (1:24:07.960)
like what Spotify has does not give me control
Lex Fridman (1:24:10.880)
over what the algorithm is trying to optimize for.
François Chollet (1:24:16.320)
Well, public awareness, which is what we're doing now,
Lex Fridman (1:24:19.360)
is a good place to start.
Lex Fridman (1:24:21.360)
Do you have concerns about longterm existential threats
Lex Fridman (1:24:25.960)
of artificial intelligence?
François Chollet (1:24:28.280)
Well, as I was saying,
Lex Fridman (1:24:31.040)
our world is increasingly made of information.
François Chollet (1:24:33.360)
AI algorithms are increasingly going to be our interface
Lex Fridman (1:24:36.240)
to this world of information,
Lex Fridman (1:24:37.880)
and somebody will be in control of these algorithms.
Lex Fridman (1:24:41.480)
And that puts us in any kind of a bad situation, right?
François Chollet (1:24:45.920)
It has risks.
Lex Fridman (1:24:46.880)
It has risks coming from potentially large companies
François Chollet (1:24:50.840)
wanting to optimize their own goals,
Lex Fridman (1:24:53.760)
maybe profit, maybe something else.
François Chollet (1:24:55.960)
Also from governments who might want to use these algorithms
Lex Fridman (1:25:00.720)
as a means of control of the population.
Lex Fridman (1:25:03.520)
Do you think there's existential threat
Lex Fridman (1:25:05.000)
that could arise from that?
Lex Fridman (1:25:06.320)
So existential threat.
Lex Fridman (1:25:09.120)
So maybe you're referring to the singularity narrative
François Chollet (1:25:13.240)
where robots just take over.
Lex Fridman (1:25:15.560)
Well, I don't, I'm not terminating robots,
Lex Fridman (1:25:18.320)
and I don't believe it has to be a singularity.
Lex Fridman (1:25:21.000)
We're just talking to, just like you said,
François Chollet (1:25:24.800)
the algorithm controlling masses of populations.
Lex Fridman (1:25:28.920)
The existential threat being,
François Chollet (1:25:32.640)
hurt ourselves much like a nuclear war would hurt ourselves.
Lex Fridman (1:25:36.760)
That kind of thing.
François Chollet (1:25:37.600)
I don't think that requires a singularity.
Lex Fridman (1:25:39.480)
That requires a loss of control over AI algorithm.
François Chollet (1:25:42.560)
Yes.
Lex Fridman (1:25:43.560)
So I do agree there are concerning trends.
François Chollet (1:25:47.000)
Honestly, I wouldn't want to make any longterm predictions.
Lex Fridman (1:25:52.960)
I don't think today we really have the capability
François Chollet (1:25:56.000)
to see what the dangers of AI
Lex Fridman (1:25:58.560)
are going to be in 50 years, in 100 years.
François Chollet (1:26:01.360)
I do see that we are already faced
Lex Fridman (1:26:04.800)
with concrete and present dangers
François Chollet (1:26:08.840)
surrounding the negative side effects
Lex Fridman (1:26:11.560)
of content recombination systems, of newsfeed algorithms
François Chollet (1:26:14.960)
concerning algorithmic bias as well.
Lex Fridman (1:26:18.640)
So we are delegating more and more
François Chollet (1:26:22.240)
decision processes to algorithms.
Lex Fridman (1:26:25.080)
Some of these algorithms are uncrafted,
François Chollet (1:26:26.760)
some are learned from data,
Lex Fridman (1:26:29.360)
but we are delegating control.
François Chollet (1:26:32.920)
Sometimes it's a good thing, sometimes not so much.
Lex Fridman (1:26:36.280)
And there is in general very little supervision
Lex Fridman (1:26:39.480)
of this process, right?
Lex Fridman (1:26:41.000)
So we are still in this period of very fast change,
François Chollet (1:26:45.400)
even chaos, where society is restructuring itself,
Lex Fridman (1:26:50.920)
turning into an information society,
François Chollet (1:26:53.160)
which itself is turning into
Lex Fridman (1:26:54.520)
an increasingly automated information passing society.
Lex Fridman (1:26:58.360)
And well, yeah, I think the best we can do today
Lex Fridman (1:27:02.520)
is try to raise awareness around some of these issues.
Lex Fridman (1:27:06.040)
And I think we're actually making good progress.
Lex Fridman (1:27:07.680)
If you look at algorithmic bias, for instance,
François Chollet (1:27:12.760)
three years ago, even two years ago,
Lex Fridman (1:27:14.760)
very, very few people were talking about it.
Lex Fridman (1:27:17.040)
And now all the big companies are talking about it.
Lex Fridman (1:27:20.320)
They are often not in a very serious way,
Lex Fridman (1:27:22.360)
but at least it is part of the public discourse.
Lex Fridman (1:27:24.560)
You see people in Congress talking about it.
Lex Fridman (1:27:27.080)
And it all started from raising awareness.
Lex Fridman (1:27:31.960)
Right.
Lex Fridman (1:27:32.800)
So in terms of alignment problem,
Lex Fridman (1:27:36.080)
trying to teach as we allow algorithms,
François Chollet (1:27:39.400)
just even recommender systems on Twitter,
Lex Fridman (1:27:43.640)
encoding human values and morals,
François Chollet (1:27:48.280)
decisions that touch on ethics,
Lex Fridman (1:27:50.200)
how hard do you think that problem is?
Lex Fridman (1:27:52.600)
How do we have lost functions in neural networks
Lex Fridman (1:27:57.240)
that have some component,
Lex Fridman (1:27:58.640)
some fuzzy components of human morals?
Lex Fridman (1:28:01.080)
Well, I think this is really all about objective function engineering,
François Chollet (1:28:06.080)
which is probably going to be increasingly a topic of concern in the future.
Lex Fridman (1:28:10.520)
Like for now, we're just using very naive loss functions
François Chollet (1:28:14.640)
because the hard part is not actually what you're trying to minimize.
Lex Fridman (1:28:17.760)
It's everything else.
Lex Fridman (1:28:19.040)
But as the everything else is going to be increasingly automated,
Lex Fridman (1:28:22.840)
we're going to be focusing our human attention
François Chollet (1:28:27.040)
on increasingly high level components,
Lex Fridman (1:28:30.240)
like what's actually driving the whole learning system,
François Chollet (1:28:32.680)
like the objective function.
Lex Fridman (1:28:33.960)
So loss function engineering is going to be,
François Chollet (1:28:36.920)
loss function engineer is probably going to be a job title in the future.
Lex Fridman (1:28:40.640)
And then the tooling you're creating with Keras essentially
François Chollet (1:28:44.520)
takes care of all the details underneath.
Lex Fridman (1:28:47.040)
And basically the human expert is needed for exactly that.
François Chollet (1:28:52.720)
That's the idea.
Lex Fridman (1:28:53.920)
Keras is the interface between the data you're collecting
Lex Fridman (1:28:57.640)
and the business goals.
Lex Fridman (1:28:59.080)
And your job as an engineer is going to be to express your business goals
Lex Fridman (1:29:03.480)
and your understanding of your business or your product,
Lex Fridman (1:29:06.720)
your system as a kind of loss function or a kind of set of constraints.
Lex Fridman (1:29:11.840)
Does the possibility of creating an AGI system excite you or scare you or bore you?
Lex Fridman (1:29:19.480)
So intelligence can never really be general.
François Chollet (1:29:22.080)
You know, at best it can have some degree of generality like human intelligence.
Lex Fridman (1:29:26.400)
It also always has some specialization in the same way that human intelligence
François Chollet (1:29:30.640)
is specialized in a certain category of problems,
Lex Fridman (1:29:33.440)
is specialized in the human experience.
Lex Fridman (1:29:35.440)
And when people talk about AGI,
Lex Fridman (1:29:37.280)
I'm never quite sure if they're talking about very, very smart AI,
Lex Fridman (1:29:42.520)
so smart that it's even smarter than humans,
Lex Fridman (1:29:45.080)
or they're talking about human like intelligence,
François Chollet (1:29:48.000)
because these are different things.
Lex Fridman (1:29:49.680)
Let's say, presumably I'm oppressing you today with my humanness.
Lex Fridman (1:29:54.760)
So imagine that I was in fact a robot.
Lex Fridman (1:29:59.240)
So what does that mean?
François Chollet (1:30:01.920)
That I'm impressing you with natural language processing.
Lex Fridman (1:30:04.920)
Maybe if you weren't able to see me, maybe this is a phone call.
Lex Fridman (1:30:07.840)
So that kind of system.
Lex Fridman (1:30:10.000)
Companion.
Lex Fridman (1:30:11.120)
So that's very much about building human like AI.
Lex Fridman (1:30:15.040)
And you're asking me, you know, is this an exciting perspective?
François Chollet (1:30:18.200)
Yes.
Lex Fridman (1:30:19.440)
I think so, yes.
François Chollet (1:30:21.760)
Not so much because of what artificial human like intelligence could do,
Lex Fridman (1:30:28.000)
but, you know, from an intellectual perspective,
François Chollet (1:30:30.880)
I think if you could build truly human like intelligence,
Lex Fridman (1:30:34.120)
that means you could actually understand human intelligence,
Lex Fridman (1:30:37.240)
which is fascinating, right?
Lex Fridman (1:30:39.880)
Human like intelligence is going to require emotions.
François Chollet (1:30:42.680)
It's going to require consciousness,
Lex Fridman (1:30:44.400)
which is not things that would normally be required by an intelligent system.
François Chollet (1:30:49.720)
If you look at, you know, we were mentioning earlier like science
Lex Fridman (1:30:53.160)
as a superhuman problem solving agent or system,
François Chollet (1:30:59.600)
it does not have consciousness, it doesn't have emotions.
Lex Fridman (1:31:02.120)
In general, so emotions,
François Chollet (1:31:04.320)
I see consciousness as being on the same spectrum as emotions.
Lex Fridman (1:31:07.640)
It is a component of the subjective experience
Lex Fridman (1:31:12.280)
that is meant very much to guide behavior generation, right?
Lex Fridman (1:31:18.800)
It's meant to guide your behavior.
François Chollet (1:31:20.800)
In general, human intelligence and animal intelligence
Lex Fridman (1:31:24.520)
has evolved for the purpose of behavior generation, right?
François Chollet (1:31:29.280)
Including in a social context.
Lex Fridman (1:31:30.680)
So that's why we actually need emotions.
François Chollet (1:31:32.480)
That's why we need consciousness.
Lex Fridman (1:31:34.920)
An artificial intelligence system developed in a different context
François Chollet (1:31:38.360)
may well never need them, may well never be conscious like science.
Lex Fridman (1:31:42.800)
Well, on that point, I would argue it's possible to imagine
François Chollet (1:31:47.960)
that there's echoes of consciousness in science
Lex Fridman (1:31:51.480)
when viewed as an organism, that science is consciousness.
Lex Fridman (1:31:55.480)
So, I mean, how would you go about testing this hypothesis?
Lex Fridman (1:31:59.160)
How do you probe the subjective experience of an abstract system like science?
François Chollet (1:32:07.000)
Well, the point of probing any subjective experience is impossible
Lex Fridman (1:32:10.400)
because I'm not science, I'm Lex.
Lex Fridman (1:32:13.200)
So I can't probe another entity, it's no more than bacteria on my skin.
Lex Fridman (1:32:20.520)
You're Lex, I can ask you questions about your subjective experience
Lex Fridman (1:32:24.160)
and you can answer me, and that's how I know you're conscious.
Lex Fridman (1:32:28.440)
Yes, but that's because we speak the same language.
François Chollet (1:32:31.840)
You perhaps, we have to speak the language of science in order to ask it.
Lex Fridman (1:32:35.520)
Honestly, I don't think consciousness, just like emotions of pain and pleasure,
François Chollet (1:32:40.320)
is not something that inevitably arises
Lex Fridman (1:32:44.160)
from any sort of sufficiently intelligent information processing.
François Chollet (1:32:47.920)
It is a feature of the mind, and if you've not implemented it explicitly, it is not there.
Lex Fridman (1:32:53.920)
So you think it's an emergent feature of a particular architecture.
Lex Fridman (1:32:58.960)
So do you think...
Lex Fridman (1:33:00.320)
It's a feature in the same sense.
François Chollet (1:33:02.000)
So, again, the subjective experience is all about guiding behavior.
Lex Fridman (1:33:08.240)
If the problems you're trying to solve don't really involve an embodied agent,
François Chollet (1:33:15.120)
maybe in a social context, generating behavior and pursuing goals like this.
Lex Fridman (1:33:19.520)
And if you look at science, that's not really what's happening.
François Chollet (1:33:22.160)
Even though it is, it is a form of artificial AI, artificial intelligence,
Lex Fridman (1:33:27.920)
in the sense that it is solving problems, it is accumulating knowledge,
François Chollet (1:33:31.920)
accumulating solutions and so on.
Lex Fridman (1:33:35.040)
So if you're not explicitly implementing a subjective experience,
François Chollet (1:33:39.440)
implementing certain emotions and implementing consciousness,
Lex Fridman (1:33:44.000)
it's not going to just spontaneously emerge.
François Chollet (1:33:47.360)
Yeah.
Lex Fridman (1:33:48.080)
But so for a system like, human like intelligence system that has consciousness,
Lex Fridman (1:33:53.200)
do you think it needs to have a body?
Lex Fridman (1:33:55.840)
Yes, definitely.
Lex Fridman (1:33:56.720)
I mean, it doesn't have to be a physical body, right?
Lex Fridman (1:33:59.600)
And there's not that much difference between a realistic simulation in the real world.
Lex Fridman (1:34:03.440)
So there has to be something you have to preserve kind of thing.
Lex Fridman (1:34:06.400)
Yes, but human like intelligence can only arise in a human like context.
François Chollet (1:34:11.840)
Intelligence needs other humans in order for you to demonstrate
Lex Fridman (1:34:16.800)
that you have human like intelligence, essentially.
François Chollet (1:34:19.040)
Yes.
Lex Fridman (1:34:20.320)
So what kind of tests and demonstration would be sufficient for you
Lex Fridman (1:34:28.080)
to demonstrate human like intelligence?
Lex Fridman (1:34:30.960)
Yeah.
François Chollet (1:34:31.360)
Just out of curiosity, you've talked about in terms of theorem proving
Lex Fridman (1:34:35.600)
and program synthesis, I think you've written about
François Chollet (1:34:38.000)
that there's no good benchmarks for this.
Lex Fridman (1:34:40.480)
Yeah.
François Chollet (1:34:40.720)
That's one of the problems.
Lex Fridman (1:34:42.000)
So let's talk program synthesis.
Lex Fridman (1:34:46.320)
So what do you imagine is a good...
Lex Fridman (1:34:48.800)
I think it's related questions for human like intelligence
Lex Fridman (1:34:51.360)
and for program synthesis.
Lex Fridman (1:34:53.360)
What's a good benchmark for either or both?
François Chollet (1:34:56.080)
Right.
Lex Fridman (1:34:56.480)
So I mean, you're actually asking two questions,
François Chollet (1:34:59.200)
which is one is about quantifying intelligence
Lex Fridman (1:35:02.480)
and comparing the intelligence of an artificial system
François Chollet (1:35:06.880)
to the intelligence for human.
Lex Fridman (1:35:08.480)
And the other is about the degree to which this intelligence is human like.
François Chollet (1:35:13.440)
It's actually two different questions.
Lex Fridman (1:35:16.560)
So you mentioned earlier the Turing test.
François Chollet (1:35:19.680)
Well, I actually don't like the Turing test because it's very lazy.
François Chollet (1:35:23.200)
It's all about completely bypassing the problem of defining and measuring intelligence
Lex Fridman (1:35:28.720)
and instead delegating to a human judge or a panel of human judges.
Lex Fridman (1:35:34.160)
So it's a total copout, right?
François Chollet (1:35:38.160)
If you want to measure how human like an agent is,
Lex Fridman (1:35:43.760)
I think you have to make it interact with other humans.
François Chollet (1:35:47.600)
Maybe it's not necessarily a good idea to have these other humans be the judges.
François Chollet (1:35:53.760)
Maybe you should just observe behavior and compare it to what a human would actually have done.
François Chollet (1:36:00.560)
When it comes to measuring how smart, how clever an agent is
Lex Fridman (1:36:05.120)
and comparing that to the degree of human intelligence.
Lex Fridman (1:36:11.120)
So we're already talking about two things, right?
Lex Fridman (1:36:13.520)
The degree, kind of like the magnitude of an intelligence and its direction, right?
François Chollet (1:36:20.320)
Like the norm of a vector and its direction.
Lex Fridman (1:36:23.280)
And the direction is like human likeness and the magnitude, the norm is intelligence.
Lex Fridman (1:36:32.720)
You could call it intelligence, right?
Lex Fridman (1:36:34.080)
So the direction, your sense, the space of directions that are human like is very narrow.
François Chollet (1:36:41.040)
Yeah.
Lex Fridman (1:36:42.240)
So the way you would measure the magnitude of intelligence in a system
François Chollet (1:36:48.880)
in a way that also enables you to compare it to that of a human.
Lex Fridman (1:36:54.640)
Well, if you look at different benchmarks for intelligence today,
François Chollet (1:36:59.200)
they're all too focused on skill at a given task.
Lex Fridman (1:37:04.160)
Like skill at playing chess, skill at playing Go, skill at playing Dota.
Lex Fridman (1:37:10.720)
And I think that's not the right way to go about it because you can always
Lex Fridman (1:37:15.600)
beat a human at one specific task.
François Chollet (1:37:19.200)
The reason why our skill at playing Go or juggling or anything is impressive
Lex Fridman (1:37:23.920)
is because we are expressing this skill within a certain set of constraints.
François Chollet (1:37:28.400)
If you remove the constraints, the constraints that we have one lifetime,
Lex Fridman (1:37:32.320)
that we have this body and so on, if you remove the context,
François Chollet (1:37:36.080)
if you have unlimited string data, if you can have access to, you know,
François Chollet (1:37:40.480)
for instance, if you look at juggling, if you have no restriction on the hardware,
François Chollet (1:37:44.640)
then achieving arbitrary levels of skill is not very interesting
Lex Fridman (1:37:48.400)
and says nothing about the amount of intelligence you've achieved.
Lex Fridman (1:37:52.400)
So if you want to measure intelligence, you need to rigorously define what
Lex Fridman (1:37:57.440)
intelligence is, which in itself, you know, it's a very challenging problem.
Lex Fridman (1:38:02.960)
And do you think that's possible?
Lex Fridman (1:38:04.320)
To define intelligence? Yes, absolutely.
François Chollet (1:38:06.000)
I mean, you can provide, many people have provided, you know, some definition.
Lex Fridman (1:38:10.560)
I have my own definition.
Lex Fridman (1:38:12.000)
Where does your definition begin?
Lex Fridman (1:38:13.440)
Where does your definition begin if it doesn't end?
François Chollet (1:38:16.240)
Well, I think intelligence is essentially the efficiency
Lex Fridman (1:38:22.320)
with which you turn experience into generalizable programs.
Lex Fridman (1:38:29.760)
So what that means is it's the efficiency with which
Lex Fridman (1:38:32.800)
you turn a sampling of experience space into
François Chollet (1:38:36.720)
the ability to process a larger chunk of experience space.
Lex Fridman (1:38:46.000)
So measuring skill can be one proxy across many different tasks,
François Chollet (1:38:52.560)
can be one proxy for measuring intelligence.
Lex Fridman (1:38:54.480)
But if you want to only measure skill, you should control for two things.
François Chollet (1:38:58.720)
You should control for the amount of experience that your system has
Lex Fridman (1:39:04.960)
and the priors that your system has.
Lex Fridman (1:39:08.080)
But if you look at two agents and you give them the same priors
Lex Fridman (1:39:13.120)
and you give them the same amount of experience,
François Chollet (1:39:16.160)
there is one of the agents that is going to learn programs,
Lex Fridman (1:39:21.360)
representations, something, a model that will perform well
François Chollet (1:39:25.440)
on the larger chunk of experience space than the other.
Lex Fridman (1:39:28.720)
And that is the smaller agent.
François Chollet (1:39:30.960)
Yeah. So if you fix the experience, which generate better programs,
Lex Fridman (1:39:37.680)
better meaning more generalizable.
François Chollet (1:39:39.520)
That's really interesting.
Lex Fridman (1:39:40.560)
That's a very nice, clean definition of...
François Chollet (1:39:42.400)
Oh, by the way, in this definition, it is already very obvious
Lex Fridman (1:39:47.280)
that intelligence has to be specialized
François Chollet (1:39:49.440)
because you're talking about experience space
Lex Fridman (1:39:51.680)
and you're talking about segments of experience space.
François Chollet (1:39:54.080)
You're talking about priors and you're talking about experience.
Lex Fridman (1:39:57.200)
All of these things define the context in which intelligence emerges.
Lex Fridman (1:40:04.480)
And you can never look at the totality of experience space, right?
Lex Fridman (1:40:09.760)
So intelligence has to be specialized.
Lex Fridman (1:40:12.160)
But it can be sufficiently large, the experience space,
Lex Fridman (1:40:14.960)
even though it's specialized.
François Chollet (1:40:16.080)
There's a certain point when the experience space is large enough
Lex Fridman (1:40:19.120)
to where it might as well be general.
François Chollet (1:40:22.000)
It feels general. It looks general.
Lex Fridman (1:40:23.920)
Sure. I mean, it's very relative.
François Chollet (1:40:25.680)
Like, for instance, many people would say human intelligence is general.
Lex Fridman (1:40:29.360)
In fact, it is quite specialized.
François Chollet (1:40:32.800)
We can definitely build systems that start from the same innate priors
Lex Fridman (1:40:37.120)
as what humans have at birth.
François Chollet (1:40:39.120)
Because we already understand fairly well
Lex Fridman (1:40:42.320)
what sort of priors we have as humans.
François Chollet (1:40:44.480)
Like many people have worked on this problem.
Lex Fridman (1:40:46.800)
Most notably, Elisabeth Spelke from Harvard.
François Chollet (1:40:51.040)
I don't know if you know her.
Lex Fridman (1:40:52.240)
She's worked a lot on what she calls core knowledge.
Lex Fridman (1:40:56.000)
And it is very much about trying to determine and describe
Lex Fridman (1:41:00.640)
what priors we are born with.
François Chollet (1:41:02.320)
Like language skills and so on, all that kind of stuff.
Lex Fridman (1:41:04.720)
Exactly.
Lex Fridman (1:41:06.880)
So we have some pretty good understanding of what priors we are born with.
Lex Fridman (1:41:11.440)
So we could...
Lex Fridman (1:41:13.760)
So I've actually been working on a benchmark for the past couple years,
Lex Fridman (1:41:17.760)
you know, on and off.
François Chollet (1:41:18.640)
I hope to be able to release it at some point.
Lex Fridman (1:41:20.480)
That's exciting.
François Chollet (1:41:21.760)
The idea is to measure the intelligence of systems
Lex Fridman (1:41:26.800)
by countering for priors,
François Chollet (1:41:28.640)
countering for amount of experience,
Lex Fridman (1:41:30.480)
and by assuming the same priors as what humans are born with.
Lex Fridman (1:41:34.800)
So that you can actually compare these scores to human intelligence.
Lex Fridman (1:41:39.520)
You can actually have humans pass the same test in a way that's fair.
François Chollet (1:41:43.280)
Yeah. And so importantly, such a benchmark should be such that any amount
Lex Fridman (1:41:52.960)
of practicing does not increase your score.
Lex Fridman (1:41:56.480)
So try to picture a game where no matter how much you play this game,
Lex Fridman (1:42:01.600)
that does not change your skill at the game.
Lex Fridman (1:42:05.040)
Can you picture that?
Lex Fridman (1:42:05.920)
As a person who deeply appreciates practice, I cannot actually.
François Chollet (1:42:11.040)
There's actually a very simple trick.
Lex Fridman (1:42:16.560)
So in order to come up with a task,
Lex Fridman (1:42:19.440)
so the only thing you can measure is skill at the task.
Lex Fridman (1:42:21.760)
Yes.
François Chollet (1:42:22.320)
All tasks are going to involve priors.
Lex Fridman (1:42:24.800)
Yes.
François Chollet (1:42:25.600)
The trick is to know what they are and to describe that.
Lex Fridman (1:42:29.920)
And then you make sure that this is the same set of priors as what humans start with.
Lex Fridman (1:42:33.760)
So you create a task that assumes these priors, that exactly documents these priors,
Lex Fridman (1:42:38.560)
so that the priors are made explicit and there are no other priors involved.
Lex Fridman (1:42:42.240)
And then you generate a certain number of samples in experience space for this task, right?
Lex Fridman (1:42:49.840)
And this, for one task, assuming that the task is new for the agent passing it,
François Chollet (1:42:56.320)
that's one test of this definition of intelligence that we set up.
Lex Fridman (1:43:04.320)
And now you can scale that to many different tasks,
Lex Fridman (1:43:06.880)
that each task should be new to the agent passing it, right?
Lex Fridman (1:43:11.360)
And also it should be human interpretable and understandable
Lex Fridman (1:43:14.480)
so that you can actually have a human pass the same test.
Lex Fridman (1:43:16.880)
And then you can compare the score of your machine and the score of your human.
François Chollet (1:43:19.760)
Which could be a lot of stuff.
Lex Fridman (1:43:20.720)
You could even start a task like MNIST.
François Chollet (1:43:23.040)
Just as long as you start with the same set of priors.
Lex Fridman (1:43:28.800)
So the problem with MNIST, humans are already trying to recognize digits, right?
Lex Fridman (1:43:35.600)
But let's say we're considering objects that are not digits,
Lex Fridman (1:43:42.400)
some completely arbitrary patterns.
François Chollet (1:43:44.480)
Well, humans already come with visual priors about how to process that.
Lex Fridman (1:43:48.880)
So in order to make the game fair, you would have to isolate these priors
Lex Fridman (1:43:54.080)
and describe them and then express them as computational rules.
Lex Fridman (1:43:57.280)
Having worked a lot with vision science people, that's exceptionally difficult.
François Chollet (1:44:01.680)
A lot of progress has been made.
François Chollet (1:44:03.120)
There's been a lot of good tests and basically reducing all of human vision into some good priors.
François Chollet (1:44:08.640)
We're still probably far away from that perfectly,
Lex Fridman (1:44:10.960)
but as a start for a benchmark, that's an exciting possibility.
François Chollet (1:44:14.640)
Yeah, so Elisabeth Spelke actually lists objectness as one of the core knowledge priors.
Lex Fridman (1:44:24.800)
Objectness, cool.
François Chollet (1:44:25.920)
Objectness, yeah.
Lex Fridman (1:44:27.440)
So we have priors about objectness, like about the visual space, about time,
François Chollet (1:44:31.520)
about agents, about goal oriented behavior.
Lex Fridman (1:44:35.280)
We have many different priors, but what's interesting is that,
François Chollet (1:44:39.280)
sure, we have this pretty diverse and rich set of priors,
Lex Fridman (1:44:43.920)
but it's also not that diverse, right?
François Chollet (1:44:46.880)
We are not born into this world with a ton of knowledge about the world,
Lex Fridman (1:44:50.800)
with only a small set of core knowledge.
Lex Fridman (1:44:58.640)
Yeah, sorry, do you have a sense of how it feels to us humans that that set is not that large?
Lex Fridman (1:45:05.040)
But just even the nature of time that we kind of integrate pretty effectively
François Chollet (1:45:09.600)
through all of our perception, all of our reasoning,
Lex Fridman (1:45:12.640)
maybe how, you know, do you have a sense of how easy it is to encode those priors?
François Chollet (1:45:17.680)
Maybe it requires building a universe and then the human brain in order to encode those priors.
Lex Fridman (1:45:25.440)
Or do you have a hope that it can be listed like an axiomatic?
François Chollet (1:45:28.640)
I don't think so.
Lex Fridman (1:45:29.280)
So you have to keep in mind that any knowledge about the world that we are
François Chollet (1:45:33.040)
born with is something that has to have been encoded into our DNA by evolution at some point.
Lex Fridman (1:45:41.120)
Right.
Lex Fridman (1:45:41.440)
And DNA is a very, very low bandwidth medium.
François Chollet (1:45:46.000)
Like it's extremely long and expensive to encode anything into DNA because first of all,
François Chollet (1:45:52.560)
you need some sort of evolutionary pressure to guide this writing process.
Lex Fridman (1:45:57.440)
And then, you know, the higher level of information you're trying to write, the longer it's going to take.
Lex Fridman (1:46:04.480)
And the thing in the environment that you're trying to encode knowledge about has to be stable
Lex Fridman (1:46:13.520)
over this duration.
Lex Fridman (1:46:15.280)
So you can only encode into DNA things that constitute an evolutionary advantage.
Lex Fridman (1:46:20.960)
So this is actually a very small subset of all possible knowledge about the world.
François Chollet (1:46:25.280)
You can only encode things that are stable, that are true, over very, very long periods of time,
Lex Fridman (1:46:32.080)
typically millions of years.
Lex Fridman (1:46:33.680)
For instance, we might have some visual prior about the shape of snakes, right?
Lex Fridman (1:46:38.720)
But what makes a face, what's the difference between a face and an art face?
Lex Fridman (1:46:44.560)
But consider this interesting question.
Lex Fridman (1:46:48.080)
Do we have any innate sense of the visual difference between a male face and a female face?
Lex Fridman (1:46:56.640)
What do you think?
Lex Fridman (1:46:58.640)
For a human, I mean.
François Chollet (1:46:59.840)
I would have to look back into evolutionary history when the genders emerged.
Lex Fridman (1:47:04.000)
But yeah, most...
François Chollet (1:47:06.240)
I mean, the faces of humans are quite different from the faces of great apes.
Lex Fridman (1:47:10.640)
Great apes, right?
François Chollet (1:47:12.880)
Yeah.
Lex Fridman (1:47:13.600)
That's interesting.
François Chollet (1:47:14.800)
Yeah, you couldn't tell the face of a female chimpanzee from the face of a male chimpanzee,
Lex Fridman (1:47:22.800)
probably.
François Chollet (1:47:23.440)
Yeah, and I don't think most humans have all that ability.
Lex Fridman (1:47:26.160)
So we do have innate knowledge of what makes a face, but it's actually impossible for us to
François Chollet (1:47:33.280)
have any DNA encoded knowledge of the difference between a female human face and a male human face
François Chollet (1:47:40.320)
because that knowledge, that information came up into the world actually very recently.
François Chollet (1:47:50.560)
If you look at the slowness of the process of encoding knowledge into DNA.
Lex Fridman (1:47:56.400)
Yeah, so that's interesting.
François Chollet (1:47:57.360)
That's a really powerful argument that DNA is a low bandwidth and it takes a long time to encode.
Lex Fridman (1:48:02.800)
That naturally creates a very efficient encoding.
Lex Fridman (1:48:05.200)
But one important consequence of this is that, so yes, we are born into this world with a bunch of
François Chollet (1:48:12.800)
knowledge, sometimes high level knowledge about the world, like the shape, the rough shape of a
François Chollet (1:48:17.600)
snake, of the rough shape of a face.
Lex Fridman (1:48:20.480)
But importantly, because this knowledge takes so long to write, almost all of this innate
Lex Fridman (1:48:26.960)
knowledge is shared with our cousins, with great apes, right?
Lex Fridman (1:48:32.080)
So it is not actually this innate knowledge that makes us special.
Lex Fridman (1:48:36.320)
But to throw it right back at you from the earlier on in our discussion, it's that encoding
Lex Fridman (1:48:42.960)
might also include the entirety of the environment of Earth.
François Chollet (1:48:49.360)
To some extent.
Lex Fridman (1:48:49.920)
So it can include things that are important to survival and production, so for which there is
François Chollet (1:48:56.480)
some evolutionary pressure, and things that are stable, constant over very, very, very long time
Lex Fridman (1:49:02.880)
periods.
Lex Fridman (1:49:04.160)
And honestly, it's not that much information.
François Chollet (1:49:06.320)
There's also, besides the bandwidths constraint and the constraints of the writing process,
François Chollet (1:49:14.400)
there's also memory constraints, like DNA, the part of DNA that deals with the human brain,
Lex Fridman (1:49:21.440)
it's actually fairly small.
Lex Fridman (1:49:22.640)
It's like, you know, on the order of megabytes, right?
Lex Fridman (1:49:25.520)
There's not that much high level knowledge about the world you can encode.
François Chollet (1:49:31.600)
That's quite brilliant and hopeful for a benchmark that you're referring to of encoding
Lex Fridman (1:49:38.880)
priors.
François Chollet (1:49:39.360)
I actually look forward to, I'm skeptical whether you can do it in the next couple of
Lex Fridman (1:49:43.120)
years, but hopefully.
François Chollet (1:49:45.040)
I've been working.
Lex Fridman (1:49:45.760)
So honestly, it's a very simple benchmark, and it's not like a big breakthrough or anything.
Lex Fridman (1:49:49.920)
It's more like a fun side project, right?
Lex Fridman (1:49:53.200)
But these fun, so is ImageNet.
François Chollet (1:49:56.480)
These fun side projects could launch entire groups of efforts towards creating reasoning
Lex Fridman (1:50:04.080)
systems and so on.
Lex Fridman (1:50:04.960)
And I think...
Lex Fridman (1:50:05.440)
Yeah, that's the goal.
François Chollet (1:50:06.160)
It's trying to measure strong generalization, to measure the strength of abstraction in
Lex Fridman (1:50:12.080)
our minds, well, in our minds and in artificial intelligence agencies.
Lex Fridman (1:50:16.960)
And if there's anything true about this science organism is its individual cells love competition.
Lex Fridman (1:50:24.800)
So and benchmarks encourage competition.
Lex Fridman (1:50:26.800)
So that's an exciting possibility.
Lex Fridman (1:50:29.520)
If you, do you think an AI winter is coming?
Lex Fridman (1:50:33.520)
And how do we prevent it?
Lex Fridman (1:50:35.440)
Not really.
Lex Fridman (1:50:36.080)
So an AI winter is something that would occur when there's a big mismatch between how we
Lex Fridman (1:50:42.160)
are selling the capabilities of AI and the actual capabilities of AI.
Lex Fridman (1:50:47.280)
And today, some deep learning is creating a lot of value.
Lex Fridman (1:50:50.560)
And it will keep creating a lot of value in the sense that these models are applicable
François Chollet (1:50:56.240)
to a very wide range of problems that are relevant today.
Lex Fridman (1:51:00.000)
And we are only just getting started with applying these algorithms to every problem
François Chollet (1:51:05.120)
they could be solving.
Lex Fridman (1:51:06.320)
So deep learning will keep creating a lot of value for the time being.
François Chollet (1:51:10.160)
What's concerning, however, is that there's a lot of hype around deep learning and around
Lex Fridman (1:51:15.920)
AI.
François Chollet (1:51:16.240)
There are lots of people are overselling the capabilities of these systems, not just
François Chollet (1:51:22.000)
the capabilities, but also overselling the fact that they might be more or less, you
François Chollet (1:51:27.760)
know, brain like, like given the kind of a mystical aspect, these technologies and also
François Chollet (1:51:36.640)
overselling the pace of progress, which, you know, it might look fast in the sense that
François Chollet (1:51:43.840)
we have this exponentially increasing number of papers.
Lex Fridman (1:51:47.760)
But again, that's just a simple consequence of the fact that we have ever more people
François Chollet (1:51:52.960)
coming into the field.
Lex Fridman (1:51:54.400)
It doesn't mean the progress is actually exponentially fast.
François Chollet (1:51:58.640)
Let's say you're trying to raise money for your startup or your research lab.
François Chollet (1:52:02.720)
You might want to tell, you know, a grandiose story to investors about how deep learning
François Chollet (1:52:09.120)
is just like the brain and how it can solve all these incredible problems like self driving
Lex Fridman (1:52:14.240)
and robotics and so on.
Lex Fridman (1:52:15.760)
And maybe you can tell them that the field is progressing so fast and we are going to
Lex Fridman (1:52:19.440)
have AGI within 15 years or even 10 years.
Lex Fridman (1:52:23.040)
And none of this is true.
Lex Fridman (1:52:25.920)
And every time you're like saying these things and an investor or, you know, a decision maker
François Chollet (1:52:32.800)
believes them, well, this is like the equivalent of taking on credit card debt, but for trust,
Lex Fridman (1:52:41.680)
right?
Lex Fridman (1:52:42.480)
And maybe this will, you know, this will be what enables you to raise a lot of money,
Lex Fridman (1:52:50.160)
but ultimately you are creating damage, you are damaging the field.
Lex Fridman (1:52:54.320)
So that's the concern is that that debt, that's what happens with the other AI winters is
Lex Fridman (1:53:00.160)
the concern is you actually tweeted about this with autonomous vehicles, right?
François Chollet (1:53:04.160)
There's almost every single company now have promised that they will have full autonomous
Lex Fridman (1:53:08.960)
vehicles by 2021, 2022.
François Chollet (1:53:11.760)
That's a good example of the consequences of over hyping the capabilities of AI and
Lex Fridman (1:53:18.080)
the pace of progress.
Lex Fridman (1:53:19.280)
So because I work especially a lot recently in this area, I have a deep concern of what
François Chollet (1:53:25.200)
happens when all of these companies after I've invested billions have a meeting and
Lex Fridman (1:53:30.400)
say, how much do we actually, first of all, do we have an autonomous vehicle?
Lex Fridman (1:53:33.600)
The answer will definitely be no.
Lex Fridman (1:53:35.840)
And second will be, wait a minute, we've invested one, two, three, four billion dollars
Lex Fridman (1:53:40.560)
into this and we made no profit.
Lex Fridman (1:53:43.120)
And the reaction to that may be going very hard in other directions that might impact
Lex Fridman (1:53:49.200)
even other industries.
Lex Fridman (1:53:50.400)
And that's what we call an AI winter is when there is backlash where no one believes any
François Chollet (1:53:55.520)
of these promises anymore because they've turned that to be big lies the first time
François Chollet (1:53:59.360)
around.
Lex Fridman (1:54:00.240)
And this will definitely happen to some extent for autonomous vehicles because the public
Lex Fridman (1:54:06.000)
and decision makers have been convinced that around 2015, they've been convinced by these
François Chollet (1:54:13.360)
people who are trying to raise money for their startups and so on, that L5 driving was coming
François Chollet (1:54:19.600)
in maybe 2016, maybe 2017, maybe 2018.
Lex Fridman (1:54:22.880)
Now we're in 2019, we're still waiting for it.
Lex Fridman (1:54:27.600)
And so I don't believe we are going to have a full on AI winter because we have these
Lex Fridman (1:54:32.800)
technologies that are producing a tremendous amount of real value.
Lex Fridman (1:54:37.680)
But there is also too much hype.
Lex Fridman (1:54:39.920)
So there will be some backlash, especially there will be backlash.
Lex Fridman (1:54:44.960)
So some startups are trying to sell the dream of AGI and the fact that AGI is going to create
Lex Fridman (1:54:53.040)
infinite value.
François Chollet (1:54:53.760)
Like AGI is like a free lunch.
François Chollet (1:54:55.680)
Like if you can develop an AI system that passes a certain threshold of IQ or something,
François Chollet (1:55:02.800)
then suddenly you have infinite value.
Lex Fridman (1:55:04.400)
And well, there are actually lots of investors buying into this idea and they will wait maybe
François Chollet (1:55:14.160)
10, 15 years and nothing will happen.
Lex Fridman (1:55:17.760)
And the next time around, well, maybe there will be a new generation of investors.
François Chollet (1:55:22.560)
No one will care.
Lex Fridman (1:55:24.800)
Human memory is fairly short after all.
François Chollet (1:55:27.280)
I don't know about you, but because I've spoken about AGI sometimes poetically, I get a lot
François Chollet (1:55:34.320)
of emails from people giving me, they're usually like a large manifestos of they've, they say
François Chollet (1:55:42.000)
to me that they have created an AGI system or they know how to do it.
Lex Fridman (1:55:47.200)
And there's a long write up of how to do it.
François Chollet (1:55:48.880)
I get a lot of these emails, yeah.
François Chollet (1:55:50.560)
They're a little bit feel like it's generated by an AI system actually, but there's usually
François Chollet (1:55:57.760)
no diagram, you have a transformer generating crank papers about AGI.
Lex Fridman (1:56:06.640)
So the question is about, because you've been such a good, you have a good radar for crank
Lex Fridman (1:56:12.160)
papers, how do we know they're not onto something?
Lex Fridman (1:56:16.720)
How do I, so when you start to talk about AGI or anything like the reasoning benchmarks
Lex Fridman (1:56:24.240)
and so on, so something that doesn't have a benchmark, it's really difficult to know.
François Chollet (1:56:29.120)
I mean, I talked to Jeff Hawkins, who's really looking at neuroscience approaches to how,
Lex Fridman (1:56:35.200)
and there's some, there's echoes of really interesting ideas in at least Jeff's case,
Lex Fridman (1:56:41.520)
which he's showing.
Lex Fridman (1:56:43.280)
How do you usually think about this?
François Chollet (1:56:46.640)
Like preventing yourself from being too narrow minded and elitist about deep learning, it
François Chollet (1:56:52.880)
has to work on these particular benchmarks, otherwise it's trash.
Lex Fridman (1:56:56.720)
Well, you know, the thing is, intelligence does not exist in the abstract.
François Chollet (1:57:05.280)
Intelligence has to be applied.
Lex Fridman (1:57:07.200)
So if you don't have a benchmark, if you have an improvement in some benchmark, maybe it's
Lex Fridman (1:57:11.040)
a new benchmark, right?
François Chollet (1:57:12.400)
Maybe it's not something we've been looking at before, but you do need a problem that
François Chollet (1:57:16.640)
you're trying to solve.
Lex Fridman (1:57:17.360)
You're not going to come up with a solution without a problem.
Lex Fridman (1:57:20.000)
So you, general intelligence, I mean, you've clearly highlighted generalization.
François Chollet (1:57:26.320)
If you want to claim that you have an intelligence system, it should come with a benchmark.
François Chollet (1:57:31.200)
It should, yes, it should display capabilities of some kind.
François Chollet (1:57:35.760)
It should show that it can create some form of value, even if it's a very artificial form
François Chollet (1:57:41.840)
of value.
Lex Fridman (1:57:42.800)
And that's also the reason why you don't actually need to care about telling which papers have
François Chollet (1:57:48.800)
actually some hidden potential and which do not.
François Chollet (1:57:53.120)
Because if there is a new technique that's actually creating value, this is going to
François Chollet (1:57:59.200)
be brought to light very quickly because it's actually making a difference.
Lex Fridman (1:58:02.480)
So it's the difference between something that is ineffectual and something that is actually
François Chollet (1:58:08.160)
useful.
Lex Fridman (1:58:08.800)
And ultimately usefulness is our guide, not just in this field, but if you look at science
François Chollet (1:58:14.080)
in general, maybe there are many, many people over the years that have had some really interesting
Lex Fridman (1:58:19.440)
theories of everything, but they were just completely useless.
Lex Fridman (1:58:22.800)
And you don't actually need to tell the interesting theories from the useless theories.
Lex Fridman (1:58:28.000)
All you need is to see, is this actually having an effect on something else?
Lex Fridman (1:58:34.080)
Is this actually useful?
Lex Fridman (1:58:35.360)
Is this making an impact or not?
François Chollet (1:58:37.600)
That's beautifully put.
François Chollet (1:58:38.640)
I mean, the same applies to quantum mechanics, to string theory, to the holographic principle.
François Chollet (1:58:43.680)
We are doing deep learning because it works.
François Chollet (1:58:46.960)
Before it started working, people considered people working on neural networks as cranks
François Chollet (1:58:52.720)
very much.
Lex Fridman (1:58:54.560)
No one was working on this anymore.
Lex Fridman (1:58:56.320)
And now it's working, which is what makes it valuable.
Lex Fridman (1:58:59.120)
It's not about being right.
François Chollet (1:59:01.120)
It's about being effective.
Lex Fridman (1:59:02.560)
And nevertheless, the individual entities of this scientific mechanism, just like Yoshua
François Chollet (1:59:08.080)
Banjo or Jan Lekun, they, while being called cranks, stuck with it.
Lex Fridman (1:59:12.480)
Right?
François Chollet (1:59:12.880)
Yeah.
Lex Fridman (1:59:13.280)
And so us individual agents, even if everyone's laughing at us, just stick with it.
François Chollet (1:59:18.880)
If you believe you have something, you should stick with it and see it through.
Lex Fridman (1:59:23.520)
That's a beautiful inspirational message to end on.
François Chollet (1:59:25.920)
Francois, thank you so much for talking today.
Lex Fridman (1:59:27.600)
That was amazing.
François Chollet (1:59:28.640)
Thank you.
Lex Fridman (20:00.040)
that the number of scientists and engineers
François Chollet (20:03.760)
is in fact increasing exponentially.
Lex Fridman (20:06.520)
The amount of computational resources
François Chollet (20:08.400)
that are available to science
Lex Fridman (20:10.040)
is increasing exponentially and so on.
Lex Fridman (20:11.880)
So the resource consumption of science is exponential,
Lex Fridman (20:15.560)
but the output in terms of progress,
François Chollet (20:18.200)
in terms of significance, is linear.
Lex Fridman (20:21.000)
And the reason why is because,
Lex Fridman (20:23.120)
and even though science is regressively self improving,
Lex Fridman (20:26.000)
meaning that scientific progress
François Chollet (20:28.440)
turns into technological progress,
Lex Fridman (20:30.240)
which in turn helps science.
François Chollet (20:32.960)
If you look at computers, for instance,
Lex Fridman (20:35.280)
our products of science and computers
François Chollet (20:38.480)
are tremendously useful in speeding up science.
Lex Fridman (20:41.560)
The internet, same thing, the internet is a technology
François Chollet (20:43.840)
that's made possible by very recent scientific advances.
Lex Fridman (20:47.480)
And itself, because it enables scientists to network,
François Chollet (20:52.400)
to communicate, to exchange papers and ideas much faster,
Lex Fridman (20:55.520)
it is a way to speed up scientific progress.
Lex Fridman (20:57.440)
So even though you're looking
Lex Fridman (20:58.440)
at a regressively self improving system,
François Chollet (21:01.440)
it is consuming exponentially more resources
Lex Fridman (21:04.080)
to produce the same amount of problem solving, very much.
Lex Fridman (21:09.200)
So that's a fascinating way to paint it,
Lex Fridman (21:11.080)
and certainly that holds for the deep learning community.
François Chollet (21:14.960)
If you look at the temporal, what did you call it,
Lex Fridman (21:18.120)
the temporal density of significant ideas,
François Chollet (21:21.240)
if you look at in deep learning,
Lex Fridman (21:24.840)
I think, I'd have to think about that,
Lex Fridman (21:26.960)
but if you really look at significant ideas
Lex Fridman (21:29.040)
in deep learning, they might even be decreasing.
Lex Fridman (21:32.400)
So I do believe the per paper significance is decreasing,
Lex Fridman (21:39.600)
but the amount of papers
François Chollet (21:41.240)
is still today exponentially increasing.
Lex Fridman (21:43.440)
So I think if you look at an aggregate,
François Chollet (21:45.880)
my guess is that you would see a linear progress.
Lex Fridman (21:48.840)
If you were to sum the significance of all papers,
François Chollet (21:56.120)
you would see roughly in your progress.
Lex Fridman (21:58.640)
And in my opinion, it is not a coincidence
François Chollet (22:03.880)
that you're seeing linear progress in science
Lex Fridman (22:05.800)
despite exponential resource consumption.
François Chollet (22:07.720)
I think the resource consumption
Lex Fridman (22:10.280)
is dynamically adjusting itself to maintain linear progress
François Chollet (22:15.880)
because we as a community expect linear progress,
Lex Fridman (22:18.560)
meaning that if we start investing less
Lex Fridman (22:21.240)
and seeing less progress, it means that suddenly
Lex Fridman (22:23.600)
there are some lower hanging fruits that become available
Lex Fridman (22:26.800)
and someone's gonna step up and pick them, right?
Lex Fridman (22:31.280)
So it's very much like a market for discoveries and ideas.
Lex Fridman (22:36.920)
But there's another fundamental part
Lex Fridman (22:38.720)
which you're highlighting, which as a hypothesis
François Chollet (22:41.800)
as science or like the space of ideas,
Lex Fridman (22:45.160)
any one path you travel down,
François Chollet (22:48.160)
it gets exponentially more difficult
Lex Fridman (22:51.080)
to get a new way to develop new ideas.
Lex Fridman (22:54.720)
And your sense is that's gonna hold
Lex Fridman (22:57.640)
across our mysterious universe.
François Chollet (23:01.520)
Yes, well, exponential progress
Lex Fridman (23:03.360)
triggers exponential friction.
Lex Fridman (23:05.480)
So that if you tweak one part of the system,
Lex Fridman (23:07.440)
suddenly some other part becomes a bottleneck, right?
François Chollet (23:10.680)
For instance, let's say you develop some device
Lex Fridman (23:14.880)
that measures its own acceleration
Lex Fridman (23:17.160)
and then it has some engine
Lex Fridman (23:18.720)
and it outputs even more acceleration
François Chollet (23:20.800)
in proportion of its own acceleration
Lex Fridman (23:22.360)
and you drop it somewhere,
François Chollet (23:23.320)
it's not gonna reach infinite speed
Lex Fridman (23:25.240)
because it exists in a certain context.
Lex Fridman (23:29.080)
So the air around it is gonna generate friction
Lex Fridman (23:31.000)
and it's gonna block it at some top speed.
Lex Fridman (23:34.320)
And even if you were to consider the broader context
Lex Fridman (23:37.480)
and lift the bottleneck there,
François Chollet (23:39.840)
like the bottleneck of friction,
Lex Fridman (23:43.120)
then some other part of the system
François Chollet (23:45.120)
would start stepping in and creating exponential friction,
Lex Fridman (23:48.120)
maybe the speed of flight or whatever.
Lex Fridman (23:49.920)
And this definitely holds true
Lex Fridman (23:51.920)
when you look at the problem solving algorithm
François Chollet (23:54.960)
that is being run by science as an institution,
Lex Fridman (23:58.160)
science as a system.
François Chollet (23:59.720)
As you make more and more progress,
Lex Fridman (24:01.720)
despite having this recursive self improvement component,
François Chollet (24:06.760)
you are encountering exponential friction.
Lex Fridman (24:09.840)
The more researchers you have working on different ideas,
François Chollet (24:13.480)
the more overhead you have
Lex Fridman (24:14.880)
in terms of communication across researchers.
Lex Fridman (24:18.040)
If you look at, you were mentioning quantum mechanics, right?
Lex Fridman (24:22.920)
Well, if you want to start making significant discoveries
François Chollet (24:26.880)
today, significant progress in quantum mechanics,
Lex Fridman (24:29.680)
there is an amount of knowledge you have to ingest,
François Chollet (24:33.000)
which is huge.
Lex Fridman (24:34.080)
So there's a very large overhead
François Chollet (24:36.520)
to even start to contribute.
Lex Fridman (24:39.240)
There's a large amount of overhead
François Chollet (24:40.680)
to synchronize across researchers and so on.
Lex Fridman (24:44.040)
And of course, the significant practical experiments
François Chollet (24:48.600)
are going to require exponentially expensive equipment
Lex Fridman (24:52.160)
because the easier ones have already been run, right?
Lex Fridman (24:56.480)
So in your senses, there's no way escaping,
Lex Fridman (25:00.480)
there's no way of escaping this kind of friction
François Chollet (25:04.480)
with artificial intelligence systems.
Lex Fridman (25:08.600)
Yeah, no, I think science is a very good way
François Chollet (25:11.520)
to model what would happen with a superhuman
Lex Fridman (25:14.280)
recursive research improving AI.
François Chollet (25:16.440)
That's your sense, I mean, the...
Lex Fridman (25:18.240)
That's my intuition.
François Chollet (25:19.680)
It's not like a mathematical proof of anything.
Lex Fridman (25:23.400)
That's not my point.
François Chollet (25:24.400)
Like, I'm not trying to prove anything.
Lex Fridman (25:26.600)
I'm just trying to make an argument
François Chollet (25:27.920)
to question the narrative of intelligence explosion,
Lex Fridman (25:31.160)
which is quite a dominant narrative.
Lex Fridman (25:32.880)
And you do get a lot of pushback if you go against it.
Lex Fridman (25:35.840)
Because, so for many people, right,
François Chollet (25:39.320)
AI is not just a subfield of computer science.
Lex Fridman (25:42.200)
It's more like a belief system.
François Chollet (25:44.120)
Like this belief that the world is headed towards an event,
Lex Fridman (25:48.640)
the singularity, past which, you know, AI will become...
François Chollet (25:55.040)
will go exponential very much,
Lex Fridman (25:57.080)
and the world will be transformed,
Lex Fridman (25:58.600)
and humans will become obsolete.
Lex Fridman (26:00.840)
And if you go against this narrative,
François Chollet (26:03.880)
because it is not really a scientific argument,
Lex Fridman (26:06.920)
but more of a belief system,
François Chollet (26:08.880)
it is part of the identity of many people.
Lex Fridman (26:11.240)
If you go against this narrative,
François Chollet (26:12.600)
it's like you're attacking the identity
Lex Fridman (26:14.400)
of people who believe in it.
François Chollet (26:15.560)
It's almost like saying God doesn't exist,
Lex Fridman (26:17.640)
or something.
Lex Fridman (26:19.000)
So you do get a lot of pushback
Lex Fridman (26:21.880)
if you try to question these ideas.
François Chollet (26:24.040)
First of all, I believe most people,
Lex Fridman (26:26.520)
they might not be as eloquent or explicit as you're being,
Lex Fridman (26:29.240)
but most people in computer science
Lex Fridman (26:30.920)
are most people who actually have built
François Chollet (26:33.000)
anything that you could call AI, quote, unquote,
Lex Fridman (26:36.360)
would agree with you.
François Chollet (26:38.080)
They might not be describing in the same kind of way.
Lex Fridman (26:40.560)
It's more, so the pushback you're getting
François Chollet (26:43.960)
is from people who get attached to the narrative
Lex Fridman (26:48.080)
from, not from a place of science,
Lex Fridman (26:51.000)
but from a place of imagination.
Lex Fridman (26:53.400)
That's correct, that's correct.
Lex Fridman (26:54.760)
So why do you think that's so appealing?
Lex Fridman (26:56.920)
Because the usual dreams that people have
François Chollet (27:02.120)
when you create a superintelligence system
Lex Fridman (27:03.960)
past the singularity,
François Chollet (27:05.120)
that what people imagine is somehow always destructive.
Lex Fridman (27:09.440)
Do you have, if you were put on your psychology hat,
François Chollet (27:12.240)
what's, why is it so appealing to imagine
Lex Fridman (27:17.400)
the ways that all of human civilization will be destroyed?
François Chollet (27:20.760)
I think it's a good story.
Lex Fridman (27:22.080)
You know, it's a good story.
Lex Fridman (27:23.120)
And very interestingly, it mirrors a religious stories,
Lex Fridman (27:28.160)
right, religious mythology.
François Chollet (27:30.560)
If you look at the mythology of most civilizations,
Lex Fridman (27:34.360)
it's about the world being headed towards some final events
François Chollet (27:38.280)
in which the world will be destroyed
Lex Fridman (27:40.480)
and some new world order will arise
François Chollet (27:42.800)
that will be mostly spiritual,
Lex Fridman (27:44.920)
like the apocalypse followed by a paradise probably, right?
François Chollet (27:49.400)
It's a very appealing story on a fundamental level.
Lex Fridman (27:52.600)
And we all need stories.
François Chollet (27:54.560)
We all need stories to structure the way we see the world,
Lex Fridman (27:58.160)
especially at timescales
Lex Fridman (27:59.960)
that are beyond our ability to make predictions, right?
Lex Fridman (28:04.520)
So on a more serious non exponential explosion,
François Chollet (28:08.840)
question, do you think there will be a time
Lex Fridman (28:15.000)
when we'll create something like human level intelligence
François Chollet (28:19.800)
or intelligent systems that will make you sit back
Lex Fridman (28:23.800)
and be just surprised at damn how smart this thing is?
François Chollet (28:28.520)
That doesn't require exponential growth
Lex Fridman (28:30.160)
or an exponential improvement,
Lex Fridman (28:32.120)
but what's your sense of the timeline and so on
Lex Fridman (28:35.600)
that you'll be really surprised at certain capabilities?
Lex Fridman (28:41.080)
And we'll talk about limitations and deep learning.
Lex Fridman (28:42.560)
So do you think in your lifetime,
Lex Fridman (28:44.480)
you'll be really damn surprised?
Lex Fridman (28:46.600)
Around 2013, 2014, I was many times surprised
François Chollet (28:51.440)
by the capabilities of deep learning actually.
Lex Fridman (28:53.960)
That was before we had assessed exactly
Lex Fridman (28:55.920)
what deep learning could do and could not do.
Lex Fridman (28:57.880)
And it felt like a time of immense potential.
Lex Fridman (29:00.600)
And then we started narrowing it down,
Lex Fridman (29:03.080)
but I was very surprised.
François Chollet (29:04.360)
I would say it has already happened.
Lex Fridman (29:07.120)
Was there a moment, there must've been a day in there
François Chollet (29:10.800)
where your surprise was almost bordering
Lex Fridman (29:14.360)
on the belief of the narrative that we just discussed.
François Chollet (29:19.440)
Was there a moment,
Lex Fridman (29:20.800)
because you've written quite eloquently
François Chollet (29:22.400)
about the limits of deep learning,
Lex Fridman (29:23.960)
was there a moment that you thought
Lex Fridman (29:25.760)
that maybe deep learning is limitless?
Lex Fridman (29:30.000)
No, I don't think I've ever believed this.
Lex Fridman (29:32.400)
What was really shocking is that it worked.
Lex Fridman (29:35.560)
It worked at all, yeah.
Lex Fridman (29:37.640)
But there's a big jump between being able
Lex Fridman (29:40.520)
to do really good computer vision
Lex Fridman (29:43.400)
and human level intelligence.
Lex Fridman (29:44.920)
So I don't think at any point I wasn't under the impression
François Chollet (29:49.520)
that the results we got in computer vision
Lex Fridman (29:51.280)
meant that we were very close to human level intelligence.
François Chollet (29:54.080)
I don't think we're very close to human level intelligence.
Lex Fridman (29:56.040)
I do believe that there's no reason
Lex Fridman (29:58.520)
why we won't achieve it at some point.
Lex Fridman (30:01.760)
I also believe that it's the problem
François Chollet (30:06.400)
with talking about human level intelligence
Lex Fridman (30:08.560)
that implicitly you're considering
François Chollet (30:11.240)
like an axis of intelligence with different levels,
Lex Fridman (30:14.360)
but that's not really how intelligence works.
François Chollet (30:16.720)
Intelligence is very multi dimensional.
Lex Fridman (30:19.480)
And so there's the question of capabilities,
Lex Fridman (30:22.480)
but there's also the question of being human like,
Lex Fridman (30:25.560)
and it's two very different things.
François Chollet (30:27.040)
Like you can build potentially
Lex Fridman (30:28.280)
very advanced intelligent agents
François Chollet (30:30.640)
that are not human like at all.
Lex Fridman (30:32.640)
And you can also build very human like agents.
Lex Fridman (30:35.240)
And these are two very different things, right?
Lex Fridman (30:37.840)
Right.
François Chollet (30:38.760)
Let's go from the philosophical to the practical.
Lex Fridman (30:42.240)
Can you give me a history of Keras
Lex Fridman (30:44.240)
and all the major deep learning frameworks
Lex Fridman (30:46.440)
that you kind of remember in relation to Keras
Lex Fridman (30:48.480)
and in general, TensorFlow, Theano, the old days.
Lex Fridman (30:52.040)
Can you give a brief overview Wikipedia style history
Lex Fridman (30:55.400)
and your role in it before we return to AGI discussions?
Lex Fridman (30:59.120)
Yeah, that's a broad topic.
Lex Fridman (31:00.720)
So I started working on Keras.
Lex Fridman (31:04.920)
It was the name Keras at the time.
François Chollet (31:06.240)
I actually picked the name like
Lex Fridman (31:08.320)
just the day I was going to release it.
Lex Fridman (31:10.200)
So I started working on it in February, 2015.
Lex Fridman (31:14.800)
And so at the time there weren't too many people
François Chollet (31:17.240)
working on deep learning, maybe like fewer than 10,000.
Lex Fridman (31:20.320)
The software tooling was not really developed.
Lex Fridman (31:25.320)
So the main deep learning library was Cafe,
Lex Fridman (31:28.800)
which was mostly C++.
Lex Fridman (31:30.840)
Why do you say Cafe was the main one?
Lex Fridman (31:32.760)
Cafe was vastly more popular than Theano
François Chollet (31:36.000)
in late 2014, early 2015.
Lex Fridman (31:38.920)
Cafe was the one library that everyone was using
François Chollet (31:42.400)
for computer vision.
Lex Fridman (31:43.400)
And computer vision was the most popular problem
François Chollet (31:46.120)
in deep learning at the time.
Lex Fridman (31:46.960)
Absolutely.
François Chollet (31:47.800)
Like ConvNets was like the subfield of deep learning
Lex Fridman (31:50.440)
that everyone was working on.
Lex Fridman (31:53.160)
So myself, so in late 2014,
Lex Fridman (31:57.680)
I was actually interested in RNNs,
François Chollet (32:00.600)
in recurrent neural networks,
Lex Fridman (32:01.760)
which was a very niche topic at the time, right?
François Chollet (32:05.800)
It really took off around 2016.
Lex Fridman (32:08.640)
And so I was looking for good tools.
François Chollet (32:11.080)
I had used Torch 7, I had used Theano,
Lex Fridman (32:14.800)
used Theano a lot in Kaggle competitions.
François Chollet (32:19.320)
I had used Cafe.
Lex Fridman (32:20.840)
And there was no like good solution for RNNs at the time.
François Chollet (32:25.840)
Like there was no reusable open source implementation
Lex Fridman (32:28.640)
of an LSTM, for instance.
Lex Fridman (32:30.000)
So I decided to build my own.
Lex Fridman (32:32.920)
And at first, the pitch for that was,
François Chollet (32:35.440)
it was gonna be mostly around LSTM recurrent neural networks.
Lex Fridman (32:39.960)
It was gonna be in Python.
François Chollet (32:42.280)
An important decision at the time
Lex Fridman (32:44.280)
that was kind of not obvious
François Chollet (32:45.440)
is that the models would be defined via Python code,
Lex Fridman (32:50.360)
which was kind of like going against the mainstream
François Chollet (32:54.400)
at the time because Cafe, Pylon 2, and so on,
Lex Fridman (32:58.000)
like all the big libraries were actually going
François Chollet (33:00.600)
with the approach of setting configuration files
Lex Fridman (33:03.520)
in YAML to define models.
Lex Fridman (33:05.560)
So some libraries were using code to define models,
Lex Fridman (33:08.840)
like Torch 7, obviously, but that was not Python.
François Chollet (33:12.280)
Lasagne was like a Theano based very early library
Lex Fridman (33:16.680)
that was, I think, developed, I don't remember exactly,
François Chollet (33:18.640)
probably late 2014.
Lex Fridman (33:20.240)
It's Python as well.
François Chollet (33:21.200)
It's Python as well.
Lex Fridman (33:22.040)
It was like on top of Theano.
Lex Fridman (33:24.320)
And so I started working on something
Lex Fridman (33:29.480)
and the value proposition at the time was that
François Chollet (33:32.520)
not only what I think was the first
Lex Fridman (33:36.240)
reusable open source implementation of LSTM,
François Chollet (33:40.400)
you could combine RNNs and covenants
Lex Fridman (33:44.440)
with the same library,
François Chollet (33:45.440)
which is not really possible before,
Lex Fridman (33:46.920)
like Cafe was only doing covenants.
Lex Fridman (33:50.440)
And it was kind of easy to use
Lex Fridman (33:52.560)
because, so before I was using Theano,
François Chollet (33:54.440)
I was actually using scikitlin
Lex Fridman (33:55.680)
and I loved scikitlin for its usability.
Lex Fridman (33:58.320)
So I drew a lot of inspiration from scikitlin
Lex Fridman (34:01.560)
when I made Keras.
François Chollet (34:02.400)
It's almost like scikitlin for neural networks.
Lex Fridman (34:05.600)
The fit function.
François Chollet (34:06.680)
Exactly, the fit function,
Lex Fridman (34:07.960)
like reducing a complex string loop
Lex Fridman (34:10.800)
to a single function call, right?
Lex Fridman (34:12.880)
And of course, some people will say,
François Chollet (34:14.880)
this is hiding a lot of details,
Lex Fridman (34:16.320)
but that's exactly the point, right?
François Chollet (34:18.680)
The magic is the point.
Lex Fridman (34:20.280)
So it's magical, but in a good way.
François Chollet (34:22.680)
It's magical in the sense that it's delightful.
Lex Fridman (34:24.960)
Yeah, yeah.
François Chollet (34:26.160)
I'm actually quite surprised.
Lex Fridman (34:27.640)
I didn't know that it was born out of desire
François Chollet (34:29.600)
to implement RNNs and LSTMs.
Lex Fridman (34:32.480)
It was.
François Chollet (34:33.320)
That's fascinating.
Lex Fridman (34:34.160)
So you were actually one of the first people
François Chollet (34:36.040)
to really try to attempt
Lex Fridman (34:37.960)
to get the major architectures together.
Lex Fridman (34:41.000)
And it's also interesting.
Lex Fridman (34:42.760)
You made me realize that that was a design decision at all
François Chollet (34:45.160)
is defining the model and code.
Lex Fridman (34:47.360)
Just, I'm putting myself in your shoes,
François Chollet (34:49.920)
whether the YAML, especially if cafe was the most popular.
Lex Fridman (34:53.200)
It was the most popular by far.
François Chollet (34:54.720)
If I was, if I were, yeah, I don't,
Lex Fridman (34:58.480)
I didn't like the YAML thing,
Lex Fridman (34:59.560)
but it makes more sense that you will put
Lex Fridman (35:02.840)
in a configuration file, the definition of a model.
François Chollet (35:05.720)
That's an interesting gutsy move
Lex Fridman (35:07.200)
to stick with defining it in code.
François Chollet (35:10.040)
Just if you look back.
Lex Fridman (35:11.600)
Other libraries were doing it as well,
Lex Fridman (35:13.480)
but it was definitely the more niche option.
Lex Fridman (35:16.320)
Yeah.
François Chollet (35:17.160)
Okay, Keras and then.
Lex Fridman (35:18.360)
So I released Keras in March, 2015,
Lex Fridman (35:21.520)
and it got users pretty much from the start.
Lex Fridman (35:24.160)
So the deep learning community was very, very small
François Chollet (35:25.800)
at the time.
Lex Fridman (35:27.240)
Lots of people were starting to be interested in LSTM.
Lex Fridman (35:30.600)
So it was gonna release it at the right time
Lex Fridman (35:32.440)
because it was offering an easy to use LSTM implementation.
François Chollet (35:35.560)
Exactly at the time where lots of people started
Lex Fridman (35:37.680)
to be intrigued by the capabilities of RNN, RNNs for NLP.
Lex Fridman (35:42.280)
So it grew from there.
Lex Fridman (35:43.920)
Then I joined Google about six months later,
Lex Fridman (35:51.480)
and that was actually completely unrelated to Keras.
Lex Fridman (35:54.920)
So I actually joined a research team
François Chollet (35:57.080)
working on image classification,
Lex Fridman (35:59.520)
mostly like computer vision.
Lex Fridman (36:00.680)
So I was doing computer vision research
Lex Fridman (36:02.320)
at Google initially.
Lex Fridman (36:03.640)
And immediately when I joined Google,
Lex Fridman (36:05.520)
I was exposed to the early internal version of TensorFlow.
Lex Fridman (36:10.520)
And the way it appeared to me at the time,
Lex Fridman (36:13.920)
and it was definitely the way it was at the time
François Chollet (36:15.720)
is that this was an improved version of Theano.
Lex Fridman (36:20.760)
So I immediately knew I had to port Keras
François Chollet (36:24.720)
to this new TensorFlow thing.
Lex Fridman (36:26.800)
And I was actually very busy as a noobler,
François Chollet (36:29.800)
as a new Googler.
Lex Fridman (36:31.600)
So I had not time to work on that.
Lex Fridman (36:34.520)
But then in November, I think it was November, 2015,
Lex Fridman (36:38.680)
TensorFlow got released.
Lex Fridman (36:41.240)
And it was kind of like my wake up call
Lex Fridman (36:44.560)
that, hey, I had to actually go and make it happen.
Lex Fridman (36:47.320)
So in December, I ported Keras to run on top of TensorFlow,
Lex Fridman (36:52.200)
but it was not exactly a port.
François Chollet (36:53.320)
It was more like a refactoring
Lex Fridman (36:55.280)
where I was abstracting away
François Chollet (36:57.920)
all the backend functionality into one module
Lex Fridman (37:00.480)
so that the same code base
François Chollet (37:02.320)
could run on top of multiple backends.
Lex Fridman (37:05.080)
So on top of TensorFlow or Theano.
Lex Fridman (37:07.440)
And for the next year,
Lex Fridman (37:09.760)
Theano stayed as the default option.
François Chollet (37:15.400)
It was easier to use, somewhat less buggy.
Lex Fridman (37:20.640)
It was much faster, especially when it came to audience.
Lex Fridman (37:23.360)
But eventually, TensorFlow overtook it.
Lex Fridman (37:27.480)
And TensorFlow, the early TensorFlow,
Lex Fridman (37:30.200)
has similar architectural decisions as Theano, right?
Lex Fridman (37:33.960)
So it was a natural transition.
François Chollet (37:37.440)
Yeah, absolutely.
Lex Fridman (37:38.320)
So what, I mean, that still Keras is a side,
Lex Fridman (37:42.960)
almost fun project, right?
Lex Fridman (37:45.280)
Yeah, so it was not my job assignment.
François Chollet (37:49.040)
It was not.
Lex Fridman (37:50.360)
I was doing it on the side.
Lex Fridman (37:52.240)
And even though it grew to have a lot of users
Lex Fridman (37:55.840)
for a deep learning library at the time, like Stroud 2016,
Lex Fridman (37:59.600)
but I wasn't doing it as my main job.
Lex Fridman (38:02.480)
So things started changing in,
François Chollet (38:04.760)
I think it must have been maybe October, 2016.
Lex Fridman (38:10.200)
So one year later.
Lex Fridman (38:12.360)
So Rajat, who was the lead on TensorFlow,
Lex Fridman (38:15.240)
basically showed up one day in our building
François Chollet (38:19.240)
where I was doing like,
Lex Fridman (38:20.080)
so I was doing research and things like,
Lex Fridman (38:21.640)
so I did a lot of computer vision research,
Lex Fridman (38:24.640)
also collaborations with Christian Zighetti
Lex Fridman (38:27.560)
and deep learning for theorem proving.
Lex Fridman (38:29.640)
It was a really interesting research topic.
Lex Fridman (38:34.520)
And so Rajat was saying,
Lex Fridman (38:37.640)
hey, we saw Keras, we like it.
François Chollet (38:41.040)
We saw that you're at Google.
Lex Fridman (38:42.440)
Why don't you come over for like a quarter
Lex Fridman (38:45.280)
and work with us?
Lex Fridman (38:47.280)
And I was like, yeah, that sounds like a great opportunity.
François Chollet (38:49.240)
Let's do it.
Lex Fridman (38:50.400)
And so I started working on integrating the Keras API
François Chollet (38:55.720)
into TensorFlow more tightly.
Lex Fridman (38:57.320)
So what followed up is a sort of like temporary
François Chollet (39:02.640)
TensorFlow only version of Keras
Lex Fridman (39:05.480)
that was in TensorFlow.com Trib for a while.
Lex Fridman (39:09.320)
And finally moved to TensorFlow Core.
Lex Fridman (39:12.200)
And I've never actually gotten back
François Chollet (39:15.360)
to my old team doing research.
Lex Fridman (39:17.600)
Well, it's kind of funny that somebody like you
François Chollet (39:22.320)
who dreams of, or at least sees the power of AI systems
Lex Fridman (39:28.960)
that reason and theorem proving we'll talk about
François Chollet (39:31.680)
has also created a system that makes the most basic
Lex Fridman (39:36.520)
kind of Lego building that is deep learning
François Chollet (39:40.400)
super accessible, super easy.
Lex Fridman (39:42.640)
So beautifully so.
François Chollet (39:43.800)
It's a funny irony that you're both,
Lex Fridman (39:47.720)
you're responsible for both things,
Lex Fridman (39:49.120)
but so TensorFlow 2.0 is kind of, there's a sprint.
Lex Fridman (39:54.000)
I don't know how long it'll take,
Lex Fridman (39:55.080)
but there's a sprint towards the finish.
Lex Fridman (39:56.960)
What do you look, what are you working on these days?
Lex Fridman (40:01.040)
What are you excited about?
Lex Fridman (40:02.160)
What are you excited about in 2.0?
François Chollet (40:04.280)
I mean, eager execution.
Lex Fridman (40:05.760)
There's so many things that just make it a lot easier
François Chollet (40:08.440)
to work.
Lex Fridman (40:09.760)
What are you excited about and what's also really hard?
Lex Fridman (40:13.640)
What are the problems you have to kind of solve?
Lex Fridman (40:15.800)
So I've spent the past year and a half working on
François Chollet (40:19.080)
TensorFlow 2.0 and it's been a long journey.
Lex Fridman (40:22.920)
I'm actually extremely excited about it.
François Chollet (40:25.080)
I think it's a great product.
Lex Fridman (40:26.440)
It's a delightful product compared to TensorFlow 1.0.
François Chollet (40:29.360)
We've made huge progress.
Lex Fridman (40:32.640)
So on the Keras side, what I'm really excited about is that,
Lex Fridman (40:37.400)
so previously Keras has been this very easy to use
Lex Fridman (40:42.400)
high level interface to do deep learning.
Lex Fridman (40:45.840)
But if you wanted to,
Lex Fridman (40:50.520)
if you wanted a lot of flexibility,
François Chollet (40:53.040)
the Keras framework was probably not the optimal way
Lex Fridman (40:57.520)
to do things compared to just writing everything
François Chollet (40:59.760)
from scratch.
Lex Fridman (41:01.800)
So in some way, the framework was getting in the way.
Lex Fridman (41:04.680)
And in TensorFlow 2.0, you don't have this at all, actually.
Lex Fridman (41:07.960)
You have the usability of the high level interface,
Lex Fridman (41:11.040)
but you have the flexibility of this lower level interface.
Lex Fridman (41:14.480)
And you have this spectrum of workflows
François Chollet (41:16.800)
where you can get more or less usability
Lex Fridman (41:21.560)
and flexibility trade offs depending on your needs, right?
François Chollet (41:26.640)
You can write everything from scratch
Lex Fridman (41:29.680)
and you get a lot of help doing so
François Chollet (41:32.320)
by subclassing models and writing some train loops
Lex Fridman (41:36.400)
using ego execution.
François Chollet (41:38.200)
It's very flexible, it's very easy to debug,
Lex Fridman (41:40.160)
it's very powerful.
Lex Fridman (41:42.280)
But all of this integrates seamlessly
Lex Fridman (41:45.000)
with higher level features up to the classic Keras workflows,
François Chollet (41:49.440)
which are very scikit learn like
Lex Fridman (41:51.560)
and are ideal for a data scientist,
François Chollet (41:56.040)
machine learning engineer type of profile.
Lex Fridman (41:58.240)
So now you can have the same framework
François Chollet (42:00.840)
offering the same set of APIs
Lex Fridman (42:02.880)
that enable a spectrum of workflows
François Chollet (42:05.000)
that are more or less low level, more or less high level
Lex Fridman (42:08.560)
that are suitable for profiles ranging from researchers
François Chollet (42:13.520)
to data scientists and everything in between.
Lex Fridman (42:15.560)
Yeah, so that's super exciting.
François Chollet (42:16.960)
I mean, it's not just that,
Lex Fridman (42:18.400)
it's connected to all kinds of tooling.
François Chollet (42:21.680)
You can go on mobile, you can go with TensorFlow Lite,
Lex Fridman (42:24.520)
you can go in the cloud or serving and so on.
François Chollet (42:27.240)
It all is connected together.
Lex Fridman (42:28.960)
Now some of the best software written ever
François Chollet (42:31.880)
is often done by one person, sometimes two.
Lex Fridman (42:36.880)
So with a Google, you're now seeing sort of Keras
François Chollet (42:40.800)
having to be integrated in TensorFlow,
Lex Fridman (42:42.840)
I'm sure has a ton of engineers working on.
Lex Fridman (42:46.800)
And there's, I'm sure a lot of tricky design decisions
Lex Fridman (42:51.040)
to be made.
Lex Fridman (42:52.200)
How does that process usually happen
Lex Fridman (42:54.440)
from at least your perspective?
Lex Fridman (42:56.800)
What are the debates like?
Lex Fridman (43:00.720)
Is there a lot of thinking,
Lex Fridman (43:04.200)
considering different options and so on?
Lex Fridman (43:06.880)
Yes.
Lex Fridman (43:08.160)
So a lot of the time I spend at Google
Lex Fridman (43:12.640)
is actually discussing design discussions, right?
François Chollet (43:17.280)
Writing design docs, participating in design review meetings
Lex Fridman (43:20.480)
and so on.
François Chollet (43:22.080)
This is as important as actually writing a code.
Lex Fridman (43:25.240)
Right.
Lex Fridman (43:26.080)
So there's a lot of thought, there's a lot of thought
Lex Fridman (43:28.120)
and a lot of care that is taken
François Chollet (43:32.280)
in coming up with these decisions
Lex Fridman (43:34.160)
and taking into account all of our users
François Chollet (43:37.160)
because TensorFlow has this extremely diverse user base,
Lex Fridman (43:40.680)
right?
François Chollet (43:41.520)
It's not like just one user segment
Lex Fridman (43:43.120)
where everyone has the same needs.
François Chollet (43:45.480)
We have small scale production users,
Lex Fridman (43:47.640)
large scale production users.
François Chollet (43:49.520)
We have startups, we have researchers,
Lex Fridman (43:53.720)
you know, it's all over the place.
Lex Fridman (43:55.080)
And we have to cater to all of their needs.
Lex Fridman (43:57.560)
If I just look at the standard debates
François Chollet (44:00.040)
of C++ or Python, there's some heated debates.
Lex Fridman (44:04.000)
Do you have those at Google?
François Chollet (44:06.000)
I mean, they're not heated in terms of emotionally,
Lex Fridman (44:08.080)
but there's probably multiple ways to do it, right?
Lex Fridman (44:10.800)
So how do you arrive through those design meetings
Lex Fridman (44:14.040)
at the best way to do it?
François Chollet (44:15.440)
Especially in deep learning where the field is evolving
Lex Fridman (44:19.280)
as you're doing it.
Lex Fridman (44:21.880)
Is there some magic to it?
Lex Fridman (44:23.600)
Is there some magic to the process?
François Chollet (44:26.240)
I don't know if there's magic to the process,
Lex Fridman (44:28.280)
but there definitely is a process.
Lex Fridman (44:30.640)
So making design decisions
Lex Fridman (44:33.760)
is about satisfying a set of constraints,
Lex Fridman (44:36.080)
but also trying to do so in the simplest way possible,
Lex Fridman (44:39.920)
because this is what can be maintained,
François Chollet (44:42.240)
this is what can be expanded in the future.
Lex Fridman (44:44.920)
So you don't want to naively satisfy the constraints
François Chollet (44:49.120)
by just, you know, for each capability you need available,
Lex Fridman (44:51.880)
you're gonna come up with one argument in your API
Lex Fridman (44:53.960)
and so on.
Lex Fridman (44:54.800)
You want to design APIs that are modular and hierarchical
Lex Fridman (45:00.640)
so that they have an API surface
Lex Fridman (45:04.080)
that is as small as possible, right?
Lex Fridman (45:07.040)
And you want this modular hierarchical architecture
Lex Fridman (45:11.640)
to reflect the way that domain experts
François Chollet (45:14.560)
think about the problem.
Lex Fridman (45:16.400)
Because as a domain expert,
François Chollet (45:17.880)
when you are reading about a new API,
Lex Fridman (45:19.840)
you're reading a tutorial or some docs pages,
François Chollet (45:24.760)
you already have a way that you're thinking about the problem.
Lex Fridman (45:28.200)
You already have like certain concepts in mind
Lex Fridman (45:32.320)
and you're thinking about how they relate together.
Lex Fridman (45:35.680)
And when you're reading docs,
François Chollet (45:37.200)
you're trying to build as quickly as possible
Lex Fridman (45:40.280)
a mapping between the concepts featured in your API
Lex Fridman (45:45.280)
and the concepts in your mind.
Lex Fridman (45:46.800)
So you're trying to map your mental model
François Chollet (45:48.880)
as a domain expert to the way things work in the API.
Lex Fridman (45:53.600)
So you need an API and an underlying implementation
François Chollet (45:57.040)
that are reflecting the way people think about these things.
Lex Fridman (46:00.120)
So in minimizing the time it takes to do the mapping.
François Chollet (46:02.880)
Yes, minimizing the time,
Lex Fridman (46:04.680)
the cognitive load there is
François Chollet (46:06.560)
in ingesting this new knowledge about your API.
Lex Fridman (46:10.920)
An API should not be self referential
François Chollet (46:13.160)
or referring to implementation details.
Lex Fridman (46:15.520)
It should only be referring to domain specific concepts
François Chollet (46:19.160)
that people already understand.
Lex Fridman (46:23.240)
Brilliant.
Lex Fridman (46:24.480)
So what's the future of Keras and TensorFlow look like?
Lex Fridman (46:27.560)
What does TensorFlow 3.0 look like?
Lex Fridman (46:30.600)
So that's kind of too far in the future for me to answer,
Lex Fridman (46:33.720)
especially since I'm not even the one making these decisions.
François Chollet (46:37.800)
Okay.
Lex Fridman (46:39.080)
But so from my perspective,
François Chollet (46:41.240)
which is just one perspective
Lex Fridman (46:43.200)
among many different perspectives on the TensorFlow team,
François Chollet (46:47.200)
I'm really excited by developing even higher level APIs,
Lex Fridman (46:52.360)
higher level than Keras.
François Chollet (46:53.560)
I'm really excited by hyperparameter tuning,
Lex Fridman (46:56.480)
by automated machine learning, AutoML.
François Chollet (47:01.120)
I think the future is not just, you know,
Lex Fridman (47:03.200)
defining a model like you were assembling Lego blocks
Lex Fridman (47:07.600)
and then collect fit on it.
Lex Fridman (47:09.200)
It's more like an automagical model
François Chollet (47:13.680)
that would just look at your data
Lex Fridman (47:16.080)
and optimize the objective you're after, right?
Lex Fridman (47:19.040)
So that's what I'm looking into.
Lex Fridman (47:23.040)
Yeah, so you put the baby into a room with the problem
Lex Fridman (47:26.480)
and come back a few hours later
Lex Fridman (47:28.760)
with a fully solved problem.
François Chollet (47:30.960)
Exactly, it's not like a box of Legos.
Lex Fridman (47:33.560)
It's more like the combination of a kid
François Chollet (47:35.920)
that's really good at Legos and a box of Legos.
Lex Fridman (47:38.800)
It's just building the thing on its own.
François Chollet (47:41.520)
Very nice.
Lex Fridman (47:42.680)
So that's an exciting future.
François Chollet (47:44.160)
I think there's a huge amount of applications
Lex Fridman (47:46.080)
and revolutions to be had
François Chollet (47:49.920)
under the constraints of the discussion we previously had.
Lex Fridman (47:52.640)
But what do you think of the current limits of deep learning?
François Chollet (47:57.480)
If we look specifically at these function approximators
Lex Fridman (48:03.840)
that tries to generalize from data.
François Chollet (48:06.160)
You've talked about local versus extreme generalization.
Lex Fridman (48:11.120)
You mentioned that neural networks don't generalize well
Lex Fridman (48:13.280)
and humans do.
Lex Fridman (48:14.560)
So there's this gap.
Lex Fridman (48:17.640)
And you've also mentioned that extreme generalization
Lex Fridman (48:20.840)
requires something like reasoning to fill those gaps.
Lex Fridman (48:23.960)
So how can we start trying to build systems like that?
Lex Fridman (48:27.560)
Right, yeah, so this is by design, right?
François Chollet (48:30.600)
Deep learning models are like huge parametric models,
Lex Fridman (48:37.080)
differentiable, so continuous,
François Chollet (48:39.280)
that go from an input space to an output space.
Lex Fridman (48:42.680)
And they're trained with gradient descent.
Lex Fridman (48:44.120)
So they're trained pretty much point by point.
Lex Fridman (48:47.160)
They are learning a continuous geometric morphing
François Chollet (48:50.520)
from an input vector space to an output vector space.
Lex Fridman (48:55.320)
And because this is done point by point,
François Chollet (48:58.960)
a deep neural network can only make sense
Lex Fridman (49:02.200)
of points in experience space that are very close
François Chollet (49:05.880)
to things that it has already seen in string data.
Lex Fridman (49:08.520)
At best, it can do interpolation across points.
Lex Fridman (49:13.840)
But that means in order to train your network,
Lex Fridman (49:17.360)
you need a dense sampling of the input cross output space,
François Chollet (49:22.880)
almost a point by point sampling,
Lex Fridman (49:25.240)
which can be very expensive if you're dealing
François Chollet (49:27.160)
with complex real world problems,
Lex Fridman (49:29.320)
like autonomous driving, for instance, or robotics.
François Chollet (49:33.240)
It's doable if you're looking at the subset
Lex Fridman (49:36.000)
of the visual space.
Lex Fridman (49:37.120)
But even then, it's still fairly expensive.
Lex Fridman (49:38.800)
You still need millions of examples.
Lex Fridman (49:40.920)
And it's only going to be able to make sense of things
Lex Fridman (49:44.240)
that are very close to what it has seen before.
Lex Fridman (49:46.880)
And in contrast to that, well, of course,
Lex Fridman (49:49.160)
you have human intelligence.
Lex Fridman (49:50.160)
But even if you're not looking at human intelligence,
Lex Fridman (49:53.240)
you can look at very simple rules, algorithms.
François Chollet (49:56.800)
If you have a symbolic rule,
Lex Fridman (49:58.080)
it can actually apply to a very, very large set of inputs
François Chollet (50:03.120)
because it is abstract.
Lex Fridman (50:04.880)
It is not obtained by doing a point by point mapping.
François Chollet (50:10.720)
For instance, if you try to learn a sorting algorithm
Lex Fridman (50:14.000)
using a deep neural network,
François Chollet (50:15.520)
well, you're very much limited to learning point by point
Lex Fridman (50:20.080)
what the sorted representation of this specific list is like.
Lex Fridman (50:24.360)
But instead, you could have a very, very simple
Lex Fridman (50:29.400)
sorting algorithm written in a few lines.
François Chollet (50:31.920)
Maybe it's just two nested loops.
Lex Fridman (50:35.560)
And it can process any list at all because it is abstract,
François Chollet (50:41.040)
because it is a set of rules.
Lex Fridman (50:42.240)
So deep learning is really like point by point
François Chollet (50:45.160)
geometric morphings, train with good and decent.
Lex Fridman (50:48.640)
And meanwhile, abstract rules can generalize much better.
Lex Fridman (50:53.640)
And I think the future is we need to combine the two.
Lex Fridman (50:56.160)
So how do we, do you think, combine the two?
Lex Fridman (50:59.160)
How do we combine good point by point functions
Lex Fridman (51:03.040)
with programs, which is what the symbolic AI type systems?
Lex Fridman (51:08.920)
At which levels the combination happen?
Lex Fridman (51:11.600)
I mean, obviously we're jumping into the realm
François Chollet (51:14.680)
of where there's no good answers.
Lex Fridman (51:16.880)
It's just kind of ideas and intuitions and so on.
François Chollet (51:20.280)
Well, if you look at the really successful AI systems
Lex Fridman (51:23.080)
today, I think they are already hybrid systems
François Chollet (51:26.320)
that are combining symbolic AI with deep learning.
Lex Fridman (51:29.520)
For instance, successful robotics systems
François Chollet (51:32.520)
are already mostly model based, rule based,
Lex Fridman (51:37.400)
things like planning algorithms and so on.
François Chollet (51:39.400)
At the same time, they're using deep learning
Lex Fridman (51:42.200)
as perception modules.
François Chollet (51:43.840)
Sometimes they're using deep learning as a way
Lex Fridman (51:46.000)
to inject fuzzy intuition into a rule based process.
François Chollet (51:50.920)
If you look at the system like in a self driving car,
Lex Fridman (51:54.560)
it's not just one big end to end neural network.
François Chollet (51:57.240)
You know, that wouldn't work at all.
Lex Fridman (51:59.000)
Precisely because in order to train that,
François Chollet (52:00.760)
you would need a dense sampling of experience base
Lex Fridman (52:05.160)
when it comes to driving,
François Chollet (52:06.200)
which is completely unrealistic, obviously.
Lex Fridman (52:08.880)
Instead, the self driving car is mostly
François Chollet (52:13.920)
symbolic, you know, it's software, it's programmed by hand.
Lex Fridman (52:18.360)
So it's mostly based on explicit models.
François Chollet (52:21.640)
In this case, mostly 3D models of the environment
Lex Fridman (52:25.840)
around the car, but it's interfacing with the real world
Lex Fridman (52:29.520)
using deep learning modules, right?
Lex Fridman (52:31.440)
So the deep learning there serves as a way
François Chollet (52:33.440)
to convert the raw sensory information
Lex Fridman (52:36.080)
to something usable by symbolic systems.
François Chollet (52:39.760)
Okay, well, let's linger on that a little more.
Lex Fridman (52:42.400)
So dense sampling from input to output.
François Chollet (52:45.440)
You said it's obviously very difficult.
Lex Fridman (52:48.240)
Is it possible?
Lex Fridman (52:50.120)
In the case of self driving, you mean?
Lex Fridman (52:51.800)
Let's say self driving, right?
François Chollet (52:53.040)
Self driving for many people,
Lex Fridman (52:57.560)
let's not even talk about self driving,
François Chollet (52:59.520)
let's talk about steering, so staying inside the lane.
Lex Fridman (53:05.040)
Lane following, yeah, it's definitely a problem
François Chollet (53:07.080)
you can solve with an end to end deep learning model,
Lex Fridman (53:08.880)
but that's like one small subset.
François Chollet (53:10.600)
Hold on a second.
Lex Fridman (53:11.440)
Yeah, I don't know why you're jumping
François Chollet (53:12.760)
from the extreme so easily,
Lex Fridman (53:14.480)
because I disagree with you on that.
François Chollet (53:16.280)
I think, well, it's not obvious to me
Lex Fridman (53:21.000)
that you can solve lane following.
François Chollet (53:23.400)
No, it's not obvious, I think it's doable.
Lex Fridman (53:25.840)
I think in general, there is no hard limitations
François Chollet (53:31.200)
to what you can learn with a deep neural network,
Lex Fridman (53:33.680)
as long as the search space is rich enough,
François Chollet (53:40.320)
is flexible enough, and as long as you have
Lex Fridman (53:42.240)
this dense sampling of the input cross output space.
François Chollet (53:45.360)
The problem is that this dense sampling
Lex Fridman (53:47.720)
could mean anything from 10,000 examples
François Chollet (53:51.120)
to like trillions and trillions.
Lex Fridman (53:52.840)
So that's my question.
Lex Fridman (53:54.360)
So what's your intuition?
Lex Fridman (53:56.200)
And if you could just give it a chance
Lex Fridman (53:58.720)
and think what kind of problems can be solved
Lex Fridman (54:01.880)
by getting a huge amounts of data
Lex Fridman (54:04.240)
and thereby creating a dense mapping.
Lex Fridman (54:08.000)
So let's think about natural language dialogue,
François Chollet (54:12.480)
the Turing test.
Lex Fridman (54:14.000)
Do you think the Turing test can be solved
Lex Fridman (54:17.000)
with a neural network alone?
Lex Fridman (54:21.120)
Well, the Turing test is all about tricking people
François Chollet (54:24.440)
into believing they're talking to a human.
Lex Fridman (54:26.880)
And I don't think that's actually very difficult
François Chollet (54:29.040)
because it's more about exploiting human perception
Lex Fridman (54:35.600)
and not so much about intelligence.
François Chollet (54:37.520)
There's a big difference between mimicking
Lex Fridman (54:39.680)
intelligent behavior and actual intelligent behavior.
François Chollet (54:42.080)
So, okay, let's look at maybe the Alexa prize and so on.
Lex Fridman (54:45.360)
The different formulations of the natural language
François Chollet (54:47.480)
conversation that are less about mimicking
Lex Fridman (54:50.520)
and more about maintaining a fun conversation
François Chollet (54:52.800)
that lasts for 20 minutes.
Lex Fridman (54:54.720)
That's a little less about mimicking
Lex Fridman (54:56.200)
and that's more about, I mean, it's still mimicking,
Lex Fridman (54:59.080)
but it's more about being able to carry forward
François Chollet (55:01.440)
a conversation with all the tangents that happen
Lex Fridman (55:03.640)
in dialogue and so on.
Lex Fridman (55:05.080)
Do you think that problem is learnable
Lex Fridman (55:08.320)
with a neural network that does the point to point mapping?
Lex Fridman (55:14.520)
So I think it would be very, very challenging
Lex Fridman (55:16.280)
to do this with deep learning.
François Chollet (55:17.800)
I don't think it's out of the question either.
Lex Fridman (55:21.480)
I wouldn't rule it out.
François Chollet (55:23.240)
The space of problems that can be solved
Lex Fridman (55:25.400)
with a large neural network.
Lex Fridman (55:26.920)
What's your sense about the space of those problems?
Lex Fridman (55:30.080)
So useful problems for us.
Lex Fridman (55:32.560)
In theory, it's infinite, right?
Lex Fridman (55:34.800)
You can solve any problem.
François Chollet (55:36.200)
In practice, well, deep learning is a great fit
Lex Fridman (55:39.800)
for perception problems.
François Chollet (55:41.800)
In general, any problem which is naturally amenable
Lex Fridman (55:47.640)
to explicit handcrafted rules or rules that you can generate
François Chollet (55:52.200)
by exhaustive search over some program space.
Lex Fridman (55:56.080)
So perception, artificial intuition,
François Chollet (55:59.320)
as long as you have a sufficient training dataset.
Lex Fridman (56:03.240)
And that's the question, I mean, perception,
François Chollet (56:05.360)
there's interpretation and understanding of the scene,
Lex Fridman (56:08.400)
which seems to be outside the reach
François Chollet (56:10.280)
of current perception systems.
Lex Fridman (56:12.960)
So do you think larger networks will be able
François Chollet (56:15.920)
to start to understand the physics
Lex Fridman (56:18.280)
and the physics of the scene,
François Chollet (56:21.080)
the three dimensional structure and relationships
Lex Fridman (56:23.400)
of objects in the scene and so on?
Lex Fridman (56:25.560)
Or really that's where symbolic AI has to step in?
Lex Fridman (56:28.320)
Well, it's always possible to solve these problems
François Chollet (56:34.480)
with deep learning.
Lex Fridman (56:36.800)
It's just extremely inefficient.
François Chollet (56:38.560)
A model would be an explicit rule based abstract model
Lex Fridman (56:42.000)
would be a far better, more compressed
François Chollet (56:45.240)
representation of physics.
Lex Fridman (56:46.840)
Then learning just this mapping between
François Chollet (56:49.080)
in this situation, this thing happens.
Lex Fridman (56:50.960)
If you change the situation slightly,
François Chollet (56:52.720)
then this other thing happens and so on.
Lex Fridman (56:54.760)
Do you think it's possible to automatically generate
Lex Fridman (56:57.440)
the programs that would require that kind of reasoning?
Lex Fridman (57:02.200)
Or does it have to, so the way the expert systems fail,
François Chollet (57:05.360)
there's so many facts about the world
Lex Fridman (57:07.120)
had to be hand coded in.
Lex Fridman (57:08.960)
Do you think it's possible to learn those logical statements
Lex Fridman (57:14.600)
that are true about the world and their relationships?
Lex Fridman (57:18.200)
Do you think, I mean, that's kind of what theorem proving
Lex Fridman (57:20.360)
at a basic level is trying to do, right?
François Chollet (57:22.680)
Yeah, except it's much harder to formulate statements
Lex Fridman (57:26.160)
about the world compared to formulating
François Chollet (57:28.480)
mathematical statements.
Lex Fridman (57:30.320)
Statements about the world tend to be subjective.
Lex Fridman (57:34.200)
So can you learn rule based models?
Lex Fridman (57:39.600)
Yes, definitely.
François Chollet (57:40.920)
That's the field of program synthesis.
Lex Fridman (57:43.640)
However, today we just don't really know how to do it.
Lex Fridman (57:48.040)
So it's very much a grass search or tree search problem.
Lex Fridman (57:52.400)
And so we are limited to the sort of tree session grass
François Chollet (57:56.800)
search algorithms that we have today.
Lex Fridman (57:58.560)
Personally, I think genetic algorithms are very promising.
Lex Fridman (58:02.760)
So almost like genetic programming.
Lex Fridman (58:04.360)
Genetic programming, exactly.
Lex Fridman (58:05.560)
Can you discuss the field of program synthesis?
Lex Fridman (58:08.840)
Like how many people are working and thinking about it?
François Chollet (58:14.560)
Where we are in the history of program synthesis
Lex Fridman (58:17.960)
and what are your hopes for it?
François Chollet (58:20.720)
Well, if it were deep learning, this is like the 90s.
Lex Fridman (58:24.600)
So meaning that we already have existing solutions.
François Chollet (58:29.120)
We are starting to have some basic understanding
Lex Fridman (58:34.280)
of what this is about.
Lex Fridman (58:35.480)
But it's still a field that is in its infancy.
Lex Fridman (58:38.000)
There are very few people working on it.
François Chollet (58:40.440)
There are very few real world applications.
Lex Fridman (58:44.480)
So the one real world application I'm aware of
François Chollet (58:47.640)
is Flash Fill in Excel.
Lex Fridman (58:51.680)
It's a way to automatically learn very simple programs
François Chollet (58:55.080)
to format cells in an Excel spreadsheet
Lex Fridman (58:58.200)
from a few examples.
François Chollet (59:00.240)
For instance, learning a way to format a date, things like that.
Lex Fridman (59:02.800)
Oh, that's fascinating.
François Chollet (59:03.680)
Yeah.
Lex Fridman (59:04.560)
You know, OK, that's a fascinating topic.
François Chollet (59:06.280)
I always wonder when I provide a few samples to Excel,
Lex Fridman (59:10.480)
what it's able to figure out.
François Chollet (59:12.600)
Like just giving it a few dates, what
Lex Fridman (59:15.960)
are you able to figure out from the pattern I just gave you?
François Chollet (59:18.480)
That's a fascinating question.
Lex Fridman (59:19.760)
And it's fascinating whether that's learnable patterns.
Lex Fridman (59:23.320)
And you're saying they're working on that.
Lex Fridman (59:25.520)
How big is the toolbox currently?
Lex Fridman (59:28.200)
Are we completely in the dark?
Lex Fridman (59:29.520)
So if you said the 90s.
Lex Fridman (59:30.440)
In terms of program synthesis?
Lex Fridman (59:31.720)
No.
Lex Fridman (59:32.360)
So I would say, so maybe 90s is even too optimistic.
Lex Fridman (59:37.720)
Because by the 90s, we already understood back prop.
François Chollet (59:41.080)
We already understood the engine of deep learning,
Lex Fridman (59:43.960)
even though we couldn't really see its potential quite.
François Chollet (59:47.280)
Today, I don't think we have found
Lex Fridman (59:48.520)
the engine of program synthesis.
Lex Fridman (59:50.400)
So we're in the winter before back prop.
Lex Fridman (59:52.880)
Yeah.
François Chollet (59:54.160)
In a way, yes.
Lex Fridman (59:55.720)
So I do believe program synthesis and general discrete search
🔗 相关节目