The following is a conversation with Francois Chollet, his second time in the podcast.
He's both a world-class engineer and a philosopher in the realm of deep learning and artificial
intelligence.
This time, we talk a lot about his paper titled, On the Measure of Intelligence, that discusses
how we may define and measure general intelligence in our computing machinery.
Quick summary of the sponsors, Babel, Masterclass, and Cash App.
Take the sponsor links in the description to get a discount and to support this podcast.
As a side note, let me say that the serious, rigorous, scientific study of artificial general
intelligence is a rare thing.
The mainstream machine learning community works on very narrow AI with very narrow benchmarks.
This is very good for incremental and sometimes big incremental progress.
On the other hand, the outside the mainstream, renegade, you could say, AGI community, works
on approaches that verge on the philosophical, and even the literary, without big public benchmarks.
Walking the line between the two worlds is a rare breed, but it doesn't have to be.
I ran the AGI series at MIT as an attempt to inspire more people to walk this line.
Keep mind and open AI for time and still, on occasion, walk this line.
Francois Chollet does as well.
I hope to also.
It's a beautiful dream to work towards and to make real one day.
If you enjoy this thing, subscribe on YouTube, review it with 5 stars on Apple Podcast, follow
on Spotify, support on Patreon, or connect with me on Twitter at Lex Freedman.
As usual, I'll do a few minutes of ads now and no ads in the middle.
I try to make these interesting, but I give you time stamps so you can skip.
But still, please do check out the sponsors by clicking the links in the description.
It's the best way to support this podcast.
This show is sponsored by Babel, an app and website that gets you speaking in a new language
within weeks.
Go to babel.com and use co-lex to get 3 months free.
They offer 14 languages, including Spanish, French, Italian, German, and yes, Russian.
Daily lessons are 10 to 15 minutes, super easy, effective, designed by over 100 language
experts.
Let me read a few lines from the Russian poem Noch Ulytso Fanaar Abteka by Alexander
Blok that you'll start to understand if you sign up to Babel.
Noch Ulytso Fanaar Abteka, Vesmyzynyi Itusklii Svet, Zhevii Shchoh Chetvirt Vekka, Vsebudit
Tak, Ishoda, Ned.
Now I say that you'll start to understand this poem because Russian starts with a language
and ends with a vodka.
Now the latter part is definitely not endorsed or provided by Babel and will probably lose
me this sponsorship, although it hasn't yet.
But once you graduate with Babel, you can enroll in my advanced course of late night
Russian conversation over vodka.
No app for that yet.
So get started by visiting babel.com and use code LEX to get 3 months free.
This show is also sponsored by Masterclass.
Sign up at masterclass.com slash lex to get a discount and to support this podcast.
When I first heard about Masterclass, I thought it was too good to be true.
I still think it's too good to be true.
For $180 a year, you get an all-access pass to watch courses from to list some of my favorites,
Chris Hadfield on Space Exploration, hope to have him in this podcast one day, Neil
Lugrass Tyson on Scientific Thinking and Communication, Neil too, Will Wright, Creator of SimCity
and Sims on Game Design, Carl Santana on Guitar, Karek Asparov on Chess, Daniel Lagrano on
Poker and many more.
Chris Hadfield explaining how rockets work and the experience of being watched at the
space alone is worth the money.
By the way, you can watch it on basically any device.
Once again, sign up at masterclass.com slash lex to get a discount and to support this podcast.
This show, finally, is presented by Cash App, the number one finance app in the App Store.
When you get it, use code LEX Podcast, Cash App lets you send money to friends, buy bitcoin
and invest in the stock market with as little as $1.
Since Cash App allows you to send and receive money digitally, let me mention a surprising
fact related to physical money.
Of all the currency in the world, roughly 8% of it is actually physical money.
The other 92% of the money only exists digitally and that's only going to increase.
So again, if you get Cash App from the App Store, Google Play and use code LEX Podcast,
you get $10 and Cash App will also donate $10 the first, an organization that is helping
to advance robotics and STEM education for young people around the world.
And now here's my conversation with Francois Chalet.
What philosophers, thinkers or ideas had a big impact on you growing up and today?
So one author that had a big impact on me when I read his books as a teenager was Jean
Piaget, who is a Swiss psychologist, is considered to be the father of developmental psychology
and he has a large body of work about basically how intelligence develops in children.
And so it's very old work, like most of it is from the 1930s, 1940s, so it's not quite
up to date.
It's actually superseded by many newer developments in developmental psychology.
But to me, it was very, very interesting, very striking and actually shaped the early
ways in which I started thinking about the mind and the development of intelligence as
a teenager.
His actual ideas or the way he thought about it or just the fact that you could think about
the developing mind at all?
I guess both.
Jean Piaget is the author that's introduced me to the notion that intelligence and the
mind is something that you construct throughout your life and that the children construct
it in stages.
And I thought that was a very interesting idea, which is, of course, very relevant to AI,
to building artificial minds.
Another book that I read around the same time that had a big impact on me, and there was
actually a little bit of overlap with Jean Piaget as well, and I read it around the same
time, is Jeff Hawkins' On Intelligence, which is a classic, and he has this vision of the
mind as a multiscale hierarchy of temporal prediction modules.
And these ideas really resonated with me, like the notion of a modular hierarchy of
potentially compression functions or prediction functions.
I thought it was really, really interesting, and it shaped the way I started thinking about
how to build minds.
The hierarchical nature, which aspect.
Also he's a neuroscientist, so he was thinking of actual, basically talking about how our
mind works.
Yeah, the notion that cognition is prediction was an idea that was kind of new to me at
the time, and that I really loved at the time.
And yeah, the notion that there are multiple scales of processing in the brain.
The hierarchy.
Yes.
This is before deep learning.
These ideas of hierarchies in here, I've been around for a long time, even before On
Intelligence, I mean, they've been around since the 1980s.
And yeah, that was before deep learning, but of course, I think these ideas really found
their practical implementation in deep learning.
What about the memory side of things?
I think he was talking about knowledge representation.
Do you think about memory a lot?
One way you can think of neural networks as a kind of memory, you're memorizing things,
but it doesn't seem to be the kind of memory that's in our brains, or it doesn't have the
same rich complexity, long-term nature that's in our brains.
Yes.
The brain is more for sparse access memory, so that you can actually retrieve very precisely
like bits of your experience.
The retrieval aspect, you can like introspect.
You can ask yourself questions, I guess.
You can program your own memory.
And language is actually the tool you used to do that.
I think language is a kind of operating system for the mind.
And you use language, well, one of the uses of language is as a query that you run over
your own memory, use words as keys to retrieve specific experiences or specific concepts,
specific thoughts.
Like language is a way you store thoughts, not just in writing, in the physical world,
but also in your own mind.
And it's also how you retrieve them.
Imagine if you didn't have language, then you would have to, you would not really have
a self internally triggered way of retrieving past thoughts.
You would have to rely on external experiences.
For instance, you see a specific site, you smell a specific smell, and it brings up memories,
but you would not really have a way to deliberately access these memories without language.
Well, the interesting thing you mentioned is you can also program your memory.
You can change it probably with language.
Using language, yes.
Well, let me ask you a Chomsky question, which is like, first of all, do you think language
is like fundamental, like, there's turtles, what's at the bottom of the turtles?
They don't go, it can't be turtles all the way down, is language at the bottom of cognition
of everything as like language, the fundamental aspect of like what it means to be a thinking
thing?
No, I don't think so.
I think language is...
You disagree with Noam Chomsky?
Yes.
Language is a layer on top of cognition.
So it is fundamental to cognition in the sense that to use a computing metaphor, I see
language as the operating system of the brain, of the human mind.
And the operating system is a layer on top of the computer.
The computer exists before the operating system, but the operating system is how you
make it truly useful.
And the operating system is most likely Windows, not Linux, because language is messy.
Yeah, it's messy and it's pretty difficult to inspect it, introspect it.
How do you think about language?
We use actually sort of human interpretable language, but is there something deeper that's
closer to like logical type of statements?
What is the nature of language, do you think?
Is there something deeper than like the syntactic rules we construct?
Is there something that doesn't require utterances or writing or so on?
Oh, you're asking about the possibility that there could exist languages for thinking that
are not made of words?
Yeah.
Yeah, I think so.
So the mind is layers, right?
And language is almost like the uttermost, the uppermost layer.
But before we think in words, I think we think in terms of emotion in space and we think
in terms of physical actions.
And I think babies in particular probably express these thoughts in terms of the actions
that they've seen or that they can perform and in terms of the, in terms of motions of
objects in their environment before they start thinking in terms of words.
It's amazing to think about that as the building blocks of language, so like the kind of actions
and ways the babies see the world as like more fundamental than the beautiful Shakespearean
language you construct on top of it.
And we probably don't have any idea what that looks like, right?
Like what?
Because it's important for them trying to engineer it into AI systems.
I think visual analogies and motion is a fundamental building block of the mind.
And you actually see it reflected in language, like language is full of special metaphors.
And when you think about things, I consider myself very much as a visual thinker.
You often express these thoughts by using things like visualizing concepts into the space or
like you solve problems by imagining yourself, navigating a concept space.
I don't know if you have this sort of experience.
You said visualizing concept space, so like, so I certainly think about, I certainly visualize
mathematical concepts, but you mean like in concept space, visually you're embedding
ideas into three-dimensional space, you can explore with your mind essentially.
You should be more like 2D, but yeah.
2D?
You're a flatlander.
You're, okay, no, I do not.
I always have to, before I jump from concept to concept, I have to put it back down on
paper.
It has to be on paper.
I can only travel on 2D paper, not inside my mind.
You're able to move inside your mind.
And even if you're writing like a paper, for instance, don't you have like a spatial representation
of your paper?
Like you visualize where ideas lie topologically in relationship to other ideas, kind of like
a subway map of the ideas in your paper.
Yeah, that's true.
I mean, there is, in papers, I don't know about you, but it feels like there's a destination.
There's a key idea that you want to write that, and a lot of it is in the fog, and you're
trying to kind of, it's almost like, what's that called when you do a path planning search
from both directions, from the start and from the end, and then you find, you do like shortest
path, but like, you know, in game playing, you do this with like a star from both sides.
And you see where we're on the join.
Yeah.
So you kind of do, at least for me, I think like, first of all, just exploring from the
start, from like, first principles, what do I know?
What can I start proving from that, right?
And then from the destination, if I use their backtracking, like, if I want to show some
kind of sets of ideas, what would it take to show them and kind of backtrack, but like,
yeah, I don't think I'm doing all that in my mind, though, like I'm putting it down
on paper.
Do you use mind maps to organize your ideas?
No.
Yeah, I like mind maps.
I'm that kind of person.
Let's get into this.
Because it's...
I've been so jealous of people.
I haven't really tried it.
I've been jealous of people that seem to like, they get like this fire of passion in their
eyes because everything starts making sense.
It's like Tom Cruise in the movie was like moving stuff around some of the most brilliant
people I know use mind maps.
I haven't tried really.
Can you explain what the hell a mind map is?
I guess a mind map is a way to make kind of like the mess inside your mind to just put
it on paper so that you gain more control over it.
That's the way to organize things on paper and as kind of like a consequence of organizing
things on paper, it's starting to be more organized inside your mind.
What does that look like?
Do you have an example?
What's the first thing you write on paper?
What's the second thing you write?
I mean, typically, you draw a mind map to organize the way you think about the topic.
You would start by writing down the key concept about that topic, like you would write intelligence
or something, and then you would start adding associative connections.
What do you think about when you think about intelligence?
What do you think are the key elements of intelligence?
Maybe you would have language, for instance, and you would have motion, and so you would
start drawing notes with these things, and then you would see what do you think about
when you think about motion and so on, and you would go like that like a tree.
Is it a tree or a tree most or is it a graph too, like a tree?
It's more of a graph than a tree, and it's not limited to just writing down words.
You can also draw things, and it's not supposed to be purely hierarchical, right?
The point is that you can start, once you start writing it down, you can start reorganizing
it so that it makes more sense, so that it's connected in a more effective way.
But I'm so OCD that you just mentioned intelligence and language and motion, I would start becoming
paranoid that the categorization isn't perfect, like that I would become paralyzed with the
mind map, that like this may not be, so like even though you're just doing associative
kind of connections, there's an implied hierarchy that's emerging, and I would start becoming
paranoid that it's not the proper hierarchy, so you're not just, one way to see mind maps
is you're putting thoughts on paper, it's like a stream of consciousness, but then you
can also start getting paranoid, well, if it's just the right hierarchy.
Sure, but it's a mind map, it's your mind map, you're free to draw anything you want,
you're free to draw any connection you want, and you can just make a different mind map
if you think the central node is not the right node.
I suppose there's a fear of being wrong.
If you want to organize your ideas by writing down what you think, which I think is very
effective, like how do you know what you think about something if you don't write it down?
If you do that, the thing is that it imposes much more syntactic structure over your ideas,
which is not required with a mind map, so a mind map is kind of like a lower level,
more freehand way of organizing your thoughts.
Once you've drawn it, then you can start actually voicing your thoughts in terms of
paragraphs.
It's a two-dimensional aspect of layout too, right?
Yeah.
It's a kind of flower, I guess, you start, there's usually, you want to start with a
central concept?
Yes.
And then you move out.
Typically, it ends up more like a subway map, so it ends up more like a graph, a topological
graph.
Without a root note.
Yeah, so like in a subway map, there are some notes that are more connected than others
and there are some notes that are more important than others, right?
So there are destinations, but it's not going to be purely like a tree, for instance.
Yeah, it's fascinating to think if there's something to that about the way our mind
thinks.
By the way, I just kind of remembered obvious thing that I have probably thousands of documents
in Google Doc at this point that are bullet point lists, which is you can probably map
a mind map to a bullet point list.
It's the same.
No, it's not.
It's a tree.
It's a tree.
Yeah.
So I create trees, but also they don't have the visual element, like I guess I'm comfortable
with the structure.
It feels like the narrowness, the constraints feel more comforting.
If you have thousands of documents with your own thoughts in Google Docs, why don't you
write some kind of search engine, like maybe a mind map, a piece of software, a mind mapping
software where you write down a concept and then it gives you sentences or paragraphs
from your thousand Google Docs document that match this concept.
The problem is it's so deeply unlike mind maps, it's so deeply rooted in natural language.
So it's not, it's not semantically searchable, I would say, because the categories are very,
you kind of mentioned intelligence, language and motion.
They're very strong, semantic, like it feels like the mind map forces you to be semantically
clear and specific.
The bullet points list I have are sparse, disparate thoughts that poetically represent
a category like motion as opposed to saying motion.
So unfortunately, that's the same problem with the internet.
That's why the idea of semantic web is difficult to get.
Just language on the internet is a giant mess of natural language that's hard to interpret.
Which, so do you think there's something to mind maps as, you actually originally brought
it up as what we're talking about, kind of cognition and language.
Do you think there's something to mind maps about how our brain actually deals, like,
think reasons about things?
It's possible, I think it's reasonable to assume that there is some level of topological
processing in the brain, that the brain is very associative in nature.
And I also believe that a topological space is a better medium to uncode thoughts than
a geometric space.
So I think...
What's the difference in a topological and a geometric space?
If you're talking about topologies, then points are either connected or not.
So a topology is more like a subway map.
And geometry is when you're interested in the distance between things.
And in subway maps, you don't really have the concept of distance, you only have the
concept of whether there is a train going from station A to station B.
And what we do in deep learning is that we're actually dealing with geometric spaces.
We are dealing with concept vectors, word vectors, that have a distance between them.
The distance is a fun one.
In terms of dot product.
We are not really building topological models, usually.
I think you're absolutely right.
Distance is a fundamental importance in deep learning.
I mean, it's the continuous aspect of it.
Yes.
Because everything is a vector.
And everything has to be a vector, because everything has to be differentiable.
If your space is discrete, it's no longer differentiable, you cannot do deep learning
in it anymore.
Well, you could, but you could only do it by embedding it in a bigger, continuous space.
So if you do topology in the context of deep learning, you have to do it by embedding
your topology in a geometry.
Yeah.
Well, let me zoom out for a second.
Let's get into your paper on the measure of intelligence that, did you put it on 2019?
Yes.
Okay.
November.
November.
Yeah.
Remember 2019?
That was a different time.
Yeah.
I remember.
I still remember.
It feels like a different, different, different world.
You could travel or you could, you know, actually go outside and see friends.
Yeah.
Let me ask the most absurd question.
I think there's some non-zero probability there'll be a textbook one day, like two
one years from now on artificial intelligence or it'll be called like just intelligence
because humans will already be gone and it'll be your picture with a quote, you know, one
of the early biological systems would consider the nature of intelligence and there'll be
like a definition of how they thought about intelligence, which is one of the things you
do in your paper on measure intelligence is to ask like, well, what is intelligence?
And how to test for intelligence and so on.
So is there a spiffy quote about what is intelligence?
What is the definition of intelligence according to François Chollet?
Yeah.
So do you think the super intelligent AIs of the future will want to remember us the
way we remember humans from the past?
And do you think they will be, you know, they won't be ashamed of having a biological
origin?
No, I think it would be a niche topic.
It won't be that interesting, but it'll be, it'll be like the people that study in certain
contexts like historical civilization that no longer exists, the Aztecs and so on.
That's how it'll be seen.
And it'll be study in the also the context on social media, there'll be hashtags about
the atrocity committed to human beings when, when the, when the robots finally got rid
of them.
Like it was a mistake.
You'll be seen as a, as a giant mistake, but ultimately in the name of progress and it
created a better world because humans were over consuming the resources and the, we're
not very rational and we're destructive in the end in terms of productivity and putting
more love in the world.
And so within that context, there'll be a chapter about these biological systems.
It seems to have a very detailed vision of that feature.
You should write a sci-fi novel about it.
I'm working, I'm working on a sci-fi novel currently, yes, self-published, yeah.
The definition of intelligence.
So intelligence is the efficiency with which you acquire new skills at tasks that you did
not previously know about that you did not prepare for, right?
So it is not intelligence is not skill itself.
It's not what you know, it's not what you can do.
It's how well and how efficiently you can learn new things.
New things.
Yes.
The idea of newness there seems to be fundamentally important.
Yes.
So you would see intelligence on display, for instance, whenever you see a human being
or an AI creature, adapt to a new environment that it does not see before, that its creators
did not anticipate.
When you see adaptation, when you see improvisation, when you see generalization, that's intelligence.
In reverse, if you have a system that when you put it in a slightly new environment,
it cannot adapt, it cannot improvise, it cannot deviate from what it's hard-coded to
do or what it has been trained to do, that is a system that is not intelligence.
There's actually a quote from Einstein that captures this idea, which is, the measure
of intelligence is the ability to change.
I like that quote.
I think it captures at least part of this idea.
You know, there might be something interesting about the difference between your definition
and Einstein's, I mean, he's just being Einstein and clever.
But acquisition of new ability to deal with new things versus ability to just change,
what's the difference between those two things?
So just changing itself.
Do you think there's something to that?
Just being able to change.
Yes, being able to adapt, so not change, but certainly change its direction, being able
to adapt yourself to your environment.
Whatever the environment is.
That's a big part of intelligence, yes.
And intelligence is most precisely how efficiently you're able to adapt, how efficiently you're
able to basically master your environment, how efficiently you can acquire new skills.
And I think there's a big distinction to be drawn between intelligence, which is a process,
and the output of that process, which is skill.
So for instance, if you have a very smart human programmer that considers the game of
chess and that writes down a static program that can play chess, then the intelligence
is the process of developing that program.
But the program itself is just encoding the output artifact of that process.
The program itself is not intelligent.
And the way you tell it's not intelligent is that if you put it in a different context,
you ask it to play go or something, it's not going to be able to perform well with that
human involvement.
Because the source of intelligence, the entity that is capable of that process is the human
programmer.
So we should be able to tell a difference between the process and its output.
We should not confuse the output and the process.
It's the same as do not confuse a road building company and one specific road.
Because one specific road takes you from point A to point B. But a road building company
can take you from, can make a path from anywhere to anywhere else.
Yeah, that's beautifully put.
But it's also to play devil's advocate a little bit.
It's possible that there is something more fundamental than us humans.
So you kind of said the programmer creates the difference between the choir of the skill
and the skill itself.
There could be something like you could argue the universe is more intelligent.
Like the deep, the base intelligence that we should be trying to measure is something
that created humans should be measuring God or what the source of the universe as opposed
to like there could be a deeper intelligence, there's always deeper intelligence.
You can argue that, but that does not take anything away from the fact that humans are
intelligence.
And you can tell that because they are capable of adaptation and generality.
And you see that in particular in the fact that humans are capable of handling situations
and tasks that are quite different from anything that any of our evolutionary ancestors has
ever encountered.
So we are capable of generalizing very much out of distribution if you consider our evolutionary
history as being in a way of trying data.
Of course evolutionary biologists would argue that we're not going too far out of the distribution.
We're mapping the skills we've learned previously and desperately trying to jam them into these
new situations.
I mean, there's definitely a little bit of that, but it's pretty clear to me that we're
able to, you know, most of the things we do any given day in our modern civilization are
things that are very, very different from what our ancestors a million years ago would
have been doing in a given day.
And our environment is very different.
So I agree that everything we do, we do it with cognitive building blocks that we acquire
over the course of evolution, right?
And that anchors our cognition to certain contexts, which is the human condition very
much.
But still, our mind is capable of a pretty remarkable degree of generality far beyond
anything we can create in artificial systems today, like the degree in which the mind can
generalize from its evolutionary history, can generalize away from its evolutionary history
is much greater than the degree to which a deep learning system today can generalize
away from its trained data.
And like the key point you're making, which I think is quite beautiful is like we shouldn't
measure, if we talk about measurement, we shouldn't measure the skill.
We should measure like the creation of the new skill, the ability to create that new
skill.
But it's tempting, like, it's weird because the skill is a little bit of a small window
into the system.
So whenever you have a lot of skills, it's tempting to measure the skills.
Yes.
I mean, the skill is the only thing you can objectively measure, but yeah.
So the thing to keep in mind is that when you see skill in a human, it gives you a strong
signal that that human is intelligent because you know they weren't born with that skill
typically.
Like, you see a very strong chess player, maybe you're a very strong chess player yourself.
I think you're saying that because I'm Russian and now you're prejudiced, you assume all
that, I'm biased.
So if you see a very strong chess player, you know they weren't born knowing how to
play chess.
So they had to acquire that skill with their limited resources, with their limited lifetime.
And you know, they did that because they are generally intelligent.
And so they may as well have acquired any other skill, you know, they have this potential.
And on the other hand, if you see a computer playing chess, you cannot make the same assumptions
because you cannot just assume the computer is generally intelligent.
The computer may be born knowing how to play chess in the sense that it may have been programmed
by a human that has understood chess for the computer and that has just encoded the output
of that understanding in a static program.
And that program is not intelligent.
So let's zoom out just for a second and say like, what is the goal of the on the measure
of intelligence paper, like what do you hope to achieve with it?
So the goal of the paper is to clear up some longstanding misunderstandings about the way
we've been conceptualizing intelligence in the AI community and in the way we've been
evaluating progress in AI.
There's been a lot of progress recently in machine learning and people are, you know,
extrapolating from that progress that we are about to solve general intelligence.
And if you want to be able to evaluate these statements, you need to precisely define what
you're talking about when you're talking about general intelligence.
And you need a formal way, a reliable way to measure how much intelligence, how much
general intelligence a system processes.
And ideally this measure of intelligence should be actionable.
So it should not just describe what intelligence is, it should not just be a binary indicator
that tells you the system is intelligent or it isn't.
It should be actionable, it should have explanatory power, right?
So you could use it as a feedback signal.
It would show you the way towards building more intelligent systems.
So at the first level, you draw a distinction between two divergent views of intelligence.
As we just talked about, intelligence is a collection of task-specific skills and a
general learning ability.
So what's the difference between kind of this memorization of skills and a general learning
ability?
We talked about it a little bit, but can you try to link around this topic for a bit?
Yes, so the first part of the paper is an assessment of the different ways we've been
thinking about intelligence and the different ways we've been evaluating progress in AI.
And this tree of cognitive sciences has been shaped by two views of the human mind.
And one view is the evolutionary psychology view in which the mind is a collection of
fairly static, special purpose ad hoc mechanisms that have been hard-coded by evolution over
our history as a species for a very long time.
And early AI researchers, people like Marvin Minsky, for instance, they clearly subscribed
to this view.
And they saw the mind as a kind of collection of static programs similar to the programs
they would run on like mainframe computers.
And in fact, I think they very much understood the mind through the metaphor of the mainframe
computer because that was the tool they were working with.
And so you had this static program, this collection of very different static programs operating
over a database like memory.
And in this picture, learning was not very important.
Learning was considered to be just memorization.
And in fact, learning is basically not featured in AI textbooks until the 1980s with the rise
of machine learning.
It's kind of fun to think about that learning was the outcast, like the weird people working
on learning.
The mainstream AI world was, I mean, I don't know what the best term is, but it's non-learning.
It was seen as reasoning would not be learning-based.
Yes, it was considered that the mind was a collection of programs that were primarily
logical in nature.
And that's all you needed to do to create a mind was to write down these programs.
And they would operate over knowledge, which would be stored in some kind of database.
And as long as your database would encompass everything about the world and your logical
rules were comprehensive, then you would have a mind.
So the other view of the mind is the brain as sort of blank slate.
Right?
This is a very old idea.
You find it in John Locke's writings.
This is the Tabulaza.
And this is this idea that the mind is some kind of information sponge that starts empty,
that starts blank, and that absorbs knowledge and skills from experience.
So it's a sponge that reflects the complexity of the world, the complexity of your life experience,
essentially.
Everything you know and everything you can do is a reflection of something you found
in the outside world, essentially.
So this is an idea that's very old, that was not very popular, for instance, in the 1970s.
But that gained a lot of vitality recently with the rise of connectionism in particular
deep learning.
And so today, deep learning is the dominant paradigm in AI.
And I feel like lots of AI researchers are conceptualizing the mind via a deep learning
metaphor, like they see the mind as a kind of randomly initialized neural network that
starts blank when you're born, and then that gets trained via exposure to training data
that requires knowledge and skills, exposure to training data.
By the way, it's a small tangent.
I feel like people who are thinking about intelligence are not conceptualizing it that
way.
I actually haven't met too many people who believe that a neural network will be able
to reason, who seriously think that, rigorously, because I think it's an actually interesting
worldview.
And we'll talk about it more, but it's been impressive what neural networks have been
able to accomplish.
And it's an eye to me, I don't know, you might disagree, but it's an open question
whether like scaling size eventually might lead to incredible results to us mere humans
will appear as if it's general.
I mean, if you ask people who are seriously thinking about intelligence, they will definitely
not say that all you need to do is like the mind is just in your network.
However, it's actually a view that's very popular, I think in the deep learning community,
that many people are kind of conceptually intellectually lazy about it.
Right.
It's it.
But I guess what I'm saying exactly right is I haven't met many people, and I think
it would be interesting to meet a person who is not intellectually lazy about this particular
topic and still believes that neural networks will go all the way.
I think it's probably closest to that would there are definitely people who argue that
current deep learning techniques are already the way to general artificial intelligence
and that all you need to do is to scale it up to all the available train data.
And that's if you look at the the waves that open the eyes, GPT-3 model is made, you see
echoes of this idea.
So on that topic, GPT-3 similar to GPT-2 actually have captivated some part of the imagination
of the public.
There's just a bunch of hype of different kind that's, I would say it's emergent.
It's not artificially manufactured.
It's just like people just get excited for some strange reason.
In the case of GPT-3, which is funny that there's, I believe a couple of months delay
from release to hype, maybe I'm not historically correct on that, but it feels like there was
a little bit of a lack of hype and then there's a phase shift into hype.
But nevertheless, there's a bunch of cool applications that seem to captivate the imagination
of the public about what this language model that's trained in unsupervised way without
any fine tuning is able to achieve.
So what do you make of that?
What are your thoughts about GPT-3?
Yeah.
So I think what's interesting about GPT-3 is the idea that it may be able to learn new
tasks in after just being shown a few examples.
So I think if it's actually capable of doing that, that's novel and that's very interesting
and that's something we should investigate.
That said, I must say I'm not entirely convinced that we have shown it's capable of doing that
but it's very likely given the amount of data that the model is trained on that what it's
actually doing is better matching a new task you give it with a task that it's been exposed
to in its training data.
It's just recognizing the task instead of just developing a model of the task.
But there's, sorry to interrupt, there's a parallel as to what you said before, which
is it's possible to see GPT-3 as like the prompts it's given as a kind of SQL query
into this thing that it's learned, similar to what you said before, which is language
is used to query the memory.
So is it possible that neural network is a giant memorization thing, but then if it's
gets sufficiently giant, it'll memorize sufficiently large amounts of thing in the world or intelligence
becomes a querying machine.
I think it's possible that a significant chunk of intelligence is this giant associative
memory.
I definitely don't believe that intelligence is just a giant associative memory, but it
may well be a big component.
So do you think GPT-3, 4, 5, GPT-10 will eventually, like what do you think, where's the ceiling?
Do you think you'll be able to reason?
No, that's a bad question.
Like, what is the ceiling is the better question?
Well, what is going to scale?
How good is GPT-N going to be?
So I believe GPT-N is going to improve on the strength of GPT-2 and 3, which is it will
be able to generate ever more plausible text in context.
It's monotonic.
Even a prompt.
An increasing performance.
Yes.
If you train a bigger model on more data, then your text will be increasingly more context
aware and increasingly more plausible in the same way that GPT-3 is much better at generating
plausible text compared to GPT-2.
But that said, I don't think just getting up the model to more transformer layers and
more trained data is going to address the flaws of GPT-3, which is that it can generate
plausible text, but that text is not constrained by anything else other than plausibility.
So in particular, it's not constrained by factualness or even consistency, which is why
it's very easy to get GPT-3 to generate statements that are factually untrue or to generate statements
that are even self-contradictory, because its only goal is plausibility and it has no
other constraints.
It's not constrained to be self-consistent, for instance.
And so for this reason, one thing that I thought was very interesting with GPT-3 is that you
can put it in mind the answer it will give you by asking the question in a specific way,
because it's very responsive to the way you ask the question, since it has no understanding
of the content of the question.
And if you have the same question in two different ways that are basically adversely engineered
to produce a certain answer, you will get two different answers, two contradictory answers.
It's very susceptible to adversarial attacks, essentially.
Potentially, yes.
So in general, the problem with these models, these generative models, is that they have
very good degenerative plausible text, but that's just not enough.
I think one avenue that would be very interesting to make progress is to make it possible to
write programs over the latent space that these models operate on, so that you would
rely on these self-supervised models to generate a sort of flag pool of knowledge and concepts
and common sense.
And then you would be able to write explicit reasoning programs over it.
Because the current problem with GPT-3 is that it can be quite difficult to get it to
do what you want to do.
If you want to turn GPT-3 into products, you need to put constraints on it.
You need to force it to obey certain rules, so you need a way to program it explicitly.
Yeah, so if you look at its ability to do program synthesis, it generates, like you
said, something that's plausible.
Yeah, so if you try to make it generate programs, it will perform well for any program that
it has seen it in its training data, but because program space is not interpretative, right?
It's not going to be able to generalize two problems it hasn't seen before.
Now, that's currently, do you think, sort of, an absurd, but I think useful, I guess,
intuition builder is, you know, the GPT-3 has 175 billion parameters.
Program brain has about a thousand times that, or more, in terms of number of synapses.
Do you think, obviously, very different kinds of things, but there is some degree of similarity.
Do you think, what do you think GPT-3 will look like when it has 100 trillion parameters?
Do you think our conversation might be in nature different, like, because you've criticized
GPT-3 very effectively now?
Do you think?
No, I don't think so.
So the beginning with the bottleneck with scaling up GPT-3, GPT models, genetic pretrain
transformer models, is not going to be the size of the model, or how long it takes to
train it.
The bottleneck is going to be the trained data, because OpenEye is already training GPT-3
on a crawl of basically the entire web, right, and that's a lot of data.
So you could imagine training on more data than that, like Google could train on more
data than that, but it would still be only incrementally more data.
And I don't recall exactly how much more data GPT-3 was trained on compared to GPT-2,
but it's probably at least, like, a hundred, or maybe even a thousand X, don't have the
exact number.
You're not going to be able to train a model on a hundred more data than what you're already
doing.
So that's brilliant.
So it's easier to think of compute as a bottleneck, and then arguing that we can remove that
bottleneck.
We can remove the compute bottleneck, I don't think it's a big problem.
If you look at the pace at which we've improved the efficiency of deep learning models in
the past few years, I'm not worried about training time bottlenecks, or model size
bottlenecks.
The bottleneck in the case of these generative transformer models is absolutely the trained
data.
What about the quality of the data?
So the quality of the data is an interesting point.
The thing is, if you're going to want to use these models in real products, then you want
to feed them data that's as high quality as factual, I would say, and biased as possible,
but there's not really such a thing as unbiased data in the first place.
But you probably don't want to train it on Reddit, for instance, sounds like a bad plan.
So from my personal experience, working with large scale deep learning models, so at some
point I was working on a model at Google that's trained on like 350 million labeled
images.
It's an image classification model.
It's a lot of images that's like probably most publicly available images on the web
at the time.
And it was a very noisy data set because the labels were not originally annotated by hand
by humans.
They were automatically derived from tags on social media or just keywords in the same
page as the image was found and so on.
So it was very noisy.
And it turned out that you could easily get a better model, not just by training, like
if you train on more of the noisy data, you get an incrementally better model, but you
very quickly eat diminishing returns.
On the other hand, if you train on a smaller data set with higher quality annotations,
quality that are annotations that are actually made by humans, you get a better model.
And it also takes less time to train it.
Yeah, that's fascinating.
It's the self-supervised learning.
There's a way to get better doing the automated labeling.
Yeah, so you can enrich or refine your labels in an automated way.
That's correct.
Do you have a hope for, I don't know if you're familiar with the idea of a semantic web.
Because the semantic web, just what people are not familiar, is the idea of being able
to convert the internet or be able to attach semantic meaning to the words on the internet.
The sentence is the paragraphs to be able to convert information on the internet or
some fraction of the internet into something that's interpretable by machines.
That was kind of a dream for, I think the semantic web papers in the 90s, it's kind
of the dream that the internet is full of rich, exciting information.
Even just looking at Wikipedia, we should be able to use that as data for machines.
And so far-
The information is not, it's not really in a format that's available to machines.
So no, I don't think the semantic web will ever work simply because it would be a lot
of work to provide that information in structured form.
And there is not really any incentive for anyone to provide that work.
So I think the way forward to make the knowledge on the web available to machines is actually
something closer to unsupervised deep learning.
Yeah.
The GPT-3 is actually a bigger step in the direction of making the knowledge of the web
available to machines than the semantic web was.
Yeah.
Perhaps in a human-centric sense, it feels like GPT-3 hasn't learned anything that could
be used to reason.
But that might be just the early days.
Yeah.
I think that's correct.
I think the forms of reasoning that you see it perform are basically just reproducing
patterns that it has seen in string data.
So of course, if you're trained on the entire web, then you can produce an illusion of reasoning
in many different situations, but it will break down if it's presented with a novel
situation.
That's the open question between the illusion of reasoning and actual reasoning, yeah.
Yes.
The power to adapt to something that is genuinely new.
Because the thing is, even imagine you had, you could train on every bit of data ever
generated in the history of humanity.
It remains, that model would be capable of anticipating many different possible situations,
but it remains that the future is going to be something different, like, for instance,
if you train a GPT-3 model on data from the year 2002, for instance, and then use it today,
it's going to be missing many things, it's going to be missing many common sense facts
about the world.
It's even going to be missing vocabulary and so on.
Yeah, it's interesting that GPT-3 even doesn't have, I think, any information about the coronavirus.
Yes.
Which is why a system that's, you tell that the system is intelligent when it's capable
to adapt.
So intelligence is going to require some amount of continuous learning.
It's also going to require some amount of improvisation.
It's not enough to assume that what you're going to be asked to do is something that
you've seen before, or something that is a simple interpolation of things you've seen
before.
Yeah.
In fact, that model breaks down for even very tasks that look relatively simple from
a distance, like L5 self-driving, for instance.
Google at the paper a couple years back showing that something like 30 million different road
situations were actually completely insufficient to train a driving model.
It wasn't even L2, right?
And that's a lot of data.
That's a lot more data than the 20 or 30 hours of driving that a human needs to learn to
drive given the knowledge they've already accumulated.
Well, let me ask you on that topic, Elon Musk, Tesla autopilot, one of the only companies,
I believe, is really pushing for a learning-based approach.
You're skeptical that that kind of network can achieve level four?
L4 is probably achievable, L5 is probably not.
What's the distinction there?
Is L5 is completely, you can just fall asleep?
Yeah, L5 is basically human level.
Well, driving, you have to be careful, saying human level, because that's the most-
Yeah, they're all kinds of drivers.
Yeah, that's the clearest example of cars will most likely be much safer than humans
in many situations where humans fail.
It's the vice versa question.
I'll tell you, the thing is the amounts of train data you would need to anticipate for
pretty much every possible situation you learn content in the real world is such that
it's not entirely unrealistic to think that at some point in the future, we'll develop
a system that's trained on enough data, especially provided that we can simulate a lot of that
data.
We don't necessarily need actual cars on the road for everything.
But it's a massive effort.
And it turns out you can create a system that's much more adaptive, that can generalize much
better if you just add explicit models of the surroundings of the car.
And if you use deep learning for what it's good at, which is to provide perceptive information.
So in general, deep learning is a way to encode perception and a way to encode intuition.
But it is not a good medium for any sort of explicit reasoning.
And in AI systems today, strong generalization tends to come from explicit models, tend to
come from abstractions in the human mind that are encoded in program form by a human engineer.
These are the abstractions that can actually generalize, not the sort of weak abstraction
that is learned by a neural network.
And the question is how much reasoning, how much strong abstractions are required to solve
particular tasks like driving?
That's the question.
Or human life, existence.
How much strong abstractions does existence require?
But more specifically on driving, that seems to be a coupled question about intelligence.
How much intelligence, like how do you build an intelligence system?
And the coupled problem, how hard is this problem?
How much intelligence does this problem actually require?
So we get to cheat, right, because we get to look at the problem.
It's not like you get to close our eyes and completely new to driving.
We get to do what we do as human beings, which is for the majority of our life, before we
ever learn quote unquote to drive, we get to watch other cars and other people drive.
We get to be in cars, we get to watch, we get to see movies about cars, we get to observe
all this stuff.
And that's similar to what neural networks are doing, it's getting a lot of data.
And the question is how many leaps of reasoning genius is required to be able to actually effectively
drive?
For example, driving, I mean, sure, you've seen a lot of cars in your life before you
learn to drive, but let's say you've learned to drive in Silicon Valley, and now you rent
a car in Tokyo, well, now everyone is driving on the other side of the road, and the signs
are different, and the roads are more narrow and so on.
So it's a very, very different environment.
And a smart human, even an average human, should be able to just zero shot it, to just
be operational in this very different environment right away, despite having had no contact
with the novel complexity that is contained in this environment.
And that is novel complexity is not just interpolation over the situations that you've encountered
previously, like learning to drive in the US, right?
I would say the reason I ask is one of the most interesting tests of intelligence we
have today, actively, which is driving in terms of having an impact on the world.
Like when do you think we'll pass that test of intelligence?
So I don't think driving is that much of a test of intelligence, because again, there
is no task for which skill at that task demonstrates intelligence, unless it's a kind of meta tasks
that involves acquiring new skills.
So I don't think I think you can actually solve driving without having any real amount
of intelligence.
For instance, if you really did have infinite train data, you could just literally train
an end-to-end deep learning model that does driving, provided infinite train data.
The only problem with the whole idea is collecting a data set that's sufficiently comprehensive
that covers the very long tail of possible situations you might encounter.
And it's really just a scale problem.
So I think there's nothing fundamentally wrong with this plan, with this idea.
It's just that it strikes me as a fairly inefficient thing to do, because you run into this scaling
issue with diminishing returns, whereas if instead you took a more manual engineering
approach where you use deep learning modules in combination with engineering an explicit
model of the surrounding of the cars, and you bridge the two in a clever way.
Your model will actually start generalizing much earlier and more effectively than the
end-to-end deep learning model.
So why would you not go with the more manual engineering oriented approach?
Even if you created that system, either the end-to-end deep learning model system that's
trained on infinite data, or the slightly more human system, I don't think achieving
L5 would demonstrate a general intelligence or intelligence of any generality at all.
Again, the only possible test of generality in AI would be a test that looks at skill
acquisition over unknown tasks.
So for instance, you could take your L5 driver and ask it to learn to pilot a commercial
airplane, for instance.
And then you would look at how much human involvement is required and how much training
data is required for the system to learn to pilot an airplane.
And that gives you a measure of how intelligent the system really is.
Yeah.
I mean, that's a big leap.
I get you.
But I'm more interested as a problem, I would see, to me, driving is a black box that can
generate novel situations at some rate, like what people call edge cases.
So it does have newness that keeps being like we're confronted, let's say, once a month.
It is a very long tail.
Yes.
It's a long tail.
That doesn't mean you cannot solve it just by training as a school model on a lot of
data.
Huge amount of data.
It's really amount of scale.
But I guess what I'm saying is if you have a vehicle that achieves L5, it is going to
be able to deal with new situations.
Or I mean, the data is so large that the rate of new situations is very low.
Yes.
That's not intelligent.
So if we go back to your definition of intelligence, it's the efficiency with which you can adapt
to new situations, to truly new situations, not situations you've seen before, not situations
that could be anticipated by your creators, by the creators of the system, but through
new situations.
The efficiency with which you acquire new skills.
If you require, in order to pick up a new skill, you require a very extensive training
dataset of most possible situations that can occur in the practice of that skill, then
the system is not intelligent.
It is mostly just a lookup table.
Likewise, if in order to acquire a skill, you need a human engineer to write down a bunch
of rules that cover most or every possible situation, likewise, the system is not intelligent.
The system is merely the output artifact of a process that happens in the minds of the
engineers that are creating it.
It is encoding an abstraction that's produced by the human mind, and intelligence would
actually be the process of autonomously producing this abstraction.
If you take an abstraction and you encode it on a piece of paper or in a computer program,
the abstraction itself is not intelligent.
This intelligent is the agent that's capable of producing these abstractions.
It feels like there's a little bit of a gray area, because you're basically saying that
deep learning forms abstractions, too, but those abstractions do not seem to be effective
for generalizing far outside of the things that you've already seen, but generalize a
little bit.
Yeah, absolutely.
No deep learning does generalize a little bit.
Modernization is not binary, it's more like a spectrum.
There's a certain point, it's a gray area, but there's a certain point where there's
an impressive degree of generalization that happens.
I guess exactly what you were saying is intelligence is how efficiently you're able to generalize
far outside of the distribution of things you've seen already.
It's both the distance of how far you can, how new, how radically new something is, and
how efficiently you're able to deal with that.
You can think of intelligence as a measure of an information conversion ratio.
Imagine a space of possible situations, and you've covered some of them, so you have some
amount of information about your space of possible situations that's provided by the
situations you already know, and that's on the other hand also provided by the prior
knowledge that the system brings to the table or the prior knowledge that's embedded in
the system.
The system starts with some information about the problem, about the task, and it's about
going from that information to a program, what you would call a skill program, a behavioral
program that can cover a large area of possible situation space.
Essentially the ratio between that area and the amount of information you start with is
intelligence.
A very smart agent can make efficient users of very little information about a new problem
and very little prior knowledge as well to cover a very large area of potential situations
in that problem, without knowing what these future new situations are going to be.
One of the other big things you talk about in the paper, we've talked about it a little
bit already, but let's talk about it some more as the actual tests of intelligence.
If we look at human and machine intelligence, do you think tests of intelligence should
be different for humans and machines, or how we think about testing of intelligence?
Are these fundamentally the same kind of intelligences that we're after and therefore
the tests should be similar?
If your goal is to create AIs that are more human-like, then it will be super valuable,
obviously, to have a test that's universal, that applies to both AIs and humans, so that
you could establish a comparison between the two that you could tell exactly how intelligent
in terms of human intelligence a given system is.
Like I said, the constraints that apply to artificial intelligence and to human intelligence
are very different, and your test should account for this difference.
Because if you look at artificial systems, it's always possible for an experimenter to
buy arbitrary levels of skill at arbitrary tasks, either by injecting a hard-coded prior
knowledge into the system via rules and so on that come from the human mind, from the
minds of the programmers, and also buying higher levels of skill just by training on
more data.
For instance, you could generate an infinity of different Go games, and you could train
a Go playing system that way, but you could not directly compare it to human Go playing
skills, because a human that plays Go had to develop that skill in a very constrained
environment.
They had the limited amount of time, they had the limited amount of energy, and of course,
they started from a different set of priors, they started from innate human priors.
So I think if you want to compare the intelligence of two systems, like the intelligence of an
AI and the intelligence of a human, you have to control for priors.
You have to start from the same set of knowledge priors about the task, and you have to control
for experience, that is to say, for training data.
So what's priors?
So priors is whatever information you have about a given task before you start learning
about this task.
And how's the difference from experience?
Well, experience is acquired.
Right.
For instance, if you're trying to play Go, your experience with Go is all the Go games
you've played or you've seen, or you've simulated in your mind, let's say.
And your priors are things like, well, Go is a game on a 2D grid, and we have lots of
hard-coded priors about the organization of 2D space, and the rules of how the dynamics
of the physics of this game in this 2D space.
And the idea that you have what winning is.
Yes, exactly.
And other board games can also share some similarities with Go, and if you've played
these board games, then with respect to the game of Go, that would be part of your priors
about the game.
Well, it's interesting to think about the game of Go is how many priors are actually
brought to the table.
When you look at self-play, reinforcement learning-based mechanisms that do learning,
it seems like the number of priors is pretty low.
Yes.
But you're saying you should be...
There is a 2D special priors in the cabinet.
Right.
But you should be clear at making those priors explicit.
Yes.
So, in part, I think if your goal is to measure a human-like form of intelligence, then you
should clearly establish that you want the AI you're testing to start from the same set
of priors that humans start with.
Right.
So, I mean, to me personally, but I think to a lot of people, the human side of things
is very interesting.
So testing intelligence for humans.
But what do you think is a good test of human intelligence?
Well, let's do a question that Psychometrics is interested in, and there's an entire subfield
of psychology that deals with this question.
So what's Psychometrics?
The Psychometrics is the subfield of psychology that tries to measure, quantify aspects of
the human mind.
And in particular, cognitive abilities, intelligence, and personality traits as well.
So what are, might be a weird question, but what are the first principles of Psychometrics
that operates on, you know, what are the priors that brings to the table?
So it's a field with a fairly long history.
So, you know, psychology sometimes gets a bad reputation for not having very reproducible
results.
And so Psychometrics has actually some fairly slightly reproducible results.
So the ideal goals of the field is, you know, tests should be reliable, which is a notion
tied to reproducibility.
It should be valid, meaning that it should actually measure what you say, what you say
it measures.
So for instance, if you're saying that you're measuring intelligence, then your test results
should be correlated with things that you expect to be correlated with intelligence,
like success in school or success in the workplace and so on, should be standardized, meaning
that you can administer your tests to many different people in some conditions.
And it should be free from bias, meaning that, for instance, if your test involves the English
language, then you have to be aware that this creates a bias against people who have English
as their second language or people who can't speak English at all.
So of course, these principles for creating Psychometric tests are very much 90 all.
I don't think every Psychometric test is really either reliable, valid, or offered from bias,
but at least the field is aware of these weaknesses and is trying to address them.
So it's kind of interesting, ultimately, you're only able to measure, like you said previously,
the skill, but you're trying to do a bunch of measures of different skills that correlate,
as you mentioned, strongly with some general concept of cognitive ability.
Yes, yes.
So what's the G factor?
So right, there are many different kinds of tests of intelligence, and each of them is
interested in different aspects of intelligence, you know, some of them will deal with language,
some of them will deal with special vision, maybe mental rotations, numbers, and so on.
When you run these very different tests at scale, what you start seeing is that there
are clusters of correlations among test results.
So for instance, if you look at homework at school, you will see that people who do well
at math are also likely statistically to do well in physics.
And what's more, there are also people who do well at math and physics are also statistically
likely to do well in things that sound completely unrelated, like writing in English, for instance.
And so when you see clusters of correlations in statistical terms, you would explain them
with a latent variable, and the latent variable that would, for instance, explain the relationship
between being good at math and being good at physics would be cognitive ability, right?
And the G factor is the latent variable that explains the fact that every test of intelligence
that you can come up with results on this test end up being correlated.
So there is some single unique variable that explains these correlations, that's the G factor.
So it's a statistical construct.
It's not really something you can directly measure, for instance, in a person.
But it's there.
But it's there.
It's there, it's there at scale.
And that's also one thing I want to mention about psychometrics, like, you know, when
you talk about measuring intelligence in humans, for instance, some people get a little bit
worried, they will say, you know, that sounds dangerous, maybe that sounds potentially discriminatory
and so on.
And they are not wrong.
And the thing is, personally, I'm not interested in psychometrics as a way to characterize
one individual person, like, if, if I get your psychometric personality assessments
or your IQ, I don't think that actually tells me much about you as a person, I think psychometrics
is most useful as a statistical tool.
So it's most useful at scale.
It's most useful when you start getting test results for a large number of people and you
start cross-correlating these test results, because that gives you information about the
structure of the human mind, particularly about the structure of human cognitive abilities.
So at scale, psychometrics paints a certain picture of the human mind.
And that's interesting.
And that's what's driven to AI, the structure of human cognitive abilities.
Yeah, it gives you an insight into, I mean, to me, I remember when I learned about G-Factor,
it seemed, it seemed like it would be impossible for it to be real, even as a statistical variable,
like, it felt kind of like astrology, like it's like wishful thinking about psychologists.
But the more I learned, I realized that there's some, I mean, I'm not sure what to make about
human beings, the fact that the G-Factor is a thing, that there's a commonality across
all of human species, that there doesn't need to be a strong correlation between cognitive
abilities.
That's kind of fascinating, actually.
Yeah, so human cognitive abilities have a structure, like the most mainstream theory
of the structure of cognitive abilities is called CHC theory, it's a cattle home, carol,
it's the name of the three psychologists who contributed key pieces of it.
And it describes cognitive abilities as a hierarchy with three levels.
And at the top, you have the G-Factor.
Then you have broad cognitive abilities, for instance, fluid intelligence, right?
That encompasses a broad set of possible kinds of tasks that are all related.
And then you have narrow cognitive abilities at the last level, which is closer to task
specific skill.
And there are actually different theories of the structure of cognitive abilities that
just emerge from different statistical analysis of IQ test results, but they all describe
a hierarchy with a kind of G-Factor at the top.
And you're right that the G-Factor is, it's not quite real in the sense that it's not
something you can observe and measure, like your height, for instance, but it's really
in the sense that you see it in a statistical analysis of the data, right?
One thing I want to mention is that the fact that there is a G-Factor does not really mean
that human intelligence is, in general, in a strong sense, does not mean human intelligence
can be applied to any problem at all, and that someone who has a high IQ is going to
be able to solve any problem at all.
It's not quite what it means, I think.
One popular analogy to understand it is the sports analogy.
If you consider the concept of physical fitness, it's a concept that's very similar to intelligence
because it's a useful concept, it's something you can intuitively understand.
Some people are fit, maybe like you, some people are not as fit, maybe like me.
But none of us can fly.
Absolutely.
Even if you're very fit, that doesn't mean you can do anything at all in any environment.
You obviously cannot fly, you cannot serve at the bottom of the ocean and so on.
If you were a scientist and you wanted to precisely define and measure physical fitness
in humans, then you would come up with a battery of tests, like you would have running
100 meter, playing soccer, playing table tennis, swimming, and so on.
If you ran these tests over many different people, you would start seeing correlations
and test results.
For instance, people who are good at soccer are also good at sprinting.
You would explain these correlations with physical abilities that are strictly analogous
to cognitive abilities.
And then you would start also observing correlations between biological characteristics, like maybe
lung volume is correlated with being a fast runner, for instance.
In the same way that there are neurophysical correlates of cognitive abilities.
And at the top of the hierarchy of physical abilities that you would be able to observe,
you would have a g-factor, a physical g-factor, which would map to physical fitness.
And as you just said, that doesn't mean that people with high physical fitness can fly.
It doesn't mean human morphology and human physiology is universal.
It's actually super specialized.
We can only do the things that evolve to do.
You could not exist on Venus or Mars or in the void of space or the bottom of the ocean.
So that said, one thing that's really striking and remarkable is that our morphology generalizes
far beyond the environments that we evolved for, like in a way you could say we evolved
to run after prey in the savanna, right?
That's very much where our human morphology comes from.
And that said, we can do a lot of things that are completely unrelated to that.
We can climb mountains, we can swim across lakes, we can play table tennis.
I mean, table tennis is very different from what we were evolved to do, right?
So our morphology, our bodies, our sense and motor affordances have a degree of generality
that is absolutely remarkable, right?
And I think cognition is very similar to that.
Our cognitive abilities have a degree of generality that goes far beyond what the mind was initially
supposed to do, which is why we can play music and write novels and go to Mars and do all
kinds of crazy things.
But it's not universal in the same way that human morphology and our body is not appropriate
for actually most of the universe by volume in the same way you could say that the human
mind is not really appropriate for most of potential problem space by volume.
So we have very strong cognitive biases, actually.
That means that there are certain types of problems that we handle very well and certain
types of problems that we are completely inadapted for.
So that's really how we interpret the g-factor.
It's not a sign of strong generality.
It's really just the broadest cognitive ability.
But our abilities, whether we are talking about sensory motor abilities or cognitive
abilities, they remain very specialized in the human condition, right?
Within the constraints of the human cognition, they're general.
Yes, absolutely.
But the constraints, as you're saying, are very limited in terms of limiting.
So we evolved our cognition and our body evolved in very specific environments.
Because our environment was so valuable, fast-changing, and so unpredictable, part of the constraints
that drove our evolution is generality itself.
So we were, in a way, evolved to be able to improvise in all kinds of physical, cognitive
environments, right?
And for this reason, it turns out that the minds and bodies that we ended up with can
be applied to much, much broader scope than what they were evolved for, right?
And that's truly remarkable.
And that's a degree of generalization that is far beyond anything you can see in artificial
systems today, right?
That's it.
It does not mean that human intelligence is anywhere universal.
Yeah, it's not general.
It's a kind of exciting topic for people, even outside of artificial intelligence, IQ tests.
I think it's Mensa, whatever.
There's different degrees of difficulty for questions.
We talked about this offline a little bit, too, about sort of difficult questions.
What makes a question on an IQ test more difficult or less difficult, do you think?
So the thing to keep in mind is that there's no such thing as a question that's intrinsically
difficult.
It has to be difficult with respect to the things you already know and the things you
can already do, right?
So in terms of an IQ test question, typically it would be structured, for instance, as a
set of demonstration input and output pairs, right?
And then you would be given a test input, a prompt, and you would need to recognize
or produce the corresponding output.
And in that narrow context, you could say a difficult question is a question where the
input prompt is very surprising and unexpected given the training example.
Just even the nature of the patterns that you're observing in the input prompt.
For instance, let's say you have a rotation problem.
You must rotate the shape by 90 degrees.
If I give you two examples and then I give you one prompt, which is actually one of the
two training examples, then there is zero generalization difficulty for the task.
It's actually a trivial task.
You just recognize that it's one of the training examples and you produce the same answer.
Now if it's a more complex shape, there is a little bit more generalization, but it remains
that you are still doing the same thing at this time as you were being demonstrated at
the training time.
A difficult task does require some amount of test time adaptation, some amount of improvisation.
So consider, I don't know, you're teaching a class on quantum physics or something.
If you wanted to kind of test the understanding that students have of the material, you would
come up with an exam that's very different from anything they've seen, like on the Internet
when they were cramming.
On the other hand, if you wanted to make it easy, you would just give them something that's
very similar to the mock exams that they've taken, something that's just a simple interpolation
of questions that they've already seen.
And so that would be an easy exam.
It's very similar to what you've been trained on, and a difficult exam is one that really
probes your understanding because it forces you to improvise.
It forces you to do things that are different from what you were exposed to before.
So that said, it doesn't mean that the exam that requires improvisation is intrinsically
hard because maybe you're a quantum physics expert.
So when you take the exam, this is actually stuff that, despite being new to the students,
it's not new to you.
So it can only be difficult with respect to what the test taker already knows and with
respect to the information that the test taker has about the task.
So that's what I mean by controlling for priors what the information will bring to the table.
And experience, which is the training data, so in the case of the quantum physics exam,
that would be all the course material itself and all the mock exams that students might
have taken online.
Yeah, it's interesting because I've also, I sent you an email and I asked you, I've
been, this is just this curious question of what's a really hard IQ test question.
And I've been talking to also people who have designed IQ tests as a few folks on the internet.
It's like a thing.
People are really curious about it.
First of all, most of the IQ tests they designed, they like religiously protect against the
correct answers.
Like you can't find the correct answers anywhere.
In fact, the question is ruined once you know, even like the approach you're supposed to
take.
That said, the approach is implicit in the training examples.
So if you release the training examples, it's over.
Which is why in ARC, for instance, there is a test set that is private and no one has
seen it.
No, for really tough IQ questions, it's not obvious.
It's not because the ambiguity.
Like it's, I mean, we'll have to look to them, but like some number sequences and so on.
It's not completely clear.
So like you can get a sense, but there's like some, you know, when you look at a number
sequence, I don't know, like your Fibonacci number sequence, if you look at the first
few numbers, that sequence could be completed in a lot of different ways.
And you know, some are, if you think deeply, are more correct than others.
Like there's a kind of intuitive simplicity and elegance to the correct solution.
Yes.
I am personally not a fan of ambiguity in test questions actually, but I think you can
have difficulty without requiring ambiguity simply by making the test require a lot of
extrapolation over the training examples.
But the beautiful question is difficult, but gives away everything when you give the training
example.
Basically yes, meaning that, so the tests I'm interested in creating are not necessarily
difficult for humans because human intelligence is the benchmark that's supposed to be difficult
for machines in ways that are easy for humans.
Like I think an ideal test of human and machine intelligence is a test that is actionable that
highlights the need for progress and that highlights the direction in which you should
be making progress.
I think we'll talk about the RARC challenge and the tests you've constructed and you
have these elegant examples.
I think that highlight, like this is really easy for us humans, but it's really hard
for machines.
But on the, you know, the designing an IQ test for IQs of like a higher than 160 and
so on, you have to say, you have to take that and put out steroids, right?
You have to think like, what is hard for humans?
And that's a fascinating exercise in itself, I think.
And it was an interesting question of what it takes to create a really hard question
for humans because you again have to do the same process as you mentioned, which is, you
know, something basically where the experience that you have likely to have encountered throughout
your whole life, even if you've prepared for IQ tests, which is a big challenge, that
this will still be novel for you.
Yeah.
I mean, novelties are requirements.
You should not be able to practice for the questions that you're going to be tested on.
That's important because otherwise what you're doing is not exhibiting intelligence, what
you're doing is just retrieving what you've been exposed before.
It's the same thing as a deep learning model.
If you train a deep learning model on all the possible answers, then it will ace your
test in the same way that, you know, a stupid student can still ace the test if they cram
for it.
They memorize, you know, a hundred different possible mock exams and then they hope that
the actual exam will be a very simple interpolation of the mock exams.
And that student could just be a deep learning model at that point.
But you can actually do that without any understanding of the material.
And in fact, many students pass the exams in exactly this way.
And if you want to avoid that, you need an exam that's unlike anything they've seen that
really probes their understanding.
So how do we design an IQ test for machines and intelligent tests for machines?
All right.
So in the paper, I outline a number of requirements that you expect of such a test.
And in particular, we should start by acknowledging the priors that we expect to be required in
order to perform the test.
So we should be explicit about the priors, right?
And if the goal is to compare machine intelligence and human intelligence, then we should assume
human cognitive priors, right?
And secondly, we should make sure that we are testing for skill acquisition ability,
skill acquisition efficiency in particular, and not for skill itself, meaning that every
task featured in your test should be novel and should not be something that you can anticipate.
So for instance, it should not be possible to brute force the space of possible questions,
right, to pre-generate every possible question and answer.
So it should be tasks that cannot be anticipated, not just by the system itself, but by the
creators of the system, right?
Yeah.
You know what's fascinating?
I mean, one of my favorite aspects of the paper and the work you do, the ARC challenge,
is the process of making priors explicit.
Just even that act alone is a really powerful one of like, what are, it's a really powerful
question to ask of us humans, what are the priors that we bring to the table?
So the next step is like, once you have those priors, how do you use them to solve a novel
task, but like just even making the priors explicit is a really difficult and really
powerful step.
And that's like visually beautiful and conceptually philosophically beautiful part of the work
you did with, and I guess continue to do probably with the paper and the ARC challenge.
Can you talk about some of the priors that we're talking about here?
Yes.
A researcher has done a lot of work on what exactly are the knowledge priors that are
innate to humans is Elizabeth Spelke from Harvard.
So she developed the core knowledge theory, which outlines four different core knowledge
systems.
So systems of knowledge that we are basically either born with or that we are hardwired to
acquire very early on in our development.
And there's no, there's no strong distinction between the two, like if you are primed to
acquire a certain type of knowledge in just a few weeks, you might as well just be born
with it, it's just part of who you are.
And so there are four different core knowledge systems, like the first one is the notion
of objectness and basic physics, like you recognize that something that moves currently,
for instance, is an object.
So we intuitively naturally, innately divide the world into objects based on this notion
of coherence, physical coherence.
And in terms of elementary physics, there's the fact that objects can bump against each
other and the fact that they can occlude each other.
So these are things that we are essentially born with or at least that we are going to
be acquiring extremely early because we're really hardwired to acquire them.
So a bunch of points, pixels that move together on objects are partly the same object.
Yes.
I mean, that, like, I don't smoke weed, but if I did, that's something I could sit like
all night and just like think about, remember right first in your paper, just objectness,
I wasn't self-aware, I guess, of how that particular prior, that that's such a fascinating
prior that, like...
That's the most basic one, but...
Objectness, just identity, just objectness, and it's very basic, I suppose, but it's
so fundamental.
It is fundamental to human cognition.
And the second prior, that's also fundamental, is agentness, which is not a real world, a
real world, but so agentness.
The fact that some of these objects that you segment your environment into, some of these
objects are agents.
So what's an agent?
It's basically it's an object that has goals.
That has what?
That has goals.
They're capable of pursuing goals.
So for instance, if you see two dots moving in a roughly synchronized fashion, you will
intuitively infer that one of the dots is pursuing the other.
So that one of the dots is, and one of the dots is an agent, and its goal is to avoid
the other dot.
And one of the dots, the other dot is also an agent, and its goal is to catch the first
dot.
Pelki has shown that babies as young as three months identify agentness and goal-directedness
in their environment.
Another prior is basic geometry and topology, like the notion of distance, the ability to
navigate in your environment and so on.
This is something that is fundamentally hardwired into our brain.
It's in fact backed by very specific neural mechanisms, like for instance grid cells and
plate cells.
So it's something that's literally hard-coded at the neural level in our hippocampus.
And the last prior would be the notion of numbers, like numbers are not actually a cultural
construct.
They are intuitively, innately able to do some basic counting and to compare quantities.
So it doesn't mean we can do arbitrary, arithmetic counting, like counting one, two, three, then
maybe more than three.
You can also compare quantities if I give you three dots and five dots.
You can tell the side with five dots has more dots.
So this is actually an innate prior.
So that said, the list may not be exhaustive, so Spelke is still pursuing the potential
existence of new knowledge systems, for instance, knowledge systems that we deal with social
relationships.
Yeah, which is much less relevant to something like ARC or IQ test.
Right.
So there could be stuff that's like you said, rotation, symmetry, is it really interesting?
It's very likely that there is, speaking about rotation, that there is in the brain a hard-coded
system that is capable of performing rotations.
One famous experiment that people did in the, I don't remember who it was exactly, but
in the 70s was that people found that if you asked people, if you give them two different
shapes and one of the shapes is a rotated version of the first shape and you ask them,
is that shape a rotated version of the first shape or not?
What you see is that the time it takes people to answer is linearly proportional to the
angle of rotation.
So it's almost like you have in somewhere in your brain a turntable with a fixed speed.
And if you want to know if two objects are rotated versions of each other, you put the
object on the turntable, you let it move around a little bit and then you stop when
you have a match.
And that's really interesting.
So what's the arc challenge?
So in the paper outline, all these principles, that's a good test of machine intelligence
and human intelligence should follow.
And the arc challenge is one attempt to embody as many of these principles as possible.
So I don't think it's anywhere near a perfect attempt, right?
It does not actually follow every principle, but it is what I was able to do given the
constraints.
So the format of arc is very similar to classic IQ tests, in particular Raven's Progessive
Metruses.
Ravens?
Yeah, Raven's Progessive Metruses.
I mean, if you've done IQ tests in the past, you know what it is probably, at least you've
seen it, even if you don't know what it's called.
And so you have a set of tasks, that's what they're called.
And for each task, you have training data, which is a set of input and output pairs.
So an input or output pair is a grid of colors, basically.
The size of the grid is variable.
The size of the grid is variable.
And you're given an input, and you must transform it into the proper output, right?
And so you're shown a few demonstrations of a task in the form of existing input output
pairs, and then you're given a new input.
And you must provide, you must produce the correct output.
And the assumptions in arc is that every task should only require core knowledge priors,
should not require any outside knowledge.
So for instance, no language, no English, nothing like this.
No concepts taken from our human experience, like trees, dogs, cats, and so on.
So only tasks that are, reasoning tasks that are built on top of core knowledge priors.
And some of the tasks are actually explicitly trying to probe specific forms of abstraction,
right?
Part of the reason why I wanted to create arc is I'm a big believer in, you know, when
you're faced with a problem as murky as understanding how to autonomously generate abstraction in
a machine, you have to co-evolve the solution and the problem.
And so part of the reason why I designed arc was to clarify my ideas about the nature of
abstraction, right?
And some of the tasks are actually designed to probe bits of that theory.
And there are things that are turned out to be very easy for humans to perform, including
young kids, right?
But turned out to be near impossible for machines.
Whatever you learn from the nature of abstraction, from designing that, can you clarify what
you mean one of the things you wanted to try to understand was this idea of abstraction?
Yes.
So clarifying my own ideas about abstraction by forcing myself to produce tasks that would
require the ability to produce that form of abstraction in order to solve them.
Got it.
Okay.
So, and by the way, just to, I mean, people should check out, I'll probably overlay if
you're watching the video part, but the grid input output with the different colors on
the grid, that's it, that's that means a very simple world.
But it's kind of beautiful.
It's very similar to classic acutas, like it's not very original in that sense.
The main difference with acutas is that we make the priors explicit, which is not usually
the case in acutas.
So you might get explicit that everything should only be built onto of core knowledge
priors.
I also think it's generally more diverse than acutas in general.
And it's, it perhaps requires a bit more manual work to produce solutions because you have
to click around on a grid for a while.
Just the grids can be as large as 30 by 30 cells.
So how did you come up if you can reveal with the questions, like what's the process
of the questions?
Was it mostly you that came up with the questions?
What, how difficult is it to come up with a question?
Like is this scalable to a much larger number?
If you think, you know, with acutas, you might not necessarily want it to or needed to be
scalable with machines, as possible, you could argue that it needs to be scalable.
So there are a thousand questions, a thousand tasks, including the test set, the private
test set.
I think it's fairly difficult in the sense that a big requirement is that every task
should be novel and unique and unpredictable.
Right.
But you don't want to create your own little world that is simple enough that it would
be possible for a human to reverse and generate and write down an algorithm that could generate
every possible arc task and their solution.
So in a sense, that would completely invalidate the test.
So you're constantly coming up with new stuff.
You need, yeah, you need a source of novelty, of unfakeable novelty.
And one thing I found is that as a human, you are not a very good source of unfakeable
novelty.
And so you have to base the creation of these tasks quite a bit.
There are only so many unique tasks that you can do in a given day.
So I mean, it's coming up with truly original and new ideas.
Did psychedelics help you at all?
No, it's okay.
But I mean, that's fascinating to think about, like, so you would be like walking or something
like that.
Are you constantly thinking of something totally new?
Yes.
I mean, this is hard.
This is hard.
Yeah.
I mean, I'm not saying I've done anywhere near perfect job at it.
There is some amount of redundancy and there are many imperfections in arc.
So that said, you should consider arc as a work in progress.
It is not the definitive state where the arc tasks today are not definitive states of the
test.
I want to keep refining it.
In the future, I also think it should be possible to open up the creation of tasks to
broad audience, to do crowd sourcing.
That would involve several levels of filtering, obviously, but I think it's possible to apply
crowd sourcing to develop a much bigger and much more diverse arc data set that would
also be free of potentially some of my own personal biases.
Is there always need to be a part of arc that the test is hidden?
Yes.
Absolutely.
It is impressive that the test that you're using to actually benchmark algorithms is not
accessible to the people developing these algorithms, because otherwise, what's going
to happen is that the human engineers are just going to solve the tasks themselves and
encode their solution in program form.
But that, again, what you're seeing here is the process of intelligence happening in
the mind of the human, and then you're just capturing its crystallized output.
But that crystallized output is not the same thing as the process generated.
It's not intelligent.
So, by the way, the idea of crowd sourcing it is fascinating.
I think the creation of questions is really exciting for people.
I think there's a lot of really brilliant people out there that love to create these
kinds of stuff.
Yeah.
One thing that surprised me that I wasn't expecting is that lots of people seem to actually
enjoy ARC as a kind of game, and I was really seeing it as a test, as a benchmark of fluid
general intelligence, and lots of people, including kids, are just enjoying it as a
game.
So, I think that's encouraging.
Yeah.
I'm fascinated by it.
There's a world of people who create IQ questions.
I think that's a cool activity for machines and for humans.
People, humans are themselves fascinated by taking the questions like measuring their
own intelligence.
I mean, that's just really compelling.
It's really interesting to me, too.
It helps.
One of the cool things about ARC, you said, is kind of inspired by IQ tests or whatever,
it follows a similar process.
But because of its nature, because of the context in which it lives, it immediately forces
you to think about the nature of intelligence as opposed to just a test of your own.
It forces you to really think.
I don't know if it's within the question, inherent in the question, or just the fact
that it lives in a test that's supposed to be a test of machine intelligence.
Absolutely.
As you solve ARC tasks as a human, you will be forced to basically introspect how you
come up with solutions, and that forces you to reflect on the human problem-solving process
and the way your own mind generates abstract representations of the problems it's exposed
to.
I think it's due to the fact that the set of core knowledge priors that ARC is built
upon is so small.
It's all a recombination of a very, very small set of assumptions.
Okay.
So what's the future of ARC?
So you held ARC as a challenge, as part of like a Kegel competition, Kegel competition.
What do you think?
Do you think that's something that continues for five years, 10 years, like just continues
growing?
Yes, absolutely.
So ARC itself will keep evolving, so I've talked about cloud sourcing, I think that's
a good avenue.
Another thing I'm starting is I'll be collaborating with folks from the psychology department at
NYU to do human testing on ARC.
I think there are lots of interesting questions you can start asking, especially as you start
coordinating machine solutions to ARC tasks and the human characteristics of solutions.
Like for instance, you can try to see if there's a relationship between the human-perceived
difficulty of a task and the machine-perceived, yes, and exactly some measure of machine-perceived
difficulty.
Yeah.
It's a nice playground in which to explore this very difference.
It's the same thing as we talked about the autonomous vehicles.
The things that could be difficult for humans might be very different than the things that
do.
Yes, absolutely.
And formalizing or making explicit that difference in difficulty may teach us something fundamental
about intelligence.
So one thing I think we did well with ARC is that it's proving to be a very actionable
test in the sense that machine performance on ARC started at very much zero initially
while humans found actually the tasks very easy.
And that alone was like a big red flashing light saying that something is going on and
that we are missing something.
And at the same time, machine performance did not stay at zero for very long actually.
Within two weeks of the Kaggle competition, we started having a non-zero number.
And now the state of the art is around 20% of the test set solved.
And so ARC is actually a challenge where our capabilities start at zero, which indicates
the need for progress.
But it's also not an impossible challenge.
It's not accessible.
You can start making progress basically right away.
At the same time, we are still very far from having solved it.
And that's actually a very positive outcome of the competition is that the competition
has proven that there was no obvious shortcut to solve these tasks.
Right.
Yeah, so the test held up.
Yeah, exactly.
It was the primary reason to the Kaggle competition is to check if some clever person was going
to hack the benchmark that did not happen, right, like people who are solving the task
are essentially doing it.
Well, in a way, they're actually exploring some flaws of ARC that we will need to address
in the future, especially they're essentially anticipating what sort of tasks may be contained
in the test set, right?
Right, which is kind of, yeah, that's the kind of hacking, it's human hacking of the
test.
Like I said, you know, with the state of the art, it's like 20% we're still very, very
far from human level, which is closer to 100% and so, and I do believe that, you know,
it will take a while until we reach a human parity on ARC and that by the time we have
human parity, we will have AI systems that are probably pretty close to human level in
terms of general fluid intelligence, which is, I mean, they're not going to be necessarily
human like, they're not necessarily, you would not necessarily recognize them as, you know,
being an AI, but they would be capable of a degree of generalization that matches the
generalization performed by human fluid intelligence.
Sure.
I mean, this is a good point in terms of general fluid intelligence to mention in your paper,
you describe different kinds of generalizations, local, broad, extreme, and there's a kind of
hierarchy that you form.
So when we say generalizations, what are we talking about?
What kinds are there?
Right.
So generalization is a very old idea, I mean, it's even older than machine learning.
In the context of machine learning, you say a system generalizes if it can make sense
of an input it has not yet seen.
And that's what I would call a system-centric generalization, generalization with respect
to novelty for the specific system you're considering.
So I think a good test of intelligence should actually deal with developer-aware generalization,
which is slightly stronger than system-centric generalizations.
So developer-aware generalization would be the ability to generalize to novelty or uncertainty
that not only the system itself has not access to, but the developer of the system could not
have access to either.
That's a fascinating meta definition.
So like the system is basically the edge case thing we're talking about with autonomous
vehicles.
Yes.
The developer nor the system know about the edge cases in my encounter.
So it's up to the system should be able to generalize the thing that nobody expected,
neither the designer of the training data nor obviously the contents of the training
data.
That's a fascinating definition.
So you can see generalization, degrees of generalization as a spectrum.
And the lowest level is what machine learning is trying to do, is the assumption that any
new situation is going to be sampled from a static distribution of possible situations
and that you already have a representative sample of the distribution.
That's your training data.
And so in machine learning, you generalize to a new sample from a known distribution.
And the ways in which your new sample will be new or different are ways that are already
understood by the developers of the system.
So you are generalizing to known unknowns for one specific task.
That's what you would call robustness.
You are robust to things like noise, small variations and so on.
Or one fixed known distribution that you know through your training data.
And a higher degree would be flexibility in machine intelligence.
So flexibility would be something like an L5 cell driving car or maybe a robot that can
pass the coffee cup test, which is the notion that you would be given a random kitchen somewhere
in the country and you would have to go make a cup of coffee in that kitchen.
So flexibility would be the ability to deal with unknown unknowns.
So things that could not, dimensions of variability could not have been possibly foreseen by the
creators of the system within one specific task.
So generalizing to the long tail of situations in cell driving, for instance, would be flexibility.
So you have robustness, flexibility, and finally you would have extreme generalization, which
is basically flexibility, but instead of just considering one specific domain like driving
or domestic robotics, you're considering an open-ended range of possible domains.
So a robot would be capable of extreme generalization if, let's say, it's designed and trained for
cooking, for instance.
And if I buy the robot and if it's able to teach itself gardening in a couple of weeks,
it would be capable of extreme generalization, for instance.
So the ultimate goal is extreme generalization.
So creating a system that is so general that it could essentially achieve a human skill
parity of arbitrary tasks and arbitrary domains with the same level of improvisation and adaptation
power as humans when it encounters new situations.
And it would do so over basically the same range of possible domains and tasks as humans
and using essentially the same amount of training experience of practice as humans would require.
That would be human level, extreme generalization.
So I don't actually think humans are anywhere near the optimal intelligence bound if there
is such a thing.
So I think for humans or in general?
In general.
I think it's quite likely that there is a hard limit to how intelligent any system can
be.
But at the same time, I don't think humans are anywhere near that limit.
Yeah, last time, I think we talked, I think you had this idea that we're only as intelligent
as the problems we face, sort of intelligence that we are bounded by the problem.
In a way, yes, we are bounded by our environments and we are bounded by the problems we try
to solve.
Yeah.
Yeah.
What do you make of Neuralink and outsourcing some of the brain power, like brain computer
interfaces?
Do you think we can expand our, augment our intelligence?
I am fairly skeptical of Neuralink interfaces because they're trying to fix one specific
bottleneck in human mission cognition, which is the bandwidth bottleneck input and output
of information in the brain.
And my perception of the problem is that bandwidth is not, at this time, a bottleneck
at all, meaning that we already have sensors that enable us to take in far more information
than what we can actually process.
Well, to push back on that a little bit, to sort of play devil's advocate a little bit,
is if you look at the internet, Wikipedia, let's say Wikipedia, I would say that humans,
under the advent of Wikipedia, are much more intelligent than...
Yes.
I think that's a good one.
But that's also not about, that's about externalizing our intelligence via information
processing systems.
The external information processing system, which is very different from brain computer
interfaces.
Right.
But the question is whether if we have direct access, if our brain has direct access to
Wikipedia without having...
Your brain already has direct access to Wikipedia.
It's on your phone, and you have your hands and your eyes and your ears and so on to access
that information.
And the speed at which you can access it...
Is bottlenecked by the cognition.
I think it's already fairly close to optimal, which is why speed reading, for instance,
does not work.
The faster you read, the less you understand.
But maybe it's because it uses the eyes.
So maybe...
I don't believe so.
I think the brain is very slow.
It's...
It's the fastest things that happen in the brain at the level of 50 milliseconds.
Forming a conscious thought can potentially take entire seconds.
Right.
And you can already read pretty fast.
So I think the speed at which you can take information in, and even the speed at which
you can output information, can only be very incrementally improved.
Maybe there's a point.
I think that if you're a very, very fast typer, if you're a very trained typer, the speed
at which you can express your thoughts is already a speed at which you can form your thoughts.
Right.
So that's kind of an idea that there are fundamental bottlenecks to the human mind.
But it's possible that everything we have in the human mind is just to be able to survive
in the environment, and there's a lot more to expand.
Maybe you said the speed of the thought.
So I think augmenting human intelligence is a very valid and very powerful avenue, right?
And that's what computers are about.
In fact, that's what all of culture and civilization is about.
Your culture is externalized cognition, and we rely on culture to think constantly.
Yeah.
I mean, that's another way.
Yeah.
Not just computers, not just phones and the internet.
I mean, all of culture, like language, for instance, is a form of externalized cognition.
Books are obviously externalized cognition.
Yeah.
That's great.
And you can scale that externalized cognition far beyond the capability of the human brain.
And you could see civilization itself is it has capabilities that are far beyond any
individual brain, and we keep scaling it because it's not rebound by individual brains.
It's a different kind of system.
Yeah.
And that system includes non-human, non-humans, first of all, includes all the other biological
systems, which are probably contributing to the overall intelligence of the organism,
and then the computers are part of it.
Non-human systems probably not contributing much, but AIs are definitely contributing
to that.
Like Google search, for instance, is a big part of it.
Yeah.
Yeah.
A huge part.
A part that we can't probably introspect.
Like how the world has changed in the past 20 years, it's probably very difficult for
us to be able to understand until, of course, whoever created the simulation wherein is
probably doing metrics measuring the progress, there was probably a big spike in performance.
They're enjoying this.
So what are your thoughts on the Turing test and the Lobner Prize, which is one of the
most famous attempts at the test of artificial intelligence by doing a natural language open
dialogue test that's judged by humans as far as how well the machine did?
So I'm not a fan of the Turing test itself or any of its variants for two reasons.
So first of all, it's really coping out of trying to define and measure intelligence
because it's entirely outsourcing that to a panel of human judges.
And these human judges, they may not themselves have any proper methodology.
They may not themselves have any proper definition of intelligence.
They may not be reliable.
So the Turing test is already failing one of the core psychometrics principles, which
is reliability because you have biased human judges.
It's also violating the standardization requirement and the freedom from bias requirement.
And so it's really a copeout because you are outsourcing everything that matters, which
is precisely describing intelligence and finding a standard on test to measure it.
You're outsourcing everything to people.
So it's really a copeout.
And by the way, we should keep in mind that when Turing proposed the imitation game, it
was not meaning for the imitation game to be an actual goal for the field of AI, an actual
test of intelligence.
It was using the imitation game as a thought experiment in a philosophical discussion in
his 1950 paper.
It was trying to argue that theoretically, it should be possible for something very much
like the human mind, indistinguishable from the human mind to be encoded in a Turing
machine.
And at the time, that was a very daring idea.
It was stretching credulity.
But nowadays, I think it's fairly well accepted that the mind is an information processing
system and that you could probably encode it into a computer.
So another reason why I'm not a fan of this type of test is that the incentives that it
creates are incentives that are not conducive to proper scientific research.
If your goal is to convince a panel of human judges that they're talking to a human, then
you have an incentive to rely on tricks and prestidigitation in the same way that, let's
say, you're doing physics and you want to solve teleportation.
And what if the test that you set out to pass is you need to convince a panel of judges
that teleportation took place and they're just sitting there and watching what you're
doing?
And that is something that you can achieve with David Copperfield could achieve it in
his show at Vegas.
And what he's doing is very elaborate, but it's not actually physics.
It's not making any progress in our understanding of the universe.
To push back on that as possible, that's the hope with these kinds of subjective evaluations
is that it's easier to solve it generally than it is to come up with tricks that convince
a large number of judges.
In practice, when it turns out that it is very easy to deceive people in the same way
that you can do magic in Vegas, you can actually very easily convince people that they're talking
to a human when they're actually talking to an algorithm.
I just disagree with that.
I think it's easy, I would push, it's not easy, it's doable.
It's very easy because-
I wouldn't say it's very easy though.
We are biased, like we have theory of mind.
We are constantly projecting emotions, intentions, agent-ness.
Agent-ness is one of our core innate priors, right?
We are projecting these things on everything around us.
If you paint a smiley on a rock, the rock becomes happy in our eyes.
And because we have this extreme bias that permits everything we see around us, it's
actually pretty easy to trick people.
I agree with that, I so totally disagree with that.
You brilliantly put the anthropomorphization that we naturally do, the agent-ness of that
word.
Is that a real word?
No, it's not a real word.
I like it.
But it's a useful word.
It's a useful word.
Let's make it real.
It's a huge help.
But I still think it's really difficult to convince, if you do like the Alexa Prize formulation
where you talk for an hour, there's formulations of the test you can create where it's very
difficult.
I like the Alexa Prize better because it's more pragmatic, it's more practical.
It's actually incentivizing developers to create something that's useful as a human
mission interface.
So that's slightly better than just the imitation.
Your ideas like a test would hopefully help us in creating intelligent systems as a result.
If you create a system that passes it, it'll be useful for creating further intelligent
systems.
Yes, at least.
Yeah.
I mean, just to kind of comment, I'm a little bit surprised how little inspiration people
draw from the Turing test today.
The media and the popular press might write about it every once in a while, the philosophers
might talk about it, but most engineers are not really inspired by it.
And I know you don't like the Turing test, but we'll have this argument another time.
There's something inspiring about it, I think.
As a philosophical device in a philosophical discussion, I think there is something very
interesting about it.
I don't think it is in practical terms, I don't think it's conducive to progress.
And one of the reasons why is that I think being very human-like, being undistinguishable
from a human is actually the very last step in the creation of machine intelligence.
That the first ARs that will show strong generalization that will actually implement human-like broad
cognitive abilities, they will not actually be able to look anything like humans.
Human-likeness is the very last step in that process.
And so a good test is a test that points you towards the first step on the ladder, not
towards the top of the ladder.
To push back on that, I usually agree with you on most things.
I remember you at some point tweeting something about the Turing test not being counterproductive
or something like that.
And I think a lot of very smart people agree with that, a computation speaking not very
smart person disagree with that, because I think there's some magic to the interactivity
with other humans.
So to play devil's advocate on your statement, it's possible that in order to demonstrate
the generalization abilities of a system, you have to show your ability, in conversation
show your ability to adjust, adapt to the conversation through not just as a standalone
system, but through the process of the interaction, the game theoretic, where you really are changing
the environment by your actions.
So in the ARC challenge, for example, you're an observer, you can't scare the test into
changing.
You can't talk to the test.
You can't play with it.
So there's some aspect of that interactivity that becomes highly subjective, but it feels
like it could be conducive to generalizability.
I think you make a great point, the interactivity is a very good setting to force a system to
show adaptation, to show generalization.
That said, you're at the same time, it's not something very scalable because you rely
on human judges.
It's not something reliable because the human judges may not do.
You don't like human judges.
Basically, yes.
And I think so.
I love the idea of interactivity.
I initially wanted an ARC test that had some amount of interactivity where your score on
a task would not be one or zero if you can solve it or not, but would be the number of
attempts that you can make before you hit the right solution, which means that now you
can start applying the scientific method as you solve ARC tasks that you can start formulating
hypothesis and probing the system to see whether the observation would match the hypothesis
or not.
It would be amazing if you could also, even higher level than that, measure the quality
of your attempts, which of course is impossible, but again, that gets subjective.
How good was your thinking?
How efficient was?
So one thing that's interesting about this notion of scoring you as how many attempts
you need is that you can start producing tasks that are way more ambiguous, right?
Because with the different attempts, you can actually probe that ambiguity, right?
So that's in a sense, which is how good can you adapt to the uncertainty and reduce the
uncertainty?
Yes, it's half fast.
Is the efficiency with which you reduce uncertainty in program space, exactly.
Very difficult to come up with that kind of test, though.
Yeah, so I would love to be able to create something like this.
In practice, it would be very, very difficult, but yes.
I mean, what you're doing, what you've done with the ARC challenge is brilliant.
I'm also not, I'm surprised that it's not more popular, but I think it's picking up
like it.
It does its niche.
Yeah.
What are your thoughts about another test that I talked with Marcus Hutter?
He has the Hutter Prize for Compression of Human Knowledge, and the idea is really sort
of quantify and reduce the test of intelligence purely to just the ability to compress.
What's your thoughts about this intelligence as compression?
I mean, it's a very fun test because it's such a simple idea, like you're given Wikipedia,
basically English Wikipedia, and you must compress it.
And so it stems from the idea that cognition is compression, that the brain is basically
a compression algorithm.
This is a very old idea, it's a very, I think, striking and beautiful idea.
I used to believe it.
I eventually had to realize that it was, it was very much a flawed idea.
So I no longer believe that it's compression, that cognition is compression.
But I can tell you what's the difference.
So it's very easy to believe that cognition and compression are the same thing because,
so Jeff Hawkins, for instance, says that cognition is prediction.
And of course, prediction is basically the same thing as compression, right?
It's just including the temporal axis.
And it's very easy to believe this because compression is something that we do all the
time very naturally.
We are constantly compressing information.
We are constantly trying, we have this bias towards simplicity.
We're constantly trying to organize things in our mind and around us to be more regular.
So it's a beautiful idea, it's very easy to believe.
There is a big difference between what we do with our brains and compression.
Compression is actually kind of a tool in the human cognitive tool kit that is used
in many ways.
But it's just a tool, it is a tool for cognition, it is an art cognition itself.
And the big fundamental difference is that cognition is about being able to operate in
future situations that include fundamental uncertainty and novelty.
So for instance, consider a child at age 10.
And so they have 10 years of life experience, they've gotten pain, pleasure, rewards and
punishment in a period of time.
If you were to generate the shortest behavioral program that would have basically run that
child over these 10 years in an optimal way, the shortest optimal behavioral program given
the experience of that child so far, well, that program, that compressed program, this
is what you would get if the mind of the child was a compression algorithm essentially, would
be utterly unable, inappropriate to process the next 70 years in the life of that child.
So in the models we build of the world, we are not trying to make them actually optimally
compressed.
We are using compression as a tool to promote simplicity and efficiency in our models.
But they are not perfectly compressed because they need to include things that are seemingly
useless today, that have seemingly been useless so far.
But that may turn out to be useful in the future because you just don't know the future.
And that's the fundamental principle that cognition, that intelligence arises from is
that you need to be able to run appropriate behavioral programs except you have absolutely
no idea what sort of context, environment, situation they're going to be running in.
And you have to deal with that uncertainty, with that future novelty.
So an analogy that you can make is with investing, for instance.
If I look at the past 20 years of stock market data, and I use a compression algorithm to
figure out the best trading strategy, it's going to be you buy Apple stock, then maybe
the past few years you buy Tesla stock or something.
But is that strategy still going to be true for the next 20 years?
Well, actually, probably not, which is why if you're a smart investor, you're not just
going to be following the strategy that corresponds to compression of the past, you're going to
be following, you're going to have a balanced spot for you, because you just don't know
what's going to happen.
We're also going to be able to anticipate totally new things.
I mean, I guess in that same sense, the compression is analogous to what you talked about, which
is like local or robust generalization versus extreme generalization.
It's much closer to that side of being able to generalize in a local sense.
That's why as humans, when we are children, in our education, so a lot of it is driven
by place, driven by curiosity, we are not efficiently compressing things.
We're actually exploring.
We are retaining all kinds of things from our environment that seem to be completely useless,
because they might turn out to be eventually useful.
That's what cognition is really about, and what makes it antagonistic to compression
is that it is about hedging for future uncertainty.
Officially.
That's antagonistic to compression.
Yes.
Officially hedging.
Cognition leverages compression as a tool to promote efficiency and simplicity in our
models.
Just as I said, make it simpler, but not however that code goes, but not too simple.
You want to compression simplifies things, but you don't want to make it too simple.
Yes.
A good model of the world is going to include all kinds of things that are completely useless,
actually, just in case.
Because you need diversity in the same way that in your portfolio, you need all kinds
of stocks that may not have performed well so far, but you need diversity.
The reason you need diversity, because fundamentally, you don't know what you're doing, and the
same is true of the human mind, is that it needs to behave appropriately in the future,
and it has no idea what the future is going to be like.
But it's not going to be like the past, so compressing the past is not appropriate because
the past is not proactive of the future.
Yes.
History repeats itself, but not perfectly.
I don't think I asked you last time the most inappropriately absurd question.
We've talked a lot about intelligence, but the bigger question from intelligence is of
meaning.
Intelligent systems are kind of goal-oriented.
They're always optimizing for goal.
You look at the Hutter Prize, actually, there's always a clean formulation of a goal, but
the natural questions for us humans, since we don't know our objective function, is
what is the meaning of it all.
The absurd question is, what Francois Chollet, do you think is the meaning of life?
What's the meaning of life?
Yes, that's a big question.
I think I can give you my answer, or at least one of my answers.
One thing that's very important in understanding who we are is that everything that makes up
ourselves, that makes up who we are, even your most personal thoughts, is not actually
your own.
Even your most personal thoughts are expressed in words that you did not invent and are built
on concepts and images that you did not invent.
We are very much cultural beings, we are made of culture.
What makes us different from animals, for instance.
Everything about ourselves is an echo of the past, an echo of people who lived before
us.
That's who we are.
In the same way, if we manage to contribute something to the collective edifice of culture,
a new idea, maybe a beautiful piece of music, a work of art, a grand theory, a new words,
maybe, that something is going to become a part of the minds of future humans, essentially
forever.
So everything we do creates repulse that propagates into the future.
And that's in a way, this is our path to immortality, is that as we contribute things
to culture, culture in turn becomes future humans.
And we keep influencing people thousands of years from now.
So our actions today create repulse and these repulse, I think, basically sum up the meaning
of life.
Like in the same way that we are, the sum of the interactions between many different
repulse that came from our past, we are ourselves creating repulse that will propagate into
the future.
And that's why we should be, this seems like perhaps an old thing to say, but we should
be kind to others during our time on Earth because every act of kindness creates repulse
and in reverse, every act of violence also creates repulse.
And you want to carefully choose which kind of repulse you want to create and you want
to propagate into the future.
And in your case, first of all, beautifully put, but in your case, creating repulse into
the future human and future AGI systems.
Yes.
It's fascinating.
Our successors.
I don't think there's a better way to end it, Francois, as always for a second time
and I'm sure many times in the future, it's been a huge honor.
You're one of the most brilliant people in the machine learning, computer science, science
world.
Again, it's a huge honor.
Thanks for talking today.
It's been a pleasure, thanks a lot for having me.
We appreciate it.
Thanks for listening to this conversation with friends, wash away and thank you to our sponsors
Babel, masterclass and cash app.
Click the sponsor links in the description to get a discount and to support this podcast.
If you enjoy this thing, subscribe on YouTube, review it with 5,000 up a podcast, follow
on Spotify, support on Patreon or connect with me on Twitter at Lex Friedman.
And now let me leave you with some words from Renee Descartes in 1668, an excerpt of which
Francois includes in his On the Measure of Intelligence paper.
If there were machines which bore a resemblance to our bodies and imitated our actions as
closely as possible for all practical purposes, we should still have two very certain means
of recognizing that they were not real men.
The first is that they could never use words or put together signs as we do in order to
declare our thoughts to others.
For we can certainly conceive of a machine so constructed that it utters words and even
utters words that correspond to bodily actions causing a change in its organs.
But it is not conceivable that such a machine should produce different arrangements of words
so as to give an appropriately meaningful answer to whatever is said in its presence
as the dullest of men can do.
Here Descartes is anticipating the Turing test and the argument still continues to this
day.
Secondly, he continues, even though some machines might do some things as well as we
do them, or perhaps even better, they would inevitably fail in others, which would reveal
that they are acting not from understanding, but only from the disposition of their organs.
This is an incredible quote, for whereas reason is a universal instrument which can be used
in all kinds of situations.
These organs need some particular action, hence it is for all practical purposes impossible
for a machine to have enough different organs to make it act in all the contingencies of
life in the way in which our reason makes us act.
That's the debate between mimicry memorization versus understanding.
So thank you for listening, and hope to see you next time.