Transcription
What we see today in LM is cleverness.
This is not limitation. This is today's
machine. We will be the pets of the new
species that we are generating today.
>> That's pets. P a playful agent. It's
very dangerous. We thought it could be
trillion times more powerful than us.
>> It could be the end of the human race.
>> It could be. Yeah. This illusion of free
will is guides our life. Without it, we
wouldn't be a human being. Machine will
make us feel that we are making
decision.
>> It means that we are not free.
>> That's right.
>> Judea Pearl is one of the unsung fathers
of artificial intelligence and a laurate
of the touring award.
>> Judea Pearl.
>> He taught computers to ask why, not just
tally how many or what. His ideas help
doctors, economists, and engineers
uncover causes rather than mere
coincidences. In this conversation, he
explains why he believes his work is the
missing link on the path to building a
conscious machine.
>> The studies of the causal inference is a
missing link.
>> It's a great honor having you here. You
divide thinking into three levels,
seeing patterns, making changes and
imaging what if AI understand only
patterns.
>> No. No.
What you mean statistics or AI that is
built on statistics only like uh deep
learning?
Yeah, deep learning is an example of a
sophisticated AI system built on
statistics
going from sample to distribution to
expectation but it doesn't give you the
understanding of the world. So AI uh
certain segments of AI that have become
very popular are pure statistics and
have no causal component to them but not
the entirety of AI. In my field since
1988
I would say life was causality.
It wasn't
divorced of causality. Everything was
causality
>> for you. It's only statistical machines
or something more.
>> Well, there are several groups in AI
that are doing different things and they
are basing their research on different
paradigms.
certain machine learning the way it was
developed was
um
statistics on steroid
but not all AI there is an AI done and
people are doing um pattern recognition
I'm sorry but people are doing logic
people are doing constraint satisfaction
people are doing uh uh reasoning,
agents, agent reasoning, multi- aent and
all these are different kind of
paradigms.
So that's not the most popular has been
machine learning and that's a statistic
on steroid but not all AI belongs to
that paradigm. So I I my paradigm has
been totally different. If you include
me in AI,
I have divorced statistics already in
the 19 when was it?
1980. The moment the Beijian network
were invented.
>> Why? We noticed something very strange
that whenever
expert use Beijian network
they always assign we always
define the parents to be the causes and
the uh child to be the effect.
What do I mean by causes and effect?
what we normal language will designate
as cause and effect. But you see in
Beijian network it's a statistical
device. So it doesn't distinguish with
cause and effect. You can go from put
your effects as your parents and the
cause to be the child. Patient network
will not discriminate. They don't force
you to go from causes to effect but we
happen to do it expert constantly and
people have noticed that in various area
I think David Hecerman was the first to
discover it. How come we always go from
parents to from causes to effect and
never reconfigure the Beijian network to
go from effects to cause?
Asking that question made you think that
there is something about cause and
effect which is not captured by the
statistics
and that is the idea of invariance.
the idea of um intervention and the
counterfactual. So that took off because
of the
um observation
that human being tend to describe the
world in terms of interaction between
coders and effect and it means something
to them different than just statistical
relationship
statistical correlation.
Do you believe that
we we can have a close consciousness
inside AI black box, LLM black box or
that future AGI black box?
>> In the future or AGI we we are going to
have
conscious machines
have very clear idea of their software.
>> That's what we can expect conscious
machine.
>> Absolutely. Yes, with individuality,
personality
variations from one machine to other
depending on what
depending on the experience that the
machine has went through in its
lifetime. They will have different
personalities
depending on the instruction get giving
them from other machines.
Yeah, they'll have different
personality. They have different
understanding of themsel in relation to
their environment and they will be able
to talk like you and I as if we have
consciousness. So we won't be able to
tell the difference between a conscious
having machine and machine like they are
today. We do not have conscious.
>> Your theory
is
missing link to create conscious
machine. Absolutely.
The studies of the causal inference is a
missing link in in
machine learning without which you would
not have a model not only of yourself. A
model of the world.
Machines today do not have a model.
Before the advent of causal inference,
machines did not have a model of the
world.
machine had
condensed and parsimmonious model of the
statistics
of statistical data that came from the
world but not of the how the world
works. We didn't have something that we
call narrative. What is a narrative?
People fight about narrative not about
the data. They agree on data. Okay. But
what about what is a narrative? They
have different name for that world model
and they have a world of
>> that's true scientists for decades are
focused on fetting AI more and more data
>> that's
but and believing that more data will
make them more intelligent right that's
that's a fallacy because you have to get
to the narrative level machine must have
a paradigm of other world and I'm
switching because even scientists switch
from one terminology to another
kun talked about paradigm paradigm
change okay other people in in the
social science they talk about narrative
okay what do the Israel and Palestinian
fight about the narrative not about the
data so the idea of having a a narrative
of a a narrative of it to explain the
world.
That idea
was denied mathematical or formal
representation.
We couldn't talk about it. We couldn't
put it in a machine. Now we have it.
It's called the structural causal model
which is built on only on one notion.
Every variable listens to all the
others. every variable in the universe
is listening to all the other listening.
And if the um barometer listens to the
atmospheric pressure, it's not true that
the atmospheric pressure necessarily
listens to the barometer. Okay? It's
unidirectional.
Science has been imprisoned
by the equality sign which is symmetric.
If f is equal to m a then a is equal to
f / m.
Speaking about Newton law, right? You
cannot distinguish between the two. But
you and I know that the acceleration is
caused by the force and not the other
way around. That the force is caused by
the acceleration. No,
science has been deprived of a an
important
symbol. A symbol that will distinguish
between
cause and effect and effect and cause
directionality. Now we have it. So I
feel great. We can do so much more
today. We have a paradigm for how things
work.
Once you said that that maybe free will
doesn't exist.
>> Correct.
>> But we made it for intelligent.
It's necessary for human intelligent.
>> Not for intelligence.
>> What does it mean? It it's necessary for
everything we do for moral behavior
for
improving our performance.
This illusion of free will
guides our life. Without it, we wouldn't
be a human being. It per permeates every
piece of knowledge.
>> Could you explain?
>> Yeah. illusion of free will because of
>> biology. We are biological machines.
>> We are biological machine. Not only
that, I can say we are deterministic
biological machines. Okay. If we remove
quantum mechanics from the discussion.
So the every action we taking is
determined by neural activation
of our uh of the neurons in our brain
which is determined by previous
excitation of other neurons in our brain
and the uh input that we get. Some of it
comes from cosmic radiation. Some of it
comes from uh what we are being told
>> all factors around
>> all factors around. So what whether or
not I'm touching my nose or I don't
touch my nose is not my option. It's
determined by the inputs that my finger
received and that determined by the
input that the brain has received in the
past. Okay. So I have no option and I
have a very vivid sensation that I do
have the option and you have the option
of touching or not touching you know so
this is the problem of free will
philosophical problem is called the
scandal of philosophy because they
haven't done any progress since
Aristotle till today
but when we come to computer
it a different it's a very even the
strongest scand in the sense that when I
talk about computer you cannot claim
that like the human brain is guided by
divine forces or you cannot claim that
it is the
controlled by quantum mechanical
noise or uncertainty no we have a finite
state deterministic
computers. Does it or doesn't it have
free will? Does it have the sensation of
having options?
On one hand, it doesn't because
everything that every step of
calculation in the machine is determined
by the initial program that you wrote by
initial code every step
>> in this meaning future is determined by
the past. Yeah. So if you know the path
and in in the computer case, all you
need to know is the code, the code and
the input, right? You know the tool and
you determine the behavior of the
machine for the next for the future.
>> It means that we are not free.
>> That's right. On the other hand,
if the machine acts like us, answer
question like us, communicate with each
other as if they have option and one
machine tells the other you shouldn't
have done it. What does it mean you
shouldn't? You problem me, right? So
there is a here in interplay between
implicit knowledge and explicit
knowledge. implicit knowledge.
A chess playing machine has no options.
They just follow the rule of the game
and the rule of the program.
At the same time, the explicit knowledge
in chess playing machine involves
uristics, involves alphabeta search,
involves all kind of things that we are
you and I communicate with. the goodness
of a board position whether or not yeah
these are explicit knowledge.
How good is this board position relative
to the other? We just pass it through a
fourstep look ahead and get the results.
So this is explicit knowledge. In terms
of the explicit knowledge, yes, we have
options and yes, we have choices
and we can teach each other and the
machine can learn. What does it mean to
learn? Let's look at the chess learning
machine. Machine that learn to play
better chess by playing with themselves,
right? What do they do? They transmit
information from explicit level to from
implicit level as the rule of the game
and explicit.
What what table do I have?
What is the function that determines the
strength of the board? Okay, this is the
explicit explicit notion related
somewhat to um Janimani
system one and system two. Okay. But now
in talk in terms of the explicit level
of knowledge, we do have options and we
do have free will.
What does consciousness differ from free
will?
>> Your consciousness is made.
>> Consciousness is a blueprint of its own
software.
>> Absolutely. You got it. Did I say it?
Yes, I did say it someplace.
Consciousness is nothing else but a
blueprint, a crude blueprint of our
software
and we cannot have a total print of our
software because that will violate a
tooling halting problem. We don't know
whether our program will terminate. So
how can we determine what predict the
results of our software? No, but we have
a partial knowledge of our software. And
this partial knowledge is a good um good
enough for certain task
not good for others
like like the halting problem but it's
good enough to
take aspirin to uh take
to act in the world to cross the street
to go to a doctor to to decide which
doctor you should have. So it's good
enough for everyday tasks. It may not.
Yeah, I I say it already. So that is
consciousness having this feeling that
this is mine. This is my software. I
have the option to touch my nose. I have
I can or cannot do a certain task. This
is part of consciousness.
I know that before I even try, you give
me a problem in math and I can tell you
whether I have the capability of solving
it without even knowing, without even
trying. If you give me two equations
with two unknown, I know I can solve it.
Okay? So having a representation of my
capabilities, that's part of
consciousness. It's part of my software
and so on and so on. Having
a model of yourself this is
consciousness
and we don't have to go to more than
that to to neural science we don't have
to go to
quantum mechanic and all this
>> to be to be honest I don't feel
>> that different between free will and
consciousness
>> imagine an agent
which has never been asked to take any
action. Free will has to do with action.
Yeah. I I feel like I want to do
something. I feel I can do something. I
I feel I have an option. Assuming you
are just passive observer of the world.
You can still have consciousness
without any action, without agency. Just
observe and say the stars are going this
way and I'm going a different way. I am
the idea of assigning a symbol to
yourself and saying this is me.
This is my hand.
>> How we can make AI machine that
understand how effect look like?
Just look at the
every system which is based on causal
calculus and you find an AI machine that
that is that understand cause and
effect. And what do I mean by understand
code and effect? I go even further. What
do I mean? It understand how things work
or it understand a domain
or a topic. Understanding means being
able to answer questions on all three
levels of the causal ladder.
>> But we know that something is missing
for me. That that page is the web of
associations for AI is a simple vector.
Who is AI again? What system you're
talking about? And who is you? Are you
statistician or are you a human being?
If you are human being, you want to
understand things. It's not enough to
say what I observe here is a pattern of
associations among pixels in my eyes.
No, that's not the way we talk and
that's not how we we communicate and
understand things. No, we talk very
clearly in terms of
what if uh what if Israel did not attack
Iran
nuclear facility or what if
Trump
did not give the okay for Israel to do
that. We ask what if not constantly
any articles in the New York Times that
you about politics or about healthcare
is laden with whatif questions is not
laden with pattern of associations.
People don't even
commu people not do not understand
associations. They make mistakes when
you talk about associations. No, we are
not association machines. We are cause
and effect machines.
>> So how we should call that large
language models?
>> Oh, that's a different one. The large
language models
are statistical machine built on text
generated by from human being who are
causal machines. Okay, you and I are
qual machine. We write articles. We load
the articles on the internet and this
becomes the training set for the large
language model. Large me language model
take the written text again written by
causal machine. These are human beings
that wrote those days. Okay. And
interpolate them statistically.
So sometimes they appear to do causal
reasoning because the authors of those
texts are causal machines. We
that's how you should look at it. But if
you exercise
large language model on data generated
not by human being but by nature.
All you have is statistics
on steroid.
>> Oh, one model from open AI,
abstractional thinking, free reasoning.
What does it mean for you?
>> In just words, not mean anything until
you tell me what question you can answer
by each one of those virtues.
Abstraction model. Tell me what you can
answer. If you have abstraction model
versus not having abstraction model.
Give me any virtue in in any text or in
any reasoning discussion.
This define the question that you can
answer with it as opposed to without it.
>> What makes us human?
>> Compared with what? With non-human.
compared with animals
>> comparing to the AI.
>> Well, we don't know AI yet. AI has not
finished its
course yet. We still
>> comparing to that systems
>> to L&M.
>> Yeah.
>> Or comparing to the next generation AI
which may be general artificial general
intelligence which may be different in
architecture in in capability
relative to what we see today. So what
AI?
>> Okay. So uh for me today AI
means the huge the biggest large
language models
>> created by the biggest companies from
Google Gemini
through open AI uh meta and companies
like that. So we'll talk about the GL
later. uh when we look at LLM that all
interfaces around the world what makes
us human
comparating to that LLM systems
abstractional thinking our abstractional
thinking
>> we have a model of the world
>> physical right understanding of of the
physical law
>> not necessarily just the quality
qualitative nature what the qualitative
um nature of how things work in the
world a child doesn't have the maxwell
equation right
>> you mean dimensions of the reality
around
>> the fact that if I turn the switch on
the light will turn on uh if I press
this button then the television will
start working I know how to operate I'm
I'm looking at the child a 10year-old
child can do many things here that a
robot find very hard to do or very hard
to learn. And a child can do it by learn
by observing father or mother doing
them, imitating them very quickly
and the ability to emulate
your
teach your father is required as a model
of the world. Look for instance the
ability to emulate the invention of a
bow and an arrow.
>> Y
>> that is it. It takes imitation but it
also takes
um an additional component to understand
what makes um the string straight. What
makes what bends the branch of the tree
and
ability to pass it on to the next
generation with very few instruction.
Just watch me and you can tell it to
your children and we have a language to
do that. Even the primitive people had a
very good spoken language to transmit
that skill to next generation. That what
made us human care and we are to a large
extent we are born with that
transmittance language.
Open AI is ignoring you. Google is
ignoring you and your study. Meta is
ignoring you your study.
>> Are they ignoring
really but because they don't what
advertise it? No, that's my question to
you.
>> No, I don't feel like I'm being ignored.
I feel like they are doing their spill
and in time they recognize that they're
missing a component and they will come
and wise up.
I remember uh few few weeks ago I talked
to Jeffrey Hinton.
>> Yeah. uh so-called godfather of of AI
Nobel this year Nobel prize winner um
and what he said that human will be
second not the first intelligent of
earth do you agree with that
>> it's second in terms of capability
>> yeah I believe so I believe so that if
you let the
general AI and develop Eventually
we are going to have a computers that
would have all the capabilities of a
human being but much more powerful
capabilities to
consciousness and to have free will and
to have a model of the world and but
it's very dangerous because it will be
trillion times more powerful than us. So
we'll be second in terms of
capabilities. Yeah,
>> it could be the end of the human race.
>> It could be. Yes. Well, that's what the
chicken
uh thought when they saw human being
coming. We going to be second and they
are second. Chicken and dogs and cats
are pets of us. We control the world
today. And eventually it's quite capable
that we will be the pets of the new
species that we are generating today.
Bats.
>> Pets. P E T. Pets.
>> Animals.
>> Animals. No. Play a dog.
>> A playful agents. Okay.
That our destiny will be determined by
the will and I use the word will and
purpose of the machine.
Yeah. The machine will make us feel feel
that we are making decision. But those
decision will be tuned to the desires
of the machine. So it will make us feel
that it's good for us to serve the
master
like dogs are being taught today. It's
good for them to be playful and to
behave in a certain way to pacify their
master. So yeah, we'll be the pets of
some other species. Yeah, I think it's
it's we don't want today. It looks scary
and we have to protect ourselves against
that. If we really do not want it, we
might enjoy it, you know.
>> Is it too late to change something?
>> Yeah, I'm not too late. No, it's not too
late, but we don't know how to do it. We
do not have the capability
to prevent it from happening, to control
it, to control its development. It's
scary. We do not know how. And anybody
who tells you a here's a scheme to turn
the plug off,
it's I do not believe that we have the
capability of controlling this
uncontrolled species. Is any question
about AI development
uh data system develop development
uh that keeps you awake at night?
It doesn't keep me awake at night for
fear. It picks me awake at night
by curiosity.
I'm driven by one um desire. I want to
understand myself. It's a nice puzzle
and that's why I work in AI. That is a
scientific curiosity uh game
and so I'm aware of night because
puzzles come to my mind that I cannot
solve and I can see maybe if I think
about it I'll solve them. So but not
fear. Why not fear? Because I know that
I don't have the capability of stopping
the way the path is going.
>> Nobody has.
>> Nobody has. Nobody has except who knows
maybe
if government regulation becomes really
dictatorial
like in North Korea and they stop all
research on AI
that it's the only way we can stop it
today. So I know I don't believe in
regulation that regulation will do it.
>> What do you think what we can expect
from AI development in the next 5 years?
>> We are going to understand ourselves
better
and that's for me is a great excitement
>> on leaders AGI
>> toward toward AGI. Yes. Like like free
will for instance, we'll know what it
takes for the mind to declare that this
action was deliberate and that action
was spontaneous. It was where does it
register? What kind of computational
resources are being exercised to enable
us to say I did it out of my free will.
It was deliberate action and that was
not that was just a either spontaneous
or it was thoughtless reaction.
I remember I talked to Roger Penrose.
>> Yes.
>> Uh and he proposed to call AI not
artificial intelligent but artificial
cleverness. That the cleverness is a
better word to explain what we are
doing.
through that what we see today in LLM is
cleverness. You go and combine things
that other people written in a nice way
and you present it as if it was product
of your thought. The clever okay but
this is not limitation. This is today's
machine. There is no basic impediment to
having intelligent machine. He sees
impediment. I don't see them right math
machine can be not only clever but
intelligent
and he didn't define the touring test
the distinguish between cleverness and
intelligent what is a touring test
at what point we'll say yes this machine
is really intelligent what
I I'm looking for a touring test do we
know more
about AI black box.
How does it look like internal
internal life deep inside that black
box?
>> You mean a general intelligence?
>> Yeah.
>> Here, here's one.
>> You mean that human brain is a black
box?
>> No, no, no. It's only a existence proof
that it exist that it it is realizable
but it it's not it doesn't tell you how
it works inside we don't know how it
works inside but I can envision very
clearly you have a model of the world
and you exercise it and you keep on
improving it
>> but how is it possible that we've
created something that we have no idea
how does it work inside.
>> We haven't created it yet, but I can
envision how it's going to be
architectured.
So, what what is the problem? Would you
like to me to tell you how the machine
going to work?
How is it possible that AI was created
by humans and blackbox exist that
questions exist? It's like a internal
life there inside. That's my question
about that that phenomena.
>> Why is it not possible that a species
like human being will create another
species that looks like it or that looks
even better more intelligent than it?
Why is it so puzzling?
Is it look I give you an example
elementary particle created us chemistry
created us chemistry is stupid things
right created intelligent machines like
us
and we are not puzzled
we live with that okay but when human
being create more powerful agents
we suddenly we become puzzled Why
can we expect huge scale of unemployment
because of AI development?
It's quite possible quite possible
human being will have other things to do
for instance to continue and explore
either the capabilities of themselves or
explore the capability of machines or
the future machines. And the question is
who is going to pay their salaries,
right?
Who would benefit from this exploration?
It's a really tough question.
>> If you could change one single thing,
what it would be regarding AI
development
>> of the NLM dimension?
>> Yeah.
>> Well, think about how you build a model
of the world.
That is a very it's a necessary
component that is missing from today's
machine.
>> How we can create a model of the world?
>> You and I have it and the machine can be
programmed to acquire one. We do have in
our in the child's software. The child
is born with the um
>> because I am biological machine that
process data collected by my own senses.
>> Correct. But you also have the capacity
to
um the framework in which to fill in the
missing part to create a model of the
world. You are born with the capacity to
create a model of the world. And now
depends on what environment you live in,
you'll get a different model. If you
live in the Amazon jungle, you'll feel
in a different
part differently than if you are born in
Sborn or born to a intelligent family in
Poland. But how we can digitalized
uh world
if we have no idea for example
how gravity was born.
We have no idea what connect quantum
world and classical world.
>> Are you really worried about it?
So many people in the world are worried
about the lack of general theory of
gravitation.
>> But it doesn't matter. It determine
everything around.
>> No, it doesn't bother me and it doesn't
bother the people who are going to work
in the morning and are living their life
happily and fulfilled.
Why?
Just the fact that some physicists are
bothered by this mathematical deficiency
bothers ordinary people. No, we can
still live fulfilled life without that.
>> It doesn't bother me.
>> When you look at the future, what's
around the corner?
I in
I look at at this
shortterm future. I'm interested for
instance to get a automated scientist.
It would be a nice thing for me to have
a automated scientist that can conduct
experiments, reason about them, see
what's missing, conduct, figure out
which next experiment to conduct and
what to conclude with it. I like to see
personalized medicine. That's something
which I really concrete and I would like
I think it's very close to being
realized
that that we don't reason in ter of
population of patients but we look at
the patient him or herself. Okay. and um
these kind of short-term advances that
are within reach in the next 10 years. I
don't think I'm going to be here more
than 35 years. I tell you why. But
so this is my horizon. In 35 years, I
would like to see personalized medicine
realized on a computer.
>> Do you remember the moment when you
received an touring award?
>> Yes. Yes.
when I was getting a phone call.
>> Yeah.
>> Yeah. I thought it was from some crack.
No, Irish, what do you call them? Irish.
I forgot the name. They had
sweep stakes.
>> I've gotten this call before. I didn't
believe it. So, I was surprised. Yes. I
was very happy to communicate with the
lady.
>> And the summary was for your incredible
input for AI development. Fundamental
>> fundamentations of AI. I forgot how they
phrase it primarily in um reasoning
about uncertainty
and reasoning about cause and effect and
a calculus for cause and effect.
>> Yeah. Tell me something about for
yourself the most important study,
the most important paper for you.
>> Mhm. Generated by my group.
>> Yeah. Generated by you and your group.
Yeah.
>> Yeah.
But looking in retrospect,
I think it was the agitimization of
counterfacturers.
The idea that we can
provide machine with the right
algorithms, the right assumptions to
answer counterfactual questions. the way
you and I answer them.
>> How this paper changed the rule have
changed the world.
>> Everything follows from that. Yes. You
take this paper, you have this algorithm
and everything follows from that from
interventions to
code of fact to
um credit and blame regret all these
notions follow from that. I don't have
to do anything. It follows from one
equation.
It contains the entire wisdom of causal
inference.
If you buy one equations, everything is
mathematics.
Straight mathematics.
>> Thank you for your time.
>> No, thank you for entertaining me.