Transcription
Biology can be an inspiration [music]
like our brains they have some
commonalities with language models and
they can be translated [music] into
ideas and AI.
>> Wukos Kaiser is one of the most
influential computer scientists in
history. The paper he co-authored
>> attention [music] is all you need
>> is the most important machine learning
publication in [music] decades.
>> Your paper changed the world. You change
[music] the world.
>> Our paper. Yes. It has also become one
of the most frequently cited scientific
papers of the 21st century. It triggered
a revolution in artificial [music]
intelligence leading to the development
of large language models and generative
AI.
>> These things [music] only operate on
text. How would they ever know about all
the other [music] aspects of life?
>> For the past 5 years, he has been one of
the key researchers at Open AI.
>> [music]
>> To me as a researcher it's a miracle
that these patterns that these machines
found in the language are so similar to
the ways humans think. Then [music] we
may have this question. Okay. So what
why does it even need us? Right? So that
was the that was the miracle of deep
learning.
>> Wukash honor having you here. Attention
is all you need. Wukash.
>> Yes. That that that's a paper we wrote.
And
>> your paper changed the world. You
changed the world.
>> Our paper. There's eight of us.
>> Um
>> and yes, it it is certainly a paper that
did change the AI industry in the world.
We certainly did not expect it to change
that much at the time when it was
written, but I guess that's how it goes.
You never know. Do you do you do you
agree that this paper
opened new era of AI development?
Yes, I think it did. Former certainly
brought it all up. And I I do think
though it's important to you know every
time I get asked this a lot to go back
in my mind to the time when we were
doing it. And you know at that time it
felt like well this is another technical
step that just should be researched and
done. It didn't feel very different from
many other technical steps. And then as
it as as we were seeing that it works
really well it got more and more
exciting. But even when it was
published, we did not think, oh, you
know,
you just take this, train it on all of
the internet and it's going to be I
mean, you know, you could think this as
as a hope or a dream,
but you I was not convinced that that
this will happen in a few years without
like adding other things. And I think
it's a big lesson to anything you do as
a researcher in in in the current world
that sometimes rarely but but sometimes
you know you change one thing that seems
fairly technical you know at that time
the move from RNN's with encoder decoder
attention to this non-recurrent
architecture with one attention it's
like okay it will be faster it gets
better translation for you but you don't
see that it may work really well and and
just
lead to a new wave of companies and
things like that. So I I think it's a
very very positive thing and that
working on something in research that
seems like just a technical improvement,
a good improvement but maybe not so big
actually may become something big just
because just because sometimes things
work better than you think. Often times
they work worse than you think. But
>> do we know how AI works?
>> Do we know how AI works? I I think we do
to a large extent. I mean, of course, we
there's a lot we don't know, but there
is a lot we do know. What we don't know.
Well, if you if you train a really big
model on a lot of data, all of the human
internet and this form or another, then
of course
what we do know like if you train a
small model on a very curated data set,
we can kind of figure out what are the
patterns that go in there. I mean, we
know every number inside, right? We can
look we can understand things like
induction heads there. There's a lot we
can understand about small models. when
[clears throat] it becomes a huge model
trained on huge amounts of data,
we don't understand why it answers
exactly this way. Like we can try to
trace it to parts of the data, but
there's just too much of it. The model's
huge. It has built a lot of
representations and associations and
like then we lose the understanding of
what's happened.
>> That's also another question about your
feeling about current stage of AI
development.
what's uh for you the current where we
are regarding to AI development today
what's the biggest challenge for you
what is the biggest question regarding
>> so I I think we will talk about this a
lot you know so to I I I try to think
about AI development
and as this path of making more and more
powerful models learnable this is like
my my model that I it's a very
simplified thing Of course, CA
development has many axis and and you do
many things.
So, so, so when I started, there were
RNNs. We made transformers which are a
more powerful model than RNNs and and
they're very learnable. And then I think
the big step recently was reasoning
models, which are even more powerful,
but they are a bit harder to train.
They're not trained with gradient
descent. We use reinforcement learning.
Maybe we'll talk more about it. So I
think we like reasoning models give you
almost ultimate power because you can
think for arbitrarily long almost if you
have the memory and computers to do
that.
So I think in this sense we have reached
an amazing stage but they they still
lack a little bit like for now they're
only trained on verifiable rewards. So
you need to curate the data. They're
very sequential still. It's a little bit
like RNN were very sequential. They
should definitely be more parallel. So
so there is a bunch of improvements we
need to do to this. But I do think this
paradigm will lead us to to models that
that generalize almost as good as
humans. So they will be very powerful.
They will be very trainable and and
these models will do amazing things that
we are starting. So reasonable models
are great in the verifiable domains like
mathematics and code and we see that
they're starting to do fairly amazing
things.
They start helping prove new
mathematical theorems. They write code
almost professionally. So we will see
these amazing things come to other
domains. Um, as I say, there's still
research problems to be solved. And this
with research, you never know if it's
one year or five years. It's always hard
to say. Um, but
but in these domains where where this
current paradigm works well, where where
we train at, we see astoundingly amazing
things.
>> [sighs]
>> So when this comes to all domains we
will have
another forms of breakthroughs at least
in science that that I think will be
very surprising too. Isn't the biggest
mystery how AI reason from something
small to something massive? I mean
generalizations problem. Well, it it is
a little bit of a mystery, but it is
it a mystery how this happens for
people, right? We we learn, children
learn later in life, we learn,
scientists learn, they reason from a few
experiments to something massive. So, we
have some understanding how this works.
We know how this works with with humans.
And the reasoning models are actually
in their design since they reason in
language they mimic
this like very elaborate talking part of
reasoning very well. I think the the
reasoning on like more latent spaces is
still a bit of a mystery. Maybe we'll
get there. But yeah, what does it truly
mean that LLM models generalizes, not
just memorizes? You use chat GPT. Yeah,
I guess very often you you ask it a
question and it gives you an answer
>> and you know there were language models
before. So the the first language models
people just took engrams. I don't know
if so so you basically take every three
words that appear on the internet and
then when you see two words you say well
in all of that appeared before what
would be the third word that you would
put here a long long time ago before
deep learning was so popular Google even
published like a huge database of all
scraped bagrams triagramgrams programs
from the internet and you can so it's a
language model you can start generating
words like that the problem is that it
doesn't give you very sensible answers.
So you can improve it. You can do more
words. You you can do other things,
right? You can you can do trees.
>> You mean faking?
>> No, no. These are language models. Those
are language models pre-neural network
that people
trained on also the whole internet of
sorts. Training is maybe a big word.
They basically collected the data and
they would generate you sentences even
some maybe fluent sentences reasonably
but but they would not give you answers
that you would consider interesting
right and it's similar if you trained
like small transformers on the internet
it gives you even fluent sentences but
but you wouldn't want to talk to it
right it it's something happened at
scale when it became big that this model
became interesting for for you to talk
with it.
So, so that's surprising in many ways,
right? But so, so it's surprising or let
me go. It's surprising for me as an ML
researcher because it clearly showed
that the size of your model, how much
can it compute, really affects how it
learns and what it outputs later.
Because it's very obvious that if you
make a model bigger, you can make a
database bigger, right? It will memorize
what you give to it. So maybe your model
can memorize all of internet. That's
great. But your question that you're
asking today is not in that database.
Unless you're saying hi, then maybe the
answer is memorized. But and you need to
realize language is so rich that all the
sentences we say are new.
It it's very few like formulaic things
like what's the weather today? Okay,
that that's maybe in the database, but
the sentence I'm saying right now has
not been said before most probably or or
maybe the sequence of two sentences,
right? Very quickly, you're in the realm
of things that cannot be memorized
because they were never said before. But
on the other hand, often times they're
not that hard, right? They're fairly
understandable if you have some whatever
you'd call understanding, even very
surface level understanding. [gasps]
So as we built more powerful models,
there was this risk that they'll
memorize everything. But when you talk
to it with something that's even new
only on the surface, they will know
nothing.
They will just answer you whatever some
sequence of words that
kind of fits but but but don't really
make any sense.
So the first mystery or or miracle of
deep learning is that this didn't
happen. Deep learning the models became
more powerful. They trained on more data
but they also generalize to what we say.
So clearly somewhere in the
representation some of these concepts
that were developed are not that
dissimilar from what we develop as
humans in our brains. So that that's a
big surprise, right? Because it could be
that the models would answer something
totally different, something that makes
sense in some alien form to the models,
but totally doesn't make sense to us.
Now, this still sometimes happens,
right? When the models makes mistakes,
it sometimes feels like they go on a
path that that's totally weird and alien
to us. So there is certainly something
in in this intelligence that develops
inside this model if if you want to call
it like that that's very different from
our intelligence because it goes on you
can see it when they make mistakes they
feel very weird they they feel like
humans don't make mistakes this way
right but there is also a lot which we
see in chat GPD because we chat with it
and it often makes sense so there is
also a lot that is similar to to to the
we learn and think from this data. So,
so that's a first surprise that the that
the that just making this bigger
training it a lot.
>> Why that was happen right? Why that was
happen?
>> Why why why did this happen? It could in
principle
mathematic you know it it was bigger so
it would find some patterns. It it had
to generalize in some mathematical sense
but maybe it could have generalized in
like a very nonhuman way but it doesn't.
It it it it learns in a way similar to
us which is quite beautiful and makes it
possible to you know to use it.
>> Maybe generalization it's about
computing power.
>> It it certainly is. So we we know the
more powerful the models we have
the better they generalize. So the more
they can answer questions.
>> That's what we know.
>> Yes.
But it's it's interesting to think about
it that
say 15 years ago maybe 20 15 20 some
something like that people had in in the
theory of learning this was like when
deep learning was still not very popular
people [clears throat] considered
generalization there were some
mathematical definitions but the
consensus is at least as I understood it
and of course some people disagreed with
it was that if a model is very powerful
and you have very little data, it's
going to just memorize it and not
generalize at all. And if you make it
generalize through some like minimizing
some something to to make it generalize,
then basically this process of
minimizing it, this training process is
so complicated and timeconuming that you
will never be able to practically
compute it. So, so there was a lot of
like yeah it's not going to work, right?
And then deep learning came and it
showed look here you do just gradient
descent. You take an architecture with a
few tweaks and and you need to work a
little on optimizing it. But here you
optimize it and and it learns and it
generalizes. So that was the that was
the miracle of deep learning, right? It
worked for vision then it worked with
RNN's for language and and when when I
this is more or less when I started when
uh when I had this paper about RNN's
learning to translate it was to me that
was like wow it's actually working it's
you can take some data and it's learning
to translate that's quite great
and and I even thought okay maybe that
is the end right maybe that's the
ultimate you just train big RNN this is
going to work.
But then
I fairly quickly came to this
realization that RNNs can't be this
architecture because they're not that
powerful. In particular, without
attention, they have a very limited
state, right? They have only this vector
of numbers. That's their whole memory.
So they can't do like paragraphs and
paragraphs because they have to squeeze
everything into this vector and it just
becomes not enough. Right? So, so you
need this was what led to transformers
and and even earlier even the encoder
decoder attention had this motivation to
do long sentences simply so so you can
>> tell me something more about the vectors
because
look at the chair I'm sitting on the
chair for my brain for my mind that
chair it's number of associations and
vectors and my brain
will generalize that picture creating
picture of of the chair. How does it
work deep inside AI blackbox?
>> We don't really know how the brain
works. I think that that's a good
statement. But we do know how neural
networks work because we built them. And
yes, they indeed have vectors,
>> but we know that we are living in a
vectoral space, vectoral reality. Well,
Sure, it's it's a way of describing it.
>> But the the the question for neural
networks was not so much if there are
vectors, of course, they're all based on
vectors,
but the question is how many, how large,
and how do they interact with each
other.
So, so in RNN, you had this one vector
that basically described all your memory
of the previous state and it got updated
every time you saw a word, right? and
and was very fixed like say 256 numbers
or or 1,000 numbers got updated every
step but but that was your memory and it
described everything. Now in
transformers you have a vector for
everything you've seen before. You you
you basically don't forget you you
remember everything you've seen [snorts]
and then you come and then you can think
so so this new thing that that's
attention. This new thing I've seen how
does it relate to everything before? Can
I can I find things in the past that are
similar to to some parts of it that are
related? So, so, so that feels more like
we think right we we think a lot with
associations and attention is like a
very simple mathematical method to say
what's associated with what right it
comes from alignment and translation
which is exactly like what what things
do align what things are associated
together. So, so that that that seems to
be like a huge benefit to to this
architecture that instead of storing
everything in one place and forcing
itself to create one vector that that
represents it, it says, well, we'll
create vectors too. We have feed forest
layers, but we'll also use associations
to everything we've seen before.
>> Do LMS truly have their own world model?
>> Yes, of course they do. Apparently, it's
more of a language model than a world
model. But, but you know, and as far as
their world is language, they have a
world model. Now, the new LLMs, as you
probably know, are also trained on
images. And they start being trained on
videos. So, their their model is now not
just of language, but also of images and
audio. And it starts to be a model of
video. It's not yet the model of our
world truly
because they I mean the videos we train
on are very short. The
the image representations are not yet
fully there. So so so so I think there
and also they don't move and act in the
world. So so that there there's still
some parts missing that are very natural
to us that we just don't train the
models on because also we don't have the
data. But
but but the question is is this just a
data problem or or is this a model
problem and you know we'll only really
know when we start training them. But
um you you you just said that
generalization problem it's about
computing power but maybe there is
something more
uh maybe LLMs they are dead end for any
real AI maybe generalization pro
solution for that problem it's not about
you know computing power but unbody AI
how to make AI that understand reality
that understand physical rights that
understand dimensions of the reality
around
right so so you see the the kind of
miracle that I was talking about that we
trained models on the text of the
internet and they started talking to us
oftentimes as if they were human and we
didn't even tell them to be human right
we just trained them on text and you can
ask chat GPT you know
how hard is it to climb a big mountain
and it will tell you hard but small
mountain is less hard how does it know
that right it has never seen a mountain
even I mean the model just trained on
text so so it for for a long time we we
had this huge worry that like these
things only operate on text how would
they ever know about all the other
aspects of life and also why would they
even be similar to us if if they you
know they don't move. They they don't
have many things that our brains have.
But it turns out they discover these
things as patterns in the sea of text
that is called the internet. And that is
absolutely amazing because it it shows
that if you have enough computational
power to to process patterns not just
surface level but but go deeper in them.
So so that's the part where the models
need to be powerful. the very surface
level models, they don't they just don't
have the power to discover that. But
when they have the power, it turns out
if you have enough data, these things
get discovered even without,
you know, some things that maybe make
them much easier to be discovered for us
because our, you know, we have evolved.
We're human. We we don't start like the
models on random weights of of some
plane. We so and our brain certainly
works very differently than the models
and at least in many ways
but we get to a similar things in in in
in a large space. So so that's that
that's a very beautiful thing that means
if you generalize powerfully you do
discover all these things that we
consider concepts and and that we talk
about. So that's why I'm not that
worried. You know it seems the models
they learn to use this computation to to
to
make sense of the world and to make
sense of the world they discover the
concepts that we use but still I think
that we are speculating a little bit
regarding generalizations problems. So
have you ever thought that maybe LLMs
are dead end for real AI?
>> Well I I don't think so. I mean it it
depends a little bit how narrowly or
broadly you define LLMs
like like I do believe that that in the
future of yeah you should train LLMs on
videos for example then they're not a
model of just language right they're
model of of the world
um I don't know if the LLM is still
appropriate they're become more than
language models but they're still I I
mean people still call them LLMs
these days.
>> Generalization problem
is for you the biggest challenge for
current AI.
>> Yes. I I think not reinforce men.
>> Well, well, it's related because what we
figured out is that not alignment.
>> Mhm. No, no. I I we figured out that if
we allow ourselves reasoning models and
train them with reinforcement learning,
they are much more computationally
powerful. So if we can teach them, they
generalize better. They are the better
class of models. So to train them
though, we do need to figure out
reinforcement learning. So so so it's
related to to reinforcement learning.
Now
alignment is is a bit of a different
problem, right? It's uh
it's a bit of a society problem. It's so
what you want the models to align with.
Uh but then
the the point is if if you have a model
that gets smarter and smarter, is it
easier and easier to align? If so,
great. Right? We'll just make them very
smart. [snorts] And I think there is
some hope that that that's how it works.
What do you think about human reasoning?
Does human reasoning fit entirely within
computable functions?
Oh, well, I think it's computable. I I
think it's also very different from what
we currently call reasoning in in LLMs,
but that's okay.
I always believe language models are
quite different from the human brain and
and in many ways actually inferior to it
but then they compensate with other
things like they can have more data and
for the reasoning in LLMs it's currently
a little weird because it goes token by
token by token. It feels like our
reasoning is more you know not
everything we reason about is in words
but maybe that's okay. Maybe maybe the
models will just catch up, right? It's
>> what's about human socal human intuition
is just a set of unhor undocumented
horroristic for you. Well, we have it in
our brain.
Some process happens there. And I'm
asking you about that because I remember
my interview with John Hopfield. And I
remember I asked him, "Do you regret
something? If you could turn back time,
would you do this?" And I remember his
answer. And his answer
that was yes. I regret that
AI
is being was being developed
as a part of computer science not as a
part of biology
and uh don't you think that
thinking when when we think about human
reasoning
we are you know
it's like a
we forget about that biology very often
that we are missing something
maybe that biology it's you know the
missing link
>> between human reasoning and that
alternative deep inside AI I would go
much further than what you say I would
say we're missing a lot of things so
there is biology can be an inspiration
like our brains they have some
commonalities with language models
because you know otherwise we wouldn't
be talking to Chad GPT but they have a
lot of things that are very different a
lot of things that still will be very
surprising to us and probably a lot of
these things could be translated into
improvements in AI too so it's a it's a
amazing area of research right I I never
thought transformers should be the end
of research you know [laughter]
[snorts] but on the other hand if you
look at the AI industry These days, it's
also abundantly clear that given the
data, these models can find the patterns
in the data and just use their
computational powers to exploit them.
And it's maybe not the fastest way. It's
maybe not the you know with human brain
maybe they would need even less data
but the way they are we can push them to
use reasonably little data to be
reasonably fast and that just gives us
huge opportunities in this line. So I
don't think this line is like the only
line and we should certainly research a
lot of other things because there are
exciting things in biology and they can
be translated into ideas in AI
>> but maybe maybe it's too late to adopt
biology
to AI. Maybe it's too late for that.
>> Well why would it be too late? I think
AI is just beginning. No,
>> because of that fundamental basics
architectures that was only know
inspired partly
uh by human brain well but I I feel like
to in many domains right you if you have
a technique you know it's I think it was
a year ago when we introduced reasoning
models and transformer is like not eight
years ago in in in many fields a year is
nothing and 8 years is short, right? And
and we can still push this technique and
get like 2x, 5x, 10x improvements every
year. So of course people do that,
right? It it's it's like if you have
something that works and you can make it
work better and better and better. So so
you push it. It's it's very natural. But
but you know if you think of plane
engines if you made a plane engine
that's 30% better or 70% better people
would be like wow that's great you know
and in AI it's like barely matters it
needs to be three times better five
times better every year right but but
this will I don't want to say hit a wall
because it's not a wall every every
single technique reaches some ceiling of
its progress and then you need a new
technique right this is this is how
progress
moves, right? You It was the same with
reasoning models, you know, we we push
transformers and you could see on math
that like of course the predicting the
next token thing if you have, you know,
okay, this is my problem. No, John had
three balls, he sold two balls. How many
ball does he have? Okay, maybe maybe you
can make a big model that will just say
one. But if you have a problem from a
math olympiad, it will not just give you
the token. That's the answer because
there is a lot you need to do computing
and thinking before that and if it's a
bigger model maybe it has a tiny bit
better chance but it's just not going to
work right. You still push the models
because they can chat with you because
they can do other things but you start
recognizing that of course they have
there are some things they will not be
able to do. So you push a new paradigm
like reasoning and it it's not an you
know we started working on reasoning
much before people stop I mean people
are still scaling up pre-training models
but we now have reasoning and I think
there will be other paradigms that I
think in this case they may not
be necessary for like I I don't think
there will be like exact examples of
what reasoning can't do because I think
these models are so computationally
powerful that with enough data you can
train them to do almost anything.
But maybe, you know, you'll they'll
maybe and I think they will generalize
because they're so powerful, but maybe
they'll generalize in ways that are like
weird to humans
and so maybe you'll want to make them
more humanlike in some way and for that
you'll need other techniques. We we'll
see. The the point is it's it's easiest
to see what's a problem with the
paradigm when it reaches its ceiling,
right? When it's in this rapid
development stage, then it's very hard
to say what doesn't work because you
haven't gotten there yet. And reasoning,
I mean, 01 preview was a year ago. So,
so that's very young. the the first
academic papers on all of this are only
started to appear because with deepse
one it's not even a year since people
can truly run experiments on it at a
larger scale. I'm asking you if
the worst scenario
in the next 10 years maybe 20 years when
AI will decide from that time I will be
against the humans it's possible
I the current AI is not for or against
anything
um
I think it's more like humans could make
the AI be for or against something but
but in that sense AI is a tool for some
humans that want to
>> what's for you
the biggest risk when we look at AI
development now I I do think if if you
think of social media for example right
or um
so what turned out to possibly be a
problem right is that even very
well-meaning algorithms They'll just
show you some fun videos you want to
see. These are your friends. They
somehow interact with the society and
then people get addicted to things or or
like or not or or or or the information
that you get to see is somehow not just
a fragment of the reality that that you
forget about other parts. So I think
risks like that
are very important with AI, right? a
billion people chat with chat GPT every
months or so, right? That that that's a
lot of people and and you know, we I
think the whole company tries to make
chat very truthful and and and
helpful and and you know, not
not too psychopantic and and like the
best for the people. But of course,
it's it's it's a computer program in the
end. it
for for some people you know what you
talk to it may be helpful for others it
may be unhelpful but that's okay you
know you read an unhelpful website you
get an unhelpful advice it's okay
but for some other people it will be
unhelpful but they will not get the
helpful part because they'll spend too
much time with it or or something like
that right that that that's my biggest
worry that when you interact with a
billion people
That's a huge number of people with a
lot of diverse backgrounds and you may
be doing something that's really not
appropriate for some groups of these
people. Not not appropriate in the sense
of like you know explicit or so that you
can filter out but but more sub in more
subtle ways not appropriate. And you
will kind of never know until until you
see some bad things happening because
[gasps]
you'll never think that some subtle ways
these models are like think of how chat
GPT is already changing the language,
right? I I never heard the word delve
until
I think chat started using it.
>> You're a good point.
>> And now people use it, right? And and
okay, I mean it's fine. people say del
you know that's maybe not a big problem
but but maybe something like that could
happen and people would start doing a
thing
and that could be bad right like how we
can define AI today is an LLM for you
simple tool simple thing that we can use
in a single case or something more like
you know alternative
intelligence on earth
>> definitely as a tool at this point, but
as a so
I still think it's a
to me as a researcher, it's a miracle
that these patterns that these machines
found in the language are so similar to
the ways humans think that that we can
talk to this machine that only read text
on the internet and it talks in clearly
similar concepts. most of the time and
then sometimes goes off and and they're
not that similar.
This is
this is amazing to me as a as a
researcher. So so so so from this wealth
of text there are clearly patterns there
that that we use. So it also says
something about us that we that we use
the same concepts. I I I I'm think
this is an important part of the LLM
development.
But it's it's a fact of research. Not
not saying that these LLMs are new
intelligence on Earth, but they show us
that that this text we created has all
of these patterns, right? That they're
there. What we can expect in the next
years from that patterns, abstractional
thinking, more abstractional thinking,
free reasoning, subjectivity. Well, you
know, as I told you, I I think about
this as like more and more powerful
models being learnable. So, we've now
learned we've now figured out how to
train models that are powerful enough to
do like some mathematics and coding.
So, they figured out like strategies
that mathematicians use to to solve
certain problems.
So, so they're like very smart humans
that that that do this. It's it's a
little bit like in chess, right? There
was a point where models figured out how
to play without
>> now we are way past that right chess
engines can play way better than any
human. So, so as we progress our our our
our
reasoning models, language models will
find patterns in in mathematics, in
science, maybe even broader in the world
that we just don't see because
no I mean humans operate in groups. So
luckily in groups as as scientists we
can do a bit more but but at some point
the models will show us like look this
is a strategy you can solve things this
way and no human has ever thought about
it. Right? This happened in games and
chess and go. This will happen probably
in mathematics in in the coming years
that there will be parts of mathematics
where where we'll be like wow this is
like a beautiful way to prove something
or that's a real feels like a real
discovery.
Um, now when this starts happening in in
broader science and and and in other
parts of the world, we may be astounded
like because the models may find
connections that we're just missing, but
when they show them to us, we'll be
like, well, indeed, this is so so so
hopefully they will be like someone very
smart at times showing us
showing us things in the world that that
we have missed. and they'll probably
also at least the way they're built
today be able to explain them to us. So,
so that's beautiful that you know chess
players have benefited dramatically from
chess AI these days they so so hopefully
human science and research will benefit
dramatically from the AI in the same way
now I think the way AI is built right
now it will still have this non-human
streak that that you know it will show
us these beautiful things but then and
some other things it will just be
totally wrong and it will be very
obvious to us that it's wrong and even
though it's that super smart, it it will
somehow not see it because it has a
different construction than than us.
That that would be very good, right? My
fear is that maybe as it learns
everything, it will also improve in
these ways and then we may have this
question. Okay, so what why does it even
need us, right? Have you ever think
about AI that feel pain like humans?
Because you know pain is in our brain.
Well, yes, there is a, you know, we have
a body that that no AI, even robots come
nowhere close to having a body like
ours, right? Don't have hormones. They
don't have soft skin. Maybe this is one
of the scenario AI that will feel the
pain
like humans. Well, you know, you can
talk with Chad GPT about pain. It will
tell you sensible things mostly, right?
But of course it has never felt pain. So
so so so physically,
>> right? But but this is a this is a
subtle but possibly important difference
between someone who you know can know
much about something but actually never
experienced it.
You you you have the same difference in
people not for pain maybe but
you know there were sometimes debates
you know there were poets who wrote
about war because there was a war and
now you are translating this poetry and
you have never experienced a war can you
can you translate it fully can you
really
>> give it truth well but but
>> because there is something more
>> deeper
>> right but but in many senses you can
because there's many translations ations
of war poetry by people who were never
at war and and they're still great,
right? So, but but but but you see like
we face this question a lot with humans.
No, no two people have the same
experiences. So, when we talk there's
always something I'm missing that you
experience, something that you
experience that I'm missing
it. The AI is missing huge parts of
human experience like all of pain, love,
feelings, right? But then it has read
all of the internet. So it has much more
it much more words that describe it. So
it gives it some form of representations
for that. Um
you know to which extent I I I think
robotics will be a big breakthrough in
AI because the next revolution
well one you know I think there'll be
many revolutions in parallel as the
years move on.
But but certainly you know models that
are allowed to act that can act in a
physical world not just observe it on
videos or images
they they will develop representations
that will be much more human in some
sense because you know it's as I said
it's it's absolutely astonishing that
models that don't can even talk to us so
understandably right so so we should
absolutely build machines that act in
the physical world so They can learn how
this is right how is it to lift
something folded all of that right it's
it's for humans and animals this is the
most basic things we we learn way before
language right or or at the same time as
we learn language but our machines never
do this so so so that's great right then
robots of course will still not feel
pain they don't have bodies like ours
but you know it feels like almost a
smaller
scale thing than the fact that that that
lang language models for now don't even
move ever, right? Or
so so so so we will build machines that
move. Will we build machines that are
hungry?
Maybe they'll want to recharge, you
know. Then maybe one day we'll build
ones that actually feel some form of
pain.
Or maybe we won't. Maybe it will turn
out that it's just not necessary or
practical.
>> What keeps you awake at night? No.
Regarding AI development, it's not
[laughter] question about your dreams,
but you know it's a metaphor.
>> No, I have I have little kids. They keep
me awake at night. I
>> But when we look at the science, what
keeps awake at night? Well, that's a
question. if you could you know find
answer for just only one question about
reality about fundamental AI um about
science what it would be
well I at this point I'm most excited
about
AI like being a huge catalyst for human
science like for our models and
reasoning models finding and helping
human scientists to understand improve
our science
>> and we're starting to see it in
mathematics but but but we'd like to
improve all of science.
Um
I do believe that the current generation
of reasoning models will need some
improvements to do that. Um it's it's
not very clear you know how big do the
improvements need to be how much can we
just scale up and and how much it will
work. I think the promise here is
amazing, right? If if this really works,
if it really helps scientists, if if it
suggests things that are truly better,
you know, we could we could see the
science flourishing much more. That that
would be amazing. Um I think it's within
reach, but but it needs work. I I you
know, I like to think about problems
that are not too far in the future that
it's a little bit more clear what may be
needed. On the other hand, you know, as
I as I say, the models are quite like
human, but also sometimes
especially when they make what we call
mistakes, right? For them, maybe they're
not they're very different. So, we don't
know how this will like there is some
chance that that we will scale up. We
will put our best and and they will not
be as helpful as we hope.
So, so that would be very interesting
for me as a researcher because we see
these these places where maybe we need
new ideas. Um, but maybe we can just
scale you know these ones will work with
human scientists. Maybe like chat GPD
sometimes tells stupid things doesn't
seem to be a problem for most people. So
maybe we can just have a huge impact in
science
wi without like fixing any of this. If
if it becomes a real problem then it's
even more interesting to me as a
researcher because I'll see like okay
now it's very clear where we need to put
like new research ideas
and and the these things will you know
should become clear in the next year
that that is very exciting that we're at
this level where where we can hope to
have real impact or find things where
where we need to put beautiful Thank you
very Thank you very