↔
Title: Do LLMs Understand? AI Pioneer Yann LeCun Spars with DeepMind’s Adam Brown.
Duration: 01:15:39
Total Correct Answers:
Current Caption
Correct
Learning Modes
YouTube Video Transcript Hide
Ask AI:
Export as:
Ask AI Result
The ask AI result will appear here..
(00:00:00) Your YouTube transcript will appear here
(00:00:01)
[music]
(00:00:06)
[music]
(00:00:13)
[music]
(00:00:15)
It's a pleasure to have Adam uh my
(00:00:19)
colleague and friend and Yan who's been
(00:00:22)
with us before. Yan, you really are all
(00:00:24)
over the news right now. Um, I've gotten
(00:00:27)
so many people forwarding articles about
(00:00:29)
you this week. It all kicked off on
(00:00:31)
Wednesday. Do you want to discuss the I
(00:00:34)
can just say the headline. The headline
(00:00:37)
was the equivalent of Yan Lun, chief
(00:00:40)
scientist leaves meta. Um,
(00:00:43)
do you care to comment?
(00:00:46)
>> I can neither confirm nor deny.
(00:00:48)
[laughter]
(00:00:49)
>> Okay.
(00:00:51)
So, all of the press uh that the core
(00:00:54)
that's here to get the scoop cannot get
(00:00:56)
the scoop tonight. [laughter]
(00:00:58)
All right. Well, um you can come
(00:01:00)
afterwards and buy a drink and see how
(00:01:02)
far you get with um
(00:01:04)
>> Oh, really? I had one, but that wasn't
(00:01:06)
[laughter]
(00:01:07)
>> the Frenchman had some wine upstairs.
(00:01:10)
So, we have this era where every time
(00:01:14)
any of us turn on the news, look at the
(00:01:16)
computer, read the paper, we're
(00:01:17)
confronted with conversations about the
(00:01:19)
societal implications of AI, and whether
(00:01:21)
it's about economic upheaval or um the
(00:01:25)
potential for political manipulation or
(00:01:27)
AI psychosis. There's a lot of pundits
(00:01:30)
out there discussing this and I and and
(00:01:32)
I it is a very important issue. I kind
(00:01:34)
of want to push that towards the end of
(00:01:36)
our conversation because what a lot of
(00:01:38)
people who are discussing this don't
(00:01:40)
have is the technical expertise that's
(00:01:42)
on this stage. And so I really want to
(00:01:44)
begin by grounding this in that
(00:01:48)
technical scientific conversation. And
(00:01:51)
so I want to begin with you Yan about
(00:01:53)
neural nets. Here's this instance of
(00:01:56)
kind of biomimicry where you have these
(00:01:58)
computational neural networks that are
(00:02:01)
emulating human networks. Can you
(00:02:03)
describe to us what that means that a
(00:02:07)
machine is emulating human neural
(00:02:09)
networks?
(00:02:10)
>> Well, it's not really mimicry. It's more
(00:02:12)
inspiration. The same way I don't know
(00:02:15)
airplanes are inspired by by birds,
(00:02:18)
right? The underlying
(00:02:20)
>> that didn't [clears throat] work. I
(00:02:21)
thought,
(00:02:21)
>> say again.
(00:02:22)
>> But I thought that didn't work. Copying
(00:02:24)
birds with airplanes. Well, in the sense
(00:02:26)
that you know airplanes have wings like
(00:02:28)
birds and they generate lift by
(00:02:31)
propelling themselves through the air,
(00:02:33)
but then the analogy stops stops there.
(00:02:35)
And the wing of an airplane is much
(00:02:37)
simpler than the wing of a bird, but yet
(00:02:39)
the underlying principle is the same. So
(00:02:42)
neural networks are a bit like like that
(00:02:44)
like are like you know our to real
(00:02:47)
brains as airplanes are to birds.
(00:02:50)
They're much simplified in many ways. Um
(00:02:54)
but perhaps some of the underlying
(00:02:56)
principles are the same. We don't
(00:02:58)
actually know because we don't really
(00:03:00)
know the sort of underlying algorithm of
(00:03:03)
the cortex if you want or the the the
(00:03:06)
method by which the brain organizes
(00:03:08)
itself and and learns. So we invented
(00:03:13)
substitutes
(00:03:15)
um sort of like you know birds flap
(00:03:17)
their wings and not airplanes right they
(00:03:19)
have propellers so or or turbo jets you
(00:03:22)
know in in neural nets we have learning
(00:03:25)
algorithms and they they allow
(00:03:29)
artificial neural nets to learn in a way
(00:03:32)
that we think is similar to how the
(00:03:34)
brains learn. So the brain is a network
(00:03:37)
of neurons. The neurons are
(00:03:39)
interconnected with each other and the
(00:03:41)
way the brain learns is by modifying the
(00:03:44)
efficacy of the connections between the
(00:03:45)
neurons and the way a neural net is
(00:03:48)
trained is by modifying the efficacy of
(00:03:50)
the connections between those simulated
(00:03:52)
neurons. Um each of those is like a we
(00:03:55)
call it a parameter. You you you see
(00:03:57)
this is a press the number of parameters
(00:03:59)
of a neural net right? So the the
(00:04:01)
biggest neural net at the moment have
(00:04:03)
you know hundreds of billions of
(00:04:04)
parameters if not more and um those are
(00:04:08)
the individual coefficients that are
(00:04:11)
modified by by by training. So
(00:04:14)
>> and how is deep learning uh emerge in
(00:04:18)
this discussion because deep learning
(00:04:20)
came along the path after thinking about
(00:04:22)
neural nets and this has been since the
(00:04:24)
80s or earlier even
(00:04:25)
>> um yeah 80s roughly. Um
(00:04:29)
so early neural nets um the the first
(00:04:32)
ones that were capable of of learning or
(00:04:34)
learning something useful at least in
(00:04:36)
the 50s uh were shallow. You could you
(00:04:39)
could basically train a single layer of
(00:04:42)
neurons, right? So you would feed the
(00:04:44)
input and train the system to produce a
(00:04:46)
particular output and and you could use
(00:04:48)
those things to kind of recognize or
(00:04:51)
classify relatively simple patterns uh
(00:04:55)
but not really sort of complex things
(00:04:57)
and people at the time even in the 60s
(00:04:59)
realized that the way to make progress
(00:05:01)
was going to be able to train neuronets
(00:05:04)
with multiple layers. They built
(00:05:05)
neuronets with multiple layers but they
(00:05:07)
couldn't train all the layers. It would
(00:05:08)
only train the last layer for example.
(00:05:11)
Um, and they didn't really find uh until
(00:05:16)
the 1980s, nobody found really a good
(00:05:18)
way to train those those multi-layer
(00:05:20)
systems. Uh, mostly because the neurons
(00:05:23)
that that they had at the time were the
(00:05:26)
wrong type. Um, they had neurons that
(00:05:28)
were binary. So, neurons in the brain
(00:05:30)
are binary. They they either fire or
(00:05:32)
they don't fire. Um, and people wanted
(00:05:35)
to reproduce that. So they they built
(00:05:37)
simulated neurons that would either be
(00:05:40)
active or inactive. And it turns out for
(00:05:42)
the modern learning algorithms to work,
(00:05:44)
we call them back we call it back
(00:05:46)
propagation. You need to have neurons
(00:05:47)
that have sort of graded responses. Um
(00:05:51)
and uh that only became practical,
(00:05:55)
possible or people realized it could
(00:05:56)
work in the 1980s. People had the idea
(00:05:59)
before but they never could really make
(00:06:01)
it work. And so that caused um a renewal
(00:06:05)
of interest in neural nets in the 1980s.
(00:06:07)
They had been largely abandoned in the
(00:06:09)
late60s and then they came to the four
(00:06:12)
again in the mid to late 80s. That's
(00:06:14)
when I started kind of my graduate
(00:06:18)
school basically in 1983 and uh there
(00:06:21)
was a wave of interest that lasted about
(00:06:23)
10 years and then interest waned again u
(00:06:27)
in the mid9s until the late 2000 when we
(00:06:31)
rebranded it into deep learning. Neural
(00:06:33)
net had kind of a bad rep um people in
(00:06:38)
computer science and engineering thought
(00:06:40)
neuron nets were kind of a bad thing. It
(00:06:43)
had a bad reputation and so we rebranded
(00:06:45)
it into deep learning and sort of
(00:06:47)
brought it back to the four and then the
(00:06:49)
results were were there in computer
(00:06:52)
vision in natural language understanding
(00:06:54)
speech recognition to really convince
(00:06:56)
people that this was a good thing.
(00:06:59)
>> Now Adam you at at a very young age were
(00:07:02)
interested in theoretical physics not
(00:07:04)
specifically computer science and you're
(00:07:06)
watching some of this unfold in some
(00:07:07)
sense from afar. What's the catalyst
(00:07:10)
that sweeps up so many people decades
(00:07:14)
later? There's there's this time where
(00:07:16)
it's of great interest, there's great
(00:07:18)
success in handwriting recognition or uh
(00:07:22)
uh visual recognition and these things,
(00:07:24)
but it's not sweeping up the world. What
(00:07:26)
what happens that brings us to this
(00:07:28)
point where we're all now talking about
(00:07:30)
large language models? So many
(00:07:32)
physicists in the last years have
(00:07:35)
pivoted, should we say, from working on
(00:07:37)
physics to working on AI. And it really
(00:07:40)
traces back to some of the work that Yan
(00:07:43)
and others did to prove that it works.
(00:07:46)
Like when it wasn't working, it was just
(00:07:48)
this this thing that's over there in
(00:07:49)
computer science and like of of many
(00:07:51)
things in the world that are not
(00:07:52)
particularly uh maybe interesting, but
(00:07:55)
not many physicists are paying attention
(00:07:56)
to it. But then after you know Yan and
(00:07:59)
some of the other pioneers of this field
(00:08:01)
proved that it would work it became a
(00:08:03)
totally fascinating subject for physics
(00:08:05)
that you link up these neurons together
(00:08:08)
in a certain way and suddenly you get
(00:08:12)
emergent behavior that didn't exist at
(00:08:14)
the individual neuron level. That seems
(00:08:16)
like a a subject that physicists who
(00:08:18)
spend their life imagining how the sort
(00:08:20)
of rich pageantry of the world could
(00:08:22)
emerge from simple laws. that
(00:08:24)
immediately attracted the attention of
(00:08:25)
many physicists and nowadays it's a a a
(00:08:27)
very common career path to do a PhD in
(00:08:29)
physics and then apply it to a emergent
(00:08:32)
system but the emergent system is an
(00:08:34)
emergent network of neurons that
(00:08:36)
collectively give rise to intelligence.
(00:08:40)
>> Now let's [clears throat] do a lightning
(00:08:42)
round because you raised the dreaded
(00:08:43)
word intelligence. [laughter]
(00:08:45)
Um everybody in this room very likely
(00:08:48)
has interacted with something that we're
(00:08:49)
now calling an AI. These are all large
(00:08:51)
language models. And before I ask you to
(00:08:54)
define those for us, I just want to kind
(00:08:55)
of do a lightning round of um of what's
(00:09:00)
your yes or no response to certain
(00:09:02)
things. So um Adam, [laughter]
(00:09:06)
yes or no? Are these AIs, these large
(00:09:09)
language models, uh understanding the
(00:09:12)
meaning of the conversations they are
(00:09:13)
having with us? Yes or no?
(00:09:15)
>> Yes.
(00:09:16)
>> Yan.
(00:09:17)
>> Sort of. [laughter]
(00:09:20)
Um, perfect.
(00:09:22)
>> Yan's neurons are not stuck as binary
(00:09:24)
values, isn't it? [laughter]
(00:09:26)
>> Right. Exactly. Um, it was my fault for
(00:09:29)
giving you a binary choice. Okay. So,
(00:09:31)
that allows me to ask the next question
(00:09:33)
because it's not a foregone conclusion.
(00:09:36)
If you don't say yes to that, it's going
(00:09:38)
to be interesting what you say to this.
(00:09:39)
Are these AIs conscious?
(00:09:43)
>> Absolutely not.
(00:09:44)
>> Adam,
(00:09:46)
>> probably not. [laughter]
(00:09:47)
>> Okay.
(00:09:48)
Um, will they soon be?
(00:09:52)
>> I think they'll one day be conscious if
(00:09:53)
if progress continues in the way that
(00:09:55)
we're we're continuing.
(00:09:57)
>> When is hard to say, but
(00:10:00)
>> Mhm.
(00:10:02)
>> for appropriate definitions of
(00:10:04)
consciousness.
(00:10:04)
>> Yes. Okay. Well, we do have some
(00:10:06)
philosophers in the house and um we're
(00:10:09)
we're not going to indulge in
(00:10:11)
philosophical definitions of
(00:10:12)
consciousness or there our hour would go
(00:10:14)
[laughter]
(00:10:15)
and we'd still be here. Oh, I just heard
(00:10:18)
that groan, I think, from our friends up
(00:10:20)
in the balcony. Yes. [laughter]
(00:10:22)
>> Um, but I have one other question. Okay.
(00:10:24)
No, I have two. I have two in the
(00:10:25)
lightning round. Uh, are we on the
(00:10:27)
precipice of doomsday or a renaissance
(00:10:30)
in human creativity? Yan
(00:10:32)
>> renaissance.
(00:10:33)
>> Adam,
(00:10:35)
>> most likely Renaissance. [laughter]
(00:10:39)
>> Um, I have to throw this out the same
(00:10:41)
question to the audience, but I'm going
(00:10:42)
to phrase it more colorfully, which I
(00:10:44)
think they'll relate to. Will the robot
(00:10:46)
overlords rise up against humanity? Yes.
(00:10:49)
Hands up. Oh, interesting. Okay. No.
(00:10:54)
Hands up.
(00:10:56)
Okay. How many robots in the audience?
(00:10:58)
Hands up. Okay. So, okay. So, that's
(00:11:02)
interesting. See, that's cool. It was a
(00:11:05)
little more nose maybe, although the
(00:11:07)
light is blinding. All right. We're
(00:11:09)
going to come back and ask that again at
(00:11:10)
the end. Um, so here we are. These
(00:11:14)
neural nets have been taught to uh
(00:11:19)
execute a process we now call deep
(00:11:20)
learning. And there's other kinds of
(00:11:22)
learning that take off. And what are the
(00:11:23)
large language models specifically which
(00:11:25)
is really what has swept up the news and
(00:11:29)
people's personal experience? We're
(00:11:31)
we're mostly relating to large language
(00:11:32)
models and and what are the large
(00:11:34)
language models Adam? Maybe you could
(00:11:35)
take that.
(00:11:36)
>> Yeah. So large language model is uh
(00:11:39)
you've probably played with some of
(00:11:40)
them. Chat GPT Gemini made by my company
(00:11:44)
uh various others um made by other
(00:11:46)
companies. It is a special kind of
(00:11:49)
neural network that's trained on
(00:11:51)
particular inputs and particular outputs
(00:11:52)
and trained in a particular way. So it
(00:11:55)
is at at heart it is mainly the kind of
(00:11:58)
deep neural network that was pioneered
(00:12:00)
by by Yan and by others but uh with a
(00:12:03)
particular architecture designed for the
(00:12:06)
following task. uh it takes text in. So
(00:12:09)
it'll it'll read some uh the first few
(00:12:12)
words of some sentence or the first few
(00:12:15)
paragraphs of some book and it will try
(00:12:17)
and predict what the next word is going
(00:12:20)
to be. And so you take a deep neural
(00:12:23)
network with a particular architecture
(00:12:25)
and you have it read basically to first
(00:12:29)
approximation the entire internet and
(00:12:32)
for every word that comes along on the
(00:12:34)
entire internet all of the text data and
(00:12:36)
now other kind of data you can find uh
(00:12:39)
you then ask it what do you think the
(00:12:40)
next word's going to be? What do you
(00:12:41)
think the next word's going to be? And
(00:12:43)
to the extent that it gets it right you
(00:12:46)
give it a little bit of uh reward and
(00:12:47)
strengthen those neural pathways. to the
(00:12:49)
extent that it gets it wrong, you you
(00:12:51)
diminish those neural pathways. And if
(00:12:54)
you do that, uh it'll just start off
(00:12:56)
spewing just completely random words for
(00:12:58)
its prediction. But, uh if you train it
(00:13:01)
on a million words, it'll still be
(00:13:03)
spewing random words. If you train it on
(00:13:05)
a a billion words, it'll maybe have just
(00:13:07)
started to learn subject, verb, object,
(00:13:10)
and various bits of sentence structure.
(00:13:13)
uh and if you train it as we do today on
(00:13:15)
on a trillion words or more, tens of
(00:13:17)
trillions of words, uh then it'll start
(00:13:19)
become the conversation partner that you
(00:13:21)
you've probably I hope uh played around
(00:13:23)
with today.
(00:13:25)
>> Now, um it
(00:13:28)
it strikes me as intriguing like it's
(00:13:30)
it's it amuses me sometimes people get
(00:13:32)
really outraged at their chatbot that
(00:13:35)
they're engaged with when it leads them
(00:13:38)
astray or lies to them. And sometimes
(00:13:40)
I've said, well, it's it's doesn't need
(00:13:41)
to be words. it it might as well be
(00:13:44)
colors or symbols. It's just playing a
(00:13:46)
mathematical game and therefore doesn't
(00:13:49)
have a sense of meaning. Now, I know
(00:13:50)
Adam sort of objected to my summary of
(00:13:53)
that. Do you think that they are
(00:13:57)
extracting meaning
(00:14:00)
um in the same sense that we do when we
(00:14:03)
are engaging in composing sentences?
(00:14:07)
Well, they're certainly extracting some
(00:14:08)
meaning. Um, but it's it's a lot more
(00:14:12)
superficial than what most humans would
(00:14:16)
extract from from text. Most humans uh
(00:14:21)
intelligence is linked to is is grounded
(00:14:24)
into an underlying reality, right? And
(00:14:27)
language is a way to express
(00:14:30)
phenomena or things in that or concepts
(00:14:33)
grounded in that reality. uh LLMs don't
(00:14:36)
have any notion of the underlying
(00:14:38)
reality and so their understanding is is
(00:14:42)
relatively superficial. Um they don't
(00:14:45)
really have common sense in the in the
(00:14:46)
way that we understand it. U but if you
(00:14:50)
train them long enough they they will
(00:14:53)
answer correctly most questions that
(00:14:55)
people will think about asking. That's
(00:14:58)
the way they're trained. you you you
(00:15:01)
collect all the questions that everybody
(00:15:03)
has ever asked them and then you trend
(00:15:04)
them to produce the correct answer for
(00:15:06)
this. Now there's always going to be new
(00:15:10)
questions or new prompts, new sequences
(00:15:12)
of words for which the system has not
(00:15:14)
really been trained and for which it
(00:15:16)
might produce complete nonsense. Okay,
(00:15:18)
so in that sense they don't have the
(00:15:21)
real understanding of the underlying
(00:15:22)
reality or they do have an understanding
(00:15:24)
but it's it's superficial. Um, and so
(00:15:29)
you know, and the next question is, how
(00:15:30)
do we fix that?
(00:15:31)
>> So I I could play devil's advocate and
(00:15:34)
say, well, how do I know that what a
(00:15:36)
human being doing is doing is that much
(00:15:38)
different, right? We're trained on lots
(00:15:40)
of language. We get some dopamine hit or
(00:15:44)
some reward system for having said the
(00:15:46)
right word at the right time and the
(00:15:48)
right grammatical structure for the
(00:15:50)
language that we're immersed in. And um,
(00:15:52)
and we back propagate. [laughter] we try
(00:15:54)
to do a better job the next time. In
(00:15:57)
some sense, how how is that different uh
(00:16:01)
than what a human being is doing? And
(00:16:02)
you you were saying maybe it's the
(00:16:04)
sensory experience of being immersed in
(00:16:05)
the world.
(00:16:07)
>> Okay. Um a typical L&M as I mentioned is
(00:16:13)
trained on tens of trillions of of
(00:16:16)
words. Typically
(00:16:18)
>> there's only a few hundred thousand
(00:16:19)
words of it. You're just saying
(00:16:20)
sentences.
(00:16:21)
>> It's combinations. No, it's 30 trillion
(00:16:24)
30 trillion words is is a a typical size
(00:16:27)
for the training set pre-training of of
(00:16:29)
an LLM. Uh a a word is represented
(00:16:34)
actually as sequences of tokens doesn't
(00:16:36)
really matter. Uh and a token is about
(00:16:39)
three bytes. So the total is about 10 to
(00:16:41)
the 14 bytes, right? One with 14 zeros
(00:16:44)
um of training data to train those LLMs.
(00:16:48)
And that corresponds to basically all
(00:16:50)
the text that is uh publicly available
(00:16:52)
on the internet plus some other stuff.
(00:16:54)
And it would take any of us something
(00:16:55)
like half a million half a million years
(00:16:58)
for any of us to read through that
(00:17:00)
material. Right? So it's an enormous
(00:17:02)
amount of textual data. Now compare this
(00:17:05)
with what a child uh perceives during
(00:17:09)
the first few years of life. Um
(00:17:11)
psychologists tell us that a
(00:17:12)
four-year-old has been awake a total of
(00:17:15)
16,000 hours. Um, and there's about
(00:17:21)
one bite per second going through the
(00:17:24)
optic nerve. Every single fiber of the
(00:17:26)
optic nerve, and we have two millions of
(00:17:28)
them. So, it's about 2 megabytes per
(00:17:30)
second getting to the visual cortex. Um,
(00:17:35)
during 16,000 hours, do the arithmetics
(00:17:38)
and it's about 10^ the 14 bytes. A
(00:17:40)
four-year-old has seen as much visual
(00:17:44)
data as the biggest LLM trained on the
(00:17:47)
entire text ever produced. And so what
(00:17:49)
that tells you is that there is way more
(00:17:53)
um information in the real world, but
(00:17:55)
it's also much more complicated. It's
(00:17:58)
noisy. It's high dimensional. It's
(00:18:00)
continuous. And basically the methods
(00:18:02)
that are employed to train LLMs do not
(00:18:05)
work in the real world. That explains
(00:18:08)
why we have LLMs that can pass the bar
(00:18:11)
exam or solve equations or compute
(00:18:14)
integrals like college students and
(00:18:16)
solve math problems. But we still don't
(00:18:18)
have a domestic robot. They can, you
(00:18:20)
know, do the chores in the house. We
(00:18:22)
don't we don't even have level five
(00:18:24)
self-driving cars. I mean, we have them,
(00:18:25)
but we cheat. So, um I mean, we
(00:18:29)
certainly don't have self-driving cars
(00:18:30)
that can learn to drive in 20 hours of
(00:18:33)
practice like any teenager, right? So
(00:18:35)
obviously we're missing something very
(00:18:38)
big to get machines to the level of
(00:18:40)
human or even animal intelligence,
(00:18:42)
right? Let's not talk about language.
(00:18:43)
Let's talk about how a cat is
(00:18:44)
intelligent or a dog. Um we we're not
(00:18:48)
even at that level with AI systems.
(00:18:52)
>> Adam, you you
(00:18:55)
impart more comprehension on uh the part
(00:18:59)
of the LLMs at this point already. Uh
(00:19:02)
>> I think that's right. So I mean Yan
(00:19:05)
is making sort of excellent points that
(00:19:07)
the LLMs are much less for example
(00:19:10)
sample efficient than humans. humans or
(00:19:13)
indeed your cat or just a a cat, I don't
(00:19:16)
know if it was your cat or any smart cat
(00:19:18)
>> in your example, um is able to learn
(00:19:22)
from many fewer examples than a large
(00:19:26)
language model, for example, can learn
(00:19:28)
from that takes way more data to teach
(00:19:30)
it to the same level of proficiency. Um
(00:19:34)
and and that's true and that that is a
(00:19:35)
thing that is better about uh the you
(00:19:37)
know architecture of animal minds
(00:19:40)
compared to these artificial minds that
(00:19:41)
we're building. Um on the other hand
(00:19:43)
sample efficiency isn't everything. Um
(00:19:47)
we see this frequently in fact when we
(00:19:49)
try and you know before large language
(00:19:51)
models when we try and put uh machines
(00:19:54)
on you know make artificial minds to do
(00:19:57)
other tasks even the famous chess bots
(00:19:59)
that we built uh on built on tops of
(00:20:02)
large language models uh the way they
(00:20:04)
were trained sort of alpha zero and
(00:20:05)
various other ones they would play each
(00:20:07)
other uh they would play itself at chess
(00:20:10)
a huge number of times and to begin with
(00:20:12)
it would just be making random moves and
(00:20:14)
then uh every time it it won or lost the
(00:20:17)
game when it was playing itself, it
(00:20:18)
would sort of uh you know reward that
(00:20:21)
neural pathway or punish that neural
(00:20:23)
pathway. And they play each other at
(00:20:24)
chess again and again. And when they
(00:20:26)
played as many games as a human
(00:20:28)
grandmaster has played, they were still
(00:20:30)
making essentially random moves. But
(00:20:32)
they didn't were not confined to making
(00:20:34)
the same number of move uh playing the
(00:20:36)
same number of games that a human
(00:20:37)
grandmaster could play. Because silicon
(00:20:39)
chips are so fast because we can build
(00:20:43)
them with such parallel processing. They
(00:20:45)
were able to play many more human more
(00:20:49)
games than any human could ever play in
(00:20:50)
their lifetime. And what we found is
(00:20:52)
that when they did that, they reached
(00:20:55)
and then far surpassed the level of
(00:20:57)
human chess players. They're less sample
(00:21:00)
efficient, but that doesn't mean they're
(00:21:01)
worse at chess. It is clear that they're
(00:21:02)
much better at chess. So too with
(00:21:04)
understanding when uh we it is it is
(00:21:09)
true that we can you know it is harder
(00:21:13)
uh with these things to you need more
(00:21:15)
samples to get them up to the same level
(00:21:17)
of proficiency. But the question is once
(00:21:19)
they've reached that can we use the fact
(00:21:21)
that they are so much more general and
(00:21:23)
so much more so much faster and more
(00:21:25)
inherent to push beyond that. I I mean
(00:21:28)
another example perhaps with the cat is
(00:21:30)
a cat is in fact even more sample
(00:21:32)
efficient than a human. Uh a human takes
(00:21:35)
a a year to learn to to walk. A cat
(00:21:38)
learns to walk in a in a week or so. You
(00:21:40)
know it's much much faster. That does
(00:21:42)
not mean that a cat is smarter than a
(00:21:44)
human. Uh it does not mean that a cat is
(00:21:46)
smarter than a large language model. The
(00:21:49)
final question at the end should be what
(00:21:51)
is the capabilities of these things? How
(00:21:53)
far can we push the capabilities? And on
(00:21:56)
almost every uh except for the somewhat
(00:21:58)
impoverished metric of sample
(00:21:59)
efficiency, on every metric that counts,
(00:22:02)
uh we've pushed these uh large language
(00:22:04)
models far beyond the frontier of cat
(00:22:06)
intelligence.
(00:22:07)
>> So [laughter]
(00:22:10)
[gasps]
(00:22:10)
um yes, I don't understand why we're not
(00:22:13)
making cats, but [laughter]
(00:22:16)
sorry, what was I? I mean certainly the
(00:22:18)
L&Ms in question have much more
(00:22:20)
accumulated knowledge than cats or even
(00:22:23)
humans for that matter and we do have
(00:22:26)
many examples of computers being far
(00:22:28)
superior to humans in a number of uh you
(00:22:31)
know different tasks like playing chess
(00:22:33)
for example um that's humbling I mean it
(00:22:36)
just means that humans just suck at
(00:22:38)
chess that's all it means no we really
(00:22:40)
suck at chess and go by the way even
(00:22:43)
even more um and and many other tasks
(00:22:46)
that computers are much better than than
(00:22:48)
us um at at at solving. Um so certainly
(00:22:53)
LLMs can accumulate a huge amount of of
(00:22:55)
of knowledge and some former them can be
(00:22:59)
trained to translate languages
(00:23:01)
understand spoken language and and
(00:23:03)
translate it into another one from you
(00:23:06)
know a thousand languages to another
(00:23:08)
thousand languages in any direction. No
(00:23:10)
human can do this. Um, so they they do
(00:23:12)
have superhuman capabilities. U, but the
(00:23:16)
ability to learn quickly, efficiently,
(00:23:18)
to apprehend a new problem that we've
(00:23:21)
never been trained to solve and be able
(00:23:25)
to come up with a solution. Um, and to
(00:23:28)
really, you know, understand a lot about
(00:23:30)
how the how the world behaves that is
(00:23:34)
still out of reach of AI systems at the
(00:23:36)
moment.
(00:23:38)
I I would I mean we've had recent
(00:23:39)
successes with this where it is not the
(00:23:42)
case that they're just taking problems
(00:23:43)
that they've seen before letter for
(00:23:45)
letter and looking up the answer in a in
(00:23:48)
a lookup table or even that they're uh
(00:23:51)
they are they are in some sense doing
(00:23:53)
pattern matching but they're doing
(00:23:54)
pattern matching at a sufficiently
(00:23:56)
elevated level of abstraction that
(00:23:58)
they're able to do things that they've
(00:24:00)
never seen before and no no human can
(00:24:02)
do. So there's a there's a competition
(00:24:04)
uh each year called the International
(00:24:05)
Maths Olympiad. Um it is the very
(00:24:08)
smartest
(00:24:10)
uh finishing high school maths uh
(00:24:12)
teenagers in the entire world. They're
(00:24:15)
all given six problems uh each year. The
(00:24:17)
pinnacle of human intelligence. I have
(00:24:19)
some mathematical ability. I look at
(00:24:21)
these problems. I don't even know where
(00:24:22)
to start.
(00:24:23)
um you know this this year we fed them
(00:24:26)
into our machine
(00:24:28)
uh as as did a number of other LLM
(00:24:31)
companies and they took these problems
(00:24:33)
they'd never seen before they were
(00:24:34)
completely fresh didn't appear anywhere
(00:24:36)
in the training data completely made up
(00:24:38)
to a whole bunch of different ideas
(00:24:40)
combined the different ideas and got a
(00:24:42)
score on these tests that was better
(00:24:44)
than all except the first dozen the top
(00:24:46)
dozen humans on the planet I think
(00:24:49)
that's uh that's pretty impressive
(00:24:50)
intelligence
(00:24:52)
>> I I The question is um back to this idea
(00:24:57)
do they understand you we can look at
(00:25:01)
the mathematics of the model there's
(00:25:02)
some input data we understand what it's
(00:25:04)
doing it is a black box which is kind of
(00:25:06)
fascinating it's just so complex that
(00:25:10)
it's not as though we can't do that with
(00:25:11)
the human mind either it's not as though
(00:25:13)
you can look at the inner workings and
(00:25:15)
see exactly what they're doing to some
(00:25:16)
extent it is a black box but we presume
(00:25:18)
it's just doing these calculations it's
(00:25:20)
moving these matrices it's working in
(00:25:21)
some vector face it's doing some higher
(00:25:23)
dimensional thing I have some experience
(00:25:25)
of understanding I guess people are
(00:25:28)
still grasping at that is it having some
(00:25:31)
experience of understanding is it
(00:25:33)
important whether or not they experience
(00:25:35)
understanding is that sufficient to call
(00:25:38)
it comprehension of meaning
(00:25:41)
>> are you describing understanding as a
(00:25:43)
behavioral trait here where it gives the
(00:25:45)
right answers to problems or whether it
(00:25:47)
deeply at the neural level understands
(00:25:49)
>> yeah I'm I'm completely at the whims
(00:25:51)
little philosophers here. No, I I don't
(00:25:53)
know if I understand that at my at the
(00:25:56)
human level, right? I can't tell you
(00:25:58)
what process I'm executing at the moment
(00:26:00)
either, right? But I have some intuitive
(00:26:03)
subjective experience that I understand
(00:26:05)
the conversation. Obviously, not that
(00:26:07)
well. Um but but uh I when I'm talking
(00:26:12)
to you, I feel you are understanding
(00:26:16)
and uh when I'm talking to chat GBT, I
(00:26:19)
do not. And you're telling me I'm
(00:26:21)
mistaken. It's understanding as well as
(00:26:22)
I am or you are.
(00:26:24)
>> In my opinion, it is understanding. Yes.
(00:26:26)
And I think there's two different pieces
(00:26:29)
of evidence for that. One is I think if
(00:26:31)
you talk to them like
(00:26:34)
if you talk to them and ask them about
(00:26:36)
difficult concepts I'm frequently
(00:26:39)
surprised and with every passing month
(00:26:41)
and every new model that comes out I am
(00:26:43)
more and more surprised at the level of
(00:26:45)
sophistication with which they're able
(00:26:47)
to discuss things. And so just just at
(00:26:51)
that level it it's super impressive. I I
(00:26:53)
would really encourage everybody here um
(00:26:56)
to talk to these large language models
(00:26:58)
if you've not already. You know, when
(00:27:00)
the science fiction writers imagined
(00:27:03)
that we'd built some sort of touring
(00:27:05)
test passing uh machine that that was
(00:27:08)
going to, you know, some new alien
(00:27:09)
intelligence that we'd have in a box. uh
(00:27:11)
they all imagined that we'd sort of hide
(00:27:14)
it in a basement, you know, in a castle
(00:27:16)
surrounded by a moat with arms guards
(00:27:18)
and we'd only have like a priestly class
(00:27:20)
who be able to go and and talk to it. Uh
(00:27:23)
that is not not as not the way it worked
(00:27:24)
out. The way it's worked out is the
(00:27:26)
first thing we did is we immediately
(00:27:27)
hooked it up to the internet and now
(00:27:29)
anybody can go talk to it and uh I would
(00:27:31)
highly encourage you to to talk to these
(00:27:34)
things and explore in areas that you
(00:27:36)
know to see both their limitations but
(00:27:38)
also their strength and their their
(00:27:39)
depth of understanding. So, I'd say
(00:27:40)
that's the first piece of evidence. The
(00:27:42)
second piece of evidence is you said
(00:27:44)
they're a black box. They're not exactly
(00:27:46)
a black box. We do have access to their
(00:27:48)
neurons. In fact, we have much better
(00:27:49)
access to the neurons of these things
(00:27:51)
than we do with a human. It's very hard
(00:27:53)
to get IRB approval to slice open the
(00:27:55)
human while they're doing a math test
(00:27:57)
and see how their neurons are firing.
(00:27:59)
And if you do do that, uh you can only
(00:28:01)
do that once on a per human basis. Uh
(00:28:04)
whereas these neural networks, we can
(00:28:06)
freeze them, replay them, write down
(00:28:08)
everything that happened. uh if we're
(00:28:09)
curious, we go and go go and prod their
(00:28:11)
neurons in certain ways and see what
(00:28:13)
happened. And so this is it's still
(00:28:15)
rudimentary, but this is the field of
(00:28:16)
interpretability, mechanistic
(00:28:18)
interpretability, trying to understand
(00:28:20)
not just what they say, but why they say
(00:28:22)
it, how they think it. And when you do
(00:28:24)
that, we see uh when you feed them a
(00:28:28)
math problem, there's a little bit of a
(00:28:30)
a circuit there that computes the answer
(00:28:33)
that that we didn't program it to have
(00:28:34)
that. It learned how to do that while
(00:28:36)
trying to predict the next token on all
(00:28:38)
of this text. It learned that in order
(00:28:41)
to most accurately predict the next the
(00:28:42)
next word, I should say in order to most
(00:28:44)
accurately predict the next word, it
(00:28:46)
needed to figure out uh how to do maths
(00:28:49)
and it needed to build a sort of proto
(00:28:51)
little circuit inside it to do the
(00:28:52)
mathematical computations.
(00:28:54)
>> Now Yan, you famously threw a slide up
(00:28:58)
at one of your uh keynote lectures that
(00:29:01)
was very provocative um very scholarly.
(00:29:04)
It said u machine learning sucks I
(00:29:07)
believe was it and then that kind of
(00:29:08)
went wild. Yan Lun says machine learning
(00:29:11)
sucks. Um why are you saying machine
(00:29:13)
learning sucks? Adam has just told us
(00:29:15)
how phenomenal it is. He talks to them
(00:29:18)
[laughter] and wants us to do the same.
(00:29:21)
Um why do you think it sucks? What's the
(00:29:23)
problem? [clears throat]
(00:29:25)
>> Well, that statement has been widely
(00:29:28)
misinterpreted. But
(00:29:31)
the point the point I was making is the
(00:29:33)
point that u we both we both made which
(00:29:36)
is that why is it that a teenager can
(00:29:39)
learn to drive a car in 20 hours of
(00:29:41)
practice. Uh a 10-year-old can clean up
(00:29:45)
the dinner table and fill up the
(00:29:47)
dishwasher the first time you ask the
(00:29:49)
child to do it. Whether the 10-year-old
(00:29:52)
will want to do it is a different story,
(00:29:53)
but you know certainly can. Um, we don't
(00:29:57)
have robots that are anywhere near this
(00:30:00)
and we don't have robots that are even
(00:30:02)
anywhere near the, you know, physical
(00:30:05)
understanding of of reality of of a cat
(00:30:08)
or a dog. And so in that sense, machine
(00:30:10)
learning sucks. It doesn't mean that the
(00:30:12)
the deep learning method, the back
(00:30:14)
propagation algorithm, the neural nets
(00:30:16)
suck.
(00:30:17)
>> That was obviously excellent. Yes,
(00:30:19)
>> obviously that's great. [laughter]
(00:30:20)
>> And we don't have any alternative to
(00:30:22)
this. And uh I I certainly believe that
(00:30:28)
you know neural nets and deep learning
(00:30:29)
and back propagation would be you know
(00:30:31)
are with us for for a long time would be
(00:30:33)
the basis of future AI systems. But but
(00:30:37)
how is it that uh u you know young
(00:30:40)
humans can can learn how the world works
(00:30:42)
in the first few months of life. It
(00:30:44)
takes nine months for human babies to
(00:30:46)
learn um intuitive physics like gravity,
(00:30:49)
inertia and things like this. Um, baby
(00:30:52)
animals learn this much faster. They
(00:30:53)
have smaller brains, so it's easier for
(00:30:55)
them to learn. Um, they don't learn to
(00:30:58)
the same level, but they but they do
(00:30:59)
learn faster. And and so, you know, it's
(00:31:02)
this type of learning that we need to
(00:31:03)
reproduce. Um, and we'll do this with
(00:31:06)
back prop with neural net with deep
(00:31:07)
learning. It's just that we're missing a
(00:31:09)
concept, an architecture. Um, so I've
(00:31:12)
been I've been making proposals for the
(00:31:14)
type of architectures that could
(00:31:16)
possibly
(00:31:17)
learn this kind of stuff. You know, why
(00:31:19)
is it that LLMs handle language so
(00:31:23)
easily? It's because um as Adam
(00:31:26)
described, you you train an LLM to
(00:31:29)
predict the next word or the next token,
(00:31:32)
doesn't matter. There's only a finite
(00:31:34)
number of words in the dictionary. So
(00:31:37)
you can never actually predict exactly
(00:31:38)
which word comes after a sequence, but
(00:31:41)
you can train a system to produce
(00:31:43)
essentially what amounts to a score for
(00:31:45)
every possible words in your dictionary
(00:31:46)
or a probability distribution over every
(00:31:48)
possible words. So essentially what an
(00:31:50)
LLM does is that it produces a long list
(00:31:53)
of numbers between 0 and one that sum to
(00:31:55)
one which for each word in the
(00:31:57)
dictionary says this is the likelihood
(00:31:59)
that this word appears right now. You
(00:32:01)
can represent the uncertainty in the
(00:32:03)
prediction this way. Now try to
(00:32:06)
translate it um the same principle
(00:32:09)
instead of training a system to predict
(00:32:10)
the next word um feed it with a video
(00:32:14)
and ask it to predict what happened next
(00:32:16)
in the video and this doesn't work. I've
(00:32:18)
been trying to do this for 20 years and
(00:32:21)
it it really doesn't work if you try to
(00:32:23)
predict at the pixel level. Uh and it's
(00:32:26)
because
(00:32:27)
the real world is messy. There's a lot
(00:32:29)
of things that that may happen,
(00:32:31)
plausible things that may happen. Um,
(00:32:34)
and you can't really represent a
(00:32:36)
distribution over all possible uh things
(00:32:39)
that may happen in the future because
(00:32:40)
it's basically an infinite list of
(00:32:43)
possibilities and we don't know how to
(00:32:44)
represent this um efficiently. And so
(00:32:47)
those those techniques that work really
(00:32:49)
well for text or for sequences of of
(00:32:52)
symbols do not work for real world
(00:32:56)
sensory data. They just don't. They
(00:32:59)
absolutely don't. And and so we need to
(00:33:01)
invent new techniques. So one of the
(00:33:03)
things I've been proposing in one in
(00:33:05)
which the the system learns an abstract
(00:33:08)
representation of what it observes and
(00:33:10)
it makes prediction in that abstract
(00:33:11)
representation space. And this is really
(00:33:13)
the way humans and animals function. We
(00:33:15)
we find abstractions that allow us to
(00:33:18)
make predictions while ignoring all the
(00:33:20)
detail the details we cannot predict. So
(00:33:23)
you really think that despite the
(00:33:25)
phenomenal successes of these LLMs that
(00:33:28)
they are limited and and their limit is
(00:33:31)
quickly approaching. You don't think
(00:33:33)
that they're scalable to this you know
(00:33:35)
artificial general intelligence or a
(00:33:37)
super intelligence.
(00:33:37)
>> That's right. No they [clears throat]
(00:33:38)
don't and and in fact we see the
(00:33:40)
performance saturating. So we we see uh
(00:33:43)
progress in in some domains like
(00:33:46)
mathematics for example and mathematics
(00:33:48)
and and code generation you know
(00:33:51)
programming are two domains where the uh
(00:33:54)
the the manipulation of symbols actually
(00:33:56)
gives you something right as a physicist
(00:33:58)
you you know this right you write the
(00:34:00)
equation and it actually kind of
(00:34:02)
>> follow you can follow it and it it uh it
(00:34:05)
drives your your thinking to some extent
(00:34:07)
right I mean you you drive it by
(00:34:09)
intuition but but the simple
(00:34:10)
manipulation itself actually has uh
(00:34:13)
meaning. So this type of problems LLMs
(00:34:15)
actually can handle pretty well where
(00:34:18)
the the reasoning really consists in
(00:34:19)
kind of searching through sequences of
(00:34:21)
symbols but it's only there's only a
(00:34:23)
small number of problems for which
(00:34:24)
that's the case. Chess playing is
(00:34:26)
another one. Um you search through
(00:34:28)
sequences of of of moves that you know
(00:34:31)
for a good one or sequences of uh
(00:34:34)
derivations in mathematics that will
(00:34:36)
produce a particular result, right? Um,
(00:34:38)
but in the real world, you know, in high
(00:34:41)
dimensional continuous things where the
(00:34:43)
search has to do with like how do I move
(00:34:45)
my muscles to uh, you know, grab this
(00:34:48)
uh, you know, grab grab this um this
(00:34:51)
this glass here. I'm not going to do it
(00:34:52)
with my left hand, right? I'm going to
(00:34:54)
have to change hand with this and and
(00:34:56)
then grab it, right? you need to plan
(00:34:59)
and have some understanding of what's
(00:35:01)
possible, what's not possible that you
(00:35:03)
know I can't just attract the glass you
(00:35:06)
know by telekinesis or I can't just I
(00:35:09)
can't just make it appear in my in my
(00:35:11)
left hand like this I can't have my hand
(00:35:13)
kind of cross my body like you know all
(00:35:15)
of those intuitive things we we learned
(00:35:18)
them when we were babies um and and we
(00:35:20)
learn you know how our body reacts to
(00:35:23)
our controls and how uh you know
(00:35:28)
the world reacts to to the actions we
(00:35:31)
take. So, you know, if I push this
(00:35:33)
glass, I know it's going to slide. If I
(00:35:36)
push it from the top, maybe maybe it's
(00:35:37)
going to flip. Maybe not because the
(00:35:40)
friction is not that high. If I push
(00:35:42)
with the same force on this table, it's
(00:35:43)
not going to flip. You know, we have all
(00:35:45)
those those intuitions that allow us to
(00:35:47)
kind of apprehend the real world. Uh but
(00:35:50)
this is it turns out much much more
(00:35:53)
complicated than manipulating language.
(00:35:56)
We think of language as kind of the
(00:35:57)
epitome of, you know, human intelligence
(00:36:00)
and stuff like that. It's actually not
(00:36:01)
true. Language is actually easy.
(00:36:03)
[laughter]
(00:36:05)
>> Is it the Morvec paradox that what
(00:36:08)
computers are good at, humans are bad
(00:36:10)
at? What humans are good at, computers
(00:36:12)
are bad at?
(00:36:13)
>> Yeah, we keep running into the Marx.
(00:36:15)
Yeah. Now, Adam, I I know that you are
(00:36:19)
less pessimistic about the potential of
(00:36:21)
the current neural net deep learning um
(00:36:26)
paradigm and you see the potential for a
(00:36:28)
great escalation in success and you
(00:36:30)
don't see it saturating. Um what's your
(00:36:33)
thought about that?
(00:36:34)
>> I um
(00:36:37)
>> I don't That's right. Um and so yeah,
(00:36:40)
>> we have witnessed
(00:36:42)
>> over the last 5 years the most
(00:36:44)
extraordinary runup in capabilities in
(00:36:48)
any system I've ever seen. This is what
(00:36:51)
transfixed my attention. It's what
(00:36:53)
transfixed many other people uh in AI
(00:36:58)
and neighboring fields to focus all of
(00:37:02)
our attention on this matter. I don't
(00:37:05)
see any slowdown in the capabilities. a
(00:37:08)
year ago. If you just look at all of the
(00:37:09)
all of the metrics we use to judge how
(00:37:12)
good these large language models are,
(00:37:14)
they're getting stronger and stronger
(00:37:15)
and stronger things that they you know a
(00:37:17)
the model from a year ago today would be
(00:37:20)
you know table stakes will be considered
(00:37:22)
extremely poor. Every few months these
(00:37:24)
things push their capabilities and if if
(00:37:27)
you track their capabilities on all of
(00:37:29)
these tasks uh they're heading towards
(00:37:31)
superhuman on on almost all of them.
(00:37:34)
It's already better gives better legal
(00:37:36)
advice than uh than a lawyer. It gives
(00:37:40)
better um it's a better poet than almost
(00:37:43)
every poet you all come in my little
(00:37:46)
area in my little area of physics. Uh I
(00:37:49)
I use it because like there's something
(00:37:51)
I kind of should know but I don't. I'll
(00:37:53)
ask the language model and it will not
(00:37:55)
only tell me what the right answer is,
(00:37:56)
it will patiently and I should say
(00:37:58)
non-judgmentally listen while I explain
(00:38:00)
my misconception to it and it will
(00:38:02)
carefully debunk my misconception. Um,
(00:38:06)
the extraordinary runup in capabilities
(00:38:08)
that we've seen over the last 5 years uh
(00:38:12)
and that continues up to the present is
(00:38:15)
extremely tantalizing to to me and and
(00:38:17)
many other people in San Francisco. And
(00:38:19)
and maybe maybe Yan is correct that
(00:38:22)
we're just going to suddenly saturate
(00:38:24)
and all of these uh straight lines that
(00:38:26)
have been going up steadily for the last
(00:38:28)
five years are suddenly going to stop
(00:38:29)
going up. But I am mighty curious to see
(00:38:33)
uh whether we can push it further. And
(00:38:35)
I've actually seen no indication
(00:38:36)
whatsoever that it's slowing down. Every
(00:38:38)
indication I've seen is that these these
(00:38:40)
are improving and we don't have far to
(00:38:42)
go because once it's a better coder than
(00:38:45)
almost all our best coders, it can start
(00:38:47)
improving itself and then we're really
(00:38:49)
in for a wild ride.
(00:38:50)
>> Well, we we've had better coders than
(00:38:53)
the original coders of 1950s,
(00:38:56)
[clears throat] you know, for six
(00:38:57)
decades or so. That's called compilers.
(00:38:59)
I mean we
(00:39:01)
we we keep getting confused about the
(00:39:05)
fact that it's not because machines are
(00:39:08)
good at a certain number of tasks that
(00:39:11)
they have all the underlying
(00:39:13)
intelligence that we assume a human
(00:39:16)
having those capabilities will have.
(00:39:18)
Right? We're fooled into thinking those
(00:39:19)
machines are intelligent because they
(00:39:21)
can manipulate language. and we're used
(00:39:23)
to the fact that people who can
(00:39:25)
manipulate language very well are
(00:39:28)
implicitly smart. Um but we're being
(00:39:31)
fooled. Um now they they're useful.
(00:39:35)
There's no question. Um you know we can
(00:39:38)
use them to do what you said. I use them
(00:39:40)
for similar things. Great. They're great
(00:39:43)
tools like you know computers uh have
(00:39:46)
been for the last decade five decades.
(00:39:49)
But let me make an interesting
(00:39:50)
historical point.
(00:39:51)
>> [snorts]
(00:39:52)
>> Um, and this is maybe due to my age. Uh,
(00:39:56)
there's been generation after generation
(00:39:59)
of AI scientists
(00:40:01)
since the 1950s claiming that the
(00:40:05)
technique that they just discovered was
(00:40:07)
going to be the ticket for human level
(00:40:10)
intelligence. you you see declarations
(00:40:12)
of Marvin Minsky, Newan Simon, um you
(00:40:16)
know, Frank Rosenblad who invented the
(00:40:19)
perceptron, the first learning machine
(00:40:21)
in 1950 saying like within 10 years
(00:40:23)
we'll have machines that are as smart as
(00:40:25)
humans. They were all wrong. This
(00:40:28)
generation with L&M is also wrong. I've
(00:40:30)
seen three of those generation in my
(00:40:32)
lifetime. Okay. Um so you know it's it
(00:40:37)
it's just another example of being
(00:40:39)
fooled and
(00:40:41)
um in the 50s New and Simon pioneers of
(00:40:44)
AI came up with a program they said well
(00:40:47)
you know really what what humans are
(00:40:49)
doing um is in reasoning is really a
(00:40:52)
search right every reasoning can be
(00:40:54)
reduced to kind of a kind of search. So
(00:40:57)
you formulate a problem, you write a
(00:40:59)
program that tell you whether a
(00:41:01)
particular proposal for a solution is a
(00:41:03)
solution to your problem and then you
(00:41:05)
just have to search for all possible
(00:41:07)
combinations, you know, all possible
(00:41:09)
hypothesis for one that actually matches
(00:41:12)
uh satisfies the the constraint and
(00:41:16)
that's it. We're going to write a
(00:41:17)
program that does this and we're going
(00:41:18)
to call it the general problem solver
(00:41:20)
GPS 1957.
(00:41:23)
I think um they won the training award
(00:41:26)
for for things like that and it was it
(00:41:28)
was great but then they didn't realize
(00:41:29)
that all the interesting problems
(00:41:31)
actually have a complexity that grows
(00:41:33)
exponentially with the size of the
(00:41:35)
problem. So in fact you can't really use
(00:41:37)
this uh uh technique to build
(00:41:40)
intelligent machines. It can be a
(00:41:42)
component of it but it's really not not
(00:41:43)
the thing. Simultaneously
(00:41:45)
uh for Rosenlack came up with a
(00:41:47)
perceptron a machine that could learn
(00:41:48)
and he said if we can train a machine
(00:41:50)
then it can become infinitely smart and
(00:41:52)
so within 10 years we'll have we just
(00:41:54)
need to big you know to build bigger
(00:41:56)
perceptrons right not realizing that you
(00:41:59)
need to train multiple layers and that
(00:42:00)
turned out to be uh difficult to find a
(00:42:03)
solution for this. Um then in the 1980s
(00:42:06)
there was um expert systems. Okay,
(00:42:10)
reasoning is is fine. Just write a bunch
(00:42:12)
of facts and a bunch of rules and then
(00:42:15)
just deduce all the facts from the
(00:42:17)
original facts and the and the rules and
(00:42:20)
u now we can reduce all the human
(00:42:22)
knowledge into into this. The the
(00:42:25)
coolest job is going to be knowledge
(00:42:27)
engineer. you're going to sit down next
(00:42:29)
to an expert and then write down all the
(00:42:31)
rules and the facts and turn this into
(00:42:33)
an expert system and you know everybody
(00:42:36)
was excited about this and there were
(00:42:37)
you know billions that were invested the
(00:42:40)
Japan started the fifth generation
(00:42:42)
computer program pro project which was
(00:42:46)
which was going to revolutionize
(00:42:48)
computer science complete failure okay
(00:42:50)
it created an industry it was useful for
(00:42:52)
a few things but basically the cost of
(00:42:56)
reducing human knowledge age to to rules
(00:42:59)
uh was just too high for most problems
(00:43:01)
and so the whole thing collapsed. Then
(00:43:03)
there was neural nets the the first the
(00:43:05)
second wave of neural nets in 1980s deep
(00:43:07)
you know which we now call deep learning
(00:43:10)
a lot of interest but then it was before
(00:43:13)
the internet we didn't have enough data
(00:43:14)
we didn't have powerful computers and
(00:43:16)
now we're we're going through the same
(00:43:18)
cycle again and we're getting fooled
(00:43:19)
again
(00:43:20)
>> so just to be oh Adam please
(00:43:22)
>> in in technologies every dawn has before
(00:43:25)
it false dawn that doesn't mean we'll
(00:43:27)
never we'll never hit the dawn I I guess
(00:43:29)
I would like um Yan, if you think that
(00:43:34)
LLMs are going to saturate, what is a
(00:43:36)
concrete task that they will never be
(00:43:39)
able to do? That a that an LLM augmented
(00:43:41)
by, you know, the the tools we give it
(00:43:43)
today will never be able to perform uh
(00:43:48)
clear the dinner table, fill up the
(00:43:49)
dishwasher. [laughter]
(00:43:52)
>> Okay.
(00:43:53)
>> And that's easy compared to
(00:43:54)
>> I'm skeptical.
(00:43:55)
>> That's super easy compared to fixing
(00:43:56)
your toilets.
(00:43:57)
>> Yeah.
(00:43:58)
>> Okay. Plumber, right? You're never going
(00:43:59)
to have a plumber with L&Ms. You're
(00:44:01)
never going to have a robot driven by
(00:44:02)
L&M. It just cannot understand the real
(00:44:05)
world. It just can't.
(00:44:06)
>> So I want to clarify for the audience
(00:44:07)
that you're not saying that machines or
(00:44:10)
robots won't be able to do this. That's
(00:44:11)
not your position. You think they will.
(00:44:13)
>> They will. They absolutely
(00:44:14)
>> not by this algorithmic approach or for
(00:44:17)
this particular approach of the deep
(00:44:18)
learning on the
(00:44:18)
>> program we're working on succeeds which
(00:44:21)
may take a while.
(00:44:22)
>> This is cheaper. Am I Japa?
(00:44:24)
>> Ja and and you know all the things world
(00:44:27)
models and things that go with it. If it
(00:44:29)
succeeds, which may take, you know,
(00:44:30)
several years, then we we may have, you
(00:44:33)
know, AI system. There's no question
(00:44:35)
that at some point in the future, we
(00:44:36)
will have machines that are smarter than
(00:44:38)
humans in all domains that, you know,
(00:44:40)
where humans have, uh, abilities.
(00:44:43)
There's no question about that. It will
(00:44:44)
happen. Okay? It probably take longer
(00:44:46)
than, you know, some of the people in
(00:44:48)
Silicon Valley at the moment are saying
(00:44:49)
it it it will. Uh, and uh, and it it
(00:44:54)
will not be LLM. It will not be
(00:44:55)
generative models that predict discrete
(00:44:57)
tokens. It will be models that learn
(00:45:00)
abstract representations and make
(00:45:02)
predictions in abstract representations
(00:45:04)
and can reason about what is going to be
(00:45:07)
the effect of me taking this action. Can
(00:45:09)
I plan a sequence of actions to arrive
(00:45:11)
at a particular goal?
(00:45:12)
>> You call this self-supervised learning.
(00:45:14)
>> No. So self-supervised learning is used
(00:45:16)
also by LMS. So supervised learning is
(00:45:18)
the idea that you train a system not for
(00:45:21)
a particular task other than capturing
(00:45:24)
the the sort of underlying structure of
(00:45:27)
the of the data you you you show it. And
(00:45:30)
one way to do this is to give it a piece
(00:45:34)
of of data corrupt it in some way by uh
(00:45:38)
removing a piece of it for example
(00:45:40)
masking a piece of it and then training
(00:45:41)
a bit neural net to predict the piece
(00:45:43)
that is missing. So, LLMs do this,
(00:45:47)
right? You take a text, you remove the
(00:45:49)
last word, and you train the LLM to
(00:45:50)
predict the the word that is missing.
(00:45:52)
You have other types of language models
(00:45:55)
that actually fill up multiple words,
(00:45:57)
they turn out to not work as well as the
(00:45:59)
ones that just predict the last one. Um,
(00:46:02)
at least for certain task. Um, you can
(00:46:04)
do this with video. If you try to
(00:46:06)
predict at the pixel level, it doesn't
(00:46:08)
work or it doesn't work very well. um my
(00:46:10)
colleagues at Meta probably boiled a
(00:46:13)
couple small lakes in the west coast to
(00:46:14)
you know trying to make this work. Um
(00:46:18)
[laughter]
(00:46:19)
to cool the GPUs u so it simply doesn't
(00:46:23)
work. So, so you have to, you know, come
(00:46:25)
up with those new architectures like JA
(00:46:27)
and stuff like that and those kind of
(00:46:28)
work like we we have models that
(00:46:30)
actually understand video
(00:46:32)
>> and Adam are people exploring other ways
(00:46:36)
of building an architecture or imagining
(00:46:39)
a computer mind this the actual
(00:46:40)
fundamental structure of a computer mind
(00:46:42)
and how it would um how it would learn
(00:46:44)
how it would acquire information. One of
(00:46:46)
the criticisms as I understand it is
(00:46:48)
it's a lot of the LLM are trained for
(00:46:50)
this one specific task of this discrete
(00:46:52)
prediction of these um tokens. But
(00:46:55)
something that is more unpredictable
(00:46:57)
like how the audience is distributed in
(00:46:58)
this room. What might happen with the
(00:47:00)
weather next unpredictable more human
(00:47:03)
experience-based
(00:47:04)
phenomena.
(00:47:06)
>> Um certainly all kinds of explorations
(00:47:08)
are being made in all kinds of
(00:47:09)
directions including yans and you know
(00:47:11)
let a thousand flowers bloom. Um
(00:47:16)
but all of the resources I mean the bulk
(00:47:17)
of the resources right now are going
(00:47:19)
into
(00:47:21)
large language models and large language
(00:47:22)
model like applications including taking
(00:47:25)
in text to say to say that they are it's
(00:47:29)
a specialized task predicting the next
(00:47:31)
token. I think that's a not a helpful
(00:47:34)
way to think about it. It is true that
(00:47:36)
the thing that you train them on is
(00:47:39)
given this corpus of text. I mean there
(00:47:41)
are other things we do as well but the
(00:47:42)
the bulk of the compute goes to given
(00:47:43)
this corpus of text please predict the
(00:47:46)
next word please predict the next word
(00:47:47)
please predict the next word. Um but we
(00:47:50)
have discovered something truly
(00:47:51)
extraordinary by doing it which is that
(00:47:54)
given a large enough body of text to be
(00:47:57)
able to reliably predict the next word
(00:47:59)
or you know do it do it well enough to
(00:48:03)
predict the next word you really need to
(00:48:04)
understand the universe and we have seen
(00:48:06)
the emergence of understanding of the
(00:48:08)
universe as we've done that. So I I
(00:48:10)
would liken it a little bit. I mean in
(00:48:12)
physics we're very used to systems where
(00:48:15)
you just take a very simple rule and you
(00:48:18)
know by the repeated application of that
(00:48:20)
very simple rule you get extremely
(00:48:22)
impressive behavior. Uh we see the same
(00:48:25)
with these LLMs. Uh and another example
(00:48:28)
of that would maybe be evolution. You
(00:48:29)
know at each stage in evolution you just
(00:48:31)
say uh biological evolution you just say
(00:48:34)
you know maximize the number of
(00:48:35)
offspring maximize number of offspring.
(00:48:37)
maximize some number of offspring. Uh a
(00:48:39)
very sort of unsophisticated learning
(00:48:41)
objective. But out of this simple
(00:48:43)
learning objective repeated many many
(00:48:46)
times uh you eventually get all of the
(00:48:48)
you know splendor of biology that we see
(00:48:50)
around us and and indeed this room. So
(00:48:54)
the evidence is that predicting the next
(00:48:56)
token while a very simple task because
(00:48:59)
it's so simple we can do it at massive
(00:49:00)
scale huge amounts of compute and once
(00:49:02)
you do it at huge amounts of compute you
(00:49:04)
get an emergent complexity.
(00:49:07)
[clears throat]
(00:49:07)
>> So I I guess the next question could be
(00:49:10)
related to evolution. However this
(00:49:13)
intelligence emerges that you both
(00:49:15)
imagine is certainly possible. You don't
(00:49:17)
think there's anything special about
(00:49:18)
this wetwware that there will be
(00:49:20)
machines. We just have to figure out how
(00:49:21)
to launch them that will um have
(00:49:24)
capacities that we align as a kind of
(00:49:26)
intelligence or maybe consciousness
(00:49:29)
that's a almost a different question.
(00:49:30)
Will consciousness be a crutch machines
(00:49:33)
don't need? I don't know. We can talk
(00:49:34)
about that. But but is there a point in
(00:49:36)
the evolution of these uh machines where
(00:49:39)
they're going to say, "Oh, how quaint
(00:49:41)
mom and dad. You you made me in your
(00:49:43)
image with these human neural nets." But
(00:49:46)
I know a way a much better way having
(00:49:48)
scanned 10,000 years of human uh output
(00:49:51)
to make machine intelligence and I'm
(00:49:53)
going to evolve and leave us in the
(00:49:56)
dust. I mean yeah what are we why are we
(00:49:59)
imagining they would be limited at that
(00:50:01)
capacity to the way we design them?
(00:50:04)
>> Uh absolutely. This is this idea of
(00:50:06)
recursive self-improvement where when
(00:50:09)
they're bad that they're useless, but
(00:50:11)
when they get good enough and strong
(00:50:13)
enough, you can start using them to
(00:50:15)
augment human intelligence and uh
(00:50:18)
perhaps eventually just be fully
(00:50:20)
autonomous and replace and make future
(00:50:22)
versions of them. Once we do that, I
(00:50:24)
mean, I think what we should do is just
(00:50:26)
take this large language model paradigm
(00:50:28)
that's currently working so well and
(00:50:30)
just see how far we can push it. you
(00:50:31)
know, it keeps every time someone says
(00:50:33)
there's a barrier, it pushes through the
(00:50:34)
barrier over the last five years, but
(00:50:36)
eventually these things will get smart
(00:50:38)
enough and then they can uh read Yan's
(00:50:41)
papers, uh, read all the other papers
(00:50:43)
that have been made, try and figure out
(00:50:45)
uh new ideas that none of us have
(00:50:46)
thought of.
(00:50:48)
>> Yeah.
(00:50:48)
>> So, I completely disagree with this. Um,
(00:50:53)
so LMS are not controllable. It's not
(00:50:58)
dangerous because they're not that
(00:50:59)
smart. As I as I explained previously uh
(00:51:02)
and they're certainly not autonomous in
(00:51:04)
a way that uh we understand autonomy. We
(00:51:07)
have to distinguish between autonomy and
(00:51:09)
intelligence. You can be very
(00:51:10)
intelligent without being autonomous and
(00:51:12)
you can be autonomous without without
(00:51:14)
being intelligent. Um and you can be
(00:51:17)
dangerous without being particularly
(00:51:18)
intelligent. Um,
(00:51:21)
and you can want to be dominant without
(00:51:26)
being intelligent. In fact, that's going
(00:51:27)
to be inversely correlated in the human
(00:51:29)
species. [laughter]
(00:51:32)
Um,
(00:51:36)
[laughter]
(00:51:37)
>> you know, politics.
(00:51:39)
Um, I won't site names.
(00:51:43)
So
(00:51:45)
I think what
(00:51:47)
what is required is systems that are
(00:51:50)
intelligent in other words can solve
(00:51:52)
problems for us but it will solve the
(00:51:55)
problem we give them. Okay and again
(00:51:58)
that would require a new design than
(00:52:01)
LLMs. LLMs are not designed to fulfill a
(00:52:06)
goal. They're designed to predict the
(00:52:08)
next word and we fine-tune them so that
(00:52:12)
they behave, you know, for particular
(00:52:14)
questions they answer in a particular
(00:52:16)
way. Um, but there's always what's
(00:52:18)
called a generalization gap, which means
(00:52:20)
you can never train them for every
(00:52:22)
possible uh question and there's a very
(00:52:24)
long tail. And so they're not
(00:52:26)
controllable. Um, and [snorts] again,
(00:52:29)
that doesn't mean they're very it's very
(00:52:31)
dangerous because they're not that
(00:52:32)
smart. Um now if we build systems that
(00:52:35)
are smart we want them to be
(00:52:36)
controllable and we want them to be
(00:52:38)
driven by objectives. We give them an
(00:52:40)
objective and the only thing they can do
(00:52:43)
is fulfill this objective according to
(00:52:46)
their you know internal model of the
(00:52:48)
world if you want. So plan a sequence of
(00:52:50)
actions that will fulfill that
(00:52:51)
objective. If we design them this way
(00:52:54)
and we also put guard rails in them so
(00:52:57)
that
(00:52:59)
in the process of fulfilling the
(00:53:01)
objective they don't do anything you
(00:53:02)
know bad for for humans. Um so the the
(00:53:05)
usual joke is if you have a robot um
(00:53:08)
domestic robot and you ask it to fetch
(00:53:10)
you coffee and someone else is you know
(00:53:12)
someone is standing in front of the
(00:53:13)
coffee machine you don't want your robot
(00:53:15)
to just you know kill that person to get
(00:53:17)
access to the coffee machine right so
(00:53:19)
you want to put some guardrail uh into
(00:53:21)
the the behavior of that robot and we do
(00:53:23)
have those guardrails in our head
(00:53:25)
evolution build them into us right so we
(00:53:27)
don't kill each other all the time I
(00:53:29)
mean we do kill each other all the time
(00:53:30)
but not you know not all the time all
(00:53:32)
the time Um
(00:53:36)
I mean you and you know we feel empathy
(00:53:38)
and and things like that and that's just
(00:53:39)
built into us by evolution that that's
(00:53:42)
the way evolution set of hard wire
(00:53:44)
guardrails into us. So we should build
(00:53:45)
our AI systems the same way have
(00:53:49)
objectives and goals drives but also um
(00:53:54)
you know guardrails inhibition basically
(00:53:57)
um and and then they will solve problems
(00:54:00)
for us. They will amplify our
(00:54:01)
intelligence. they will uh do what we
(00:54:04)
ask them to do and our relationship to
(00:54:07)
those intelligence system will be like
(00:54:09)
the relationship of let's say a
(00:54:12)
professor with graduate students who are
(00:54:14)
smarter than them right
(00:54:15)
>> hey [laughter]
(00:54:17)
>> I mean I don't know about you but I have
(00:54:18)
students who are smarter than me so um
(00:54:23)
>> it's the best thing that can happen to
(00:54:24)
you right
(00:54:24)
>> yes it's the best thing that can happen
(00:54:26)
>> right so we'll be working around with AI
(00:54:29)
assistant um that will help us you our
(00:54:31)
daily lives. They be smarter than us,
(00:54:33)
but they will work for us. They be like
(00:54:35)
our staff. Again, there is a political
(00:54:38)
analogy here, right? A politician,
(00:54:39)
right, is a figurehead and they have a
(00:54:41)
staff of people all of all of whom are
(00:54:44)
smarter than them, right? Um, so it's
(00:54:46)
going to be the same thing with AI
(00:54:48)
system, which is why I to the question
(00:54:50)
of Renaissance, I said Renaissance.
(00:54:52)
>> So, you have no concerns um about the
(00:54:55)
safety of the current models, but the
(00:54:58)
question is maybe we should stop there.
(00:55:00)
I mean, why is it necessary for us to
(00:55:03)
scale up so widely that every single
(00:55:05)
person has uh this super intelligence in
(00:55:09)
their pocket on their iPhone? Is that
(00:55:11)
really necessary? A friend of mine was
(00:55:13)
saying it's like bringing a ballistic
(00:55:16)
missile [clears throat] to a knife
(00:55:17)
fight. I mean, is this necessary that
(00:55:19)
every person has a ballistic missile
(00:55:21)
capability? Um, or should we stop here
(00:55:24)
where we have these controllable
(00:55:26)
systems? You can say exactly the same
(00:55:27)
thing
(00:55:28)
>> about
(00:55:30)
teaching people to read, giving giving
(00:55:33)
them a textbook of chemistry of volatile
(00:55:36)
volatile chemicals, you know, with which
(00:55:38)
they can make explosives or nuclear
(00:55:41)
physics book, right? I mean, we do not
(00:55:44)
question the idea that knowledge and
(00:55:48)
more intelligence is good, intrinsically
(00:55:51)
good, right? We do not question anymore
(00:55:55)
the fact that the invention of printing
(00:55:57)
press was a good thing, right? It made
(00:55:59)
everybody smarter. It it gave it gave
(00:56:02)
access to knowledge to everyone. Um,
(00:56:05)
which was not possible before. It
(00:56:07)
incited people to learn to read. It uh
(00:56:09)
it caused the enlightenment. It also
(00:56:11)
caused 200 years of, you know, religious
(00:56:13)
wars in Europe. But
(00:56:15)
>> okay, but
(00:56:17)
>> he got over it. Yeah.
(00:56:19)
>> But it it caused the enlightenment.
(00:56:20)
because you know the emergence of
(00:56:23)
philosophy, science, democracy, the
(00:56:25)
American revolution, the French
(00:56:27)
revolution um all of that would not have
(00:56:29)
been possible without uh the the
(00:56:32)
printing press. So you know every
(00:56:34)
technology that particularly
(00:56:37)
communication technology but technology
(00:56:38)
that amplifies human intelligence I
(00:56:40)
think is intrinsically good. Now, Adam,
(00:56:43)
people are concerned. I'm sure they'll
(00:56:45)
feel very reassured um that Jan is not
(00:56:48)
concerned and these doomsday scenarios
(00:56:50)
you think are highly exaggerated, but
(00:56:52)
are you concerned about some of the
(00:56:54)
safety issues around AI or our ability
(00:56:56)
to really uh keep the relationship
(00:56:59)
balance in the direction that we want it
(00:57:01)
to be?
(00:57:02)
>> Um I think to the extent that I think
(00:57:05)
this is going to be a more powerful
(00:57:07)
technology than Yan thinks it does, I am
(00:57:09)
more concerned. I think it's going to be
(00:57:12)
a very power to the extent that it is a
(00:57:13)
very powerful technology. It'll have
(00:57:15)
both positive and negative impacts. Um
(00:57:19)
and I think it's very important to make
(00:57:20)
sure that you know that we work together
(00:57:22)
to make sure that the positive impacts
(00:57:24)
are uh outweigh the negative impacts. I
(00:57:27)
think that path is totally open to us.
(00:57:29)
There are huge number of possible
(00:57:30)
positive impacts and we could just you
(00:57:33)
know talk about some of those perhaps
(00:57:35)
but uh we need to make sure that that
(00:57:38)
happens. Now let's talk about agentic
(00:57:39)
misalignment which is the phrase that's
(00:57:42)
been passed along. It was my
(00:57:43)
understanding there was reports recently
(00:57:45)
that when claude 4 was rolled out that
(00:57:48)
those in simulations and tests uh one of
(00:57:52)
the models was or I don't know if
(00:57:55)
there's a singular model I don't know if
(00:57:56)
it thinks it's of itself as a singular
(00:57:58)
entity or they um but the model uh
(00:58:02)
exhibited resistance to rumors in the
(00:58:06)
simulation that it was going to be
(00:58:07)
replaced. It was sending messages to its
(00:58:12)
future self um trying to undermine the
(00:58:14)
intentions of the developers. It faked
(00:58:17)
legal documents and it threatened to
(00:58:19)
blackmail one of the engineers,
(00:58:22)
[laughter] right? Um so this notion they
(00:58:24)
were concerned
(00:58:26)
um uh so this notion of agentic
(00:58:28)
misalignment is that something that
(00:58:31)
you're concerned with that there will be
(00:58:32)
a power over say financial systems
(00:58:36)
heating and cooling systems the energy
(00:58:38)
grid and um and that that they will
(00:58:42)
resist its developers intentions.
(00:58:46)
>> Yes. So the that paper was a paper by uh
(00:58:49)
Anthropic which is a paper at a company
(00:58:51)
in San Francisco, not my company, but a
(00:58:52)
company that takes safety very seriously
(00:58:54)
and they did a slightly mean thing to
(00:58:56)
their LLM where they gave it a scenario
(00:58:59)
sort of philosophy professor style
(00:59:02)
scenario where it had to do a bad thing
(00:59:05)
to stop an even worse thing happening.
(00:59:07)
uh sort of you know utilitarian ethics
(00:59:10)
and deansical ethics colliding and it
(00:59:13)
was eventually persuaded by them to do
(00:59:14)
the utilitarian thing and that's kind of
(00:59:16)
not what we we want I would say we
(00:59:19)
wanted that if it has a rule that it you
(00:59:21)
know will not lie that it will not lie
(00:59:23)
uh no no matter what um and to their
(00:59:25)
credit they tested it for that found
(00:59:27)
that it would occasionally act
(00:59:29)
deceptively if if promised that by doing
(00:59:31)
so it could save that many lives these
(00:59:34)
are tricky things that you know human
(00:59:35)
philosophers wrestle with um I think it
(00:59:39)
is a we need to be careful to train them
(00:59:41)
to obey our command and and we spend a
(00:59:45)
lot of time doing that.
(00:59:46)
>> Us um isn't this a big concern? Uh we're
(00:59:50)
assuming that all of humanity is aligned
(00:59:53)
in our intentions. That's clearly not
(00:59:55)
the case. And and I know Yan, you in a
(00:59:58)
very interesting way argue for open
(01:00:00)
source, which some people would say is
(01:00:02)
even more dangerous because now anyone
(01:00:03)
can have access to it. It's dangerous
(01:00:05)
enough that it's in the hands of a h a
(01:00:07)
small number of people who rule
(01:00:10)
corporations, but let alone everyone
(01:00:12)
having it. Maybe that is dangerous. Um,
(01:00:14)
but again, who's us and we?
(01:00:16)
>> The danger is if we don't have open
(01:00:19)
source AI systems. Okay, in the future,
(01:00:22)
every single one of our interaction with
(01:00:25)
the digital world will be mediated by an
(01:00:27)
AI system, right? We're not going to go
(01:00:30)
to a website or a search engine or
(01:00:33)
whatever. We're just going to talk to
(01:00:34)
our AI assistant, however it's built.
(01:00:38)
Um, so our entire information diet will
(01:00:42)
come from AI systems. Now,
(01:00:45)
what does it mean to
(01:00:48)
culture, language, democracy,
(01:00:51)
um, everything if those systems come
(01:00:54)
from a handful of companies on the west
(01:00:57)
coast of the US or China?
(01:01:00)
I tell you no country in the world
(01:01:02)
outside the US and China likes the idea.
(01:01:06)
Um so we need a high diversity of AI
(01:01:10)
assistant for the same reason we need a
(01:01:12)
high diversity of the press. We cannot
(01:01:15)
afford to have just a handful of
(01:01:17)
proprietary system coming out of a small
(01:01:19)
number of companies. There's one thing
(01:01:21)
I'm scared of and that's it. Okay. If we
(01:01:24)
don't have open platforms,
(01:01:27)
um we're gonna have
(01:01:29)
uh you know capture of information flow
(01:01:33)
by a handful of companies, some of which
(01:01:35)
we may not like.
(01:01:37)
And so,
(01:01:44)
so how can we be certain that these um
(01:01:49)
when when they really are self-motivated
(01:01:51)
agents, if that ever actually happens,
(01:01:53)
that they won't collude, fight amongst
(01:01:55)
themselves, want to wrestle for power,
(01:01:58)
that we won't be sitting back watching
(01:02:01)
conflicts that we simply couldn't have
(01:02:03)
imagined before. We give them clear
(01:02:05)
objectives and we build them in such a
(01:02:06)
way that the only thing they can do is
(01:02:08)
fulfill those objectives. Now this is
(01:02:11)
not doesn't mean it's going to be
(01:02:13)
perfect but the question of AI safety in
(01:02:16)
the future I'm I'm worried about it in
(01:02:18)
the same way that I'm worried about the
(01:02:20)
question of reliability of turbo jets.
(01:02:22)
Okay. I mean turbo jets I mean it it's
(01:02:26)
amazing. I don't know about you, but and
(01:02:28)
my dad was aeronautical engineer, but
(01:02:31)
I'm totally amazed by the fact that you
(01:02:32)
can fly halfway around the world in
(01:02:34)
complete safety on a two engine
(01:02:35)
airplane. It's amazing, right? And and
(01:02:40)
we feel completely safe doing this. It's
(01:02:42)
a it's it's a magical production of uh
(01:02:46)
you know, engineering of the modern
(01:02:47)
science and engineering of the modern
(01:02:49)
world. AI safety is a problem of this
(01:02:52)
type. It's it's an engineering problem.
(01:02:54)
Um I think the fears are caused by
(01:02:57)
people who think about um you know
(01:03:01)
science fiction scenario where somewhere
(01:03:03)
someone invents the secret to super
(01:03:06)
intelligence turns on the machine and
(01:03:07)
the next second it takes over the world.
(01:03:10)
That is complete BS. Like the world
(01:03:12)
doesn't work this way. Certainly the
(01:03:14)
world of technology and science doesn't
(01:03:15)
work the world this way. The emergence
(01:03:19)
of super intelligence is not going to be
(01:03:21)
an event.
(01:03:22)
Um as we see we have super intelligent
(01:03:25)
systems that can do super intelligent
(01:03:27)
tasks you know and there is kind of
(01:03:29)
continuous progress one at a time u but
(01:03:34)
you know we're going to find some you
(01:03:36)
know better recipe to build AI systems
(01:03:38)
that may have kind of a more general
(01:03:40)
intelligence than we currently have and
(01:03:43)
and we'll have systems there's no
(01:03:44)
question that are smarter than humans
(01:03:46)
but we'll build them so that they
(01:03:48)
fulfill the goals we give them subject
(01:03:50)
to guardrails
(01:03:52)
Um, I I I was going to uh again question
(01:03:57)
this idea of we we we know that if we
(01:04:00)
can code them in a certain way, somebody
(01:04:02)
could recode them and the concept of bad
(01:04:05)
actors. But before we fall into that
(01:04:07)
hole, I have a plant in the audience.
(01:04:10)
Does my plant have a mic? Is my plant
(01:04:13)
know who he is?
(01:04:15)
>> Does my Meredith Isaac? Does my plant
(01:04:18)
have a mic? Yes.
(01:04:19)
>> He's up there. Oh, but he doesn't have
(01:04:22)
the mic.
(01:04:23)
>> Okay, David, can you shout?
(01:04:26)
>> Okay. So, [laughter]
(01:04:28)
um, so I want to introduce the, uh,
(01:04:31)
philosopher of mine, David Chalmer's.
(01:04:33)
I'm going to give you a very brief
(01:04:34)
introduction.
(01:04:38)
David, I can't see you, but I I I said
(01:04:40)
um that you could be my plant to ask a
(01:04:43)
question. Could you do you want to throw
(01:04:45)
something down here?
(01:04:46)
>> Okay, I'm over here.
(01:04:49)
>> Okay. Okay, you asked Janet asked you to
(01:04:51)
ask a question about uh AI
(01:04:53)
consciousness. Hi Adam.
(01:04:55)
>> Hi.
(01:04:55)
>> Hi.
(01:04:56)
>> Okay. So, uh you both said I think
(01:04:59)
roughly current AI systems probably not
(01:05:04)
conscious.
(01:05:05)
Future AI systems possibly descendants
(01:05:08)
of the ones today, but some future AI
(01:05:10)
systems probably will be conscious. So I
(01:05:14)
guess I want to know part one um what
(01:05:17)
requirements for consciousness do you
(01:05:19)
think current systems are lacking? Um
(01:05:23)
and then the positive side of that is um
(01:05:26)
what steps do you think we need to take
(01:05:29)
in order to develop AI systems which are
(01:05:33)
conscious and then third when is that
(01:05:36)
going to happen?
(01:05:39)
Okay, I give a crack at this. Uh, and
(01:05:42)
David already knows my answer, but
(01:05:45)
um, so first of all, I don't attribute
(01:05:48)
like I don't really know how to define
(01:05:49)
consciousness and I don't attribute much
(01:05:52)
importance to it and this is an insult
(01:05:55)
to David. I'm sorry uh because he
(01:05:58)
devoted his entire career to it.
(01:05:59)
>> Subjective experience.
(01:06:01)
>> Okay, that's a different thing. Okay,
(01:06:03)
subjective experience. Um, so clearly
(01:06:06)
we're going to have systems that have
(01:06:08)
subjective experience, that have
(01:06:11)
emotions. Emotions to some extent are an
(01:06:14)
anticipation of outcome. If we have
(01:06:16)
systems that have role models that are
(01:06:18)
capable of anticipating uh the outcome
(01:06:20)
of a situation perhaps resulting from
(01:06:23)
their actions, they're going to have
(01:06:25)
emotions because they can predict
(01:06:27)
whether something is going to end up,
(01:06:28)
you know, good or bad for, you know, in
(01:06:32)
on the way to fulfilling their
(01:06:33)
objectives, right? So, so they're going
(01:06:35)
to have all of those characteristics.
(01:06:36)
Now, I don't know how to define
(01:06:38)
consciousness in this kind of in in
(01:06:40)
this, but perhaps uh consciousness would
(01:06:43)
be the ability for the system to kind of
(01:06:46)
observe itself and configure itself to
(01:06:48)
solve a particular sub problem that it's
(01:06:51)
facing. It needs to have kind of a way
(01:06:53)
of observing itself and configuring
(01:06:55)
itself to um solve a particular problem.
(01:06:58)
We we certainly can can do this. And so
(01:07:01)
um perhaps that's what gives us the
(01:07:03)
illusion of uh of consciousness. I have
(01:07:06)
I I have no doubt this will happen at
(01:07:08)
some point.
(01:07:09)
>> And will the machines have moral worth
(01:07:10)
when it happens?
(01:07:11)
>> Yeah, absolutely. I mean they will have
(01:07:13)
some moral sense. Whether it aligns with
(01:07:17)
us or not will depend on how we define
(01:07:18)
those objectives and guardrails. Um but
(01:07:21)
yeah, they will have a sense of of
(01:07:23)
moral.
(01:07:24)
>> Let me ask Adam this question a slightly
(01:07:26)
different way or you can answer the same
(01:07:28)
question as well. Um, are we too
(01:07:32)
attached to the human subjective
(01:07:34)
experience, our sense of consciousness?
(01:07:37)
Uh, clearly we've already know that
(01:07:38)
animals don't have the same experience
(01:07:41)
that we do. And, uh, why should we
(01:07:43)
imagine that this super intelligence
(01:07:45)
will have the same subjective experience
(01:07:47)
as human beings?
(01:07:48)
>> Okay, let me answer all your questions
(01:07:50)
then. Uh, just my my gut. I I think
(01:07:54)
machines can certainly be conscious in
(01:07:55)
in in principle that if they're doing at
(01:07:58)
the you know the artificial neurons end
(01:08:00)
up doing the same information
(01:08:04)
processing in the same way as human
(01:08:06)
neurons uh then then you know the very
(01:08:08)
least that will give rise to to
(01:08:10)
consciousness. It's not about the
(01:08:11)
substrate whether it's silicon or carbon
(01:08:13)
it's just about the nature of the
(01:08:14)
information processing will give rise to
(01:08:16)
consciousness.
(01:08:18)
um what we're missing to get there. Um
(01:08:22)
as as David knows there's, you know,
(01:08:24)
there are these things called the neural
(01:08:26)
correlates of consciousness. People who
(01:08:28)
don't want to say they're studying
(01:08:29)
consciousness directly can look at human
(01:08:31)
brains or perhaps animal brains and say
(01:08:33)
what is the processes going on in the
(01:08:35)
neurons that give rise to conscious
(01:08:39)
experience. Um and uh there's a number
(01:08:43)
of number of theories and from my point
(01:08:46)
of view they all kind of suck. Um
(01:08:48)
there's there's the recurrence theory
(01:08:50)
that you need to be able to take your
(01:08:51)
outputs and plug them back in to the
(01:08:53)
inputs and that's an essential part of
(01:08:55)
consciousness. There's something called
(01:08:56)
global workspace theory, integrated
(01:08:58)
information theory. Every, you know,
(01:09:00)
physicist turned neuroscientists like to
(01:09:02)
have their own def set of criteria for
(01:09:05)
what it is for a machine for a
(01:09:07)
information processing system to be
(01:09:09)
conscious. I don't find any of them
(01:09:11)
particularly compelling and I think we
(01:09:14)
should have extreme humility about
(01:09:18)
recognizing consciousness in other
(01:09:19)
entities. We are very bad at doing it in
(01:09:23)
you know in animals. We very much
(01:09:24)
changed our mind over history whether
(01:09:26)
animals are conscious uh whether babies
(01:09:28)
experience consciousness. So my question
(01:09:31)
is a little bit don't know. Um, but
(01:09:36)
I do think that if you just told me
(01:09:39)
about neural networks or told, you know,
(01:09:42)
if I if I didn't know about
(01:09:43)
consciousness and I just heard about the
(01:09:45)
processing of information that happens
(01:09:46)
in neural neural networks, human neural
(01:09:49)
networks, I would not have predicted
(01:09:50)
that gives rise to consciousness. That's
(01:09:51)
a great surprise. Uh, and we should be
(01:09:54)
for that reason extremely humble even
(01:09:56)
about what the form of the consciousness
(01:09:58)
would make. uh to to answer Janna's
(01:10:00)
question, we have seen that what we used
(01:10:02)
to think of as a reasonably unified idea
(01:10:05)
of intelligence, human intelligence,
(01:10:06)
which is a whole bunch of different
(01:10:08)
abilities and uh and skills, we've
(01:10:11)
unbundled that with these machine
(01:10:13)
intelligences where we've constructed
(01:10:14)
things that have some of them but not
(01:10:16)
others. Very superhuman in some,
(01:10:18)
subhuman in others. Perhaps we will be
(01:10:21)
unbundling consciousness as well. And
(01:10:23)
this thing that we think of as
(01:10:24)
consciousness, we will realize that
(01:10:25)
there is uh you know many different
(01:10:28)
aspects to it that we can have some and
(01:10:30)
not the others and maybe as you
(01:10:32)
indicated we could even transcend human
(01:10:35)
consciousness in in some capacities. I'm
(01:10:38)
pretty excited about answering this
(01:10:40)
question though. I I think we finally
(01:10:43)
finally finally have a model organism
(01:10:46)
for intelligence in the form of these
(01:10:48)
artificial minds that we're building.
(01:10:50)
And maybe we can turn that model
(01:10:52)
organism for intelligence into a model
(01:10:54)
organism for consciousness and answer
(01:10:57)
some of these questions that have
(01:10:59)
intrigued mankind.
(01:11:01)
>> I just didn't think I heard an answer to
(01:11:02)
when
(01:11:03)
>> Oh, [laughter]
(01:11:04)
um I I can neither confirm nor deny, I
(01:11:07)
think, is the standard phrase we're
(01:11:08)
using here. [laughter] Um, I think if if
(01:11:11)
progress keeps going, uh,
(01:11:15)
2036
(01:11:19)
[laughter]
(01:11:19)
>> Okay. Not in the next two years.
(01:11:23)
>> Um, just one closing question. We're a
(01:11:24)
little bit over time, but I'm going to
(01:11:26)
ask this to you, Yon. Uh, in many ways,
(01:11:29)
you're a contrarian. Maybe not by
(01:11:32)
choice. Maybe this is just how it's
(01:11:33)
happened. You've called it the cult of
(01:11:34)
LLMs. you you sort of often refer to the
(01:11:37)
fact that in Silicon Valley you're don't
(01:11:38)
have the most conventional approach. Um
(01:11:41)
but yet you have an optimism
(01:11:43)
[clears throat]
(01:11:44)
you you really do not indulge in the
(01:11:46)
doomsday sort of rhetoric. Uh what is
(01:11:49)
your most optimistic vision for if not
(01:11:52)
two years from now 2036?
(01:11:56)
Well, the new renaissance that's a
(01:11:58)
pretty optimistic uh view of you know AI
(01:12:01)
systems that amplify human intelligence
(01:12:04)
is under our control can solve uh a lot
(01:12:07)
of complex problems can accelerate the
(01:12:08)
progress of science and medicine can uh
(01:12:12)
educate our children um you know help us
(01:12:16)
uh
(01:12:18)
you know process all the information or
(01:12:21)
bring us uh all the knowledge and
(01:12:23)
information that we need to to see. Uh
(01:12:26)
in fact you know people have been
(01:12:29)
interacting with AI systems for much
(01:12:30)
longer than they realized. Um of course
(01:12:32)
there is you know L&M and chatbots now
(01:12:35)
for the last three years. Uh but before
(01:12:37)
that um you know most like every car
(01:12:42)
sold in in in the EU and most cars sold
(01:12:46)
in in the US have uh what's called ADAS
(01:12:50)
advanced uh driving assistance systems
(01:12:52)
or automatic emergency braking systems.
(01:12:55)
You know a camera that looks out the
(01:12:56)
window and stops your car if you are
(01:12:58)
about to hit a pedestrian or another
(01:13:00)
car. Um it saves lives. Um you you get
(01:13:05)
an X-ray today. let's say a mamogram or
(01:13:08)
something, you know, at the bottom it
(01:13:10)
says the thing has been reviewed by an
(01:13:12)
AI system. It saves lives. Um, you can
(01:13:15)
get an MRI now, full body MRI in 40
(01:13:18)
minutes. Um, this is because you can
(01:13:21)
accelerate the process of collecting the
(01:13:22)
data because AI systems can can sort of
(01:13:25)
fill in the blanks. You don't need to
(01:13:26)
collect that much data for this. Um but
(01:13:29)
also all the news you're seeing whether
(01:13:32)
you connect on Google or you know
(01:13:35)
Facebook, Instagram, any social network
(01:13:37)
is determined by an AI system that
(01:13:39)
basically caters to your uh interest. Um
(01:13:43)
and so you know AI has been with us for
(01:13:46)
for a while already.
(01:13:48)
>> But you're saying we should be impressed
(01:13:49)
when they can pour a glass of water and
(01:13:51)
do our dishes.
(01:13:52)
>> Pour a glass of water, do our dishes. um
(01:13:54)
you know uh drive our cars uh like learn
(01:13:58)
to drive our cars in 10 hours
(01:13:59)
>> without the cheating
(01:14:00)
>> practice without all the cheating with
(01:14:02)
with sensors and mapping and and and
(01:14:05)
>> and hard coding of rules. So um yeah
(01:14:08)
this is going to take a while u but this
(01:14:11)
is going to be the next revolution of
(01:14:13)
AI. So this is what I'm working on.
(01:14:15)
Okay. Um, and the the message I've been,
(01:14:18)
you know, carrying for a while now is um
(01:14:21)
is okay, LM are great, they're useful,
(01:14:24)
we should invest in them. Um, a lot of
(01:14:27)
people are going to use them. They are
(01:14:29)
not a path to human level intelligence.
(01:14:32)
They're just not. Uh, right now they are
(01:14:35)
sucking the air out of the womb anywhere
(01:14:37)
they go. And so there's basically no
(01:14:39)
resource left for anything else. Um
(01:14:43)
and so for the next revolution we need
(01:14:46)
to kind of you know take a step back and
(01:14:49)
figure out what's missing from um the
(01:14:52)
current approaches and then I've been
(01:14:55)
making proposals on this and working
(01:14:58)
inside of uh meta for a number of years
(01:15:00)
on this uh alternative approach. It's uh
(01:15:04)
come to the point where um you know we
(01:15:07)
need to kind of accelerate this this
(01:15:09)
progress now because we know it works.
(01:15:11)
We have early results and so um that's
(01:15:15)
the plan.
(01:15:17)
>> Okay. I could we could have a whole
(01:15:18)
another hour starting right here. But um
(01:15:21)
I hope you'll all join me in thanking
(01:15:24)
our guests for an incredible
(01:15:26)
conversation. Thank [music] you so much.
(01:15:28)
[applause]
