Kind: captions
Language: en
Welcome to the explainer. Today we are
diving head first into one of the
biggest, most fundamental debates
happening in artificial intelligence
right now. And it all boils down to a
really simple question. Does AGI,
artificial general intelligence, the
holy grail of the entire field, even
exist? Now, this isn't just some
abstract philosophical argument. This is
a real deal clash between two absolute
titans of AI. We're talking the head of
AI at Meta versus the CEO of Google
DeepMind. and how they answer this
question could literally change the
future for all of us. So, let's break
this down. So, it all kicked off with
this this one sentence from Yan Lun who
is, you know, one of the original
godfathers of AI and it's an absolute
bombshell. He didn't just say AGI is far
away. He said the very concept, the
thing everyone's racing towards doesn't
exist. Now, you're probably thinking,
wait, what? How can one of the key
architects of our AI future say that its
ultimate goal is basically a fantasy?
It's such a wild counterintuitive
statement that it forces you to stop and
rethink everything you thought you knew
about what intelligence even means. And
yeah, that's the billion dollar
question, isn't it? What is Yan Lun
really getting at here? And why does
Deis Sabis, the guy running Google
DeepMind, think he's, and this is a
direct quote, plain incorrect? This is
so much more than a war of words. It's a
fundamental disagreement about the path
we should be on. Is the future of AI
some single godlike all- knowing system?
Or is it something completely different
like a whole society of brilliant but
specialized AIs? The answer to that
question changes everything. Okay, so to
really get to the bottom of this, here's
our game plan. First, we're going to set
the stage and introduce you to the two
heavyweights in this intellectual boxing
match. Then, we're diving deep, and I
mean deep, into Yan Lun's case for why
everything is specialized. After that,
we'll unpack the incredibly powerful
rebuttal from Dimisabus before we circle
back to see how Lun responds. And
finally, we'll tie it all together to
figure out what this whole debate really
tells us about the search for
intelligence. So, let's properly meet
our two main characters. In one corner,
you have Yan Lun. He's the chief AI
scientist at Meta, a Turing Award
winner, which is like the Nobel Prize
for computing. and his work on
convolutional neural networks is
literally the reason computers can see
and recognize things in images today.
And in the other corner, you've got Deis
Hosabus, the co-founder and CEO of
Google DeepMind. This is the lab that
gave us Alph Go, the AI that did the
impossible and beat the world's best Go
player. So yeah, these aren't just
commentators. These are two of the
people actively building our future. And
they see the end goal in completely
different ways. So from a 30,000 ft
view, here's how their arguments stack
up. Luna is basically saying the whole
idea of general intelligence is a
mirage. It's a complete illusion. He
believes that every single form of
intelligence we've ever seen, including
our own, is incredibly profoundly
specialized. Hacabus, on the other hand,
comes back and says, "No, no, no. You're
confusing general intelligence with
something totally different, a
theoretical universal intelligence." He
thinks the human brain's amazing
flexibility is all the proof we need
that true generality is very, very real.
So, let's dig into Lun's argument first,
cuz it's a fascinating place to start.
All right. So, the absolute core of Yan
Lun's argument is this really
provocative idea. What we humans call
our general intelligence is just an
illusion. We walk around thinking we're
these amazing all-purpose problem
solvers, right? But in his view, we're
actually one of the most specialized
species on this planet. We've just
gotten so used to our own little
backyard of problems that we've mistaken
it for the entire universe of
possibility. So, here's Lacun's logic.
Our brains aren't some magic all-purpose
machine. Far from it. They are a toolkit
that has been fine-tuned over millions
of years for a very, very specific job,
surviving on the African savannah.
That's why we are absolutely brilliant
at certain things. We can navigate a
complex 3D world without even thinking
about it. We can instinctively predict
how a thrown object will travel. And
maybe most importantly, we are masters
at understanding the incredibly complex
social cues of other humans. But the
moment you step outside that
evolutionary box, our supposed
generality just completely falls apart.
He points out that we are objectively
laughably bad at chess compared to an AI
like Stockfish. Most of us are terrible
at complex mathematics or highfrequency
stock trading. Lun's big point is that
we're living inside our own bubble. We
think we're general because we can solve
all the problems we happen to encounter,
but the problems we encounter are by
definition the ones our specialized
brains are built to handle. And there's
this perfect realworld example that just
drives this point home. You may have
seen this video of a chimpanzee at a
research institute in Japan. The chimp
sits in front of a touchcreen and
numbers 1 through nine flash on the
screen for just a fraction of a second
before being hidden by white squares.
And then without even pausing, the chimp
just flawlessly taps the squares in the
right order. 1 2 3 4. It's an incredible
display of short-term photographic
memory. And you know what? Most humans
who try it are absolutely terrible. We
just can't keep up. Lacun uses this to
show that intelligence isn't some single
ladder with humans sitting at the top.
It's more like a huge mountain range
with lots of different peaks. And we're
on one peak and the chimp is on another.
We're not better overall. We're just
different. We are specialists just like
every other creature out there. So, this
is the big takeaway from Lacun's side of
things. Our feeling of being these
generalpurpose thinkers is basically a
cognitive bias. We've mistaken our very
particular specialized toolkit for a
universal one. For him, trying to build
an AGI is like trying to create a
generally athletic animal. I mean, what
does that even mean? Is it as fast as a
cheetah, as strong as a gorilla, as good
a swimmer as a dolphin? No. Nature
doesn't build generalists. It builds
specialists. It's a really powerful
argument, but Damus Hassaba says it's
completely wrong. That it's all based on
a fundamental misunderstanding. His
counterargument flips this whole thing
on its head. and we are getting into it
right after this. And hey, if you're
finding this deep dive valuable, now is
the perfect time to subscribe so you
don't miss our future explainers. Okay,
so let's switch gears and look at Dimma
Assabus' rebuttal. And honestly, it is
surgical. It's incredibly precise. He
basically argues that Yan Lun isn't
really arguing against general
intelligence at all. He says Lun has
built up a straw man, a completely
different impossible concept that he
calls universal intelligence. And
understanding this one distinction is
the absolute key to his entire argument.
This slide right here lays it all out.
On the left, we've got general
intelligence. Habis defines this as a
flexible adaptive learning architecture.
The magic is that it can learn a huge
range of things it wasn't specifically
programmed to do. And his star example,
of course, is the human brain. Then on
the right, you have universal
intelligence. This is a hypothetical
basically godlike system that could
solve any conceivable problem and solve
it perfectly optimally every time.
Hassabis is quick to point out that no
serious AI researcher actually believes
this is possible. It's a fantasy. So his
argument is this. When Lacun says, look,
humans have limits and weaknesses. He's
not disproving general intelligence.
He's just stating the obvious fact that
we don't have universal intelligence.
And to back this all up, Hosibus brings
in a really fundamental concept from
computer science called the no free
lunch theorem. Now, this is a real
mathematical proof and it basically says
that there is no single master algorithm
that is the best at solving all possible
problems. It's a fundamental law of
trade-offs in the universe. If you
design a system to be amazing at task A,
you are by mathematical necessity making
it worse at some other tasks. It's a
guarantee that specialization has to
happen in any system that actually
exists in the real world. The flying car
analogy makes this crystal clear, right?
Think about it. You can build an amazing
car optimized for the road. It's fast.
It's efficient. It corners perfectly. Or
you can build an amazing plane optimized
for the sky. But if you try to build one
single vehicle that does both, what do
you get? You get a terrible car and a
terrible plane. It's inevitable. You
simply cannot escape the trade-offs. The
no free lunch theorem is just the
mathematical universal version of that
simple truth. So, here is where Habis
pulls off this incredible intellectual
judo move. He completely turns Lun's
argument against itself. He starts by
saying, "You're right, Yan. The no free
lunch theorem is real, and that means
any real world system like our brains
must have some specialization." But, and
this is the absolute crucial point. He
says that specialization is not proof
that the underlying architecture isn't
general. In fact, it's the opposite. He
argues that specialization is the
natural consequence of a general
learning system adapting to a specific
environment with limited time and
energy. And this leads him to this
beautiful conclusion. The human brain is
a system that is specialized towards
generality. It's a general purpose
learning machine that has simply become
an expert on the problem set of being a
human. But Hassimus isn't done. He then
brings out what you could call his heavy
artillery, the touring machine argument.
Now, for anyone who's not familiar, a
touring machine is a theoretical concept
from Alan Turing back in the 1930s. It's
a simple abstract model of a computer,
but it's incredibly powerful. It can in
theory compute anything that is
computable as long as you give it enough
time and memory. It's the foundational
idea behind every single computer,
laptop, and smartphone on Earth.
Habibus' claim here is just profound. He
argues that the human brain and also
today's big AI models like GPT are what
he calls approximate touring machines.
What that means is at their very core,
they possess a universal generalpurpose
computational architecture. They are
fundamentally capable of learning pretty
much anything that can be computed. So
all the limits we see, the reason we
can't solve every problem instantly,
those aren't flaws in the design. They
are just practical real world
constraints. We have finite memory. We
have finite processing speed. We have a
finite amount of data to learn from. But
the generality, he insists, is baked
right into the architecture itself. And
he brilliantly brings this all back to
Lun's own example of chess. Lun says our
being bad at chess proves we're
specialized. Hosibus fires back that the
very fact that a brain which evolved to
throw spears and find berries can even
invent and play a game as abstract and
logical as chess is the ultimate proof
of its generality. That incredible
ability to repurpose a survival machine
for pure abstract thought is for Hospice
the very definition of general
intelligence. It's a really powerful
point. But Lun is still not buying it.
And his response takes us from the clean
theoretical world of computer science
into the absolutely mind-bending scale
of physical reality. And if you want to
keep going down this rabbit hole with
us, you know what to do. Hit that
subscribe button. So, how in the world
does Yan Lun respond to that, to this
incredibly elegant Turing machine
argument? Well, what's so interesting is
that he basically agrees with the
theoretical point, but then he argues
that it's completely irrelevant in the
real world. His rebuttal shifts the
entire debate. It's no longer about
computer science theory. It's about the
staggering, almost impossible to
comprehend scale of reality itself and
our tiny, tiny place in it. Lun comes
back and essentially says, "Look, I
think we're just arguing about words
here." When Habisas calls the brain
general, Lun has a problem with that
word because he feels it implies a
scope, a capability that is just wildly
misleading. He says, "Sure, okay, in
some abstract theoretical sense, a human
brain is a touring machine. But in
practice, the real world constraints are
so profound, so astronomically huge that
to use the label general is to
completely miss the point of what's
actually going on." And to make this
tangible, he uses the example of our own
vision. Think about it. Your optic
nerve, the cable connecting your eye to
your brain, has about a million little
fibers sending signals. You can think of
each fiber as a simple onoff switch. So,
Lon asks a really interesting question.
How many different possible ways could a
brain be wired up to process those 1
million inputs? And the answer is, well,
it just breaks your brain. The number of
possible ways to process that visual
data is 2 to the power of two to the
power of a million. It's a number so
ridiculously large it's basically
meaningless. We're talking about a 10
followed by roughly 300 trillion zeros.
Just to give you some perspective, the
total number of atoms in the entire
observable universe is a 10 with about
80 zeros. Lacun's point here is that the
space of all possible ways to see is
nearly infinite. And our brain, even
with its 100red trillion connections,
can only explore a tiny infinite decimal
fraction of that space. It's like a
single grain of sand on a beach the size
of the entire cosmos. And this leads him
to a point that Albert Einstein once
made. The real miracle isn't that the
universe is so complex. The miracle is
that we can understand any of it at all.
And according to Lun, the only reason we
can comprehend anything is because our
brains have evolved to be hypers
specialized for the very tiny, very
structured, very non-random slice of
reality that we happen to live in. So
this is the final crucial point for Lon.
Most of the universe, most of that vast
computational space of possibilities
would look like pure random static noise
to us. We have a word for it. We call it
entropy. and we just ignore it because
our brains literally aren't equipped to
process it. We live on this tiny little
island of understandable patterns
surrounded by an infinite ocean of what
is to us total chaos. So from his
perspective, to call our intelligence
general when it's confined to this one
tiny island is an act of incredible
arrogance. The constraints aren't just a
minor detail. They are the single most
defining feature of our intelligence.
Okay, so we've been through the
arguments, the rebuttals, the counter
rebuttals. We have Lun saying we're
hypers specialized creatures living in a
world that's just too vast for us to
truly comprehend. And we have Hosabus
saying no, we have a general
architecture that's just constrained by
the laws of physics. So what's really
going on here? Where is the real
disagreement? It turns out it's not
really about the facts. It's a deep
philosophical difference in perspective
about what matters more, the theoretical
potential of a system or its practical
real world limits. I think this analogy
just captures the whole debate
perfectly. Imagine a Swiss Army knife.
Deus Hosabibus looks at it and says,
"Wow, look at this incredible
generalpurpose tool. It has a knife. It
has a corkcrew. It has scissors. Its
amazing flexibility to tackle all these
different problems is its greatest
strength." Yan Lun looks at that exact
same knife and says, "This is a highly
specialized tool. Compared to a
professional chef's knife or a dedicated
corkcrew or a pair of industrial shears,
it's pretty mediocre at everything. And
if you compare it to the infinite set of
all possible tools that could exist,
like a plasma cutter or a molecular
assembler, it's practically useless.
See, they're both looking at the exact
same object, but they are framing it in
a completely different context. And this
final table really brings it all home.
Deis Hassabis is focused on the
potential of the architecture. He sees a
touring machine that in theory has the
potential to learn anything. So for him,
the answer is yes, it's general. It just
has some practical constraints. Yan Lun
is focused on the practical limits
themselves. He sees those constraints as
being so massive, so overwhelming that
they completely negate the label
general. For him, the answer is no,
because calling it general is just
misleading. Hassabis sees a master key
that could theoretically open any lock,
even if it's a bit clunky. Lun sees one
very specific key that just happens to
fit the one tiny slice of reality we can
actually perceive. And so, that is the
question that we want to leave you with.
Is our brain a master key? A truly
general architecture capable of adapting
to anything from quantum physics to
playing chess? Or is it just one very
specific, very intricate key that
happens to be a perfect fit for the only
lock we've ever known? The answer to
that question isn't just a war of words
between two of AI's most important
thinkers. It will define the entire
direction of our quest to build truly
intelligent machines.