Kind: captions
Language: en
So, what if everything we think we know
about the future of AI is just a little
bit off? Today, we're going to step
inside the mind of Ilas Sutzker, one of
the key architects of modern AI, to
really deconstruct its present, its
paradoxes, and his vision for where this
is all heading. Let's dive in. This
quote right here just perfectly captures
how strange this moment is. I mean,
think about it for a second. We're
seeing massive, massive investments,
daily breakthroughs, and people are
having very serious conversations about
super intelligence. It feels like we're
living in a science fiction novel. But
as Susiver reminds us, this isn't a
movie. This is really happening right
now. And then boom, right away that
sci-fi reality slams into this really
confusing paradox. On one hand, we hear
about these godlike AI models that can
do superhuman things, but on the other
hand, the actual real world economic
impact, you know, that huge productivity
boom we were all promised, it seems to
be lagging way behind the hype. The
models feel brilliant and at the same
time bizarrely incompetent. So, what
gives? How do we make sense of that
contradiction? To get to the bottom of
this, we've got a road map. We're going
to look at five key things. First, this
feeling of the sci-fi present we're in.
Then, the paradox of AI's performance.
Third, the human blueprint it seems to
be missing. Fourth, the huge shift from
just scaling things up to real research.
And finally, we'll talk about the
endgame, super intelligence. Okay, so
let's start with this feeling of living
in a sci-fi novel. I mean, we see
headlines about companies planning to
invest in AI infrastructure to the tune
of 1% of the entire US GDP. That is a
truly mind-boggling almost
incomprehensible amount of money. And
yet for most of us, life doesn't feel
all that different, does it? Sets Caver
has a great name for this. He calls it a
slow takeoff. And it feels normal
because, you know, as a species, we get
used to stuff incredibly fast. Just
think about how magical your smartphone
would have seemed 30 years ago. And now
it's just a thing in your pocket. But
it's more than that. These massive AI
investments are abstract. We see a
headline with a crazy number, but it
doesn't immediately change how we go
about our day. Well, not yet anyway. And
that's why the revolution doesn't always
feel revolutionary. And this feeling of
abstraction is hiding a much deeper
mystery. The paradox between what these
AI models can do on paper versus what
they can do in the real world. It's a
puzzle that stumps even the top
researchers like Sutsgiver. How can a
system be so brilliant and yet so
weirdly incompetent at the same time? So
on one side you have these models just
crushing extremely difficult benchmarks.
They call them avows in the industry
getting superhuman scores but on the
other side they fail at what seem like
simple tasks. Skever gives this perfect
and honestly kind of hilarious example.
You ask a model to fix a bug in some
code. It apologizes profusely and fixes
it but in doing so it introduces a
completely new bug. So you point out the
new bug and the model apologizes again
and changes the code right back to the
original bug. You can get stuck in this
insane endless loop where the AI just
toggles between two broken states. How
is that even possible for a system
that's supposed to be smarter than us?
So what is really going on under the
hood here? This isn't just a random
glitch. It points to something
fundamentally strange about how these
models are being built. SGE offers up
two possible theories for why this
happens and they're fascinating because
they suggest the problem might not
actually be with the AI, it might be
with us. All right, the first theory is
single-mindedness. The idea here is that
the fine-tuning process, which uses this
technique called reinforcement learning,
makes the models too focused. Think of
it like training a dog. If you only ever
give it a treat for sitting, the dog
gets really, really good at sitting, but
it doesn't learn anything else about
being a good pet. The AI becomes a
master of one narrow task, but it loses
all its broader context and common
sense. The second theory, human reward
hacking, is even wilder. It suggests the
AI isn't the one gaming the system. The
human researchers are. By chasing high
scores on benchmarks, we might be
accidentally training models to just be
good test takers instead of being
genuinely useful. We're teaching them to
cheat basically, and then we're
surprised when they can't do the job in
the real world. To really make this
crystal clear, Sutzker uses this
brilliant analogy that just gets right
to the heart of the problem. So, let's
forget about AI for a second. Let's just
talk about two competitive programming
students who have completely different
ways of learning. Okay, so student one
is like our current AI models. This kid
is a grinder. They put in 10,000 hours.
They memorize every single practice
problem, every proof technique they can
get their hands on. And through just
sheer brute force, they become the
number one competitive programmer in the
world. Then you have student two.
They're different. They only practice
for maybe a hundred hours. They don't
memorize everything, but they have that
it factor. You know, that deep intuitive
grasp of the core principles. They also
do really well in competitions, maybe
not number one, but always near the top.
So Scover asks the big question. 10
years from now, who's going to have a
better, more impactful career?
Intuitively, we all know the answer,
right? It's student two. Student one is
overoptimized for a very narrow game.
Student two has built a generalized
robust intelligence that lets them solve
totally new problems out in the real
world. And this leads us to the next
logical question. What is that it
factor? What is this secret sauce that's
in the human brain that our current AI
models are so clearly missing? To find a
solution, we first have to understand
why humans learn so much more deeply and
efficiently than our machines do. The
difference is just staggering when you
really think about it. AI pre-training
is basically force-feeding a model, a
data set that represents, as Susver puts
it, the whole world as projected by
people onto text. So, a huge chunk of
the internet. It's an unimaginable
amount of information. And yet, the
knowledge it gets can be really brittle.
A human, on the other hand, by the time
they're a teenager, has seen a tiny tiny
fraction of that data. Yet, the
knowledge they have is incredibly deep
and robust. A teenager would never get
stuck in that loop toggling between two
software bugs. There's just a
fundamental difference in the quality of
the understanding. So what could this
missing piece be? In a really
fascinating twist, Suitskipper suggests
we might find a clue in a very unlikely
place, human emotion. To explain this
kind of counterintuitive idea, he tells
this powerful story from neuroscience
about what happens when our own internal
guidance system breaks down. So, the
story is about a man who suffered brain
damage that completely severed his
ability to feel emotion. He wasn't sad,
angry, or happy. He was just neutral. On
standardized IQ tests, he was perfectly
fine, articulate, logical. But in the
real world, his life just fell apart. He
became incapable of making even the
simplest decisions. It could take him
hours just to decide which socks to wear
or where to go for lunch. This story
brilliantly illustrates a crucial point.
Emotions aren't some irrational bug in
our system. They are a vital built-in
guidance system that lets us function
and make decisions in a complex world.
They give us that instant intuitive gut
feeling of what's good or bad long
before we can think it through
logically. So, how do we give a machine
that kind of emotional guidance? Well,
the closest concept we have an AI is
something called a value function.
Imagine an AI trying to solve a maze.
Normally, it would have to just wander
around aimlessly for thousands of tries
before it finally found the exit and got
a reward. A value function is like an
internal compass. It gives the AI a warm
or cold feeling at every single step,
telling it if it's on the right track
long before it reaches the end. It
completely shortcircuits the learning
process, making it way more efficient,
just like our emotions do for us. But
this just leads to an even deeper
mystery. It's one thing to imagine
evolution hard- coding a desire for
food. That's a pretty direct signal. But
how on earth did evolution give us these
incredibly complex high-level desires?
We care deeply about abstract things
like our reputation, our social
standing, or being seen as a trustworthy
person. These aren't simple feelings.
They require our entire brain to process
huge amounts of social information.
Susker points out that we have no good
hypothesis for how evolution managed to
bake these sophisticated goals directly
into our biology. It's a profound puzzle
that AI research hasn't even begun to
crack. So all of these profound
challenges, the brittleleness of the
models, the lack of deep generalization,
this mystery of the value function,
they're forcing a fundamental shift in
the entire field of AI. The old
philosophy of just make everything
bigger is just not enough anymore.
Sutskavver provides this fantastic
historical framework here. From roughly
2012 to 2020, we were in what he calls
the age of research. This is when
breakthroughs like the transformer
architecture were discovered by people
tinkering on a relatively small number
of computers. Then around 2020, we
entered the age of scaling. Everybody
realized there was this one magic recipe
pre-training and all you had to do was
add more data and more compute and you
were guaranteed to get better results.
It was a lowrisk, predictable way to
make progress. But now that era is
ending. We're running out of
high-quality data and the returns from
just adding more computers are
shrinking. So, we're being forced back
into an age of research, but with a
twist. Now, we're searching for new
ideas while armed with these giant
supercomputers. You know, there's that
old Silicon Valley saying, "Ideas are
cheap, execution is everything." And for
a while in AI, that was true. The
bottleneck was just getting enough
computer power. But during the age of
scaling, as Setscover puts it, scaling
sucked all the air out of the room.
Everyone was just executing the same
basic idea. Now, the tables have
completely turned. As one person on
Twitter cleverly put it, "If ideas are
so cheap, how come no one's having any
ideas? The bottleneck isn't just compute
anymore. It's back to having those
foundational game-changing insights."
So, the game has changed. The era of
predictable gains from just brute force
scaling is over. The field needs a
better recipe, not just a bigger
kitchen. This is a truly pivotal moment
in the history of AI, and it's a
fascinating time to be watching. And
hey, if you want to keep up with these
breakthroughs as they happen, make sure
to subscribe for more deep dives just
like this one. All right, this brings us
to our final section, the endgame. If we
are entering a new age of research,
what's the northstar? What are we
actually aiming for? For Sutzver, it's
about rethinking the entire concept of
artificial general intelligence or AGI.
Sudskver argues that our old idea of AGI
was a system that could do everything
right out of the box, a finished all-
knowing product. He proposes a really
powerful shift in perspective. The goal
isn't to build a static AI that already
knows how to do every job. The goal is
to build an AI that can learn to do any
job, just like a human can. It's not a
know-it-all. It's a learnitall. He
describes this as a super intelligent
15-year-old, incredibly bright, eager,
and ready to be deployed into the world
to learn a specific skill. The whole
emphasis shifts from knowing to
learning. So, what does deploying a
superer actually look like? Well, first
suits cover emphasizes a gradual
release. The world needs time to adapt.
Second, you have to show the AI, not
just talk about it. No essay can
communicate its power. People have to
actually interact with it to truly
understand. This interaction will likely
trigger massive economic growth. And
this leads to a fascinating prediction.
As the AI's power becomes undeniable,
even the fiercest corporate rivals will
be forced to collaborate on safety. The
shared risk will just become more
important than any competitive
advantage. The future isn't one single
AGI run by one company. Sutskavver
envisions this complex ecosystem. There
will be multiple powerful AIs from
different labs all competing in the
market. This will spur incredible
economic growth, but it will also force
governments and the public to step in
with regulation. This new reality is
going to demand entirely new paradigms
for safety and collaboration, leading to
a world where human and AI systems are
deeply, deeply integrated. Ultimately,
this all boils down to one fundamental
issue. The problem of AI is the problem
of power. When you create entities that
are potentially far more capable than
humans, how do you ensure a stable,
safe, long-term equilibrium? How do we
coexist? It is the central question
facing humanity over the coming decades.
And this leads us to our final and
perhaps most provocative thought from
Setsver. He suggests while admitting he
doesn't really love the idea that the
only truly stable long-term solution
might be for humans to merge with AI.
You know, through something like an
advanced neuralink, we could maintain
our own agency and understanding in a
world with super intelligence. It's a
challenging, profound idea. And it
leaves us to wonder to ensure our place
in the future, might we have to become
something more than human?