Kind: captions
Language: en
If you follow tech at all, or you know,
frankly, if you just exist in the world
today, it is absolutely impossible to
escape the constant drum beat of AI
progress. I mean, every single day,
right, there's another headline, another
mind-blowing demo that just seems to
defy what's possible. It can honestly
feel like we're all strapped into this
unstoppable high-speed train just
rocketing towards a future where AI does
everything. The pace is, it's just
dizzying. And it's not just the
headlines. We're hearing these
incredibly bold claims from the biggest
names in the game. You've got Mark
Zuckerberg saying meta will have AI that
can do the work of mid-level engineers
in like a year. Tech leaders,
influencers, they're all saying every
new model means AGI, artificial general
intelligence is right around the corner.
We're talking months, not years. It all
feeds this feeling that human level AI
isn't some sci-fi dream anymore. It's an
imminent reality. And what really
cements this idea, what makes it so hard
to argue with are the benchmarks. These
are the hard data points. First, it was
passing the bar exam, then the SATs,
then suddenly it was acing coding
challenges that would take a human
engineer hours to solve. And now, now
we're seeing these models solve PhD
level math and physics problems that
would stump most human experts. I mean,
when you see a machine ASA test designed
to measure the absolute peak of human
intellect, you can't help but ask, are
we about to be replaced? But what if
that whole story, the unstoppable train,
AGI being just around the corner, is
just fundamentally wrong? What if the
very yard stick we're using to measure
all this progress is completely flawed?
Let's really dig into that. Because
according to AI researcher Aiden Meyer,
whose work we're diving into today, this
whole idea of an all- knowing AGI isn't
just hype. He says it's a lie. It's
A fundamental misdirection
that's leading the entire field and all
of us down the completely wrong path.
So, how on earth could that be true with
everything we're seeing? Well, in this
deep dive, we're going to break down
this really powerful counterintuitive
argument. We're going to start by
deconstructing that myth of unstoppable
progress and show why we might be
measuring the wrong thing entirely.
Then, we'll explore a new and I think
much more exciting vision for what AGI
could be. unpack the three essential
keys to get there and finally rethink
what the future of AI should even look
like. Okay, so this brings us right to
the core of the whole argument. We are
measuring the wrong thing. How can an AI
ace all these impossible exams and still
not be on the path to real general
intelligence? Well, it all comes down to
a basic misunderstanding of what we're
actually testing. Our benchmarks from
the LSAT to coding competitions are
missing half of the equation. This slide
splits the very idea of intelligence in
two. And this is so key. On one side,
you have applying knowledge. This is all
about performance on a fixed task. It's
what every single AI benchmark today
tests. Can you answer this question? Can
you solve this problem? Can you pass
this exam? And yeah, our current models
are amazing at this. But on the other
side is the much much more important
skill, acquiring knowledge. the ability
to learn something genuinely new, to
adapt, to grow. And that is the part we
completely ignore. And look, this isn't
some new definition of the word. Just
look up intelligence in the dictionary.
It's the ability to acquire and apply
knowledge and skills. It's two parts,
both equally important. By focusing only
on the apply part, we've basically
created the world's most sophisticated
parrots. They are incredible at
repeating and remixing the vast ocean of
human knowledge they were trained on,
but they're almost completely incapable
of learning anything new after that
training is done. Okay, so if the goal
of building this all- knowing digital
oracle is basically a dead end, what's
the alternative? Well, the vision Meer
proposes is a pretty radical shift.
Instead of building an agent that knows
everything, he says we should be
building an agent that can learn
anything, a lifelong learner. So, what
does that actually look like? This new
AGI wouldn't be some static database you
just ask questions. It would be a
dynamic agent. It would have its own
goals and it would decide what it needs
to pay attention to in order to achieve
them. It would teach itself, learn from
interacting with you and then use that
new knowledge to keep learning on its
own. So think less of a tool and more of
a partner that you actually grow with.
And this is a perfect example of the
realworld limits of what we have now.
Meyer talks about asking ChatGpt for
ideas for his YouTube channel, and the
suggestions are always just terrible.
They're bland. They're generic, and they
have nothing to do with his unique
style. And that's because the AI can't
actually learn him. It can't look at
their past conversations, get his
context, and create something new. It
can only give him a remix of what's
already popular on YouTube from its
training data. It can apply knowledge
about YouTube video ideas, but it can't
acquire new knowledge about Eden Meyers
video ideas. You know what's wild? This
whole idea of focusing on learning isn't
some futuristic concept. It's actually
going back to the very roots of AI.
Early research way back in the 30s and
50s was obsessed with creating systems
that could learn and adapt, like a
little robotic rat figuring out a maze
through trial and error. The goal was to
copy how animals learn continuously.
It's only in our modern era of big data
and giant computers that we've shifted
to this train once on the entire
internet then deploy a frozen model
approach. We've traded the ability to
learn for the ability to know. Okay, so
if we accept this new vision of AGI as a
lifelong learner, how do we actually
build it? Well, Meyer lays out three
critical and mostly neglected areas of
research that are absolutely essential
to make this happen. These are the three
keys that could unlock a very different
and honestly much more powerful future
for AI. And hey, if you're finding this
kind of deep dive valuable, make sure
you're subscribed for more explainers
that cut through the hype just like this
one. The first key is the most direct
one, continual learning. It's simple,
really. The agent should never stop
learning. Right now, AI models have two
separate phases. A massive offline
training phase and then an online
deployment phase where its knowledge is
basically frozen in time. Continual
learning says those should be the same
thing. Learning and doing should happen
all at once forever. Now, you might be
thinking, hold on, but current models
can learn through things like
fine-tuning or in context learning. But
these are more like band-aids than real
solutions. When you try to fine-tune a
model with new data, you run into this
huge problem called catastrophic
forgetting. So, imagine you teach a
model to be an expert on Shakespeare.
Then you try to teach it about Jane
Austin. In the process, it will
literally overwrite and forget
everything it knew about Shakespeare.
It's not adding knowledge, it's
replacing it. And over time, this also
leads to a loss of plasticity where the
model just gets worse and worse at
learning anything new. As for in context
learning, that's just cramming for a
test. The information disappears the
second the conversation moves on. These
are not paths to lifelong learning. All
right, this brings us to key number two,
which is a little more abstract, but
just as important for building something
that thinks like a human. To really
learn about the world, an AI has to
experience it not as a jumbled mess of
data, but as a single continuous,
unbroken stream, one moment flowing into
the next, creating a personal timeline
of cause and effect. Now, let's contrast
that with how we actually train large
language models today. It is literally
the exact opposite. We scrape billions
of random disconnected bits of text from
all over the internet. A paragraph from
a poem, a form post about fixing a car,
a cooking show transcript. Then we toss
all of it into a giant digital blender,
shuffle it up, and feed it to the model
in these random huge batches. There is
no timeline, no narrative, no journey of
discovery. In training this way has huge
consequences. It's why a concept like
episodic memory, your memory of the
events in your own life, is basically
meaningless for an AI. To understand why
that matters, Meer explains how the very
idea for his research came to be. It
wasn't a generic prompt. It was his
unique single stream of experiences. A
specific interest he had led to a
certain grad program which led to
specific conversations which were
influenced by a startup he worked on
years ago. That chain of events, his
personal episodic memory, created
something new. An AI trained on a random
batch of the internet could never come
up with that because it has no personal
history to pull. And that brings us to
the third and final key. This is all
about changing how our models get
smarter. Right now, they get smarter
with more and more data. What we need
are algorithms that get better with more
compute. Basically, more thinking time.
What's so interesting here is the
difference in philosophy. Current
scaling only works if you have more
compute and way more data. It's
incredibly inefficient. I mean, we've
scripted basically the whole public
internet and big tech companies are
already worried about running out of new
data. That's not a data problem. It's an
algorithm problem. The real goal should
be algorithms that scale with compute
alone. Think of it like this. Current AI
is a student who can only get a better
grade by reading more textbooks. The
goal is a student who can get a better
grade by thinking more deeply about the
same textbook. Given more processing
power, this ideal agent could pull more
wisdom and new ideas from the exact same
set of experiences. It's the difference
between thinking deeper, not just
reading fast. So, when you put all these
pieces together, the bad benchmarks, the
forgotten history of AI, these three
keys to learning, it paints a totally
different picture of the future. The
current north star of the field, this
chase for an all- knowing oracle trained
on all human data, it's not just
ambitious, it's the wrong goal entirely.
We need a new north star. So what's the
big takeaway here? It's that our current
benchmarks are just measuring static
performance, not the dynamic process of
intelligence. True intelligence at its
core is about the ability to acquire new
knowledge. So our goal shouldn't be an
all- knowing agent, but a general
purpose learning agent. And to get
there, the field has to shift its focus
to solving these hard fundamental
problems. Continual learning, single
stream experience, and scaling with
comput. This really leaves us with a
fundamental choice about the kind of
future we want to build. Are we chasing
this sci-fi fantasy of an all- knowing
oracle that just spits out answers? A
tool that kind of replaces thought, or
are we trying to create a true lifelong
learning partner, an agent that can
grow, adapt, and discover new things
right alongside us? The path we choose
from here is going to define the next
era of AI. If this kind of analysis, the
kind that cuts through the noise and
questions the core assumptions of a
field, is something you find valuable,
make sure you subscribe. We'll be
bringing you a lot more explainers that
go beyond the headlines to give you a
deeper, more nuanced view of the tech
that's shaping our world. Thanks for
joining this deep dive.