All right. So, Google DeepMind just
dropped this absolutely massive 57-page
paper. And honestly, the title alone is
kind of wild. It's called from AGI to
ASI. Not how to get to AGI or when AGI
will arrive. No, they're literally
[music] saying AGI is the starting
point. They're already looking past it,
which is kind of insane when you think
about how much everyone's been obsessing
over just reaching human-level AI.
The team behind this isn't some random
research group either. We're talking
Shane Legg, who literally co-founded
DeepMind with Demis Hassabis and Mustafa
Suleyman.
>> [music]
>> He's their chief AGI scientist. And then
there's Marcus Hutter, his doctoral
supervisor and the guy who invented the
AIXI theory. So, basically, 14 of the
top minds in AI getting together to map
out what happens after we hit
human-level intelligence. That's the
conversation they're having now. And
here's something I found absolutely
fascinating. The very first section of
this paper isn't even called
introduction. It's called summary
instructions. And get this,
it's written for AI. Like, they're
literally giving instructions to future
AI assistants that might be called upon
to summarize this report. They're
telling these AIs to make sure they
clarify the definitions, not compress
the lists, and judge whether the
conclusions hold up over time. This is
the first time in academic history where
authors are assuming AI will be reading
their paper on behalf of humans. That
alone tells you something about where we
are right now. So, let's talk about what
they're actually defining here. Because
the definitions matter a lot. AGI,
according to this paper, is basically a
system that performs at roughly the
median human level across most cognitive
[music] tasks. Not the smartest person
in the room, just your average person.
If an AI can reason, learn, plan,
communicate, use tools, and adapt to new
situations at that level, it's AGI.
Pretty straightforward. But ASI is where
things get crazy. Artificial
superintelligence isn't just about
beating one human expert at one task.
It's about a system that can outperform
tens of thousands of top experts working
together, well coordinated, for an
entire decade on a single problem. Think
about that for a second. We're not
talking about competing with one person
or even one company. We're talking about
matching the output of an entire
professional research field or [music] a
massive corporation going all in for 10
years. And this needs to happen across
virtually every domain. That's the bar
for ASI. There's also this third level
they mentioned called universal AI or
AXI, which is basically the theoretical
absolute ceiling of intelligence. It's
mathematically proven, but uncomputable,
meaning we can only approach it from
below, never actually reach it. Kind of
like the speed of light in physics. Now,
here's where the paper gets really
interesting. They lay out four main
pathways from AGI to ASI. And honestly,
each one is kind of terrifying in its
own way. The first pathway is just pure
scaling. More compute, bigger models,
more data. This is basically what got us
here in the first place. Over the last
decade, the compute used for the largest
machine learning training runs has been
growing exponentially, and algorithmic
efficiency has been improving at the
same time. So, we're not just throwing
more hardware at the problem, we're
getting better at using the hardware we
already have. The paper actually runs
this thought experiment. Let's say when
AGI first arrives, it's super expensive
and only a thousand instances can run
globally. But with a 10 times annual
growth rate, you'd have 10,000 instances
after 1 year, and after 5 years [music]
you'd have a hundred million instances.
And here's the kicker, a hundred million
AGIs at human level isn't just a hundred
million separate workers.
These things can share knowledge
instantly, communicate at incredibly
high bandwidth, copy themselves
perfectly, and coordinate [music] in
ways humans simply can't. They don't
need meetings or emails or time to
explain concepts to each other. One
instance figures something out and
potentially all 100 million know it
immediately. That collective
intelligence could easily qualify as
ASI, even if each individual unit is
still at the human level. You're
basically looking at a digital
civilization that thinks hundreds of
times faster than us, but then there's
the data wall problem. Current AI
systems learn from human-generated
content, text, code, images, [music]
videos, scientific papers, all of it.
But we're not producing high-quality
human data at the same exponential rate
[music] that AI models are growing.
Eventually, you run out of good stuff to
train on. The paper doesn't say this
will definitely stop progress, though.
There are workarounds like synthetic
data, simulations, [music] self-play,
reinforcement learning, and having AI
systems improve their own outputs
through search and then training on
those improved results. [music]
The question is whether that generated
data is good enough, because if you
naively train on AI-generated content,
things can degrade fast. The second
pathway is algorithmic paradigm shifts.
This is where AI doesn't just get
bigger, it gets fundamentally different.
Right now, we're dominated by
transformer-based [music]
models trained on massive data sets,
then refined with instruction tuning and
reinforcement learning. That's taken us
incredibly far, but a lot of researchers
think it's still missing key ingredients
for true AGI, things like robust
long-term planning, continual learning,
persistent memory, better world models,
and the ability to operate reliably in
completely open-ended [music]
environments. A real paradigm shift
could mean new architectures entirely,
new training methods, new memory
systems, new forms [music] of reasoning,
maybe even new hardware like
neuromorphic chips or analog computing.
The problem is that paradigm shifts are
basically impossible to predict before
they happen. If we knew exactly what the
next breakthrough [music] would be, it
wouldn't really be a surprise anymore.
But if it does happen, all the forecasts
based purely on scaling current systems
would become wrong almost overnight.
Then there's the third pathway, and this
one gets closest to the classic
intelligence explosion idea, recursive
self-improvement. The loop is simple. AI
helps improve AI research, which
produces better AI, which then helps
even more with AI research. This doesn't
have to mean one dramatic moment where a
model rewrites its own code. It can be
much more gradual and distributed. AI
systems could help write better
algorithms, discover better
architectures,
>> [music]
>> design more efficient chips, improve
manufacturing processes, curate better
data sets, generate better synthetic
examples, create better simulations, and
generally improve all the infrastructure
around AI development. The paper makes
this interesting comparison to human
evolution. We didn't just improve
through individual intelligence. We
built language, writing, institutions,
science, markets, education systems,
entire civilizations. A single human
isn't that impressive, but a
civilization is. The question is whether
AI systems can build their own version
of that, but way faster, because code
can be edited faster than DNA [music]
changes, data can be copied faster than
books can be printed. Specialists can be
spawned and trained faster than humans
can be educated. But recursive
self-improvement is also one of the
least understood pathways. [music] Maybe
it explodes exponentially. Maybe it
fizzles out. Maybe AI helps researchers
a lot, but progress still slows because
experiments are expensive, hardware
takes time to manufacture, energy
becomes a constraint, or the next
breakthrough ideas are just genuinely
hard to find. Even digital researchers
can't skip every bottleneck. If you need
to build a new AI chip, that happens in
the physical world. If you need to run a
biology experiment, it still takes real
time.
The fourth pathway might actually be the
most underrated one in the whole paper,
ASI through multi-agent collectives.
Instead of asking whether one AI model
becomes superintelligent, ask whether a
massive group of AI agents can become
superintelligent together. Humans
already do this. A corporation solves
problems no individual employee could
solve alone. A scientific field produces
knowledge no single researcher could
produce. But human group intelligence is
slow and messy. Communication is
limited, coordination is hard,
organizations become bureaucratic,
knowledge gets siloed. AI collectives
[music] could be completely different.
They could share information at speeds
we can't even imagine. They could
duplicate specialists instantly. They
could coordinate through software. They
could run thousands of parallel
experiments. They could form temporary
teams for specific problems, then
dissolve and reconfigure. They could use
market-like systems or centralized
planning in ways humans can't manage
because our communication bandwidth is
way too low. So, ASI might not look like
one giant mind at all. It might look
like a vast digital organization, a
swarm, a self-organizing research
ecosystem, or a supercompany made of
agents.
Now, after laying out these four
pathways, the paper gets into what they
call frictions. Basically, the things
that could slow everything down or stop
it entirely. There are six main ones.
First is the data wall we already talked
about. Not enough high-quality training
data to keep scaling forever.
Second is resource constraints. The
physical stuff like energy, chips, rare
materials, data centers, [music]
cooling systems, manufacturing capacity.
If capabilities require exponentially
larger infrastructure, the world might
just struggle to build it fast enough.
Third is the possibility that the
current neural network paradigm simply
isn't sufficient for AGI or ASI, no
matter how much you scale it. Fourth is
that research itself gets harder as
fields mature.
Low-hanging fruit disappears. Progress
requires more effort and more complex
ideas. Fifth is what they call the
abstraction barrier. Current AI systems
learn mostly from human abstractions,
the concepts, categories, language, and
structure we already use. But major
scientific breakthroughs often require
inventing completely new abstractions,
new ways of thinking about reality. The
worry is that AI trained mainly on human
representations might become excellent
at manipulating existing concepts, but
weaker at discovering fundamentally new
ones from scratch. And sixth is
deliberate slowdown, political and
social factors. If AI creates accidents,
enables misuse, destabilizes labor
markets, or triggers public backlash,
governments might slow development
through regulation, licensing
requirements, capability caps, or other
restriction. The paper's point isn't
that any of these definitely stops ASI.
It's that we genuinely don't know. Each
bottleneck could be a minor speed bump,
or it could be an absolute wall. Whether
it becomes one or the other depends on
how fast the counterforces improve.
There's also this important reality
check buried in the paper. ASI is not
omnipotent. Even a superintelligence
would still face fundamental limits.
Physics doesn't stop applying.
Information can't travel faster than
light. Computation costs energy.
Physical systems take time to
manipulate. Some problems are chaotic,
unpredictable, or fundamentally hard, no
matter how smart you are.
Complexity theory still matters. [music]
Logic still has limits. So people need
to stop jumping from ASI to magical
thinking, like instant cures for
everything or perfect control over
reality. ASI could be far beyond human
intelligence and still be constrained by
computation, energy, uncertainty, time,
and the physical world. The deeper
message here is uncertainty, not
ignorance, but genuine uncertainty about
which pathway dominates and where
progress plateaus. Maybe scaling
continues and gets us there. Maybe
scaling hits limits, but paradigm shifts
unlock the next jump. Maybe recursive
improvement becomes the main driver.
Maybe multi-agent collectives turn
human-level systems into superhuman
organizations. Or maybe all four happen
simultaneously, compounding each other.
Or maybe several major bottlenecks hit
at once and progress becomes slower and
more uneven than current trends suggest.
What makes this paper significant is
that it's forcing a conversation shift.
We need to stop treating AGI as a single
finish line. If AGI arrives, the next
question won't be are we done?
It'll be what does this system make
possible next? Because a human-level AI
isn't just another human.
It's a digital intelligence that can be
copied, accelerated, coordinated,
specialized, connected to tools, placed
inside organizations, and potentially
used to build better versions of itself.
>> [music]
>> We might be entering a period where
intelligence itself becomes an
industrial process. And once that
happens, the pace of change may no
longer be limited by how fast humans can
learn, organize, or invent. AGI might
just be the moment the real race begins.
So, what do you think is the real wall
between AGI and ASI? Compute, energy,
data, or something deeper? Drop your
thoughts in the comments.
Subscribe for more. Thanks for watching,
and I'll catch you in the next one.
>> Mhm.