[0:02] All right, so Google DeepMind just
[0:04] dropped this absolutely massive 57page
[0:07] paper and honestly the title alone is
[0:09] kind of wild. It's called From AGI to
[0:12] ASI, not how to get to AGI or when AGI
[0:16] will arrive. No, they're literally
[0:18] saying AGI is the starting point.
[0:21] They're already looking past it, which
[0:23] is kind of insane when you think about
[0:24] how much everyone's been obsessing over
[0:27] just reaching human level AI. The team
[0:29] behind this isn't some random research
[0:32] group either. We're talking Shane Le who
[0:34] literally co-founded Deep Mind with Deus
[0:36] Hassabis and Mustafa Sullean. He's their
[0:39] chief AGI scientist. And then there's
[0:41] Marcus Hutter, his doctoral supervisor
[0:44] and the guy who invented the AIXI
[0:46] theory. So basically 14 of the top minds
[0:49] in AI getting together to map out what
[0:51] happens after we hit human level
[0:53] intelligence. That's the conversation
[0:55] they're having now. And here's something
[0:57] I found absolutely fascinating. The very
[0:59] first section of this paper isn't even
[1:01] called introduction. It's called summary
[1:04] instructions. And get this, it's written
[1:07] for AI. Like, they're literally giving
[1:10] instructions to future AI assistants
[1:12] that might be called upon to summarize
[1:14] this report. They're telling these AIs
[1:16] to make sure they clarify the
[1:18] definitions, not compress the lists, and
[1:20] judge whether the conclusions hold up
[1:22] over time. This is the first time in
[1:24] academic history where authors are
[1:26] assuming AI will be reading their paper
[1:28] on behalf of humans. That alone tells
[1:30] you something about where we are right
[1:32] now. So let's talk about what they're
[1:34] actually defining here because the
[1:36] definitions matter a lot. AGI according
[1:39] to this paper is basically a system that
[1:42] performs at roughly the median human
[1:44] level across most cognitive tasks. Not
[1:46] the smartest person in the room, just
[1:48] your average person. If an AI can
[1:51] reason learn plan communicate use
[1:54] tools, and adapt to new situations at
[1:57] that level, it's AGI. Pretty
[1:59] straightforward. But ASI is where things
[2:02] get crazy. Artificial super intelligence
[2:05] isn't just about beating one human
[2:07] expert at one task. It's about a system
[2:09] that can outperform tens of thousands of
[2:11] top experts working together,
[2:13] well-coordinated, for an entire decade
[2:15] on a single problem. Think about that
[2:17] for a second. We're not talking about
[2:19] competing with one person or even one
[2:21] company. We're talking about matching
[2:23] the output of an entire professional
[2:25] research field or a massive corporation
[2:28] going allin for 10 years. And this needs
[2:31] to happen across virtually every domain.
[2:34] That's the bar for ASI. There's also
[2:36] this third level they mention called
[2:38] universal AI or AIXI, which is basically
[2:41] the theoretical absolute ceiling of
[2:43] intelligence. It's mathematically proven
[2:46] but uncomputable, meaning we can only
[2:49] approach it from below, never actually
[2:51] reach it. Kind of like the speed of
[2:52] light in physics. Now, here's where the
[2:55] paper gets really interesting. They lay
[2:57] out four main pathways from AGI to ASI.
[3:01] And honestly, each one is kind of
[3:03] terrifying in its own way. The first
[3:05] pathway is just pure scaling. More
[3:07] compute, bigger models, more data. This
[3:10] is basically what got us here in the
[3:12] first place. Over the last decade, the
[3:14] compute used for the largest machine
[3:16] learning training runs has been growing
[3:18] exponentially, and algorithmic
[3:20] efficiency has been improving at the
[3:22] same time. So, we're not just throwing
[3:24] more hardware at the problem. We're
[3:26] getting better at using the hardware we
[3:27] already have. The paper actually runs
[3:30] this thought experiment. Let's say when
[3:32] AGI first arrives, it's super expensive
[3:35] and only a,000 instances can run
[3:37] globally. But with a 10 times annual
[3:39] growth rate, you'd have 10,000 instances
[3:42] after one year. And after 5 years, you'd
[3:44] have a 100 million instances. And here's
[3:46] the kicker. A 100 million AGIS at human
[3:49] level isn't just 100 million separate
[3:51] workers. These things can share
[3:53] knowledge instantly, communicate at
[3:55] incredibly high bandwidth, copy
[3:58] themselves perfectly, and coordinate in
[4:00] ways humans simply can't. They don't
[4:02] need meetings or emails or time to
[4:04] explain concepts to each other. one
[4:06] instance figures something out and
[4:08] potentially all hundred million know it
[4:10] immediately. That collective
[4:12] intelligence could easily qualify as ASI
[4:15] even if each individual unit is still at
[4:17] the human level. You're basically
[4:19] looking at a digital civilization that
[4:21] thinks hundreds of times faster than us.
[4:23] But then there's the data wall problem.
[4:25] Current AI systems learn from human
[4:27] generated content, text, code, images,
[4:30] videos, scientific papers, all of it.
[4:32] But we're not producing highquality
[4:34] human data at the same exponential rate
[4:36] that AI models are growing. Eventually,
[4:39] you run out of good stuff to train on.
[4:41] The paper doesn't say this will
[4:43] definitely stop progress, though. There
[4:45] are workarounds like synthetic data,
[4:47] simulations self-play reinforcement
[4:50] learning, and having AI systems improve
[4:53] their own outputs through search and
[4:55] then training on those improved results.
[4:57] The question is whether that generated
[4:59] data is good enough. Because if you
[5:01] naively train on AI generated content,
[5:04] things can degrade fast. The second
[5:06] pathway is algorithmic paradigm shifts.
[5:09] This is where AI doesn't just get
[5:10] bigger, it gets fundamentally different.
[5:13] Right now, we're dominated by
[5:14] transformer-based models trained on
[5:16] massive data sets, then refined with
[5:19] instruction tuning and reinforcement
[5:20] learning. That's taken us incredibly
[5:22] far, but a lot of researchers think it's
[5:24] still missing key ingredients for true
[5:26] AGI. things like robust long-term
[5:29] planning, continual learning, persistent
[5:31] memory, better world models, and the
[5:34] ability to operate reliably in
[5:36] completely open-ended environments. A
[5:38] real paradigm shift could mean new
[5:40] architectures entirely, new training
[5:42] methods, new memory systems, new forms
[5:44] of reasoning, maybe even new hardware
[5:46] like neuromorphic chips or analog
[5:49] computing. The problem is that paradigm
[5:51] shifts are basically impossible to
[5:53] predict before they happen. If we knew
[5:55] exactly what the next breakthrough would
[5:57] be, it wouldn't really be a surprise
[5:59] anymore. But if it does happen, all the
[6:02] forecasts based purely on scaling
[6:04] current systems would become wrong
[6:06] almost overnight. Then there's the third
[6:08] pathway, and this one gets closest to
[6:10] the classic intelligence explosion idea,
[6:12] recursive self-improvement. The loop is
[6:15] simple. AI helps improve AI research,
[6:17] which produces better AI, which then
[6:20] helps even more with AI research. This
[6:22] doesn't have to mean one dramatic moment
[6:23] where a model rewrites its own code. It
[6:26] can be much more gradual and
[6:27] distributed. AI systems could help write
[6:30] better algorithms, discover better
[6:32] architectures, design more efficient
[6:34] chips, improve manufacturing processes,
[6:36] curate better data sets, generate better
[6:39] synthetic examples, create better
[6:41] simulations, and generally improve all
[6:43] the infrastructure around AI
[6:45] development. The paper makes this
[6:47] interesting comparison to human
[6:49] evolution. We didn't just improve
[6:51] through individual intelligence. We
[6:53] built language, writing, institutions,
[6:56] science, markets, education systems,
[6:59] entire civilizations. A single human
[7:02] isn't that impressive, but a
[7:04] civilization is. The question is whether
[7:06] AI systems can build their own version
[7:08] of that, but way faster. Because code
[7:11] can be edited faster than DNA changes.
[7:14] Data can be copied faster than books can
[7:16] be printed. Specialists can be spawned
[7:18] and trained faster than humans can be
[7:20] educated. But recursive self-improvement
[7:22] is also one of the least understood
[7:23] pathways. Maybe it explodes
[7:25] exponentially. Maybe it fizzles out.
[7:28] Maybe AI helps researchers a lot, but
[7:30] progress still slows because experiments
[7:32] are expensive. Hardware takes time to
[7:34] manufacture. Energy becomes a
[7:36] constraint. Or the next breakthrough
[7:38] ideas are just genuinely hard to find.
[7:40] Even digital researchers can't skip
[7:42] every bottleneck. If you need to build a
[7:44] new AI chip that happens in the physical
[7:46] world, if you need to run a biology
[7:49] experiment, it still takes real time.
[7:51] The fourth pathway might actually be the
[7:53] most underrated one in the whole paper,
[7:55] ASI through multi-agent collectives.
[7:58] Instead of asking whether one AI model
[8:00] becomes super intelligent, ask whether a
[8:02] massive group of AI agents can become
[8:05] super intelligent together. Humans
[8:07] already do this. A corporation solves
[8:10] problems no individual employee could
[8:11] solve alone. A scientific field produces
[8:14] knowledge no single researcher could
[8:15] produce. But human group intelligence is
[8:17] slow and messy. Communication is
[8:20] limited. Coordination is hard.
[8:22] Organizations become bureaucratic.
[8:24] Knowledge gets siloed. AI collectives
[8:26] could be completely different. They
[8:28] could share information at speeds we
[8:29] can't even imagine. They could duplicate
[8:31] specialists instantly. They could
[8:33] coordinate through software. They could
[8:35] run thousands of parallel experiments.
[8:37] They could form temporary teams for
[8:39] specific problems, then dissolve and
[8:42] reconfigure. They could use marketlike
[8:44] systems or centralized planning in ways
[8:47] humans can't manage because our
[8:48] communication bandwidth is way too low.
[8:51] So ASI might not look like one giant
[8:53] mind at all. It might look like a vast
[8:56] digital organization, a swarm, a
[8:58] self-organizing research ecosystem, or a
[9:01] super company made of agents. Now, after
[9:04] laying out these four pathways, the
[9:06] paper gets into what they call
[9:08] frictions. Basically, the things that
[9:10] could slow everything down or stop it
[9:12] entirely. There are six main ones. First
[9:15] is the data wall we already talked
[9:17] about. Not enough high-quality training
[9:19] data to keep scaling forever. Second is
[9:21] resource constraints, the physical stuff
[9:24] like energy, chips, rare materials, data
[9:27] centers, cooling systems, manufacturing
[9:29] capacity. If capabilities require
[9:32] exponentially larger infrastructure, the
[9:34] world might just struggle to build it
[9:36] fast enough. Third is the possibility
[9:38] that the current neural network paradigm
[9:40] simply isn't sufficient for AGI or ASI,
[9:43] no matter how much you scale it. Fourth
[9:45] is that research itself gets harder as
[9:47] fields mature. Lowhanging fruit
[9:49] disappears. Progress requires more
[9:51] effort and more complex ideas. Fifth is
[9:54] what they call the abstraction barrier.
[9:57] Current AI systems learn mostly from
[9:59] human abstractions. The concepts,
[10:02] categories, language, and structure we
[10:04] already use. But major scientific
[10:07] breakthroughs often require inventing
[10:09] completely new abstractions, new ways of
[10:12] thinking about reality. The worry is
[10:14] that AI trained mainly on human
[10:16] representations might become excellent
[10:18] at manipulating existing concepts but
[10:21] weaker at discovering fundamentally new
[10:23] ones from scratch. And sixth is
[10:25] deliberate slowdown political and social
[10:28] factors. If AI creates accidents,
[10:31] enables misuse, destabilizes labor
[10:34] markets, or triggers public backlash,
[10:36] governments might slow development
[10:38] through regulation, licensing
[10:39] requirements, capability caps, or other
[10:42] restrictions. The paper's point isn't
[10:44] that any of these definitely stops ASI.
[10:47] It's that we genuinely don't know. Each
[10:49] bottleneck could be a minor speed bump
[10:51] or it could be an absolute wall. Whether
[10:54] it becomes one or the other depends on
[10:56] how fast the counterforces improve.
[10:58] There's also this important reality
[11:00] check buried in the paper. ASI is not
[11:03] omnipotent. Even a super intelligence
[11:05] would still face fundamental limits.
[11:07] Physics doesn't stop applying.
[11:09] Information can't travel faster than
[11:11] light. Computation costs energy.
[11:14] Physical systems take time to
[11:16] manipulate. Some problems are chaotic,
[11:18] unpredictable, or fundamentally hard. No
[11:21] matter how smart you are, complexity
[11:23] theory still matters. Logic still has
[11:26] limits. So people need to stop jumping
[11:29] from ASI to magical thinking like
[11:31] instant cures for everything or perfect
[11:34] control over reality. ASI could be far
[11:36] beyond human intelligence and still be
[11:39] constrained by computation, energy,
[11:41] uncertainty, time, and the physical
[11:43] world. The deeper message here is
[11:46] uncertainty. not ignorance but genuine
[11:49] uncertainty about which pathway
[11:51] dominates and where progress plateaus.
[11:54] Maybe scaling continues and gets us
[11:56] there. Maybe scaling hits limits but
[11:58] paradigm shifts unlock the next jump.
[12:00] Maybe recursive improvement becomes the
[12:02] main driver. Maybe multi-agent
[12:04] collectives turn human level systems
[12:06] into superhuman organizations. Or maybe
[12:09] all four happen simultaneously,
[12:11] compounding each other. or maybe several
[12:13] major bottlenecks hit at once and
[12:15] progress becomes slower and more uneven
[12:18] than current trends suggest. What makes
[12:20] this paper significant is that it's
[12:22] forcing a conversation shift. We need to
[12:24] stop treating AGI as a single finish
[12:26] line. If AGI arrives, the next question
[12:29] won't be are we done? It'll be what does
[12:31] this system make possible next? Because
[12:33] a human level AI isn't just another
[12:36] human. It's a digital intelligence that
[12:38] can be copied, accelerated, coordinated,
[12:42] specialized, connected to tools, placed
[12:44] inside organizations, and potentially
[12:47] used to build better versions of itself.
[12:49] We might be entering a period where
[12:50] intelligence itself becomes an
[12:52] industrial process. And once that
[12:54] happens, the pace of change may no
[12:56] longer be limited by how fast humans can
[12:58] learn, organize, or invent. AGI might
[13:01] just be the moment the real race begins.
[13:03] So, what do you think is the real wall
[13:05] between AGI and ASI? Compute, energy,
[13:09] data, or something deeper? Drop your
[13:11] thoughts in the comments. Subscribe for
[13:14] more. Thanks for watching, and I'll
[13:16] catch you in the next one.