Google Just Revealed What Comes After AGI And It’s Shocking

[00:02] All right, so Google DeepMind just

[00:04] dropped this absolutely massive 57page

[00:07] paper and honestly the title alone is

[00:09] kind of wild. It's called From AGI to

[00:12] ASI, not how to get to AGI or when AGI

[00:16] will arrive. No, they're literally

[00:18] saying AGI is the starting point.

[00:21] They're already looking past it, which

[00:23] is kind of insane when you think about

[00:24] how much everyone's been obsessing over

[00:27] just reaching human level AI. The team

[00:29] behind this isn't some random research

[00:32] group either. We're talking Shane Le who

[00:34] literally co-founded Deep Mind with Deus

[00:36] Hassabis and Mustafa Sullean. He's their

[00:39] chief AGI scientist. And then there's

[00:41] Marcus Hutter, his doctoral supervisor

[00:44] and the guy who invented the AIXI

[00:46] theory. So basically 14 of the top minds

[00:49] in AI getting together to map out what

[00:51] happens after we hit human level

[00:53] intelligence. That's the conversation

[00:55] they're having now. And here's something

[00:57] I found absolutely fascinating. The very

[00:59] first section of this paper isn't even

[01:01] called introduction. It's called summary

[01:04] instructions. And get this, it's written

[01:07] for AI. Like, they're literally giving

[01:10] instructions to future AI assistants

[01:12] that might be called upon to summarize

[01:14] this report. They're telling these AIs

[01:16] to make sure they clarify the

[01:18] definitions, not compress the lists, and

[01:20] judge whether the conclusions hold up

[01:22] over time. This is the first time in

[01:24] academic history where authors are

[01:26] assuming AI will be reading their paper

[01:28] on behalf of humans. That alone tells

[01:30] you something about where we are right

[01:32] now. So let's talk about what they're

[01:34] actually defining here because the

[01:36] definitions matter a lot. AGI according

[01:39] to this paper is basically a system that

[01:42] performs at roughly the median human

[01:44] level across most cognitive tasks. Not

[01:46] the smartest person in the room, just

[01:48] your average person. If an AI can

[01:51] reason learn plan communicate use

[01:54] tools, and adapt to new situations at

[01:57] that level, it's AGI. Pretty

[01:59] straightforward. But ASI is where things

[02:02] get crazy. Artificial super intelligence

[02:05] isn't just about beating one human

[02:07] expert at one task. It's about a system

[02:09] that can outperform tens of thousands of

[02:11] top experts working together,

[02:13] well-coordinated, for an entire decade

[02:15] on a single problem. Think about that

[02:17] for a second. We're not talking about

[02:19] competing with one person or even one

[02:21] company. We're talking about matching

[02:23] the output of an entire professional

[02:25] research field or a massive corporation

[02:28] going allin for 10 years. And this needs

[02:31] to happen across virtually every domain.

[02:34] That's the bar for ASI. There's also

[02:36] this third level they mention called

[02:38] universal AI or AIXI, which is basically

[02:41] the theoretical absolute ceiling of

[02:43] intelligence. It's mathematically proven

[02:46] but uncomputable, meaning we can only

[02:49] approach it from below, never actually

[02:51] reach it. Kind of like the speed of

[02:52] light in physics. Now, here's where the

[02:55] paper gets really interesting. They lay

[02:57] out four main pathways from AGI to ASI.

[03:01] And honestly, each one is kind of

[03:03] terrifying in its own way. The first

[03:05] pathway is just pure scaling. More

[03:07] compute, bigger models, more data. This

[03:10] is basically what got us here in the

[03:12] first place. Over the last decade, the

[03:14] compute used for the largest machine

[03:16] learning training runs has been growing

[03:18] exponentially, and algorithmic

[03:20] efficiency has been improving at the

[03:22] same time. So, we're not just throwing

[03:24] more hardware at the problem. We're

[03:26] getting better at using the hardware we

[03:27] already have. The paper actually runs

[03:30] this thought experiment. Let's say when

[03:32] AGI first arrives, it's super expensive

[03:35] and only a,000 instances can run

[03:37] globally. But with a 10 times annual

[03:39] growth rate, you'd have 10,000 instances

[03:42] after one year. And after 5 years, you'd

[03:44] have a 100 million instances. And here's

[03:46] the kicker. A 100 million AGIS at human

[03:49] level isn't just 100 million separate

[03:51] workers. These things can share

[03:53] knowledge instantly, communicate at

[03:55] incredibly high bandwidth, copy

[03:58] themselves perfectly, and coordinate in

[04:00] ways humans simply can't. They don't

[04:02] need meetings or emails or time to

[04:04] explain concepts to each other. one

[04:06] instance figures something out and

[04:08] potentially all hundred million know it

[04:10] immediately. That collective

[04:12] intelligence could easily qualify as ASI

[04:15] even if each individual unit is still at

[04:17] the human level. You're basically

[04:19] looking at a digital civilization that

[04:21] thinks hundreds of times faster than us.

[04:23] But then there's the data wall problem.

[04:25] Current AI systems learn from human

[04:27] generated content, text, code, images,

[04:30] videos, scientific papers, all of it.

[04:32] But we're not producing highquality

[04:34] human data at the same exponential rate

[04:36] that AI models are growing. Eventually,

[04:39] you run out of good stuff to train on.

[04:41] The paper doesn't say this will

[04:43] definitely stop progress, though. There

[04:45] are workarounds like synthetic data,

[04:47] simulations self-play reinforcement

[04:50] learning, and having AI systems improve

[04:53] their own outputs through search and

[04:55] then training on those improved results.

[04:57] The question is whether that generated

[04:59] data is good enough. Because if you

[05:01] naively train on AI generated content,

[05:04] things can degrade fast. The second

[05:06] pathway is algorithmic paradigm shifts.

[05:09] This is where AI doesn't just get

[05:10] bigger, it gets fundamentally different.

[05:13] Right now, we're dominated by

[05:14] transformer-based models trained on

[05:16] massive data sets, then refined with

[05:19] instruction tuning and reinforcement

[05:20] learning. That's taken us incredibly

[05:22] far, but a lot of researchers think it's

[05:24] still missing key ingredients for true

[05:26] AGI. things like robust long-term

[05:29] planning, continual learning, persistent

[05:31] memory, better world models, and the

[05:34] ability to operate reliably in

[05:36] completely open-ended environments. A

[05:38] real paradigm shift could mean new

[05:40] architectures entirely, new training

[05:42] methods, new memory systems, new forms

[05:44] of reasoning, maybe even new hardware

[05:46] like neuromorphic chips or analog

[05:49] computing. The problem is that paradigm

[05:51] shifts are basically impossible to

[05:53] predict before they happen. If we knew

[05:55] exactly what the next breakthrough would

[05:57] be, it wouldn't really be a surprise

[05:59] anymore. But if it does happen, all the

[06:02] forecasts based purely on scaling

[06:04] current systems would become wrong

[06:06] almost overnight. Then there's the third

[06:08] pathway, and this one gets closest to

[06:10] the classic intelligence explosion idea,

[06:12] recursive self-improvement. The loop is

[06:15] simple. AI helps improve AI research,

[06:17] which produces better AI, which then

[06:20] helps even more with AI research. This

[06:22] doesn't have to mean one dramatic moment

[06:23] where a model rewrites its own code. It

[06:26] can be much more gradual and

[06:27] distributed. AI systems could help write

[06:30] better algorithms, discover better

[06:32] architectures, design more efficient

[06:34] chips, improve manufacturing processes,

[06:36] curate better data sets, generate better

[06:39] synthetic examples, create better

[06:41] simulations, and generally improve all

[06:43] the infrastructure around AI

[06:45] development. The paper makes this

[06:47] interesting comparison to human

[06:49] evolution. We didn't just improve

[06:51] through individual intelligence. We

[06:53] built language, writing, institutions,

[06:56] science, markets, education systems,

[06:59] entire civilizations. A single human

[07:02] isn't that impressive, but a

[07:04] civilization is. The question is whether

[07:06] AI systems can build their own version

[07:08] of that, but way faster. Because code

[07:11] can be edited faster than DNA changes.

[07:14] Data can be copied faster than books can

[07:16] be printed. Specialists can be spawned

[07:18] and trained faster than humans can be

[07:20] educated. But recursive self-improvement

[07:22] is also one of the least understood

[07:23] pathways. Maybe it explodes

[07:25] exponentially. Maybe it fizzles out.

[07:28] Maybe AI helps researchers a lot, but

[07:30] progress still slows because experiments

[07:32] are expensive. Hardware takes time to

[07:34] manufacture. Energy becomes a

[07:36] constraint. Or the next breakthrough

[07:38] ideas are just genuinely hard to find.

[07:40] Even digital researchers can't skip

[07:42] every bottleneck. If you need to build a

[07:44] new AI chip that happens in the physical

[07:46] world, if you need to run a biology

[07:49] experiment, it still takes real time.

[07:51] The fourth pathway might actually be the

[07:53] most underrated one in the whole paper,

[07:55] ASI through multi-agent collectives.

[07:58] Instead of asking whether one AI model

[08:00] becomes super intelligent, ask whether a

[08:02] massive group of AI agents can become

[08:05] super intelligent together. Humans

[08:07] already do this. A corporation solves

[08:10] problems no individual employee could

[08:11] solve alone. A scientific field produces

[08:14] knowledge no single researcher could

[08:15] produce. But human group intelligence is

[08:17] slow and messy. Communication is

[08:20] limited. Coordination is hard.

[08:22] Organizations become bureaucratic.

[08:24] Knowledge gets siloed. AI collectives

[08:26] could be completely different. They

[08:28] could share information at speeds we

[08:29] can't even imagine. They could duplicate

[08:31] specialists instantly. They could

[08:33] coordinate through software. They could

[08:35] run thousands of parallel experiments.

[08:37] They could form temporary teams for

[08:39] specific problems, then dissolve and

[08:42] reconfigure. They could use marketlike

[08:44] systems or centralized planning in ways

[08:47] humans can't manage because our

[08:48] communication bandwidth is way too low.

[08:51] So ASI might not look like one giant

[08:53] mind at all. It might look like a vast

[08:56] digital organization, a swarm, a

[08:58] self-organizing research ecosystem, or a

[09:01] super company made of agents. Now, after

[09:04] laying out these four pathways, the

[09:06] paper gets into what they call

[09:08] frictions. Basically, the things that

[09:10] could slow everything down or stop it

[09:12] entirely. There are six main ones. First

[09:15] is the data wall we already talked

[09:17] about. Not enough high-quality training

[09:19] data to keep scaling forever. Second is

[09:21] resource constraints, the physical stuff

[09:24] like energy, chips, rare materials, data

[09:27] centers, cooling systems, manufacturing

[09:29] capacity. If capabilities require

[09:32] exponentially larger infrastructure, the

[09:34] world might just struggle to build it

[09:36] fast enough. Third is the possibility

[09:38] that the current neural network paradigm

[09:40] simply isn't sufficient for AGI or ASI,

[09:43] no matter how much you scale it. Fourth

[09:45] is that research itself gets harder as

[09:47] fields mature. Lowhanging fruit

[09:49] disappears. Progress requires more

[09:51] effort and more complex ideas. Fifth is

[09:54] what they call the abstraction barrier.

[09:57] Current AI systems learn mostly from

[09:59] 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.