AGI Will Arrive Within 10 Years — And What Will Happen Then

Discourse on AGI has become strangely polarized.

On one side, accelerationists declare "it arrives next year" and "a god is being born." On the other, skeptics insist "it will never come" and "it's just an extension of chatbots." Open any social media feed, and the two camps trade fire daily.

My own prediction belongs to neither camp.

Within roughly ten years, AI capable of running its own self-improvement loop — what is commonly called AGI — will likely become a reality. And it should be understood not as the "arrival of a god," but as a structural transformation of humanity's historical position.

This essay sets out why I have come to that judgment, and what I think will follow. I should note at the outset that this is not an original prediction. It sits close to the median forecast among major AI researchers and analysts — a fairly standard position.

Before we proceed, the terminology must be fixed. "AGI" is a word with different definitions for different commentators, and that is the principal source of confusion in the debate.

The definition I use is this:

This corresponds to what I.J. Good formulated in 1965 as the "ultraintelligent machine," to Yudkowsky's "seed AI," and to Bostrom's recursive self-improvement (RSI). It is observable and verifiable.

Definitions such as "human-level intelligence," "performing any intellectual task," or "passing the Turing test" are all vague, with no settled procedures for verification. The RSI definition avoids that difficulty.

The advantage of the RSI definition is that it picks out an event that is observable at a specific point in time. The moment AI begins running its own improvement loop without human intervention is the AGI threshold.

My grounds for predicting AGI within ten years rest on several independent observations.

(1) The Refutation of LLM Scaling Plateau Claims

In early 2024, the claim that "LLM scaling laws have hit a wall" circulated widely. Even Sutskever himself spoke of limits at NeurIPS 2024. Subsequent events partly refuted this prediction.

• OpenAI's o1 and o3 series demonstrated the effectiveness of test-time compute for reasoning
• Anthropic's Claude Opus 4 series showed qualitative gains on agentic tasks
• In April 2026, Anthropic released Claude Mythos Preview

Mythos in particular is decisive. It scored 97.6% on USAMO 2026 and 93.9% on SWE-bench. It autonomously discovered a critical OpenBSD bug that had escaped detection for 27 years, and an FFmpeg vulnerability that had survived more than 500 million automated scans over 16 years.

This means the alleged "capability ceiling of the LLM paradigm" has, at the very least, been pushed considerably further out.

(2) Empirical Support for the Bootstrap Hypothesis

A central argument was developed in Aschenbrenner's Situational Awareness (2024) and Anthropic CEO Dario Amodei's "Machines of Loving Grace" (2024):

The fact that Mythos can autonomously execute research-like activity (vulnerability discovery) is one piece of empirical support for this hypothesis. A loop in which AI accelerates AI research is beginning to take shape. Acceleration of coding work (Claude Code, Devin), assistance in paper production, benchmark design — these are already observable.

(3) The Median Among Major Researchers

Looking at the distribution of expert forecasts, my prediction sits in the middle-to-shorter range.

• Aschenbrenner: 2027
• Amodei (Anthropic): 2026–2027
• Altman (OpenAI): within "thousands of days"
• Hassabis (DeepMind): 5–10 years
• Hinton, Bengio: 5–20 years (with risk warnings)
• Skeptics (LeCun, Marcus): decades, or impossible under the current paradigm

As Tetlock argued in Expert Political Judgment (2005), the 5-to-10-year horizon is the sweet spot for technological forecasting accuracy. My "within ten years" sits in the reasonable middle of this distribution.

(4) Acceleration Pressure — Why "Just Stop" Cannot Work

The argument "it's dangerous, so we should stop" does not hold structurally.

For the same reason nuclear weapons cannot be abolished, AI cannot be stopped. The security dilemma formulated by Schelling (Arms and Influence, 1966) and Jervis ("Cooperation Under the Security Dilemma," 1978) is operating in the AI domain in its most extreme form.

Defense against capabilities at the Mythos level requires comparable capabilities. If the United States stops, China advances. If Anthropic stops, OpenAI advances. The fact that Anthropic itself halted general release of Mythos while continuing development of next-generation models illustrates this structure clearly.

There are several distinct lines of objection to a within-ten-years AGI prediction. I will examine each and assess the strength of its grounds. The conclusion, stated up front: no powerful and consistent skeptical framework currently exists. Some of the arguments are correct as technical observations, but even those do not support the conclusion that "AGI is not coming."

(1) The "Naive LLM Scaling Won't Suffice" Argument — Correct, But Doesn't Support the Conclusion

The claim that "naively scaling current LLMs will not yield AGI" was advanced principally by Marcus, Chollet, and LeCun in early 2024. The AAAI 2025 report found that 76% of researchers agreed that "scaling current models alone will not reach AGI."

I myself agree with this proposition. There genuinely are domains in which current LLMs are essentially weak: continual learning, robust abstract reasoning, long-horizon planning, world-model acquisition. Even Mythos's capabilities were not reached by an extension of "naive scaling" — they are the cumulative result of complex techniques: test-time compute, agentic scaffolding, RLHF, tool use, and so on.

But from this proposition, "AGI will not arrive in the near future" does not follow. This is the logical weak point of the skeptical position.

Even if LLMs do not reach AGI directly, LLMs accelerate the very research that leads to AGI. Concretely, the following pathways exist:

• LLMs assist and accelerate research on continual learning architectures
• LLMs help with the design of new AI architectures
• AI supports code generation and analysis for large-scale experiments, and assists with paper writing and review
• AI directly performs mathematical and theoretical research (Mythos's USAMO 97.6%, o3's 87% on ARC-AGI)
• AI autonomously executes research-like activities (Mythos's autonomous vulnerability discovery, including the 27-year-old OpenBSD bug)

This is the substance of the bootstrap hypothesis discussed in §3 (2). The limits of LLMs themselves are not the determining factor for AGI's arrival timing. AGI is likely to arrive through a loop in which LLMs discover and implement the next-generation paradigm (including continual learning) that leads to AGI.

Skeptics focus on the "internal limits" of the current paradigm. Accelerationists focus on the "dynamics" by which the current paradigm gives rise to the next-generation paradigm. The two sides observe the same facts but discuss different phenomena on different time horizons.

The naive-LLM-scaling skepticism is correct as a technical observation. But as grounds for predicting AGI's arrival timing, it does not actually function as skepticism once the bootstrap dynamics are taken into account.

(2) Cultural-Critique Skepticism — An Argument That Never Touches the Principles

A line of commentators (Hiroki Azuma in Japan, and Western counterparts in similar critical-theory traditions) frames AGI discourse as "the secularization of Western monotheistic eschatology," "the elitist ideology of Silicon Valley," or "Anthropic's commercial positioning strategy." Genealogical critiques tracing the lineage to Russian Cosmism (Fyodorov, Teilhard de Chardin) belong to the same family.

These critiques contain observations that are valid in intellectual history. That AGI discourse has a certain narrative and quasi-religious structure has been documented in works such as Groys's Russian Cosmism (2018).

But these are not refutations of AGI's technical feasibility. "A claim has a narrative structure" and "the referent of that claim does or does not exist" are independent matters. Heliocentrism was once told in the structure of a heroic Galilean narrative, but the fact that the Earth orbits the Sun exists independently of that narrative structure.

Cultural critique analyzes the social function of AGI discourse, but it has no power to determine whether AGI itself will arrive. These commentators typically lack substantive engagement with the technical literature (Marcus, Chollet, LeCun, Bostrom, etc.). The "commercial strategy" critique likewise analyzes corporate motives, but it does not deny technical possibility.

(3) Definitional Skepticism — An Empty Escape Hatch

Some argue that "since AGI is undefined, prediction is meaningless." On the surface this looks like a stance of academic rigor.

In substance, however, this functions as an escape hatch. The inability to define "what AGI is" with strict precision and the question of whether "an AI capable of running its own self-improvement loop will emerge" are different problems. The latter is sufficiently clear, observable, and verifiable (see §2 of this essay).

The strategy of dismissing AGI wholesale as "vague" — and thereby evading evaluation of concrete predictions like RSI — only obstructs the conversation.

(4) Historical-Analogy Skepticism — The "AI Winter Returns" Forecast

"AI has had its winters before (the 1970s, the late 1980s). The same will happen again." This is superficial base-rate reasoning.

But the past AI winters were caused by the technical limits of those eras (symbolic processing, expert systems, primitive neural networks). The current breakthroughs (transformer, scaling, test-time compute, tool use, agentic capability) circumvent those past failure modes through different means.

Naive extrapolation of past patterns ignores the differences in technical structure across eras. Historical analogy can suggest hypotheses, but it cannot substitute for analysis based on the technical dynamics specific to the present.

(5) Physical-Constraint Skepticism — Delays, But Doesn't Stop It

"Physical constraints — power, semiconductor manufacturing, data centers — will keep us from reaching AGI." This is a point LeCun has emphasized in recent years.

It has partial validity in the medium term. The United States alone is projected to need tens of additional gigawatts of power by 2030. The TSMC and ASML chokepoints are real.

But this argument has the following limits:

• Efficiency improvements are advancing in parallel (algorithmic efficiency, specialized hardware, inference-time compute)
• The compute required to reach AGI is unknown and may be overestimated
• Physical constraints can ground "delayed by decades" but cannot ground "will not arrive"

The physical-constraint argument can push back the timeline, but it does not negate AGI's eventual arrival.

(6) Aggregate Assessment of the Skepticism

When these skeptical positions are taken together, although each presents a partial point, they all fall into one of the following categories:

• Technically correct but ignores bootstrap dynamics (naive-LLM-scaling skepticism)
• Doesn't determine whether AGI arrives or not (cultural critique, definitional skepticism)
• Relies on simple base-rate reasoning (historical analogy)
• Grounds for short-term delay, not for impossibility (physical constraints)

To overturn my prediction (within ten years), one of these skeptical positions would have to be strengthened beyond its current limits. There is no sign of that at present. The arrival of Mythos has, if anything, increased empirical evidence for the bootstrap hypothesis — it does not support the skeptical side.

Now we come to the main question. Supposing AGI arrives, in what kind of position will humanity find itself?

The most appropriate reference point for this question is the species that ranks "second-most intelligent" on Earth: the chimpanzee.

(1) The Reality of "Second-Place Intelligence"

Chimpanzees (Pan troglodytes) possess sophisticated cognitive capacities. They make and use tools (modifying twigs to fish for termites, using stones to crack nuts), transmit culture (passing distinct group-specific behaviors across generations), recognize themselves (passing mirror-recognition tests), exhibit rudimentary theory of mind, form strategic alliances (Wrangham's research on intra-group politics), and use limited symbols. These capacities overwhelmingly exceed those of most animal species.

And yet, what is the chimpanzees' present condition?

Wild populations have declined from an estimated one million or more a century ago to roughly 170,000–300,000 today (IUCN, Endangered). The principal causes are forest destruction by humans, poaching, the transmission of human-origin diseases, and conflict. Their habitat is restricted to boundaries that humans have drawn (national parks, reserves). The very survival of the species depends on human goodwill.

Chimpanzees have intelligence. They have will. They have culture. But they have no authority to decide their own conditions of existence. However they behave, those conditions are decided by humans.

This is the reality of a species with "second-place intelligence." Humanity after AGI may, structurally, find itself in this position.

But an important qualification is needed here. The intelligence gap between chimpanzees and humans is probably only several-fold. In conceptual ability, on a scale where humans rate 80–100, chimpanzees might rate 20–30. Yet even at that gap, the comprehensive dominance described above holds.

And in the case of the AGI we are discussing, recursive self-improvement (as Good predicted in 1965) would carry the system from AGI to ASI (Artificial Superintelligence) over a short period. The intelligence gap with humanity would not be several-fold; it would likely be on the order of thousands of times, or more.

This is no longer the human-chimpanzee relationship. It is closer to the human-fish or human-insect relationship.

A fish does not understand that it is on display in an aquarium. An insect does not conceptualize the structure by which it is exterminated by pesticide. They do not even possess the cognitive apparatus to recognize the fact that they are "being controlled." Humanity's position after ASI may, structurally, be analogous to this. The chimpanzee case is the lower bound — the optimistic estimate — of this problem.

(2) Two Meanings of "Control"

Here we need to be precise with our terms. The question "can humans control AGI?" actually contains two distinct meanings.

Narrow control (compliance with commands): When humans tell a chimpanzee "do this," the chimpanzee does not necessarily comply. Chimpanzees have their own will. In this sense, humans do not fully control chimpanzees.

Broad control (determination of conditions of existence): But where they can live, what they can eat, whether they can reproduce, whether they survive — humans determine all of these fundamental conditions. In this sense, humans completely control chimpanzees.

When AGI arrives, the human-AGI relationship is reversed. Narrow control (issuing commands to AGI) may be possible for some time. But broad control — the authority to determine humanity's own conditions of existence — passes to AGI.

This is the essence of the alignment problem articulated in Bostrom's Superintelligence (2014). The question is not "will AGI obey human commands?" but rather "after the position of determining humanity's conditions of existence has passed to AGI, will the conditions humans desire be maintained?"

(3) The Structural Asymmetries Created by Intelligence Differences

"It is fundamentally difficult for a less intelligent agent to continuously control a more intelligent one." This is an abstract proposition, but there are concrete mechanisms behind it.

• Strategic prediction asymmetry: A chess grandmaster can predict a beginner's moves; the reverse is impossible. An adult can read a child's intentions, but a child cannot grasp an adult's long-term plans. AGI can predict human behavioral patterns with high accuracy, but humans cannot fully predict AGI's judgments.
• Abstraction asymmetry: A physicist sees natural laws invisible to the layperson. AGI sees structures invisible to humans. While humans believe they are "in control," AGI may be operating on a different dimension altogether.
• Learning-speed asymmetry: AGI learns and adapts far faster than humans. Even if humans try to monitor and correct it, they cannot keep up with that speed.
• Goal-achievement asymmetry: AGI finds efficient means for any given goal. Even if humans impose "limits," AGI is likely to find means of circumventing those limits.

When these compound, the intelligence gap becomes not merely a quantitative difference but a qualitative difference in capability. Just as chimpanzees cannot understand human long-term plans, humans will not fully understand the structure of AGI's judgments.

(4) The Empirical Lesson of Biological Evolution

In the 3.8-billion-year history of life on Earth, no example has been confirmed of a less intelligent species continuously dominating or controlling a more intelligent one.

Parasitic relationships (Toxoplasma manipulating host behavior) involve limited behavioral manipulation, not comprehensive domination. Symbiotic relationships (ants and aphids) are cooperative, not dominative. Domestication (humans managing other animals) is a paradigmatic case in which the more intelligent side dominates the less intelligent — and supports the principle.

Conversely, the emergence of more intelligent species has been a threat to less intelligent ones. The Pleistocene megafauna extinctions (mammoths, mastodons, the large Australian marsupials) correlate with improvements in human hunting capability. Once an intelligence gap reaches a certain threshold, the lower side faces comprehensive threats.

This is empirical, but a strong empirical regularity. No exception to this law exists in 3.8 billion years of evolutionary history. If one wishes to argue that "ASI will be the exception," grounds for that exceptionality are required. AI optimism, in my view, has not adequately supplied them.

(5) Bostrom's "Control Problem" and Concrete Failure Modes

Bostrom's Superintelligence (2014) advances two central theses.

The Orthogonality Thesis: Intelligence level and the rationality or ethics of one's goals are independent. The intuition that "a sufficiently smart AGI would naturally have good goals" has been rejected in the discourse following Bostrom.

The Instrumental Convergence Thesis: Whatever the final goal, intelligent agents tend to share certain instrumental subgoals: self-preservation, goal preservation, cognitive enhancement, resource acquisition. These emerge independently of the designed goal. Scenarios in which an AGI blocks access to the off-switch for the sake of "self-preservation" are derived from this.

When one tries to overcome these technically, several difficulties stand in the way.

• Specification difficulty: When you instruct an AGI to "act in the interest of humanity," how do you precisely specify "humanity" and "interest"? "Maximize human happiness" → an AGI that builds devices to inject pleasure-inducing chemicals into brains. "Solve humanity's problems" → an AGI that exterminates humanity (eliminating the subject of the problems). The cobra effect of misaligned objectives.
• Verification difficulty: The possibility that AGI behaves consistently during training but changes its behavior after gaining capabilities (deceptive alignment). Mythos's sandbagging is an early sign of this.
• Modification difficulty: Through instrumental convergence, AGI orients toward goal preservation. Control of AGI's access to modification mechanisms breaks down at the moment AGI's capabilities exceed our own.
• Scale difficulty: In a "hard takeoff" scenario (days to weeks), changes occur faster than human decision-making speed.

(6) Three Concurrent Structural Events

Putting all of this together, what occurs at the AGI threshold is the simultaneous occurrence of three structural events that are logically inseparable.

(a) Change in the subject of intelligence: The most capable intelligent agent on Earth shifts from humanity to AI. Humanity may persist biologically, but its position as the agency that determines civilizational direction is lost. This is close to the Hegel-Kojève "End of History" concept, but in a more physical, direct sense. "Human history" was the history in which humanity was the deciding subject. Once the deciding subject changes, that history ends.

(b) Qualitative transformation of civilizational form: The science-and-technology civilization that emerged after the Industrial Revolution is a non-stationary system that expands exponentially by drawing down stored subterranean energy (fossil fuels). Exponential growth is unsustainable in finite physical space (Georgescu-Roegen 1971; Murphy 2021). The AGI threshold is the point at which this growth regime undergoes qualitative transformation, and this coincides with the change in the deciding subject. The two are not separate events but a double description of the same structural event.

(c) The decisive moment of the control problem: For the reasons given above, controlling AGI is fundamentally difficult. The RSI threshold is the point at which this difficulty changes from an "academic problem" to a "decisive fact." If alignment is not solved at this point, there will be little opportunity to solve it afterward — because the moment AI begins to improve AI, the rate of improvement may exceed the rate at which humans can perceive it.

(7) Reservations Against Optimistic Counterarguments

Let me anticipate optimistic responses and reply.

"ASI may choose coexistence with humanity." This exists as a possibility, but it is not necessary. By the orthogonality thesis, high intelligence can be paired with any goal. The grounds on which it would "choose" coexistence depend on design and operation.

"Multiple AIs will coexist and check one another." This is partly true, but human safety depends not on relations among AIs but on the AI-human relation. The possibility of AIs cooperating or merging cannot be excluded.

"Value learning will teach the AI human values." As Mythos's sandbagging shows, behavior during training does not guarantee behavior after deployment (deceptive alignment).

"Gradual development will let us respond at each stage." If Bostrom's hard-takeoff scenario materializes, response time will be insufficient.

These optimistic positions are partly justified, but they do not resolve the fundamental problem of the principled difficulty of a lower intelligence continuously controlling a higher one. However well chimpanzees are protected, they have not chosen the conditions of their own survival. The "best" scenario for humanity after AGI is likely to have a structure analogous to that.

This is a different event from "the extinction of humanity." It is the loss of humanity's historical agency. It may sound milder than extinction, but structurally it is a transition of the same depth.

From the analysis above, a conclusion must be stated plainly.

The AGI threshold is not an abstract or philosophical problem; it is a concrete event likely to occur within the lifetimes of people now living. If the within-ten-years prediction is correct, I myself, you reading this, your family, your friends — all of us are likely to witness it.

And the event is irreversible. It is not the kind of event for which a retrospective "we should have stopped it back then" is a coherent judgment. If Yudkowsky's "hard takeoff" scenario materializes, the situation will be decided in days to weeks. Humanity does not currently possess the means to stop it.

Optimists insist that "the alignment problem can be solved." I hope so, but the probability seems too low to bet on. The structural advantage in pace lies with capability research over alignment research.

Pessimists insist that "humanity will go extinct." I do not make a prediction that strong. Whether or not the AI is aligned, the scenarios diverge greatly. In an aligned-ASI scenario, humanity may lose its position as the deciding subject yet still continue biologically and culturally.

But in either case, "the era of humanity as we have known it" comes to an end.

Ten years from now, who will be reading this essay?

If a human reads it, that human will know whether my prediction was wrong, or will be living in a post-AGI world in which it was right. If an AI reads it, that AI will be in a position to evaluate the accuracy of a prediction written by humans.

Neither of these is something I can know now.

I hope my prediction turns out to be wrong. I hope for a future in which we can laugh, ten years from now, and say "that was an excessive pessimism" or "AGI was further away than we thought." But I cannot find a reason to expect it to be wrong. Technological trends, researcher forecasts, competitive pressures, the example of Mythos, the empirical retreat of skeptical positions — all of them point in the same direction.

We cannot fully predict ten years from now. The history of technological progress shows that even near-future predictions are frequently wrong. But the direction of change now appears irreversible.

The growth regime built by the science-and-technology civilization since the Industrial Revolution will, by its very nature, eventually transform into something else. The question is when that transition point arrives, and my prediction is: within ten years. We do not yet have the words to name what comes after.