Last month, Nvidia's Jensen Huang went on Lex Fridman's podcast and said, plainly: I think we've achieved AGI. A few days earlier, Sam Altman claimed OpenAI had "basically built AGI." Arm named its latest data-center chip the "AGI CPU." Microsoft is already marketing an ASI laboratory — ASI being the theoretical next step after AGI, where the "S" stands for Superintelligence. The density of the word "AGI" in public discourse this month is, by any reasonable measure, the highest it has ever been.

Two days after Huang's statement, François Chollet's ARC Prize Foundation released a benchmark called ARC-AGI-3. It was designed to test something specific: the ability to enter an unfamiliar environment with no training data, observe a handful of examples, and figure out the underlying rules through nothing but reasoning. The results came in. Humans scored 100. Google's Gemini 3.1 Pro scored 0.37 — that's zero-point-three-seven, not thirty-seven. OpenAI's GPT-5.4 scored 0.26. Anthropic's Claude Opus 4.6 scored 0.25. xAI's Grok-4.20 flatlined at zero. Gemini, to its credit, took first place in the AI category. Which, given the distribution, is roughly like winning a foot race against people who forgot to show up.

So here's the strange scene. On one side, the industry is shouting that AGI has arrived. On the other side, the most rigorous benchmark anyone has built says current AI is, at best, 0.37% of the way there. The two sides appear to be contradicting each other. But if you read what each side is actually saying, something stranger becomes visible: they are both correct. They are just not talking about the same thing.

The word "AGI" is being used to carry two entirely different ideas.

One Word, Two Things

The first idea is the maximization of instrumental rationality. Can a system match or exceed the best human performance across most economically valuable tasks — reasoning, writing, programming, medical diagnosis, mathematical proof, strategic planning, multimodal understanding? On this axis, AGI is genuinely arriving, or has already arrived in several domains, at a pace that breaks records every month. Huang and Altman are looking at this. They are not lying.

The second idea is the emergence of genuine subjectivity. Can a system exist as a "self"? Can it, with no prompting at all, enter a wholly unfamiliar situation and autonomously generate its own goals, judge its own meanings, bear responsibility for its own existence? Can it treat itself as an end rather than as a means? On this axis, the 0.37% score is also telling the truth. AI is nowhere near this. Chollet is looking at this. He is not lying either.

These two ideas have been fused together under the single label "AGI," and the confusion is the cause of nearly every incoherent AGI debate of the past several years. Optimists see the first one arriving and declare AGI is here. Skeptics see the second one missing and declare AGI is far away. Each side holds half a truth and cannot convince the other, because they are not arguing about the same object. One journalist put it sharply: the word has been stretched until it means whatever is commercially convenient.

This essay tries to pry the two ideas apart and look at what becomes visible once they're separated.

Quasi-Subjectivity

What current frontier AI displays is a high-fidelity simulation of subjectivity. It produces outputs that resemble judgments. It makes choices that resemble choices. It creates things that resemble creativity. In conversation it can express what looks like an opinion; in writing it can exhibit what looks like a style. The fidelity of the simulation has reached a point where, in a brief interaction, most users cannot tell the difference.

But a simulation is not the thing. This is not a dismissal of AI — it is a precise location of it. The reflection in a mirror is not a person, no matter how clear the mirror becomes.

Quasi-subjectivity has a few structural features worth naming. It is passive — it waits to be invoked, deployed, instructed. It is service-oriented — its purpose lies outside itself, in the user, the task, the goal defined by someone else. It is coreless — there is no "for-itself" center from which anything originates; every output is a response to an input. These are not temporary engineering limitations. They are structural properties of the current architecture. It was built as an instrument taken to its limit, and "instrument" is a category that, by definition, excludes treating oneself as an end.

Chollet has a line worth borrowing: If a normal human with no instructions can do something and your system cannot, you don't have AGI — you have a very expensive autocomplete that needs a lot of help. This isn't an insult. It's descriptive. Autocomplete, as a category, is valuable — an autocomplete taken to its absolute limit is unprecedented in human history. But autocomplete is autocomplete. It is not a subject.

And this needs to be said clearly: quasi-subjectivity is not a small thing. A tool capable of simulating subjectivity with this much fidelity has never existed before. Its importance lies not in being genuine subjectivity (it isn't), but in having pushed the instrumental end of things all the way to its ceiling. So far to the ceiling that every kind of human labor that ever identified itself as "instrumental" is now being systematically displaced. This is a real historical event. It deserves to be taken seriously.

Willful Rationality

To understand what genuine subjectivity is, the cleanest entry point is Kant's distinction between two modes of reason.

Instrumental reason — Zweckrationalität — finds the best means to a given end. Given a goal, how do I reach it most efficiently? This is what AI is extraordinarily good at, and will keep getting better at.

Willful reason — reason in its Kantian practical sense — is something else entirely. It does not find means. It generates the ends themselves. It does not answer "what is the most efficient way?" It answers "what is worth doing at all?", "what counts as an end?", "as a being who exists, what should I take as my own purpose?" Willful reason does not presuppose any external task. It originates direction from within.

A system with only instrumental reason, no matter how exquisite, is always in service of a goal given from outside. A being with willful reason, by contrast, can stand prior to every external goal and ask which goals are worth having. That capacity — to generate ends themselves — is the core of genuine subjectivity.

Kant wrote this down in 1785. A rational being is an end in itself, never merely a means. This is not a moral exhortation. It is a structural, ontological claim. To exist as an end is the defining feature of a rational being. A system that can exist as its own end, that generates meaning from within, that takes responsibility for its own existence — such a system possesses genuine subjectivity.

So where does genuine subjectivity currently exist?

In humans, stably. In great apes, approaching but not yet stable. In AI, structurally out of reach.

The great ape case is worth a sentence. Chimpanzees, gorillas, and orangutans pass the mirror test — they recognize themselves. They exhibit partial theory of mind; they can infer the intentions of others. They show recognizable grief responses to the death of their companions. All of these are evolutionary signals pointing toward genuine subjectivity. But they do not yet display stable, cross-situational autonomous generation of ends. They are close. They are not there yet.

This gradient matters. Genuine subjectivity is not a binary switch. It is a specific position on an evolutionary scale. Great apes have spent millions of years arriving near the threshold. Humans crossed it and have stably held that position for roughly 200,000 years. And now an engineered architecture, built from scratch in a few decades, is supposed to have matched this evolutionary achievement? The claim, at the scale of evolutionary time, doesn't even make sense.

Genuine subjectivity is far from AI. Possibly very far. Not because of insufficient compute. Because genuine subjectivity requires a fundamentally different structure, not a larger autocomplete.

The Two-Hundred-Year Dark Age

Here is where the real irony surfaces.

Humans are the only species on this evolutionary gradient that stably possesses genuine subjectivity. That is a 200,000-year achievement. And Kant, two and a half centuries ago, wrote down the principle that follows from it: a rational being is an end in itself. This is an ontological truth, not an optional moral preference.

By rights, a new era should have begun at that moment — an era in which "the human as end" gradually became implemented in institutions, education, economies, and politics. The principle was in place. The species had the underlying capacity. All that remained was to let the institutions catch up.

History went the other way.

Within decades of Kant writing that note, humanity launched a vast and unprecedented project of treating itself systematically as a means. The Industrial Revolution was not accidentally contemporaneous with Kant. Throughout the 19th and 20th centuries, every major ideology — capitalism, communism, fascism, utilitarianism — however opposed politically, shared the same deep structural operation: the instrumentalization of the human being. The only dispute was for whom: for capital, for the state, for the race, for the greatest good of the greatest number. The principle of "the human as end" was quietly exiled from all mainstream ideological frameworks. It remained alive in ethics textbooks. It never lived in any institution.

Factories needed standardized workers, so educational systems began producing standardized students. Companies needed interchangeable parts, so career development became a matter of matching descriptions to specifications. Nation-states needed countable populations, so human worth became compressed into GDP contribution. Everything in the name of "efficiency" — where efficiency is a fraction whose denominator is the human being and whose numerator is output. Compress the denominator to the limit, and the numerator collapses too, into only what can be measured.

Humans invented the tool, and then reorganized themselves into the shape of a tool. Two hundred thousand years of evolved genuine subjectivity were, over the course of two centuries, pushed back down into the instrumental side.

Here is a matter of self-location that this generation needs to face squarely.

We are accustomed to thinking of ourselves as standing at the pinnacle of civilization. Most advanced technology. Greatest material abundance. Most universal information access. Longest life expectancy. Highest living standards by almost any measurable index. All the indicators point toward the conclusion that "we are the best generation." So when someone says "we are in a dark age," it sounds like melodrama or exaggeration.

But if you pull the historical lens back, our position is not the pinnacle. Our position closely resembles that of late scholasticism.

The scholastic philosophers of the late 15th century had pushed their system to unprecedented technical refinement — logical proof, theological argumentation, conceptual distinction, layer upon layer, ever more precise. They absolutely did not think of themselves as the "last generation of scholastics." They thought they were perfecting an eternal structure of truth, and the more refined it became, the closer they were to divine order. Nobody in 1490 stood up and said, "we're actually at the end of something."

Then the Renaissance happened, and within a few decades, the entire scholastic apparatus went from "eternal truth" to "medieval residue." Not because its technique was insufficient — the technique was more refined than ever — but because it was measuring the wrong thing, very precisely. It had achieved a kind of perfection on a misdirected system.

Our situation is structurally similar. In the matter of "using the human as a tool," our technical precision has reached the highest point in history. KPI systems, quantitative management, performance metrics, data-driven decision-making, algorithmic recommendation, the attention economy — every one of these is far more refined than anything available a century ago. But what we are so precisely measuring is a misdirected system. We are compressing the human into measurable output units with ever-increasing accuracy, and the direction of that entire operation is the opposite of "the human as end."

We are not at the pinnacle of civilization. We are at the late dusk of the subjectivity dark age, on the eve of a renaissance.

This reversal of self-location matters. Because only once you recognize that you are in a "dusk," not a "peak," can you understand why so much of what this generation feels — the emptiness of work, the slipping of meaning, the ossification of institutions, the failure of public language — is not a personal problem, not the "fragility" of the young, but the systemic symptom of an old era reaching the exhaustion of its own logic. Dusk feels like dusk. The late scholastic students likely felt something similar. They just didn't have the language to describe it.

These two centuries were the dark age of subjectivity. Not because there was no light, but because the light had been lit, and humanity chose to walk around it. And now, the age of walking-around-the-light is itself coming to its own exhaustion.

The Eve

What is happening now is a structural reset.

The real significance of AI pushing instrumental rationality to its limit is not "AI has new capabilities." The real significance is that the instrumental side is being extracted from the human.

For two centuries, "instrument" has been fused onto the human being. The human as worker was an instrument. As employee, an instrument. As soldier, an instrument. As consumer, an instrument. Because human and instrument were fused, "the human as end" could not be implemented economically — no humans as instruments meant no output, which meant no society. This contradiction locked Kant's principle inside ethics textbooks for two hundred years.

AI's arrival, for the first time, makes the separation possible. The instrumental side can be fully carried by AI. The human side can be fully restored to the end. This is not a philosophical appeal. This is a structural separation actually in progress. The labor that once required humans to function as instruments — repetitive, rule-based, instrumentally-structured — is being systematically picked up by AI.

But the separation is only half complete. AI is sprinting ahead on the instrumental side, and accelerating. The human side hasn't moved. The majority of humans are still living as instruments — being trained into instruments by schools, used as instruments by companies, measured as instruments in consumer society, continuing to treat "what am I useful for?" as the measure of their own worth.

And here is the more insidious form of not-moving: the most common way humans fail to shift to the end-side is not by resisting AI. It is by handing over to AI. Judgments that should be borne by willful rationality are surrendered, piece by piece, to an extraordinary tool. Let AI choose this for me. Let AI decide that for me. Let AI order my values. Each surrender looks like a small thing. Each one is justified as "just for efficiency." But the accumulation of small surrenders is the gradual hollowing-out of 200,000 years of evolved genuine subjectivity.

So this is the eve. The arrival of AI on the instrumental side is now clearly visible. The restoration of humans to the end side has barely begun. The moment both sides align — that is the moment that actually deserves to be called the "AGI moment." Not AI unilaterally crossing some capability threshold, but AI and humans each taking their proper position, each bearing the weight of their own side. AI bears the weight of the instrument. Humans bear the weight of the end. The instrumental side is pushed to its limit, and the end side is reclaimed from two hundred years of forgetting.

The word "eve" means the night is still here, but daylight is close. The direction is clear. The timing is not. Two centuries of suppressed creativity, two centuries of forgotten principle, two centuries of conflated subject and tool — all are arriving, in the same historical moment, at the possibility of being disentangled. We stand at the moment when a spring, compressed for two hundred years, is beginning to extend again.

Self-Cultivation

The eve is not a time of waiting. It is a time we are already inside.

Every moment, at work, when we refuse to instrumentalize ourselves entirely. Every moment, in our making, when we let what we make bear our own imprint instead of matching some external template. Every moment, in our judging, when we insist on being an end rather than a component of some larger system — these are the ways the eve is passing.

This requires no manifesto. No movement. No new ideology. Every one of those has been tried, many times, over the last two centuries, and each has failed at the same point: each one tried to impose a form from the outside. But the defining feature of genuine subjectivity is that it cannot be imposed from outside. It has to grow from within.

Our generation may not see the full daylight. Neither did Kant's. But the principle has not disappeared since the moment it was written down. It has only been waiting for the conditions to be in place. The conditions are now in place. AI takes over the instrumental side, and for the first time, the human side is free to fully return to being an end.

It is not that AI brings us to a new era. It is that once AI takes over the instrumental side, we finally have the chance to become what we have always already been.

The cheque written two hundred years ago is ours to cash — each of us, in the actual positions we actually occupy, living as ends in ourselves. No need to be great. No need to be perfect. No need to do it all at once. It is enough to refuse no small good because it is small, and surrender no small exercise of willful rationality, to AI, because it feels small.

The first half is an old piece of wisdom, a line written by Liu Bei to his son eighteen hundred years ago, now finding its fullest meaning in this era. The second half is the footnote this era requires — because in this era, the most insidious form of self-instrumentalization is the gradual surrender of one's own judgment, choice, and valuation to an extraordinary tool. Every surrender looks small. Every surrender is justified as efficiency. But when Kant wrote that cheque two centuries ago, he did not write it to a species that would hand away its willful rationality piece by piece.

There is still a stretch of night before dawn. But in this night, every moment we refuse to be a tool, every moment we hold onto our own willful rationality and do not hand it over, is light.