AGI Has Arrived — From Where?

The arrival of AGI can be defined later. The transfer of judgment cannot.

1. “AGI has arrived.”

After OpenAI released its new model, GPT-6 Astra, NVIDIA CEO Jensen Huang posted a short, forceful sentence.

“AGI has arrived.”

Huang wrote that Astra had been trained on “~100K+ NVIDIA Grace Blackwell NVLink72,” noted that only four years separated ChatGPT, o1, and Astra, and added that “400K GPUs” were coming online next. There is no need to reinterpret the hardware count. What matters is the shape of the statement: a steep rise in compute and capability, ending with the declaration that AGI had arrived.

I do not want to decide here whether Jensen Huang is right or wrong. His sentence left me with a different question.

From where?

Where does one have to be standing for AGI to look as though it has already arrived?

We are all looking at the same technology, but not from the same place. And perhaps AGI is not, as we have long imagined, a single line that everyone crosses at the same moment.

2. Three people looking at the same technology

Imagine three people looking at the same AI system.

The first builds the computing infrastructure behind it.

What this person sees is a curve: more compute, larger training runs, and work that once needed several systems now done by one model.

From that vantage point, AGI may look less like a distant event than a point already passed.

The second person builds the model itself and sees something different: remarkable capability, and also what remains unresolved.

A model can outperform humans on complex problems and still fail in unexpected ways, or miss a judgment that would be obvious to a person.

OpenAI describes Astra as its most intelligent and aligned model to date, built for longer and more demanding work—operating computers, writing software, doing research, completing professional tasks.

Then there is the third person: the one receiving the consequences.

To this person, GPU counts and benchmark scores may not matter.

If their account has changed, if a file has been modified, if a task has been carried out on their behalf, the questions are far simpler.

What happened?

And who allowed it?

All three are looking at the same technology.

For the builder of compute, AGI arrives as a curve.

For an organization, as productivity.

For the person affected by its actions, as real-world consequences.

Perhaps AGI does not arrive for everyone at the same time.

3. A horizon, not a line

For a long time, we have imagined AGI as a line separating what is not yet AGI from what is.

Real progress is not that clean. AI became better at language and reasoning; it learned to interpret images and speech, to use tools, to operate software, to carry a task across many steps. None of these arrived on the same day, and the boundary kept moving.

So at exactly what point should we say, “This is AGI”?

A researcher might ask how well a system generalizes to unfamiliar problems. A company might ask whether it can complete economically valuable work on its own. A user asks something plainer: “Has this AI actually started doing the work I used to do?” These questions are simply measuring different things.

That is why I have come to think of AGI less as a line and more as a horizon. A horizon depends on where you stand: from one position it looks close, from another still far away. And when we walk forward, it moves with us. What looks like the standard for AGI today may not feel sufficient tomorrow.

If the AGI debate refuses to end, it may be because we are using the same word while standing in different places, looking at different horizons.

4. Looking back from the future

OpenAI President Greg Brockman offered a different way of thinking about time. Asked whether Astra might be the model that marks the arrival of AGI, he replied, “I think it might be about this model,” and closed the briefing with, “Welcome to the AGI era.” He was not fixing a definition so much as describing how this moment might look from the future.

That perspective matters because AGI may be recognized after the fact. A transition can be real before everyone agrees on the date, or even on the name. We may eventually choose 2026, or some much later model. Definitions take time.

We may spend years deciding when AGI arrived.

But we cannot spend those years deciding when governance of AI should begin.

Here two kinds of time part ways. The definition of AGI can be settled in hindsight. Actions taken in the world cannot.

5. A name and an event are not the same thing

AGI is a name we give to a capability, and a name requires a definition.

How general must it be? How independently must it act? Is human level enough?

Move the criterion and the location of AGI moves with it. One person can call today’s AI AGI; another can say something essential is still missing.

Action is different.

If a file has been deleted, it has been deleted.

If access to an account has changed, it has changed.

If money has moved, it has moved.

If a command has been sent, it has already been sent.

Whether or not we call the AI that did it AGI does not change the event.

AGI is a name we attach to capability. Action is an event that remains in the world.

A name can remain disputed. A consequence cannot.

This matters because the absence of a definition can become a reason to postpone a different decision: how much judgment, and how much authority to act, we are willing to hand to a system.

We may grant AI the authority we would reserve for AGI before we have agreed to call it that.

From that moment the problem is no longer a word.

It is who holds the initiative in judgment.

6. Describing a capability and granting it authority are not the same act

This brings me back to Huang’s declaration. I am not arguing that he is wrong. Someone who has watched the computing power behind AI expand at close range might naturally see this moment as a historic threshold, and the people who build the models know their capabilities better than anyone.

But knowing a capability best and deciding how much authority to grant it do not have to be the same act.

We already understand this structure elsewhere. Imagine that the person who prepared a company’s books then audited them and reported, “I checked, and there is no problem.” The trouble is structural rather than moral: when the one who records and the one who verifies are the same, verification loses its independence. That is why accounting and auditing are separate. Auditing was not created because society decided not to trust people. It was created because trust grows stronger when the structure itself makes verification possible.

OpenAI itself makes the same distinction. The company has said Astra is the first model to reach the Critical level for cybersecurity under its Preparedness Framework: given the right tools and access, by its own description, it can find unknown security flaws and develop new attack methods against well-protected systems without a human directing each step. Its response was not “it is that powerful, so let everyone use it.” It strengthened its safeguards—isolation, monitoring, blocking—and limited advanced cyber capabilities according to who is asking and for what purpose.

The company that built the model is treating “what can it do” and “what will we allow it to do” as separate questions—a far more practical distinction than the label AGI. Capability is close to a technical fact. Permission is a social, organizational, and ethical judgment. And the less independent judgment there is between the two, the more easily an increase in capability becomes an increase in authority.

7. While we look for AGI, authority can move

This is what concerns me most. The lack of a shared metric for AGI does not buy us time. It may do the opposite, because while we fail to locate AGI, the distribution of authority between humans and AI keeps changing.

The change is not only in what AI can do. The order of judgment is changing too.

A person used to decide and a system executed.

Increasingly, the AI forms the proposed decision, selects the tools, sets the sequence, and shows a person the result.

Push that further and the person no longer judges; the person checks a judgment already made.

The shift does not happen in one dramatic moment, which is exactly why it is hard to notice.

We hand over one task because it is convenient, another because it is efficient, one more because the system seems to do it better.

Each looks small. None of them, by itself, feels like the moment authority changed hands.

But when enough accumulate, much of what people once judged directly may already be judged first by AI—and we may still be debating, “But can we really call this AGI?”

That is why “Has AGI arrived?” is not enough. A question comes before it.

While we are still unsure whether this is AGI, what have we already handed over?

8. Uncertainty is not permission to wait

In my last two editorials I dealt with other sides of this problem. I argued that an AI’s ability to remember something about a person does not give it the authority to use that memory, and that an AI’s becoming capable of an action does not give it the legitimate authority to perform it. Memory is not authority. Capability is not authority.

The AGI debate reveals a problem that comes earlier still. We may grant real authority to a capability before we have agreed on what it is. If so, waiting for certainty before we prepare reverses the order. Precisely because AGI is hard to identify, preparation is needed sooner. We can go on arguing where the line belongs. What we are delegating can already be seen.

Uncertainty may remain around the definition of AGI. It must not put human authority on indefinite hold.

Uncertainty is not permission to wait.

9. What matters is now

Perhaps in a few years we will say that Jensen Huang was right, and record 2026 as the year AGI arrived.

Or we will conclude that today’s systems were not yet AGI.

I do not know which, and I do not think we need to settle it today.

AGI may really be like a horizon. It looks nearer or farther depending on where we stand.

But that is exactly why something else matters.

Now.

While we are still locating AGI, parts of judgment and execution are already moving from humans to AI.

That movement does not wait for us to agree on a definition.

What we will entrust to AI, how far humans will keep the initiative in judgment, at which moments a person must step back in—these are not decisions that can wait until the name has been settled.

Perhaps AGI depends on where we stand.

But what matters is now.

by SeongHyeok Seo AAIH Insights Editorial Writer