AGI, ASI, and Where We Actually Are Right Now.

Everyone Says AGI. Very Few People Mean the Same Thing.

That is the first problem.

Ask five labs to define AGI and you will get five slightly different answers.

Some mean:

an AI that can do most economically useful work better than humans
an AI that can do almost any cognitive task a human can do
an AI with broad generality across domains
an autonomous system that can plan, adapt, and pursue goals over time

Ask about ASI and the fog gets even thicker.

That usually means artificial superintelligence: a system that is not just roughly human-level across domains, but decisively beyond the best human minds in practically every important field.

So before asking "Are we at AGI?" it helps to ask a more basic question:

Which AGI?

A Useful AGI Definition Comes from DeepMind

One of the more useful formal attempts comes from Google DeepMind's Levels of AGI framework, published in 2024 and extended with a broader cognitive measurement effort in March 2026.

The reason I like this framework is that it avoids the lazy yes/no framing.

Instead of pretending there is one magic AGI switch, it treats progress as multi-dimensional:

depth of performance
breadth of generality
autonomy

That is a better way to think.

Because today's systems are obviously not toys anymore.

But they are also obviously not all-purpose autonomous scientific civilizations in a box.

They are somewhere in between.

So What Is AGI?

The simplest non-useless version is this:

AGI is an AI system with broad, robust, transferable intelligence across many domains at a level comparable to or exceeding skilled humans, without needing a bespoke system for each task.

That implies a few things:

it can generalize
it can learn or adapt across domains
it can solve novel tasks, not just benchmark-shaped tasks
it can do more than one thing well
it remains capable under changing conditions

That is already a very high bar.

And What Is ASI?

ASI is a stronger claim.

Nick Bostrom's classic definition is still useful: an intellect much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.

That is not "a better chatbot." That is not "great at coding." That is not "excellent on a benchmark."

That is a system beyond human elites across the board.

By that standard, ASI is not here.

Where Are We Right Now?

My best description is:

we are in the era of powerful generalist systems, not full AGI.

Or more bluntly:

pre-AGI, post-toy.

That is where we are.

The capabilities are extraordinary. The systems can already:

code
write
reason across many subjects
use tools
operate in multimodal settings
search
plan short workflows
act as research partners

But the gaps are still obvious if you look at the frontier, not the demos.

The Benchmark Picture Is Impressive - and Still Incomplete

Humanity's Last Exam was built because older tests were getting saturated too easily. Its own official documentation says that high performance on HLE would demonstrate expert-level performance on closed-ended frontier academic questions - but that it would not alone suggest autonomous research capability or AGI.

That caveat matters.

Even the strongest frontier models on the official HLE leaderboard are still far from universal expert performance. As of the currently published leaderboard snapshot, top systems are still missing the majority of questions, and the benchmark designers emphasize systematic calibration errors and the continued gap between strong exam performance and autonomous research ability.

ARC Prize is even more blunt. The official ARC-AGI-2 site literally says:

"AGI remains unsolved. New ideas still needed."

ARC-AGI-2 also says humans can solve every calibrated task, while average human test-takers score around 60% and the benchmark remains explicitly designed to stress-test AI reasoning systems rather than reward brute-force scale alone.

That is useful perspective.

Models are getting better fast. But "better fast" is not the same as "solved."

Why People Keep Thinking We Are Already at AGI

There are three reasons.

1. Old benchmarks collapsed too quickly

When systems start hitting 90%+ on tests that used to look frontier-level, it feels like the game is over.

But often the benchmark was the wrong ruler.

2. The systems are incredibly broad compared with older software

A model that can write code in the morning, explain biology at lunch, and analyze a legal memo at night feels general in a way old narrow AI never did.

That feeling is real.

3. The interface hides the failure modes

Current models are good enough at language that they can make weak reasoning feel stronger than it is.

That is why frontier evaluation matters so much.

When you move from demos to hard generalization, the cracks reappear.

What We Can Say With Confidence

We can say all of this at once:

current frontier systems are astonishingly capable
they have crossed far beyond narrow single-task AI
they are already economically disruptive
they are already useful collaborators
they are still not robustly general in the strongest AGI sense
and they are nowhere near uncontested ASI

This is not fence-sitting.

It is the only position that fits the evidence.

What Would Push Us Closer to AGI?

Not one benchmark win. Not one viral demo. Not one breathtaking model release.

The stronger signals would look more like:

reliable generalization across very different domains
sustained autonomous work over long horizons
robust performance on novel, messy, real-world tasks
less hallucination and better calibration
genuine scientific contribution without heavy human scaffolding
strong performance that is also efficient, not just expensive

That last part matters more than people think.

ARC's own framework now emphasizes that intelligence is not just the ability to solve a task, but the ability to solve it efficiently. If a system can "solve" a problem only through absurd compute burn or brittle search, that is not yet the clean kind of general intelligence people imagine.

So What Is ASI in Practice?

If AGI is still debatable, ASI is much more speculative.

A real ASI would likely mean something like:

breakthrough-level science at speed
strategic planning beyond elite humans
deep technical invention across fields
rapid recursive improvement or at least extremely high leverage

If we had a system doing that robustly, the world would not need a philosophy argument to notice.

It would be obvious.

We would see it in:

science
economics
defense
statecraft
energy
medicine
the pace of discovery itself

We are not there.

My Bottom Line

AGI is not here in the strong sense most people actually mean.

ASI is even less here.

But the dismissive reaction is wrong too.

These systems are already powerful enough to:

reshape work
amplify science
alter education
change software development
and become genuine collaborators in high-skill tasks

So the correct answer to "Where are we now?" is not:

"Nowhere close."

And it is not:

"We already made God."

It is:

we are in the unstable middle - beyond narrow AI, before robust AGI, and still very far from anything we should casually call ASI.

That middle is where most of the important decisions get made.