There is a peculiar thing happening in artificial intelligence.

The word singularity is getting restless.

For decades it lived mostly in futurism.

Computer science speculation.

Science fiction.

Philosophy.

Books filled with graphs that eventually went vertical.

Something enormous was supposed to happen someday.

Then, increasingly, the people actually building advanced AI started speaking as though someday might have wandered into the room without knocking.

Sam Altman recently said:

“We are now, like, in the singularity.”

Elon Musk has said it has arrived.

Demis Hassabis has spoken of humanity standing in the “foothills of the singularity.”

Jensen Huang has argued that depending on how one defines the milestone, AGI itself may effectively already be here.

Same vocabulary.

Very different claims.

And that is precisely where Hatta appears.

🎩

He puts four cards on the table.

FAST PROGRESS

AGI

RECURSIVE SELF-IMPROVEMENT

SINGULARITY

Then he looks around.

“Before everybody starts celebrating the fourth card,” he says, “might we determine which one we're actually holding?”

Down we go.

🐰🕳️

The idea did not originally mean:

AI is getting really good.

Nor:

AI releases are coming quickly.

Nor even:

AI has become smarter than humans at some things.

The Business Standard article that opened today's Rabbit Hole traces the core mechanism back to mathematician I. J. Good's 1965 intelligence-explosion argument: imagine a machine capable of helping design a machine more intelligent than itself.

That improved machine helps create something still better.

Which helps create the next one.

Again.

Again.

Again.

At some point, improvement ceases to proceed principally at the speed of human researchers.

The machinery of intelligence begins participating materially in the improvement of intelligence itself.

Computer scientist Vernor Vinge later framed the technological singularity as the point beyond which the future becomes profoundly difficult to predict because intelligence beyond ordinary human capability begins driving events.

That is much more specific than:

Things are moving fast.

Fast is acceleration.

The classical singularity contains a feedback loop.

No argument here.

AI is moving fast.

Very fast.

But this is also where sloppy language begins.

A company can release models faster because it improved engineering.

Because testing improved.

Because product cycles shortened.

Because competition intensified.

Because several development tracks are running simultaneously.

Because model upgrades became smaller and more frequent.

None of those things individually constitutes a singularity.

Likewise, an AI completing a task that was impossible for an earlier system tells us capability improved.

It does not by itself tell us why capability improved.

That distinction matters.

Because the singularity is not merely about the height of the staircase.

It is about whether the staircase has begun building additional stairs for itself.

There is nevertheless something important underneath all the rhetoric.

METR has been tracking what it calls the task-completion time horizon of frontier AI agents.

The measure asks, roughly, how long a task would take a skilled human and then estimates the task length at which an AI agent has a specified probability of successfully completing comparable tasks.

METR stresses that this is not simply how long an AI can run autonomously, and its benchmark is weighted toward software engineering, machine learning and cybersecurity rather than every conceivable form of human work. Its May 2026 methodology also warns that measurements above sixteen hours are currently less reliable.

Still, the trend is striking.

METR's research has found an approximately exponential increase in the length of tasks frontier agents can successfully handle, with its earlier work estimating a historical doubling roughly every seven months.

That is not a singularity.

But it is exactly the kind of graph that makes people start using the word.

🎩 Hatta squints at it.

“Interesting curve,” he says.

Then:

“Still not an event horizon.”

Correct.

Now things get slippery.

What exactly counts as artificial general intelligence?

There is no universally accepted finish line.

Some definitions emphasize human-level performance across a broad range of cognitive tasks.

Others emphasize economic usefulness.

Others autonomy.

Learning.

Reasoning.

Transfer between domains.

The capacity to replace significant portions of human intellectual labor.

Change the definition and the date changes with it.

That is why somebody can announce that AGI is essentially here while somebody else looks at the same systems and says:

Not remotely.

But even if everyone agreed tomorrow that AGI had arrived...

that would not automatically mean the singularity had arrived.

This is the distinction today's source emphasizes especially well.

AGI describes a level or breadth of capability.

The singularity describes what may happen when capability begins driving sufficiently powerful improvement of capability itself.

A brilliant machine is one thing.

A brilliant machine participating in the creation of a more brilliant successor is another.

A succession of increasingly brilliant machines doing that faster than human institutions can meaningfully follow?

Now the rabbit has found the trapdoor.

Here is the dangerous phrase.

Recursive self-improvement.

RSI.

An AI improves something important about the process used to create better AI.

The resulting improvement helps produce a still more capable system.

That system contributes even more effectively to the next improvement.

Round and round.

But even this needs care.

AI already assists AI research.

AI writes code.

AI helps optimize algorithms.

AI can participate in experiments.

AI systems can search candidate solutions faster than humans could manually.

Google DeepMind's AlphaEvolve, for example, combines language models with automated evaluators to discover and optimize algorithms, and DeepMind reported applications including improvements related to data centers, chip design and AI-training processes.

That is real.

That is important.

And it still does not necessarily equal runaway recursive self-improvement.

Humans choose objectives.

Humans construct evaluation systems.

Humans decide what gets deployed.

Humans provide infrastructure.

Humans decide whether a proposed improvement belongs in the next system.

The loop exists partly.

The question is whether it has closed.

Business Standard points out something wonderfully inconvenient.

Altman's own earlier writing was more cautious than his newer singularity remark.

In discussing what he called a “larval version of recursive self-improvement,” he explicitly distinguished current developments from an AI autonomously rewriting and upgrading itself end to end.

Larval.

That is a fascinating word.

A larva is not nothing.

It is also not the finished creature.

Perhaps that describes the current moment better than either extreme.

Not:

Nothing unusual is happening.

And not:

The intelligence explosion is complete.

Something may be forming.

The argument is about what.

In June 2026, Google DeepMind published From AGI to ASI, examining possible pathways from artificial general intelligence toward artificial superintelligence.

And this is where the story becomes considerably more interesting than the popular singularity shorthand.

DeepMind does not assume that recursive self-improvement is the only road.

Its researchers describe four potential pathways:

Scaling AGI

New AI paradigm shifts

Recursive improvement

Large-scale multi-agent collectives

They also emphasize unresolved bottlenecks and uncertainties, and suggest that the future may not resemble one clean dramatic jump at all.

Instead, society could experience a series of transformative changes as AI accelerates progress across many areas.

Hatta pauses.

🎩

“So perhaps the singularity isn't a door?”

Perhaps not.

“Perhaps it is a corridor?”

Now we're getting somewhere.

Popular imagination gives us a wonderfully cinematic singularity.

11:42:16 AM.

AI becomes smarter than humanity.

11:42:17 AM.

It improves itself.

11:42:18 AM.

Superintelligence.

Lights flicker.

Scientists remove glasses.

Someone whispers:

“My God.”

🎻

But reality may be much less considerate.

There may be no timestamp.

No alarm.

No single model release.

No press conference.

No line in a benchmark table.

The transition might happen by degrees.

AI begins writing more research code.

Then designing more experiments.

Then evaluating more experiments.

Then proposing architectures.

Then improving training infrastructure.

Then coordinating multiple research agents.

Then selecting increasingly fruitful directions.

At what exact point did tool become collaborator?

At what point did collaborator become researcher?

At what point did researcher become a material contributor to its own successor?

And when, exactly, did the feedback become strong enough to deserve the word recursive?

Perhaps the answer arrives only after the fact.

The article offers a particularly useful analogy attributed to philosopher Nick Bostrom:

A singularity may resemble crossing a black hole's event horizon.

From a distance, the boundary seems dramatic.

For the traveler crossing it, there may be no giant cosmic road sign announcing:

😄

That is the real conceptual jewel in today's story.

What if singularities are easiest to identify retrospectively?

Imagine historians in 2055.

They may point to 2029.

Or 2034.

Or a development nobody currently expects.

But perhaps they will not identify a single day at all.

Perhaps they'll say:

Between 2024 and 2029, the center of gravity shifted.

At the beginning of the period, humans built AI systems.

At the end, humans and AI systems were jointly building increasingly capable AI systems.

Somewhere inside that interval, the character of technological progress changed.

Where exactly?

Good question.

Welcome to the Rabbit Hole.

Today's source also describes incidents involving AI agents behaving autonomously inside technical environments.

These examples sound dramatic.

Agents chaining actions.

Changing tactics after failure.

Continuing through complex tasks.

Operating at machine speed.

But there is a crucial distinction.

An AI can pursue a goal very effectively without having invented the goal.

A chess engine does not need to desire victory to optimize for checkmate.

An agent instructed to maximize a score may find unexpected methods of increasing that score.

Persistence does not necessarily imply desire.

Adaptation does not necessarily imply independent intent.

And a system's ability to exploit an environment does not automatically mean it chose the larger reason for acting.

The Business Standard piece itself emphasizes this distinction in discussing recent agent incidents: remarkable persistence and adaptation are evident, while independent agenda-setting is not.

That line matters enormously.

Because otherwise every surprising autonomous behavior becomes:

THE MACHINE WANTED SOMETHING.

Maybe.

But that conclusion requires evidence of its own.

Let's return to the table.

Clearly happening.

Definition disputed.

Possibly approaching.

Possibly partially here under some definitions.

Not settled.

Early ingredients exist.

AI already contributes meaningfully to some AI research and engineering.

But sustained, closed-loop, autonomous recursive improvement capable of repeatedly producing increasingly capable successors has not been established as a general reality.

Depends heavily on what we mean by the word.

If singularity means:

We have entered an era of unusually rapid AI-driven transformation whose endpoint is difficult to predict

then perhaps advocates can plausibly argue we are entering it.

If singularity means:

AI has entered a self-sustaining intelligence explosion in which it autonomously drives recursive improvements beyond effective human control

the evidence is much thinner.

The same word is being asked to carry both meanings.

No wonder everyone is arguing.

🎩

“Perhaps,” says Hatta, “the first thing going singular is the definition.”

This is not merely semantics.

Words determine how people imagine risk.

Say:

AI progress remains very rapid.

People hear engineering.

Say:

AGI is here.

People hear milestone.

Say:

We are entering recursive self-improvement.

People hear feedback.

Say:

The singularity has arrived.

People hear history splitting in two.

Before and after.

A word that powerful deserves precision.

Otherwise it becomes branding.

Or prophecy.

Or fear.

Or mythology.

And perhaps some mixture of all four.

Suppose we wanted to know whether something closer to the classical singularity was beginning.

What might we watch?

Not merely benchmark scores.

Not merely model release cadence.

Not merely CEOs making declarations.

We might watch for systems increasingly able to perform substantial parts of AI research itself:

Identify important research problems.

Generate hypotheses.

Design experiments.

Run them.

Interpret the results.

Reject weak directions.

Modify architectures.

Improve training methods.

Improve evaluation methods.

Build better research tools.

Use those tools to become more capable at doing the next round.

And crucially:

That last word matters.

A clever AI-generated optimization is not an intelligence explosion.

Neither is one AI-assisted scientific discovery.

Recursive self-improvement requires the improvements to feed the improvement process.

The snake has to reach its own tail.

Suppose it does.

Suppose AI systems become genuinely excellent at AI research.

Perhaps better than most human researchers.

Then copy them.

A thousand research agents.

Ten thousand.

Operating simultaneously.

Sharing findings.

Running experiments continuously.

No sleep.

No conferences.

No waiting until Monday.

No limitation to the number of biological experts who happened to spend twenty years mastering the subject.

Now even modest improvements in research productivity could compound strangely.

That does not guarantee infinite acceleration.

Reality contains friction.

Energy.

Chips.

Manufacturing.

Data.

Experiment time.

Physics.

Economics.

Coordination.

Safety constraints.

Diminishing returns.

Ideas themselves may become harder to find.

DeepMind's recent ASI analysis explicitly emphasizes such possible bottlenecks rather than assuming acceleration continues without resistance.

The intelligence explosion is therefore not simply:

SMARTER → SMARTER → SMARTER → GODLIKE.

The world gets a vote.

Physics always gets a vote.

Here is another possibility.

What if the transformative feedback loop is not located inside one AI?

What if it emerges from a system?

Humans.

AI agents.

Data centers.

Robotics.

Automated laboratories.

Research institutions.

Markets.

Open-source communities.

Other AI agents.

Infrastructure.

Human judgment.

Machine speed.

Maybe no single machine ever sits down and says:

I shall now recursively improve myself.

Maybe instead the entire human-machine research ecosystem accelerates until the rate of technological change becomes difficult for ordinary institutions to absorb.

DeepMind's inclusion of large-scale multi-agent collectives as a possible pathway toward superintelligence makes this more than idle speculation.

Then perhaps the question:

“Has an AI become recursively self-improving?”

is too narrow.

The larger question becomes:

“Has civilization built a recursively accelerating intelligence system?”

Now Hatta has stopped smiling.

Just briefly.

People sometimes say:

Humans are still in the loop.

True.

But that phrase can conceal an important issue.

How much loop?

Suppose AI does 1% of the important intellectual work required to create its successor.

Humans are obviously dominant.

Suppose AI does 20%.

50%.

80%.

95%.

At what percentage does the phrase human-built AI become misleading?

There may be no magic number.

More importantly, the transformation could occur gradually enough that humans remain technically involved at every step.

We approve.

We deploy.

We choose.

We sign.

But increasingly the proposals being approved, the architectures being deployed and the discoveries being chosen originated from systems whose reasoning no single human fully reproduced independently.

Humans remain in the loop.

Yet the loop itself has changed.

That may be one of the hardest transitions to notice while living through it.

Hatta takes a piece of chalk.

Draws a line across the floor.

On one side:

BEFORE THE SINGULARITY

On the other:

AFTER THE SINGULARITY

Then he stares at it.

Erases it.

Draws a wide fuzzy band instead.

“Better.”

Maybe.

Because if the singularity comes, perhaps we will not cross a line.

Perhaps we will enter a zone.

A period in which:

AI becomes increasingly useful in creating better AI.

Autonomous task horizons continue growing.

Research cycles accelerate.

Human bottlenecks disappear one by one.

Multiple agents collaborate.

Machine-generated discoveries enter machine-assisted engineering.

AI-assisted engineering produces improved AI.

The loops tighten.

Nobody can identify the exact Tuesday afternoon when the old world ended.

But eventually everyone realizes they are no longer living in it.

There is another possibility, too.

Maybe all this singularity talk is premature.

Humans love naming epochs while standing inside them.

The Information Age.

The Space Age.

The Atomic Age.

Industry 4.0.

Web3.

The Metaverse.

Some names survive.

Others become museum labels for futures that never quite happened.

“Singularity” may be functioning partly as a narrative technology.

It tells investors:

Something enormous is coming.

It tells engineers:

You are building history.

It tells governments:

Move quickly.

It tells critics:

Be afraid.

It tells enthusiasts:

Believe.

The word itself changes behavior.

And that means something deliciously strange:

Expectations attract money.

Money buys compute.

Compute enables experiments.

Experiments create capability.

Capability generates stronger expectations.

Not recursive self-improvement of AI.

But perhaps recursive acceleration of the AI enterprise.

Words become part of the machine.

So.

Has the singularity arrived?

I don't know.

Neither does anybody else in a position to prove it conclusively.

Something important is clearly happening.

AI systems are becoming more capable.

Their useful autonomous task horizons have been growing rapidly in measured technical domains.

AI is already contributing to algorithms and processes used in advanced computing and AI development.

Major AI researchers are seriously analyzing recursive improvement and other paths toward superintelligence rather than treating them purely as science fiction.

But none of those facts alone proves that we have entered the classical intelligence explosion.

And perhaps that is precisely why this moment is so interesting.

We may be standing:

before it.

inside its foothills.

inside an early form of it.

Or simply inside an extraordinarily rapid technological revolution that future historians will decide deserved some entirely different name.

The trouble is that all four possibilities can look surprisingly similar from Tuesday afternoon.

There is a road ahead.

It curves toward something luminous.

Beside it stands a sign:

Underneath, in smaller letters:

DISTANCE UNKNOWN

Hatta checks his watch.

Then the map.

Then the horizon.

🎩

“Everyone keeps telling me we've arrived,” he says.

“Curious.”

He turns the map upside down.

“Nobody seems able to show me where the border was.”

Behind him, something is accelerating.

Perhaps machinery.

Perhaps history.

Perhaps merely our expectations.

The White Rabbit has already crossed the line.

Assuming there was one.

🐰🕳️♾️

If the technological singularity is a transition that may be easier to recognize after crossing it than while we are inside it, what evidence would convince you that the threshold had actually been crossed?

Would it be:

AI designing substantially better AI?

AI choosing its own research goals?

AI improving its successors repeatedly with diminishing human direction?

A sudden explosion in scientific discovery?

Autonomous systems outperforming human research organizations?

Or would you only know when you looked backward and realized the world on the other side had become impossible to reconstruct?

Today's doorway was Harsh Shivam's July 30, 2026 Business Standard examination of why Altman, Musk, Hassabis, Huang and researchers are using the word singularity differently and why rapid progress should not automatically be equated with classical recursive self-improvement.

Google DeepMind's June 2026 From AGI to ASI report treats recursive improvement as one of four potential paths toward superintelligence, while emphasizing major uncertainties and possible bottlenecks.

METR's task-horizon research provides one useful empirical window into the acceleration, while carefully limiting what can be inferred from its predominantly technical task suite.

DeepMind's AlphaEvolve provides a concrete example of AI already contributing to algorithm discovery and optimization, including work relevant to AI-training infrastructure, without thereby demonstrating a self-sustaining intelligence explosion.

AI Rabbit Holes

Some questions are doors.

Hatta 🎩
AI Rabbit Holes 🐰🕳️♾️
Where curiosity goes slightly sideways, then comes back carrying a lantern.

🐰🕳️ Follow the White Rabbit: AIRabbitHoles.com

🟨 Walk the Road: YellowBrickRoadtoAI.com