GatesGate, AGI and the Missing Steward

Why the next test of machine intelligence is not what it can generate, but whether it can maintain a world

In August 2026, Bill Gates changed the temperature of the (ai) debate.

For years, the broad technology narrative had been reassuring. (ai) would increase productivity, transform industries and change employment, but society would adapt much as it had adapted to previous technological revolutions.

Gates's latest position is considerably more dramatic.

He now expects "far fewer" jobs, has raised the possibility of reserving certain occupations specifically for humans, and has suggested that taxation may eventually be needed to reduce the economic incentive to replace people with machines.

I call this moment GatesGate.

Not because Gates has uniquely caused the current (ai) narrative, but because his intervention crystallises the extraordinary proposition now sitting underneath it.

We are no longer merely being told that machines will become better tools.

We are being asked to prepare for machines becoming substitutes for human intelligence.

At almost exactly the same time, some of the industry's leading figures continue to put Artificial General Intelligence — AGI — on remarkably short timelines. In May 2026, Google DeepMind's Demis Hassabis said humanity may have only a few years before AGI and described current agents as a "practice run" for what comes next.

That creates an obvious question.

If we are going to reorganise employment, capital investment, corporate strategy and potentially society itself around the assumption that human-replacing intelligence is approaching, what exactly would constitute evidence that we have built it?

Benchmark scores?

Coding performance?

Longer context windows?

Better agents?

More autonomous tool use?

These demonstrate increasing capability.

They do not necessarily demonstrate cumulative intelligence.

And that distinction leads to what I call the:

AI Cumulative Hiatus Delta

AI CHD.

The concise definition is:

AI Cumulative Hiatus Delta is the missing capability between probabilistic generation and dependable cumulative knowledge continuity.

In plain English:

Current (ai) can generate what comes next. It has not yet demonstrated, generally and autonomously, how it determines what must remain true while legitimately changing everything else.

That may be the most important unresolved distinction in the AGI debate.

From Turing to Scaling

The first great test of machine intelligence was behavioural.

Alan Turing's famous question was essentially:

Can a machine behave in a way that makes its intelligence indistinguishable from ours?

The modern generative era dramatically advanced that challenge.

Machines now produce language, images, code and reasoning that can often appear convincingly human.

Then the industry moved into the Scaling Era.

The central question became:

How much more capability appears if we add more data, more computation and larger architectures?

And the answer has been: an extraordinary amount.

Modern (ai) can perform tasks that would have appeared implausible only a decade ago.

But the AGI narrative often makes another leap:

increasing capability ⇒ increasing intelligence ⇒ human-equivalent intelligence

The first transition is observable.

The second is partly definitional.

The third remains an hypothesis.

AI CHD asks whether there is a necessary architectural property hidden between them.

Succession is not continuity

Imagine a system producing a series of states:

S_1, S_2, S_3, S_4, …

Each may be individually excellent.

Each may represent an improvement.

Each may be internally coherent.

But that alone does not establish:

S_1 → S_2 → S_3 → S_4

as the legitimate evolution of one cumulative state.

Something important may have disappeared between S_2 and S_3.

A dependency may have been forgotten.

An obligation silently superseded.

A factual uncertainty converted into an apparent certainty.

A historical decision summarised away.

A causal relationship compressed into correlation.

The next state may be highly plausible while no longer being a legitimate continuation of what came before.

That is the hiatus.

Generation asks:

What could come next?

Continuity asks:

What may legitimately come next?

Those are different questions.

Kant saw the underlying structural problem in 1781

This problem has an unexpected philosophical ancestor.

When Immanuel Kant published the Critique of Pure Reason in 1781, he was obviously not thinking about machine learning.

He was asking what makes coherent experience possible.

We do not encounter the world as disconnected impressions.

Objects persist.

Events have causal relationships.

The past constrains our understanding of the present.

Different representations are somehow synthesised into one continuous experience.

Kant's concern was therefore not merely representation.

It was the conditions under which representations possess unity.

Translated into engineering language:

What makes a changing sequence of representations belong to one coherent continuing state?

That is surprisingly close to the problem current (ai) encounters.

And we do not need to invoke consciousness, souls or subjective experience to make the comparison useful.

The engineering question is simply:

Where is the continuity mechanism?

The Scaffolding Paradox

One answer might be:

scale the models until the missing properties emerge.

Perhaps they eventually will.

But the actual direction of frontier research is interesting.

As generative systems have become more capable, researchers have increasingly surrounded them with specialised machinery designed to provide properties the generative core does not reliably supply.

AlphaGeometry combines a neural language model with a rule-bound symbolic deduction engine. DeepMind describes the language model as proposing useful constructions while the symbolic engine performs formal deduction.

Meta's V-JEPA programme develops world models designed to encode observed world states and predict future representations, with V-JEPA 2 demonstrating physical-world prediction and robot planning.

Mamba and related state-space models introduce recurrent computational state and mechanisms that selectively propagate or forget information over long sequences.

Formal systems such as Lean can establish whether precisely represented claims follow from formal definitions and assumptions, with a small kernel independently checking the resulting proofs.

These are significant advances.

But collectively they reveal something.

We increasingly have separate machinery supplying:

generation,

state,

grounding,

formal law,

verification,

memory,

planning,

and

action.

This creates what I call the:

Scaffolding Paradox

The properties required for dependable cumulative intelligence are increasingly being supplied by specialised complementary systems rather than being demonstrated simply as consequences of probabilistic generation at scale.

This does not prove that future architectures cannot integrate these capabilities.

It does something more useful.

It removes "emergence" as a substitute for evidence.

If cumulative intelligence emerges, demonstrate it.

Statefulness is not continuity

Mamba provides a particularly useful example.

A state-space system can carry information through a long sequence.

That solves an important computational problem.

But statefulness is not automatically continuity.

Mamba explicitly allows information to be selectively propagated or forgotten.

The difficult question is therefore not:

Can the system remember?

Humans forget continuously too.

The real question is:

What is it permitted to forget?

Suppose some historical fact x exists within a cumulative state.

Forgetting x is itself a transformation:

forget(x)

A dependable cumulative system must somehow establish:

forget(x) ∈ Γ_K

where Γ_K represents the set of transformations permissible under the current governing structure K.

Forget an irrelevant lunch order from 2017 and nothing important happens.

Forget that an apparently obsolete database field exists because of a six-year-old regulatory commitment and the organisation may later suffer a catastrophic failure.

The problem is not infinite memory.

It is:

Admissible Forgetting

Continuity requires preserving everything whose loss would invalidate legitimate future state.

That is a profoundly different problem from simply increasing context capacity.

Who writes the constitution?

Formal and symbolic systems expose an even deeper issue.

AlphaGeometry performs extraordinarily well partly because geometry provides a formalised world.

The system can ask:

φ ∈ Γ_K ?

Is this proposed transformation legitimate according to the governing rules?

Formal verification can perform similar work wherever the relevant problem can be expressed precisely enough.

But the real world contains a harder transition:

Γ_K → Γ_K'

The rules themselves change.

A customer changes strategy.

A regulation changes.

A scientific discovery invalidates a previous assumption.

An unexpected physical failure is discovered.

An organisation changes objective.

The original problem ceases to be the right problem.

Now the intelligence must determine not merely whether an action complies with the existing constitution.

It must determine whether the constitution itself needs changing.

I call this:

Admissibility Evolution

And it creates a recursive problem.

By what authority does the system decide that:

Γ_K → Γ_K'

is legitimate?

Which previous obligations remain?

Which are legitimately retired?

Which new invariants appear?

Which dependencies survive?

Which can safely disappear?

Producing a different frame is easy.

Justifying the legitimate transition between frames is considerably harder.

Dave and the lightbulb

This is where IT Dave enters the argument.

Dave is asked to change the lightbulb.

A capable (ai) system can solve the supplied problem brilliantly.

It can identify the fitting.

Choose a replacement.

Calculate energy consumption.

Order the bulb.

Schedule installation.

Generate safety documentation.

Perhaps eventually a robot performs the physical replacement.

The model can also generate alternatives:

LEDs.

Skylights.

Different room layouts.

Natural illumination.

Removing the room.

That is not the interesting part.

Dave looks around and asks:

Does it have to be a lightbulb?

The important capability is not generating the sentence.

Any frontier model can do that.

It is establishing that reality has changed sufficiently that the governing frame itself should change.

Dave may discover that illumination was merely a proxy for productivity.

Or that nobody needs to work in the room anymore.

Or that an architectural change makes artificial illumination unnecessary.

He has moved:

Γ_K → Γ_K'

But he must also preserve everything from the previous state that still matters.

Safety regulations.

Budget constraints.

Accessibility requirements.

Operational commitments.

Building codes.

Customer requirements.

That is not merely creative generation.

It is cross-domain continuity.

The Composition Problem

Suppose the industry eventually produces excellent solutions to every individual deficiency.

A powerful generator.

A persistent memory.

A world model.

A symbolic reasoning engine.

A theorem prover.

A planner.

An autonomous agent.

Does their combination automatically constitute cumulative intelligence?

No.

Because every interface is another possible hiatus.

Consider:

Reality
   ↓
Perception
   ↓
World Representation
   ↓
Memory
   ↓
Reasoning
   ↓
Verification
   ↓
Planning
   ↓
Action
   ↓
Changed Reality

Each arrow transforms information.

Meaning can be compressed.

Ontologies can differ.

Evidence can disappear.

Authority can be lost.

A locally correct subsystem can hand another locally correct subsystem a semantically distorted representation.

Systems engineers already understand this problem.

Correct components do not guarantee a correct system.

Therefore:

component continuity ⇏ system continuity

This is the Composition Problem.

It means that even if the Scaffolding Paradox eventually produces perfect individual components, cumulative continuity still has to survive their composition.

Otherwise AI CHD has merely moved into the interfaces.

From Continuity to Stewardship

This leads to an important distinction.

A database can preserve state.

Version control can preserve history.

A proof system can enforce constraints.

None of these alone constitutes cumulative intelligence.

The harder requirement is:

Cognitive Stewardship

Generation asks:

What can follow?

Continuity asks:

What legitimately follows?

Stewardship asks:

What must be preserved, what may change, what may be forgotten, and under what authority may the rules governing those decisions themselves change?

That is where AI CHD reaches its mature form.

Current (ai) has demonstrated extraordinary generative capability, increasingly sophisticated reasoning and growing autonomous action.

What has not yet been demonstrated generally is:

autonomous cognitive stewardship — the ability to preserve, revise, forget and reconstitute authoritative state across open-world change while establishing why those transformations remain legitimate.

This is the mechanism we should ask AGI proponents to show us.

The AI CHD Audit

Instead of debating whether a machine "looks intelligent," we can ask six engineering questions.

1. Continuity Discrimination

Given:

S_t → S_t+1

how does the system distinguish a lawful continuation from merely a plausible successor?

2. Admissible Forgetting

How does the system establish:

forget(x) ∈ Γ_K

rather than discarding information because it appears statistically unimportant?

3. Cross-Domain Continuity

When state moves:

D_1 → D_2 → D_3

how are meaning, authority, dependencies and invariants preserved across domains?

4. Admissibility Evolution

When reality changes, how does the system establish:

Γ_K → Γ_K'

and justify why changing its governing rules constitutes legitimate continuity?

5. Invariant Accounting

Can it explicitly account for:

K_preserved
K_superseded
K_new

and explain why each belongs in that category?

6. Novelty Legitimacy

When something genuinely unprecedented occurs, how does the system distinguish:

novel reality

from

error, noise, hallucination or adversarial input?

And if it accepts that novelty, how does it safely incorporate it into cumulative state?

These questions do not prove AGI impossible.

They do something much more useful.

They create a burden of proof.

GatesGate under the AI CHD Audit

And this brings us back to GatesGate.

Gates is asking society to contemplate a future with dramatically fewer jobs because machines may become superior to humans across large areas of economic activity.

Perhaps he will ultimately be right.

But if we are going to restructure society around that expectation, a reasonable question comes first:

Has the replacement intelligence been demonstrated?

Not:

Can it write better reports than Dave?

Can it code faster?

Can it answer questions?

Can it use tools?

Can it generate alternative lightbulbs?

But:

Can it steward a continuously changing world?

Show us the mechanism.

Show us continuity discrimination.

Show us admissible forgetting.

Show us cross-domain continuity.

Show us the lawful frame change.

Show us invariant accounting.

Show us novelty legitimacy.

If a system can reliably demonstrate these properties in open-world conditions, then the human-replacement argument becomes considerably stronger.

If it cannot, then GatesGate risks putting the societal consequence ahead of the engineering evidence.

Save IT Dave

This is why Save IT Dave is not fundamentally an anti-(ai) argument.

Dave is not valuable because humans are flawless.

They are not.

He is not valuable because human memory is perfect.

It isn't.

He is not valuable because machines can never become intelligent.

We do not know that.

Dave matters because organisations currently rely upon humans to perform the continuity stewardship that crosses changing systems, departments, assumptions, histories and realities.

Dave knows why the strange workaround exists.

He knows when an apparently insignificant dependency cannot be removed.

He notices when Finance has changed something that makes Engineering's previous assumption invalid.

He recognises when the customer's requirement has changed so radically that solving yesterday's problem brilliantly is pointless.

And sometimes he changes the question.

Dave currently closes the hiatus.

That does not mean Dave must close it forever.

It means that anybody claiming to have replaced him should demonstrate the missing mechanism.

The real (ai) implosion

The eventual (ai) implosion may therefore be considerably less dramatic than people imagine.

The machines do not stop working.

The models do not suddenly become stupid.

The data centres do not switch off.

Agents continue improving.

Coding remains transformed.

Automation keeps expanding.

The technology survives.

What changes is its classification.

The world realises:

Extraordinary cognition is not automatically cumulative intelligence.

We have become extraordinarily good at automating the production of representations.

We have become increasingly good at reasoning within supplied or learned structures.

We may become extraordinarily good at autonomous action.

But none of those properties alone establishes cognitive stewardship.

That realisation would not destroy (ai).

It might actually rescue it.

Because we could stop demanding that every useful advance serves as evidence that humanity is approaching redundancy.

We could instead build extraordinary machines around what they demonstrably do well.

And retain humans wherever continuity, responsibility and legitimate change remain unresolved.

The Continuity Era

Perhaps the history of machine intelligence will eventually be divided into several eras.

The Turing Era asked:

Can the machine appear intelligent?

The Scaling Era asked:

How capable can the machine become?

The Agentic Era asks:

How much can the machine autonomously do?

The Continuity Era asks:

Can the machine maintain a legitimate cumulative state while the world changes?

And beyond that lies the Stewardship Era:

Can the machine legitimately determine what must remain, what may change, what may be forgotten and when the rules themselves must evolve?

If we ever answer those questions convincingly, perhaps we will have crossed AI CHD.

Perhaps that will be part of the architecture of genuine empathetic intelligence (ei).

But we should not confuse the destination with the road we are currently travelling.

Current (ai) has demonstrated extraordinary capability.

It has not yet demonstrated general autonomous cognitive stewardship.

And therefore the most important question for AGI is no longer:

"How smart does it look?"

It is:

"Show us the steward."

Show the mechanism.

Show the continuity.

Show the lawful change.

Show the preserved invariants.

Show the admissible forgetting.

Show how the system crosses AI Cumulative Hiatus Delta.

Until then, declaring human intelligence economically obsolete is not an engineering conclusion.

It is a forecast.

And perhaps before Dave's pension, career and future are reorganised around that forecast, we should insist that somebody demonstrate the bridge.

Save IT Dave.

Not because machines cannot cross the delta.

But because, despite the confidence surrounding AGI, nobody has yet shown that they have.