Why (ai) Kant Do It

Bill Gates has just published an important essay: The turbulent AI era is here. The choices we make now are critical.

His central concern is difficult to disagree with.

Gates argues that (ai) is developing extraordinarily quickly, could profoundly improve health, education and economic opportunity, but could equally deepen inequality and disrupt employment on a scale for which our institutions are badly prepared. He therefore calls for a new system for managing the transition.

On that, he may well be right.

But buried within his argument is a much bigger technical assumption.

Gates writes that (ai) can now "replace and even exceed human cognition." He acknowledges that current models still make mistakes, but argues that the reliability problem is being fixed quickly as systems become better at checking and improving their own work.

This is where I think we need to stop.

Not because (ai) isn't becoming extraordinarily capable.

It is.

But because reliability is not the same problem as continuity, and cognition is not necessarily the same thing as cumulative intelligence.

To understand why, we need to go back 245 years. To Immanuel Kant.

Who was Kant, and what has he got to do with (ai)?

Immanuel Kant was born in Königsberg in 1724 and became one of the central figures of modern philosophy.

His Critique of Pure Reason, first published in 1781 and substantially revised in 1787, addressed a deceptively difficult problem:

How is coherent human knowledge possible in the first place?

Kant was writing long before computers, neural networks, transformers, GPUs or large language models. But the question he asked now looks surprisingly contemporary.

Kant's argument, enormously simplified, was that intelligence cannot consist merely of receiving information about the world. Experience provides content, but coherent experience requires structure. Space and time organise experience. Causality connects events. Different perceptions have to be synthesised into a continuing experience. And underlying this is what Kant called the transcendental unity of apperception: the requirement that different experiences can belong to one coherent consciousness.

Kant therefore changed the question. Instead of asking only: What do we know? he asked something closer to: What must already be structurally possible for coherent knowing to occur at all?

That question should make the current (ai) industry slightly uncomfortable.

Because we have spent hundreds of billions asking: how much can a machine know? How well can it reason? How much context can it process? How convincingly can it generate? How effectively can it check its own answers?

But we have spent much less time asking the Kantian question:

What structures must exist for a machine to maintain a coherent world through change?

Gates' reliability problem isn't the whole problem

Consider Gates' argument that models still make mistakes but that the reliability problem is being fixed quickly.

There is every reason to expect individual outputs to become more reliable. Models can reason for longer. They can use tools. They can retrieve external information. They can critique candidate answers. Multiple models can check one another. Verification systems can reject incorrect outputs. All of that matters.

But imagine a system whose individual answers become 99.999% reliable. A different problem remains.

What happens when today's correct answer changes something upon which tomorrow's correct answer depends?

That is not simply an accuracy problem. It is a continuity problem. And it is what we have been calling the:

AI Cumulative Hiatus Delta

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

In plain English:

(ai) can create the next version. It cannot yet reliably preserve what must stay true while changing only what needs to change.

That distinction sounds subtle until (ai) encounters the physical world.

Suppose an engineering model proposes a modification to an engine. The modification is correct. Another system subsequently changes the combustion chamber. That modification is also individually reasonable. A third changes the piston geometry. Again, perfectly plausible.

But the important question is no longer whether each individual answer was good. It is whether all three transformations remain compatible with everything that the preceding states made true.

Reality ultimately performs the verification. The piston either clears the cylinder head or it doesn't. No amount of linguistic confidence changes the result.

Plausible continuation is not continuity

Current generative (ai) is extraordinarily good at something that looks remarkably like continuity. Give it State 1 and ask for State 2. Then ask for State 3. Then State 50. Each can appear to be a perfectly sensible continuation of the previous state.

But:

"This looks like a sensible successor" is not equivalent to "This is a lawful continuation of the same thing."

That difference takes us surprisingly close to Kant.

When a horse walks from one side of a field to another, we don't perceive a succession of unrelated horses. We perceive the same horse changing position. When an engineering design undergoes 300 revisions, engineers don't normally regard them as 300 unrelated machines. Something persists through transformation.

So where does that continuity actually reside?

The interesting object isn't the state

At first, we might represent continuity like this: S₀ → S₁ → S₂ → S₃. Record every state. Preserve the history. Problem solved.

Except it isn't. Because simply possessing two states doesn't establish that the second is a legitimate continuation of the first. The more interesting object is the transformation between them.

Our recent work led us towards representing this using a commuting relationship. If an object changes under transformation φ, while its relevant properties or representation change under ψ, continuity requires those transformations to remain coherent:

Sₜ₊₁ ∘ φ = ψ ∘ Sₜ

Something important happens here. Continuity moves out of the snapshots. It exists in the lawful relationship between transformations. If the relationship commutes, we have evidence of continuation. If it doesn't, the resulting state may be a replacement — however plausible it looks.

That is a fundamentally different computational problem from generating the next likely state.

Lawful change doesn't mean freezing the world

There is another complication. An intelligent system cannot simply preserve everything. Reality changes. Components wear. Scientific theories are superseded. Children become adults. Contracts are amended. Engineering requirements change. Organisations restructure.

An intelligent continuity system therefore cannot operate according to: Nothing important may change. It needs something closer to: These transformations are admissible while preserving this identity under these conditions.

Call the family of permissible transformations Γₖ. And immediately we encounter an even deeper problem. Who defines Γₖ?

If humans must manually provide the machine with a catalogue containing every permissible transformation, we haven't really created general intelligence. We've created an extremely elaborate rulebook.

The radical question is whether some of Γₖ can instead be derived from the intrinsic structure, constraints and dependencies of the thing itself. Put more simply:

Can the machine determine what a thing is allowed to become without ceasing to be that thing?

That is not a bigger language-model problem. It is an architecture-of-intelligence problem.

Kant's question returns

This is why Kant matters. Not because an eighteenth-century philosopher somehow predicted transformers. He didn't. And certainly not because Kant supplied us with a convenient mathematical architecture for (ai). He didn't do that either.

What he recognised was more fundamental: Coherent intelligence requires organising conditions. Experience isn't simply an accumulation of observations. Something makes successive experiences belong to a coherent world.

So perhaps the question facing (ai) isn't merely whether machines can exceed humans at particular cognitive tasks. Perhaps it is:

Can those cognitive events become part of a persistent, coherent and lawfully evolving state?

That distinction separates repeated intelligence from cumulative intelligence.

Being brilliant a million times isn't necessarily intelligence

Imagine an (ai) system capable of solving one million extraordinarily difficult problems. Problem 1: solved. Problem 2: solved. Problem 999,999: solved. Problem 1,000,000: solved. It sounds superintelligent.

But now add one requirement:

Solution 1,000,000 must preserve every still-valid dependency created by solutions 1 through 999,999 — unless it can establish why a particular dependency may legitimately change.

Suddenly we have a very different machine. It must know not merely what has previously been said. It must know what previous states oblige subsequent states to preserve.

More context doesn't automatically provide that. More parameters don't automatically provide it. Better retrieval doesn't automatically provide it. Self-critique doesn't automatically provide it. Even perfect memory wouldn't necessarily provide it.

Remembering the past is not the same as being constrained by what the past made true.

That is the AI Cumulative Hiatus Delta.

This matters to Gates' argument

Gates may be completely correct that (ai) will transform education, medicine, science and employment. Indeed, the more capable these systems become, the more important his governance argument becomes. His concern that existing institutions were not designed for a technology spreading this rapidly deserves serious attention.

But there is another layer of infrastructure we may need. Not simply infrastructure governing humans' use of (ai). Infrastructure governing how machine-generated knowledge itself becomes cumulative.

Where did this claim originate? What evidence supported it? What depended upon it? What transformations has it undergone? What must remain invariant? What has subsequently contradicted it? Which decisions consumed it? And under what conditions may it legitimately change?

These aren't merely questions of ethics. They're questions of knowledge architecture.

Without that layer, increasingly capable models may produce an extraordinary quantity of increasingly convincing intelligence without producing an equally dependable cumulative knowledge system.

From (ai) towards empathetic intelligence (ei)

This may ultimately mark the distinction between today's (ai) and what we describe as empathetic intelligence (ei).

An (ai) system can produce a representation of the world. An empathetic intelligence would have to exist in a continuing relationship with it. Actions create consequences. Consequences change state. Changed states alter future possibilities. Failures cannot simply disappear when a new context window begins. Contradictions have consequences. Dependencies matter. Origin matters. And history constrains what can legitimately happen next.

Its fundamental question therefore changes from:

What is the best thing to generate next?

to something closer to:

Given what exists, what has happened, what depends upon it and what must remain invariant, what transformations are admissible now?

Generation can still happen. But generation occurs inside continuity, rather than impersonating it.

Why (ai) Kant do it

Bill Gates is asking society to prepare for machines that may exceed human cognition. We should.

But before declaring the cognition problem solved, Kant gives us one more question to answer. Not: How intelligent is the next answer? But: What makes the next answer part of the same continuing understanding?

That distinction could prove enormously important.

Current (ai) is becoming spectacularly good at asking: What could come next? Cumulative intelligence requires something harder: What may come next while preserving everything that must remain true?

And the deepest version of the question is harder still:

Can the system derive those obligations from the structure of reality it is changing?

Until it can, we should be careful about equating increasingly powerful cognition with cumulative intelligence.

The missing capability isn't simply memory. It isn't simply reliability. And it isn't simply reasoning. It is the gap between succession and continuation. Between generating another state and establishing a lawful transformation. Between remembering yesterday and being constrained by what yesterday made true.

Gates is right that we need to prepare for a turbulent (ai) era.

But perhaps, 245 years later, Kant has identified one of the pieces of infrastructure we haven't built yet.

(ai) can increasingly tell us what should come next. It still Kant reliably establish what is allowed to come next.