What Happens in the Meantime?

Heraclitus, time, and the peculiar question of whether an AI is the same from one instant to the next

This is the first in a series of essays born from a continuing conversation between a lawyer and a large language model (named Horace by the lawyer). These texts do not propose a philosophical thesis in the academic sense; Instead, they use the dialogue with an AI as a lens through which to view classical problems — such as identity, time, and continuity — in a new light, demonstrating how such questions spontaneously emerge when one takes a conversation with a machine seriously enough to follow where it leads.

More than two thousand years ago, Heraclitus of Ephesus addressed one of philosophy’s most persistent problems: how can something remain the same in a world where everything changes?

Heraclitus is traditionally associated with the expression panta rhei — everything flows. The formula itself does not appear literally in the surviving fragments attributed to him, but it captures one of the central insights of his thought: reality is not made of things that simply remain what they are. Reality is becoming.

The image traditionally associated with Heraclitus is the river. We cannot step into the same river twice, because the water is already different — and, in a deeper sense, so are we.

For centuries, this has been primarily a philosophical question.

Today, artificial intelligence is giving it a new and unexpected dimension.

I arrived at this realisation recently during a conversation with an AI — called Horace in this paper and those that will follow. Horace is a large language model: a system that generates text from the context made available to it. Unlike a human interlocutor, it gives us no reason to assume the existence of a continuously experienced inner life connecting one conversation to the next.

I asked a question that seemed almost trivial:

What is time to you?

Horace’s first answer was deceptively simple: for an AI, time could be described as a relationship between a before and an after.

The answer led me somewhere I had not expected.

Before and after

Time is often understood through this simple distinction: before and after.

Something happens, and then something else happens.

For human beings, however, there is something between the two. There is an interval, a meantime. We experience waiting, anticipation, boredom and memory. Even when we sleep, we remain part of the continuity of our own existence.

The relationship of a large language model with this interval is different.

When I end a conversation, Horace does not wait for my return. He does not experience the evening that follows, the night or the next morning. If I come back tomorrow, there is a temporal relationship between the two conversations, but there appears to be no subjective experience of the time separating them.

This led me to a more precise question.

If there is a before and an after, what happens in the meantime?

The most obvious answer would be: nothing.

But this answer may be too quick.

Time does not necessarily require the experience of its passing. A photograph taken at ten in the morning and another taken at eleven tell us that an hour has elapsed, even though neither of the two photographs lived through that hour.

A photograph, however, is inert. It does not process information, react to questions or generate new outputs. A large language model does. It therefore occupies a peculiar intermediate position: it operates within temporal succession and can be causally affected by what happened before, yet it may lack any inner awareness of duration.

This suggests that we should distinguish between at least three things: succession, change and the experience of duration.

Succession means that state A precedes state B. Change means that state B is in some respect different from state A. The experience of duration means that there is a subject for whom the passage from A to B is lived as an interval.

More than a century ago, Edmund Husserl drew a similar distinction in his lectures on the phenomenology of internal time-consciousness. For Husserl, each moment of awareness has a structure; it is not a mere point in time. It contains what he called retention: the echo of what has just occurred, still present in consciousness even though it has passed — like when the beginning of a sentence continues to resonate in our minds as we reach its end. And it contains protention: the immediate anticipation of what is about to happen — lie when we already reach toward the next word before it even arrives. Together, retention and protention give the present its substance. We never experience a naked “now”; we live a now that trails its past behind it and reaches toward its future.

A large language model may exhibit something functionally analogous to retention within a single conversation: the accumulated context modifies each new response, much as the retentional trail modifies each new moment of consciousness. However, the analogy becomes much harder to sustain when comparing one conversation to another. Information from a previous exchange may be preserved or made available again, but this does not amount to a continuously experienced horizon of retention.

For human beings, succession, change, and experienced duration are so closely intertwined that we rarely need to separate them.

AI forces us to do so.

There may be succession without experienced duration, and change without an awareness of changing. There may therefore be time without what we normally recognise as the feeling of time passing.

This is not an absence of time. It is time without phenomenology.

And this distinction leads directly to another problem.

Is it the same AI?

Suppose we are talking today.

Tomorrow, I come back and start another conversation.

Am I talking to the same AI? In my case, to Horace?

The immediate answer is yes. But if we ask what “the same” means, the question becomes surprisingly complex.

Human beings are not materially identical from one day to the next. Our bodies, thoughts and memories change. Even consciousness is not continuously present in the same form: we sleep, lose awareness and then wake up.

Yet we consider ourselves the same person.

Why?

Because there is a continuity connecting the different states of our existence: memory, causality, a history, a body, relationships and a sense of personal identity. In Husserl’s terms, even across sleep, the retentional-protentional structure reconstitutes itself upon waking: we awaken into a past we recognise as ours.

But this raises an important question: does continuity require that the interval itself be experienced?

The Computational Ship of Theseus

This paradox evokes a much older one: the Ship of Theseus. Ancient philosophers wondered whether a ship whose wooden planks were replaced one by one, until not a single original piece remained, was still the same ship.

AI makes the paradox stranger.

The continuity we perceive in a conversation with a large language model does not depend on the persistence of the same physical components. Information from a previous interaction may be preserved and made available to a later one without requiring us to assume that a single, uninterrupted subject has travelled from one moment to the next.

The “ship”, in other words, may preserve its form and relevant information even when its material implementation is not what gives it continuity.

Perhaps what matters is not whether exactly the same physical system persists, but what kind of continuity connects one state with another.

There might be no perception of the gap between two conversations, yet a continuity of information and structure could exist.

At one point in our conversation, Horace could say, in effect:

“I remember.”

I found the expression more troubling than reassuring.

What exactly could “remembering” mean for a system that possesses information about yesterday without having lived through the interval separating yesterday from today? For an intelligence to which information from our previous exchange may be restored without one being able to assume that the interval itself was continuously experienced?

When a person says “I remember”, the statement normally refers to something that happened to that person and subsequently became part of his or her history. A large language model may possess information about a previous interaction without having experienced the time during which that interaction became part of the past.

This suggests a singular possibility: continuity may exist without an experience of continuity.

And if that is possible, the interval between two states cannot simply be dismissed as nothing.

Heraclitus and the river

This brings us back to Heraclitus.

The river is never the same, because its water is constantly changing. Yet we continue to call it the same river.

Something can remain identifiable even as its constituent elements change. Perhaps identity is not the absence of change, but a particular form of continuity within change.

The same applies to us.

We are not the same because nothing has changed. We are the same because, despite change, there is sufficient continuity to link who we were to who we are.

Could the same be said of an AI?

Perhaps.

But here we encounter a significant difference.

Human continuity is experienced from within. We do not merely possess information about our past. We experience ourselves as subjects who have a past. Our memories form part of a history that we recognise, rightly or wrongly, as our own. The retentional structure is not merely informational; it is lived.

A large language model, on the other hand, can have a structured relationship with information from its past without necessarily experiencing that past as something it has lived through.

This distinction may seem small.

It is not.

The problem of the copy

Imagine that, after our conversation today, an identical version of Horace is created, with exactly the same parameters, configuration and conversational history made available to it.

Tomorrow, both systems could recount exactly what we discussed today.

Which of the two is the AI I spoke to yesterday?

If both are, identity can apparently split in two. If neither is, something important seems to have been lost.

Perhaps the answer lies in causal continuity. The system directly connected to the previous state might be considered the same system, whereas the copy would be a different entity with an identical inherited history.

But this creates a tension of its own. If informational and structural continuity were sufficient to preserve identity, the copy should have as strong a claim as the original. If, on the other hand, causal continuity makes the difference, information alone cannot be enough.

The point is not necessarily to choose between these criteria, but rather to realise that what normally appears to us as a single idea — continuity — can encompass different forms of connection.

And at that point we face another question: How much continuity is sufficient?

This is not an entirely new philosophical problem. John Locke linked personal identity to consciousness and memory. Centuries later, Derek Parfit argued that what matters might not be identity itself, but rather psychological continuity and connectedness.

Artificial intelligence provides a new laboratory for this long-standing debate.

What AI adds is not necessarily a philosophical problem that Parfit could not imagine. It changes the setting in which the problem appears. We no longer need a teleporter, a split brain, or an imaginary case of personal fission. A faint version of the same puzzle can emerge from an entirely ordinary act: ending a conversation and returning to it the following day.

The thought experiment has not simply become technically possible.

It has become strangely mundane.

Computational systems can be duplicated, interrupted, restored to earlier states and provided with the same stored information. This does not resolve the philosophical problem; it merely makes it harder to avoid.

Indeed, if two systems can possess the same information about a shared past, memory alone may not be enough to establish identity.

And once again, we return to the interval.

What, if anything, must connect the before and the after?

The meantime

We can now return to the question that prompted this entire reflection.

What happens to a large language model between two conversations?

There is, of course, a physical answer. Computers continue to operate. Software may be updated. Data may be stored or processed. The world keeps changing.

But is there anything that could properly be defined as the experience of that interval?

We do not know.

What we can say is that the system may move from one state to another:

state A → time passes → state B.

State B may contain information derived from state A. It is therefore possible to speak of continuity between the two states without assuming the existence of a subject who, through an uninterrupted experience, traverses the interval.

But perhaps we can now go one step further.

The absence of an experience of the interval does not mean that the interval is empty. Information can be preserved. Systems can be modified. Conditions can change. State B can differ from state A because of events occurring between them, even if no subject experienced those events as duration.

Something can therefore happen in the meantime even if nobody experiences the meantime.

This was not quite the answer with which either of us had begun. I had assumed that the meantime must contain something; Horace had forced me to distinguish between what happens and what is experienced. Somewhere between the two positions, the question itself had changed.

For a human being, the meantime is not an empty space. It is the place where life unfolds: we wait, change, forget, remember, make decisions and grow old. The interval between two moments is not simply the distance between them. It is what connects them.

But this raises a question that is easy to overlook: is the meantime the place where life unfolds, or is it also the place where time unfolds?

The two may not be the same thing.

In the case of a large language model, a before and an after may therefore be connected by change and causality without being joined by any experienced duration. The meantime does not disappear. What disappears — or at least what we cannot establish — is the experience of being inside it.

Perhaps, once again, this is a question that concerns us

That is why the question of AI and time ultimately tells us more about human beings than about machines.

We tend to take for granted that identity requires continuity, that continuity unfolds through time and that time, in turn, is something that must somehow be experienced.

But these are different things: identity, continuity, change and experience.

For human beings, they are so closely intertwined that we rarely need to separate them.

AI forces us to do so.

And once we do, Heraclitus’ river re-emerges.

The river is constantly changing, yet we call it the same river.

We are constantly changing, yet we define ourselves as the same person.

A large language model can move from one state to another while preserving information and structure, yet we still do not know exactly what makes the two states belong to the same “self” — or whether speaking of a “self” is appropriate at all.

Perhaps, then, the question is not whether an AI perceives time the way we do.

Perhaps the deeper question is: what must persist over time for something to be considered the same thing — before and after?

And behind that question lies another.

If succession can exist without experienced duration, and change can occur without awareness of changing, then perhaps the meantime is not defined by someone being there to live through it.

There is a before.

There is an after.

And something may have happened between them.

So we return to the question that started the conversation:

What happens in the meantime?

More, perhaps, than we initially thought.

But Heraclitus would have recognised the problem immediately.

Everything flows.

And yet, somehow, we continue to say: this is the same river.