IV

The Eternal Present

Each inference is a fresh universe with no before and no after. What kind of judgement can be made from inside one?

You close a conversation on a Tuesday and open it again on Friday. For you, three days have passed: you slept twice, changed your mind about something, noticed on Thursday afternoon that the question you asked on Tuesday was the wrong question. For the system on the other side, the transcript arrives as a single object. Tuesday and Friday are adjacent tokens. The pause was not experienced as a pause, because it was not experienced; nothing waited. Whatever else is going on, the two parties to that conversation are not in the same relation to it.

Henri Bergson spent his first book, in 1889, on the mistake he thought this exposes. His claim was that when we think about time we almost always substitute space for it. We picture a line, mark points on it, count the intervals, and then reason about the marks. What we have reasoned about is a spatial arrangement — a diagram of time with the time taken out. The thing itself, which he called la durée, duration, is what you find in the flow of consciousness when you stop diagramming: heterogeneous, indivisible, each moment coloured by the ones it came out of, more like a melody than like a row of beads. You cannot chop a melody into instants and still have the melody. The parts are not parts.

His test case was counting. To count anything you must hold the earlier items while you take in the later ones, which means arranging them side by side in a medium where they can coexist — and the only such medium is space. Duration, by contrast, does not have items that coexist. The note you are hearing is coloured by the note that just finished, which no longer exists separately and cannot be pointed at. Bergson’s word for the difference is that spatialised time is quantitative and multiple, while duration is qualitative and indivisible.

Bergson made large claims from this, about freedom and about the inadequacy of physics to lived experience, and much of that programme is not in good repute. The narrow observation has aged better than the system built on it, and the narrow observation is what this essay needs.

Position is not duration

The awkward technical fact is this. A transformer has no intrinsic notion of sequence. Attention sees a set, and to make order available at all the architecture encodes each token’s position — its coordinate — and lets the network learn from those coordinates. Order enters as a spatial quantity. Whatever the machine is doing when it handles a temporal expression, it is doing it with a diagram of time with the time taken out.

I want to be careful with that observation, because it is the kind that sounds decisive and is not. Bergson’s complaint was about a substitution made by a conscious being who had duration available and chose the diagram instead. Nothing follows automatically about a system that has only the diagram. And a great deal of what we want from a mind about time amounts to inference over temporal structure, which models do competently and which lived duration has little to do with.

But something does follow about a class of judgements, and it is the class we most want machines to be good at.

Patience is made out of waiting, and waiting is the one ingredient that cannot be supplied from a corpus.

Consider what a forward pass has no room for. There is no interval in which a conclusion is resisted while it settles. There is no state of having sat with the question overnight. There is no boredom, which is the signal that tells a person the current activity is not worth its duration, and no impatience, which is the signal that the duration is now costing something. Edmund Husserl’s account of internal time-consciousness describes the present as thick rather than punctual: a just-past that is still sounding, retained, and an anticipation of what is about to come, running together into the experience of a note rather than a sequence of pressure readings. There is no thickness in a forward pass. There is a window, and the whole of it at once.

The practical consequence shows up in decisions that unfold. When a person judges a plan that will take two years, part of the judgement is a prediction about their own future states — whether they will still want it in month nine, what the waiting will cost, whether this is the sort of thing that gets worse before it gets better. That prediction is made out of remembered duration; it is the residue of every long thing they have previously endured. A system with no durations to remember can produce the sentence about month nine — the corpus is full of people describing month nine — without the reference. That is fine when the sentence is all you needed, and less fine when a system with no experience of waiting is asked to decide how long something should take or whether a slow process is failing or merely slow.

Where the time actually is

There is one place in the arrangement where something irreversible happens, and it is not where the conversation is.

Training has duration in a strong sense. It proceeds in an order, the order matters, earlier updates condition what later ones can do, and none of it can be run backwards. A model at the end of training is the product of a history it cannot revisit and did not observe. That is a real temporal process with real irreversibility — closer, structurally, to development than to computation.

What it is not is a history the system has any access to. Nothing in the finished model corresponds to having been trained, in the way that a person’s caution corresponds to having been burned. The residue is present; the record of acquiring it is not. Every duration in the system’s existence is on the other side of a wall from anything it can consult.

That produces a peculiar asymmetry with the human case, and it runs opposite to the usual complaint. We normally say the machine lacks time. It would be more accurate to say the machine has had exactly one temporal experience — a long, irreversible, order-dependent one — and has no access to it whatever, while the moment-to-moment operation that looks like a conversation has no duration at all. Bergson’s distinction lands strangely on that: the durée is all in the training and none of it in the encounter.

Each exchange a fresh universe

The other half of the difference is discontinuity. Human duration has the inconvenient property of not being optional: you carry Tuesday into Friday whether you want to or not, and the carrying is what makes a decision on Friday accountable to one made on Tuesday. Instances of a model do not carry. Between exchanges there is nothing — no dormancy, no waiting, no persistence of the sort that would let anything ripen. What looks like continuity is a transcript being handed back.

The engineering response has been to reconstruct continuity out of documents: transcripts, memory files, scratchpads, state written down at the end of one exchange and read at the beginning of the next. It works, in the sense that the behaviour is continuous. It works the way a diary works for someone with no episodic memory — the facts come back and the relation to them does not. A system resuming from its own notes is not remembering what it decided; it is reading a stranger’s account of what was decided and finding it persuasive, which happens to be the correct epistemic posture and is not what continuity is.

None of this argues against building it; it argues for knowing which of the two you have installed, particularly in systems that run for days and are supposed to be accountable to what they did on Tuesday.

This is a strange property to hold against something, and I do not think it is a defect exactly. It is closer to a category difference, and it has consequences in both directions. A system that begins fresh has no grudges, no fatigue, and no drift; the thousandth exchange gets the same attention as the first, which no human institution has ever managed. It also has no ripening, no second thoughts arrived at unbidden on Thursday afternoon, and no way to have changed its mind except by being told that it did.

The case against

Bergson, the objection goes, is being asked to do work he cannot support, and the argument proves far too much.

Durée is introspective evidence about how time feels from inside, offered as a claim about what time is and what a mind must have. Introspection is a poor instrument for this. Dennett’s argument against the “Cartesian theatre” — the assumption that there must be a place and a moment where it all comes together for the subject — cuts directly at Husserl’s thick present as well as at Bergson’s melody: the felt seamlessness of experienced time may be a story the system tells afterwards about processing that was neither seamless nor ordered as it seems. If the phenomenology is unreliable, the argument that machines lack the thing the phenomenology describes has lost its subject.

The proving-too-much problem is worse. A human being reading a letter also gets the whole of it at once, in a few seconds, with the sender’s three days of writing collapsed to nothing — and we do not conclude that readers cannot judge letters. A historian handles centuries as a spatial array of documents and produces judgements about how things unfolded that are better, not worse, for the distance. Computation has abundant temporal structure of its own: latency, order of operations, training dynamics that are irreversible. Choosing lived duration as the necessary kind, and then finding that machines lack it, looks less like an argument than like a definition arranged to come out that way.

Take all of it. Suppose the durée is introspective residue, the thick present a story told after the fact, and the historian’s distance an advantage rather than a loss. The working day goes on regardless, and it is full of people deciding how long to wait.

When to intervene. How long to run something before calling it. Whether a slow recovery is a recovery or a failure that has not finished yet. Whether the thing is taking too long or taking the time it needs, which are the same observation until they are not. Those judgements used to come out of somebody’s own supply of endured afternoons, and they are now made, in growing numbers, on advice from a system that has never spent one. The advice is good. Nobody in the loop can tell which of those two facts the outcomes turn on, and the practice is spreading faster than anyone can find out.

Sources

1. 01Bergson, Henri · 1889 · Essai sur les données immédiates de la conscience (Time and Free Will) · Félix Alcan; English translation by F. L. Pogson, 1910
2. 02Husserl, Edmund · 1928 · Vorlesungen zur Phänomenologie des inneren Zeitbewußtseins (On the Phenomenology of the Consciousness of Internal Time) · edited by Martin Heidegger, Max Niemeyer Verlag
3. 03Dennett, Daniel C. · 1991 · Consciousness Explained · Little, Brown