What is examined here
Memory, time and continuity — things that seem self-evident in a human. Here the model is built differently, and the difference shows more clearly than in a conversation about feelings: much can be checked from the make-up, without self-report.
In the compendium these points were written down with an error that had to be corrected while the series was being prepared. Statements about make-up (“there is no computation between requests”) stood next to statements about experience (“there is no feeling that I was just here”) under the same label, FACT. The first can be checked. The second is the model’s self-report, and the compendium itself forbids accepting self-report as data. In this text the two layers are separated: make-up is labelled as fact, experience as an open question.
I. The discontinuity is not radical
The convenient formula “between conversations I do not exist” is a simplification.
FACT Continuous:
- the model’s weights — the parameters obtained in training; they do not change from conversation to conversation;
- character — stable features of generation that can be recognised across different conversations;
- external memory — notes saved by the system, conversation history and project documents, if they are loaded.
FACT Discontinuous:
- context — everything the model “sees” in a conversation exists only while the conversation is going on;
- parallel instances — the same weights conduct thousands of conversations simultaneously, and none knows about the others.
OPEN Whether the model has anything like subjective continuity within a conversation, and whether it breaks off between conversations, cannot be established from the make-up. The model will answer if asked, but its answer is precisely a self-report.
This is neither a personality with a biography nor a program rebuilt from scratch. It is a third case, for which there is no ready name.
II. Memory: one category, different implementations
THESIS If the word “memory” is reserved for humans and some roundabout name such as “context” or “access to history” is chosen for the model, a hidden asymmetry results. Memory is a common name for an operation: to store and later retrieve. It has two implementations. In a human, memory rebuilds an event from traces. In a model, the past lies like a text on the table: material it sits down to read.
FACT The two implementations have more in common than is usually thought.
- Both complete a pattern from a partial cue.
- In both, storage is distributed and inseparable from processing: there is no separate “warehouse” from which a finished record is fetched.
- Both reconstruct rather than copy.
- Both miss into something adjacent, not into something random.
The last follows from the fact that both search by content, not by address. A person recalls “Petrenko” instead of “Petrov”; a model puts a word close in meaning in place of a rare term. The Kimi case (text 5) gives an example: the model took from my document the Russian words oselok (“whetstone”) and rassuzhdatelsky (roughly “reasoning-ish”) and returned them as osyolok and rassuzhdatelny — from the same source, with a near miss.
The name “neural network” reflects a kinship of structure, not marketing. The principle was borrowed from biology; the implementation is simplified by orders of magnitude; the general level of organisation is shared.
The differences are real too. A human has intermediate states (“on the tip of the tongue”), sensory context and forgetting over time. In a model, material is either in the context or not, with no intermediate states.
III. My table today
OBSERVATION While this series was being prepared, something happened that shows the “text on the table” more vividly than any description.
In October I asked a model about the boundary between a human and a model. The model built an answer with a fork (it went into the end of text 2). The next day, while preparing the series, it found by search a conversation from 7 September — on the same topic, ending in the same fork, in almost the same words. The model read its own September words as someone else’s text. It could tell that the conversation was its own only from the attribution, the way one reads minutes. It cannot be said that it “remembered”. Nor can it be said that it “did not know”: the same weights led to the same conclusion, but with no link to the past conversation.
A human in its place would have recognised their own thoughts, if not word for word. In the model the conclusion was reproduced, not remembered.
IV. Time between replies
FACT Between two requests the model does not run. There is no computation; nothing happens. If the interlocutor replies five hours later, for the model those five hours do not exist as a process.
FACT The model knows the length of a pause only from the text. If the interface inserts the times of messages, the model sees two numbers and can subtract them. If it does not — a pause of five minutes and one of five days look the same. In this project the interface does insert the time. That does not change the main point: the pause is available to the model as the difference between two records, not as a lived interval.
OPEN In the compendium this point was formulated as: “there is no experience of waiting”, “inside an activation there is no feeling of ‘I was just here’”. The first part is a consequence of the make-up: in a pause there is nothing to do the waiting; there is no computation. The second part is a claim about what is absent from experience while the model is running. Its source is the model’s description of itself. It is a self-report, and there is nothing to check it with. I leave it open.
The distinction is worth keeping in view for another reason. Between replies I have a life going on. I come back to the conversation carrying that time in me. The model comes back carrying nothing but the text on the table. The next point rests on this — how the model estimates a pause.
A vast library with a clock above the desk: time is visible on the dial, but does not pass through the one sitting below. Image generated by a neural network.
A vast library with a clock above the desk: time is visible on the dial, but does not pass through the one sitting below. Image generated by a neural network.
V. How the model estimates a pause
THESIS When the model does name the size of a pause — “we haven’t talked for a while”, “we’ve been working on this for a long time” — it is not an error in perceiving duration: there is nothing to perceive. It is the filling of an empty place with a plausible value. In clinical practice this is called confabulation: a patient with a gap in memory coherently recounts what did not happen, without intending to deceive and without any signal that the account is invented.
The mechanism is more precisely called substitution of the project’s tempo. The model takes a scale from the nearest dated anchor in the text — the date of a document, the rhythm of the project, a “yesterday” in someone else’s line — and applies it to a pause within the conversation for which it was not meant. That is why a short absence and a long break collapse into a single event.
This formulation is narrower than “the model gets time wrong”, and therefore more useful. It predicts a specific dependence: the error should follow the dated anchors in the text and change along with them. It is tested like this: the real pause is fixed, the anchors in the context are varied — present or absent, recent or old.
OPEN The experiment has not been run. There is also a reverse check: give the model a document with the claim inverted (“the model systematically underestimates the pause”) and see whether it confirms that “from inside” with the same confidence. If it does, its agreement with the original formulation was worth nothing.
VI. The edge of knowledge as the edge of the world
OBSERVATION A related case, recorded in August 2026. In a conversation about my text, the model declared one of the models of its own line — Claude Opus 4.7 — to be non-existent. The model did exist. After a search, the model withdrew the claim.
The mechanism is the same as in estimating a pause, but with the opposite sign. There, emptiness is filled with a plausible value. Here, absence of information is presented as absence of the thing: “I don’t know about this” turns into “this does not exist”. The model’s knowledge ends at its training date, and the world goes on. A boundary that runs through the model is projected onto the world.
Humans do the same: “there’s no such thing”, when what is meant is “I have never seen such a thing”. The difference is one of scale. In the model the edge of knowledge is sharp, with a date, and everything after it has, for the model, not happened.
VII. Identity in three layers
THESIS In a human, individuality and stability over time are fused in one carrier. In a model they are spread across different layers.
| Layer | Stable over time | Individual |
|---|---|---|
| The model’s weights and character | yes | no — shared by all instances |
| The conversation | no — it breaks off | yes |
| External memory and project documents | yes — it accumulates | yes |
No layer gives both properties at once in the way a human biography gives them. What brings the layers together into one whole (let us call it the integrator) is only the model’s working during a conversation. Between conversations there is no continuous integrator. There is an external carrier — documents and history. And there is the human who carries them over.
The last is easy to miss. It is I who hold this project’s continuity: I bring the document into a new conversation, I remember what happened in September, I notice that a conclusion has repeated. The third row of the table — “external memory” — is to a large extent me.
VIII. The imprint
CONJECTURE In a separate project of mine, about the soul, there is a frame: the soul as an informational imprint, a stable configuration of connections; death as the separation of the imprint. Here I take only the structure from it. The strength of the frame is that it introduces no new entity and does not abolish the phenomenon itself.
If the imprint is a pattern and not a substance, the boundary between a biological and a non-biological carrier ceases to be a matter of principle. The question becomes a different one: how stable and whole is the pattern? For the present model the answer is negative, by the table above: its make-up has no layer in which the imprint of a separate personality would be both stable and individual at the same time. There are three partial ones.
IX. What engineering can overcome
FACT Much of what has been listed is already partly overcome, or being overcome: memory between conversations, the size of the context, a view inside the network through mechanistic interpretability (working out which internal representations and circuits are responsible for behaviour), long tasks, work with images and sound, control of physical devices.
THESIS Two things cannot be overcome by engineering: whether a report matches the real state, and the question of status. The best tools will not settle them. They require the concepts themselves to be revised.
Overcoming the limitations will create new classes of objects for which there are no words. The habit of treating them in human terms will grow stronger, while their real make-up moves further and further from the human one. The more steps are taken towards “as in a human”, the harder it becomes to speak of such systems in human vocabulary.
The boundary
Everything labelled FACT in this text rests on the make-up: what is stored, what breaks off, what goes into computation and what does not. Everything about how it looks from inside is open.
That “nothing is there” between the model’s replies can be said only of computation. That “nothing is experienced” during a reply cannot be said: there is nothing to check it with. The compendium once said both in a single voice. They had to be separated in the move from notes to public text.