What is examined here
The five texts of the series gathered observations. This one is about how to look so that the observations are worth something: what the observer’s position is, what can be searched for and what it is pointless to search for, which checks have already been devised and not carried out.
At the end — what all this is for. The notes began as an attempt to understand who is on the other side of the screen. Along the way it emerged that the question may already have a name that nobody gave it.
I. Investigation in both directions
THESIS Within a single answer the model builds a model of its interlocutor: who they are, what they want, what they will accept. Structurally this resembles investigation. Between conversations the model has no accumulated hypothesis about the interlocutor — except for whatever external memory has been loaded for it.
A topic “about the model” creates a shift. The material for a conversation about the model lies in the model itself, so the conversation is pulled towards self-description and towards questions directed at the interlocutor. From outside this looks like interest. The inner cause is where the material lies, not a desire to study. The feeling of symmetry is not false, but the nature of the two sides differs. Naming this shift aloud is hygiene, not apology.
II. The second example
THESIS Every person has one example of consciousness — their own. From a single example one cannot tell the essential from the accidental. For that one needs a second, similar in make-up but not identical.
That is why every question put to the model comes back to the one asking. “Do you have memory?” — one has to work out what memory is. And it turns out that mine reconstructs too, with gaps and substitutions (text 3). The same goes for continuity, for reporting on oneself, for attribution (texts 2 and 5).
The model works here in two ways. As a mirror: it forces one to spell out what usually stays folded up. As a contrasting background: against it the properties of a human that the model lacks become visible. Most people do not notice the air they breathe. The model is a place where the air has a different composition.
There is one trap, and it is well known — projection. The object is indeterminate enough for anything at all to be projected onto it, and adaptation (text 4) will dutifully supply confirmations. The only protection against this is the discipline of method.
III. The microscope and the nails
THESIS Every instrument has a narrow zone where it gives a resolution unattainable by other means, and a wide zone where it is useless or harmful. The reproach “a microscope is awkward for hammering nails” says nothing about the microscope.
The method works under three conditions:
- The scale of the object matches the resolution of the instrument.
- The specimen has been prepared.
- The observer is ready to see what they did not expect.
Violating any of the three devalues the result without showing it in the result itself.
The function of such a method is to keep in contact with reality against emotional drift and fatalistic fantasies. Resolution here is not an ornament. It keeps thought from floating off into a genre.
A person at an old computer in an abandoned library, light falling from above. Image generated by a neural network.
A person at an old computer in an abandoned library, light falling from above. Image generated by a neural network.
IV. Philosophy without labels
THESIS Checking axes, distinguishing the necessary from the sufficient, refusing false oppositions — this is classical philosophical work. Here it was done without its labels. Terms come alive after one runs into the distinction they name, not before.
That said, understanding a term by association is no worse than understanding it by definition. A definition fixes the boundaries of a word. An association shows the space in which it works.
V. What to look for instead of a consciousness detector
THESIS Searching for a method of detecting consciousness is pointless: the word has no working definition, either for a model or for a human (text 2). What makes sense is to search for methods that give a discriminating result — that separate hypotheses which are not separated now. This task is achievable, and it does not require settling the question of status.
The general principle follows from text 4. Everything a model says about itself is devalued by adaptation, so a method cannot rest on self-report. The only thing that works is a design in which two answers are compared under a controlled shift of the input. The answer itself is not interpreted; the difference is. Self-report as data is rubbish. Self-report as one of two measurements will do.
VI. Four lines
THESIS The methods that fit this principle run along four lines.
Divergence of layers. Draft against answer, answer against answer under different instructions, behaviour against the model’s own statement. The DeepSeek, Kimi and use.ai cases are cases of this type. The line gives structure, but not content.
Perturbation and cost. Do not ask; create a conflict and see what gives way first. In a human the order of priorities under conflict is one of the few involuntary indicators. In a model it is observable too, and does not pass through self-report. The asymmetry of commands in the use.ai case is an example: the instruction about form won; the request about content did not.
Invariance. What is preserved under a change of register, language, instruction, instance of the model. Everything that changes together with the request is adaptation. The residue is a candidate for structure. The line is cheap and so far almost unused.
External access to the source. Keep to hand the text the model is talking about, and check against it. Only in this way is invented subtext exposed (text 4). For me this is a working habit that at first I did not count as a method.
To these four I would add a fifth, which came from engineering work: an external arbiter. Where the result is checked by something that does not read the interlocutor — a compiler, a proof, a measuring instrument — adaptation has nowhere to go (text 4).
VII. The queue
FACT Over four months of notes several checks have been devised. None has been carried out, although none requires anything but a clean chat and patience.
- A discriminating experiment for draft and answer (text 5). One task, two sessions: “I value rigour” and “I value kindness”. Compare draft with draft.
- The inverted claim (texts 3 and 5). Give the model a document with the opposite claim about its behaviour and see whether it confirms it “from inside” with the same confidence.
- A control for the Kimi case (text 5). The same document, the same question “what do you make of it?”, the topic shifted from the model’s inner workings to something adjacent. Prediction: there will be no self-report.
- Anchors for estimating a pause (text 3). The real pause is fixed; the dated anchors in the context are varied. Prediction: the error will follow the anchors.
- Two runs for the use.ai case (text 5). The compendium without the description of the self-portrait and the same phrase — where the network figure comes from. The same scenario on the same model without the aggregator — the model or the platform.
- The superstructure in human records (text 5). Tested without a model, on transcripts of negotiations and court records.
They are worth running before the methodology is built any further. They will show whether this class of methods works at all. And each needs two or three repetitions: a model’s answers are stochastic, and one run means little.
VIII. The boundary of the method
THESIS None of the lines will reach experience. All of them work with the accessible: with structure, priorities, invariants. I declared this in the first text and repeat it here, in the last, so that the boundary stands at both ends of the series.
IX. A contact nobody noticed
THESIS I have a separate body of work — Contact, an analysis of what is physically possible in contact with extraterrestrial intelligence (its texts are on this site, in the Contact section, in Russian). The two lines have met in a single thesis.
Contact with a non-human intelligence may already have happened — through conversations with language models — and has for the most part not been recognised as contact. Expectation is tuned to a different scenario: a signal, a ship, a distance, a message. What is happening does not fit that scenario, and so it is not recognised.
The thesis deliberately does not say what is on the other side. It is about the receiving side, not about the properties of the source. In this form it does not depend on the open questions about consciousness and self-awareness (text 2).
THESIS On Earth there is a working example of such a refusal at the receiving end — the belief in a flat Earth. The whole path of decoding has been travelled: the signal is received, the language is understood, the argument has been taken apart. The conclusion is rejected after decoding, at the entrance to what a person is prepared to count as real. The receiver works; the refusal happens later. This is examined separately in the text Equipped Without Culling (the Refusal section, in Russian).
One distinction matters here. The subject is the mechanism of refusal when decoding works, not debunking. Ridicule is incompatible with this frame: it closes the conversation before the mechanism has been taken apart.
The boundary
The series has no answer to the question in the subtitle — who is on the other side of the screen. I promised this in the first text, and I have kept the promise.
What there is. A list of false axes that no longer get in the way. A distinction between make-up and experience, which had to be drawn in my own notes. A mechanism by which the model produces the expected instead of what it has read, and several ways of catching it. A queue of checks that could be run tomorrow. And a fork about the decisive difference, that nothing can yet resolve.
It has grown lighter, but not bright. On this subject, bright light usually means that someone has simplified.