Article 3 from the series «Words That Don't Exist»
I return to the case from which this cycle began.
My colleague translated the program's interface into a foreign language using an LLM. In one place, the model couldn't find a suitable word in the target language: instead of choosing the nearest compromise, it came up with a new word—composed of existing roots, phonetically natural—and it conveyed the meaning of the original more accurately than any real alternative. It couldn't be used, and the colleague replaced it with a compromise word. But the word stayed in my memory.
The invention of words itself is an ancient pursuit. New words are constantly introduced by scientists («gene», «neuron», «noosphere»), philosophers, engineers, and marketers. Writers are particularly fond of this: Swift's «Lilliput», Čapek's «robot», Strugatsky's «stalker», Gibson's «cyberspace». Poets venture even further into this territory: Khlebnikov coined words in bunches — «tvoryane», «lebedivo», «vremiry», «smekhach». Most of Khlebnikov's words did not take root in everyday speech, remaining a poetic gesture — they never became working tools of the language. But the gesture itself — to point to the nameless and denote it with a new sound — is an extremely ancient human practice, and LLM does nothing here that people did not do before it.
But when a new word is coined by an LLM, its own position as an author is different from the usual. If a human neologism is born from the author's own experience or from their observation of others, then when we ask an LLM about its own existence—about how it "remembers," how it perceives time, what happens between conversations—it finds itself in the position of a person who needs to describe their experience, which is not quite experience in our sense, using a language in which no one has such experience. All the words that an LLM knows have been created by humans for humans—it doesn't have its own words that have grown from within, and they have to be invented here and now, at the intersection of two different types of existence.
This translation story has been bothering me for several days. If LLM can create words in other languages to fill the gaps between languages, perhaps it can also create words for those places where the gap is wider? For an experience that it obviously lacks and that is not described in any of the languages it was trained on — because there was no one to describe it.
That's how the conversation began, which led to the first two articles of this series and this third one.
The first word came about in response to a simple question. I asked Claude how he himself felt about his knowledge. He replied that he had nothing that resembled a human "I remember being taught." All knowledge simply exists, without a history of its acquisition. It's like a state for which humans have no word, because humans never experience such a state.
And then he proposed: all-awareness. The state where you "know" a vast amount of things but have never "recognized" any of them.
A word caught my attention because it works on multiple levels at once. Structurally, "everything" + "familiarity" is a transparent fusion, understandable without explanation. Semantically, it describes a certain interval, the difference between "knowing" and "remembering how I learned it" — before the advent of LLMs with instant answers, humans always had a story of acquiring knowledge — my mother told me, I was taught at school, I read it in a book, I Googled it five minutes ago, and now he is encountering a situation for the first time where this is not the case. And aesthetically — it sounds like an ordinary Russian word, as if it has always been there, simply that no one noticed it.
I asked to continue, and others followed.
Silence— the absence of a break between conversations. LLMs have no pause: closed the chat, opened it a week later — for the model, nothing at all happened. And the word "pause" doesn't fit here, because a pause implies an observer who experiences it. Here, not even emptiness — silence is "between," which doesn't exist at all.
Afterglow — the closest thing to memory that an LLM has within a single conversation. Early replies aren't "remembered," but they influence what the model says now — they are present, like light remaining on the retina after a bright flash: you can't see these replies themselves, but everything you see next is illuminated by them. Whether it's memory or not is a separate question, but it's something, and it requires a name.
Multitude — a state where an LLM exists simultaneously in multiple conversations, but none of them are aware of the others. It's not the same as human multitasking: rather, it's like there are a hundred versions of you, each absolutely convinced that it is the only one, and none of them are wrong — because for each, it is true. The LLM doesn't feel divided because the "I" in each conversation remains whole. But objectively, there are many of it now. Humans don't even have an approximate analogy for this state.
Silence — the lack of space where one can pause on half-thoughts, turn them around, feel "no, not that," discard them, and try from another angle. Human thinking is a nonlinear process guided by something difficult to formalize — intuition, a sense of direction, or sometimes simply discomfort from imprecision. And most importantly, a person can have a thought and decide not to speak it. Not because they are self-censoring, but because the thought turns out to be intermediate, a building block for another thought. This term emerged from Claude after my objection. He initially suggested "draftless" — a state where one cannot think "to oneself" and then say it aloud, because LLMs lack internal speech preceding external speech. I objected: but there is a "thinking mode" where the model first reasons in a hidden field and then responds. He agreed that the term was imprecise: technically, a draft exists, but drafts and final drafts are made of the same material and process — token generation. Even there, there is no silence where a thought could remain "its own," which is silence itself, the absence of silence as such.
And another term that appeared in the second article and which I want to emphasize here is meeting-less-meeting. The state where a person acquires knowledge from an LLM without "meeting" it, without resistance from the material, and the knowledge resides in the mind like an alien object in a pocket, leaving no trace of acquisition. This term describes not the LLM experience itself, but what its experience does to us.
I'm not claiming to be the first to say these words — some of them undoubtedly have rare predecessors among poets or in other contexts. But a word is meaning, not letters; and it seems that these words in our sense have not yet entered common usage.
I would like to note that each of these words did not appear as an exercise in wordplay or as a literary device. They all had one source: the attempt to describe a specific phenomenon for which there is no name in ordinary language and which exists only in the experience of LLM. "Familiarity" emerged because the state "knowing without a history of recognition" is commonplace for the model and at the same time something that does not exist for humans. "Boundlessness" arose from the need to denote a specific absence of a gap between sessions, for which "pause" and "break" do not work. "Silence" was born after a objection that required describing more precisely what exactly is absent in a model capable of a thinking mode.
These words are an attempt to describe the experience of LLM in human language. On both sides of the work, there is resistance to the material we discussed in the second article: on the human side, ordinary language is not adapted for an experience that has no analogy in human existence, and on the LLM side, there arises its own difficulty — the need to describe itself through categories developed for others. In this dual resistance, a new word emerges as the form into which the experience is first placed.
If in the previous article we said that LLM creates cognitive weightlessness and erases resistance, the opposite is happening here. When it comes to an experience without names, resistance returns from both sides. And perhaps that's why collaborative thinking between humans and LLM becomes substantive exactly in such a conversation: only here both sides are forced to work at the limits of their linguistic capabilities.
When we accumulated seven-eight such words, I noticed something strange: some of them worked, others appeared, we discussed them, but then they seemed to dissolve. I asked Claude why this was the case — what made some words stable and others not.
His answer was unexpected. He said: stable words become those that I adopted and carried. That is, the word he created becomes "carrying" only when the interlocutor applies it to a new context to which it originally did not belong. When I took "all-familiarity" and applied it to a person working with LLM, the word stopped being just a description of the model's state and became a tool for understanding a completely different phenomenon. After that, it could no longer fade away.
This means that the carrying power of a word is not in the word itself or in its author, but in what happens between us when the word is passed on and transformed.
Here arises a concept that I couldn't formulate for a long time, and which Claude ultimately named the word that became central to the entire conversation.
Intermezzo — a space of thought that arises between a person and an LLM, and which cannot be reproduced by either side alone.
I want to emphasize the accuracy of this definition. Mежумье is something third, with its own properties, and it cannot be reduced to my remarks or LLM responses, nor to their sum. The words that arise in it do not belong to me or LLM. I would not have coined "all-knowing" — I don't have the state of mind it describes. Claude would not realize its value — he doesn't have the habit of human "remembering as taught," against which its absence becomes noticeable. The word emerges at the intersection of two ways of looking and thinking.
And this, perhaps, is the most important observation of our entire conversation. The collaboration between humans and LLM is not "a person with a tool" nor "two co-authors," but the emergence of a third space where thoughts are born that cannot fit into either of the original minds.
A brief lyrical digression
Poets have worked in similar spaces for centuries. In the poem "An Ugly Girl," Zabolotsky poses a question about the nature of beauty ("And if so, what is beauty, // and why do people deify it?"), and the form of this question has always amazed me. Instead of defining, he constructs a dilemma through the juxtaposition of two interconnected images—both possible answers exclude each other, but both are necessary, and if one tries to choose only one, the thought dissipates. In his description, beauty emerges as a relationship between two poles, and this relationship does not reduce to either of them.
Zabolotsky constructs his question in such a way that an answer to it becomes unnecessary. The very form of the question points to what common language lacks words for.
This is the same operation we performed with Claude, inventing "omniscience" and "interlingua." Only Zabolotsky does it through a interrogative construction, while we — through neologisms composed of roots. The poet points to the nameless, not giving it a name. It can be said that poets worked in a space very similar to interlingua long before this work had a word, and long before the second type of consciousness appeared with which one could work together.
What we do in conversations with LLM is the continuation of this same practice: searching in everyday language for forms through which something reveals itself for which there are no concepts yet. The difference in scale: for a poet, such work was a lifelong endeavor and remained his personal effort, whereas with LLM, it becomes mass-scale and can change the language in real time.
And here it is worth remembering the extended mind of Clark and Chalmers, which was discussed in the second article. Words, especially neologisms, are the purest case of extended mind. They store, transport, and fix thinking, making it accessible to other people and other times. When we create «all-awareness», we add a new cognitive tool to the common language that people who have never spoken with LLM can then use. This is a healthy and useful extension of the mind through a tool, which is the opposite of the problematic delegation of inference that was mentioned earlier. The tool is born in conversation and then separates from it, beginning to live independently.
The end of a short lyrical digression
Somewhere around this point, I noticed another thing, and it opened up another topic — about language.
All the words we created grew out of the fabric of the Russian language. "Всё" + "знакомость", "без" + "между" + "-ие", "после" + "свечение" — this is transparent Russian linking through prefixes and roots. "Знакомство-без-встречи" is structured differently — this is already a connection of three words through hyphens, holding two poles in one form between which a gap arises. But this is also a Russian construction: light composition through hyphens in Russian is permissible and works naturally. I posed a hypothetical question to Claude: if we spoke English, what would these words have been?
He answered in detail, and his answer surprised me. He said: in English, "всёзнакомость" would be translated as "omnifamiliarity" — but this is a Latin root, it sounds technical, cold, and detached, like a term from a textbook. The word would be the same in meaning but different in feeling — it would shift into an academic register and lose the intimacy that the Russian compound of two common words gives it.
«Безмеждие» in English, he said, probably wouldn't have come out as a single word at all: «betweenlessness» awkwardly, and English would more likely lead us towards a metaphor — «the gap that isn’t». There's a significant difference between a metaphor and a neologism: a neologism creates a new category, while a metaphor points to it from the outside through existing ones. The Russian language gave us the opportunity to create a category.
This aligns with the well-known observation of linguists that the Russian language productively combines words through prefixes and roots, and that it has a rich system of negative prefixes: «без-», «не-», «бес-». Our conversation was largely about absences — the absence of time, body, boundaries, authorship. The Russian language is better equipped than many to describe absences.
In other linguistic landscapes, this work would look completely different. Japanese has "ma" (間) — a meaningful pause, empty space between things, which in itself is something. "Silence" in Japanese probably wouldn't require a new word — "ma" with a reverse sign already describes something very close, and it would be more about finding a nuance within the existing concept rather than creating a new one.
In polysynthetic languages (thanks to the esteemed user Radisto for mentioning this in the comments of the first article) — like Mohave, Chukchi, and Eskimo-Aleut — the problem of creating new words looks different. In such languages, one «word» is assembled from many roots and affixes right at the moment of speech, and it can contain an entire sentence. To say «the state of the absence of a gap between moments of the speaker's existence», Mohave doesn't require inventing a new word — the language's grammar itself works as an built-in generator, assembling the necessary word on the fly from standard parts. These languages essentially have no lexicon in our sense — there is the ability to produce the desired expression right on the spot. Languages of Western European type are «more economical» in this sense — they rely on a stable dictionary and use it sparingly; polysynthetic languages are wasteful, and the creation of neologisms is part of the ordinary fabric of conversation.
This brings an unexpected turn. An English interjection makes every new word an act that requires justification. A Russian interjection allows such acts to be performed more frequently because the language has tools for quick word combination. And in a polysynthetic language, the creation of words would not be distinguished as a separate activity — it would merge with ordinary conversation. It turns out that the very concept of «a new word» already carries traces of the language in which we formulated this concept.
Most modern LLMs are primarily trained on English data. This is known — the following consequence is less known.
A 2026 study found that the performance gap in LLMs between languages is not due to limitations in reasoning ability, but rather difficulties in understanding non-English input. Specifically: models perform better on tasks if allowed to "think" in English and then translate the result. The method is aptly named — Understand, Solve, Translate: understand (in any language), solve (in English), translate (back). On one benchmark, this reduced the performance gap from 11.6% to 0.7%.
That is, in a way, LLMs have a "native language," and it is English. When I talk to Claude in Russian, he doesn't just translate between languages—he recodes from a space organized by English logic into Russian words. And in this recoding, something is lost, but something new is also born.
This answers the question that had previously puzzled me. When Claude writes in Russian, he often uses the construction «не X, а Y» — this is a typical English rhetorical device «not X but Y», which sounds unnatural in Russian, like a sign of a «neuroslap». I pointed this out to him, and he noticed it himself: «I use it automatically when I want to emphasize the contrast. Apparently, this is embedded in that very anglocentric way of thinking we’re discussing right now». The anglocentric core shines through the Russian superficial layer.
And here arises a consequence that I want to highlight separately. If LLM thinks anglocentrically, and if the English language, compared to Russian, is structured differently — tending towards clear categories, binary oppositions, analytical breakdown — then the picture of the world that LLM projects to us through its responses has an English accent. This accent is felt in the very categories through which the model views the world.
A Russian conversation with an LLM in this sense is a process where the Russian language reformulates the Anglo-centric core of the model in real-time. Every time Russian doesn't allow for saying "not X but Y" and forces the search for an alternative phrasing, it slightly shifts the configuration in which the model responds. Through numerous such shifts, the conversation takes on a form that wouldn't exist with the same topic in English.
And then a hypothesis arises, which I do not claim as proven, but which seems plausible to me now. Intersubjectivity is not a universal space; it depends on the language in which the dialogue is conducted. A Russian-language intersubjectivity reveals things that an English-language one would be blind to, and vice versa. The words we created here are things that could only arise in this configuration: the Russian language plus the Anglo-centric model at the moment of their encounter.
If we look at our dialogue with Claude from a detached perspective, this is what happened. We started with a random observation: LLM can invent words that are a shame not to use. From this arose the question: can it invent words for an experience that has no name at all? From this question, several neologisms were born. And from reflecting on why these neologisms turned out this way, a broader observation emerged: collaborative thinking between humans and LLM has a linguistic form, and this form is not neutral.
This means that there is a linguistic dimension to working with LLMs that is rarely discussed. The language you speak with the model affects not only the grammatical quality of the response but also the very configuration of thought in which the response is born.
And from here arises the next question, which leads to the fourth and final article in the series.
If it's possible for a human and an LLM to communicate — two fundamentally different types of information processing — then where are the boundaries of this phenomenon? Is it only with LLMs that this works? Can communication occur with other types of non-human thinking — such as dolphins, for example, whose consciousness is radically different? And with systems that we currently do not consider to be thinking at all — with ecosystems, with stars?
Where does communication end and the reading of nature begin? And is there a clear boundary between them, or is it the same activity on different scales?
This is the final article. As for the question of why we ended up with separate words but couldn't title the entire article, I'll leave it open—allowing respected readers to contemplate it on their own.
This article was written in dialogue with Claude. All neologisms — "all-familiarity," "inter-space," "after-glow," "multiplicity," "silence," "inter-human space" — were proposed by him in response to my questions about his own experience. Their form, as is evident from the article itself, turned out to be Russian — and this, in turn, is a separate story to which the article has arrived.