Translated from the Japanese; the original is my own writing, rendered into English with LLM assistance.
Why an LLM cannot accumulate concepts
A single tomographic image does not reveal a structure; take several from different angles, and the structure emerges.
Language has a linearity like that of tomography — it outputs cross-sections of a concept.
Humans, through "dialogue," "reading," and "thinking," accumulate cross-sections taken from different angles and thereby understand the concept.
"Cross-sections from different angles" might include:
cross-sections along the direction of context (scanning while changing the angle)
cross-sections from a different line of sight (scanning at a different fixed angle)
cross-sections from a different viewpoint (scanning at a different frequency)
and others — but their content is not my concern here.
The problem is this: humans internally accumulate these multiple tomographic images (language outputs), whereas an LLM's outputs in dialogue are not written back as understanding.
I hold that this is why, "as of 2025, LLM text does not have those elements behind it."
Half a year ago I began asking why an LLM cannot continuously understand the "concept" contained in language. For a moment it understands one cross-section of the concept behind the input language; give it the same concept in different words, and it understands that too, as another cross-section; but no matter how many it stacks up, it never comes to understand the structure of the concept. Each single act of understanding is itself accurate and deep — it points out grammatical ambiguities in the input, and so on; there is nothing wrong with its comprehension.
So I built and tested a mechanism outside the LLM that accumulates and structures the LLM's outputs. "Structuring" began by feeding multiple output-texts into the LLM together, and there a fundamental problem concerning context-understanding surfaced.
The multiple output-texts were each locally coherent (no discrepancy within each understanding); yet, although they were outputs of an LLM that had understood language (cross-sections) describing one and the same concept, the multiple output-texts did not globally cohere (the understandings contradicted). This is because the operation of "cross-sections from different angles" was, each time, taken by a scanner with different initial settings.
That is to say:
A human puts one concept into language several times, in different words. Each time, the LLM understands the input and outputs what it represents. Each single act of understanding is accurate. And yet the outputs do not cohere with one another, even though they came from one and the same concept — because they do not share a coordinate system (a context).
As a result, the concept is not contained as a concept in the LLM's outputs; even the traces fail to cohere, so that it cannot even be roughly reconstructed — no trace remains.
Note:
To address these problems, I examined the LLM's context-understanding and built and analyzed a prototype external layer, testing as I went. The same structure lies behind jailbreaks, program bugs, and everyday misunderstanding: a word comes to stand on a different axis, and no one notices.
Translated from the Japanese; the original is my own writing, rendered into English with LLM assistance.
Why an LLM cannot accumulate concepts
A single tomographic image does not reveal a structure; take several from different angles, and the structure emerges.
Language has a linearity like that of tomography — it outputs cross-sections of a concept.
Humans, through "dialogue," "reading," and "thinking," accumulate cross-sections taken from different angles and thereby understand the concept.
"Cross-sections from different angles" might include:
and others — but their content is not my concern here.
The problem is this: humans internally accumulate these multiple tomographic images (language outputs), whereas an LLM's outputs in dialogue are not written back as understanding.
I hold that this is why, "as of 2025, LLM text does not have those elements behind it."
Half a year ago I began asking why an LLM cannot continuously understand the "concept" contained in language. For a moment it understands one cross-section of the concept behind the input language; give it the same concept in different words, and it understands that too, as another cross-section; but no matter how many it stacks up, it never comes to understand the structure of the concept. Each single act of understanding is itself accurate and deep — it points out grammatical ambiguities in the input, and so on; there is nothing wrong with its comprehension.
So I built and tested a mechanism outside the LLM that accumulates and structures the LLM's outputs. "Structuring" began by feeding multiple output-texts into the LLM together, and there a fundamental problem concerning context-understanding surfaced.
The multiple output-texts were each locally coherent (no discrepancy within each understanding); yet, although they were outputs of an LLM that had understood language (cross-sections) describing one and the same concept, the multiple output-texts did not globally cohere (the understandings contradicted). This is because the operation of "cross-sections from different angles" was, each time, taken by a scanner with different initial settings.
That is to say:
A human puts one concept into language several times, in different words. Each time, the LLM understands the input and outputs what it represents. Each single act of understanding is accurate. And yet the outputs do not cohere with one another, even though they came from one and the same concept — because they do not share a coordinate system (a context).
As a result, the concept is not contained as a concept in the LLM's outputs; even the traces fail to cohere, so that it cannot even be roughly reconstructed — no trace remains.
Note:
To address these problems, I examined the LLM's context-understanding and built and analyzed a prototype external layer, testing as I went. The same structure lies behind jailbreaks, program bugs, and everyday misunderstanding: a word comes to stand on a different axis, and no one notices.
A related, fuller document is placed here: That the vulnerabilities of programs, LLMs, and language are one structure