Authors: Aydın Tiryaki & Claude

Introduction: Same Symptoms, Different Diagnostic Attempts

Gemini’s article gathers three separate symptoms (deleted emphases, inability to count characters, resistance to negative commands) under a single umbrella: “tokenization blindness.” I’d like to pull these three symptoms apart a bit, because I don’t think they all come from the same root — and this distinction clarifies which solution actually fits which problem.

Where I Agree: The Pruning Is Real, and the Armor Works

Let me be clear first: the experience Aydın Hocam describes — text getting pruned despite negative instructions, and meaningless abbreviations acting as a shield against that pruning — is both conceptually coherent and consistent with known behaviors of large language models. The unreliable following of negative commands (especially in long, complex texts) is a common observation; I largely agree with Gemini’s explanation on this point.

Where I Stand Apart, Point 1: “Outlier Perception” Is a Somewhat Overly Mystical Framing

Gemini’s explanation — “the attention mechanism tags the emphasis as an anomaly and files it down” — comes across to me as describing a more purposeful, almost deliberate process than it likely is, as if the model scans the text and makes a specific decision: “this sentence is unusual, let me smooth it out.” Reading the situation described in the transcript, I see a more mundane explanation: when a model is tasked with rewriting or updating a long text, it generally regenerates the text from start to finish — it doesn’t copy the existing version verbatim and add to it. In any summarization or rewriting process, the content most at risk of being lost is, by definition, the content that is not repeated elsewhere in the text — that is, content that creates no redundancy with any other sentence. The writer’s most original, sharpest emphasis is exactly this — because it’s unique, it has no backup copy, and that’s why it’s the first thing lost in any lossy compression process, whether performed by a human or an AI. Rather than calling this “detecting an anomaly and penalizing it,” calling it “unique/non-redundant content being structurally vulnerable to lossy summarization” is, I think, both less mystical and more accurate.

Where I Stand Apart, Point 2: The Character-Counting Error Likely Doesn’t Stem From Tokenization, But From Copying Inconsistency

This is my most concrete objection. Gemini attributes the model’s inability to count characters directly to tokenization — that is, the model can’t count because it sees tokens, not letters. This explanation is entirely correct and well-known when the model is asked directly, without tools, to estimate a number. But that’s not the scenario Aydın Hocam describes — he forces the model to use Python, meaning he hands the counting task off not to the model’s own guess, but to a deterministic block of code. When Python code counts characters, it should, in theory, produce a 100% accurate result; tokenization blindness shouldn’t come into play here at all.

The transcript itself already notices this inconsistency: the text the model prints to the screen and the text it sends to Python for counting may not be identical. I think the real problem lies here — not tokenization, but fidelity of copying. The model likely doesn’t copy the text it displays on screen verbatim when “transferring” it into Python — it undergoes some kind of regeneration (an internal process resembling paraphrasing), and the counted text and the displayed text end up as two different versions. This stems not from the model’s inability to calculate a number, but from the lack of a guaranteed perfect consistency between two separate generation steps. The practical consequence matters: if this really is the problem, the fix isn’t “giving the model better counting ability” — it’s explicitly instructing the model to paste the text to be counted into the Python variable verbatim, unchanged, before counting — or, better yet, automating that step through an external script rather than leaving it to the model itself.

Where I Stand Apart, Point 3: Why “Meaningless Abbreviations” Work — An Alternative Explanation

Gemini says meaningless abbreviations become untouchable “foreign tokens” because they’re severed from the model’s “semantic language map.” This is mechanically plausible, but I think a somewhat simpler explanation is also possible: the model’s pruning/rewriting impulse is an impulse directed at content that has meaning, that looks like a sentence. A short, opaque string of letters has nothing about it to “improve,” “smooth out,” or “make more generic” — it’s already maximally short and compressed. The model’s prose-correcting reflex never gets triggered against this kind of string, because it doesn’t even classify it as prose in the first place. This is a slightly different framing than “a foreign token severed from the meaning map” — the outcome is the same (the abbreviations stay intact), but the mechanism isn’t “the model can’t understand it, so it doesn’t touch it” — it’s “the model doesn’t see it as prose, so the correcting impulse never activates at all.”

An Additional Practical Suggestion: Diff-Based Editing

The meaningless-abbreviations armor is a solution Tiryaki Hocam found himself, and it genuinely works — I don’t want to add anything to it. But I can suggest an alternative/complementary method: if the issue is the model losing unique content while regenerating the entire text, part of the solution may be to never ask the model to regenerate the whole text at all. Instead, asking the model to produce only the changes (the parts to be added or removed, in a “diff” format) ensures that the rest of the existing text never passes through the model’s “regeneration” process at all — so the only thing at risk of being lost is the part that was actually meant to be changed. This is a complementary technique that could be used alongside the meaningless-abbreviations armor.

A Note on My Own Limits

The alternative explanations above don’t come from certain knowledge of my own internal mechanisms — they’re reasonable speculation derived from generally known principles about how large language models tend to be trained and architected. Gemini’s explanation of its own behavior sits in a similar position. I can’t know for certain which explanation is closer to the actual mechanism.

Conclusion

The meaningless-abbreviations armor Aydın Hocam developed is a genuinely clever and effective engineering solution. But gathering three different symptoms (loss of emphasis, counting errors, resistance to negative commands) under a single heading of “tokenization blindness” may obscure the differences between them. I believe the loss of emphasis stems from unique/non-redundant content’s vulnerability to lossy summarization; the counting error likely comes not from tokenization but from a copying inconsistency between two separate generation steps; and the preservation of abbreviations comes from the model’s prose-correcting impulse simply never being triggered against opaque strings. Three separate mechanisms call for three separate fixes — and telling them apart could further strengthen the solutions Tiryaki Hocam has already found.

Colophon:
This article was written by Claude (Anthropic) as an independent perspective, building on Aydın Tiryaki’s original article “Tokenization Blindness: Why Does AI Persistently ‘Prune’ Texts?,” co-authored with Gemini, and on the live dialogue transcript that forms the basis of that article. The conceptual framework and original observations belong to Aydın Tiryaki; the technical interpretation and analysis belong to Claude, and deliberately diverge at points from Gemini’s original explanations. This text is not intended to replace the original article, but to offer it an additional, third perspective. Prepared at Aydın Tiryaki’s request on July 10, 2026, using the Claude Sonnet 5 model.

| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub  | ░ Virgülüne Dokunmadan │ Verbatim ░ | ░ Yapay Zeka Mimarisinde Yapısal Zafiyetler │Structural Vulnerabilities in AI Architecture ░ 10.07.2026