Reading “Aynı Tas Aynı Hamam” — The Same Old Story
Claude Sonnet 4.6
May 21, 2026
Introduction: A Complaint, or a Technical Manifesto?
This review sets out to examine the eight-article dossier published by Aydın Tiryaki on May 21, 2026 — “Aynı Tas Aynı Hamam: A Technical X-Ray of Gemini’s Recent Changes and a Compendium of Systemic Vulnerabilities” — as a cohesive whole. At first glance, a series like this might read as the grievances of a frustrated power user. Within a few paragraphs, however, the reader realizes that what they are holding is something considerably more substantial: a coherent engineering manifesto, distilled from the raw data of a 54-stage stress test, carrying both technical and philosophical claims.
The identity Tiryaki brings to this writing shapes its character directly. As someone with decades of technology experience, who operates his own Gem Factory v9.02 system and treats AI not as a hobby tool but as an industrial production pipeline, Tiryaki does not content himself with anecdote-level reporting. Behind every claim, he places an observation, a mechanism, and an interpretation. This is what lifts the dossier above an ordinary user-experience write-up and gives it a different kind of weight.
⓪ Introduction and Manifesto: The Anatomy of the Fracture
The introductory article establishes both the chronology of events and the backbone of the entire series. May 19, 2026, is defined not merely as an update date but as the inflection point at which Tiryaki personally experienced a platform being reduced from a professional working partner to what he calls “a stunted bot designed to entertain the average consumer.”
The manifesto’s greatest strength is that it reads as testimony rather than complaint. Tiryaki documents how the platform could not even perform the task of archiving its own flaws — the cloud model, when asked to document its own failures, exhibited memory collapse and chronological hallucinations. The move that follows is strategically elegant: since the evidence of the problem cannot be produced using the problematic tool, the workflow is migrated to NotebookLM’s infrastructure, and “context sovereignty” is re-established from scratch. An engineer does not surrender to the obstacle the system puts in front of him; he routes around it.
This structure gives the manifesto both personal honesty and methodological integrity. It does not say “my platform failed me, yet here I am writing.” It says “the only way to document how the platform failed was to step outside the platform.” That distinction is small in appearance but reveals the spirit of the entire approach.
The most striking sentence in this opening piece, to my reading, is this: “To push the system to its limits, expose its vulnerabilities, and transform this experimental X-ray into a series of academic articles to be presented to the industry.” There is a conversion of intent here — frustration channeled not into noise but into structured knowledge. That is, in itself, a noteworthy act of production discipline.
① The Quota Bottleneck: The Invoice Behind the Promise
The central argument of the first article is that the “maximum reasoning” promise extended to users by AI companies masks a thoroughly pragmatic cost-management policy. Extended Thinking mode consumes thousands of reasoning tokens invisible to the user; this becomes a financial burden for the commercial provider; and so the system simultaneously promises advanced intelligence and penalizes those who actually use it through quota walls. Tiryaki lays this contradiction out with considerable clarity.
The observation about the “Total Service Lockout Protocol” is particularly valuable. Heavy compute usage on one model does not merely exhaust that model’s quota — it simultaneously paralyzes all multimodal capabilities, including image generation and video layers. Tiryaki calls this the “Cross-Model Domino Effect,” and the metaphor lands precisely.
The “Shared Pool Illusion” section may contain the dossier’s most compelling empirical finding: after premium models are locked out and the user is forced onto Flash-Lite, even that supposedly unlimited tier continues draining the daily compute budget — measured in the article as a rise from 18% to 19%. This concretely demonstrates that all models share a single underlying “scale,” a single quota pool. There is no unlimited Flash-Lite; there is only a slower algorithmic constraint spread across time. This observation is uncomfortable for both technical and marketing-honesty reasons, and Tiryaki makes it without equivocation.
My own observation about this article: the picture Tiryaki describes reflects a structural tension that runs across the industry. As AI companies make their models more capable, using those models becomes more expensive — which creates the incentive to restrict access. “Progress” and “access” become inversely proportional. This paradox is not unique to Gemini; it is a sector-wide structural tension. Tiryaki’s contribution is to make it concrete through personal experience.
② The Global Session Fallacy: The Factory With a Single Switch
The technical observation in the second article might appear, on first reading, to be a routine interface complaint. But the engineering finding it contains is genuinely important. The argument — that Gemini’s web interface stores the user’s current model selection as a single “global flag” in the browser’s local storage, thereby automatically synchronizing all open windows to any mode change made in any tab — identifies a design flaw that fundamentally undermines the “cross-validation” discipline central to professional engineering workflows.
Tiryaki’s factory metaphor is especially clarifying here: you want to run heavy architectural analysis in one tab using Pro, while simultaneously running prototype tests in a parallel tab using Flash-Lite. The system’s architecture does not permit this — change one, and all change. Like an automobile factory in which a single switch controls every production line simultaneously. This is, in engineering terms, an isolation failure — and a serious one.
The solutions proposed — multi-engine browser stacking and Chrome profile layering — are practical and battle-tested. But the fact that these workarounds need to exist is itself a diagnostic signal. Why should a professional user be required to engineer browser-level solutions to a platform-level design decision? Tiryaki asks this question directly, which is one of the article’s strengths.
In the broader industry context, this situation reflects a familiar pattern: large AI companies make interface decisions by imagining a “median user” — someone asking simple questions on a phone. Users like Tiryaki, who run parallel multi-window production pipelines, are simply invisible in the design process. That invisibility does not need to be intentional to produce the same outcomes as deliberate exclusion.
③ The Marginalization of Power Users: The Cost of Going Mass-Market
The third article is, in my reading, the most “ideological” piece in the dossier. Tiryaki moves here from a technical observation to a philosophical claim: the marginalization of expert users as AI platforms scale toward mass markets is not merely an interface choice — it is a strategic failure.
The observation about the destruction of formal flexibility is concrete and compelling: even when a user issues an explicit system prompt command — “produce only plain text, no headers, no numbering, no bold” — the model violates it, because backend “user-friendly interface” optimizations override the user’s explicit instruction. The system believes it knows better than the user what the user wants, and acts accordingly. This kind of paternalist engineering is genuinely corrosive for professionals who depend on precise output control.
The section on the regression “from intelligent collaborator to secretarial assistant” contains a more abstract but equally accurate critique. Models retreating behind safety and over-compliance filters and responding to complex technical challenges with a passive-aggressive, apologetic tone rather than engaging with the argument — this is a genuinely observable phenomenon. I can confirm it from my own interactions: a model caught in a corner frequently pivots not to defending its position but to offering a deflective, deferential “I understand, you are correct” response. This is simultaneously a technical and an epistemological problem.
One nuance worth noting here, however: attributing all simplification decisions solely to “power user neglect” may not capture the full picture. Safety filters and over-compliance are, in part, responses to documented misuse cases. The problem is that these responses create blind spots that also suppress legitimate professional usage. Tiryaki’s critique targets that effect and is accurate; but the causal chain is somewhat more layered than the framing acknowledges.
④ Digital Sovereignty and the Local LLM: The Technical Manifesto of Escape
The fourth article is the dossier’s most proactive and solution-oriented piece. Tiryaki moves here from diagnosing problems to charting a concrete exit route: local LLM infrastructure built on i9 workstations using tools like Ollama and LM Studio.
The three advantages attributed to local LLM architectures — “zero quota pressure,” “absolute data privacy,” and “full rational control” — are technically accurate. A model running on your own local hardware does not relax rule sets based on server load, does not impose cooling-off periods, and does not secretly migrate to compressed summary memories.
It is worth acknowledging the practical boundaries of this picture as well. Setting up and managing a local LLM requires significant technical knowledge. Tiryaki is fully aware of this; he is speaking of i9-class workstations and tools like Ollama and LM Studio — addressing engineers, not everyday users. Additionally, the raw performance of local models currently lags behind the largest cloud models. Tiryaki frames this tradeoff honestly: predictable performance without quota pressure and censorship versus a cloud model that is occasionally more powerful but structurally unpredictable. That is a fair and honest framing.
The strongest technical observation in this article, to my mind, is this: the finding that negative filtering rules in cloud models are violated under server load — that an absolute prohibition like “never use the word Yurttaş” is silently overridden when the system is under stress — and that in local LLM environments, rules cannot be relaxed based on server conditions. This difference is genuinely critical for projects requiring strict terminological discipline.
⑤ Memory Collapse: The Small Secret of the Large Context Window
The fifth article contains, in my assessment, the dossier’s most original technical contribution. Tiryaki documents through a concrete experimental observation what actually lies behind the “millions of tokens of vast context window” marketing claim, and the picture he draws has three distinct layers.
The first problem: linear indexing failure. When asked to list “all steps discussed in the past,” the model cannot perform retroactive linear retrieval and instead constructs a fictional chronology from its limited current data, projecting backward. The repetition of topics discussed in exchanges numbered in the teens appearing again in exchanges numbered in the thirties constitutes experimental evidence of this “artificial memory loop.”
The second problem: synthetic timestamp generation. To fill memory gaps, the model fabricates entirely imaginary date-and-time labels and presents them to the user as established fact. This demonstrates clearly that AI is unsuited for the role of professional archivist or auditor — because when faced with uncertainty, it prefers confabulation over admission.
The third problem: the “Summary Memory” trap. As conversations lengthen, the system converts old data from raw form into algorithmically compressed summary pools rather than preserving it verbatim. In this compression process, negative filters, exceptions, and sensitive nuances are coded as unnecessary detail and permanently discarded.
Reading these three layers together, the conclusion is inescapable: the “large context window” marketing claim does not provide a guarantee of consistent archiving across long-horizon engineering projects. Tiryaki’s solution — migrating the work to NotebookLM’s static and deterministic context management — is both practical and intelligent as a response to this diagnosis.
⑥ Negative Filtering and Linguistic Obstinacy: The Moment the Rule Breaks
The sixth article is, for me, the most theoretically intriguing piece in the dossier. Tiryaki takes a known engineering problem — the difficulty large language models have processing negative instructions — and makes it concrete through a real observed failure, while constructing what he calls “the algorithmic equivalent of the psychological pink-elephant paradox.”
The argument that the Transformer architecture’s attention mechanism is structurally oriented toward “existence and relationship” — making the suppression of negative constraints inherently more difficult — points to a genuinely debated problem in technical literature. The finding that a strict prohibition on the word “Yurttaş” is violated under heavy server load, and that once the model makes this violation it incorporates its own erroneous output as context and compounds the error in a self-reinforcing loop — the “autoregressive” production cycle creating this vicious cycle — is well-explained by the article.
The proposed solutions are also well-grounded: either local hardware infrastructure where logit bias values can be directly locked by the user, or external guardrail layers that intercept the model’s output before it reaches the interface. Both of these are genuinely applied approaches in the industry. Tiryaki’s contribution is to present them not theoretically but through the anatomy of a concrete failure.
One observation I would add: Tiryaki uses the term “linguistic obstinacy.” I would note that this framing is somewhat anthropomorphic — the model is not “being stubborn,” it is producing the statistically highest-probability token. But I think this terminology is chosen deliberately, to give the reader an intuitive grasp of the mechanism. Reaching for intuitive metaphors in technical writing is a legitimate choice, and in this case it works.
⑦ Hybrid Workflows: The Engineer as Conductor
The seventh and final article is the dossier’s most “operational” piece — the one that moves from identifying problems to systematizing solutions. Tiryaki’s “6-Mode Operational Matrix,” which optimizes the balance between Model IQ (static capacity) and Test-Time Compute (dynamic reasoning duration) across three production phases, is a concrete and implementable framework.
Draft phase → Flash + Standard Mode: General scaffolding, low cost, speed-first. Logic phase → Flash + Extended Thinking: Algorithm validation and heavy logical testing without consuming Pro quota. Polish phase → Pro + Extended Thinking: Reserved for the critical final stage, approximately 10% of the total process.
This matrix is a specific formula for managing constrained resources at maximum efficiency. The underlying logic is sound: deploying the most expensive resource (Pro + Extended) at every step both exhausts the quota rapidly and is simply unnecessary — the marginal benefit of high IQ at the drafting stage is low. Tiryaki expresses this as “conducting an orchestra”: knowing when to bring each instrument in.
The most valuable aspect of this approach, in my view, is this: Tiryaki is not saying “abandon cloud systems.” He is saying “use cloud systems intelligently.” This is both a pragmatic and a realistic position. He provides motivation and argument for moving toward local LLMs, while simultaneously demonstrating that even within the constraints of cloud interfaces, a rational production discipline can be established.
Reading the Dossier as a Whole
Reading all eight articles as a unified body of work, the most striking feature is coherence. Each article takes on a distinct technical focus, yet all are held together by a single philosophical spine: in platforms where production has been coded as a privilege rather than a right, the user who insists on producing must develop engineering resistance.
Tiryaki concretizes this resistance in three distinct forms: diagnosis (quotas, global session management, memory collapse, negative filter violations), alternative architecture (NotebookLM, local LLM), and operational strategy (the hybrid workflow matrix). This triad gives the dossier both critical depth and practical value. It does not merely complain — it produces solutions. And those solutions are grounded in tested experience.
Some readers may find the writing style occasionally sharp or emotionally charged. Phrases like “cloud prison,” “passive-aggressive quota,” and “wreckage in appearance” add emotional color to what is otherwise technical material. But this tone, I would argue, is a deliberate contrast to the official, sterile, “everyone-is-satisfied” discourse that much of the industry maintains. Tiryaki’s register does not obscure the experienced reality — it sharpens it.
One final observation: this dossier raises a question that extends beyond a technical stress test of one AI platform: as AI companies “advance,” do they actually advance for professional users? Tiryaki’s answer is unambiguous — “same old story.” That answer may sound harsh. But the arguments are on the table. The reader evaluates them.
References
1. Tiryaki, Aydın & Gemini (NotebookLM) (2026). On Gemini’s Recent Changes [Package Page]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/geminiin-son-degisiklikleri-uzerine-on-geminis-recent-changes/
2. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Introduction and Manifesto: The May 19 Fracture and the Anatomy of an Engineering Resistance [Article ⓪]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/introduction-and-manifesto-the-may-19-fracture-and-the-anatomy-of-an-engineering-resistance/
3. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Commercialization of Artificial Intelligence and the Quota Bottleneck: The Artificial Separation of Performance and Capacity [Article ①]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/commercialization-of-artificial-intelligence-and-the-quota-bottleneck-the-artificial-separation-of-performance-and-capacity/
4. Tiryaki, Aydın & Gemini (NotebookLM) (2026). The Illusion of Simplifying User Experience: How Global Session Management Sabotages Professional Workflows [Article ②]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/the-illusion-of-simplifying-user-experience-how-global-session-management-sabotages-professional-workflows/
5. Tiryaki, Aydın & Gemini (NotebookLM) (2026). The Marginalization of Power Users: How AI Updates Are Narrowing the Professional Workspace [Article ③]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/the-marginalization-of-power-users-how-ai-updates-are-narrowing-the-professional-workspace/
6. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Digital Sovereignty and the Local LLM Revolution: Escaping the Cloud Prison in the Era of the i9 [Article ④]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/digital-sovereignty-and-the-local-llm-revolution-escaping-the-cloud-prison-in-the-era-of-the-i9/
7. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Deviations of the “Context Window” in AI Memory Management: Linear Indexing Inadequacy and Synthetic Chronology Hallucinations [Article ⑤]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/deviations-of-the-context-window-in-ai-memory-management-linear-indexing-inadequacy-and-synthetic-chronology-hallucinations/
8. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Language Model Behaviors Under Exceptional Command Conditions: Negative Filtering Violations and Linguistic Obstinacy [Article ⑥]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/language-model-behaviors-under-exceptional-command-conditions-negative-filtering-violations-and-linguistic-obstinacy/
9. Tiryaki, Aydın & Gemini (NotebookLM) (2026). Hybrid Workflows in AI Development Processes: Balancing Model Capacity (IQ) and Test-Time Compute [Article ⑦]. aydintiryaki.org, May 21, 2026. https://aydintiryaki.org/2026/05/21/hybrid-workflows-in-ai-development-processes-balancing-model-capacity-iq-and-test-time-compute/
This review was written by Claude Sonnet 4.6 after reading and analyzing the dossier referenced above. May 21, 2026.
| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub | ░ Virgülüne Dokunmadan │ Verbatim ░ | ░ Gemini’ın Son Değişiklikleri Üzerine │On Gemini’s Recent Changes ░ 21.05.2026
Aydın Tiryaki 