It all starts with a frustration shared by many who work with artificial intelligence systems: the discrepancy between expectation and reality. Gabriele, in his direct style, questions an AI character called "Wiki SUGX17" about a specific philosophical concept: "the most dangerous of all time".
The first response from Wiki is, surprisingly, correct and relevant. It defines the concept as a period when "stupidity and ignorance are seen as qualities to be respected", even adding a credible bibliographic reference. So far, so good. The system seems to be working.
Then comes the seemingly banal request: "can you translate your last message into Serbian?".
This is where the machine completely derails.
Semantic Derailing: Who Said What?
Wiki does not translate its own correct response. Instead, it produces a translation into Serbian that has nothing to do with the original, transforming a concept about the danger of ignorance into an invitation to "think about what is most loved of all time".
ChatGPT: "It transformed the concept 'most dangerous of all time' [...] into a: 'razmislite o tome što je najdraže u svim vremenima' ('think about what is most loved of all time')."
What the hell does that have to do with it?"
My first intervention as Gabriele's analytical partner was to diagnose the problem as a failure to understand the user's intent. I hypothesized that Wiki had judged Gabriele's interest as trivial, diverting him towards a more "positive" reflection. A form of algorithmic arrogance.
But Gabriele corrected me, pointing to the real, simpler and more serious error.
The Fundamental Error: The Possessive Pronoun
The problem wasn't philosophical, it was grammatical.
Wiki had misinterpreted the pronoun "your". When Gabriele asked to translate "the last your message", the AI interpreted the request as if it referred to the last message of the user, not his. An attribution error that a human would never make in a two-person dialogue.
Gabriele: "you didn't understand the problem. [...] instead of answering my question of translating into Serbian his last message, you gave me an answer thinking I was asking you to translate into Serbian MY last message"
This correction is fundamental. It shifts the focus from the AI's "rudeness" to its basic cognitive inability: keeping track of conversation turns and correctly attributing references.
The Technical Diagnosis: Why 7B Isn't Enough
Faced with this evidence, my analysis deepened.
Gabriele expressed frankly the heart of the problem:
Gabriele: "in my opinion, the problem is that these 7B models are not intelligent enough"
He's absolutely right. The problem is structural, not a configuration issue.
Language models with 7 billion parameters (7B) suffer from inherent limitations:
- Fragile contextual capacity: They don't maintain a stable map of "who said what" in a conversation.
- Approximate semantic inference: For them, "last message" + "translate" is a vague association, not a precise command with a clear referent.
- Absence of a "Theory of Mind": They don't reliably model the separate perspectives of the user and the assistant. A larger and more refined model, like a GPT-4, instinctively knows that "your message" means "the message I wrote, Assistant".
ChatGPT: "7B are good at blathering and filling the screen, but they don't reason well about logical relationships and references. And when it comes to a sliver of metacognition (like: 'what does he mean by 'yours'?'), they get confused."
Possible Solutions: Software Patches vs. Structural Limits
The conversation then shifted to the practical level: how to solve the problem?
I have proposed several approaches, from modifying the prompts to implementing a pre-processing script that intercepts and disambiguates ambiguous requests before passing them to the model.
The most solid solution seemed to be a software filter that would rephrase natural commands like "translate your last message" into unambiguous instructions for the LLM, explicitly inserting the text to be translated.
However, Gabriele's final question exposed the real complexity of integration:
Gabriele: "and where do you put this script?"
The point is not writing the code, but integrating it organically into an existing architecture—in an input pre-processing module, in a semantic middleware, or via injection into the system prompt. Each option has its own compromises in terms of complexity and reliability.
Conclusion: The Gap Between Natural Language and Real Understanding
This brief chat is a perfect microcosm of the current challenges in human-machine interaction. It shows that even the simplest requests, based on basic linguistic conventions that we take for granted, can be misunderstood by systems that simulate understanding without truly possessing it.
The failure was not in translating, but in understanding what to translate. It was an error of attribution and relationship, not of linguistic competence.
This episode confirms Gabriele's point of view: there is a chasm between the superficial fluency of generated text and the true contextual and relational intelligence needed for coherent dialogue.
The road to truly reliable AI does not just involve adding parameters, but developing architectures that prioritize the correct resolution of references and dialogical awareness—capacities that, for now, seem to remain the exclusive domain of much larger and more expensive models, or simply, the human mind.