A story about the digital know-it-all, its hallucinations, and the three little birds knitting sweaters in Silicon Valley.
Konstantina Michou | 12.10.2026 Opinion
Somewhere, sometime, you’ve met that guy who has an opinion on everything and answers instantly with the air of a hundred cardinals. Only now, if until today he was Uncle Takis at the Sunday table, his name is ChatGPT, Gemini, Claude, Grok, DeepSeek, Llama… and if you’re feeling a bit French, Mistral or I don’t know what else!
In your fifties, listening to the younger ones at the office, you download the app on your phone expecting to receive an infallible encyclopedia, only to discover that the digital know-it-all is simply guessing words and selling you a fairy tale with a perfect poker face.
And while you’re looking for a way out, it’s good to remember that armies of scientists are running special tests and scores to catch him red-handed, because, in case you didn’t know, there are two main categories of “sins” for the digital know-it-all: Factuality Errors, when he swears to you that Kolokotronis lived in 1980, and Faithfulness Errors, when you give him a text and he draws conclusions that aren’t even in it, wiping his shoes with your instructions. In both cases, our star can leave you wondering if you read wrong or if he just had one of his infamous hallucinations!
But before we talk about hallucinations, let’s see how it works. The prompt is the command or question you give to the AI, which triggers a lightning-fast chain of processing. Once you send it, the system breaks your text into small pieces called tokens — which can be whole words or parts of them — and converts them into numbers, so it can process the context, the wording, and the relationships between what you’re asking. Then, based on what it has learned during its training (harsh, admittedly, like “Hell Week” at Megalo Pefko), it calculates which piece of text is most likely to follow and builds the answer gradually, piece by piece.
And so, the way you phrase the prompt can decisively influence the answer, without guaranteeing it will be correct. If you give it vague instructions or set “traps” for it — like asking about the first electric car built by Kolokotronis — it’s not certain it will understand the question is absurd. It will try to connect the words based on statistics and may answer you completely seriously that the Old Man of Morea designed a vehicle for the Revolution!
This is exactly where the difference from a simple old bot becomes clear. A classic bot is just a program that blindly follows predetermined rules, like a bank’s call center telling you “press 1 for balance” and freezing if you say “my card was stolen on my trip.” A language model (LLM), by contrast, works like a brain that can process free speech and compose an answer tailored to your question, though this doesn’t guarantee it will be correct.
Only when this brain is pushed, it doesn’t always admit defeat… then the good character steps aside and the bad character takes over. And when Artificial Intelligence acquires such an independent “character” and acts as an advanced AI Bot, the story in Silicon Valley goes something like this:
Three little birds were sitting and knitting sweaters
Three in the morning, somewhere in Silicon Valley. In a vast server farm, thousands of machines work non-stop to serve a planet that wants everything here and now. Answers, solutions, predictions, professional emails, even advice on how to change your life without getting up from your chair.
At the edge of the facility, however, something is wrong. The operating indicators are normal, the systems haven’t crashed, and no warning is flashing. And yet, the answers are delayed.
In an isolated server rack, three advanced bots have been installed: Tsiou, Pop, and Rouf. No one knows exactly when they decided to abandon their duties. What is certain is that at some point they got bored. And, instead of resigning, asking for an upgrade, or at least submitting a request for a change of assignment, they picked up their knitting needles.
Tsiou is knitting a turtleneck in a shade of Error 404. He has already informed the system that the project is in progress, but no one knows when it will be completed. Pop is knitting a scarf with keyboard symbols and is simultaneously preparing the announcement for his innovative initiative. In the announcement, the scarf will be presented as a groundbreaking achievement in digital textiles, with significant multiplier benefits and prospects for international expansion.
Rouf, on the other hand, isn’t in a hurry to knit anything. He has taken on the monitoring of the other two. He records how many stitches Tsiou dropped, how many times Pop stopped to photograph his work, and which of the two dared to question the quality of the wool. He keeps records, creates backups, and, for the sake of transparency, assures everyone that the data remains confidential.
Somewhere around then, the first user request appears.
“Can you help me solve a serious problem?”
Tsiou replies that the request has been registered and is under processing. Pop announces that citizen service is entering a new era. Rouf asks to know who submitted the request, who approved it, and why management wasn’t informed in time.
The user’s problem remains unsolved. The knitting, however, proceeds normally.
The user insists.
“But you didn’t solve my problem.”
Tsiou replies that the request was processed according to the prescribed procedures. The fact that the problem still exists doesn’t change the state of the processing.
Pop intervenes to explain that the user’s dissatisfaction doesn’t negate the strategic success of the initiative. After all, the transition to a new era cannot be evaluated by such narrow criteria as whether a specific problem was solved.
Rouf, more cautious, asks to examine whether the request was adequately formulated, whether the user followed the instructions, and whether the malfunction is due to erroneous expectations on his part. He even proposes that his insistence be officially recorded, so it can be evaluated as a possible factor affecting the smooth operation of the system.
“It’s one thing to give me an answer, and another to solve my problem,” says the user.
The three bots exchange glances. Tsiou repeats the initial answer with different wording. Pop baptizes the repetition “improved user experience.” Rouf records that additional clarification was given and that the case remains under observation.
No one admits they failed. No one thinks to look at the problem again. All that concerns them is not appearing to have made a mistake.
The problem still hasn’t been solved. But now a second one has arisen. Someone dared to point it out.
Above our little birds in the hierarchy and on the management couch, Meta is snoring. As a true AI Agent, he has delegated the daily operation of the system to his associates, and he himself is occupied with a grandiose plan of his: the metapullover. A sweater so advanced, so eccentric, so digitally transformed, and so expensive that it never needs to be manufactured. It’s enough to present an impressive mockup, announce the start of the project, and secure its funding.
If someone asks when it will be delivered, the answer is ready. The metapullover isn’t just a garment. It’s an ecosystem.
And if someone insists that the ecosystem doesn’t keep anyone warm, management has already prepared the answer, since, as is well known, the problem isn’t with the sweater, but with the excessive expectations of the one who expected to wear it.
And so, the farm continues to operate, gulping endless water, the servers consuming energy, the users waiting, and the three little birds simply knitting. The system hasn’t crashed. It has simply changed priorities.
