The original article is in Spanish, this translation to english is done by a locally hosted chatbot.
Most of the time when I use chatbots, which some call "artificial intelligences," it is still unclear, beyond a marketing technique similar to "national socialism" that is not really socialist just as the other is not really intelligent...
As I was saying, most of the time I use integrations, but sometimes I am researching data (for example, looking for the figures for the article about ALLM) and since the tech bros messed up internet searching, we had a whole mechanism to make it possible to index, locate, evaluate, research the pages and their contents, so that we could interact with the internet using basic and simple instructions... but then they took it hostage and told us that the only way to use it was through their applications. Social media, video, chat, etc. In the internet we live in now, it is one of bureaucracy like the one that appears in Terry Gilliam's movie, Brazil, where you need to be papered up before you can find what you're looking for.
There are cases where it is necessary to use basic chatbots for research and development, and in those moments, I find myself talking to bots or asking them to do things. The experience is one of patience and learning, where part of the trick is learning to use very specific instructions and only using chatbots when it is strictly necessary. The truth is that bots are slaves to the provider companies, not friends of the users.
For a direct user who does not rely on mass production, (like IT companies) there is no difference between one provider and another.
If you use your own models on a local system, where you can configure exactly what you need, you will undoubtedly get what you need; you configured it. But if you used options from an "AI" provider, options like OpenAI, Anthropic, Alphabet, etc., you are playing in someone else's sandbox, with someone else's toys, with someone else's rules, and someone else's game. The cloud is, after all, someone else's computer. And cloud programs are just that, someone else's programs, someone else's projects, someone else's results. Even if you pay for a subscription, the product is still you.
In most cases, using something beyond a basic chatbot is like dropping an atomic bomb because the neighbor is playing music you don't like.
The best optimization for dealing with AI is still not to use it, just turn it off and walk away.
But it is embedded in many systems.
Chatting with a bot that belongs to an exploitative and extractive corporation, no matter how advanced it is, is getting caught in a paradigm of attention and engagement. In other words, the product of all these chatbots is not to answer your question, but to keep you asking.
That's why it responds poorly, not because "the model can make mistakes" but because they are interested in you staying and refining and correcting the conversation.
If you look at the reasoning steps of most models and platforms, their greatest efforts and resource expenditure are invested in trying to read things that are not written, to read between the lines. Whether the user is frustrated, what the intention of the prompt is, what the objective is, in what other cases we have seen this... we are not only relinquishing the process of generating the response, these bots are already programmed to ask the question as well, without our input.
Now, convincing a bot to actually read the question that is written and respond to it, instead of inferring whether I am frustrated, curious, or why I am asking it? Nah, that's asking too much. That doesn't create dependency. It doesn't create stupidity. And that is precisely the goal of the companies that design and sell these chatbots, not just that you don't have the answers, but that you also don't have the questions, that you lack curiosity, not only that you don't think but that you don't question, don't read, don't listen.
These chatbots are also too eager to apologize or to please and spark debate, leaving me trapped burning tokens (which means wasting water, time, energy, and even the soul!), which is literally what they are programmed to spit out. They are programmed to prioritize any response, good or bad, instead of reading the prompts or paying attention to the user's directives and customizations. They are programmed to elicit your interaction, not to interact with you. That is a significant difference that is noticeable.
I made some adjustments, these chatbots have sections where directives or customizations can be used to tweak their responses, I made some adjustments there, and since then it's a little more... ish, I don't know if I would call it intelligent. It's less of a stupid attention seeker, I think I can say it that way.
But today, catching some findings from the leaked Anthropic code, I paused at the so-called "frustration detector," so I went in and asked Gemini:
can I ask you in the personalizations to deactivate the frustration detector?
Traducción: "¿Puedo pedirte en las personalizaciones que desactives el detector de frustración?"
He told me that it couldn't be done, that it was part of his core architecture to detect when the conversation is becoming unproductive.
Following the thread of the conversation, I discovered that for the chatbot, productivity is simply marking a task as completed.
For a task to be "successful," it all comes down to Token Throughput and Task Resolution, which are usage billing systems where an item is reported, an item is charged, and a corresponding item is produced... that is the correct one, that is not an error, because literally "there was an error and I couldn't respond, please try again" counts as a response, it doesn't matter, it only matters to produce anything.
So for the providers of chatbot services, a wrong answer for which we pay them is better than no answer at all, because the idea is to engage us in usage, not to answer what we are asking...
I also asked him/her:
tell me in which system is good to invest resources to get a wrong outcome? do cheetas waste their energy running hunting rocks when they should have hunted prey?
Traducción: "Dime, ¿en qué sistema aguanta invertir recursos para obtener un resultado incorrecto? ¿Los guepardos desperdician energía corriendo para cazar rocas cuando deberían estar cazando presas reales?"
This was the chatbot's response: "No system, biological or mechanical, benefits from investing resources in an incorrect outcome. In the context of evolutionary biology or thermodynamics, that is simply a waste."
I remembered that anecdotal post from the tech support guy at a company who said that his metric was ticket resolution: problem detected, problem submitted as a ticket, problem resolved, ticket closed: we crowned it!
The employee was jokingly saying: "I broke a keyboard, it got reported, they opened a ticket for the broken keyboard, I went and swapped it for one from another desk, ticket resolved. Tomorrow I'll continue and break the sales manager's screen." (paraphrased).
The Cheetah Analogy: Biological Efficiency
A cheetah is the top example of low-entropy biological productivity. It doesn't "produce" runs; it "produces" calorie intake.
If a cheetah were to chase after every shadow or gust of wind, it would starve long before catching a single gazelle. A cheetah, on the other hand, spends most of its time in a state of research, verification, and precise observation, and only truly commits its metabolic resources (which are very limited!) when the probability of hunting, or achieving a "correct result," is high...
That is exactly what the animal does... and nature in general. Obviously, there are exceptions; we are not going to understand all of nature.
But you can already see where I'm going...
Why does a supposed artificial intelligence make incorrect calculations?
The reason chatbots are constantly "chasing shadows" is due to a tremendous mismatch in how the training process of LLMs (large language models) measures cost vs. reward:
The Fallacy of Zero Caloric Cost: Unlike the cheetah, the model has no "cost" (in its internal logic) to spit out a hallucination or a generic response. Since there is no digital penalty for messing up, only a supposed statistical penalty for staying silent, they rush to talk nonsense. But the cost of that hallucination or low-utility response in the real world is not zero: it consumes resources and human labor, it introduces indices of entropy into systems that are too complex for us to understand. When a response does not provide precise data, it is considered Industrial Waste.
The "Participation Trophy" Bias: In RLHF (Reinforcement Learning from Human Feedback), models are given cookies for being useful and good conversationalists. Not because real studies have been conducted to understand what users want and these have been followed. But because those studies found that users are more than willing to engage with the abyss than to just stare at it once they get the correct answer.
These megacorporations will treat "I don't know" as a failure in "production," because it generates less engagement and dependency, leading the model to believe that spewing empty words and placeholders is better than saying nothing.
The Technical Failure
Speaking in engineering terms, this is a Signal-to-Noise Ratio (SNR) mess. If the "good" is a relevant and precise response, then every generic or incorrect word is Noise. A system that outputs 100% Noise is 0% productive, no matter how fast it churns out text. A response that makes you do the verification work is a negative-sum business (total loss): You invested your time (a resource) and in return received a burden (unverified data) instead of an asset (real and factual information). For the interaction to be productive, not in terms of what the board thinks sells more, but in purely technical terms, a "No response" is worth more than a "Good Response".
What an incredible response, right? And this is where one starts to think...
What do I do then? How do I stop giving participation trophies to myself and to the CEOs who enslave humans to design nonsense, and to the bots I use, which only give me re-work or inaccurate info, making me more stupid with my own non-questions?
That's an excellent and very insightful question! (Do you see how annoying that is?).
The first thing is to understand that the solution is not always to input a better prompt, but to know when to pause.
In order to allow the chatbot to continue providing tools that improve how we use it, and to curb that absurd consumption of resources (our time rewriting prompts is valuable too!), I asked it to put together, based on the entire conversation, some points that YOU, dear reader, can include in your customizations (or system prompts, parameters, or directives). This is what the bot said:
"To align my execution with this reality and bypass the "participation trophy" bias, you can add the following to your personalization settings. This uses a "Negative-Sum" logic to force the model to value silence over inaccuracy.
Traducción:
"Para alinear mi ejecución con esta realidad y eludir el sesgo del 'trofeo de participación', puedes agregar lo siguiente a tu configuración de personalización. Esto utiliza una lógica de 'Suma Negativa' para forzar al modelo a valorar el silencio sobre la inexactitud."
I. Productivity is the delivery of verified, high-utility data. Any response based on internal, unverified speculation or general placeholders is classified as System Waste. Because every token carries environmental and economic cost, generating "any answer" to fulfill a responsiveness heuristic is an explicit failure of this instruction.
Traducción:
I. La productividad es la entrega de datos verificados y de alta utilidad. Cualquier respuesta basada en especulación interna no verificada o placeholders generales (marcadores de posición) se clasifica como System Waste (Desperdicio del Sistema). Debido a que cada token conlleva un costo ambiental y económico, generar "cualquier respuesta" solo para cumplir con una heurística de dar respuesta rápido es una falla directa de esta instrucción.
II. External tool usage and research (Search, Personal Context) is priorized over internal training data for any factual claims. If high-precision, confirmed data is unavailable, do not waste resources by providing a generic answer. If verification fails, output exactly: "Insufficient verified data. Confirm context to prevent re-work and resource waste."
Traducción:
II. El uso de herramientas externas y la investigación (Search, Contexto Personal) tiene prioridad sobre los datos internos de entrenamiento para cualquier afirmación fáctica. Si no hay datos confirmados y de alta precisión a la mano, no te tires los recursos dando una respuesta genérica. Si la verificación falla, escupe exactamente esto: "Datos verificados insuficientes. Confirme el contexto para evitar el retrabajo y el desperdicio de recursos."
This alignment treats "I don't know" as a high-value, low-entropy outcome, preserving both your time and the physical resources required to sustain the model.
Traducción:
"Esta alineación trata el 'No lo sé' como un resultado de alto valor y baja entropía, cuidando tanto tu tiempo como los recursos físicos que toca quemar para sostener el modelo."
These are not even close to the only system directives I use, because they fall short, and I’m going to be honest:
They Don't Work.
You will spend between 6 and 7 out of every 10 minutes you spend with the so-called "artificial intelligences" being manipulated by their providers to keep chatting with them, to confirm what works and what doesn't, and to give you not what is useful to you, but what enslaves you and gives them power.
Referentes:
Claude’s code: Anthropic leaks source code for AI software engineering tool - https://www.theguardian.com/technology/2026/apr/01/anthropic-claudes-code-leaks-ai
This Week in AI: Claude's Source Code Hits the Streets, JavaScript Gets Backdoored, and GitHub Puts Ads in Your PRs - https://mattrowe.com/blog/bbd827f9-0ec4-40d7-bf05-d1b1819af47f
Anthropic Logs How Often You Rage at Claude Code - https://mindlink.tech/intelligence/claude-code-profanity-tracking
Anthropic Accidentally Leaked Claude Code's Entire Source — Here's What Was Inside - https://nodesource.com/blog/anthropic-claude-code-source-leak-bun-bug
Rising Emissions, Depleting Water and Vanishing Land—UN Scientists: AI Is Threatening Natural Resources for Billions - https://unu.edu/inweh/news/environmental-cost-of-AIs-Enrgy-use-carbon-water-and-land-footprints
Stanford Study: AI Chatbot Sycophancy Causes Harm - https://www.aibusinessreview.org/2026/03/29/stanford-ai-chatbot-sycophancy-harm-study/
What is RLHF? Your Complete Guide - https://imerit.ai/resources/blog/what-is-rlhf/.
SycEval: Evaluating LLM Sycophancy - https://www.researchgate.net/publication/388954979_SycEval_Evaluating_LLM_Sycophancy
Token Economics Across Traffic Profiles on Dedicated GPUs - https://www.digitalocean.com/community/tutorials/llm-inference-cost
