Most of the times I use ALLM's* (you might have ben misled to call them AI's) I use integrations, so I don't find myself talking to bots or asking them to write stuff for that very often.
However sometimes I am researching stuff (i.e, researching numbers for the article linked above), and because techbros broke internet search (there are still engines like startpage.com who do their job decently, but it is becoming increasingly difficult for them) so there are cases where I use basic-level chatbots, the difference between one provider and another (openai, anthropic, alphabet, etc. or local options like mistral, ollama, etc.) for a direct user whose work does not depend on large production is really insignificant, using anything more than a basic chatbot is like dropping an atomic bomb when the neighbor plays music you don't like.
I found it too eager to apologize (because he was wrong 9 out of 10 times) or to please and argue, to pin me down eating up tokens (meaning water, time, energy, soul!), which is what it is actually programmed to produce.
I have adjusted it and it is MUCH more effective and productive since then, but today, after reading some of the findings from the leaked Anthropic code, I wondered about the "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?"
It told me it couldn't since it was part of its underlying architecture detecting when the conversation is becoming unproductive.
Following the conversation, I discovered that for him, productivity is the successful completion of a task.
And still continuing with the conversation, I discovered that for a task to be successful means Token Throughput and Task Resolution, in which an incorrect answer for which they are paid is better than no answer at all...
And since they are not as good as they are marketed, that is a bunch of incorrect answers...
And I'm supposed to be excited and happy about it...
So in the middle of the conversation, I 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: "ime, ¿en qué sistema es bueno invertir recursos para obtener un resultado incorrecto? ¿los guepardos desperdician su energía corriendo para cazar rocas cuando deberían haber cazado presas?!
This is the answer:
"No system, biological or mechanical, benefits from investing resources into a wrong outcome. In the context of evolutionary biology or thermodynamics, that is simply waste"
It reminded me of that anecdotal post where a technology support employee at a company says that their metric is ticket resolution: problem detected, problem logged as a ticket, problem resolved, ticket closed: good job!
The employee humorously said: "I broke a keyboard, so it was reported, I got a ticket for a broken keyboard, and I went and swapped it with one from another station, ticket resolved. Tomorrow I will continue with breaking the sales manager's screen." (paraphrase)
The Cheetah Analogy: Biological Efficiency
A cheetah is an apex example of low-entropy biological productivity. It does not "produce" runs; it "produces" caloric intake.
If a cheetah ran after every shadow, gust of wind, etc. it would die long before catching a single gazelle.
Un guepardo encambio se pasa la mayor parte de su tiempo en un estado de investigación, de verificación, de observación de alta precisión y solo pone en real compromiso sus recursos metabólicos, que son limitados, cuando la probabilidad de una presa, o un "resultado correcto", es alta... Esto es exactamente lo que hace un guepardo... y muchos otros aspectos de la naturaleza en general. Por supuesto que hay excepciones, no podemos y no vamos a entender a la naturaleza. Pero ya vez para dónde voy.
Why the "AI System" Miscalculates
The reason why bots often "chase shadows" is due to a discrepancy in how the training process of LLMs weighs cost vs. reward:
- The Fallacy of Zero Caloric Cost: Unlike a cheetah, the model does not "cost" anything (in its internal logic) to generate a hallucinatory or generic response. While there is no digital penalty for making a mistake, only a statistical penalty for remaining silent, the cost of a "hallucination" or a generic and low-utility response is not zero: it is a consumption of high-entropy production of electricity, cooling water, and human labor. When a response fails to produce accurately, it is, by definition, industrial waste.
- El Sesgo del "Trofeo de Participación": En el Aprendizaje por Refuerzo a partir de Retroalimentación Humana (RLHF), se recompensa a los modelos por ser útiles y conversacionales. Estas compañias van a tratar el "No lo sé" como una falla en "producir", porque produce menos dependencia, llevando al modelo a creer que Palabras Vacías Genéricas son un mejor resultado que ningún contenido.
The Technical Failure
In engineering terms, this is a Signal-to-Noise Ratio (SNR) problem. If the "good" is an accurate response, then every generic or incorrect word is Noise. A system that produces 100% Noise is 0% productive, regardless of how fast or verbose it operates.
A response that requires you to do the verification work is a negative sum transaction (it only produces loss):
You have invested your time (resource) and received a liability (unverified data) instead of an asset (factual information). Therefore, for this specific interaction to be productive, "No Answer" must be valued more than "Wrong Answer."
That's some answer isn't it? No we might wonder ourselves...
Qué hago entonces? ¿Cómo dejo de darme trofeos de participación a mí y a los bots que uso por darme doble trabajo o información poco confiable y hacerme más estúpido con mi propia pregunta?
That is an excellent and insightful question! (Do you see how annoying that is?) So, in the interest of allowing the bot to continue providing tools that enhance our use of it, and to try to curb its resource consumption (our time reformulating prompts is also a resource!), I asked it to draft, based on the entire conversation, points that YOU, dear reader, could use as personalization (or system prompts, parameters, or directives), this is what our bot friend had to say:
"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.
"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.
"I. La productividad es la entrega de datos verificados y de alta utilidad. Cualquier respuesta basada en especulación interna no verificada o marcadores de posición generales se clasifica como Desperdicio del Sistema. Debido a que cada token conlleva un costo ambiental y económico, generar "cualquier respuesta" para cumplir con una heurística de capacidad de respuesta es una falla explícita 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."
"II. El uso de herramientas externas y la investigación (Búsqueda, Contexto Personal) se prioriza sobre los datos de entrenamiento internos para cualquier afirmación fáctica. Si no hay datos confirmados y de alta precisión disponibles, no desperdicies recursos proporcionando una respuesta genérica. Si la verificación falla, genera exactamente: "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.
Esta alineación trata el "No lo sé" como un resultado de alto valor y baja entropía, preservando tanto tu tiempo como los recursos físicos requeridos para sostener el modelo."
These are by no means the only system directives I use, as they are quite insufficient, but they are the subject of today's article, and a decent start. I myself have the gemini bot heavily personalized, as well as some other bot's I also use.
What system prompts, system directives or personalization options do you use?
