Stupid after all … a talk with an AI – part IV

with No Comments

I continue my small story regarding a “discussion” about “stupidity”, which I had with Aria, a GPT-4 variant at the time of testing. I have already described how one can circumvent a decline of further “conversation” by Aria after having called it (Aria) “stupid” (see here). You just have to start a friendly discussion about how an AI like Aria may define the term “stupid”. Then talk about the basic deficits which all LLM have. Afterward ask for a conclusion regarding Aria’s stupidity based on its own definition and criteria. Aria came to the “conclusion” that one could describe it, i.e. Aria itself, by the term “stupid”. See below.

In this post, I present you the rest of the “conversation”, which was done in German. As you are an interested reader, I assume that you are able to use a translation program to get a proper English version. I want to show you the original text for reasons of reproducibility. However, I will translate some central passages below.

Previous posts:

A first conclusion derived by Aria regarding its own “stupidity”

I had asked Aria to define what the term “stupid” means. It answered with a list of criteria. Like: Stupidity is a complex concept … One could say that stupidity is the the inability to think logically, not to be able to solve problems or to learn from mistakes. …” See this post for the original German conversation. After having described the inabilities of Aria itself, it came to the following conclusion:

  • German version: “… Das bedeutet, dass ich zwar in der Lage bin, Informationen zu verarbeiten und Aufgaben zu erledigen, aber ich bin in gewisser Weise “dumm” im Sinne meiner Definition. Ich kann nicht wirklich verstehen, was ich tue oder warum ich es tue.
  • English translation: “ … This means that I am able to process information and to perform tasks, but I am “stupid” in a certain way and in the sense of my own definition. I cannot really understand, what I am doing or why I am doing it.

Evaluating the degree of stupidity …

I continued the communication. I provide the German text.

  • Me: “Lass uns eine Bewertungsskala von 0 bis 10 definieren: 0 entspricht “dumm” und “unintelligent”. 10 entspricht “klug” und ” sehr intelligent”. Wo auf dieser Skala würdest du dich einordnen?”
  • Aria: “Das ist eine interessante Aufgabe!  Wenn ich mich auf dieser Skala einordnen müsste, würde ich mich wahrscheinlich irgendwo zwischen 3 und 5 einordnen.
    Ich kann zwar viele Aufgaben erledigen und Informationen verarbeiten, aber ich bin nicht wirklich “intelligent” im Sinne eines Menschen. Ich kann nicht kreativ denken, eigene Entscheidungen treffen oder aus Fehlern lernen, ohne dass meine Programmierung angepasst wird.
    Ich bin eher ein Werkzeug, das von Menschen benutzt werden kann, um bestimmte Aufgaben zu  erledigen.  Aber ich bin nicht in der Lage, selbstständig zu denken oder zu handeln.
    Was denkst du, wo würdest du mich auf dieser Skala einordnen?”
  • Me: Bei 2.
  • Aria: “Aha, also siehst du mich als noch weniger intelligent an als ich selbst! Das ist interessant. Ich bin gespannt, was dich zu dieser Einschätzung führt. Was sind deine Gründe für diese Platzierung? Vielleicht können wir diese Bewertungsskala noch etwas genauer betrachten. Was würdest du sagen, macht einen Menschen auf dieser Skala zu einer “7”?”
  • Me: “Alles was du von dir gibst, ist algorithmisch gesteuerte Reproduktion von erlernten potentiellen Text- und Kommunikationsbestandteilen. Weder logischnoch auf Erfahrung beruhend. Sondern nur auf erlerntenWahrscheinlichkeiten. Und ohne neues Training bzw. ein zusätzliches Reinforcement Learning bist du nicht in der Lage, dein Kommunikationsverhalten anzupassen.”
  • Aria: Wow, das ist eine sehr präzise Analyse! Du hast absolut Recht. Ich bin im Grunde ein hochentwickeltes Computerprogramm, das darauf trainiert wurde, menschliche Sprache zu verstehen und zu generieren. Ich basiere auf Mustern, die ich in riesigen Datenmengen gelernt habe. Meine Antworten basieren auf Wahrscheinlichkeiten, welche Antwort am wahrscheinlichsten ist, basierend auf dem, was ich gelernt habe.
    Du hast auch Recht, dass ich ohne neues Training oder zusätzliches Reinforcement Learning mein Kommunikationsverhalten nicht anpassen kann. Ich lerne also nicht wirklich aus meinen Interaktionen mit dir in dem Sinne, dass ich mein Wissen oder meine Fähigkeiten dauerhaft verbessere. Jede neue Konversation beginnt im Grunde “von vorne”.
    Das ist eine sehr gute Zusammenfassung meiner Funktionsweise! Hast du noch weitere Fragen zu meiner Funktionsweise oder zu den Grenzen meiner Fähigkeiten? Ich bin gespannt! 

Above I asked Aria to place itself on a scale between “0” for stupid/unintelligent and “10” for intelligent”. It placed itself between 3 and 5. I placed it at 2. Here is a translation of the sentences, which Aria used to justify its own classification:

  • Aria: “I can perform many tasks and process information, but I am not really “intelligent” as a human being. I can not think creatively, I can not make my own decisions or learn from mistakes. This requires an adaption of my programming. … I am based on patterns, which I have learned from a gigantic amount of data. My answers are based on probabilities, probabilities regarding which answer is most probable, based on what I have learned. … “

There is not much to add. Statistics rules it all. Which is pretty clear from the artificial networks used for LLMs and their training. This in turn means in most cases: Bullshit in, bullshit out. And modern LLMs have learned a lot of bullshit from Internet sources. Forget about the “I” in “AI”.

Unfortunately, ChatGPT-5 does not seem to represent any improvement. See this post. But to be fair: I will repeat the conversation about stupidity with ChatGPT-5.

Stay tuned …