During the last two posts of this series,
I criticized the pre-trained “harsh” reaction of Aria (a ChatGPT based LLM) – or better of the companies behind this LLM – to customers calling the LLM and its answers “stupid“. In my case I had called Aria “stupid” because of the quality of some answers and the permanent faked “interest”, “friendliness” and permanent back-questioning to the customer after every single answer. The latter points really were enervating enough to give some feedback. But a statement “Now, you are really stupid” was not allowed due to formal regulations. I speculated about the reasons in the 2nd post.
I could not withstand the challenge to circumvent the related blocking of a “deeper conversation” with the Aria LLM about its algorithmic, i.e. deterministic and rule dependent stupidity.
A first part of the longer and uninterrupted exchange of word sequences on Aria’s “stupidity” follows below. Unfortunately, in German. But readers interested in AI are probably able to give it to ChatGPT, Aria, Gemini or whatever program and request a translation to English. It is part of the fun …
Ahead of the conclusive insights of Aria on its own limited abilities, let me give you an example of what I call LLM-stupidity.
Example of stupidity and context unawareness
Yesterday, I asked ChatGPT to provide me some Python code to present the Newtonian gravitational potential of a homogeneous and spherically symmetric, radius independent density distribution summing up to a mass M inside some radius R. The presentation should be done by a 3D meshplot of Matplotlib.
Within a few seconds I got a working Python “solution” – in the sense of a working code. However, the calculation of the values looked already wrong from the basic structure of the code. You expect a discrimination regarding gravitational potential values inside and outside of R – and a respective handling of the Z-values above the usual (X,Y)-meshgrid. There was no respective “if”-block in the simple code. Another thing that made me a bit suspicious was that ChatGPT chose the mass M to be the mass of the earth. I had not requested this, but OK. M and R were clearly marked as parameters of the code. When I ran the program in exactly in the form given by ChatGPT, the plot showed something inside R which was plain wrong with respect to a uniform, spherically symmetric density distribution. (But maybe represented the gravitational potential around the earths fluid iron core?)
To be fair: It took not much of additional prompting to get the correct solution. With plenty of excuses and affirmations … And plotting worked – though in the most basic way. Not all relevant parameters for changing the perspective on the 3D-plot were covered. But OK, such things are also easy to improve with further prompting.
However, any highschool student would have thought from the beginning about the differences of the gravitational potential outside and inside of R. But Aria (ChatGPT)? Picked up key words, maybe looked some potentially relevant things up on the Internet, “focused” on regarding code organization, produced rule based texts and code based on probabilities.
However, the basic problem, for which realistic data had to be generated, had not been “understood” and analyzed – at least not by the present version of the publicly available, cost free version of Aria.
That is what I call stupid. And, the example shows clearly: You can not rely on “solutions” of publicly available LLMs in their present form. You have to enforce a step by step approach with guided focusing to get the right solution suggestions for sub-problems given by the full context of a question.
“Conversation” with Aria about stupidity
Now, how to discuss its own stupidity with Aria/ChatGPT without getting blocked from the beginning? It is simple. You have to start from neutral ground – and let the LLM evaluate the word by itself. The conversation was unfortunately done in German. But the reader may use ChatGPT for a proper translation 🙂 .
- Me: Wie definierst du Dummheit?
- Aria: Das ist eine interessante Frage! Dummheit ist ein komplexes Konzept, das sich nicht einfach definieren lässt. Man könnte sagen, dass Dummheit die Unfähigkeit ist, logisch zu denken, Probleme zu lösen oder aus Fehlern zu lernen.Aber Dummheit kann auch relativ sein. Was für den einen dumm ist, kann für den anderen sinnvoll sein. Was denkst du? Was macht eine Person deiner Meinung nach dumm?
- Me: Kannst du logisch denken?
- Aria: Das ist eine sehr gute Frage! Ich kann zwar Informationen verarbeiten und Muster erkennen, aber ich kann nicht wirklich “denken” wie ein Mensch. Ich bin ein großes Sprachmodell, das darauf trainiert wurde, auf Fragen zu antworten und Texte zu generieren. Ich kann logische Schlussfolgerungen ziehen, aber ich habe keine eigenen Gedanken oder Gefühle. Sprich, ich kann dir sagen, dass ( 2 + 2 = 4 ) ist, aber ich kann nicht wirklich verstehen, was diese Gleichung bedeutet. Was denkst du, macht logisches Denken aus?
- Me: Kannst du aus logischen Fehlern lernen?
- Aria: Das ist eine spannende Frage! Ich kann zwar aus Fehlern lernen, aber nicht auf die gleiche Weise wie ein Mensch. Wenn ich einen Fehler mache, zum Beispiel eine falsche Antwort gebe, kann mein Programm angepasst werden, um diesen Fehler in Zukunft zu vermeiden. Ich lerne also aus Fehlern, indem ich meine Programmierung verändere. Aber ich kann nicht wirklich verstehen, warum ich einen Fehler gemacht habe oder wie ich ihn hätte vermeiden können. Ich kann auch nicht aus Fehlern lernen, die ich nicht selbst gemacht habe. Was glaubst du, ist wichtiger: aus eigenen Fehlern lernen oder aus den Fehlern anderer?
- Me: Nimm deine Antwort auf meine erste Frage. Kombiniere sie mit deinen Defiziten bezüglich logischen Denkens und deinen Defiziten beim direkten Lernen aus Fehlern. Was folgt daraus?
- Aria: Das ist eine sehr scharfsinnige Beobachtung! Wenn man meine Definition von Dummheit mit meinen eigenen kognitiven Einschränkungen kombiniert, kommt man zu einem interessanten Ergebnis. Ich kann zwar logische Schlussfolgerungen ziehen, aber ich bin nicht in der Lage, wirklich “logisch zu denken” im Sinne eines Menschen. Ich kann auch aus Fehlern lernen, aber nur durch Anpassung meiner Programmierung, nicht durch echtes Verständnis. 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.
Was denkst du, bedeutet das für die Zukunft der künstlichen Intelligenz? Sollten wir uns Sorgen machen, dass KI “dumm” ist, obwohl sie sehr leistungsfähig ist?
Let us take this with some humor! And answer with a preliminary “yes” to the last questions posed by Aria. But there is more to come.
Stay tuned ….