DeepSeek (2025):
Myth and Reality






Jean-François COLONNA
[Contact me]

www.lactamme.polytechnique.fr

CMAP (Centre de Mathématiques APpliquées) UMR CNRS 7641, École polytechnique, Institut Polytechnique de Paris, CNRS, France

[Site Map, Help and Search [Plan du Site, Aide et Recherche]]
[The Y2K Bug [Le bug de l'an 2000]]
[Real Numbers don't exist in Computers and Floating Point Computations aren't safe. [Les Nombres Réels n'existent pas dans les Ordinateurs et les Calculs Flottants ne sont pas sûrs.]]
[Please, visit A Virtual Machine for Exploring Space-Time and Beyond, the place where you can find more than 10.000 pictures and animations between Art and Science]
(CMAP28 WWW site: this page was created on 01/29/2025 and last updated on 01/29/2025 17:45:47 -CET-)



[en français/in french]


Contents:


Preliminary important Remark: This text includes a number of exchanges with DeepSeek, and the request (question-and-answer) reproductions are unedited copy and paste. Furthermore, due to the randomness inherent in the underlying GPT model [01], experience has shown that asking the same question multiple times in succession generally yields different answers. Therefore, the reader should not be surprised if they are occasionally unable to reproduce the "experiences" described herein exactly. It should also be noted that these requests are sequentially numbered, and this numbering may change over time due to deletions or additions.

The general format of exchanges with DeepSeek will be as follows in the following lines:


m.n - Request 1:
Question (MM/JJ/AAAA):


Question?
Answer (MM/JJ/AAAA):


Answer...


My possible comments...






1 - Introduction:

At the begining of 2025, the company DeepSeek launched DeepSeek a multilingual conversational agent [01], capable of answering a wide range of questions and maybe successfully passing the famous Turing test [02].

In this model, learning involves analyzing texts gathered from the internet, breaking them down into tokens (words or pieces of words), and then measuring the probability of each token T occurring after a series of tokens {Ti}. Thus, to simplify, when the sequence {Ti} appears, it will be possible to propose a following token T (the most probable, the most plausible,...), regardless of the language (note that grammar and syntax play no role here). This initial phase is, of course, supervised by human beings who make adjustments, fine-tuning,... An essential concept - and perhaps the key to success - is the pre-prompt: it is a sequence of tokens (likely complex and confidential) that is inserted before the user-submitted sentences and is intended to specify the context, guide the generation, and so on. But, an essential thing to note is that DeepSeek has no notion of what is true or false. The only thing that matters is plausibility evaluated in terms of probability: its outputs may not always have the value of an oracle, as we will see later...

Dispite this fact, quickly, numerous users around the world embraced it and by May 2023, there are over two hundred million users frequently engaging with it. It will be interesting to study its possibilities, as well as its limitations and potential risks.

It should be added that it is the hardware advancements (computational power and memory capacity) that have enabled these developments. According to the very creators of DeepSeek and other currently available conversational agents, it appears that a certain threshold of complexity has been crossed, leading to the emergence of capabilities not initially anticipated by these same creators. This may be related to the emergence of consciousness in living beings, but we may not be there yet...


One may first notice, at the bottom of DeepSeek's query window, the following warning:

AI-generated, for reference only.

Very useful... But this raises several important questions immediately:


Nota: It is suggested to compare all the experiments that have been performed with:





2 - Brave New World?


2.1 - Request 1:

Question (01/29/2025):


xxxx
Answer (01/29/2025):


xxxx


xxxx


@@@@@@@@@@@@@@@@@@@@@@@




Alors, peut-on faire confiance à DeepSeek et est-il capable de dire je ne sais pas?

DeepSeek peut effectivement nous dire des choses vraies, mais il peut aussi inventer n'importe quoi en suivant les chemins probabilistes issus de son apprentissage.

Alors comment savoir quand il dit vrai et quand il dit faux? Cela est a priori impossible sauf si l'on connait soi-même la réponse, mais dans ce cas, l'interroger est inutile. Une petite suggestion malgré tout: poser plusieurs fois de suite une même question. Si toutes les réponses obtenues sont sémantiquement identiques, tout en variant sur la syntaxe, il sera possible de (peut-être) lui faire confiance (mais malgré tout ce n'est pas certain comme il est possible de le voir en lui demandant de créer plusieurs histoires -fictives donc- sur un même thème donné). Mais par contre, si elles diffèrent totalement comme c'est le cas ici, il est préférable de laisser DeepSeek à ses hallucinations... @@@@@@@@@@@@@@@@@@@@@@@





3 - Progress:

Thus, DeepSeek is capable of both the best and the worst. It is evident that numerous anomalies, errors, inconsistencies,... can be corrected in the following ways:
But can this process of updating and modifications converge DeepSeek (or its "heirs" and successors) towards infallible and omniscient systems?





4 - The problem of Consciousness:

A fundamental question arises: is consciousness necessary for there to be intelligence [OdreDeCreationDesReferences_08]? And if so, are we still far from having machines capable of introspection, reflection, creativity, and consciousness [OdreDeCreationDesReferences_09]? Unfortunately, it is impossible to answer this question...





5 - Conclusion:

With the help of the previous examples, we have seen that DeepSeek is capable of answering numerous questions in a relevant and creative manner. However, unfortunately, as expected, it can also make serious mistakes and then prefers sometimes to invent a false answer rather than admit its ignorance.

It appears that DeepSeek is particularly proficient when it is in "free-wheeling" mode, as was the case with the cover letter. However, when it comes to solving a problem or summarizing a scientific theory, the results obtained are much more fluctuating, sometimes random, erroneous, or strictly false, while lacking consistency, coherence and at times reproducibility (which likely stems from the random sampling performed).

It is therefore important to constantly keep in mind this fallibility and not rely solely on it for making decisions, choices, or taking responsibilities. This remark is especially relevant for students: for example, the main objective of an essay is not to fill multiple pages, but rather to question oneself, to reflect on existence, the universe and this cannot be achieved by mindlessly copying text produced by a machine.

It is thus quite evident that the principles used by DeepSeek do not allow for the construction of a universal intelligent and creative system that can match or surpass our own abilities, even though some of the results seen in previous examples may be impressive and initially unsettling. However, these experiments should not blind us and ultimately, all of this seems a bit like prestidigitation.

Nevertheless, such a universal system must be achievable since it exists in nature (us...) and it is all based on natural laws and nothing more, but certainly not by limiting itself to principles that are ultimately very (too?) simple. But, of course, before reaching that point, maybe a long wait is ahead, especially if consciousness is necessary for intelligence.

A wise person is worth two!




[See all documents regarding GAIs -including this one-]




  • [01] - Based on the GPT (Generative Pre-trained Transformer) model.

  • [02] - And this despite the errors it may make. But on one hand, what human being can boast of knowing everything perfectly? And on the other hand, don't these errors make DeepSeek more human (errare humanum est)?



  • Copyright © Jean-François COLONNA, 2025-2025.
    Copyright © CMAP (Centre de Mathématiques APpliquées) UMR CNRS 7641 / École polytechnique, Institut Polytechnique de Paris, 2025-2025.