I feel you’re wasting your time here. Some people seem to be under the impression it’s the year 1990 or 1950 and we’re talking about markov chain chatbots. The stochastic parrot argument would certainly apply there. But we’re talking about something else here.

And it’s also a fairly common misconception that AI somehow has to be intelligent in the same way a human is. And by using the same methods. But it really doesn’t work that way. That’s why we put the word “Artificial” in front of “Intelligence”.

But this take gets repeated over and over again and I don’t really know why we need to argue about how maths and statistics are a part of our world, how language and perception work and who is dehumanizing themselves… The scientific approach is to define intelligence, come up with some means of measuring it, and then do it… And that’s what we’ve done. We can get rid of the perception part of language. We can measure how “intelligent” entities can memorize and recall facts, combine them, transfer and apply knowledge… That’s not really a secret… I mean obviously it seems to be misunderstood or hyped or whatever by lots of people. But we also (in theory) know some of the facts about AI and what it can and can not do and how that relates to the vague concept of intelligence.

Go to one of these “reasoning” AIs. Ask it to explain its reasoning. (It will!) Then ask it to explain its reasoning again. (It will!) Ask it yet again. (It will gladly do it thrice!)

Then put the “reasoning” side by side and count the contradictions. There’s a very good chance that the three explanations are not only different from each other, they’re very likely also mutually incompatible.

“Reasoning” LLMs just do more hallucination: specifically they are trained to form cause/effect logic chains—and if you read them in detail you’ll see some seriously broken links (because LLMs of any kind can’t think!)—using standard LLM hallucination practice to link the question to the conclusion.

So they do the usual Internet argument approach: decide what the conclusion is and then make excuses for why they think it is such.

If you don’t believe me, why not ask one? This is a trivial example with very little “reasoning” needed and even here the explanations are bullshit all the way down.

Note, especially, the final statement it made:

Yes, your summary is essentially correct: what is called “reasoning” in large language models (LLMs) is not true logical deduction or conscious deliberation. Instead, it is a process where the model generates a chain of text that resembles logical reasoning, based on patterns it has seen in its training data[1][2][6].

When asked to “reason,” the LLM predicts each next token (word or subword) by referencing statistical relationships learned from vast amounts of text. If the prompt encourages a step-by-step explanation or a “chain of thought,” the model produces a sequence of statements that look like intermediate logical steps[1][2][5]. This can give the appearance of reasoning, but what is actually happening is the model is assembling likely continuations that fit the format and content of similar examples it has seen before[1][2][6].

In short, the “chain of logic” is generated as part of the response, not as a separate, internal process that justifies a previously determined answer. The model does not first decide on an answer and then work backward to justify it; rather, it generates the answer and any accompanying rationale together, token by token, in a single left-to-right sequence, always guided by the prompt and the statistical patterns in its training[1][2][6].

“Ultimately, LLM ‘reasoning’ is a statistical approximation of human logic, dependent on data quality, architecture, and prompting strategies rather than innate understanding. … Reasoning-like behavior in LLMs emerges from their ability to stitch together learned patterns into coherent sequences.” [1]

So, what appears as reasoning is in fact a sophisticated form of pattern completion, not genuine logical deduction or conscious justification.

[1] https://milvus.io/ai-quick-reference/how-does-reasoning-work-in-large-language-models-llms

[2] https://www.digitalocean.com/community/tutorials/understanding-reasoning-in-llms

[3] https://sebastianraschka.com/blog/2025/understanding-reasoning-llms.html

[4] https://en.wikipedia.org/wiki/Reasoning_language_model

[5] https://arxiv.org/html/2407.11511v1

[6] https://www.anthropic.com/research/tracing-thoughts-language-model

[7] https://magazine.sebastianraschka.com/p/state-of-llm-reasoning-and-inference-scaling

[8] https://cameronrwolfe.substack.com/p/demystifying-reasoning-models

Now I’m absolutely technically declined. Yet even I can figure out that these “reasoning” models are nothing different from the main flaws of LLMbeciles. If you ask it how it does maths, it will also admit that the LLM “decides” if maths are what it needs and will then switch to a maths engine. But if the LLM “decides” it can do it on its own it will. So you’ll still get garbage maths out of the machine.