There was an interesting debate where Chomsky took a position on intelligence being rooted in symbolic reasoning and Asimov asserted a statistical foundation (ah, that was not intentional ;).
LLM designs to date are purely statistical models. A pile, a morass of floating point numbers and their weighted relationships, along with the software and hardware that animates them and the user input and output that makes them valuable to us. An index of the data fed into them, different from a Lucene or SQL DB index made from compsci algorithms & data structure primitives. Recognizable to Azimov's definition.
And these LLMs feature no symbolic reasoning whatsoever within their computational substrate. What they do feature is a simple recursive model: Given the input so far, what is the next token? And they are thus enabled after training on huge amounts of input material. No inherent reasoning capabilities, no primordial ability to apply logic, or even infer basic axioms of logic, reasoning, thought. And therefore unrecognizable to Chomsky's definition.
So our LLMs are a mere parlor trick. A one-trick pony. But the trick they do is oh-so vastly complicated, and very appealing to us, of practical application and real value. It harkens back to the question: What is the nature of intelligence? And how to define it?
And I say this while thinking of the marked contrast of apparent intelligence between an LLM and say a 2-year age child.
That's not true, symbols emerge out of the statistics. Just look at the imagenet analysis that identified distinct concepts in different layers, or the experiments with ablation in LLMs.
They may not be doing strict formal logic, but they are definitely compressing information into, and operating using, symbols.
Which argues much symbolic manipulation is formulaic, and not inductive reasoning or indicative of intelligence.
Sentence parsing with multiple potential verb-noun-adjective interpretations are an example of old, Chomsky made fruit flies like a banana famous for a reason.
(without the weights and that specific sentence programmed in, I would be interested exactly how the symbol models cope with that, and the myriad other examples)
To me the interesting idea is the followup question: Can you do complex reasoning without intelligence?
LLM's seem to have proven themselves to be more than a one-trick-pony. There is actually some resemblance of reasoning and structuring etc.. No matter if directly within the LLM, or supported by computer code. E.g it can be argued that the latest LLMs like Gemini 2.5 and Claude 4 in fact do complex reasoning.
We have always taken for granted you need intelligence for that, but what if you don't? It would greatly change our view on intelligence and take away one of the main factors that we test for in e.g. animals to define their "intelligence".
> E.g it can be argued that the latest LLMs like Gemini 2.5 and Claude 4 in fact do complex reasoning.
They most definitely don't. We attach symbolic meaning to their output because we can map it semantically to the input we gave it. Which is why people are often caught by surprise when these mappings break down.
LLMs can emulate reasoning, but the failure modes show that they don't. We can get them to be coincidentally emulating reasoning well enough long enough to fools us, investors and the media. But doubling down on it hoping that this problem goes away with scale or fine tuning is proving more and more reckless.
Humans aren't infallible and make mistakes in reasoning as well. What is fundamentally different about the mistakes we make versus the mistakes that Claude or Gemini make? Haven't LLM's even been shown to make the same posthoc rationalizations of mistakes that we as humans do all the time?
Unless you're pulling humans out of the streets at random and asking them questions or to do work, I guess you also shouldn't do that with statistical models of random human language.
I think we are ignoring that the statistical aspect of our ability to reason effectively and to apply logic was predicated on the deaths of millions of our ancestors. When they made the wrong decision, they likely didn't reproduce. When they made the right decision, that particular configuration of their cortical substrate was carried forward a generation. The product of this cross-generational training could have easily led to non-intelligence, and often does, but we have survivor's bias in our favor.
> Chomsky took a position on intelligence being rooted in symbolic reasoning and Asimov asserted a statistical foundation
I dunno if people knew it at that time, but those two views are completely equivalent.
Perhaps the next question we are asking is "what happens if you give a statistical model symbolic input" and the answer appears to be, you get symbolic output.
Even more strangely, the act of giving a statistical model symbolic input allows it to build a context which then shapes the symbolic output in a way that depends on some level of "understanding" instructions.
We "train" this model on raw symbolic data and it extracts the inherent semantic structure without any human ever embedding in the code anything resembling letters, words, or the like. It's as if Chomsky's elusive universal language is semantic structure itself.
>There was an interesting debate where Chomsky took a position on intelligence being rooted in symbolic reasoning and Asimov asserted a statistical foundation (ah, that was not intentional ;).
Chomsky vs Norvig
> and very appealing to us
Yes, because anthropomorphism is hardwired into our biology. Just two dots and an arc triggers a happy feeling in all humans. :)
> of practical application and real value
That is debatable. So far no groundbreaking useful applications have been found for LLMs. We want to believe, because they make us feel happy. But the results aren't there.