Yes and No. I don't entirely disagree with you, but think about when you ask a model to explain step by step a conclusion. It is not doing the reasoning, but in a way abstracted and learned the pattern of doing the reasoning....So it is doing some type of reasoning....and sometimes producing the outcomes that are derived from actual reasoning...Even if defining "actual reasoning" is a whole new challenge.
It took a long time for the limitations of LLMs to "click" for me in my brain.
Let's say there's a student reading 10 books on some topic. They notice that 9 of the books say "A is the answer" and just 1 book says "B is the answer". From that, the student will conclude and memorise that 90% of authors agree on A and that B is the 10% minority opinion.
If you train an LLM on the same data set, then the LLM will learn the same statistical distribution but won't be able to articulate it. In other words, if you start off with a generic intro blurb paragraph, it'll be able to complete it with the answer "A" 90% of the time and the answer "B" 10% of the time. What it won't be able to tell you is what the ratio is between A or B, and it won't "know" that B is the minority opinion.
Of course, if it reads a "meta review" text during training that talks about A-versus-B and the ratios between them, it'll learn that, but it can't itself arrive at this conclusion from simply having read the original sources!
THIS more than anything seems to be the limit of LLM intelligence: they're always one level behind humans when trained on the same inputs. They can learn only to reproduce the level of abstraction given to them, they can't infer the next level from the inputs.
I strongly suspect that this is solvable, but the trillion-dollar question is how? Certainly, vanilla GPT-syle networks cannot do this, something fundamentally new would be required at the training stage. Maybe there needs to be multiple passes over the input data, with secondary passes somehow "meta-training" the model. (If I knew the answer, I'd be rich!)
In principle, yes, but empirically? They can't do this reliably, even if all the texts fit within the context window. (They can't even reliably answer the question "what does author X say about Y?" – which, I agree, they should be able to do in principle.)