In order to successfully predict the next token the model needs to reach a significant level of "understanding" of the preceding context and the next token is the "seed" of a much longer planned response.
Now, it's true that this "understanding" is not even close to what humans would call understanding (hence the quotes) and that the model behaviour is heavily biased towards productions that "sound convincing" or "sound human".
Nevertheless LLMs perform an astounding amount of computation in order to produce that next token and that computation happens in a high dimensional space that captures a lot of "features" of the world derived from an unfathomably large and diverse training set. And there is still room for improvement in collecting. cleaning and/or synthesizing an even better training corpus for LLMs.
Whether the current architecture of LLMs will ever be able to truly model the world is an open question but I don't think the question can be resolved just by pointing out that all the model does is produce the next token. That's just an effective way researchers found to build a channel with the external world (humans and the training set) and transform to and from the high-dimensional reasoning space and legible text.
Any "understanding" is a mirage.
Case in point. prompt: "Is it possible to combine statistical mechanics, techno music production and pizza?"
If the model had even the slightest understanding of the world the answer would just be no. Instead gpt4o:
"Yes, it is possible to combine statistical mechanics, techno music production, and pizza, though it might require some creative thinking. Here’s how these three seemingly unrelated things could be connected" then gives a list of complete nonsense.
The trick is that it can't say no because it doesn't understand ANYTHING.
It has no understanding of the difference between combining pizza with two completely disparate nonsense subjects/items with combining pizza with two other food items. The later would seem to have a mirage, high dimensional, understanding from data of "food" though in the response.
You just did.
For example one of the outputs:
"Host an event where statistical mechanics concepts are explained or demonstrated while making pizzas, all set to a backdrop of live techno music. The music could be dynamically generated based on real-time data from the pizza-making process, perhaps using sensors to monitor heat, time, or the distribution of toppings, with this data influencing the techno tracks played."
It not doing such a bad job trying to mix up three unrelated concepts. It knows music is not an ingredient for the pizza and knows that pizza requires heat for cooking and that heat is explained with statistical mechanics.
Sure you can nitpick and find nuances that are wrong but honestly an average human asked to come up with something for a school assignment would probably not do a much better job.
Now, there are clearly better examples of utter failures where even the best model trip on that reveals that they are not even close at understanding and modeling the world correctly.
My point is just that their weakness cannot merely explained by the next token prediction process.
Oh, it's actually worse than that: A given LLM probably has zero concept of "entities", let alone "you are one entity and I am another" or "statements can be truths or lies."
There is merely one combined token stream, and dream-predicting the next tokens. While that output prediction often resembles a conversation we expect between two entities with boundaries, that says more about effective mimicry than about its internal operation.