If anything their capability of abstract inductive reasoning is way beyond the average human given how much better LLMs are at solving math problems, it's the paradox that they can do complicated reasoning before they can do intuitive peep-pe-boo.
Words are also not necessary for LLMs, hear me out: they could be trained mostly on video showing how to solve problems and would work almost the same (in principle). That's because both word tokens and video will end up in the same latent space which is where real consciousness and thinking takes place.
Because these models are now also trained on visual data, so they have a common abstract language in the latent space for different kinds of modalities. It's perfectly reasonable to ask the model if it can associate an image of a server with its own existence. In fact it once saw an open process and said "that's me"
If it turns out to be true, it would open the door a bit for connecting Indo-European languages with Semitic languages. In the beginning of the last century it was believed that these were related. Later this came out of vogue. How could they have been so wrong initially? Because both languages families were entangled, as now there is genetic evidence that both languages spread from very close to the Caucasus. It's probably old news for most but in the last 15 years it became clear that Europe was completely resettled, once by Anatolians and then partly by Indo-Europeans. The language of the Anatolians is still unknown.
Most importantly, the reinforcement loop is used during training. I don't agree with Sutton's original hypothesis, but it holds even less after reinforcement learning.