One well-supported theory for how humans learn language and knowledge is Relational Frame Theory (RFT)[0]. Essentially, when we learn that an elephant is larger

One well-supported theory for how humans learn language and knowledge is Relational Frame Theory (RFT)[0]. Essentially, when we learn that an elephant is larger than a mouse, we're learning a relationship ("more") and a dimension ("size"), which can be reversed ("smaller") or generalized (we can figure out what's larger than an elephant or what's smaller than a toaster). I think one theory is that the large number of parameters in LLMs allow them to go beyond mere association and actually model relational frames, which lets them generalize knowledge in an analogous way to humans. Maybe this behavior is not emergent in smaller networks (which could explain why animals with fewer neurons don't process language as well). Disclaimer: I'm not an academic researcher or scientist of any kind, this is just a hunch. [0]: https://en.wikipedia.org/wiki/Relational_frame_theory https://en.wikipedia.org/wiki/Relational_frame_theory