[synapsomorphy]

When a model reads "dog" it associates the patterns it gleaned from the text and images about dogs - its past 'experiences'. What is the difference in kind between that and animal understanding?

LLMs do have language-agnostic understandings in their latent space. "Dog" and "Perro" have largely the same representation (depending on the model. I believe more advanced ones show this more strongly?) as does a picture of a dog. I'm not sure if that's exactly the form of understanding you're referring to?

I agree the human text/images corpus is very small compared to evolution's millions of years of learnings from interacting with the environment, which is why I'm excited for RLing LLMs because it opens up the same data trove.