The issue with bigger LLM or even MMMs is that the bigger they are the more they are cramming and regurgitating the training data, and that opens up to lawsuits.

Making NNs generalize the way humans do it is still a hard problem.