> Despite the tremendous success, existing machine learning models still fall short of human-like systematic generalization

Keep in mind paper was published before ChatGPT was released

I think it still stands

Am I wrong in impression that this resembles MoE at some grade, and that implementation of these interoperable systems under one hut will cost significantly more then the MoE?

> NSR achieves 100% generalization accuracy on SCAN and PCFG

How generic is this? Seems like something specialized.