This exercise seems like highlighting the need for a cheap (under use), all-purpose framework for an efficient function approximator (as bare ML is too costly).

This exercise seems like highlighting the need for a cheap (under use), all-purpose framework for an efficient function approximator (as bare ML is too costly). They are trying LLMs for this purpose, but maybe the structure of an optimal architecture should be studied.