Or just … put a calculator in the middle of the LLM and let it learn how to use it. Why are you jumping through so many hoops to help a language model learn an incredibly inefficient way to inaccurately do math in its head? The whole point of machines is to do things in ways that are better than what we do in our heads.

I bet they’ll pull a https://en.wikipedia.org/wiki/Embrace,_extend,_and_extinguish

this is sooo complicated

It is, objectively, significantly more complicated, though.