You got this wrong, human brain does this because it doesn't have a long active memory.

But I can give you pen and paper, and you can use the actual algorithm to do the calculations instead of relying on multiple heuristics.

Human brain uses heuristics because it can't calculate long numbers. But LLMs, don't know the alogrithm itself. That is what the paper is about.

If I teach you newtons laws of motion, you can then apply them to any sized object, but what LLMs, does is that it might be able to apply to a few small sized object, but might mess up with other objects.