I recently wondered; how can one teach AI to answer mathematical questions?
The problem is that large language models (LLM) learn solutions by heart. So, if you ask them what 1+2 is, they will answer 3, because this pops up everywhere.
However, if you ask what 1,559,993,117 + 33,492,224 is, they will generally not find the right answer.
How does one solve this? Well, AI found the solution itself!
ChatGPT is using code to compute answers. That is the right alternative, instead of learning mathematical solutions by heart. You (or AI) need to understand the fundamental logic of the problem.