I recall that asking gpt3 to write python code to compute the answer instead of trying to come up with the answer itself worked a ton better.

You do have to execute the python yourself somehow still, but even with just a super simple interpreter you can make your own math-enhanced LLM ai assistant