A calculator uses some type of finite precision arithmetic internally. If you run afoul of the limits of this arithmetic system, it may very confidently give you a wrong answer!

Some calculators will confidently state incorrect answers to questions like:

(10^15 + 7.2 − 10^15) * 100

LLMs are notoriously bad at math but they’re not LMMs so that shouldn’t be surprising.

If you want an LLM to do math you just ask it to write a program with tests.

How about Pentium II floating point arithmetic?