It's interesting how the formulas for the conversion are correct, but the result keeps being missed.
It’s just a text completion model. It doesn’t actually evaluate computations, despite it being common to find it giving correct answers for some computation queries
That's true. It does not acknowledge this limitation when communicating with the user, even when I prompted the model with suggestions.