Why Generative AI Struggles with Math (and How to Fix It)

Might seem obvious to some, but after trying for half an hour for llama.cpp/gemma3-1b to correctly tell me how many 11mm tall sugar cubes I should stack to get to the moon (the model insisted that 11 mm == 0.011 km for its calculations), this might shed a light on why this seemingly easy task is not for LLMs.