“Large language models are inherently not good at math because they are probabilistic technologies.”