Large language models are ‘bad at maths’ and can only predict answers based on training data, according to Zhejiang University researcher Wu Yiquan.