Can large language models (LLMs) truly grasp mathematical concepts, or are they just excellent statistical parrots? A new research paper, "Can Large Language Models put 2 and 2 together? Probing for Entailed Arithmetical Relationships," delves into this question, exploring whether LLMs can reason about numbers implied by everyday objects. The researchers investigated if LLMs could deduce, for example, that a bird has fewer legs than a tricycle has wheels, without explicitly stating the number of legs or wheels. The results reveal that while LLMs are getting better at factual recall with each iteration, their mathematical reasoning remains limited. They often stumble when dealing with the concept of zero, likely due to the scarcity of negative statements in training data (we don't often say things like "a bird has zero wheels"). Furthermore, the study highlights a fascinating bias: LLMs tend to hallucinate numbers related to the core characteristics of an object. For instance, when asked about a bicycle, they might incorrectly associate it with the number two (because of two wheels), even when querying about unrelated features. This suggests that LLMs rely heavily on statistical correlations rather than true understanding. While LLMs can sometimes produce correct answers, this research reinforces the idea that they are not genuinely reasoning. They excel at pattern matching and statistical inference, but struggle with the kind of symbolic manipulation required for true mathematical reasoning. This raises important questions about the limitations of current AI approaches and the need for new methods that incorporate symbolic reasoning alongside statistical learning. The future of AI likely lies in hybrid models that combine the strengths of both approaches, enabling machines to not just compute, but truly understand.