The discussion around the LIMO framework highlights the potential of large language models (LLMs) to perform mathematical reasoning more efficiently. The premise is that these models, having been pretrained on extensive internet data, inherently possess a range of skills related to reasoning. However, the challenge lies in eliciting this knowledge effectively. Comments suggest that with minor adjustments and a few training examples, LLMs can enhance their reasoning capabilities, similar to how adding specific prompts can improve generative models' output quality. This raises questions about the nature of reasoning patterns—whether they can be summarized effectively within a compact model and categorized similarly to knowledge patterns in image generation. There is an emerging interest in the 'pedagogy of LLMs,' suggesting a shift towards understanding how to train and evaluate these models for optimal reasoning performance. Overall, this reflects a trend towards recognizing the latent skills within LLMs and exploring how to unlock and utilize them efficiently, alongside debates about the inherent limitations of model generalization, particularly in theorem proving.