Analogies, like "Oxygen is to Gas as Aluminum is to Metal," are fundamental to human reasoning. They allow us to connect seemingly disparate concepts based on shared relationships. But can artificial intelligence truly grasp these subtle links? A new study explores this question by testing how well large language models (LLMs) can solve proportional analogies. Researchers crafted a massive dataset of 15,000 multiple-choice analogy questions and tested nine different LLMs using various prompting strategies. They found that even the best-performing models only achieved around 55% accuracy, highlighting the persistent difficulty LLMs face with this form of reasoning. Interestingly, simply stuffing the models with more general knowledge didn't improve performance. Instead, the most effective strategy involved "targeted knowledge prompting," where the models were given hints about the specific semantic relationships underlying the analogies, mimicking how humans might approach the problem. This implies that for LLMs to excel at analogical reasoning, they need more than just vast data; they require a deeper understanding of how concepts relate to each other. This study not only reveals the limitations of current AI but also points toward new avenues for developing models that can reason more like us. Future research might involve creating algorithms that can automatically extract relevant relationships for analogy solving, a critical step towards more human-like AI reasoning.