Can AI make sense of scientific texts even if they're read backward? A fascinating new study explores this question, training large language models (LLMs) on both forward and backward neuroscience literature. Surprisingly, the backward-trained models performed almost as well as their forward-trained counterparts on a challenging neuroscience benchmark, even surpassing human expert accuracy in some cases. This raises intriguing questions about how LLMs learn and process information. While humans rely on the inherent structure and flow of language, these AI models seem capable of extracting predictive patterns regardless of the text's order. The backward models did exhibit higher perplexity, suggesting they found the reversed text more challenging to process, analogous to how humans struggle with backward speech. This research highlights that LLMs, while incredibly powerful, don't learn like humans. Their strength lies in identifying patterns, even in data that violates human cognitive constraints. This makes them versatile tools for diverse applications, but it also suggests caution when interpreting their success as mirroring human understanding.