Can you read this sentence if the middle letters of some words are scrambled? You probably can, thanks to a phenomenon called Typoglycemia. Humans rely on context, word shapes, and letter patterns to decipher jumbled text. But what about AI? A fascinating new research area, dubbed "LLM Psychology," explores this very question. Researchers are using Typoglycemia as a tool to peek into the "minds" of large language models like GPT-4. By scrambling text at different levels – letters, words, and even sentences – they're testing the limits of AI comprehension. The results are intriguing. While LLMs do surprisingly well at understanding jumbled text, showing a degree of resilience like humans, their success stems from statistical pattern recognition rather than genuine comprehension. Interestingly, just like people have unique cognitive fingerprints, these experiments reveal that each LLM has its own distinct "cognitive pattern." This innovative approach, dubbed "TypoBench," uses existing datasets, making it easier for researchers to compare the performance of different models in these scrambled text scenarios. This opens up a new way of evaluating AI models and pushes us closer to truly understanding how these powerful tools process language. The findings also raise interesting questions about the gap between apparent human-like intelligence and the underlying statistical mechanics at play within AI.