Large language models (LLMs) like ChatGPT can write poems and code software, yet they often stumble on seemingly simple tasks like counting the letters in a word. Why is such a basic skill so challenging for AI? A new study dives into this puzzle, examining why LLMs struggle with something a child can easily do. Researchers tested a range of LLMs, from open-source models like LLaMA to proprietary giants like GPT-4, giving them a letter-counting challenge across thousands of words. Surprisingly, word frequency in the training data had little impact on accuracy. Instead, the complexity of the counting task itself emerged as the key factor. LLMs generally excelled at identifying letters *within* a word, but faltered when a letter appeared multiple times. For instance, counting the 'r's in 'strawberry' proved tricky. The findings suggest the challenge isn't about recognizing letters, but the actual *computation* of counting, particularly when a letter appears more than once within the same word or across different tokens. Interestingly, tokenization, often blamed for these counting errors, wasn't the primary culprit. This research highlights a fundamental difference between how humans and LLMs process language. While we learn letters as building blocks, LLMs often work with larger chunks of text (tokens), potentially hindering their grasp of basic letter counts. Further research is needed to fully understand these limitations and bridge the gap between AI’s impressive language abilities and its struggles with fundamental counting skills, ultimately paving the way for more robust and reliable AI systems.