Tokenization explains real LLM quirks
It's tempting to think of tokenization as a solved, invisible implementation detail once you're working with a modern LLM, but it directly explains behavior that otherwise looks like a strange failure. When people notice an LLM struggling to count the letters in a word, or making mistakes on arithmetic with multi-digit numbers, the root cause is usually tokenization, not a lack of intelligence. The model doesn't see individual characters or digits the way a human reading the text does, it sees whatever chunks the tokenizer produced. The word 'strawberry' might become two or three tokens, and a number like '48756' might be split into '487' and '56'. The model is reasoning over those chunks, not the raw characters underneath them.
Once you know this, these quirks stop looking mysterious. Asking an LLM to count letters is really asking it to reason about something that got obscured by the tokenization step before the model ever saw the text as characters. This is a direct, practical payoff of understanding tokenization: it turns 'the AI is being weirdly bad at this simple thing' into a specific, explainable mechanism, and it explains why techniques like spelling a word out with spaces or hyphens between each letter often help, that forces each letter into its own token.