Why LLM’s can’t count the R’s in the word “Strawberry”

Understanding Why Large Language Models Struggle with Counting Letters: The Case of “Strawberry”

In the AI community, it’s common to see discussions and jokes about large language models (LLMs) like GPT-3 or GPT-4 failing at simple tasks—such as counting the number of R’s in the word “Strawberry.” But what underlies this limitation? Why do these advanced models often stumble over such straightforward queries?

At their core, LLMs process text by first dividing the input into manageable segments known as “tokens.” These tokens can be words, parts of words, or even individual characters, depending on the model’s tokenizer. Once tokenized, each piece is transformed into a numerical representation called a “vector.” These vectors serve as the model’s internal language comprehension tool, feeding into subsequent layers that generate predictions or responses.

The crucial point is that LLMs are not inherently designed to perform character-by-character counting or precise letter recognition. Since their training is based on vast datasets of text, their internal representations focus on capturing semantic and syntactic patterns rather than exact letter positioning. As a result, the vector summaries do not preserve granular details about individual characters, making tasks that require counting or specific character recognition inherently challenging.

This explains why asking an LLM to count the number of R’s in “Strawberry” often results in inaccuracies—even though the task seems simple to humans. It highlights an important distinction: while these models excel at understanding context and generating coherent language, they are not reliable at tasks requiring meticulous character-level operations.

If you’re interested in a visual explanation of this concept, you can explore this detailed diagram here: https://www.monarchwadia.com/pages/WhyLlmsCantCountLetters.html. (Please note that image sharing is restricted on some platforms, so visiting the link is recommended for a deeper understanding.)