Claude and other formidable AI products tell us that the word “strawberry” contains the letter “r” twice.
Writing essays and solving equations is a task that large language models (LLMs) can do quickly. More quickly than a human can open a book, they can synthesize terabytes of data.
Nevertheless, there are instances when these ostensibly intelligent AIs make such glaring mistakes that the incident becomes a widely shared meme, and we all applaud, relieved that perhaps we still have some time before we have to submit to our new AI rulers.
The fact that large language models are unable to comprehend the notions of letters and syllables points to a more important reality that we frequently overlook: these objects are not human brains. They don’t think the same way as us. They are not human, nor are they even very similar to humans.
Transformers are a type of deep learning architecture that forms the foundation of most LLMs. Text is divided into tokens by transformer models, which may or may not be entire words, syllables, or letters.
“LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding. When it sees the word ‘the,’ it has this one encoding of what ‘the’ means, but it does not know about ‘T,’ ‘H,’ ‘E,” Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta said, TechCrunch reported.
This is a result of the transformers’ inefficiency in processing and producing actual text. Rather, the text is transformed into numerical representations of itself, which are then contextualized to assist the AI in deriving a sensible answer. Stated differently, the AI may recognize that the tokens “straw” and “berry” combine to form the word “strawberry,” but it might not comprehend that this particular combination of letters is made up of the letters “s,” “t,” “r,” “a,” “w,” “b,” “e,” “r,” and “y,” in that particular order. It cannot, therefore, determine the number of letters in the word “strawberry,” much less the number of “r”s.
Since it’s ingrained in the architecture that underpins these LLMs, fixing it is a difficult task.
