Lefty

People often assume that ChatGPT must understand words in the way humans do, and that real AI will emerge as these models get better.

However, to the model, words are just mathematical tokens represented as vectors. The relationships between those vectors allow it to predict what token is most likely to come next, but there is no understanding attached to them.

It would be a simple exercise to train a model to generate text from the end of a reply backwards, from right to left instead of left to right. It would work just as well. The direction of text generation is a design choice, not a property of meaning.