The argument “it just predicts the most likely next word” while true massively under values what it even means to predict the next word or token. Largely these predictions are based on sentences and ideas the model has trained on from its data sets. It’s pretty intelligent if you think about it. You read a text book then when you apply the knowledge or take a test you use what you read to form a new sentence in relation to the context of the question or problem. For the models “text prediction” to be correct it has to understand certain relationships between complex ideas and objects to some capacity. Yes it absolutely is not as good as human intelligence. But what it’s doing is much more advanced then text to type on your phone keyboard. It’s a step in the right direction, over hyped right now but the hype is funneling cash into research. The models are already getting more advanced. Right now half of what it says is hot garbage but it can be pretty accurate.