27/09/2024
This is how large language models generate their content.
Predicting the next word.
It’s not magic; there’s no real thinking. It is all about predicting language patterns.
Suppose you ask a GPT to describe the reasoning for counting “r”s in a sentence. The model will just spit out the most “likely” sentences to occur, given your request and the trillions of text it has “seen” (trained on).
That’s all.
The reasoning seems coherent because the AI learned how to construct sentences that "most likely seem like" a good answer to the question.
It works well for most questions because the models are so big that they're very good at predicting "the likely sequence of words" that answer your question.
If AI truly "thought" independently—from language (the most likely sentences), it would just give you, or use, a simple Python script to count the “r”s—no need for complex reasoning.
Understanding how GPTs work is very important to use AI to your advantage. You will know what AI can and can not do, when to use AI, and how/ when not to use AI.