Machines Will Not Rise: Limitations and Obstacles of Neural Networks

[sadiksojib35]

Do you know the answer? You will find the correct version at the end of the article. In the meantime, let's understand why the neural network gave us such nonsense:



The neural network is trained to continue the text. To do this, it takes the entire original text, runs it through itself, each time receiving exactly one next word (not really a word, but a token - a symbol or several symbols). Then it takes the entire text again, this time with the continuation, and repeats the process. And so on until it receives a token indicating the end of the message, or hits a limit.

The most important nuance: the neural network does india telegram database not try to "think" unless asked to do so. It writes the most probable continuation. If we ask it to write an answer and an explanation, it will generate something as similar as possible to them. But if we asked it to think, it would write an answer similar to thinking. Each word from these thoughts would pass through the neural network again and again, and this could lead to a truly correct answer.

For a long time I used this riddle as a demonstration of the fact that the neural network cannot answer some questions. But then I realized that it was I who did not know how to write instructions and prompts. How to make the neural network give the correct answer, we will learn at the end of the article, but for now let's talk about limitations.



Over the course of this year, neural networks have penetrated all spheres of human activity. But there are a few "buts". Language models have limits of applicability.