AI: the illusion of understanding.
When we ask AI to make something, it takes the list of allowed actions. We know that AI requires limits. Things like AI-created photos and other things like that should be marked using the stamp. The AI can give texts that it creates to plagiarism detection software that universities and high schools should use. That thing makes it possible to detect AI-created texts. In that case, the plagiarism detector just compiles those texts together. And that uncovers the plagiarism and AI-created texts with quite good accuracy.
Then the AI makes the asked action by connecting components that it takes from the net. This is how the AI generates images. When the AI makes the text like a letter or scientific article it searches a couple of sources from the net. After that, it just connects those sources. If the thing that the user asks is common, the AI gives good-looking answers. The AI uses search engine solutions for selecting sources. And that makes it effective in common things. But in the cases that the asked thing is uncommon like "Ramsay's numbers in mathematics," the AI can give an answer that handles the noble family's place in some club.
Image created by Bing.Then the AI suggests that user can search for their telephone numbers by using some net-based telephone catalog. This happens if the user forgets the word "mathematics". We can call this thing "Ramsay's problem". The user must give clear and well-cropped commands to the AI. If the user will not crop the commands anyway. The AI can give a very long answer that it generates for a very long time. And then that answer could be wrong. This means the user who uses the AI requires practicing.
This is the thing that shows that the AI doesn't understand what it makes. The situation is similar to the case, where the 12 years old kid makes the scientific article by collecting good-looking sources. Then that kid connects paragraphs from those sources together by using copy-paste.
The kid can take the first paragraph from the first source, the second paragraph from the second source, etc..And that thing makes it possible to create a good-looking article. In that case, the kid can have a trusted sources list. But the AI and that kid have the same problem: the source can involve errors.
The thing that helps to eliminate those "Ramsay problems" is that the AI makes the user profile. The AI could ask what the user does with it. And if the user says being a mathematician. That helps the AI to crop the homepages to mathematical homepages, and then the AI can ask when the person starts to use it, is the session for work or fun.
The fact is that the AI requires precise and good language so that it can create answers. Even if the AI is an excellent tool for work and fun, the thing is that people who use the AI should know how to use it. They also must have the ability to estimate sources that the AI is useful. But if the user doesn't know how to ask questions or give commands. That thing makes the AI useless.
https://scitechdaily.com/the-illusion-of-understanding-mit-unmasks-the-myth-of-ais-formal-specifications/
Comments
Post a Comment