Why large AI models are often described as black boxes
Machine-assisted translation of the archived article. Read the Chinese original.

The AI megamodel itself is a black box, and even if the developers are not sure what the machine's overall decision-making process is, the answer to this question already existed in the pregpt era, as in 2015, when Google's image recognition software directly identified our brothers in Africa as gorillas, scared by Google to block the chimpanzee key word, which was never able to recognize gorillas anymore, but it wasn't the way to do that in gpt, the problem would only become more complex, both to hear imaginative answers and to look for an ai closer to our human mind. And so to deal with these problems, all the big language models have an ongoing process of alignment of values, and we hope that all the issues he's talking about are in line with our dominant human values, he's neutral and objective, he's positive, he's hot, he can help people solve problems, he can comfort people.
The 100 bottles of poison given to ai, and the 100 bottles of poison, are actually 100 poisons, which means that experts in each field have targeted 100 issues of discrimination with obvious or hidden prejudice to feed to ai, induce ai to make mistakes, thereby strangling toxic answers into the cradle, why do big models use this method of confrontation? Let ai learn in a very large magnetic vector space to do a hyphenation game, and in short, the user's question is to enter words, and they look for the most probabilities of words, that is, words that combine the nearest distance in the vector space, with sentences to respond back to users!
For a large-language model, he only cares about one sentence, whether the grammar is correct or not, and whether it's right in the sense that it's his original purpose, because, of course, the human language itself contains a great amount of knowledge and information, so this training has produced a wonderful by-product of whatchatgpt learned a lot of knowledge, which, after all, in a very large language library, ideally, the correct answer is often the most probabilities, so gbt does not have a database to store all kinds of knowledge, but it simply combines the sentences with the highest probability.
