How ChatGPT works: a plain-language guide
Machine-assisted translation of the archived article. Read the Chinese original.

Let's see how the ChatGPT works. No matter where you ask it, it answers. In short, ChatGPT is a large-scale language model similar to GPT. In essence, it is calculating what should happen in the next word or sentence, which involves probabilistic issues. For example, when it says “I'm very good”, it can choose the next words: “happiness”, “health”, “emergency”, “hungry” and so on. However, context is needed to determine what follows. For example, the reference above to “good weather today” makes “I'm happy” the most likely choice in terms of probability. Each response and glossary is calculated on the basis of context-related relevance. By learning a lot of content, ChatGPT has hundreds of billions of parameters and words. When you find patterns through complex models, it creates a vast neural network. You don't have to tell it what programming is and what video scripts are because it learns. It knows what programming is about code, what video scripts should look like. So I made it write me a video script explaining ChatGPT. It finds a word from the relevance of the summary and then keeps it down.
ChatGPT is essentially a linguistic model that learns how others talk, but at least the current version of ChatGPT does not understand what it is saying. It's like a child with a special memory but no knowledge of many things, learning to talk to adults. But let's think it knows everything. That is also why its answer sounds perfect and very human, but it often makes some logical mistakes that make us feel a little stupid. That's because it's actually a language model. ChatGPT is more focused on whether or not the answer is real than the answer. In fact, GPT often invents answers, which means it doesn't know what it's talking about, but it's there to talk to you.
It also involves many ethical questions, such as asking how human beings are seen and it may tell you that human beings are inferior, selfish and worst. Is that really what it thinks? It clearly does not know what it is saying or where to learn from. However, these are all problems in the current version of ChatGPT. Although it may be a simple imitation now, does it really understand or is it pure imitation when the imitation becomes more and more human-like, with 99.9 per cent of the right answer? This is actually an issue that was discussed in the Turing test paper. Rather than asking whether machines can think like humans, it is even deeper to ask whether machines can do what humans can.
I think the major breakthrough of ChatGPT is that it greatly increases the efficiency of communication between people and machines. People communicate mainly by word, while computers communicate through code. Before that, humans had always moved to computers, and everything had to be programmed, then written into a language that the computer could understand, before the computer was implemented. We have also converted the problem into several keywords for search, including searches.
However, with language models, the situation has changed. Computers can now understand human beings, we can communicate with them directly in the human language, and then it translates and implements. People think ChatGPT is amazing. You ask it and it knows everything. However, its wonder is not that it can carry out these tasks, but that it can understand your problems very accurately and extract the most appropriate information from a vast database, to translate it into human language and to tell you. This link is the most amazing place. With such a powerful interface, we can leave much more easily to the machine and have greatly improved the efficiency of our work. Imagine, if we combine ChatGPT with voice recognition systems (e.g. Siri), so that they can engage in free dialogue. A number of specialized analytical interfaces were then added, such as stock analysis, programming, computing, robotics and the visual generation component. Well, then everyone can have an assistant like Iron Man in the movie. For example, you just said it would help you calculate what the Mobius ring is, and then it'll do. I think ChatGPT opened up a lot of possibilities at once.
