Shulou(Shulou.com)11/24 Report--

This article comes from the official account of Wechat: SF Chinese (ID:kexuejiaodian), author: SF

Not long ago, OpenAI, a US artificial intelligence research and development company, released the latest chat robot ChatGPT, which has almost human language skills, heralding another revolution in the field of artificial intelligence.

/ photos by Kate Darling (Kate Darling)

The original article is published in Science focus 2023.03.

I remember the first time I saw my son interact with a large language learning model. At that time, he was only five years old, but he was already able to have a natural and smooth conversation with artificial intelligence, as if artificial intelligence were a real human being. Watching him talk to artificial intelligence, I have mixed feelings in my heart. The scene was vivid at that time, which not only reflected the achievements of human beings in the field of artificial intelligence, but also made me realize that the language model had unlimited potential: it was revolutionizing the mode of interaction between humans and machines.

In fact, the whole passage above was written by AI, and its level is about the same as that of my own handwriting. If you've ever been disappointed with customer service robots, you might wonder why all of a sudden we have artificial intelligence that can understand customer needs (such as writing the beginning of this article) and give appropriate feedback.

The ability to understand is shocking. Today, even the most advanced commercial chatbots still use predetermined answers many times. If you ask Alexa, Amazon's intelligent voice assistant, what's his favorite drink, his answer is likely to be a prefabricated answer from one of Amazon's programmers.

Different from the above principle, the ChatGPT I use is based on the generative pre-training transformation model (Generative Pre-Trained Transfer model, referred to as GPT). This model can generate its own answers in the conversation. For example, instead of saying exactly what kind of drink you like best, it may recommend that you try seabuckthorn juice.

ChatGPT, released by artificial intelligence research and development company OpenAI in November 2022, is a generative conversational AI that is subverting the rules of the game, similar to language models developed by companies such as Google and Facebook.

ChatGPT's language learning model has received billions of text training on the Internet. Based on such a large database, ChatGPT can predict exactly which word to use when outputting text. Although this is not a new algorithm, it also includes a "Transformer": by analyzing the relationship between complete sentences and individual sentences, ChatGPT can understand the context of the conversation.

Before that, chatbots often did not understand our questions or gave a logical but meaningless answer because they could not understand the context. We have long been accustomed to this.

When my husband asked ChatGPT to write an oath to propose to me in sarcastic style, ChatGPT wrote: "ruthless robotic scientist Kate Darling, although making an emotional connection is bound to be futile, please marry this uncontrollable human suitor." It is safe to say that this sentence was definitely not written in advance by artificial intelligence experts at OpenAI. ChatGPT's ability to understand the task is simply shocking.

Another breakthrough in many uncertain transformation modules is that it significantly reduces the time required to build a model. Today's technology companies have vast amounts of training data, greater numeracy than ever before, and the ability to build and train language models in less time. These conditions are ushering in a new era of conversational artificial intelligence.

Of course, the new language model also has its shortcomings. If it is used as a commercial chat robot, the autonomy of the language model cannot be too strong, at least not yet. ChatGPT is eloquent, can write poems, and can write sarcastic emails for you, but sometimes it can say something very inappropriate or even offensive.

Obviously, magical ability coexists with risk. OpenAI is really trying to adjust the functionality of ChatGPT. For example, let humans train artificial intelligence and provide feedback on its dialogue skills. Not only that, ChatGPT also contains some pre-written answers and jokes. But even so, it is not realistic to predict what ChatGPT might say in each scenario. This makes ChatGPT uncertain in many practical applications and may also cause a lot of ethical problems.

As ChatGPT said at the beginning, we have indeed made great strides in artificial intelligence, and these technological breakthroughs are likely to mean that "the language model has unlimited potential: it is revolutionizing the way humans interact with machines." However, on the road of moving forward hand in hand, human beings and technology should continue to communicate.

Introduction to the author

Kate Darling (Kate Darling) is a researcher in the Media Lab of the Massachusetts Institute of Technology, focusing on machine ethics and human-robot interaction to provide reference for future legal and policy-making.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.