ChatGPT sets off a technological frenzy
Under the "top stream", look at the joys and sorrows of artificial intelligence
Produced by Shentong Studio
Written by: Intern reporter Zhang Jiaxin, reporter Liu Yuanyuan and Chen Xi
Curator: He Yifang Linlin Shao Deqi
Currently, ChatGPT has not passed the evaluation of the Turing test.
In fact, so far, no artificial intelligence model can really pass the Turing test.
In November last year, OpenAI, an American artificial intelligence research company, released a chatbot called ChatGPT, which took the world by storm within weeks of its launch and even sparked a new global AI competition.
Elon Musk, CEO of social media Twitter, called ChatGPT "too good to be scary" after using it, and even asserted that "artificial intelligence is getting stronger and stronger, and it is not far from threatening humans."
Aaron Benanav, author of the book "Automation and the Future of Work" and a professor at Syracuse University in the United States, said that 200 years ago, people had waves of fear of automation technology. Fear that they will completely replace human labor.
At the beginning of February this year, ChatGPT launched the "hurricane" mode, which caused a new round of crazy technological upgrades, industrial restructuring, capital frenzy, and giant competition.
At first, the general public, who felt ups and downs as this wave of technological heatwave swept the world, went through an "immersive" experience. Although they did not feel fear, they still felt a little worried.
The First Steps to Realizing Conversational AI
ChatGPT has only been released for 2 months, and its monthly active users have exceeded 100 million.
You know, "100 million" is not a small goal. It took Facebook 4.5 years, WhatsApp 3.5 years, Instagram 2.5 years, and Google nearly a year to reach this figure.
Zhang Shuwu, a researcher at the Institute of Automation of the Chinese Academy of Sciences and director of the Digital Content Technology and Service Research Center, said in an interview with a reporter from Science and Technology Daily that people can regard ChatGPT as the most complex language recognition technology breakthrough after voice, image and video recognition technology, and it is also based on patterns. A new leap in the development of perceptual intelligence technology for recognition and deep learning.
Professor Xiong Deyi, head of Tianjin University's Natural Language Processing Laboratory, said that from the perspective of artificial intelligence development, this is an important milestone, and its significance is higher than that of "AlphaGo" seven years ago.
He told reporters that the "Alpha Dog" in the Go world has refreshed the general public's views on artificial intelligence technology, but it itself cannot be widely used, and it is difficult for the general public to directly contact it.
ChatGPT is completely different.
Compared with "Alpha Dog", ChatGPT can directly apply or support many scene applications and products.
If "Alpha Dog" is the "Yangchun Baixue" of technology application, then ChatGPT may have a chance to become "Xiliba People".
Indeed, after users have made unremitting efforts to explore the uses of ChatGPT, people have discovered that it is not only a chat tool, but also can write codes, love letters, papers, online consultations, scenario scripts, poems, novels, and even It is to draw up travel routes and answer brain teasers. It can be said to be "responsive to requests", and its functions are so powerful that it is unimaginable.
ChatGPT is rapidly penetrating various industries, and the general public can become users of ChatGPT, which is an important reason for its popularity.
In fact, there have been many related applications for chatting with robots before, and the one we are most familiar with is the commonly used intelligent voice assistant.
How is ChatGPT different from it?
"Smart speakers also use natural language processing technology and conversational AI technology, but language intelligence is relatively rudimentary. In many cases, we can only have simple conversations with them." Xiong Deyi said, "ChatGPT, Its underlying technology is not only completely different from the underlying technology of general smart speakers, but also much more powerful than them. Not only that, it improves the human feedback data collection form of its similar technology InstructGPT. In conversations with people, it can be based on history. content, supporting multiple rounds of dialogue in a row.”
Xiong Deyi believes that ChatGPT is the first step to realize conversational AI.
From the perspective of the Internet, it can be used as an information portal and combined with search engines; it can also be combined with office software to enhance office efficiency; from the perspective of human-computer interaction, it embodies an interactive method based on natural language, which can Combined with metaverse and digital people.
In addition, it can also act as a robot brain.
"If you look at the scale of human information development, it has made a great breakthrough in the information interface between humans and machines." Xiong Deyi said.
Proficient in all "Eighteen Martial Arts"
ChatGPT uses GPT-3 technology, the third generation of generative pre-trained Transformer (Generative Pretrained Transformer 3).
This is an autoregressive language model that uses deep learning to generate human-like text.
According to the British Broadcasting Corporation's "Science Focus" magazine, GPT-3's technology may seem simple, for example, it accepts user requests, listens to questions or prompts and answers them quickly, but in fact this technology is much more complicated to implement.
The model is trained using text databases from the Internet, including up to 570GB of data obtained from books, web texts, Wikipedia, articles, and other texts on the Internet.
More precisely, it was intensively trained on training data consisting of 500 billion words.
GPT-3.5 is a fine-tuned and optimized version of GPT-3, which is more powerful than the latter.
ChatGPT is supported by the large language model (LLM) of the GPT-3.5 architecture.
ChatGPT can also be described as an artificial intelligence model in the field of natural language processing (NLP).
This means it is a program capable of encoding, modeling and generating human language.
As a language model, after self-supervised pre-training on large-scale text data, ChatGPT needs to go through three stages of "refining".
The first is the supervised fine-tuning stage. In this stage, the human AI trainer acts as both the user and the AI assistant, generating dialogue samples in the form of self-dialogue, and ChatGPT performs supervised training on these dialogue samples.
Then, ChatGPT enters the second stage.
The main purpose of this stage is to train a language model-based reward model to score machine-generated replies. In order to train the reward model, it is necessary to randomly extract machine-generated text and sample multiple different versions. Different versions are sorted by preference, and the sorted results are used to train the reward model.
Based on the trained reward model, the third stage employs reinforcement learning techniques to further fine-tune ChatGPT.
What sets this technology apart is its ability to use the reinforcement learning techniques described above based on human feedback, enabling the generated text to match human intentions and values, among other things.
Regardless of whether they understand technology or not, netizens all over the world are competing to open their "brain holes" to test how "magic" ChatGPT is.
According to CNN, it was asked to rewrite the 1990s hit "Baby Got Back" in the style of "Canterbury Tales"; a letter was written asking ChatGPT to remove bad accounts from credit reports ; it was also asked about its fairy-tale-inspired home design proposal.
Natural language processing technology sees the light of day
After an instant hit, the concept stocks of ChatGPT also followed along the way.
According to media reports, Hanwang Technology gained 7 daily limit in 9 days.
Natural language processing research is the company's main business, and the technology behind ChatGPT is NLP.
Xiong Deyi told reporters that NLP is a branch of artificial intelligence, which was first born in machine translation. Its history is actually longer than the name of "artificial intelligence", and it has been studied for more than 70 years.
To put it simply, the goal of this technology is to enable computers or robots to have the language abilities of listening, speaking, reading, writing, and translating like humans.
"ChatGPT is not a technology achieved overnight. It is the product of a variety of technologies superimposed to form a qualitative change. It is the crystallization of the NLP field." Xiong Deyi explained that its underlying technologies include Transformer, self-supervised learning, fine-tuning, human feedback reinforcement learning ( RLHF), AI alignment, etc., have extensive research and applications in the fields of natural language processing and artificial intelligence.
"However, ChatGPT combines these technologies ingeniously. It effectively avoids the unaligned behavior of large models, takes advantage of the intelligent emergent capabilities brought by the scale of large models, and enhances the instruction learning capabilities of large models and the ability to capture user intentions, unlocking The generalization ability of the large model is enhanced. These abilities are superimposed to form the qualitative change effect of ChatGPT." Xiong Deyi said.
The US "Miami Herald" reported that a user on Twitter took a mock SAT test with the help of ChatGPT and got a score of 1020 (out of 1600).
That's only slightly below average, according to College Simply.
But when we ask ChatGPT "Can you help me pass the test", it will kindly remind us not to use it to cheat, saying: "It is important to remember that the purpose of the test is to assess your understanding and mastery of the material degree. Cheating, or relying solely on outside sources like mine, destroys the value of the test and does not promote learning or growth."
This is because ChatGPT is added to the pre-designed "moral" guidelines, which is the human feedback reinforcement learning mentioned above.
"This enables the tool to answer follow-up questions, admit mistakes, challenge incorrect premises, and deny inappropriate requests," OpenAI said.
Xiong Deyi believes that this is because RLHF effectively improves the alignment ability of large models, that is, how to make the output of AI models consistent with human common sense, cognition, needs, and values.
After being "tuned" by RLHF, ChatGPT has become highly anthropomorphic and can learn and generate natural-sounding replies.
RLHF also enables ChatGPT to understand commands and respond appropriately, making conversations more realistic.
As a result, ChatGPT has quickly become the "ceiling" among all intelligent chat robots.
It's within reach, but it can't replace humans
With the development of science and technology, artificial intelligence technology has surpassed the limits that people thought could not be reached time and time again, from the IBM "Deep Blue" computer defeating humans in chess for the first time in 1997, to the IBM artificial intelligence robot "Watson" winning intelligence The competition "Jeopardy", from "Alpha Dog" defeated humans in the Go game in 2016, to Pluribus, known as the "AI God of Gamblers", crushed human players in Texas Hold'em in 2019...
Now, ChatGPT is at your fingertips, pulling and colliding with people's usual modes and frameworks for participating in production and life, and continues to "surge" on the track beyond human beings.
Topics such as "ChatGPT triggers ten occupational crises" and "20 occupations in the future may be replaced by AI" are frequently searched.
Will it really compete with humans for jobs, or even replace humans?
The Miami Herald reported that, according to a SortList survey of 500 ChatGPT users in six different countries, “beating workers” are already thinking about what it means for their industry.
About 23 percent of employees working in software and technology fear losing their jobs to it, and more than 40 percent of employers expressed interest in "hiring" ChatGPT with its rich marketing copy, the study said.
Other areas of expertise may also see workflow changes, including mid-level writing, advertising, media planning and legal documents, according to CBS Fortune Watch.
"ChatGPT and other AI technologies have the potential to change the nature of work, but they will not replace the need for humans." Asked whether they will replace human jobs to ChatGPT, it says they are designed to assist and augment Human tasks, not replacing humans.
They complement human resources and enhance human problem-solving abilities, increasing efficiency and effectiveness.
Xiong Deyi believes that at present, the more role of artificial intelligence is to improve human efficiency and assist human beings.
At the same time, the public needs to keep pace with the times, be familiar with and master new technologies, just like learning to use computers and mobile phones. After all, these new technologies will definitely create new job opportunities.
No self-awareness and thinking ability
British philosopher Derek Parfit defines human beings as "creatures capable of giving reasons and responding to them" in his book "On Things That Matter".
ChatGPT is an artifact that can give reasons and respond.
How is it different from others?
Is it self-aware like a human being?
Computers have become more sophisticated, but they lack the real understanding that humans have developed through evolution as autonomous agents embedded in networks of social practice.
This is an experience that artificial intelligence does not have.
When asked about ChatGPT, it replied: "I am a machine learning model, not conscious or self-aware like a human. I also cannot feel emotions or experience the world like a human. However, I can process large amounts of data quickly, and The ability to understand and respond to natural language input in a manner similar to human communication."
Zhang Jiaxing, chair scientist of the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute (IDEA), said in an interview with reporters that for practitioners in the field of artificial intelligence, first of all, there is no clear definition of what is self-awareness. It is difficult to measure quantitatively like face recognition and speech recognition.
"If we vaguely define 'self-awareness', I think first, the subject must have self-awareness—it can realize that it is doing something; second, emphasize autonomy—it can spontaneously Decide what to do and what not to do. However, at present, artificial intelligence does not know what it is talking about, and it cannot autonomously decide or choose what it will say. From this perspective, there is still some distance from autonomous consciousness.” Zhang Jiaxing said.
Zhang Shuwu also believes that although ChatGPT has made breakthroughs, its answers are based on empirical judgments and predictions based on requirements. It is also a "knowledge platter" refined on the basis of prior knowledge, without logical derivations and mathematical calculations. The real cognitive function has obvious limitations.
In 1950, the brilliant computer scientist Alan Turing came up with a thought experiment he called the "imitation game."
The interviewer talks to the two subjects through a typewriter, knowing that one is a human and the other is a machine.
Turing suggested that a machine can be said to be capable of thinking if it can consistently convince interviewers that it is human.
This is the famous "Turing test".
Currently, ChatGPT has not passed the evaluation of the Turing test.
In fact, so far, no artificial intelligence model can really pass the Turing test.
But according to the New York Times, OpenAI has built a more powerful system than ChatGPT, GPT-4, which can even generate images and text, and industry rumors it can pass the test.
Whether this is the case, we will wait and see.
Or become a tool to spread misinformation
People have gradually discovered that ChatGPT is not a panacea, nor is it always correct.
It's not good at basic math, can't seem to answer simple logic questions, and even argues facts that are completely incorrect.
According to the British "Nature" magazine, some researchers emphasized that unreliability is integrated into the way LLM is constructed.
ChatGPT, powered by LLM, works by learning statistical patterns in language from vast online text databases, which contain untrue, biased or outdated knowledge, making it easy to generate false and misleading information, especially when it is not extensively The technical field of data training.
OpenAI also admits: "ChatGPT sometimes writes plausible but incorrect or even absurd answers." Especially dangerous.
The New York Times reported that NewsGuard, an American news credibility evaluation and research organization, tested ChatGPT. Researchers asked it questions full of conspiracy theories and false narratives. As a result, it wrote news, essays, and TV scripts. Write articles and compile tons of convincing, clean, unsourced content in seconds.
For disinformation purveyors, generative techniques could make disinformation cheaper and easier to produce, the researchers predict.
"This tool is going to be the most powerful misinformation tool the internet has ever seen," said NewsGuard co-CEO Gordon Kravitz. A new false narrative, it’s as if there are artificial intelligence agents helping disinformation.”
There are also concerns that ChatGPT will challenge language diversity.
The Australian "Dialogue" magazine published an article saying that the default function of ChatGPT will give priority to English writing forms created by white people.
White English speakers have long dominated many writing-intensive industries, including journalism, law, politics, medicine, computer science and academia.
The output of these individuals in the aforementioned fields is many times higher than that of their colleagues of color.
The sheer volume of text they wrote means they likely make up the bulk of ChatGPT's learning model, although OpenAI doesn't publicly disclose the source of its material.
So when users ask ChatGPT to generate content in any of these disciplines, the default output is written in the voice, style, and language of those English-speaking white males.
American philosopher, author, and cognitive scientist Daniel Dennett described the juvenile sea squirt in his 1991 book Consciousness Explained: "It wanders the ocean, looking for a suitable rock or coral as a lifelong home." Once found, the sea squirt no longer needs its brain and eats it.
It sounds cruel, but there is a worrying metaphor behind it: as artificial intelligence becomes increasingly integrated into our daily lives, humans may become too stupid by relying on "smart" machines.
Xiong Deyi reminded us to be wary of artificial intelligence weakening human thinking.
"ChatGPT is actually learned and trained from the massive language data and knowledge base of human beings. If human beings completely rely on it and give up their own language expression, language thinking, and ability training, it is very sad and terrible."
(This version of the picture is provided by Visual China)