exclusive interview｜zhang junping: artificial intelligence does not need to pursue "omnipotence", it only needs to be good at one aspect
2024-09-02
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
"teacher zhang, artificial intelligence is so advanced. if we study, we will be replaced, and if we don't study, we will still be replaced. then what's the point of studying?" at this year's shanghai book fair, a young reader asked professor zhang junping after a book sharing event with the theme of "will artificial intelligence replace humans? - the history, current situation and future of artificial intelligence."
zhang junping, a well-known expert in the field of artificial intelligence and a professor at the school of computer science and technology of fudan university, replied that mastering basic knowledge such as matrix operations is crucial to understanding artificial intelligence. only by learning these basics can we better control and optimize algorithms and understand the decision-making of artificial intelligence. he also emphasized that learning can enable people to better understand and apply artificial intelligence technology.
book cover of "a brief history of artificial intelligence".
in order to help the public understand the development of artificial intelligence, zhang junping published "a brief history of artificial intelligence" last year, which introduces the history and technological progress of ai in popular language. the book covers important milestones from early theories to modern deep learning, and looks forward to the future of ai.
the history of artificial intelligence can be traced back to the mid-20th century, and it has experienced many ups and downs. in the 1950s, pioneers such as alan turing and john mccarthy laid the foundation. despite the "ai winter", since the 21st century, with the advancement of computing power and big data, deep learning has led a new round of development in artificial intelligence.
zhang junping's book "a mistake-loving agent" published in 2019 explores the challenges and misunderstandings faced by artificial intelligence, and encourages an open mind to accept the mistakes made by agents in the learning process. at that time, chatgpt was not well known, and the concept of artificial intelligence was not popular among the public. zhang junping proudly told the paper (www.thepaper.cn) that even though technology is developing rapidly, most of the content in the book has not changed, and the views expressed are still valid.
the emergence of the "ai bubble" is due to excessive expectations
the paper:on september 17, 2018, the first world artificial intelligence conference was held in shanghai. you also began to popularize relevant knowledge to the general public outside of university classrooms in june 2018. from then until the publication of "a brief history of artificial intelligence" last year, how did you become a popular science expert in the field of artificial intelligence?
zhang junping:i first did science popularization in 2018 when science times invited me to write an introduction to generative adversarial networks. (editor's note: gan is a deep learning architecture that trains two neural networks to compete with each other to generate more realistic new data from a given training data set.)
because the content involved is quite complicated, i started to think about how to give you a popular science interpretation. there is a concept of cross-quotient, that is, the generation network and the confrontation network, which reminds me of zhou botong's left-right fighting technique in the legend of the condor heroes, that is, attacking with the left hand and defending with the right hand.
at that time, there were many versions of the explanation of deep learning in the academic circle, but i wanted to explain this issue from a popular science perspective, so i wrote "deep learning, you are the 116-year-old longevity grandma". the response was very good in the end, so i wanted to continue writing it along this line of thought. in 2019, i published "intelligent agents who make mistakes", and later began to turn to short videos.
this time, when writing "a brief history of artificial intelligence", my purpose is to sort out the history first. its rigor is far higher than the previous book.
i started doing research on artificial intelligence in 2000. i am clear about the main line of its development and know how to describe it. however, in order to write it down very accurately, i need to look up literature and make modifications based on the original content.
for example, why did the first artificial intelligence fail? this content first reminded me of many things, but then my teacher, academician lu qian, pointed out that the expert system era (from the 1970s to the 1980s, it was a knowledge-based artificial intelligence system that used expert knowledge and rules to solve problems in specific fields) was a key node that helped the first artificial intelligence slump to get out of it, and it should not be mentioned before, but should be adjusted to the end. so i re-checked the information based on his advice and adjusted this part.
writing a book may be different from doing research (writing a paper). if the book is well written or has sufficient predictions about the future, then the vitality of this book will be much stronger than that of a scientific paper. especially in the field of artificial intelligence, in many cases, only half a year after the paper is published, the (technical) performance is improved and the original article is gradually forgotten.
the paper:some people say that there is competition between china and the united states in the field of artificial intelligence. you have been to the united states for exchanges before. what will be the decisive factor in the future?
zhang junping:on the one hand, it is talent. research on scientific development shows that if the number of scientific achievements of a country accounts for more than 25% of the world's total during the same period, this country can be called a "world science and technology center." however, the world's science and technology center has been changing. for example, the first was italy, the united kingdom, then france, germany, and later the united states.
will the world's science and technology center change? will it come to china? these are all questions. if there is to be a reversal in the competition between china and the united states, i think the most obvious sign is when the world's science and technology center can come to china.
the paper:you said before that there is a trend of everyone throwing themselves into the artificial intelligence craze, and recently the term "ai bubble" has emerged. what do you think?
zhang junping:whether it is a "bubble" depends on how you define it. in fact, it can also be described as "overdraft". because in the academic world, there is no such thing as a "bubble". as long as you move forward slowly and steadily, step by step, you will be fine. but if it is a "bubble", it means that there must be an overly high expectation ahead.
for example, in this round of ai craze, the academic community may not think that the development of ai is so hot. instead, they think that under the current environment, there is still a long way to go before general ai. however, enterprises may have too high expectations for the development of ai, thinking that it is about to achieve or has already achieved certain results. if expectations are high, disappointment may also be high. because too much investment has been made, if the ideal results are not achieved, people will think that it is a bubble.
the paper:in fact, you also mentioned that it is possible that things that machines find simple may be complex for humans, and vice versa. our original intention was to let machines help us do some simple and repetitive tasks, but in fact we are also letting ai write poems or do some difficult tasks. do you think that humans will become overly dependent on artificial intelligence in the future?
zhang junping:in fact, dreyfus talked about this issue in his book "what computers can't do". he said that if humans rely heavily on artificial intelligence in the future, the possible result may not make humans smarter, but make humans super stupid.
the paper:do you agree that artificial intelligence will destroy humanity?
zhang junping:i don't agree.
innovation, regulation and data protection
the paper:you have mentioned many times before that people are overly optimistic about artificial intelligence, and some people think that artificial intelligence is not yet that "intelligent". how do we view people's expectations for artificial intelligence?
zhang junping:most researchers in artificial intelligence are optimistic. but i always think that it is very difficult for people to understand themselves. just like an ant, if it crawls on a two-dimensional plane, it will never know that it is on a two-dimensional plane, unless someone stands in a higher dimension than it and sees that it is actually crawling in a three-dimensional space. it is the same for people. it may be difficult to fully understand yourself because your dimension has been limited and there is no way to jump out of this dimension to see yourself.
i compare artificial intelligence (and its application) and humans to airplanes and birds. humans have always wanted to simulate the flight of birds. in fact, there is no need to understand all the structures of birds. we just need to focus on the angle of flight. if we do a single aspect well, we can make the airplanes fly farther and farther and carry more passengers.
the same principle applies to artificial intelligence. although it is difficult to figure out how human intelligence works in a short period of time, we can focus on one goal, such as whether facial recognition is accurate or whether the previous word is used to predict the next word in natural language processing. we only need to do one aspect well.
the paper:that is to say, we can focus more on applications rather than so-called general artificial intelligence. regarding artificial intelligence prediction, how much help do you think artificial intelligence can provide in weather forecasting, or what prospects does it have?
zhang junping:weather forecasting is still a very difficult problem. because weather is different from the images and videos we see now, it is actually the data we get, or it may be a cloud map obtained by scanning the sky with radar at a certain distance from the ground (such as 10 kilometers), and then predicting the weather based on the changes in the cloud map. but the cloud map shows the evaporation of water vapor on the ground. it is difficult to know how this state is generated and how it disappears. from the ground to the 10-kilometer space, we cannot make it full of collectors and observers. even the information on the ground we don’t know much. for example, there may be only a few hundred observation stations in shanghai, and the amount of information collected is not enough, which makes it not so easy to make weather forecasts. for typhoon forecasts, not only the (observation) range should be larger, but we rarely or even cannot make forecasts at sea. in recent years, teams including deepmind, huawei, alibaba, and fudan’s fuxi have paid special attention to weather, and there is no progress, such as short-term forecasts and local forecasts, but it is still very difficult to do a good job, especially for sudden and rare meteorological events.
the paper:in terms of governance, we have seen a lot of news reports about copyright owners such as visual artists, news media, and record companies filing lawsuits against technology companies for using their works to train generative ai systems. as both a creator and an expert in the field of ai, how do you view the issue of technological innovation and data protection?
zhang junping:this problem has not been satisfactorily solved so far, and many people currently have certain objections to ai creation. take writing a novel as an example. after writing the first few dozen chapters, an author uploads his creative outline to a certain platform, and the platform may use his outline data for ai training - filling in the content of the novel, which may result in the content supplemented by ai being not much different from what the author originally wanted to write. in a sense, it will affect some of the author's creative intentions.
therefore, as a creator, you need to handle your products and creative achievements with caution, and do not blindly upload content to the cloud, because there is a risk that the platform will use the materials uploaded by users for training.
the paper:how do you think we should strike a balance between innovation and regulation?
zhang junping:at this stage, strict supervision is not needed. from a national perspective, data is the most important thing to be supervised. the "data security law of the people's republic of china" issued a few years ago also reflects the need to be careful about data leakage.
but in terms of innovation, i think we should focus on encouragement. in fact, we don’t know what to do. because scientific research is like this. many times we don’t know how to do it, but we may find a way as we go along. this is the reality of scientific research.
the paper reporter zhang wuwei and intern zhan huijuan
(this article is from the paper. for more original information, please download the "the paper" app)
