What is the "top sky" and "standing ground" of AI? That is, not only to solve some small problems in existing applications, but also to solve large problems at the NP-hard level.
”
Author | Berries and tinctures
Edit | Cen Feng Twilight
Songshan Lake, which bears the birth of Dongguan to break through the inherent development path, is a vivid microcosm of the high-quality transformation of China's urban economy.
In the "Science and Technology Dongguan" plan launched in Dongguan, Li Guojie was the first pioneer to participate in cooperation. If you look at the business leaders in China's IT industry, his name seems to be little known. But you must have heard of China's local high-tech brands: "Dawn" and "Loongson". And Li Guojie is the layout and founder behind them. Li Guojie, who is a scholar of Hongru, seems to be very different from the image of the business legend that people talk about, and he is more like a big gentleman who is deeply involved in academia, trying to build a bridge between scientific research, technology and industry.
At present, Li Guojie is also the chief scientist of the Cloud Computing Center of the Chinese Academy of Sciences and lives in Songshan Lake in Dongguan most of the time.
In 2021, Songshan Lake sounded the clarion call of "reform, innovation, and re-departure", and "innovation" is precisely the point that Li Guojie values most.
1
Dry all day long, with the times
The scene of the 2021 GAIR Conference
Li Guojie previously wrote that the key to innovation is not only to establish a world-class university, but also to the vision, strength and vitality of scientific and technological innovation of enterprises.
He has also expressed concern about the innovative ability of domestic AI research. "Although the state attaches great importance to the development of artificial intelligence technology, in recent years, Chinese scholars have also published a large number of AI papers and patents, etc., and have achieved remarkable results in the construction of smart cities, the fight against the new crown epidemic, the preparation for the Beijing Winter Olympics and other applications, and some artificial intelligence unicorn enterprises have achieved AI landing results." But most of our research is technology-driven, paper-oriented, goal-oriented and problem-oriented. ”
As early as 2006, Li Guojie pointed out in the Journal of the Chinese Academy of Sciences: "Although the innovation ability of China's enterprises, universities and scientific research institutions is very weak, I think the weakest link in China's national innovation system is technology transfer." This is also his own profound experience of the industrialization of Shuguang computer and Loongson CPU, during his tenure as the director of the Institute of Computing of the Chinese Academy of Sciences, he also took "technology transfer" as the key to improving the innovation ability of scientific research institutions, and promoted the implementation of the knowledge innovation project of the Academy of Sciences oriented to "innovation leapfrog and sustainable development".
After Li Guojie stepped down as the director of the Institute of Computing of the Chinese Academy of Sciences in 2011, the Cloud Computing Center of the Chinese Academy of Sciences (formerly known as the Guangdong Electronics Industry Research Institute) at the forefront of reform and opening up became another base for him - in 2005, Li Guojie founded the Guangdong Electronics Industry Research Institute in Songshan Lake, Dongguan, which is the first provincial-level scientific research platform jointly organized with a national scientific research institution in Dongguan.
Guangdong Electronics Industry Research Institute
"The purpose of establishing the Guangdong Electronics Industry Research Institute is twofold, one is to promote the establishment of a number of public technical service platforms to provide technical support for the processing and manufacturing industry in Dongguan, which is in urgent need of transformation; the other is to build the institute into a platform for the technological transformation of the Chinese Academy of Sciences." From Li Guojie's talk about the reasons for the development of Dongguan, it is not difficult to see that he regards the Guangdong Electronics Industry Research Institute as a "test field" carrying his own innovation and technology transfer thinking and solutions.
Li Guojie said that the "batch" refers to the fact that just two years after the Guangdong Electronics Industry Research Institute landed in Dongguan, the "Science and Technology Dongguan" project was officially launched, and Dongguan opened the road of cooperation and co-construction with colleges and universities and testing technology institutions. A large number of public science and technology innovation platforms have successively settled in Dongguan and Songshan Lake, and the resources for scientific and technological innovation have accelerated.
In recognition of Li Guojie's contribution to Dongguan's scientific and technological upgrading and innovation over the years, in 2021, the Dongguan Municipal Government awarded Li Guojie the title of "Honorary Citizen".
Affected by the epidemic, Li Guojie lived in Dongguan for a longer time last year, which also allowed him to calm down and have more time to think about the relationship between scientific research, innovation and the transformation of scientific and technological achievements.
What really made Li Guojie's above ideas "out of the circle" was an article titled "Academician Li Guojie: Domestic AI Research Can't Stand the Sky, Can't Fall to the Ground, It's Time to Think About It" in August 2021. It was like a flat thunderclap, causing heated discussion. It also pushed Li Guojie, who has always been low-key, to the cusp of the storm, and criticism and support came pouring in.
The discussion on media and social platforms about the lack of original work in the domestic AI field and even the entire academic community seems to have never stopped. The reason is that the current academic evaluation system is generally unreasonable.
Paying too much attention to superficial and indexed things such as the number of articles, IF, citation, etc., major universities implement "rigid" standards to evaluate talents. "Thesis orientation" makes many scientific research topics and research contents come from the desire of "building behind closed doors", which is divorced from actual needs, which is not conducive to generating innovative work from 0 to 1, and wastes a lot of capital investment. The dual role of the "academic evaluation system" and the "academic unspoken rules" may be one of the reasons why AI research "cannot stand the sky and fall to the ground" pointed out in Li Guojie's article.
However, some people believe that Li Guojie's statement that "listening to the words of the jun is like listening to a seat" is correct but has no guiding value. There are also people who think that "I will also say such a thing", and there is such a comment under the discussion of the article: "If the level of the academician is given in 2021, the diagnosis of the disease that cannot stand up to the sky and cannot stand up to the ground; then he should immediately give 'how to stand up to the sky, how to stand on the ground, why can't he stand up?' Why can't it stand? 's treatment options. ”
In the face of doubts, Li Guojie kept a low profile and only issued a brief statement on a WeChat public account. In the age of information explosion, this statement is also like a stone thrown into a lake, although it stirs up ripples, the lake will eventually calm down over time.
"First of all, I feel very relieved, because it is good that young people are willing to question and have the courage to question." At the GAIR 2021 conference held at the end of 2021, Li Guojie told Leifeng Network.
As a guest, Li Guojie participated in the live discussion of the GAIR 2021 "40 Years of Parallel Computing and System Architecture" commemorative roundtable. After the meeting, Li Guojie had a dialogue with Leifeng Network, commented on the long-standing tendency of computer science and artificial intelligence theory research to "not solve problems", and further interpreted the "can't stand the sky, can't fall to the ground": AI's "top of the sky" and "landing" refers to AI not only to solve some small problems in existing applications, but also to solve NP-hard level big problems, and our current common problem in the research direction planning of artificial intelligence is that it is either not "top the sky", Either it is difficult to "land".
There's a historical reason for this, and that's where he wants to remind researchers of what they should pay attention to.
The following is the transcript of the dialogue compiled by Leifeng Network, which has made an edit that does not change the original meaning:
2
The controversy caused by "can't stand the sky, can't fall to the ground"
Leifeng Network: Let's start with an article you wrote in August 2021 that "stand tall against the sky". At that time, your article was published and caused a lot of discussion in the industry.
Li Guojie: Actually, at that time, I had already issued a statement, and I was not making a conclusion about the current state of AI research in China. The original title of the article is "The Revelation of Major Achievements in the Application of Artificial Intelligence", mainly to expound the enlightenment brought to us by AlphaFold's breakthrough in the field of biology, and the relevant newspapers believe that the title is too flat and has not been communicated, so it is changed to "Domestic AI research "can't stand the sky, can't fall to the ground", it's time to think about it."
The problem I mentioned in the article is that many universities and enterprises in China have felt the problem of "not being able to withstand the sky and not falling to the ground". The main reason for the success of AlphFold 2 is that the DeepMind team has a keen eye to believe that the protein structure prediction problem can be solved with artificial intelligence. The direction itself is forward-looking, challenging, and meaningful when addressed. The new generation of artificial intelligence major scientific and technological projects launched by China have carried out research on data intelligence, cross-media perception, group intelligence, brain-like intelligence, quantum intelligent computing, etc., and have achieved many research results, but they do not cover this type of research. So it's time to think about it. This is a reminder to think more about when to choose artificial intelligence to do, and not to "go with the flow".
AlphaFold 2 freely models protein-like targets for two examples
"Standing on top of the sky" means: in terms of technology, we must "stand up to the sky", dare to break into the "forbidden area", and do forward-looking research that others think is impossible to succeed; we must "stand on the ground" in application, and we must solve major problems in economic and national defense construction, including using artificial intelligence technology to solve challenging problems in basic research.
Lei Feng Network: Compared with the domestic research on artificial intelligence, I believe that your views were not formed overnight. What made you pay attention to ARTIFICIAL intelligence and generate these ideas?
Li Guojie: After the article was published, the young people who did AI entrepreneurship were not convinced, and I fully understood, after all, in their view, I am just an 'old man' engaged in high-performance computing, what qualifications do I have to evaluate artificial intelligence? First of all, I feel very relieved, because it is good that young people are willing to question and have the courage to question.
As I've said on many occasions, I'm one of the "hipsters" of the second wave of artificial intelligence.
In 1987, Li Guojie (left), Xia Peisu (center), and Li Zhengdao (right) discussed issues at an international academic conference
After graduating with a master's degree from the Chinese Academy of Sciences in 1981, Mr. Xia Peisu recommended me to pursue a doctorate at Purdue University in the United States to study combinatorial searches related to AI. At that time, there were few Chinese scholars in the international AI academic circle. In 1984, I presented a paper at the AAAI Conference, which was then the American Association for Artificial Intelligence (AAAI was renamed the Association for the Advance of Artificial Intelligence in 2007). The fame was not as great as it is now, and I did not meet academics from home to Austin, Texas, for the conference.
In 1985, together with my mentor, Professor Hua Yunsheng, we co-authored a self-study reference book, Computer for Artificial Intelligence Applications, which became the IEEE's best-selling publication for three consecutive years, and most scholars who were new to the field of intelligent computers at that time read this collection of papers. In the book, I did not use the term "intelligent computer", but "a computer suitable for artificial intelligence applications", at that time it was difficult to make a real "smart machine", can only be said to apply computers to artificial intelligence. In 1987, I returned to China to work, and successively served as a researcher at the Institute of Computing of the Chinese Academy of Sciences and the director of the National Intelligent Computer Research and Development Center, and also focused on high-performance computing research. But I've never stopped focusing on "artificial intelligence."
Books co-authored by Li Guojie and his mentor
Leifeng Network: It can be said that you have witnessed the growth and development of the artificial intelligence discipline in China, what was the initial situation of artificial intelligence in our country?
Li Guojie: The development of artificial intelligence in China has gone through a detour. The earliest artificial intelligence society was not in the Chinese Association for Science and Technology system, but in the social science system, under the Chinese Academy of Social Sciences. (The early story of the founding of the Chinese Society for Intelligent Engineering is described in more detail in the forthcoming "A Brief History of Chinese Intelligence (Volume I): A Tribute to legends").
At the beginning of the 863 program, I was the deputy leader of the intelligent machine expert group (306 theme). According to the opinion of the expert group, Dai Ruwei (academician of the Chinese Academy of Sciences, a well-known cybernetics and artificial intelligence expert) and I came forward to contact artificial intelligence scholars across the country and tried to establish a national large-scale artificial intelligence society, benchmarking with the world's mainstream artificial intelligence societies and belonging to the science and technology association system, but this matter could not be done. At that time, our artificial intelligence research was not in line with the mainstream artificial intelligence in the world, and it was also artificial intelligence, and everyone paid attention to things with different points of force.
Leifeng Network: You mentioned in this article that AlphaFold has made some achievements in the field of biology. How did you come up with the idea of using AlphaFold as an example?
Li Guojie: My understanding of the field of biology began with my student Bu Dongbo, who is an expert in this field. Before the advent of AlphaFold 2, many scientists at home and abroad were doing research on using computers to predict the three-dimensional folding structure of proteins. Dongbo Bu's team published a paper in the journal Nature Communications in 2020 and made a world-leading achievement in protein structure prediction. Several of his representative predictions were better than AlphaFold's, alphafold's GDT score in CASP matches was about 50 points, and Bu Dongbo was able to achieve 70 points. AlphaFold 2 later surpassed him with a score of 90.
Why use AlphaFold as an example? This is based on a basic judgment about artificial intelligence: artificial intelligence must not only imitate people, but also solve big problems. From a computer science perspective, AI should focus on NP-hard level puzzles. Our existing artificial intelligence research is either not enough to solve small problems, or it is difficult to land and difficult to be applied in actual scenarios.
3
"Artificial intelligence is used to solve big problems"
What does "top the sky" mean
Leifeng Network: You mentioned earlier that academic research on artificial intelligence is too limited to John McKinsey's definition, that is, the goal of artificial intelligence is to be "like people", and pointed out that we should break through the narrow understanding of intelligence. What does this have to do with the NP-hard level puzzles that AI is trying to solve?
Li Guojie: "Artificial intelligence like people" is a direction that has been valued by everyone, but I think another point of artificial intelligence is to "solve big problems". In particular, machine learning is used to solve significant scientific problems, that is, to "effectively solve" exponential complexity problems in polynomial time.
Exponential complexity means that the time or space (storage usage) required to solve a problem increase exponentially as the size of the problem increases. This is also what people often call the combination explosion. In computational complexity theory, a large class of problems that do not yet find polynomial-level complexity algorithms is classified as NP-hard problems. If a problem can find an algorithm of polynomial complexity, such as a sorting algorithm, etc., it can be accurately solved by directly calculating according to a certain program, and people generally do not think that it is an artificial intelligence application. The problems that artificial intelligence has to study are almost all NP-hard problems, and from its inception, it is necessary to deal with the combination explosion. In this sense, the "sky" of artificial intelligence is the combination explosion, and the so-called "top sky" is to find a clever way to overcome the combination explosion.
After more than 60 years of artificial intelligence research, we have made satisfactory progress in dealing with the combined explosion in the fields of computer vision, hearing, and machine translation, but there are still a large number of NP-hard problems waiting for us to break through in basic research and practical applications. As amino acid monomers increase, the computational complexity of protein structure prediction increases exponentially, and if searched with barbarism, the likely combination of protein structure predictions is as high as 10 hundreds of powers, which is a typical NP-hard problem. Today's "card neck" chip design problem is also an NP-hard problem. The Institute of Computing of the Chinese Academy of Sciences is exploring the replacement of "chip design" with "chip learning", which may be a way out to crack the chip design talent gap.
These are the real big problems, and if artificial intelligence research is to stand up to the sky, it must enter these "forbidden areas" that were considered impossible in the past.
Difficult to land is "not for also", "must not also be"
Leifeng Network: The two examples you mentioned above, protein structure prediction and EDA are both highly valuable problems. Do you think that as long as you focus on NP-hard level problems, you can make artificial intelligence research both top the sky and land?
Li Guojie: It is not enough, it is related to a "tradition" in the computer science community. There is a classic graduate textbook on NP, Computers and Intractability: A Guide to the Theory of NP. The first page of the book is a cartoon of two people talking and one person saying, "I can't find an algorithm that works, but neither can all these best people" (see image below). This actually represents the attitude of the computer theory community to the NP-hard problem.
To this day, 50 years later, this "tradition" continues to influence generations of scholars in the field of computer science. People are desperately trying to prove that "this problem is not NP-hard". As long as it is an NP-hard problem, there is no responsibility for 'us', and it is a very funny situation to no longer wonder if there is any way to solve the difficult problem. Other disciplines struggle to solve all kinds of problems, but computer science discusses what problems can't be solved all day long.
Only dwelling on the proof of theoretical boundaries, without trying to find a way to solve the problem, is the fundamental reason why we cannot let the difficult problem land. For decades, we have seen np-hard problems as obstacles, problems we can't solve. But with the advancement of artificial intelligence and computer technology, we have found that through heuristic search, knowledge engineering and machine learning, coupled with sufficient computing power, many NP-hard problems can be satisfactorily solved. NP-hard means that the era of the impossible is over, NP-hard simply means that there may not always be algorithms that work and are extensible, but many NP-hard problems are actually solvable for applications. The task of ai-powered scholars is to unearth possible solutions that seem impossible.
The use of artificial intelligence to solve NP-hard problems I am talking about here does not refer to "solving" in the theoretical sense.
The "P=NP" problem may not be solved for decades, but AI scholars can keep approaching this equation in practice. The computer science community used to tend to make perfect proofs of theory, perhaps because later people misunderstood Turing's intentions. Turing defines the problem of indeterminate, such as the problem of downtime, pointing out that this type of problem can never be solved with a Turing machine, which delineates the power boundary of the Turing machine, which in turn defines what a computable problem is. The result itself was great, but then many people understood turing machines in the wrong direction, clinging to the former, keen to explore what problems were "theoretically" incalculable or in an acceptable time and space, rather than actively exploring how to "actually" solve intractable problems.
This misconception stems from the failure to distinguish between "problem" and "problem instance." The "problem" that is required to be solved in computer science refers to a class of problems that contain various examples, and the "problem" that artificial intelligence applications are trying to solve is often a specific problem example. In fact, the most difficult to solve in an exponential complexity problem (class) is usually only a few of its instances, and the other instances can be solved.
The Golden Age of Machine Learning
Leifeng Network: Deep learning has been very popular in recent years, is machine learning an effective way to solve the NP-Hard problem?
Li Guojie: There is complexity in the calculation process, just like there is friction in physical motion, friction cannot be completely eliminated, and complexity cannot be completely eliminated. But friction can be reduced by changing the material and the way it moves, and the actual complexity of the solution method can also be changed by changing the way the problem is described or the way knowledge is represented. The current widely popular deep neural network description of a problem is completely different from the symbolic reasoning in the past, and the distribution of connection weights obtained by deep neural networks through machine learning is actually a new way of problem and knowledge representation, which has shown unprecedented problem solving ability.
There is a popular saying in the artificial intelligence community that deep learning has hit the ceiling. But I think there is still room for deep learning to grow, and the huge room for machine learning in a broader sense is difficult to estimate, and the next decade may be the golden age of machine learning. Machine learning, especially deep learning, is less dependent on human knowledge and can be applied to many types of NP-hard problem solving. Machine learning is highly scalable, and new discoveries may continue to be made through scaling effects. Artificial intelligence is a science that pursues "surprising" results, and I believe that there will be many "surprising" new achievements popping up in the next decade.
Man has "human intelligence" and "wit", and the scope of knowledge will be expanded to "dark knowledge" other than "explicit knowledge" and "latent knowledge". The combination of machine learning, huge computing power and existing scientific knowledge will promote scientific research to a large platform model based on artificial intelligence technology, and the depth and efficiency of scientific research will exceed the "fourth way of scientific research" that is only data-driven. The prototype of the new "fifth paradigm of scientific research" is now looming.
Artificial intelligence is inseparable from computational thinking, but it is not equivalent to computational thinking. Turing's definition of computation (execution of algorithms) is a function mapping from input to output, the result of which must be repetitive and consistent, and this "computational thinking" limits the creativity of artificial intelligence research to some extent. "Turing machine" does not refer to a "machine", but to a specific operating process or mode of use of a machine, including the division of the initial and final states. The output properties of machine learning are often based on experience and situation, and a system that is constantly learning does not repeat previous internal states. The concept of "computation" is not enough to cover all intelligent and cognitive processes. The traditional theory of computational complexity, which simply divides "easy-to-solve" and "difficult" problems, also needs to be broken.
Turing machine model
4
AI research requires strategic vision
and the perseverance of "biting and holding on"
Leifeng Network: You once said that there is still a long distance between domestic scholars and first-class scientists, such as the success of the AlphaFold2 project, you think that they are "sharp-eyed" when choosing topics. But it does not mean that you can easily find a good scientific research topic, how do you think Chinese scholars should cultivate a "keen eye" in scientific research?
Li Guojie: The so-called "lack of keen vision" refers to the fact that the scientific research projects in the layout are either incremental technological improvements, that is, they cannot stand up to the sky, or they are ideal goals that are difficult to break through for decades, that is, they cannot fall to the ground. DeepMind scholars use AI to predict protein folding structure, which fully reflects the foresight and deserves our deep consideration.
How to have a "keen eye" is a big problem in academia, and it is also the difference between so-called "masters" and "second-rate scholars". Truth is often in the hands of a few people, very few scientists can really see the direction of scientific research, and whoever gets the first major new discovery in scientific research has the fortuitousness. However, "following the current" is a relatively common phenomenon in current scientific research, and in general, chasing hot spots and following the current cannot make great achievements.
"Keen eye" is the embodiment of a person's comprehensive quality, not only scientific literacy, but also humanistic sentiments. Mr. Qiu Chengtong, a famous mathematician, said: "The originality of China's theoretical scientists is still not as advanced as that of the world, and I think an important reason is that our scientists' humanistic cultivation is still not enough, and their feelings for the truth and beauty of the natural world are not rich enough." ”
"Keen eye" is not a momentary node that refuses to go with the flow, but a longitudinal extension of the timeline, before the node is a deep insight and insight into the industry, after the node is the determination to keep the cloud open and see the moon. Achieving original major scientific research achievements not only requires learning and daring to be the first, but also needs to "bite and hold on" and persevere.
We all know Turing Award winner Geoffrey Hinton, and behind his recognition is 30 years of silent persistence. At that time, the mainstream academic community in the United States was not optimistic about deep learning, and After several turns, Hinton, whose research funds were stretched, could only go to Canada. In 2006, Hinton finally made a splash by publishing in Science. It wasn't until 2012, when Hinton and his student Alex Krizhevsky won the ImageNet Massive Visual Identity Challenge, that deep learning was noticed and shined ever since.
5
In basic research, we should pay attention to play
The role of engineering technology
Leifeng Network: You said that AlphaFold does not propose new scientific principles, it is more like an integration work. In the article, you also proposed that engineering technology is not a tool, not just the application of basic research results, but an important part of the role that can play a huge role in basic research.
Li Guojie: No. There are many people in our country who do engineering, but there is some disconnection in the use of engineering methods to solve basic scientific problems. I mean, the ability to organize dozens or even hundreds of people to collaborate on major basic research problems needs to be improved, and attention should be paid to the role of engineering technology in basic research. However, under the wave of AI, the recent engineering implementation of brushing points and brushing seems to be too heavy, and it is also worth noting that the mining of the law itself is ignored.
The AlphaFold team is a typical interdisciplinary collaborative team, with 34 authors publishing this major achievement in Nature, of which 19 are tied for the first author, including well-known scholars in the fields of machine learning, speech and computer vision, natural language processing, molecular dynamics, life sciences, high-energy physics, and quantum chemistry. The reason why proteins form a stable folded structure is that the potential energy inside the molecule drops to the lowest point, and the predictive calculation is actually an optimization of energy minimization, which involves knowledge in many fields.
AlphaFold2 does not have new discoveries in the composition mechanism of protein structure, but can make predictions much better than others faster and more accurately in engineering, which is recognized by the biological community and is currently the best solution.
The way to obtain major scientific research results is different from the past, before a person can make achievements by thinking hard, and now it needs interdisciplinary cooperation and strong engineering support to get things done, so engineering technology is now part of basic research.
6
Hard work, tranquility
Leifeng Network: If you had to use two words to describe yourself, what word would you choose?
Li Guojie: "Struggle" and "tranquility". Whether it is my personal growth experience, or the development of projects such as "Dawn" and "Loongson", if you leave the spirit of "hard work", today's achievements will no longer exist. But after I was elected an academician, my long-lost college classmates asked me what I was pursuing now. My answer was: "I am seeking tranquility". The two realms of "struggle" and "tranquility" that seem to contradict each other are unified in my mind.
Since I started high school, my life path has been bumpy, and I have never had extravagant hopes for promotion and prosperity, and I only want to pursue self-improvement in a quiet life. Lin Zexu's "The wall stands a thousand people, no desire is just; the sea is a hundred rivers, there is tolerance is great", and Zhuge Liang's "indifferent to Mingzhi, quiet and far-reaching", these two pairs of banners have always been my motto.
Actually, I'm a very ordinary person.
In my lifetime, I have not climbed to the peak of science and technology and made amazing scientific research results. I'm well aware that I'm not a particularly smart person, and I'm not particularly capable. Fortunately, I have experienced more and have more setbacks, so I don't suffer from gains and losses, and I will not give up halfway to the goal. Looking at the problem is not too disturbed by small things, there is a momentum in the heart, that is, to get things done, not willing to reach the goal.
When I say "tranquility," I don't mean what young people today often call 'Buddhism.' The 'Buddhist system' that is now popular on the Internet refers to an attitude of "no desire, no demand, and no concern for anything." I think that good things should still be pursued, but we have wasted time in the fight for fame and profit. Do not forget the original intention, do not live up to the mission, the road under the feet will be wider and wider.
References: Xiaoxiang Morning Post "Fifteen Years of Academicians of Great Powers".
For more articles on the frontiers of AI and AI technology research, please pay attention to AI technology reviews.
END