Editor’s Note: This article comes from the WeChat public account “Hatching Hand” (ID: ibushouzhi), author Li Wei, 36氪 authorized to publish. Topic: One of the top ten artificial intelligence scientists in the world, the American Academy of Sciences, the pioneer and deep-rooted pioneer of deep learning, Terence Shenovsky, with the drooping fruit being picked, more and more People have questioned whether artificial intelligence will enter the cold wave again? But few people know the true situation of the cold wave. In what aspects of artificial intelligence has undergone changes over the past 60 years, what factors will restrict its development? In what direction will it evolve? In the face of various constraints, what efforts can we make? The recent catcher ID (ibushouzhi) and the American “Four Academicians” (only three in the United States), pioneers and founders of deep learning, “Deep Learning: The Age of Intelligence The author of the book “The Core Drive Power” Terence has exchanged a lot and hopes to solve the above problems. Terence is involved in the first wave of AI, and his deep thinking about the current state of AI development is worth a look. Author / Li Wei editor / Pan Yubo unauthorized, declined to reprint one, AI development and constraints Li Wei: You participated in the research in the first wave of AI, what changes have occurred in AI so far? Terence: In the 1980s, computers were slow and expensive, so they had to be logically programmed, but today computers are a million times faster than in the past, and with deep learning, AI can be done by example. Self programming. For example, when I started to study AI, we could only solve small problems, but today we can solve problems through a deeper learning network. Li Wei: What is the nature of deep learning? Terence: The success of deep learning is based on learning highly complex statistical models in very high-dimensional parameter spaces. This makes it possible to solve real-world statistical problems such as speech recognition and natural language translation, which are also very difficult. Li Wei: What do you think are the main misunderstandings of the public about deep learning? Terence: People are worried that deep learning will take away human work. On the contrary, for humans, whether you are a white-collar worker or a doctor in the office, deep learning will make you smarter. Some work will change radically, but new jobs will also be created. In fact, whenever a new technology is produced, there will be an adjustment transition period during which unexpected results will occur. People are currently testing autonomous vehicles, but as I described in Deep Learning, it takes decades for human society to adapt to this new technology. Li Wei: What level will the final deep learning develop? Terence: Deep learning is just the first step in general artificial intelligence. It is based on our cerebral cortical structure, but there are many other areas in our brain that can be used to develop general intelligence, and we need to understand this. Li Wei: What areas of deep learning can solve problems? There are about 20 unicorns and 30 quasi-unicorn companies in China’s AI field. Nearly 80% of them are related to image recognition or speech recognition. What is the situation of American AI companies? Terence: Deep learning opens a door for artificial intelligence to communicate with people through voice, language, and vision, and AI based program-writing can’t do this because the problems are different and the programs written are huge and complex. . Deep learning can be applied to many problems with sufficient data volume, and it has a major impact on many other issues. For example, in the medical field, deep learning is better for medical diagnosis of many diseases or against the best doctors. Deep learning is also very effective in many fields of science, such as astronomy and cell biology. There are hundreds of small AI startups in the United States, and there are many hardware startups that build dedicated machine learning chips. Li Wei: Why are there so many hardware companies in the United States that produce dedicated machine learning chips? Terence: Calculating energy costs is the ultimate limit of artificial intelligence. There are 100 billion neurons in the human brain, each of which is connected to thousands of other neurons, totaling billions of synaptic connections. The power required to run the brain is 20 watts, but a petascale supercomputer that is far less powerful than the brain has a power consumption of 5 megawatts, which is 250,000 times the brain’s power consumption.At present, the largest deep learning network is just embedded in the brain part like the size of rice. As the learning system expands, the development of AI must produce more energy-efficient dedicated chips. We need a million times more hardware than now. Now we are just getting started: large technology companies like Google, Microsoft and Amazon are building more energy-efficient dedicated machine learning chips. Li Wei: Data is also an important reason for restricting the development of AI. Most of the current data is concentrated in the hands of technology giants in various countries, forming an island of data. If the data is not complete (diversity, sufficient), AI may make the wrong decision, is the blockchain one of the solutions? Terence: Blockchain technology has the potential to help us solve data access problems, one of the many solutions being explored around the world. Second, AI’s national borders and governance Li Wei: Do you think China and the United States will have an advantage in developing AI? It is conceivable that AI will have an impact on the world’s pattern, so what role should the government play in it? Terence: AI is awakening all over the world, and no country has a monopoly. In the 1980s, only a few universities had enough computer skills to conduct AI research. Today, anyone with a laptop can make an important discovery: AI has become democratized, meaning that no one can predict where the next big step will come from. The government should not obstruct this, but should strengthen international cooperation to promote this process. Li Wei: You also mentioned in China’s book “Deep Learning” that China has more data, but does it mean more data and more engineers. This scale advantage can push back the basic research level. Breakthrough or determine the route of technology? Terence: My comments in the book are based on the application of current AI technology to AI and can provide many solutions. But some problems may require a new breakthrough, and no one knows when and where the breakthrough occurred. Li Wei: Nowadays, the application of AI has obvious boundaries. What conditions do you need to play the role of artificial intelligence? Some people think that after five conditions (sufficient data, certainty, complete information, static, single-task in a specific field), it is difficult for AI to play a role. How do you think? Terence: These current limitations will eventually be overcome. Deep learning has been combined with intensive learning to reach the world level of Go. We have found many other types of learning algorithms in the brain that will greatly enhance the functionality of AI in the future. Li Wei: Indeed, the combination of deep learning and reinforcement learning has promoted the self-evolution of algorithms. If we promote the development of AI according to people’s thinking, will it make the development of AI into trouble? If the development of AI is not within the expectations of people, will it become uncontrollable in the future? Terence: All technologies can be used for both good and evil purposes. Some of the technologies that exist with us are even more risky than AI, such as nuclear weapons and biological warfare. So far, we have survived by identifying and mitigating risks. Natural evolution gives us intelligence to help us survive in an uncertain environment, and we need to save ourselves to survive in a more uncertain environment. Li Wei: This seems to have forced the management of AI to be put on the road. Terence: Because we created AI, we should be able to better control AI compared to controlling ourselves, and human beings themselves are a much worse troublemaker. AI will have many uses, and we can’t even imagine it now. We have to deal with every new application to make sure it is properly regulated. Third, the evolution and break of AI Li Wei: AI’s system is still very vulnerable and vulnerable to attack and deception. It requires a lot of data, and it is unexplainable. What is the reason for this defect? American physicist Freeman Dyson believes that there is a lot of noise in the AI direction, and the whole circle is built on the wrong idea. Because the brain is analog, and the machine is digital, the brain is used to compare images and patterns, the machine can do it but the flexibility is poor. He believes that unless we can make a simulated model, not a digital model, we can have more knowledge of the brain, and few people are working in this direction. Terence: We don’t understand how the human brain works, but even if it’s not perfect, it doesn’t stop us from using it.Our mathematical understanding of neural networks is making significant progress and can fix defects, but it takes many years. I think Freeman Dyson may not be aware of these advances. Li Wei: We have deepened our understanding of the neural network and what is the significance of the development of AI? Terence: A theoretical understanding of a technology can make a significant improvement to the technology, but it can be a long process. For example, we have experienced the development of the first manned flight of the Wright brothers in 1903 to today’s large jets. Li Wei: Of course, some people have asked questions: Who stipulates that AI must develop according to human intelligence? Terence: Nature has allowed the human brain to evolve to solve many problems, and we can learn a lot from nature. Nature solves these problems, which proves that the solution is possible. Until recently, we didn’t know much about how the brain works, but in the 21st century, great progress has been made, and reverse engineering of the brain has just begun. Li Wei: The evolution of nature really makes us realize the problems of AI at this stage, and there will be solutions in the future. But does the understanding of evolution also limit the development of AI? Commercial companies are using the intuition and feelings that evolution brings us to control, such as using the human body’s demand for sugar in the evolution to sell sweets to people, but the environment has changed (the fruit is the sweetest food in the primitive society) . Terence: Humans are not doing well enough to regulate the business world. You are pointing out the flaws in human intelligence, which may not be the best goal of AI. Li Wei: What is the best goal of AI? Terence: Everyone has their own goals. I think the ultimate goal of AI is to let humans better understand themselves and tilt the balance in a better direction – it is in the interest of all of us to help each other. Li Wei: This means that we will be symbiotic with the machine in the future. Terence: “The symbiosis” is a good description of the way humans and AI cooperate. AI will not replace us, but will make us do things smarter. The phenomenon of “symbiosis” has now taken place in the fields of medicine and biology. Li Wei: This reminds me of the scene where people and machines coexist in many science fiction movies. Rodney Brooks, a well-known robotic manufacturing expert, believes that for robots to perform their daily tasks, their higher cognitive abilities should be based on the interaction of sensory movements with the environment rather than abstract reasoning. You also mentioned that copying the body is more complicated than copying the brain. Terence: I totally agree with Rodney Brooks that we need to better understand how our bodies can help us solve problems. Since the advent of AI in 1956, we have realized that we think that easy problems, such as vision, speech, and motion control, are much more difficult than we think, and abstract reasoning is much simpler. Li Wei: Although our brain is not completely logical, it does not hinder your vision. “If the brain is based on logic, then it should be a universal intelligence across domains.” AI is based on probability statistics, and the mathematics category is based on logic. Does it prove that AI is universal? Terence: AGI is a system-level issue, and we are still solving peripheral problems. We need more time to explore how to integrate all the features that deep learning can achieve. Today, Deep Mind’s inspiring research is leading the way. Li Wei: What is the current driving force for AI development, and what is the next new power? Terence: The time from AI development to its conversion to mass market products is about 50 years. The basic findings that drive the development of AI today came from the 1980s. On the basis of progress, the current technology will take decades to mature. A revolution is taking place in the current field of neuroscience, which will lead to new insights into the brain’s ability to create intelligence, but it will take decades to turn these findings into next-generation AI. Li Wei: What changes do you think AI will have in the next 3 years? Terence: On the basis of the existing technology, the short-term progress will be gradual, and the unpredictable major advances will occur within 25 years. .