Geng Hongming: Dreyfus's artificial intelligence "alchemy"

Hubert L. Dreyfrus, a philosophy professor at the University of California, Berkeley, boldly predicted that AI has limitations and cannot handle complex problems like face recognition and chess. In the mid-to-late twentieth century, he likened artificial intelligence to the "alchemy" of wizards and warlocks. However, history seems to be moving in the opposite direction predicted by Dreyfus. In the second decade of the 21st century, the game between the Go robot AlphaGo and humans has become a hot news event, while AI face recognition has been widely used, appearing in various life scenes such as office buildings, train stations, and shopping centers. It can be seen that Dreyfus's prophecy failed. The label "alchemy," which Dreyfus derided, now seems to have been given back to himself. In the face of this event, it is necessary to revisit Kant's classic question: What can we know and know?

Dreyfus: From fame to infamy

Dreyfus was taught by the master of analytic philosophy Quinn, but he has always been fond of continental philosophers such as Heidegger and Merleau-Ponty, and his research can be described as the best of both continental philosophy and analytic philosophy. His works cover a wide range of topics, and Chinese scholars can already read "What Computers Can't Do", "On the Internet", "In the World" and so on.

As a philosopher, Dreyfus's brightest label is "anti-AI," a label that derives from some of the events in the history of ideas he has been involved in. For example, he constantly criticized early AI scientists such as Marvin Lee Minsky and Herbert Alexander Simon; Another example, in his phenomenological and existential research, he often satirizes "computational reason"; For example, he used the phenomenological tradition to reflect on the cognitive logic of AI, wrote "Artificial Intelligence and Alchemy", and provided negative opinions to the RAND Corporation, predicting the inevitable failure of this emerging discipline in the future. These events made him famous, or notorious.

Some of his books, some yet available to Chinese readers, include Mind over Machine, one of Dreyfus's masterpieces of "anti-AI," written in 1986 by philosophy major Hubert Dreyfus and mathematics major Stuart Dreyfus, which is a systematic summary of his early thought. The book reflects on the early practice of artificial intelligence such as "expert systems", and the title expresses his general attitude, "Mind over Machine", that human intelligence is superior to artificial intelligence, in a rhyming manner in English rhetoric. Dreyfus's point in this book can be summed up briefly as follows: in areas such as "chess" and "face recognition", machine intelligence is bound to fail. As far as specific prophecies are concerned, there is no doubt that history has proven Dreyfus's prophecy to fail.

Why did Dreyfus fail?

If we want to boil down its underlying causes, we need to observe an important mathematical paradigm shift—AI methodology from symbolism of expert systems to machine learning artificial neural networks, completing a holistic evolution, a change known as the "statistical revolution". The statistical revolution in the history of artificial intelligence refers to AI innovation based on artificial neural networks and machine learning, rather than on large knowledge bases and symbolic deduction, which came into being. Rosenblatt proposed the perceptron model in the sixties of the twentieth century, but it was not until the eighties of the twentieth century that the AI winter ice and snow melted, artificial neural networks ushered in a boom in academic discussion, and it was not until the twenty-first century that it reaped many specific commercial results.

From the perspective of ideological foundation, the "statistical revolution" stems from the conceptual change from symbolism to connectionism. In short, symbolism advocates the simulation of the human mind through knowledge accumulation, rule setting, logical deduction, and symbolic operations; Connectionism, on the other hand, completes the fitting of judgment and behavior through the accumulation of a large amount of data and the learning of artificial neural networks. Based on this theory, computers have made important breakthroughs in the field of chess. Implicit in the "statistical revolution" is a radical subversion of Dreyfus's early predictions, beyond Dreyfus's old imagination.

Dreyfus's specific predictions failed, but did his philosophical foundations also falter?

Science or alchemy?

Let's take a closer look and avoid Dreyfus's failure in specific predictions for a moment and consider the question: Is Dreyfus's underlying logic similar to the later "statistical revolution"?

At the level of epistemology, Dreyfus's core idea is that there is an "intuitive expertise" that is distinguished from formal knowledge, symbolic knowledge, and computational knowledge. In "Five steps from novice to expert," Dreyfus cites the example of a "bicycle" to illustrate his idea. Many of us ride bikes, know how to get on the bike, start riding, and ride to our destination, and we know how to maintain balance, adjust our rate, and dodge obstacles and vehicles along the way. However, few people can give a standardized set of cycling-related rules and expertise. Here, Dreyfus makes a major distinction: knowing how is different from know-what.

At the beginning of learning a new skill, we must first learn to recognize all kinds of objective things, to understand their characteristics, to try to remember complex rules, which are not related to the situation at the moment, and for us, their level of abstraction is no different from a foreign language or a bunch of mathematical formulas. Conversely, we don't know why we take things for granted, but not knowing why doesn't mean we can't do it well.

Starting with the bicycle example, Dreyfus extended by arguing that many of the things that humans are good at dealing with follow this rule, such as chatting and talking, how to walk, etc., and we are accustomed to it, so we also turn a blind eye. Rules are only remembered when these day-to-day problems arise, for example, when you drive in the wrong gear, when you chat and say the wrong thing.

In general, Dreyfus believes that learning does not come from calculation, not from logic, not from the systematic knowledge reserve of experts, but from "learning", including fixed response patterns formed in long-term practice. So is this similar to machine learning? A closer examination of the case of learning bicycles and thinking about Dreyfus's critique of symbolic cognition shows that Dreyfus seems to have completed the transcendence of symbolism from another perspective.

In "Logical machine and i t s l imi t s", Dreyfus sums up his idea by saying that the machine mind is symbolic, calculated, formal, rule-dependent, detached from the big world scenario, and operates itself in the micro-world; The human mind, on the other hand, is non-symbolic, non-formal, non-computational, and forgettable rules, and depends on the social-historical practices and lifestyle of the big world. This line of thinking is also the crux of Dreyfus's thinking. All computers, machines, or artificial intelligence, no matter how far they develop, as long as they are symbolically dependent, then they cannot get rid of the nature of logical machines and reasoning machines.

Of course, we need to be wary that we cannot force ourselves across disciplinary boundaries and think that Dreyfus's philosophical foundation is the same as the mathematical foundation of deep learning; or presumptuously believe that Heidegger's ghost was accidentally discovered by scientists in the eighties and nineties; Or more extremely, the failure of AI is due to scientists not understanding Heidegger. However, Dreyfus refutes that the underlying logic of the "expert system" and the underlying logic of the progress of artificial intelligence do have something in common if they are expressed in language that everyone can understand and become common sense of the public.

The problem began to get complicated, Dreyfus failed and succeeded, whether he liked to call him a "scientific prophet" or not, his predictions should no longer be labeled as "alchemy" and "crazy words", in fact, he did think seriously and seriously, and came to enlightening conclusions.

Dreyfus's Place—The Way Humanities Scholars Talk About Science

The objects of study in the natural sciences and the humanities may occasionally overlap, but in most cases, their methods are very different, the way of speech is very different, the natural sciences are the opponents of the humanities methodology, and there is a very tense relationship between them. Therefore, humanities scholars have to confront this opponent, think about it, study it, and respond to it. If we do not talk about early philosophers and scientists such as Thales and Pythagoras, or all-rounders such as Leibniz, Descartes, and Russell, those pure humanists have never stopped discussing and reflecting on science.

To summarize roughly, there are roughly four types of ways to talk about it. The first is that humanists reflect on the shortcomings of modern technology, the alienation of mechanized mass production on people, and under this line of thought, a long list of names such as Marx, Weber, Horkheimer and so on can be listed. Second, from the general methodology, talk about the limitations of scientific thinking and ignore irrationality, such as Bergson in France and Xiong Shili in Neo-Confucianism in China. The third type is a relatively neutral type, in which humanists use modern sociological methods to conduct fieldwork and sociological analysis of scientists, such as the work of the French thinker Bruno Latour. The last is the positive type, where humanities scholars use scientific methods to restore and deconstruct traditional humanities problems, such as digital humanities using statistical methods.

Therefore, it is of little significance to talk about the debate between humanities and science in general, because humanities studies have long been divided into multiple discourse types. Of the four initial categories, the first two are more likely to attract resistance from scientists because they have something in common with Dreyfus and both show the tension between science and the humanities:

In the first negative type, Marx's concept of "alienation" has now become a popular word among literary and artistic youth, and people began to work in factories, mechanized control, alienated into things, machines, and parts, this concept shows the powerlessness, helplessness and inevitable incomplete fate of individual people in the face of machines. Through the elaboration of the Frankfurt School, this question has been further reflected. In The Dialectics of Enlightenment, Horkheimer and Adorno, representatives of the Frankfurt School, position the fundamental feature of modernity as the impulse to pursue universal science. They believe that the Enlightenment has a tendency to "unify" and "rationalize" everything, and that what cannot be counted and digitized is superstition, witchcraft and human illusion that should be abandoned, and this impulse will inevitably lead the Enlightenment to its opposite, no longer giving light to the world, but making the world darker and darker.

In the second type, many thinkers, such as Heidegger, Bergson, Merleau-Ponty, Deleuze, etc., are reflecting on the limitations of the underlying logic of science. For example, Bergson has pointed out that pure logical forms of thinking cannot illuminate the true nature of life. For example, Deleuze proposed a rhizome epistemology, in which the relationship between things is not based on binary opposites, but is interconnected, changing, and diverse coexistence, which has the characteristics of connexion, hétérogénéité and multiplicité, which is the opposite of the binary tree in computer science.

Of the first and second types, the second type is easily considered pseudoscience because it reflects on specific scientific ways of thinking, and in fact, some people have already labeled the thinking of the French philosopher Deleuze similarly. However, although Dreyfus occasionally criticizes the alienation brought about by the Internet and reflects on the cognitive contributions of science as a whole, his approach is different from the above type, so he is more likely to be considered "pseudoscience" or, more intensely, "pseudoscience" within "pseudoscience". After all, when Heidegger's esoteric thinking about existence was criticized by Carnap as an "invalid metaphysical statement," Heidegger simply clings to metaphysical positions. When Heidegger talks about science and technology in The Age of the World Image, he only thinks in general, calling it the core characteristic of our time, and does not concretize, posing as a technologist, discussing the future direction of a specific technology.

This is what makes Dreyfus unique: he is discussing specific scientific progress, specific technical fields, and even predicting the future of a certain scientific field, the future of a specific technical detail, he does not reduce philosophy through science, he uses philosophy to restore specific science, he is reverse analytic philosophy. He talks less about ethical issues, nor does he discuss the general crisis of science in general, let alone a fieldworker who goes deep into the laboratory, but based on the philosophical background of phenomenology, he specifically discusses the limitations of thinking and technical dilemmas of a certain scientific and technological progress (expert system).

4. Be wary of prophecy or wary of arrogance?

In fact, one should also note Dreyfus' dual identity. First of all, he is an unquestioned philosopher. Second, he is also an early AI scholar, and he belongs to this exploratory and pioneering group. When Dreyfus's prediction failed, then, it seems inappropriate to immediately kick him into the camp of philosophers as a failure of philosophy, and if one considers his other identity, then it is also a failure of early AI scholars, although they differ in specific views.

When we jump beyond Dreyfus and re-examine the predictions of other early AI scholars, we can see that Dreyfus's predictions did not fail at the beginning, and he was not a "metaphysical ghost" in the beginning, not an alchemist. For a long time in the second half of the twentieth century, he was the undoubted victor in the arena of prophecy in which he competed against many scientists.

Let's go back to the moment when Dreyfus came to the stage of intellectual history, when the predictions of Marvin Minsky, Sima He and other famous early AI scholars failed. Since the beginning of the Dartmouth Conference in 1956, the idea of artificial intelligence has excited scientists, who are full of youth and confidence in a discipline and the entire scientific faith, believing that artificial intelligence will quickly succeed, Marvin Minsky asserted that machines can quickly complete all human tasks, and they once thought that "the human brain is a machine made of meat." And for them, it's a kind of self-fulfilling prophecy, that is, they make predictions, and at the same time use their own efforts to develop early products such as expert systems, and then use their efforts to continuously fulfill the predictions.

It was at this time that Dreyfus poured cold water on this matter, and even affected the investment in artificial intelligence in the US government and business community. RAND has an important position in the history of artificial intelligence development, due to the decline of Dreyfus's "Alchemy and Artificial Intelligence" article, but also due to the imbalance of the input-output ratio of many early artificial intelligence practices, artificial intelligence ushered in its own winter, and the early artificial intelligence practice with expert systems as the core quickly cooled, Dreyfus's prudence and conservatism made him the winner of the prophecy competition again and again, and his articles became a precious voice in such a gesture.

In this context, we can revisit prophecy. "Prophecy" is a unique act of human language, a way of trying to transcend dimensions, at a certain point in the arrow of time, the process of turning a one-dimensional point into a two-dimensional vector, so as to transcend specific limitations, a way of comparing oneself with God, a way of human beings acting for God, trying to accurately cut the picture of the future with language; Prophecy is often limited to the characteristics of the times, because there is no time-penetrating, history-independent perspective of God: people will always predict what computers can and what computers cannot, based on their own time. And this process is inevitably accompanied by prediction failure, because the scientific and historical framework has undergone earth-shaking innovations.

Assuming that the progress and development of AI technology is an overall incremental function, then its current value range (the tasks that the machine can do) is (0,100), then the prophet is very likely to use (101-105) as an example to demonstrate the limitations of the machine. As everyone knows, with the development of technology, when they are realized, this field quickly becomes a white bone field for the failure of predictions.

If Kurzweil's distinction between linear progress and exponential progress is introduced, it seems that the problem of predictions about artificial intelligence can be better understood. In The Singularity Approaching, Kurzweil points out that exponential progress, once the singularity approaches, the old way of growing quickly iterates into a new model of growth of another magnitude. This shows that Marvin Minsky is prematurely judging the coming of the exponential growth rate at the beginning of the exponential function; Dreyfus belongs to the beginning of the exponential function, which is regarded as a linear function, and even believes that when this linear function touches the complex human ability, the growth rate will gradually slow down, or even stagnate, becoming a function of constant range.

It can be seen that the probability of failure is very large. Given such a high probability of failure, there should be no double standard: when scientists predict the failure of artificial intelligence, the scientific community will consider failure to be the mother of success, and early hasty practice and thinking should also be regarded as an important part of the history of scientific development. And when philosophers predict the failure of artificial intelligence, it must be the nature of philosophical thinking that cannot predict scientific problems, and it should be labeled as a metaphysical ghost. Such an approach is clearly unreasonable. In fact, both philosophers and scientists, when a field is still full of unknowns, it is valuable to think about a problem based on the ideas of the discipline.

It is useful to revisit Carnap's reflections on prophecy, which he pointed out: "In many cases, the laws contained are not universal but statistical laws, so that prophecy will be only probable." At the same time, however, "prophecy is essential in everyday life, as in science, on which even the most trivial activities we accomplish on a daily basis are based."

Philosophers have no problem talking about science, even in specific scientific fields, and they should promote true knowledge by learning from each other and arguing within the community, even if the prediction is wrong, the value of their thinking cannot be erased. It is not prophecy, not interdisciplinary prophecy, nor philosophical prophecy that needs to be wary, but arrogance. In fact, Dreyfus's notorious cause, and what provoked the wrath of scientists, is not only phenomenological methods—his arrogance, arrogance, and arrogance, his illusion of philosophical dominance.

Science is a field that breeds arrogance and arbitrariness, and philosophy is even more so, and we should encourage Dreyfus to prophesy science, even if it is by philosophical methods, to achieve mutual benefit between philosophy and science. But it is also necessary to oppose their arrogance towards science, where philosophy and science need to construct a mutually beneficial constraint on arrogance, especially for humanists who tried to legislate everything with philosophy but had long lost the era in which philosophy could legislate for all.