Translated material
Go through the Turing Wall
Could there be intelligent machines that can play board games or recognize images only in their cozy virtual world? To become reliable and convenient assistants to people, machines need to learn to communicate and act in physical reality, like people. An article by Albert Efimov , Candidate of Philosophy, Vice President, Director of the Research and Innovation Department of Sberbank, Head. Department of Engineering Cybernetics, NUST MISIS.
The idea of an artificially intelligent humanoid being is much older than science fiction fans believe, who attribute the idea of robots to Karl Capek, who coined the term in the 1920s. The legend about a man-made creature that can do everything that a person can do comes from deep antiquity. We read about this in the parables of Zhuang Tzu from Ancient China, in the ancient Greek myths about Pygmalion and the “smart” tripods - assistants to Hephaestus. thought seriously about the “automation” of reasoning Aristotle , describing syllogisms - forms of logical constructions that serve as the elementary “building blocks” of rational thinking. Another important milestone on this path was the work of Gottfried Leibniz , who at the end of the 17th century not only laid the foundations of mathematical logic, but also talked a lot about the possibilities of algorithmic thinking.
At the end of the 17th century, Leibniz described the concept of a ratiocinator - a logical system that allows any derived concepts to be expressed in a clear and simple way based on basic elementary concepts and strict rules, performing mathematical-like operations on them. The idea of creating such a “philosophical machine” was grandiose, but due to the underdevelopment of technology it remained unrealized for many years. The first significant step towards this was taken only in the 1830s, when Charles Babbage tried to construct the Analytical Engine , the mechanical prototype of modern programmable computers. The result was not very successful, but attempts to “mechanize” reasoning did not stop.
In the 1930s , Kurt Gödel formulated and then proved incompleteness theorems, according to which no system of formal arithmetic can be both complete and internally consistent. In other words, there is no system that allows you to prove or disprove any statements. For some time this puzzled many experts. But soon Alan Turing and Alonzo Church introduced the concepts of computable (solvable in a particular system) functions and showed that all of them can be solved not using formulas, but algorithmically - for example, using a Turing machine .
Turing's thesis, in its simplest formulation [1], states that a universal Turing machine can perform any computation that a human computer can perform. This idea, amazing in its simplicity and depth, paved the way for the emergence of the first computers, which Turing himself worked on during the Second World War. At the same time, the British scientist thought about creating an “intelligent machine” - the term “artificial intelligence” had not yet appeared.
Chess and cipher
Many experts today believe that before starting to create artificial intelligence, the nature and structure of natural intelligence should be clarified. However, Turing saw the problem completely differently. He, in fact, inherited the ideas of Rene Descartes, who considered living organisms to be completely automatic beings, believing that full consciousness and thinking are characteristic only of humans. In such a picture of the world, the human mind, his intellect, are separated from the real world as part of another “sphere of consciousness.” Likewise, in the ideas of Alan Turing, intelligence was practically independent of its physical carrier.
In his famous 1950 paper, “Computing Machinery and Intelligence ,” he identified several areas representing the “highest expressions” of human intelligence that future modeling should focus on. These are language learning (and translations), games (chess, etc.), mathematics and cryptography (including riddle solving). If in these areas the activities of a computer cannot be distinguished from human ones, Turing reasoned, then their thoughts are equivalent and we can say that we are dealing with an “intelligent machine.”
Did Turing really think that the most important thing in a person is to be able to play chess, conduct sublime dialogue, or solve cryptographic riddles? This is wrong. The British mathematician was confident that to create intelligent machines with abilities comparable to humans, it is not enough to simply teach the machine to interact with the physical world. In his 1948 report to the National Physical Laboratory, Turing wrote that such a machine " would not have the power to appreciate things so important to man as food, sport, or sex ." Recognizing that by developing a machine capable of interacting with the real world, we were moving toward a more assured creation of artificial intelligence, Alan noted that in 1948 this path seemed longer and more expensive than teaching a computer to play chess.
It is worth saying that within a couple of years after this publication, “ turtles ” [3] Walter Gray’s appeared - one of the first autonomous machines of our time. Very primitive creatures demonstrated unexpectedly “intelligent” behavior and could, for example, find their charging station using phototaxis, guided by light. This complexity was born in the direct interaction of the real world with the simplest “consciousness” of robots, and if Turing had written his article later, he would probably have formulated the problem differently.
However, it was an article written more than 70 years ago that set the conceptual framework for the work of many generations of artificial intelligence developers. According to the Turing approach, high-level intellectual functions of the brain can be reproduced on the basis of an artificial system, a computer, without imitating a system in the physical world. The test described in the same article also developed such ideas.
Inside and behind the wall
In his thinking about the Turing test, he drew on the Victorian “imitation game.” According to its terms, the presenter, exchanging notes with the players, must determine which of them is a woman and which is pretending. Of course, the “judge” does not see them; he is separated from the players by a wall, impenetrable to everything except the exchange of symbolic information - notes, or, in modern terms, chat messages. Such a test can be considered as an “intelligence test” for a man who is required to imitate “feminine” (of course, in the ideas of Victorian times) behavior and reactions. The Turing test transferred this situation to a game with a computer, which must imitate a living person, remaining hidden from the judge by the same “wall”.
This wall seems to be a mandatory element of the test, because without it we will immediately see who we are dealing with. It hides the physical reality of the interlocutor, but at the same time reduces all his thinking to a limited set of processes. At the same time, even the hero of our story himself admitted that full human knowledge of the world is impossible without direct interaction with it. However, at that time, simulating tasks such as sports, food or sex seemed completely unthinkable, so the British mathematician relegated them to the indefinitely distant future, proposing a focus on games, languages and cryptography.
As a result, Turing launched a kind of race between man and machine exclusively in virtual space. At the same time, the idea of such a test stimulated the development of systems that perform certain, narrow functions - be it playing chess, translating or driving a car - better than people and even ready to replace us in one area or another. The narrowness of the capabilities of an intellectual machine was initially inherent in the paradigmatic idea of the British mathematician, limiting intelligence to only simple verbal, symbolic communications and ignoring all other modalities. Can intelligence that can play chess, chat, and solve riddles be called general? Hardly.
However, as long as a robot or computer remains separated from a person and from the world by a wall, it is not able to fully interact with them, and the true intelligence of the machine is replaced by the complexity of the functions it implements. Perhaps this is enough for a self-driving car or a chess program, but in the pursuit of general artificial intelligence, a paradigm shift is no longer necessary: we will have to “break the wall” and move to a new, post-Turing methodology [4]. Its essence is that all the elements of the “test” mentioned by Turing form a single whole and are considered as a whole: the observer-judge, the test subject (human or computer) and the questioning tool (the wall turns into an interaction environment, an interface between a machine and a person).
Verification by correspondence
For an easy explanation, you can turn to the “penpal girl” thought experiment proposed by [5] Andrey Alekseev . Let's say a young man, looking for a partner, turns to an Internet dating service. Having specified the appropriate parameters (age, lifestyle, etc.), he receives a list of users and begins correspondence with one of them. After a long virtual conversation, the young man finally asks her out on a date, only then discovering that he had been talking to the program all this time. This development of events is equivalent to the successful passage of the Turing test in its classical version by artificial intelligence. Expecting that the development of technology will continue within the framework of already existing trends once set by Turing, Alekseev suggested that “ in the near future, the “pen pal” scenario will also come true . ”
However, it can be noted that even if this scenario comes true, it does not make any practical sense. Correspondence “through the Turing wall” deprives the machine of the possibility of comprehensive useful interaction with humans and the world. To explain this, let us imagine a different scenario for the development of events in the same experiment. Let's say that when it comes to a date, the young man meets his pen pal in a cafe - she is quite alive and real. However, the conversation offline does not go well: it turns out that they do not have many common interests, and the witty and appropriate remarks that the girl gave during correspondence were automatically suggested to her by artificial intelligence. The disappointed young man returns home and writes to her that such a meeting makes him a little uncomfortable, but she responds with a quote from his favorite TV series, and the correspondence resumes.
Thus, the wall separating machine and man no longer exists and is not needed to evaluate artificial intelligence and its interaction with people. At the same time, the computer turns out to be closer and more understandable emotionally than even a live interlocutor; it is “more human than the person himself.” In this regard, one can again recall Plato with his “eternal ideas”, the realizations of which are real objects. Artificial intelligence, not limited by the Turing wall, can embody the “idea of intelligence” in the same way as a person himself does, on an equal basis and in interaction with him. A person knows who he is dealing with and understands that he is better off with a car: more interesting, more useful, more reliable.
After Turing
A real thinking machine must become the product of diverse interactions with a person and the world around him: verbal and non-verbal, occurring both in a virtual environment and in a real one. Thus, the classic Turing test covers only the areas of verbal and virtual interaction, as do Terry Winograd's circuits and most other popular tests of artificial intelligence. This is not surprising, because they are all in the paradigm set by the British genius, “behind the wall.” To break it means to enter the realm of non-verbal and real mastery of the world by artificial intelligence.
After all, we now understand that many animals have some form of consciousness, including even cephalopods. And every time thinking and its manifestations turn out to be connected with the real conditions of existence, with the physicality of a living being, its motor skills. According to our outstanding philosopher David Dubrovsky , the development of the psyche affected [6] only those organisms that actively move in the external environment. It seems that full knowledge of the world around us is, in principle, impossible without physical interaction with it. Therefore, the condition for creating “general” (“strong”) artificial intelligence will be the ability to work in different modalities and different environments. He needs an outlet in the areas of non-verbal and physical.
Examples of artificial intelligence that cope with non-verbal tasks include existing systems that play computer games, or the virtual TV presenter Elena [7] (created at the Sber Robotics Laboratory, she completely imitates a real TV presenter, including movements, emotions and gestures). However, both the first and the second do not leave the limits of the virtual. Real human interaction in the physical world is still extremely challenging. For general artificial intelligence, this is not enough: such a machine must cover all four areas of interactions and environments.
The Coming of the Techno-Umwelts
Back in the 19th century, the prominent biologist Jakob von Uexküll noted that different living beings are characterized by different spheres of worldview - umwelt . The umwelt of a butterfly is noticeably different from the umwelt of a fish or a person. By analogy with them, the four areas of interactions possible for machines are proposed [8] to be called technoumwelt . Technoumwelt is the domain of worldview, the way a machine sees the world around it. The human umwelt is familiar to each of us, and many have observed the techno-umwelt of unmanned vehicles using radars and lidars in videos.
The dimensions of human-machine interactions described above (verbal - non-verbal and virtual - physical) give four independent techno-umwelts: verbal virtual, non-verbal virtual, verbal physical and non-verbal physical. The versatility of artificial general intelligence (AGI) is only possible if the machine can move freely between all four techno-umwelts. The current generation of AI is capable of recognizing objects of different classes without prior training. This is a major achievement, but it has nothing to do with the ability to work in different techno-umwelts. Apparently, this will require the implementation of some kind of “translators” from the language characteristic of one domain of world perception into the language of another. Only then will artificial intelligence become truly multimodal, be able to solve the entire range of possible problems and fully “communicate” with humans.
Its appearance will forever change the way we interact with technology. After thousands of years of philosophical reflection and scientific and technological progress, people for the first time in history will be faced with truly “smart” things, with devices that may have even more complete and accurate knowledge about the world and about us than we do ourselves. Such a situation will require a new look at what a person and his mind are, redefining many established ideas.
These processes have already begun: today we are beginning to “dissolve” in the technologies and gadgets that surround us everywhere, and the very concept of “human” is blurring. As computers master new areas of activity, be it chess or translation, they cease to be considered the exclusive prerogative of humans. Perhaps a person is something that a machine is not yet capable of doing. Human engineers can create a machine that can independently get from point A to point B, but you need to become a philosopher to see the place where this point B is located.
2. ru.wikipedia.org/wiki/Вычислительные_машины_и_разум
3. en.wikipedia.org/wiki/Turtle_(robot)
4. intellekt-izdanie.osu.ru/arch/2020_2_74.pdf
5. artsoc.jes.su/s207751800007698-6-1/
6. ncbi.nlm.nih.gov/pmc/articles/PMC7179338/
8. iphras.ru/uplfile/zinaida/ROOTED/aspir/autoreferat/efimov/dissertatsiya_efimov_final_151020.pdf