Jeffrey Khinton's “Cross -Father” quit Google, criticized the company for inequality - and now warns the world about the danger of artificial intelligence

Here is his story

Among those who openly asked questions about the ethical use of ChatGPT, from the very beginning was the developer of neural networks, the British-Canadian scientist Jeffrey Hinton. He made a significant contribution to the development of chat bots with artificial intelligence, but in recent months has turned into a fundamental opponent of the improving use of AI. In the spring of 2023, he left Google. "Medusa" talks about what Hinton was nicknamed the "godfather of AI" and why he criticizes the technologies that he had recently helped to create.

Hinton left Google and opposed chat bots, because because of them ordinary people "will no longer know where the truth is"

In 2012, professor at the University of Toronto Jeffrey Hinton and his two students developed a system that could analyze thousands of pictures and teach themselves in reality similar objects - for example, flowers, animals or cars - with unprecedented accuracy. After the outstanding achievement, Hinton and his students, Ilya Sutskevers and Alex Krizhevsky, continued the study - and the company in which they worked on neurosetates acquired Google. It was the development of the British professor and two students that accelerated the introduction of AI-and led to the appearance of ChatGPT , Google Bard and other chat bots .

In 2018, Hinton received the Alan Turing Award with the wording "for conceptual and engineering breakthroughs that made the deep neural networks with a cornerstone in computing technology." Five years later, the scientist sharply changed his eyes on AI, left Google - and on May 1 gave an interview to The New York Times (NYT), in which he spoke about dangers in this area. The main threat of chat bots and other similar technologies, according to Hinton, is that the Internet is so filled with fake content-generated photos, videos and texts-that ordinary people "will no longer know where the truth is."

The scientist also expressed fears that AI-technologies will eventually turn from assistants to people into their replacement and leave representatives of many professions without work. “Some people believed that technology could become smarter than us, but most considered it an exaggeration,” Hinton said in an interview with Nyt. - I also thought that it was still far away. I thought that we have left from 30 to 50 years or even more. Obviously, I don’t think so anymore. ”

Hinton developed the ideas of Turing and dreamed of creating an artificial system that would learn how the human brain

Hinton was born in 1947 in the family of a respected entomologist who lived in Wimbledon (suburb of London). His great -grandfather George Bul developed a mathematical logic , and the Algebra named in his honor laid the foundations for many technologies that developed in the information era .

To the disappointment of his father, Jeffrey decided to study not insects, but people. In 1970, he graduated from the University of Cambridge with a degree in experimental psychology, but then paused in academic activity and was engaged in carpentry for a year. “Science tired me, and I decided that I want to be a carpenter,” Hinton explained . In 1972, Jeffrey became interested in the artificial intelligence program at Edinburgh University, left the dream of a carpenter’s career and moved to Scotland. There he plunged into the development of neural networks.

“Then not a single person on earth believed in this idea,” said Hinton. “She was considered a dead end even among AI researchers.”

However, the development of a young British attracted the attention of a group of cognitive psychologists from San Diego. Hinton went to the United States and began to explore the method of reverse error spread - an algorithm that allowed neural networks to study and improve. Later, the methods in the creation of which Hinton took an active part was called deep training. They allowed AI not only to perform specific actions within the framework of one algorithm, but also to expand their capabilities using external data.

“The idea was to get a device that would study the same way as the brain study,” said Hinton. - It was not my idea, Alan Turing thought of the same. He believed that the brain is an unorganized machine that uses reinforcement training (a method of machine learning, when the system studies, interacting with the environment - approx. Medusa) to change connections inside, and which is able to learn anything. He believed that this was the best approach to the study of intelligence. ”

The desire to develop a system that could study the same way as the human brain determined the entire career of Hinton. In the late 1970s and in the 1980s, he actively tried to create algorithms that could recognize sounds and images. The key role in the projects of scientists of that time was played by internal ideas, that is, information structures that the perceiving system creates and preserves as combinations of qualities inherent in external objects.

“The question of how to create internal ideas was considered the Holy Grail of the AI,” Hinton explained . However, computers half a century ago did not have a sufficient speed to generate internal representations as quickly as the human brain.

Since 1982, Hinton has held a professor’s position at the University of Carnegie - Mellon in Pittsburgh, but the policy of the American president Ronald Reagan was increasingly annoying the British. He did not like that the majority of AI's studies in the United States were sponsored by the Ministry of Defense, so in the mid-1980s Hinton moved to Canada, where he was offered positions at the Canadian Institute of Promising Research and Toronto University.

There, Hinton continued the study, despite the fact that many of his colleagues by that time were disappointed in the idea of ​​creating a student and engaged in other projects. “In the 1990s, data sets were not large enough, and computers were still not fast enough,” said Hinton. - With small data sets, other methods worked better. They did not distract the abundance of information. This was very depressed, because in the 1980s we developed a method of reverse error spread and thought that we solved all the problems. We were worried why we still could not use this technology. ”

The puzzle with artificial systems was decided with the advent of faster and more powerful computers in the 21st century. Hinton and his colleagues managed to introduce algorithms into artificial systems, which allowed them to form internal representations. The II also supplied speech patterns that allowed to perceive and interpret oral speech and written texts. The detectors of signs embedded in Hinton’s systems processed fragments of information and found connections between them, so that the learning process was noticeably accelerated.

Hinton had to engage in research due to health problems. He damaged his back as an adolescence, when he transferred the heater for his mother. Since then, the sitting position threatened the scientist hernia of the intervertebral disc, so he was used to constantly standing. “The last time I sat in 2005,” Hinton admitted in 2012. “And this turned out to be a big mistake.”

“Other scientists also had similar ideas, but he always found himself in the center of events,” the journalist of NYT and writer Kade Metz explained the contribution of Hinton to the study of AI. - He actively participated in the creation of systems that they learned to recognize first speech, and then the image. These two points have become key for technological progress and as AI is used in our time. ”

Two years ago, Hinton developed a new system, but changed the opinion of AI due to the technological race of Google and Microsoft

After the development of the neural network with unique opportunities for deep education in 2012, Hinton founded the company Dnnresearch and hired students Ilya Sutkeever and Alex Krizhevsky, who together with him were engaged in the research of AI. Despite the modest status of the company with only three official employees, the struggle broke out between large corporations for the opportunity to absorb Dnnresearch. The Chinese technological giant Baidu, Google, Microsoft and the British startup Deepmind of the British neurobiologist Demis Hassabis, launched only two years earlier, claimed cooperation with Hinton.

According to the results of the auction, Dnnresearch has passed under Google control for $ 44 million. Despite the readiness of Baidu to offer a larger amount, Hinton himself decided to stop the auction, because he did not want scientific cooperation to turn into a farce. From then until May this year, the “Cross Father of AI” continued to develop more advanced neural networks in Google.

In 2021, Hinton focused on creating a GLOM system. According to Hinton, the new development was to solve one of the main problems associated with AI, since it would have the ability to recognize the same objects from different angles. Prior to this, despite the rapid development of neural systems and the technologies associated with them, the algorithm, for example, could not recognize the same cup if it was depicted first from the side, and then from above. Similar tasks continued to motivate and inspire Hinton, despite the status and position, which he has achieved in almost a half -century career.

Another idea of ​​Hinton was to provide neural systems with “consent vectors” - algorithms that would analyze the information in a similar way, but complement each other and helped to make a more complete picture. The scientist compared a similar approach to a group of people engaged in the brainstorming: they all can agree with each other, but during the discussion, they will be able to improve individual internal ideas with a high degree of probability.

“Jeff is a very unusual thinker in the sense that he is able to combine complex mathematical concepts with biological factors and, based on this regard, formulate the theory,” said the neuro -scientist and former student Hinton Sue Becker. - Researchers of a narrower profile concentrate either on mathematical theory or on neurobiology. They have much less chance of finding the answer to the question of how people and cars can study and think. ”

When Google, Openai and other technological companies engaged in the research of AI began to develop neural networks that were studied on the basis of extensive fragments of digital text, Hinton suggested that now the cars will have more chances to recognize and generate languages. However, this method of learning, in his opinion, was still inferior to how people recognize the text and speech.

Hinton's position changed in 2022, when Google and Openai developed systems with much larger databases, than earlier. Prior to this, the scientist was proud that Google acts carefully and does not allow technology to enter the market that theoretically able to harm humanity. But in the last year, the competition between Microsoft and Google pushed both companies to release cities, the capabilities of which, according to Hinton, are not yet fully realized by the developers themselves. According to the scientist, in some aspects, neural networks have become more perfect than the human brain.

“There is a possibility that what is happening in these systems is far exceeded in complexity processes in the human brain,” the scientist noted . - Look at what happened [in the field of AI research] five years ago, and what is happening now. Imagine at what speed the changes will occur in the future. It scares. " Most of Hinton’s colleagues still consider the threat of hypothetical, but he himself is sure that in order to avoid disaster, you need to take the strict control of Google, Microsoft and other technological giants.

Hinton fears that technology will eventually erase the border between fiction and reality for most people, as well as dramatically change the situation in the labor market . Another reason for the scientist’s fears is that artificial systems can learn unpredictable behavior when extensive data arrays are analyzed. This means that it will be increasingly difficult for people to predict the mechanisms of AI function.

In an interview with NYT, after leaving Google, it is stated that Hinton regrets the work of his whole life, but comforts himself with an excuse: “If I did not do this, someone else would have done it.” In a later conversation with Mit Technology Review, the scientist clarified :

[Nyt journalist] tried to make me say that I am sorry. In the end, I said that perhaps I have small regrets. [But] I do not think that I made any wrong decisions during research [to create neural networks]. In fact, [in the 1970s and 80s] it was impossible to foresee [the current stage of the development of AI]. Until the very last moment, I thought that the existential crisis [due to the threats of AI] was still very far [from us]. In general, I have no regrets about what I did.

"Jellyfish"