By the fall of 2022, Naked Science looked around and decided it was time to publish a series of articles about AI . We start with a long article with a possible subtitle of “A Brief Introduction to Artificial Intelligence History – From Medieval Monks to Deep Learning.”
Artificial Intelligence image generated by the word "Dream Artificial Intelligence Spring"
Artificial Intelligence (AI) - Almost everyone knows these two words, but few people can give a clear definition and explain what it is. Our understanding of AI comes more from Hollywood movies than really understanding the technology behind AI. Movies, like any art, always work with the audience through emotions. The best-selling of these is fear.
001 HAL 9000 AI Computer: Space Odyssey controls the interstellar spacecraft. T-800 Cyborg in " Terminator " returns to the past and kills Sarah Conner. Recent examples – Neurological implants conquer the wearer’s consciousness in the movie “upgrade”. In the movie Ex Machina, Gynoid Ava easily manipulates programmer , invited to perform reverse Turing tests, kills his creator and escapes to the wild. There are countless examples. However, reality is far from the film.
Humans have come a long way from 80 years of searches, errors and dead ends, and each ends with a "winter of artificial intelligence" - disappointed with the ability and potential of this technology. But since the early 2010s, the world has once again experienced a "recovery" in the field of artificial intelligence. So in the cinema, they try to scare the layman—the big corporations and governments in major countries are investing billions of dollars in the development of artificial intelligence, because now it is changing everything—from scientific research to daily life.
However, even the most advanced modern development is far from the possibility of any fantasy of directors and screenwriters. As usual, ideas about the future may coincide in details, but never in general. Let’s understand what artificial intelligence is and briefly understand the major milestones in its development history.
Inventive concepts
American mathematician John McCarthy (1927-2011) first became interested in computers when he started attending a seminar on brain behavior mechanisms in 1948, which discussed whether computers can start thinking like humans. This topic fascinated him so much that later—in the summer of 1956—he organized a ten-week workshop at Dartmouth College (a private research university in New Hampshire, USA) with money from Rockefeller Foundation .
In the grant application, McCarthy stated the objectives of the workshop:
(our) study will assume that every aspect of learning or any other feature of (human) intelligence can be described so accurately (mathematically) that it is possible to create a computing device to simulate it. (We) will try to find a way to make computers use (natural human) language to form abstractions and concepts, solve various problems that only humans can solve now, and improve themselves. We believe that if a carefully selected team of scientists study these issues together in the summer, significant progress can be made on one or more of them.
The team formed by McCarthy is really impressive. There is the creator of information theory, Claude Shannon, the future star of the foundation of artificial intelligence mathematics, and the creator of framework theory, Marvin Minsky, cognitive psychologist Alan Newell , and later developed the logic theorist program, which learned to play chess and solve difficult problems after modification, and there are many, many other talented people.
Use Tesla Autopilot to drive in traffic, which controls the distance from other vehicles on the road and places the vehicle in the middle of the lane. This car is a 2017 Model X 75D with a dark interior.
As a real scientist, John McCarthy and his colleagues start by defining concepts.The first one sounds like this: "Artificial intelligence is a science and engineering activity aimed at creating intelligent ( Intelligent ) machines" . I believe this is the first time the term "artificial intelligence" has appeared in history. Later, as often happens, the names of scientific disciplines are also transferred to the names of their research and design objects—the "intelligent machines" themselves, implemented in the form of physics and algorithms.
More modern versions of this definition may sound different. For example, Artificial intelligence is one of the branches of computer science, whose task is to provide reasonable reasoning and action using computing systems and other artificial devices. Meanwhile, the technology created based on the knowledge gained within the industry is called information technology.
Well, simply put, Artificial intelligence is a field of science and technology that can automatically solve intellectual problems. If people solve a problem with their intelligence, scientists and engineers can create an artificial system to solve it in place of people. This will be an artificial intelligence system.
However, anyone can perform many intellectual tasks such as driving, playing chess, discussing plans for the day on the phone, etc. Sometimes all of these tasks can be accomplished simultaneously. But computers can't do this yet. Therefore, I believe that all we create is a system that is weak (weak) or narrow (narrow), or an artificial intelligence ) that is applied (applyed).
All of them can only complete one intellectual task - driving driverless cars on the streets of cities or playing chess to perfection. And even if the program beats chess world champion , it won't drive. But it’s easy for people to switch from one task to another and can master new skills for life.
weak artificial intelligence systems are countless. In fact, almost all engineers and scientists are engaged in their development work. Nevertheless, the main goal of the industry is to create artificial intelligence powerful ( strong () or general ( general ). Everything is not that simple here, and it is difficult to develop a single definition that satisfies everyone. This is due both to the history of this concept and to the difficulty of understanding humans and their abilities.
In fact, the concept of strong artificial intelligence was first proposed by philosopher John Sell and the concept of "Chinese Room". In short, it sounds like this: If a person who understands Chinese is placed in a closed room and then the questions written in Chinese are sent to him through a special hatch, then he will also write the answers in Chinese based on his understanding of this. But what happens, continue with Searle's thought experiments, if another person who doesn't know Chinese is placed in the same room, but at the same time he is provided with an exhaustive system of rules that allow him to form others in response to receiving some hieroglyphic sequences?
If the rules system is broad enough, a person who does not understand Chinese will give a very meaningful answer without understanding the nature of the problem. In other words, he will imitate and understand Chinese without really mastering it. This is the origin of his concept of " strong artificial intelligence ", which means a system that "acts as if it is intelligent" to distinguish it from a "really intelligent" system.
However, the original explanation of the concept of "strong artificial intelligence" can only be found in philosophical books.For example, IBM gives different definitions:
Strong artificial intelligence, also known as artificial general intelligence (AGI), is a theoretical description of a specific form of artificial intelligence, with human-like intelligence, self-awareness, and the ability to solve a wide range of intellectual problems, learn and plan for future actions .
The easiest way to explain this definition is to take the "coffee test" proposed by the co-founder of Apple as an example. and Stephen Wozniak, the creator of the company's first personal computer. This is simple, but it is still insurmountable for any system.
This image is edited by AI in response to Naked Science
The essence of coffee testing is to test how artificial intelligence can fully interact with people and the real physical environment, and how to successfully navigate in a new environment without prior training. To do this, an AI-controlled robot must be able to find a coffee, coffee machine or coffee machine in any randomly drawn kitchen, which has not been previously and its plan is not loaded into its system and prepare a drink.
There are a lot of food here—all kinds of strange environments where artificial intelligence knows nothing. The goal is to make drinks. Limitation - the existence or absence of means to achieve goals. Indeed, sometimes acknowledging that a problem cannot be solved is more proof of wisdom than repeating it in vain.
Interestingly, some Russian experts say that if robots and algorithms do not interact with the physical environment, just as humans do during ontogenesis , it is impossible to create a universal AI. With the transformation to reality, a qualitative breakthrough will occur.
Many scientists, such as Marvin Minsky, or entrepreneurs like Eron Musk , predict that general artificial intelligence will be created in the next few decades. Instead, others even now believe that it is completely impossible to create it. The greatest achievement will be the development of elements of strong artificial intelligence or narrow general artificial intelligence (narrow AGI), that is, systems that achieve outstanding results on one problem and can solve other problems but are significantly lower than the average human ability.
In the third circle
The official history of artificial intelligence usually begins 15 years before the concept itself emerges—from the 1940s. But prehistoric history can be from ancient times—the automaton of Alexander’s heron and the Anticaisera mechanism—a mechanical device that calculates the movement of celestial bodies. However, all these achievements should be attributed to the early stages of automation and computer technology development, rather than creating machines that can replace humans to solve intellectual problems.
Antikythera Mechanism
It may be fair to start the calculation from the logic machine of Raymond Lull (1235-1315). Ruhr served as a poet for a long time under the king of Aragon , but then transformed his court life into a monk's path. In 1272, during a religious ecstasy, he was visited by a divine vision—a special device that you can use to infer all the truths of the world from a limited number of general concepts.
Ruer described his machine in the paper "Great Art" (Ars Magna). In fact, it is a set of concentric circles whose rotation makes it possible to obtain various combinations of symbols and main concepts (mainly theology).
The simplest car variant consists of three circles. The first contains many Latin letters, the second - concept ("strength", "goodness", "wisdom"), and the third - attribute ("strong", "goodness", "wisdom"). And in the middle of the circle, there is a star-shaped figure connecting all sectors of the circle. In its most complex version, the device contains 14 circles and provides a combination of astronomical numbers.
Ars Magna
Ruhr's thoughts then had a significant impact on philosopher and mathematician Gottfried William Leibniz :
"In philosophy, I found a way to implement Cartesian with the help of algebra and analysis the same thing that others do in arithmetic and geometry... through combinatorial science... can be broken down into A small number of simple concepts, which can be said to be their alphabet , can be obtained from such an alphabet combination of letters, and time of proof of everything and its theoretical.
Alas, in the end, Leibniz failed to create the "alphabet of human thought" and attribute all philosophy conclusions to strict form. But in the process of trying to invent it, he proposed the famous 0 and 1 binary system. Now it is the basis of any computing device, encoding digital data and logical operations.
At the same time, Ruhr's works also attracted the interest of Russian thinkers. In the late 17th and early 18th centuries, Ars Magna was a poet and philosopher Andrei Belobotsky Translated into Russian. In Before the era of Peter the Great, Russia, his two manuscripts, "The Great Science of Raymond Ruhr" and "The Brief Science of Raymond Ruhr" were widely used.
, the father of punching card
, apparently, Semyon Korsakov, the inventor of the first Russian intelligent machine, was widely used. 2019) These books are very familiar. Kosakov invented and described in detail five similar devices for searching and sorting information. The simplest of these is the "linear endoscope with fixed parts". It uses a perforated table, moving a wooden board with spikes of different heights on the columns of the table. The columns required are determined by the coincidence of the holes in the table with the protruding spikes. The other machines he proposed, slightly more complex, work in a similar way: linear endoscopes with moving parts, plane endoscopes, ideoscopes and comparators.
Semyon Nikolaevich Korsakov Intelligent machine punch card
When explaining how endoscopy works, Korsakov cited an example of a medical diagnosis: the columns of the table contain symptoms of the disease, and the doctor sets a spike on the bar based on the symptoms observed in the patient. The column moves along the column until the correct column hits the doctor only chooses the right medicine. "The number of details considered by the device can reach hundreds," Korsakov adds. In fact, this is the world's first doctor expert decision support system!
about a century and a half later, in 1980 In the era, a famous Soviet mathematician and cybernetics , Gly Povalov, who was very interested in the history and prehistoric history of these scientific developments in our country, accidentally discovered Kosakov's publication. Povalov popularized his ideas and made Kosakov one of the founders of domestic cybernetics, and brought him a major world achievement - the first time in the history of computer science to use punch cards.
Early ideas
In fact, the official history of artificial intelligence began with cybernetics. In the initial version, cybernetics was not just that A science is worse than a metatheory that describes how to create and process information in any complex system from biological organisms to multinational corporations.
As a result, cybernetics quickly collapsed, providing fertile ground for the emergence of many fields such as computer science, biology, mathematics, management, engineering, etc. As for artificial intelligence, the longest and most successful artificial intelligence, although it has experienced several crises, has been born in the embrace of cybernetics.
from MIPT Center for Applied Artificial Intelligence Systems - PhysTech AI Expert Mikhail Burtsev From the perspective of Mikhail Burtsev, there are three ways to design intelligent machines that can perform intelligent tasks.
. Neural Network and deep learning training. Physics and software modeling based on single neurons, neural networks and brains in animals and humans.
. Symbol artificial intelligence. Based on modeler's reasoning methods and logical conclusions.
. Evolutionary planning and genetic algorithm . They are based on modeling evolution mechanisms, where algorithm must mutate and survive to obtain better solutions, experiencing the severe pressure of natural selection.
The first two methods have been actively developed since the 1940s to the 1950s. And the direction of evolution and some other directions that emerged from the 1960s to the 1990s have now entered some shadows.
The rise and fall of neural networks
In the early 1940s, two completely different people met by chance at University of Chicago and made a major discovery, laying the foundation for the development of artificial intelligence in the next few decades. Warren McCulloch 42, a successful professor, the son of a businessman, is passionate about finding “psychon” – the basic unit of neural activity. Walter Pitts is less than 18 years old, a prodigy who wrote a letter to Bertrand Russell at the age of 13, a self-study without formal education from a poor, dysfunctional family from Detroit.
However, their cooperation has proved to be very productive. In 1943, they proposed a "formal neuron" model that operates according to mathematical logic.
Most living nerve cells always have many small protrusions— dendrites (through them, signals enter cells from the outside), a body and a large protrusion—axons (through which signals leave cells). All these mechanisms work according to the principle of "all or nothing" - if the input pulse does not exceed a certain threshold, the neuron will not give the answer at all, and if the threshold is exceeded, the maximum possible response is generated.
This is how artificial intelligence neurons imagine themselves (response to edit request "neuron")
"formal neurons" are also threshold elements that run according to the principle of "all or nothing" at strictly fixed time points. It has a limited number of inputs and an output. The inputs are divided into excitability (+1), inhibitory (–1), and inhibitory. The latter can block signals from any other input. If at some point the algebraic sum of the actions received by the input reaches or exceeds the threshold (0), a signal appears at the output of the neuron.
The article they published in MathematicsBiophysicsBulk Titled "Logical Calculation of Thoughts Related to Neurological Activities" caused a sensation. It was translated and published in the Soviet Union 13 years later, and neuronal modeling aroused great interest. Russian cyberneticist Viktor Varshavsky proposes a generalized model of his threshold neurons. A team led by Nikita Pozin created an electronic model of neurons at the Institute of Control Problems of the Academy of Sciences. And the team led by Nikolai Amosov has continuously developed the idea of a network of “formal neurons” to create an “internal information model” of the outside world inside the computer. Unfortunately, due to the lack of foreign translation and weak external connections, almost all of these works are known to the world's scientific and engineering communities.
McCulloch and Pitts believe that "formal neurons" with binary input perform logical calculations, so the brain can be compared to inference machines, which has the greatest impact on Cornell psychologist Frank Rosenblatt. He worked in an aviation lab, but at the same time he dreamed of finding exoplanet for his whole life, and for this he needed an automatic tool for image recognition and classification. He invented it.
In the history of science and technology, Rosenblatt is always the father of perceptrons, and Perceptron is the direct predecessor of modern deep learning systems. Essentially, the perceptron is a neural network of single McCulloch and Pitts to modify neurons. However, this modification is very important - here the connection weight of the input signal to the output block is increased, or, more simply, the effect of each input of the neuron on the value obtained at the output.
weight allows the perceptron to learn by itself.It happens like this: For example, we want the perceptron to learn to determine the letter A and indicate the expected output: +1 means A, and -1 means any other letter in the alphabet. After that, we provide the letter image to the perceptron’s input. If an error occurs, the weight is automatically corrected (for example, if it gives +1 when it should be -1 and vice versa). If the answer is correct, no changes are made to it. Gradually, the perceptron finds the required set of weights and stops errors, clearly identifying letter A.
957, Rosenblatt proved the perceptron convergence theorem. Russian mathematician Vladimir Vapnik also made important contributions to the promotion of perceptron principles. He created the " support vector machine ", which was later widely used in machine learning. With its help, you can automatically compute a plane that most effectively separates two sets of points in Cartesian space (such as photos of dogs and cats), that is, it also solves the classification problem (in this case binary).
is solved by the basic perceptron of the "XOR problem". Threshold for all elements = 0
Despite all the success, the perceptron was not widely used to solve application problems due to the limited capabilities of electronic and computer technology at the time. They then suffered a heavy blow from mathematicians Marvin Minsky and Seymour Papert, who showed in their book Perceptron that it is likely impossible to create this type of multi-layer classifier. At that moment, their considerations (which later proved to be wrong) were accepted by faith. Neural networks were not forgotten until the mid-1980s.
Logistics, heuristics and knowledge
It all starts with the desire to write a chess program. In 1954, future Dartmoor workshop participants Allen Newell and Herbert Simon started working with John Clifford Shaw. In their work, they start from multiple places. First, programs must use empirical methods and rules derived from the analyzing how people solve problems. People engage in cognitive activities in the process of thinking. It includes operations on symbolic representations of objects around the world. Therefore, physical computing systems need to be equipped with knowledge about the world and methods of dealing with it. This is where the history of symbolic AI and heuristic programming begins.
Psychologists immediately participated in the study. They studied the chess styles of successful chess players and how people solve various problems. The team then created their own information processing language (IPL) programming language, which is probably the first language to represent data as a list of symbolic links. Three programs were written on it: the logic theorist in 1956, the general problem solver in 1957, and the NSS chess program in 1958.
"Logistic theorists" are able to prove theorems from elementary mathematics. The variable list is saved in the machine and can be combined into expressions using logical conjunctions - "AND" (connection), "OR" (disconnection), "IF ... THEN..." (hint). Next, axioms and three inference rules are given: replacement, replacement, division.
result, proves to be understood as a series of expressions, where each expression originates from a previous expression, and everything starts with axioms and known theorems and results in the desired expression. This method is called heuristic programming because on the one hand it recreates real human actions with symbols, and on the other hand it cuts off many possible sequences and expressions based on given rules.
result, "General Problem Solver" learned to deal with many difficulties, find indefinite integral and solve algebraic problems. NSS chess is pretty good. However, to do this, it is necessary to add a description of the target and an evaluation of the current and target situations in the control algorithm , as well as their differences.
In the Soviet Union, heuristic programming also attracted the attention of scientists.At the beginning of psychologists exploring thinking, such as Veniamin Pushkin, for them, heuristics became the concept of operationalization and experimental research in human experiments. Only then, a little later - cybernetics and logicians. Russian artificial intelligence expert Dmitry Pospelov, together with Pushkin, conducted a series of research on situational management—making decisions in specific situations, while taking into account many action options. Well, logician Sergei Maslov was the first to propose a method to automatically search for theorem proof in predicate calculations.
Chess program "Kaissa"
At the same time, the development of the Chess program was carried out by Mikhail Botvinnik, the world champion in the history of chess. From the late 1950s to 1990, he created his “pioneer” with a group of programmers. However, there was not much success. This was not the case with the Kaissa project created by the Institute of Control of the Soviet Academy of Sciences. In 1974, she won the world championship in other chess events.
LISP, PROLOGUE and Russian REFAL
IPL laid the foundation for another famous language used in AI systems. In 1958, John McCarthy, the author of the term “artificial intelligence”, introduced the LISP language (LISP-LISt processing) to the interested public. Like all other languages focused on solving intellectual problems, LISP is designed to deal with symbolic data that is not digital, but organized in list form, just like the executable program code itself. But most importantly, at the grammatical level, heuristics are implemented in it, such as reduction, that is, dividing any complex task into simple tasks. Expressions can be used to define functions, while symbols use many brackets, which often make the code difficult to read. Another professional language of
AI is PROLOG (PROLOG - PROgramming LOGic), which was founded in 1972 and is the pinnacle of logic programming. The traditional syllogism "Everyone will die. Socrates is a man. Therefore, Socrates in the preface is a mortal" becomes "In order to prove that Socrates is a mortal, evidence that Socrates is a human being is taken as a sub-target" - from a grammatical point of view, it looks like this:
Mortal (Socrates): - Human (Socrates)
my country has also developed a special programming language for creating symbolic AI systems. Therefore, in 1966, Valentin Turchin introduced the REFAL language ( recursive function algorithm). Today, it is the oldest functional language. Turchin assumes that it will act as an algorithmic meta-linguistic to describe the semantics of other languages, but in fact it is very convenient for processing symbolic information and therefore solving intellectual problems.
Heuristic programming using logic and operations on symbolic information has flourished for nearly 20 years. However, by the mid-1970s, it was clear to all experts that algorithmic complexity and improvements in heuristics had reached their limits. A fundamental breakthrough cannot be expected. The algorithm must be equipped with a huge knowledge base, and technical capabilities do not allow it. Coupled with the decline of perceptrons and other neural networks, this led to the first “artificial intelligence winter” of the 1970s.
Nevertheless, heuristics did not die out and were developed quite effectively until the mid-2000s. Their uses remain limited. For example, the same LISP language is used in the Autodesk product for engineers. The new revival of
neural network
Rosenblatt's perceptron reached its limit in the early 1960s. To solve the problem that is more complex than detecting a single letter, it is necessary to increase the number of layers in the neural network. A sharp question arises: Is it possible to train such a system? The popular view of scientists at that time expressed by the famous Marvin Minsky was negative. Industry insiders all feel that "winter is coming" - artificial intelligence research is beginning to be considered hopeless, with sharp decline in funds, and the topic is not only outdated, but also marginalized to a certain extent.
AI Image generated in response to the requirements of "Revival of Neural Networks" editing
Here, psychologists come again to save artificial intelligence. Among engineers and natural scientists, psychology has always been in the position of pseudoscience, and if it lost its former position in the technical knowledge community, why wouldn't they choose a rejected topic? The first "Artificial Intelligence Spring" in the mid-1980s was related to the name of David Rumelhart, a cognitive and mathematical psychologist at the University of California, San Diego, and the invention of backpropagation methods or gradient descent.
This is interesting, but even in this case, if the Soviet science at that time could be more integrated into international science and published in English, or at least systematically translated works from foreign countries, the revolution in neural networks might have begun ten years ago. Today, even computer technology historians have little to know about the names of Avtandil Kvitashvili, Genrikh Otkhmezuri, Sergey Dayan and others who experimented with multi-line perceptrons in the 1950s to 1970s.
David Rumelhart
Cybernetic Aleksey Ivakhnenko trained eight-layer neural networks as early as the early 1970s, although created based on different types of "formal neurons". In 1974, future MIPT professor Alexander Galushkin published a monograph Synthesis of Multilayer Pattern Recognition Systems, in which he describes an identification system composed of linear threshold elements organized in the form of an open-loop network. In his work, Galushkin redescribes the problem of learning multi-layer networks and networks with loop connections as gradient descent problems.
Unfortunately, it is only in the article " Learning to Characterize by David Rumelhart, Jeffrey Hinton and Ron Williams" published in Nature of It describes multi-layer neural networks and a new training method.
For example, we have a neural network consisting of three layers of neurons: five in the input layer, three in the hidden layer, and one in the output layer. We remember that there is a connection between the neuron and the weights of these connections. If there are multiple layers in the network, the weights are constructed as value vectors or gradients.
The same situation as the perceptron, the teacher sets the output value. If the output value obtained as a result of a training period does not match the given, the difference is used to update the weight in the output neuron connection. The weights of the connections in the hidden neuron and the input neurons then change according to the back propagation of the error and how each weight affects that error.
Essentially, Rumelhart and his colleagues figured out how to calculate the gradient of each weight in the network, from the output layer of known errors to the input layer, layer by layer. Therefore, neural networks can now be trained to solve complex problems and train on large amounts of data. True deep learning has emerged.
980s gave birth to a new wave of optimism. But the breakthrough discoveries of that era laid the foundation for most modern AI models, but ultimately did not give the results that everyone hoped for. There are two reasons for this result: the lack of effective learning data and weak computing power. The model was trained too slowly and it took a long time to train it, but the task was not solved in the end.
to 2000, the hype completely subsided. The second "winter of artificial intelligence" has arrived, and it did not end until the 2010s. The emergence of powerful graphics processors (GPU) for high-performance parallel computing, as well as access to large datasets with 15 million images like the ImageNet database, is divided into 22,000 categories.
In addition, in the "calm years", many new methods of neural network design emerged: convolutional neural networks (Jan Lekun), networks with long and short-term memory (Sep Hochreiter, Jurgen Schmidhuber), generative adversarial networks, autoencoders, and many others.From online theater recommendation networks and facial recognition of smartphone cameras to poultry farms and hospital operating rooms, huge breakthroughs have been made, so it is difficult to find a place without using artificial intelligence.
The powerful power of artificial intelligence
There are two leaders in the field of artificial intelligence - the United States and China. These powers are leading the way in the number of scientific publications, patents, startups, models and technologies created on artificial intelligence. With “world artificial intelligence becomes the second power”, Russia has no right to allow competing countries to go further in the competition for new technology. In fact, the situation now replicated the 1940s to the 1950s and 1950s to the 1960s, when the Soviet Union and the United States first competed in the nuclear race and then in the space race. The problem of having its own artificial intelligence technology becomes particularly acute when Russia is under unprecedented sanctions in world history.
In order to meet the challenges of the times, the Russian government has formulated a national strategy for the development of artificial intelligence in the country. The National Center for Artificial Intelligence Development under the Russian Federation Government was established. The federal project “Artificial Intelligence” is being actively implemented, one of which is the establishment of six research center universities in Russia on the basis of leading scientific and educational organizations (MIPT, NRU HSE, Skoltech, ITMO), in the leading universities of Russia, Innopolis, the Institute of Systems Programming named after V. .P. Ivannikov RAS).
OPENAI DALL-E Image generated by artificial intelligence for "endless servers"
Russia's largest company is not willing to lag behind, including Yandex, Sberbank, MTS, VK and other technology leading companies. Therefore, the Sber team trained large generative models ruGPT-3 and ruDALL-E based on the transformer architecture. For example, the latter allows you to generate images based on Russian descriptions. During the research process, Sber also created his own benchmark to evaluate the quality of multimodal algorithms solving various problems. In addition to applying artificial intelligence technology to search, Yandex recently released the largest YaLM 100B (Yet another Language Model) language model to the public, and has been testing its own driverless vehicles for several years.
The interactions with universities such as the MIPT Applied Artificial Intelligence Systems Research Center with many industrial partners and startups, allowing you to create driverless vehicles (auto and air), autonomous and voice-controlled robots with elements of powerful AI, and also equipped with advanced computer vision systems. This combination of different technologies enables robots to perform complex tasks in a variety of environments, including unfamiliar ones.
Russia has great potential to become a third world artificial intelligence power. There is only one limitation—the industry has a serious shortage of qualified talents. The state and companies are preparing to educate young people for free, including girls who want to enter the field of STEM, and everyone who is ready to change their activities, not only enter IT, but also enter the field of state-of-the-art information technology - artificial intelligence system.
