- Get link
- X
- Other Apps
"A few months ago," says Kevin Kelly of Wired, "I went to the forest campus of IBM's research laboratory in Yorktown Heights, New York, to look at the birth of a promising future of artificial intelligence. It was the house of Watson, the electronic genius who won the Jeopardy competition! in 2011. Watson was then so - the size of a bedroom, consisting of ten vertical, refrigerator-like machines that form four walls. A small cavity inside provides technicians with access to a hodgepodge of wires and cables on the backs of these machines. Also inside it is amazingly warm, as if something is living its own life. "
Today Watson is completely different. It no longer exists exclusively in the form of arrays of cabinets, but spread to cloud servers that support several hundred "instances" of artificial intelligence simultaneously. Like all cloud things, Watson is available to users around the world, they access it from the phone, from the computer, from their own servers. This type of artificial intelligence can be scaled up and down as needed. As artificial intelligence improves with how people use it, Watson always gets smarter. Everything that he studies immediately becomes available to others. And this is not one program, it is a collection of various software systems - logic, language, all this can work on different code, on different chips, in different places - which are embodied in a single stream of intelligence.
Users can connect to both intelligence directly, and through third-party applications that use the basic intelligence with the help of the cloud. Like many parents of talented children, IBM wants Watson to make a medical career, so it will not surprise anyone that medical diagnostic tools are first developed for him. Most of the previous attempts to create a diagnostic artificial intelligence have not been successful, but Watson works very well. If in simple English I list him the symptoms of a disease that I once picked up in India, he will give out a list of sores, from the most likely to the unlikely. Most likely, the cause of the disease was in Giardia - and this is the correct answer. However, such an examination is not yet available to patients; IBM gives access to Watson's intelligence to partners, helping them develop friendly user interfaces for doctors and hospitals.
"I believe that something like Watson will soon become the best diagnostician in the world - a car or a person, it does not matter," says Alan Green, chief medical officer at Scanadu, a startup who plans to create something like a medical tricorder from Star Trek on the basis of cloud artificial intelligence . "With the development and improvement of artificial intelligence, very soon the child will not need to go to the doctor for diagnosis until he reaches adulthood."
Medicine is just the beginning. All the big cloud companies, dozens of startups and other enthusiasts want to start working with Watson as a cognitive service in a crazy rush. According to Quid analysis, artificial intelligence has attracted more than $ 17 billion in investments since 2009. Only last year more than $ 2 billion was invested in 322 companies developing artificial intelligence. Facebook and Google employ scientists who join the development teams. Yahoo, Intel, Dropbox, LinkedIn, Pinterest and Twitter - all bought an artificial intelligence company last year. Private investment in that segment grew by 62% over the past four years, and this growth will be permanent.
Against the backdrop of all this activity, the image of the future artificial intelligence penetrates into the mass consciousness, and this is not always HAL 9000 - a cold, calculating and charismatic deadly machine with human consciousness. The future of artificial intelligence is more represented in the form of cheap, reliable services of the digital era, invisible and useful. A typical program will give you as many IQs as you want, but no more than you need. Like all utilities, artificial intelligence will be extremely boring, even if it transforms the Internet, the world economy and civilization. He will revive the inert objects just as electricity did it a hundred years ago. Everything that we once electrified, we will impart intellect. This new useful artificial intelligence will also complement us, people, individually and collectively, as a kind: deepen our memory, accelerate our perception. There is almost nothing impossible for him to think about, what can be new, different or interesting, and what can be done if there is a high level of intelligence. In fact, the business plan of the next 10,000 companies will be reduced to the formula "take X and add AI." And it's not bad.
In 2002, I visited a small Google party - before the company went public on an IPO, when it was engaged only in search. I started a conversation with Larry Page, co-founder of the company, who became its CEO in 2011. "Larry, I still do not understand. There are so many search companies. Free online search? What will he bring you to? " It was hard for my imagined head to imagine what the future would be, but to justify it, I can say that it was long before Google started promoting its advertising scheme, providing itself with revenue, long before the acquisition of YouTube or another large company. I was not the only active user of their search site, who thought he did not have much time left. But Page's response drove me into a stupor: "Oh, in fact we are doing artificial intelligence."
I've been thinking about this conversation for a few last years, when Google bought up 14 companies to develop robots and artificial intelligence. At first glance, it may seem that Google is expanding its portfolio in the field of artificial intelligence to improve the search capabilities, which accounts for 80% of the company's revenue. But I think that the opposite is true. Instead of using artificial intelligence to improve the search, Google uses search to improve artificial intelligence. Every time you enter a query, click on "Search" or create a link on the Web, you train artificial intelligence Google. Each of the 12.1 billion inquiries generated by 1.2 billion Google users daily teaches artificial intelligence with deep self-learning opportunities over and over again. After ten years of stable development of algorithms, studying data and with 100 times more resources, Google will finally introduce artificial intelligence. I think by 2024 Google's main product will not be search, but artificial intelligence.
And here there is where the skeptics go. For almost 60 years, researchers of artificial intelligence predicted that AI is about to, around the corner, until a couple of years ago everything again was not as far away as before. There was even a term that would describe this era of meager results: the winter of artificial intelligence. Something has changed?
Yes. Three recent achievements marked the long-awaited arrival of artificial intelligence.
1. Cheap parallel computing
Thinking is inherently a parallel process, billions of neurons are firing simultaneously, creating synchronous computing waves in the brain. To build a neural network - the primary architecture of software for artificial intelligence - it is necessary that different processes run synchronously. Each node of the neural network freely simulates the neuron of the brain, which in turn interacts with the neighboring ones, forming a signal. To understand the spoken word, the program must be able to hear all the phonemes in the aggregate; to recognize the image, it needs to see every pixel in the context of the surrounding pixels - these tasks are parallel. Until recently, a computer processor could only process one such process at a time.
The situation began to change more than ten years ago when a new type of chip appeared, a GPU that was designed specifically for visually loaded - and parallel - video games in which millions of pixels should be recalculated many times per second. It will take a special parallel computing chip, which is an addition to the motherboard of the computer. Parallel graphic chips have earned and the game industry has received a powerful push. By 2005, GPUs were produced in such quantities that they became much cheaper than they were before. In 2009, Andrew Ng of Stanford realized that the GPU chips can be used for parallel operation of neural networks.
This discovery unlocked new opportunities for neural networks, which could include hundreds of millions of connections between nodes. Typical processors need a few weeks to calculate all cascading capabilities in a neural network with hundreds of millions of parameters. Ng found that a cluster of GPUs can do the same for a day. Today, neural networks running on the GPU are used by cloud companies (including Facebook) to identify your friends on the photo, and to provide advice on content.
2. Large amounts of data
Every intellect needs to be trained. The brain of a person who is genetically programmed to classify things must see dozens of examples before he can distinguish a dog from a cat. The same is true for artificial intelligence. Even the best programmable computer should play at least a thousand games in chess before becoming a master. A number of breakthroughs in the field of artificial intelligence lies in an incredible avalanche of data collected from around the world that provide the necessary artificial intelligence training. Massive databases, self-tracking, web cookie, online traces, terabytes of disk space, decades of search queries, Wikipedia - this whole digital universe just makes AI smarter.
3. The best algorithms
Digital neural networks were invented in the 1950s, but computer engineers took decades to learn to tame the astronomically huge combinatorial relationships between a million - or hundreds of millions - of neurons. The key to the organization of neural networks has appeared in the combined layers. Take, for example, the simple task of determining that a person is a person. When there is a group of bits in the neural network that activates the pattern - the image of the eye, for example - the result moves to another level of the neural network for further searching. At the next level, two eyes must be revealed, and so on, according to the hierarchical structure that associates these two eyes with the nose. It may take a million of these nodes to form 15 levels to recognize the human face. In 2006, Jeff Hinton of the University of Toronto made a key correction to this method, which he described as "deep learning", deep learning. He was able to mathematically optimize the performance of each layer, thereby speeding up the process of superimposing these layers. The algorithms of deep learning incredibly evolved a few years later, when they were transferred to the GPU. The deep learning code itself is ineffective for performing logical thinking, but it is extremely important for all current artificial intelligence, whether it's Watson, Google's search engine or Facebook algorithms.
This ideal storm of parallel computing, large amounts of data and deep algorithms leads us to success in the field of developing artificial intelligence. And there is no reason to think that everything will stop - AI will only improve.
Cloud artificial intelligence will become a part of our daily life. But this will have its own price. Cloud computing obeys the law of increasing returns, which is sometimes called the network effect: the value of the network grows faster the more it becomes. The larger the network, the more attractive it is for new users, which makes it even bigger and more attractive, and so on. Clouds that will work on AI will obey the same principle. The more people use AI, the smarter it will be. The smarter he is, the more people will use it. After the company enters this beneficial cycle, it will grow so quickly that competitors will not be able to catch up with it. As a result, the future of AI is likely to rely on the oligarchy of two or three large commercial companies.
In 1997, the predecessor of Watson, Deep Blue, defeated incumbent chess grandmaster Garry Kasparov. After the car approved its victories with a few more matches, people began to be interested in it. You might think that this is the end of the story, but Kasparov realized that he could have played better if he had had instant access to a massive database of all the previous chess moves that Deep Blue did. If this tool is available for AI, why not give it to a person? Developing his idea, Kasparov included the concept of man-plus-machine matches, in which AI complements the chess player, and does not compete with him.
Today these so-called free-style chess matches look like mixed martial arts, when players use any favorite tricks that they want. You can play on your own, or act as a partner of a super-chess computer, just moving pieces around the board, or as a centaur, proposed by Kasparov. Such a centaur listens to AI advice, but takes the final decision on his own. In the Free Chess Championship of Freestyle Battle in 2014, only chess engines won 42 games, and centaurs - 53. Today the best chess player is the centaur Intagrand, a team of people and different chess programs.
Interesting is another. The appearance of AI does not detract from the work of chess players-people. On the contrary. Inexpensive and smart chess programs inspire people to play chess, new tournaments appear, players play better and better. Today the grandmasters are twice as many as they were when Deep Blue defeated Kasparov. The most ranked chess player of the day, Magnus Carlsen, trained with artificial intelligence and was recognized as the most "computer" of all chess players. He also has the highest rating among grandmasters of all time.
If AI can help people become good chess players, obviously, it can help us become the best pilots, doctors, judges, teachers. Most of the commercial work done by AI will be niche, narrowly focused, like translating from one language to another, but this is only the beginning. For example, AI can drive a car. In the next 10 years, 99% of the artificial intelligence with which we will interact will be autistic, but supramental experts.
In fact, it will not be the kind of intelligence that we are used to. Intellect implies responsibility, plus by intellect we mean our peculiar self-awareness, our desperate depths of introspection and self-discovery. But in the case of artificial intelligence, we want him to drive the car in silence and not be distracted. Artificial doctor Watson in the hospital should be a fan of his work. As the AI develops, we may, in every possible way, hamper the development of self-awareness in it - the most expensive AI services will be deprived of consciousness at all.
In short, we need not an intellect, but artificial brains. Unlike the general intellect, the mind is quite concrete, measurable, special. Nonhuman intellect is not a mistake, it is a feature. The main advantage of AI will be its extraterrestrial intelligence. AI will not think about food like an ordinary cook, allowing us to look at food in a new way. He will think differently about production. About clothes. About finances. About science and art. The alienation of artificial intelligence will become more valuable to us than its speed or power.
First of all, this will help us better understand what we mean by saying "intellect". In the past, we could say that the superintelligence should drive a car, defeat a person in chess or win in "Jeopardy!". But as soon as the AI achieves this, we will understand that these achievements are purely mechanical and hardly worth the candle. Every success of artificial intelligence makes us reconsider the very concept.
We do not just have to reconsider the meaning of AI - we need to reconsider the concept of man as such. Over the past 60 years, as the mechanisms have reproduced human behavior and talents that we considered unique to people, we need to think about what sets us apart from each other. Perhaps in the next decade we will experience an identity crisis, constantly asking ourselves what people need. Against this background, the almighty artificial intelligence will only oppress. The greatest benefit that artificial intelligence can bring is to help humanity in self-determination. In short, AI needs us to say who we are.
Based on Wired
The article is based on materials .
Comments
Post a Comment