|
The concept of artificial intelligence at the moment is very blurry, at the moment all and sundry feel themselves using AI - according to statistics, more than half of startups say that they use AI. In fact, this is all a hoax, there is no AI. The only thing that really managed to do was machine learning methods for image recognition. And these methods work, but they still do not solve the problem of creating intelligence.
The question is why? The answer is simple is simple - because the machines are not smart enough to teach themselves. Even the methods of neural networks are based on the fact that they are fed billions of data labeled by people, and only after that they become able to recognize the first data after billion by themselves
So if you want to teach a neural network to recognize cats in pictures, you give it a billion pictures, half of which have cats, and half do not, and for each picture you give the correct answer given in advance by people. That is, the so-called machine learning is still not machine learning, a person is involved in it - a teacher.
Of course, everyone wants a magic box with a button - you press a button, wait a couple of hours while the box is trained - and wow- in front of you skynet from Terminator films!
That is how it is described in film. And maybe someday it will be such way, but only taking into account the fact that even before pressing a button, this box will be taught by the smartest scientists of its time to all of the available methods of conducting scientific research.
We are now at this stage, we must teach the system what we know ourselves - and only after this, press the “learn further by yourself” button.
The history of technology knows few examples of the successful application of this principle - people teach the machine in the beginning.
The brightest of such examples is the only one in history - but the fantastic flight of the Soviet shuttle Buran.
Broad media preferred to forget about this flight, because it was ahead of its time by decades. And the collapse of the USSR eclipsed this engineering triumph of all mankind. But right now and precisely here, in the context of creating a real, not a fake AI, we must remember this flight.
So aboard the Buran was a real AI, completely imitating the behavior of a first-class cosmonaut pilot. Never before and never after autopilot did not land a reusable winged spacecraft which returned from space, and espesially of such gigantic dimensions as the Buran.
But we are not interested in the flight itself, but in the methods of training the onboard AI of the Buran. He was trained to fly for 10 years by professional pilots. He learned to imitate those who already know how to perfectly manage aircraft.
And we are not interested in specific algorithms - although this is undoubtedly a very interesting question - in this article we are interested in the principle itself: Buran’s AI learned to fly from living people. It was not an abstract program written by programmers — consisting of loops and logical branches — it was a system that could imitate people.
Similarly, wolf cubs learn to hunt in the wild - they look like a wolf mother does it - and repeat after it, only then they try it themselves. The domestic dog will never survive and will die of hunger in the wild forest, because no one has taught her to hunt. The wolf cub will survive.
Buran brilliantly landed at the cosmodrome with a strong side wind, and, after analyzing the situation, made an additional turn over the cosmodrome and entered the strip from an unexpected side, where the wind was not dangerous to him. The designers were not completely confident in their child, because the flight was the first and experimental. When Buran was leaving for a turn, they wanted to blow him up, fearing that a program crashed and Buran could fall on the city. But he was given a chance - and he coped brilliantly.
At the moment, one of the leaders in creating AI is China, which uses the army of cheap labor no longer for manual labor — but for training its AI.
In China, there are giant factories where people do not stand behind the machines, but instead sit at the computers and label the data. After all, a person is very good at distinguishing subjects, even a person who has a very mediocre education. But this is still the preparation of datasets for modern machine learning, which does not know how to think, but only selects successful numerical combinations on billions of test data.
And what will be the next step? The next step is the system Pangeya.
And a person looking at a picture will make a label not just a “cat” / “not a cat”. But in the most detailed detail - through the properties - describes everything - what he sees in the picture. And this is how a true AI will be created.
|
