Ten years ago, I considered the thoughts presented in this publication to be fairly commonplace. After reading the latest publications on Habré [1], I realized that this is not so.
The term “Artificial Intelligence” can be a great illustration of the concept of the precession of simulacra. Its value changes every year, depending on the mood of the market or the fashion of humanist philosophers. Decades ago, the chess program was considered an AI. Today it is commonplace engineering. In decades, Watson and Siri will be a standard component in the next framework. AI is a symbol of the unknown, as soon as we pinch off a piece of the unknown - it immediately loses its appeal.
However, if you forget the humanitarian installation and get down to business from the point of view of the technician can understand the following things.
In the definition itself is a reference to natural intelligence. Those. we define the term through another, even less defined, and which is part of the one who defines. Moreover, existing in a single copy, we do not know the intellect different from the human. And if you take into account the fashion trends in the behavioral economy [2] - possibly non-existent. Strangely enough, in our imperfect world, where people spend most of their lives in making money, they do not use rational thinking while using only sets of standard templates.
As a result, the term AI is often used as a buzzword.
')
However, we are all accustomed to working with fuzzy problem statements, because a programmer is a professional converter of a customer's hallucinations into a rigid formal system. And because the above considerations can not stop us.
Different people and different schools understand the term AI in different ways, but in the semantic spectrum I would single out two main technical trends. I will call them the Turing approach (from the Turing test) and the Godelev approach (from the Gödel machine [3]).
Gedelevsky approach is AI as a universal problem solver. There is nothing particularly complicated in the construction of such an object; the first General Problem Solver [4] was created in the late 50s. The fashionable Gödel car today is the current state of our ideas.
Imagine that we have a magic wand that fulfills our desires. But the wand unfortunately is Chinese-made, used and often buggy. Fulfills not all desires, some performs not at all as we would like, although it happens that it works as it should. The most reasonable step in such a situation is to order a simple desire from a magic wand - to make yourself better and more stable over a period of time that is no more than a given one. Perhaps she will not do it as well as we would like - but we are not limited in the number of iterations. Sooner or later the process will converge and we will get what we need.
This is a brief and very simplified description of the Gödel machine - an object that rearranges and optimizes itself for the task and does it optimally.
Strictly speaking, a universal task solver is simple enough, all that is required is to go through the search space, optimizing this search on the fly, generating and refuting the heuristics of the search and the heuristics of the search heuristics. The trivial genetic programming engine does almost the same thing and can also be rightfully called the Gedelev AI (but not the Gedelev machine).
If the Martians arrive to us tomorrow, reasonable viruses or an even more exotic form of mind and show their implementation of the Godelev AI - it will be identical to ours, since the requirements for implementation are 100% objective and do not depend on our ideas.
And this is the main difference of this direction from the Turing AI.
Turing AI is something that passes the Turing test, a machine indistinguishable from a person by social parameters. And here begins the hell in implementation. Because it is quite difficult to determine what a person is. Personality from the 19th century Turing test in our time will not pass. As well as the Chinese with Indian appraisers. In general, for the predetermined composition of the jury the Turing test will not pass the majority of the population of our planet.
Any programmer who has at least once analyzed the output of the genetic programming algorithm will understand that the human person is a jumble of introns, junk code, inherited fragments and a certain amount of working logic spread over the common mess.
To reproduce this in the code is difficult and pointless, the years of work of research teams to simulate what any woman will do in 9 months - is the price not high?
And this code will be difficult to repeat, in each individual implementation will be different.
However, the prize is also good, the market value of the Turing AI is immeasurably higher than Gedelevsky. In our closed world, we are in contact with a huge number of objects, but the most common is the other person. And his imitation, more friendly, adjusting to our mood, telling us nice things (half-and-half with advertising) will be expensive and sold quickly. It is clear to any investor that Siri is more popular than Watson.
It is important to understand that these two paradigms lead to two fundamentally different products and denote different objects, despite the fact that both are called the same term - AI. People-oriented products and products built according to objective, human-independent requirements are very different not only from the outside. Separation of approaches is found in various non-AI areas. A web designer who has drawn a website on flash animation will be very proud that the site is beautiful (for a person), attractive (for a person), but that the site is not indexed by a robot, the designer will simply not notice. And vice versa, an ugly HTML website from Web 1.0 will be perfectly understood by any parser without much strain, and even if Martians arrive with vision in the field of X-ray radiation, it is possible to transfer them an html file with preservation of meaning. Unlike visual effects, flash animations. The subjective approach is the words of Steve Jobs about the “product with a heart”. This is when we choose a programming language based on personal comfort, and not from the simplicity of mining code or code generation. This is a common dilemma.
Each approach has its advantages and disadvantages. You just have to remember that we, too, can work as Gödel machines. And that an objective approach sooner or later converges to the optimal solution. And the subjective will never converge, because There is simply no optimal solution to a subjective problem.
It is impossible to create an objective metric - two programs that pass the Turing test are better or worse. In every generation of people, in every culture, the Turing AI will be different. Well, as the TVs with a diagonal of one inch more than last year's model and with the support of the new standard super-hyper-ultra-HDTV v.314. This is a product of instant aging - right at the time of sale.
The dream of the manufacturer.
An example of the eternal precession of simulacra.
References:
1.
here
2. Behavioral economics
From the Nobel Prize in Economics for 2002 to
such a book
Or the well-known “Phryeconomics”
3. Gödel machines
point of entry
4. General problem solver - the first historical document.
very old scan