AI and MO: some trends and trends

Artificial intelligence and machine learning are two technologies the implementation of which and / or their derivatives in human society will inevitably radically change it. What trends are now especially noticeable, what scientists around the world are doing, and what is happening right now - in this article.

1. Working with natural languages

All language translation and text recognition systems began to work on the basis of trained neural networks. As a matter of fact, this is already clearly visible by the same Google Translate, the quality of translation of which is growing, one might say, every day. Apparently, there is very little time left until the results of machine processing of text become comparable with the quality of work of live translators, especially since this year many research teams are going to develop multilanguage complexes. Therefore, it’s not even very clear whether, for example, it’s worth starting to learn some Chinese right now.

2. Casts of people

Not bad progress is being made on training neural networks to simulate a specific person, his behavior and habits. So far, the main problem is that it is not completely clear what kind of data the AI ​​needs so that, as an option, he can reliably mimic a specific character, for example, in communication on the Internet so that people who know him well cannot recognize the fake. But in some aspects of this work there are already serious achievements, judging by all the AIs will soon be able to reproduce the literary style of a particular author. At this moment, all editors will have to work very hard, although on the other hand, perhaps by that time, the style of the work, along with the work itself, will become the object of copyright.

3. Computer vision

Until recently, for training neural networks to recognize objects, they had to be labeled manually, which, to put it mildly, was not very fast and not very effective. However, last year there was a breakthrough in this area, and now AI is able to learn with little or no clue. At the same time, neural networks learned how to generate images indistinguishable from photographs without special expertise, and Nvidia created a neural network that is able to do the same, but with video. Due to the fact that an increasingly important role in communicating with AI is played by “concepts” rather than “objects” (for example, the “big turtle” is a concept, and the “image area with such and such coordinates” is an object), in the course of time, soon it will be possible to create completely artificial video materials, actually "on the fingers" explaining to AI what is needed from him.

The Internet will immediately be filled with video evidence of events that never happened, and understanding what is really happening in the world will become even more difficult.

4. Practical use

The stage when AI mainly trained on solving artificial problems, such as playing Go, chess gambits or analyzing unnatural algorithms and events designed only for their training, is almost complete. The stage of introducing neural networks into real aspects of human activity has begun, from modeling exchange trends and urban traffic management to genome analysis and drug development. They are also actively working on introducing neural networks into the financial sector, AI will soon become a loyal assistant to any trader, and in the future, when training neural networks in the behavioral patterns of a particular person will become the norm, he will be able to trade for a person in the same way as he would himself - only better and 24 hours a day.

Very soon, these systems will gain enough information to optimize their areas of responsibility, and our life will begin to change dramatically in so many aspects. It’s not that it would make sense to stock up on canned goods and matches, but not all experts are sure that the neural networks are guaranteed to be safe, because very serious problems with the interpretability of AI have already begun.

5. Interpretability of AI

In recent years, colossal funds have been invested in the research and development of neural networks, so it is not surprising that tremendous progress has been made. However, it was he who created the problem, which is now no longer considered the lot of the marginal wing (so to speak) of the scientific community and brought up for discussion: the problem of interpretability of AI.

Roughly speaking, even specialists in many cases either do not understand well or do not understand at all how exactly even functionally useful neural networks process data and produce results. AI often turns into a black box, and some companies (the same Google, for example, which speaks of the extreme seriousness of this problem) are currently conducting research, the sole purpose of which is to understand one simple fact: how safe is it to continue developing AI and introducing it in real life? There’s an excellent interview on this subject with Bin Kim, a Google Brain researcher, she’s just dealing with this issue and, judging by what she says, she and her colleagues have little confidence that AI is definitely safe.

6. Ethics of AI

Well, and, in fact, the above are well understood by specialists involved in artificial intelligence, so everyone suddenly and unanimously took care of the “ethics of AI,” that is, how to make sure that this incomprehensible thing doesn’t suddenly begin to perform efficiency for some nightmarish things . However, so far, apparently, apart from the recognition by the leading players in this industry that yes, it is necessary to develop and implement some kind of special morality and ethics for neural networks, things have not budged.

In general, it's time to re-read Azimov.