In the winter of 2017, Mikhail Bilenko headed the department of machine intelligence and research of Yandex. RBC magazine asked him about the current state and prospects of AI. Most likely, the uprising of the machines we are not yet threatened.
– Let’s start with what is artificial intelligence? In the everyday sense it is a thing that can think, build inferences and so on – in this form, the artificial intelligence already exists?
– No, but this is not the goal. Artificial intelligence is both a microphone in the phone, recognizing speech, and an application with a camera that photographs the machine and recognizes its number. Even cruise control in the car is a primitive, but also an artificial intelligence. It performs a function that is required and that a person no longer needs to perform. That is, it is usually implied that the function performed by the mini-artificial intelligence requires some processing of the input and some actions or solutions at the output. The input processing function and the output solution can be relatively simple, like with cruise control: the speed is measured, the engine speed is reduced or increased. And he thinks or does not think at the same time – this is already secondary, if he perfectly keeps the car at the right speed or recognizes the voice, no matter in what noisy environment it may be pronounced or with indistinct words.
– Do you want artificial intelligence to think?
– What for? We need him to do what he needs to do. As in this case the algorithmic process occurs, it is rather secondary. Again, these are more philosophical than technical questions. What is “think”? What does “people think” mean? Again, some inputs come in, some of the designs we have, and if we do not do anything, then we call it thoughts. The system monitors the road, there is constant processing inside, and if there are no cars on the road, it simply does not take pictures of them. She thinks about the road or does not think – this is a rhetorical question. But if the time of day and lighting changes, the detectors and the work model change. Or a speech recognizer: he can monitor the noise level, and if there is noise, then include its suppression. Does he think about the noise or not? In a sense, he thinks, but it is more correct to say “reacts”.
– So we will go into philosophy. But let’s just say: the machine already knows how to learn? That is, some things she could not or did not know, but after a while she already knows and knows how.
– It’s just a function of the fact that it has the ability to process data and change actions in the algorithm. This training is in the sense that it changes the final behavior based on the received data, that is, the systems can be trained, yes. Everywhere this happens.
– There is much talk about which professions in connection with artificial intelligence may not be needed. For example, it is almost agreed that the machines will write the code themselves, without a person.
– Anyway, those machines that write the code themselves, should someone program. While it turns out that these machines are very complex and they need a lot of coders, and the labor market for developers continues to increase.
“But visionaries like Ilona Mask are already sounding the alarm, saying that we need to regulate artificial intelligence until it’s too late.”
– Visionaries have one common property: they never created artificial intelligence and systems working on it. In this sense, there is a big difference between the visionaries who came from these trenches and the visionaries who were not in them and who often have a fairly mystical idea of what is happening there.
– That is, artificial intelligence can not be a threat to people?
– I believe that the fears that Mask says are in vain. There will indeed be big changes in the labor market, this is obvious. Questions about the labor market or the ethical aspects of the application of surveillance systems are certainly relevant. And the questions about the awakening of some self-propelled intelligence that begins to displace people, in my opinion, are not relevant and will not be relevant.
– Will we have personal artificial intelligence in the near future, like in fantastic films?
– Personal helpers are usually very contextual: an assistant who reminds of an unclosed door will have some system that monitors the room, an “intelligent” house, there is a mail assistant in the mail, and so on.
“But can there be a single assistant who deals with all the problems?”
– May be. But in the end under the hood this assistant is still the sum of all these mini-assistants. In this sense, while there is no higher-level intelligence, but there is a system that allows you to integrate different assistants under one hood.
– Are you currently working on such integration?
– Everyone wants to achieve this, but in practice there is a terrible fragmentation. On the example of the same “smart” home – even manufacturers of light bulbs and thermostats can not agree among themselves until now. This is the next stage, but so far everything is at very early stages, and the integration of assistants still requires a lot of effort.
– What’s stopping you? The clash of commercial interests?
– No, it is more an engineering problem. Systems in which these assistants exist, they are quite different in order to start working together. It’s like there’s still no single charge for phones. The issue of interface compatibility, an agreement on how it all works together, because the level of complexity is very high, is the main barrier.
– A voice assistants such as Siri or Alexa – will they be able to understand a person, and not just substitute the right phrases in response? Understand metaphors and jokes?
– They will be able to communicate on more general topics. And the question is not whether they will be able to understand metaphors and jokes, or they can simply beautifully answer them. For the listener, the main thing is that they correspond appropriately, correctly, beautifully, and they understand what was said or not – here we return to the question of what it means to “think” and “understand.” Answer those scripts that are not clearly stitched in them? I want to believe that before they are close.
– Voice interfaces look like a global trend, they like everything: you do not need to type letters, you can just say something to the machine, and it will hear.
– They are very comfortable. This is an example of the fact that when a certain level of quality is reached, the technology becomes wider, more widespread and applied. That is, efficiency is the main engine in the propagation and application of something, rather than trends or something else. Wins what people think is more convenient. It seems to me that just as speech recognition has become easily accessible and inconspicuous, in the next year or two the same thing will happen with the translation.
– What exactly will change?
– Now machine translation often goes tongue-tied, wrong and requires editing, as well as until recently the recognition of speech came with gross errors that needed to be straightened. But there is a sharp improvement in the quality of translation, this is especially noticeable in languages like Russian – with complex morphology, cases, declensions, genera, forms. After a while we will stop noticing that something did not work before or worked badly because it will work very well.
– Does this apply to texts or to translation from a voice, too?
– These are related things: there is a translation of the text into text, and there is first the translation of the voice into the text. Improvements in both components add together. Of course, there will still be many problems. For example, if several people speak simultaneously, then the listener is very good at focusing on one source or on two or even snatching it, and for speech recognition programs this is still a big problem.
“So many small jumps”
– Is there any area in which to expect sharp improvements?
– Everything, that is connected with a computer sight. While this works at the level of individual applications that can identify a bottle of wine on the label and add it to the collection, identify the goods by type, or applications related to home cameras and security. Penetration of vision into different applications is fairly quick, because the quality that we begin to achieve in vision, allows it to be actively applied in different tasks. For example, visual search – you just point the phone at something and get an answer – it becomes much more popular and more used. The voice bot used to not so many people, because the quality of the voice suffered, and now it is used by many. Similarly, it will be with the search and any tasks that are solved with the help of the camera.
– What services of Yandex will you implement artificial intelligence?
– And they already have almost everything on it (see the cut). Just a lot happens, many changes are small in themselves, but they accumulate, accumulate, and all services become smarter and smarter. As “Yandex.Taxi”, which began to give advice “you better get there by taxi to this metro station, get off at another station and take a taxi further.” In the mail, learning models spread out letters to folders: purchases, tickets and so on. Weather, Translation, Zen, Maps – all these services actively use various technologies of artificial intelligence.
– Are neural networks an artificial intelligence?
– This is one of the algorithms of machine learning, part of artificial intelligence. Important training, specific models – networks or decision trees – are applied depending on the type of data. In the mail, by the way, it has long been determined by machine learning whether a cracker has entered or a real user, according to a number of factors: if it is decided that a burglar, a person is notified. There are a lot of improvements that happen completely “under the hood”, that is, some improvement in the navigator, which will lead to a reduction in the time of the routes. But a person can not notice the difference, because nothing has changed in the interface. For example, in voice recognition, we have launches continuously, the microphone improves every month.
– It seems that this is the feature of the new technological revolution – it happens somehow seamlessly, without sudden leaps.
– There are just a lot of small jumps, each of which is not big enough to be trumpeted about, and by the sum they add up to a very big advance in a small number of months or years.
Authors: Irina Yuzbekova, Valery Igumenov
With the participation of: Anna Kozhuhar
Original in Russian: RBC Magazine