"We are artificially restraining growth"

Interview with one of the founders of Prisma Alexei Moiseenkov

The Russian application Prisma, which with the help of neural networks turns any photograph into a picture, is only ten days. During this time, users with his help drew millions of paintings, and Prisma itself became one of the mandatory applications for those who are actively leading Instagram. The deputy chief editor of Medusa Alexander Polivanov and publisher Ilya Krasilchik talked with one of the founders of Prisma Alexei Moiseenkov - about how the application works and why neural networks will change our lives.

- Let's start with the current numbers. How many downloads does the application have?

- We have a million downloads in five days. We are in the storage [in AppStore] since June 11. Until the 16th there were five thousand downloads, and from them I myself made about two thousand to simply check the server. Yesterday we processed two and a half million photos.

- How are the servers doing?

- Well, bad. But we slowly stabilize, add power, increase, by some point we will become very stable, somewhere, probably by the end of the week. We had one real force majeure-three key servers were cut down, and we, as I call it, burned. We lay down for 30 minutes.

- And what happens not in Russia?

- So far we are artificially restraining growth not in Russia. Firstly, if some celebrity wants to post something with Prisma, we ask not to do this yet. Secondly, I can close the stories with my hands in Riiantime ( the AppStore makes it possible not to distribute the application in specific countries; for example, you can make it possible to download applications to users of the Russian AppStore, but not release the application in the USA-approx. Medusa ).

- How many people do Prisma?

-Four people co-founders, and someone else just helped. One of my familiar styles helped to select, for example. I myself have not been engaged in coding for a long time.

- And how much did you make the application?

- Two months.

- Do you have a plan about what to do next?

- Yes. We will clearly improve a lot. We already have a supply of 50 filters - but first we will deal with the load.

- What filters to wait?

“I don’t even know myself, but I like comic filters.” Wait that there will be some kind of comic filter.

- How is a new filter made?

-There is a neural network, more precisely, a complex of neural networks, it will have to be trained, driven through a bunch of images, then slip him an image of an artist-or a TV, anything, to slip any image to him, say what to do. And then you need to look at what the result comes out. You have done such an operation once, got some result. The result can be completely different, from completely indecent to "wow, how cool."

- The idiotic question. How do neural networks work?

- How do neural networks work? These are layers. The layer is a set of neurons, the rules for the passage of each digit. We set up the rules for passing through layers.

- Okay, we did not understand anything. Then the following question: we constantly see some amazing things related to neural networks. What has happened lately? Why is such an incredible number of projects built on neural networks right now? What has happened to this technology lately?

- The trick is that over the past year - maybe even a little more - there has been a significant breakthrough in several areas that help neurosetes work. Firstly, they began to better understand mathematics, it was the model itself, how everything works.

The second - the permissible capacities increased, it became easier to calculate. It also turned out that by the current moment a fairly narrow class of problems can be solved in the case of neural networks - and this is very good, because it is easier to configure everything to a specific task. You can, for example, take and say: I will recognize Chinese screams of fans. And for this task, you can create a very, very good quality product, it will really recognize these Chinese screams of fans, but will not recognize not screams, not Chinese-and not fans.

- That is, all tasks are very applied.

- Yes, yes. Another very important point: the neural networks solve only three categories of problems, do not extend them to everything. This is: the image - including video, speech and text. Everything else is not very good.

- Another amateurish question. And where is a person in neural networks? Here the car drew something to you. You understand: it does not fit. It turns out that you still need a person, an editor who is behind the last word.

- Suppose we recognize the Chinese cries of fans, and the car tells us that this is a Chinese scream. And you say: damn it, cannot be! The problem is that the machine cannot evaluate the general model, its parameters. They somehow tune there, but evaluating the parameters automatically is very difficult. For example, we fed the neural networks of Leo Tolstoy, she will write with a lion Tolstoy. But when you read this, you will recognize the style of Leo Tolstoy, but, for example, there will be no coherence of the text. Here we need a person.

- “Prism” is very cool, but this is a toy. Everyone admires how incredibly cool she works. What is the next stage for technology in general? How can she help us in life?

- Everything goes to a certain artificial intelligence who knows how to adapt to you. This is a general thought. And if it is more close to modern reality, then the neural network is rather a very good solution for applied tasks. For example, you are a traffic police and are looking for a car. Then you turn to neural networks that can recognize numbers. You do not need a person, you let the drone of some, he flew everything and told everything with an automatic-very quickly, one hundred percent hit, all numbers were recognized.

“Don't you think that all this leads to the total destruction of privacy?”

- It can happen. If there are cameras everywhere, it will be possible to recognize all faces. Such a scenario is possible.

- Let's get back to Prisma. If Facebook comes to you tomorrow, will you sell?

- Tomorrow he will not come.

- I came to MSQRD.

- He came to them a few months later.

- This is one and the same.

- The question is difficult. There is a probability, of course. It’s just that all this can earn on its own, but, indeed, if you think globally, then, rather, it would have looked logical with some large social player, with a large audience that somehow consumes photos or videos.

- And how much did you spend your money on the project?

- I can’t count the exact amount, but there is a little there. That is, there is no talk of some millions there. This is a little money, in the framework of salaries. In addition, we had a server partner almost immediately.

- Today it became known that the investor of Prisma became Mail.Ru. How much did the company invest in you?

“I can't talk about it.”

- Do we understand correctly that the project to Mail.ru initially has nothing to do with the fact that you have been to today an employee of Mail.ru?

- Yes.

- That is, this is an absolutely independent project?

- Certainly.

-Do you have any further plans for the use of this technology?

- Initially, we did PRISMA. But the thought, of course, is much further.

- Is that a secret?

- If you come to me in a week, I’ll tell you.

Alexander Polivanov

Ilya Dewarer