Do you think artificial intelligence is not creative? Look how he writes and how he paints | Technology

In the popular imagination, artificial intelligences (AI) were going to be dispassionate artifacts, moved only by reason, unable to intuit, joke, make mistakes or be creative. But the reality has turned out to be different: the latest AI algorithms are pretty good at doing all of that.

These days I have experimented with various AI models, massive neural networks trained to write, program or paint. They are cutting-edge algorithms, but I have been able to interact with them from home, without being an expert and dedicating some time to them. I tell you about four examples, the tip of the iceberg of a revolution in progress.

An AI that paints

I started by making a drawing with Geniverse, an AI designed to “increase your creativity”. Although it is in the testing phase, it is absurdly easy to use: you write the commands of what you want it to draw – the prompt—, you click, and its iterative algorithm converts an initial white noise into a single image that reflects what you asked for.

This was the drawing process when I asked “A boy looking at a bright landscape. The boy’s hands are burning. art station style [una web de ilustraciones]».

With the same instructions I asked him to make more drawings, which are different each time, although they repeat patterns. Here are four examples:

You can also ask Geniverse to refine a drawing. Here I took the first of the previous images, I told him that “the boy has brown hair”what to use “bright colors”, and add details. The result was this illustration below, which is better than anything I could paint.

Javier Ideami, one of the people behind Geniverse, explained to me that his plan is to make Artificial Intelligence tools that help people bring their ideas to life. They conceive it as teamwork, between person and machine, because you have to choose the commands, touch parameters, etc. Today’s AIs are like brilliant teenagers, capable but fickle, and Ideami believes that a job of the future will be to talk to them: “there will be people dedicated to engineering promptexperts in knowing how to get along with an AI, who know their training and who almost empathize with it.

After the drawing of the boy, according to asking for other images, like this one from “a girl looking at a landscape. The girl has blue hair. It is an abstract work by Kandinsky with bright colors.

This one was also pretty «landscape of a valley crossed by a river, with a herd of buffalo in the background. In watercolors. With bright colors in the sky».

Other AIs that invent objects

For my next test I used a newly published imaging model from the CompVis group at the University of Heidelberg, which is also capable of creating images from text. I asked “A Japanese Scene, a painting by Edward Hopper”and the result was these paintings almost ready to hang on the wall:

When I told him to do more complicated things, he got out of the way. here are some “paella painted by Edward Hopper”.

Isn’t it amazing? Sometimes he messes up, and there are things he was incapable of doing, but he can create really weird things, like these “kiwi-shaped chairs”.

New generative AI models are coming out all the time, like DALL E 2, which was released this week and has similar, perhaps more powerful, capabilities. I haven’t been able to test it yet, but its creators have shared some pictureslike this from “a boy and a dog looking at the stars”.

None of these images existed before. It is important to realize that these AIs are not making copies of things that they have seen. They have been trained by looking at hundreds of thousands of images and phrases, yes, but from that they have “learned” patterns and structures, in an iterative back-propagation process that has been well known since the eighties, but whose potential has been exploited in the last decade. decade. These nets can make paellas, and valleys, and buffaloes, and chairs, and Kiwis; and when they are asked to connect two of those concepts, they are somehow able to bring them together without losing the salient features of each. It is an obvious form of creativity.

Another AI writes stories

For the third example I used OpenAI’s GPT-3, a language model trained to produce human-like text. In essence, you have studied all the text on the internet to learn how to write. It is not like a repeating parrot, but it is capable of surprising things: it can summarize a text, correct grammar or translate, but you can also ask it to say which cities are mentioned or which explain a joke.

For this example, I told him this: “Write a four-sentence story about a boy and a girl who meet after 30 years.” And he gave me this back:

• Once upon a time there was a boy, who met a girl.
• They became friends.
• They were very happy together.
• They were separated when they were eleven years old.
• They met again when they were forty-one years old.

Then I asked him to add a happy ending, and he added a hackneyed one: They were so happy to see each other. they decided to get married. I told him to change the protagonists “so that they were not human, but other beings”, and exchanged them for a unicorn and a dragon. How do you know who the protagonists of an unpublished paragraph are? He understands the meaning of that concept and is able to find it in the text.

I kept giving him instructions (sometimes twice, because he failed), “remove the wedding and instead travel a lot”, “change the story to have a dialogue at the beginning”, “change the story to be a single paragraph”, “remove the phrase about being happy” Y “make concrete the places where they travel to”. He did it all:

Once upon a time there was a unicorn who met a dragon. The dragon said: “Hello, unicorn. I am a dragon. What’s your name?” The unicorn said, “My name is unicorn.” They became friends. They were separated when they were eleven years old. They met again when they were forty-one years old. They were so happy to see each other. They decided to travel to many places. They went to the moon, the sun and the stars.

The story is not much, but what is relevant in this case is that GPT-3 understands complex instructions.

Another variant of GPT-3 can convert English into computer code. Program. You can tell it “make a function in Javascript that calculates the days that have passed between two dates”. The product is called Copilot, and again it’s supposed to be a tandem, “your AI partner for programming.” the software engineer John Fontfrom the European Space Agency, told me how he uses it: “it is useful for repetitive pieces of code, boiler plate, which before you had to search for every time”. He also told me that it works better than he would have thought possible. It is not only that Copilot’s AI has trained with all the open source that exists on the internet, it is that it also looks at and understands the contextual code, the project that you just programmed, to know what you are doing, what your variables are called or what style do you use

And in the future what?

From my week of exploration I have drawn three conclusions. The first is that these generative AIs are going to be available to many people. They are expensive models (in euros, computation and kilowatts), but they run in the cloud. In addition, I have not had to program, or download applications, or use a powerful computer. The browser is enough to send them orders, wait a bit, and the neural network returns your drawing.

The second is that advances in Artificial Intelligence are going very fast. They may hit a wall, but for now they’re speeding up: none of the AIs I’ve used are three years old; and what they allow me to do was impossible then. I leave for another day an issue that floats in the environment, which is to think about its applications, the virtuous ones and the disturbing ones.

The other unknown is knowing what they will be capable of in ten years.

I have no idea, but I’m pretty sure they won’t lack creativity. On the contrary, I believe that the drawings of these primitive algorithms force us to think about our creativity. In the past we have romanticized the human process of creation, which is surrounded by myths, but we have to face the possibility that it may not be as mysterious or as unique as we feel.

I end up giving the last paragraph to the AI ​​of the GPT-3 model. I gave him the draft of the article and asked for a conclusion. This is what it tells me to add: “It does not mean that human creativity is going to disappear. On the contrary, the possibilities of creating are expected to multiply. In the future we will have more tools and we will be able to do things that we couldn’t do before. It is like the discovery of the brush: it did not erase the cave paintings, but allowed them to be much more refined”.

You can follow THE COUNTRY TECHNOLOGY in Facebook Y Twitter or sign up here to receive our weekly newsletter.