What is Generative Artificial Intelligence?

Generative artificial intelligence is a field of deep learning that enables machines to generate content such as images, videos, and text based on manually created data. 

Following the development of deep fakes, face, and car generation, the latest models are now capable of generating highly realistic text and images.

In this article, we explore the concept of generative artificial intelligence, explain how these models work, discuss their applications, and delve into the ethical considerations surrounding them. Now lets we dive into the fascinating world of generative AI and its implications.

What are the applications of generative AI?

Generative AI has an infinite number of applications. In this section, I will present the most impressive applications that are beginning to gain popularity.

Generating Images with Generative AI

One of the most talked-about applications recently is the ability of models to generate images from simple texts. I have written extensively on this topic and have even offered a tutorial on quickly generating your own images.

The output images obtained are truly impressive, and it's important to remember that we are only at the beginning of this technology.

Generating Text

In addition to image creation, generative AIs are becoming increasingly proficient in generating text. Not only can they engage in human-level discussions on various topics, but today's top models can also generate paragraphs, articles, and even entire books.

For instance, You can generate an ebook using GPT-3.

Writing Code

But that's not all!

Projects like GitHub Copilot have taken AI-generated code to new heights.

How do generative AI models work?

The most commonly used methods in generative AI are GANs, VAEs, and transformers.

GAN, or Generative Adversarial Networks

GANs are generative neural networks designed to produce realistic content based on input data. Yann LeCun considers them the most important idea in machine learning of the past decade.

GANs consist of a generator and a discriminator that compete against each other during training. The generator generates content, while the discriminator determines whether the generated content is real or fake. Through this competition, both models improve simultaneously as training progresses.

VAE, or Variational Auto-Encoders

VAEs are a variant of auto-encoders.

They have a funnel-shaped neural network architecture. The first part of the funnel called the encoder, aims to encode the input data into a small-sized vector. 

The second part called the decoder, reconstructs the input data from its encoding.

The benefit of this approach is to construct a latent space where the encodings of all input data are arranged in such a way that simple operations are possible.

With this method, it is possible to generate new data that resembles the latent space.

Recently, generative AI models have been trained using approaches like Transformers, which utilize attention mechanisms, reinforcement learning approaches, or even more traditional and less resource-intensive models like hidden Markov chains.

Examples of generative AI applications

Here are three examples of models that utilize the techniques discussed in the previous section for automated content generation.

Stable Diffusion

Stable Diffusion is an open-source model funded by Stability AI. It generates images from text descriptions. It is an open-source, more reliable, and faster version of DALL-E 2, the model proposed by OpenAI in 2022.

ChatGPT, the pinnacle of generative AI

ChatGPT is a language processing model developed by OpenAI. It uses knowledge transfer to produce responses to questions based on a large amount of previously seen text data.

ChatGPT is capable of understanding and generating text in various domains, ranging from informal conversations to more complex subjects such as science and technology.

Make-A-Video

In addition to text and images, the latest advancements allow us to consider progress in the field of text-to-video, that is, generating videos from the text.

In 2022, Meta proposed a paper called Make-A-Video, which enables the generation of short videos from text.

What about the ethical aspect?

Generative artificial intelligence models have made significant technological advancements. In many cases, the generated content is nearly as good as human-generated content.

However, there is still much work to be done in securing these models, and several ethical questions remain unanswered.

The environmental impact of training generative artificial intelligence models

Most of the recent models introduced do not present significant algorithmic advancements. They often involve reusing existing models with hundreds of billions of parameters and massive amounts of data.

Only large companies can afford to train such generative AI models. So a proposal of decentralization as a viable approach is necessary.

The other problem with the race for data and an excessive number of parameters is the environmental impact. Data centers are becoming larger, requiring more energy to operate, and this poses an increasingly uncontrollable ecological disaster.

Data labeling, bordering on slavery

Having a large amount of data is not enough to create truly powerful generative AI models. It is also necessary to associate this data with labels.

This manual labeling work is often time-consuming and labor-intensive.

What do big tech companies do in such cases? They employ underpaid workers, bordering on slavery, in countries in Africa or Asia to perform this task.

Even worse, in the case of models like ChatGPT or DALL-E, among the content that needs to be labeled, there is highly sensitive or even hardcore content.

Images of slavery, child pornography, insulting texts, and descriptions of explicit scenes are among the content that needs to be labeled.

This task helps secure the model and prevents it from being used to generate such content. However, it is unacceptable that people are paid less than 2 euros a day to review this type of content for days on end.

Conclusion

What better way to conclude this article than to let a generative AI model generate the conclusion?

Generative AI is a field of deep learning that allows machines to generate content from manual data. The latest models can generate highly realistic images and text.

There are countless applications for generative AI, including image generation, text writing, and even code generation.

The most commonly used methods are GANs, VAEs, and transformers. GANs use a generator and a discriminator to produce realistic content, while VAEs use encoding to construct a latent space for generating new data.

Generative AI applications hold promise, but they also raise important ethical questions regarding the reliability and accountability of these technologies.