Humans have made enormous progress in understanding physical reality by understanding the laws that govern the behavior of nature. However, we are a long way from understanding the nature of our intellectual capacities.
The first questions about whether a machine can think can be traced back more than two thousand years to Greek mythology. The Greeks spoke of artificially created mythological beings, which appear in stories such as Jason and the Argonauts, the bronze robot of Talos and the female robot Pandora. These entities, products of the human imagination, determine a starting point in the search for the creation of artificial life. Although they are fictitious entities, they represent an idea that shows highly challenging technical challenges that encourage us to make them a reality. These challenges have marked the path towards the goal of achieving useful Artificial Intelligence.
One of the first approaches to the creation of artificial intelligence was through the understanding of biological intelligence. The idea of obtaining a mathematical model of the neuron and connecting it to a network to obtain a neural network marked the beginning of Intelligent Systems and the area of “Machine Learning” or systems that learn by themselves.
Intelligent agents that can learn to play a game of checkers, playing against themselves and learning through trial and error, have been able to beat the programmer himself. In 1997, the IBM-built Deep Blue supercomputer defeated one of the greatest chess masters, Gari Kasparov, thus demonstrating that in certain tasks the machine can outperform man. In recent years, the English company Deep Mind created the first computer program that defeated the world champion of the game GO, known as one of the greatest challenges due to its high complexity.
In the last 12 years a breakthrough in the area of Machine Learning took place: Deep Learning. Deep learning is nothing more than the result of improvements in neural network algorithms, Big Data and advances in distributed computing. Deep learning has been a central contributor in making systems such as “Self driving cars” a reality, as neural network models trained with this technique are able to recognize images, often more accurately than a human being.
If we can be able to replace a human driver, we may be able to paint in the style of Pablo Picasso.... Generative Adversarial Networks (GAN) attempt to achieve that goal. GANs are composed of a generative and a discriminative network. The generative one forges an image by means of a given example and the discriminative network determines whether the generated image is false or real. In this process of generating and discriminating, at some point the generated image is accepted by the discriminator as real, in this case a painting in the style of Pablo Picasso.
The GPT 3 language model is a technique that aims to predict what is coming based on previous data. This model was trained with a huge amount of books and millions of web pages on the internet. GPT 3 is able to predict the next sentences of an article based on the previous ones, meaning that with a few sentences it is able to build a complete article. There are GPT 3 models capable of responding coherently and generating human-level conversations, even capable of passing the Turing test: not perceiving the difference between talking to a machine or a human being.
This long road from the time of the Greeks to the conquest of artificial intelligence and its achievements, makes us reflect on whether a machine could be able to surpass or replace a painter, a writer, a psychologist. Could it be that the one who wrote this article is not a human and was created through a GPT 3 model?
Marco Sotomayor
Faculty of Engineering
