Machines can think?

Machines can think? A simple question that can today lead us to the conquest of space, cure of ancient diseases and really renewable energies, or a mass unemployment and the end of capitalism as we know it. After all, what is and when artificial intelligence emerged, what is likely to really happen and what is just a bunch of hype trying to get your attention. And how to take the most advantage of what some foresee as a major change than the internet and mobile together. It all starts in 1943, when some scientists begin to try to replicate the structure of the human brain in neural networks. In this, we reach 1950 with Alan Turing, who asked this first question, and also develops the Turing test, a very famous one used until today, where we put a person to be the questioner and talk to two different people, one a real person and another a machine. If this person who is talking did not know how to differentiate one from the other, then the machine passed the test. Already later, in 1956, at the Dartmouth conference, they define what artificial intelligence is, with some of the founders who created this field in humanity. And with that, they define artificial intelligence is the ability of a machine to perform tasks that require human intelligence, such as learning, reasoning and solving problems. In this, in 1958, the Perceptron by Frank Rosenblatt, which is the first model of artificial neural network, still in a very different way from the current ones, but already giving some clues of what was to come, until 1959, when Arthur Samuel developed machine learning teaching a computer to play ladies, even winning the state championship at that time. In 1965, we have the first chatbot called Elisa, who was a psychotherapist who loved to ask good questions, to finally reach 1980, where deep learning, one of the techniques that we use the most today for large neural networks, begins to be developed. This resulted in artificial neural networks and also a little more advanced learning algorithms than we had before, which helps us to develop other algorithms, such as natural language processing, as the GPT chat uses, computational vision for images, videos and others, robotics and even generative for the creation of images and videos. In 1986, the back-propagation came about, and it became popular as one of the best methods to train these networks. So, instead of just feeding the machine forward, it also now starts to give feedback backwards, improving the models. This until, in 1988, one of the great winters of artificial intelligence came about, which, due to a limitation of processing and others, ends up blocking many of the developments of the time. But still, there were some interesting ones, such as this model, a demo, de uma rede convolucional que foi usado pelos Correios nos Estados Unidos e também para a leitura de cheques em grandes bancos, onde se reconheciam caracteres da tela de um computador. E é aqui que algumas das redes neurais artificiais mais famosas e mais usadas começam a ganhar forma. Algumas delas, por exemplo, a rede neurais de Feedforward, antes de existir a RetroProp, convolutional neural networks like the one we just saw for images, recurrent neural networks and deep neural networks generative, as we use for mid-journey and others, for images and videos. All of them are basically different ways of trying to rebuild our brain and how it learns to consume data. Already in 1995, it starts to talk about intelligent agents with one of the great applications in 2002 that we will already see. Because in 1997, the D Blue, from IBM, defeats Kasparov. But here's an interesting story, because despite him having achieved this feat, this was a simple and direct system, there was no neural network or deep learning, as we do today. In fact, they were classic techniques based on advanced search algorithms and heuristic evaluation functions. What does that mean? That it wasn't an intelligence that really learned to play. They only knew how to search for the answers in a faster and more optimized way, unlike what we'll see soon. In the 2000s, an explosion of machine learning and computing power emerged, which opens the doors many other things, among them different types of learning, such as supervised, non-supervised and by reinforcement. Again, there are other ways of how to train these neural networks to achieve the desired result. Some of them work better for some models and others for others. Already in 2002, the Roomba one of the first robots to clean up for artificial intelligence, which is one of the agents we have today in several houses. In 2012, one of the great marks of artificial intelligence when AlexNet won the ImageNet competition. What was that? It was 12 million images in 22 thousand different categories, where a neural network, an artificial intelligence, needed to find a small parameter which was the reference image. This was a very difficult standard and search mechanism at the time, and this one, with a convolutional neural network, winning with an error margin of almost 10% less than all the others. Already in 2016, another great advance and a great milestone now in the Go game. This is a game where there are more possible games than atoms in the observable universe. And why is this important? Because, on the contrary,