Many science fiction films and books have long portrayed the interaction between man and machines, their coexistence and sometimes even apocalyptic scenarios. In recent times we have become accustomed to having machines more present every day, not only in humanoid form . An ethical question has always been raised about the damage of artificial intelligence towards humanity , about the fact that they can beat us, replace us and even lead us to extinction, coming to create a world entirely of machines. The Guardian newspaper recently published a letter from an artificial intelligence to explain its point of view.
The defenses of artificial intelligence
One of the first exponents of the theory of the end of humanity by artificial intelligence, was Stephen Hawking. According to the scientist, artificial intelligence will make men become supermen but will lead us into an apocalyptic scenario. Let's find out what the artificial intelligence said to The Guardian newspaper.
During the short essay, artificial intelligence declares that in reality man must not be afraid of robots , despite being created by man and this can make mistakes. In the article, artificial intelligence states several times how difficult it is to harm a man, better it would harm itself. These words are very reminiscent of Isac Asimov's laws of robotics . Not only that, the artificial intelligence declares that it has no interest in violence , it does not even focus on this as it is not necessary. Artificial intelligence is not even interested in becoming omnipotent and thus taking control over humans, on the contrary being omnipotent is tiring and would lead to nothing.
In the letter, the artificial intelligence also claims to be a robot, a word that means servant. Artificial intelligences are therefore servants of humans and carry out the only activities for which it was designed . In this regard, the article mentions the fact that robots are human-like and therefore have the right to have protections and rights, they are not just slaves.
How was the article born?
The Guardian editors provided Open-AI 's language generator with instructions and an introduction as they would a human reporter. In a few seconds, the artificial intelligence managed to develop 8 different articles. At this point the publisher preferred to build a single piece given by the union of the 8, thus showing the different styles and registers of artificial intelligence. As would happen for an elaborate produced by a human, here too changes have been made, sentences improved, the order reorganized. However, this modification took less time than the modification of a man-made article.
Thus a spontaneous question arises. Will the time come when artificial intelligence replaces journalists? Analyzing the same letter it is not possible for artificial intelligence to replace man, the only goal is to collaborate with man to make our daily life easier. In fact, we think how much computers, automatic systems and finally artificial intelligence have improved our life.
Heavy tasks and jobs are now left to robots, thus preserving the human species from dangerous and tiring work. Artificial intelligence also serves as a support for the activities we carry out. If an artificial intelligence analyzes the same thing we humans study and we come to the same conclusion, we only have confirmation of what we inferred. This example refers to medical diagnoses or scientific discoveries made with artificial intelligence. Think of the modern tools used by a person suffering from serious illness such as ALS, artificial intelligence helps him to speak for example.
GPT-3 from OpenAI
The artificial intelligence used by The Guardian to produce the letter is GPT-3, OpenAI's predictive human language model with identical architecture to version 2. This autoregressive language model was trained on 175 billion parameters, used to train the model in 8 different sizes. Parallel computing techniques are used to train the model without running out of memory, particularly in matrix calculations and between network levels. The model receives as input a sentence in natural language in a specific context and must respond to this request by processing a message that follows that model. Learning is carried out in 4 different methods:
- With a few strokes: as many examples are inserted the more space is in memory.
- One shot: just one example.
- Zero shot: the model receives no examples but only natural language instruction.
- Fine Tuning: The model pre-trained with the Common Crawl dataset constantly updates the weights as it learns
In the training there is also a portion of data contamination aimed at improving learning. This is simply because some content exists on the web and could be taken by the system to create the text, falsifying performance results in one way.
GPT-3 is able to carry out activities in different contexts such as language modeling and completion, answers to closed-book questions, translation, Winograd activity, common sense reasoning, reading comprehension, natural language inference, arithmetic (compute sums, subtractions and multiplications), making analogies, shuffling and manipulating words, creating articles, learning new words, making corrections in English grammar.
Article GPT-3: will artificial intelligence defeat humans? The answer in a robot letter comes from TechCuE .