The rapid progress of learning algorithms will generate machines of intelligence comparable to ours in the decades to come. Capable of speaking and reasoning, they will have their place in a myriad of fields, such as economics, politics, and, inevitably, war. The birth of real artificial intelligence will profoundly affect the future of humanity and condition the very existence of such a fate.
Take, for example, the following quote: "Even today, research is underway to understand better what new AI programs will be able to do while remaining within the limits of today's intelligence." Most AI programs currently programmed are mainly limited to making simple decisions or performing simple operations on relatively small amounts of data. "
Perhaps you had the impression that something was wrong in this paragraph? This quote is the work of GPT-2, a language robot that I tested last summer. Developed by OpenXAI, a company in Chicago that promotes "virtuous" AI, GPT-2 is a learning algorithm based on an artificial neural network. Its entrails contain more than a billion connections simulating synapses, the junction points between neurons.
The task of the network is stupid when confronted with an arbitrary starting text, it must predict the next word. He does not "understand" the documents as a human would. But during his learning phase, he devoured astronomical quantities of texts - eight million internet pages in all - and adjusted his internal connections to anticipate word sequences better.
I wrote the first sentences of the article you are reading, then "injected" them into the algorithm by asking it to compose a suite. In particular, he spat out the paragraph cited. Admittedly, this text resembles a first-year student's efforts to remember an introductory course in machine learning, during which he would have daydreamed. But the result still contains the keywords and phrases - not wrong, really!
Intelligence is not consciousness.
The successors of these robots risk triggering a tidal wave of fake articles and reports, which will pollute the internet. It will be just one more example of programs performing feats that we thought were only for humans: playing strategy games in real-time, translating text, recommending books and movies, recognizing people in pictures or videos.The task of the network is stupid: when confronted with an arbitrary starting text, it must predict the next word. He does not "understand" the documents as a human would. But during his learning phase, he devoured astronomical quantities of texts - eight million internet pages in all - and adjusted his internal connections to anticipate word sequences better.
I wrote the first sentences of the article you are reading, then "injected" them into the algorithm by asking it to compose a suite. In particular, he spat out the paragraph cited. Admittedly, this text resembles the efforts of a first-year student to remember an introductory course in machine learning during which he would have daydreamed. But the result still contains the keywords and phrases - not wrong, really!
The successors of these robots risk triggering a tidal wave of fake articles and reports, which will pollute the internet. It will be just one more example of programs performing feats that we thought were only for humans: playing strategy games in real-time, translating text, recommending books and movies, recognizing people in pictures or videos.
Will algorithms one day write a masterpiece as successful as In search of lost time? Hard to say, but the beginnings are there. Remember that the first translation and conversation software was easy to make fun of, as it lacked finesse and precision. But with the invention of deep neural networks and the establishment of robust computing infrastructures by digital companies, computers have improved continuously, until their productions are no longer ridiculous. As we see with the game of go, chess, and poker, today's algorithms are capable of surpassing humans (as of last November, Lee Sedol, one of the greatest Go players in history, decided to retire after losing several times against the AlphaGo algorithm; he declared that it was an entity that could no longer be defeated, note). So much so that our laughter freezes: are we like Goethe's sorcerer's apprentice, having summoned helpful spirits that we can no longer control?
Much of the brain activity remains localized and, therefore, does not gain consciousness. This is the case of the neural modules which control the posture of the body or the direction of gaze. But when the activity of one or more regions exceeds a critical threshold - say, when we present to someone the image of a delicious treat - it triggers a wave of neuronal excitement which propagates through the workspace, throughout the brain. This signal then becomes accessible to a multitude of auxiliary processes, such as language, planning, the reward circuit, long-term memory, and storage in a short-term buffer. It would be the fact of disseminating this information on a global scale that would make it aware. So, sugar and fat shoot to come.
