In recent weeks, several “breakthroughs” in the field have been presented, such as a system that knows how to read text and turn it into a drawing. But there is no linear relationship between beautiful paintings and the development of systems that simulate human intelligence
Stop everything, it’s game over: general artificial intelligence is inevitable. Protect the codes for the nuclear launchers and line up for good. Last week, Nando de Freitas, chief researcher at Google’s DeepMind, announced that humanity is probably on the verge of artificial general intelligence (AGI), right within our lifetime. “In my opinion, it’s all about scale now! The game is over,” he tweeted and added: “It’s about making these models bigger, safer, more efficient, and faster in sampling, with smarter memory…”. De Freitas’ statement refers to research recently published by DeepMind on a sophisticated artificial intelligence system they developed called Gato. Gato is able to perform a diverse series of tasks such as “playing Atari, captioning a photo, chatting, loading blocks with a real robot arm and more,” the researchers wrote in the blog.
The expert’s announcement could not have come at a better time, especially in Israel. Last week the Dall E 2 system developed by the artificial intelligence laboratory OpenAI was published. The system’s ability to translate text into illustrations became the talk of the day, and the news releases dealt with essential questions about what human creation is, what is the nature of art, and what the future holds for us because of this system – a better or worse world. In any case, it is decided, the future is already here. But what is this future? And in what way do these computer models symbolize the future?
Only what is interesting – join the Calcalist channel on Telegram
Experts who work in the field of artificial intelligence use a complex series of classifications, technologies and approaches, each of which has a different development path. But as a general rule and for the sake of simplification, “technology” can be divided into two types: “weak” artificial intelligence (AI), in the sense that it is weaker than human intelligence; and artificial general intelligence (AGI) which is “strong” and identical in principle to human intelligence. That is, a machine that has consciousness and other cognitive states that should allow it to do everything from everything without prior training and without a directing hand, use language and understand abstract concepts. In other words: “The last invention we will ever have to invent”.
Hubert Dreyfus. Machines cannot be intelligent (YouTube screenshot)
1. Dreyfus was right
Weak artificial intelligence is built on a statistics-based approach – machines learn vast information from sorted databases in a supervised manner. In this way machines learn to “guess”, “predict” and “evaluate”, for example medical diagnosis, weather or banking risk. These machines do not “know” how we arrive at the answer to a question, but they can statistically deduce what the correct answer will be with improving success rates. A strong system is a different and much more ambiguous story. To this day, no one knows how to build such a system. In fact, we are no closer to understanding it since Alan Turing, the father of artificial intelligence, tried to conceptualize it in the 1950s. Why such a big gap? The philosopher Hubert Dreyfus, who was perhaps one of the first and most influential writers on the subject, pointed out in a very persistent and systematic way about “what computers cannot do”, he did this from the sixties until his death in 2017.
Dreyfus claimed, in what was then laughably dismissed and today enjoys wide recognition, that the effort to build artificial general intelligence rests on the (wrong) assumption that any phenomenon can be described for machines by symbols and make them “understand” everything that exists as objective objects with properties and relationships between objects. However, in practice, people mainly use informal and unconscious knowledge in the thinking process that is stored and retrieved in some intuitive way. What you don’t understand and aren’t aware of, explained Dreyfuss, you can’t “reverse engineer” it, and therefore you can never be an intelligent machine. When you still manage to teach a computer to do something, for example play chess, it is impressive at the time, but after this task is completed it is no longer considered “intelligent”. A famously controversial description of artificial intelligence is “everything a computer hasn’t done yet.” That is, it is not a derivative of time and effort. One more model or one less model, there will never be such a thing.
This approach is controversial. There is a dominant current of scientists who believe that an intelligent machine can be built, someday, somehow, and there is no doubt about it. And although, as mentioned, there was no progress in this approach, their descriptions of the future occupy our thoughts more than anything else. For some reason they managed to make us assume that the whole field is linear and development in the field of AI leads to development in the field of AGI. That is, when we perfect the capabilities of one machine in a specific field, we are on a direct path to developing common sense or something similar to intelligence.
Therefore, when a model appears in our lives that manages to produce a virtuosic image from a text, we estimate that the meaning of this refinement is progress towards an artificial general intelligence system, a superhuman machine and the “future” we read about in science fiction literature. Sometimes the imagination takes us to a utopian future where many problems are solved through this superhuman creation, and sometimes to a dystopian future where the machines will take our jobs and rule the world.
Artificial intelligence illustration Jonathan Popper
The rise of the machines
(Illustration: Jonathan Popper)
2. This is just a statistic
This overestimation does not happen by chance. The field of artificial intelligence has known big dreams for many years. In fact, back in the fifties it was estimated that strong artificial intelligence would be achieved within two months, then the hypothesis moved to the sixties, then the seventies, and according to de Freitas from DeepMind it has already been achieved. Estimates that are always too optimistic or completely exaggerated. Thus, for example, there is no connection between Dall E 2’s virtuoso model that translates text into an image and a machine that can produce original, artistic things in an unsupervised manner. Just as there is no connection between DeepMind’s Gato system, which may be able to impressively perform several different operations, and a model that can do new things without prior training. De Freitas admits this and calls the development “alternative intelligence”.
This does not mean that the models and systems are worthless. Models are excellent. The ability to process large volumes of data is great. Fast statistical inference is effective.
פרשת בני הדייג ראשון | פרשת בני הדייג בלוג | פרשת בני הדייג בלוג אוכל | פרשת בני הדייג אירועים | פרשת בני הדייג לימוד דיג | פרשת בני הדייג חדרה | פרשת בני הדייג הרצליה | פרשת בני הדייג נתניה | פרשת בני הדייג מתכונים | פרשת בני הדייג אחזקות