This article follows up on one published a few days ago on artificial intelligence (AI), a topic that elicits contrasting interpretations. Regarding the meaning of AI, indeed, we are faced with two diametrically opposed attitudes: there are those who believe that AI, like true intelligence, can also provide “new” results, i.e., not contained in the data previously provided to it, and those who think it simply recycles the enormous amount of information that was fed to it during the training phase, thus acting as a “rake” of information (for example, see the interview with Roberto Battiston in Alto Adige, May 5, 2024).
This second opinion is seemingly contradicted by some results that AI has recently achieved. Take, for example, the software called AlphaFold, a system developed by DeepMind (https://alphafold.ebi.ac.uk), which has managed to model, in a way that seems substantially correct, the three-dimensional structure of all the proteins encoded by a genome starting from more than 200,000 experimental three-dimensional structures present in the PDB database (Protein Data Bank, www.rcsb.org). Is this a new result? It is, in the sense that numerous structural prediction software based on previous algorithms were not able to provide the same level of response and, above all, were not able to predict the model of proteins with sequences completely different from those in the database. It is true that AlphaFold would not have achieved this result if it had not been able to use the models of all the structures determined with experimental techniques over the last sixty years, but it nonetheless provides researchers with information that was not previously available.
Similarly, it is true that ChatGPT, another example of AI, is able to translate any text into a language different from the original very efficiently. It does this by exploiting the huge amount of bilingual texts with which it was previously trained. Even in this case, an original text in Italian whose sentences are not present in the set of texts used for training is translated into English in a more than good manner, anyway better than what could be provided by a person with an average knowledge of English.
Is all this enough to define it “intelligence”? It depends. If by intelligence we mean deductive-logical capabilities, then probably yes. As far as logical reasoning capabilities are concerned, the algorithms used in neural networks that constitute the basic principle of AI seem to have already reached a more than decent level. It can be objected that they can do this because they use a large amount of information that constitutes the cultural base of the field a certain AI deals with, but isn’t that what we also do? Even humans spend a lot of their time, especially in their youth, studying, reading, informing themselves, in a word, training. When we solve new problems, we almost always do it thanks to what we have learned before.
What distinguishes us from a computer is what is defined as “self-awareness” (Federico Faggin, Silicon, 2019; Irreducible, 2022). The computer not only lacks consciousness but does not experience feelings, pleasure, suffering, joy, and so on. This is the limit of artificial intelligence (at least for now, and fortunately, there does not seem to be any progress in this sense).
(The English version of this text has been translated using an AI, ChatGTP-4).
No comment