Artificial intelligence is a vast and complex field, which can fascinate as well as intimidate. I talk about it in this interview with Alessio Figalli: one of the most important mathematicians in the world, awarded with numerous international recognitions (including the Fields Medal, the Nobel Prize for mathematicians, won in 2018) and currently Director of the Research Institute for Mathematics at the ETH Zurich.
Professor Figalli, is there a birth date for AI?
“I don’t know if you can really say that there is a precise birth date. There has certainly been a boom, for at least a decade.”
The conference held in Dartmouth in 1956 is often cited: considered by many to be the baptism of AI.
“That was clearly an important moment. But everything went a bit further, and this is perhaps the reason why that conference became so famous: because its initial expectations were largely exceeded. In fact, even up until a few years ago, what we are experiencing today would have seemed like science fiction.”
How would you explain to people who are unfamiliar with AI what AI is?
“Let’s start from the fact that almost all AI users don’t know how it works, just as most cell phone users don’t know how a cell phone works in the same way as many other objects. AI is a set of tools that perform operations by mimicking the way humans think, constantly interacting with them.”
When I think of AI, I imagine it as something that can’t stop, that doesn’t give itself a break, almost as if it were in competition with itself. Why this incessant pursuit?
“There is a need for correction, first of all. These models are not perfect and can make mistakes, so pushing forward also means moving towards improvement. They are exceptional tools, the question is knowing how to use them: if one thinks they are the solution for everything, it is a big risk.”
Has social media paved the way for AI?
“We have certainly developed familiarity with the virtual dimension thanks to them. Social media has accelerated the entire process, even in terms of giving away our information: I’m talking about statistics, general and not specific, which is what AI needs. Through social media we have given away a lot of ourselves: this is what AI feeds on.”
About our data?
“Yes. AI is more powerful and faster than us precisely because it handles an unbelievable amount of it, and not because it makes deductions or activates other intellectual pathways.”
If we humans could handle such a large amount of data, could we perform as well as AI?
“Our performance would be far superior, because we are intelligent, while AI is not intelligent, paradoxically. It is good at composing operations, but it has no intrinsic capacity for distinction. Let me give you a practical example: for a child to understand what a cow is, it needs to be explained to him only once. The child will understand immediately, he will not need to see that same cow from thousands of different angles and perspectives, all things that AI needs, which instead proceeds by associations. In other words, AI does not learn. AI categorizes, creates patterns, proceeds by accumulations and cross-referencing.”
Could we say by observation?
“Exactly. You have to understand that AI does not understand the rules, because it does not follow the usual learning phases. Let’s say you want to learn to play chess. There are two possibilities: either you assimilate the rules and then practice, or you watch all the chess games in the world until you learn those same rules by observation. AI follows this second option.”
So AI is not intelligent. How would you define it then?
“Let’s see it as a source of almost infinite knowledge, which rightly continues to impress us. However, AI cannot be compared to classical human intelligence, because this proceeds in a structurally different way. There is nothing autonomous and independent about AI: and this is a good thing, if it were not so we would be moving into a very dangerous horizon”.
Is AI close to perfect?
“No, because it is probabilistic, not deterministic. Plus it makes mistakes, which is why it needs to be trained and refined.”
Forgive me, but every time I hear about training in reference to AI, I think of a captive animal, a creature that if it escapes from its cage will cause who knows what damage.
“It always depends on our behavior, on the use we make and will make of AI. It can be insidious if left to itself or put in the wrong hands: but this has already happened in history, think of the atomic bomb.”
Did you imagine such a massive presence of AI?
“I didn’t expect such a quick leap. But I’m also a pragmatic person: now that it’s here, let’s figure out how to make it a support and not an enemy. Then: it’s true that it’s a bit everywhere, but relatively. The use we’re making of it is still moderate.”
What is its greatest danger, in your opinion?
“It tempts us to be lazy. And if we take a shortcut every time we risk losing all the intuition, all the genius that ultimately characterizes man.”
Do you think we are telling the world of AI in the best possible way? Or do we indulge in storytelling that is often fable-like and approximate?
“Maybe sometimes we shouldn’t exaggerate, avoiding being catastrophic and alarmist. The fact that information is not always correct is part of every simplification process, in which the AI story is also immersed. The real issue is whether things are distorted, the important thing is that there is a reasonable communication pact”.
What is your specific contribution to the field of AI?
“That of a theoretical mathematician. I focused on some aspects more than others, for example on how AI tries to mimic the human brain, through the functioning of its neural networks.”
Can we say that AI is made of neurons?
“Of course. And one question I asked myself was this: by increasing the number of neurons and therefore neural connections, what results do you get? If you add more and more neurons, how will this affect the machine’s learning process?”
How many neurons can AI currently control?
“Several hundred billion. ChatGPT is based on 175 billion neurons: which is more or less the same amount managed by the human brain. In addition to the number of neurons, however, we must take into account the subsequent neural connections, which are many more: in this last case we are talking about one hundred thousand billion connections”.
Can you describe an AI neuron to me?
“It’s a number, which has then been manipulated thanks to the inputs that the system receives, by means of images or other. It’s a process of constant stratification and activation.”
Can AI neurons experience emotions?
«Yes and no. Let’s say that it would be enough to train them with a sufficient amount of psychological tests to make them understand what is expected from an emotion. AI can learn what an emotion is based on large samples of data, it could also do it by studying people’s reactions to social media stimuli. So AI would not feel emotions, but it would be able to understand them».
Should we also associate concepts such as fantasy, creativity, imagination with AI?
«I wouldn’t rule it out, but again it depends on how we understand these concepts.
After all, if I ask AI to paint a portrait of someone, perhaps in a style halfway between Manet and Picasso, it will do so without any problems. Is this creative? Is this art? These are important questions that remain open, like many others.”