AI Singularity: the self-awareness of machines

How long before an artificial intelligence (AI) will equal or surpass the human one? We may be far, far away, or perhaps closer than we think.

According to Dario Amodei, researcher and CEO of Anthropic, a company that develops AI models, machines will reach a level comparable to humans as early as 2026. In 2009, Shane Legg, co-founder of Google DeepMind and USI alumnus, pointed to 2028 as a possible date for reaching that goal. According to Ray Kurzweil, one of the most influential thinkers in the field, it would be just over five years; eight years instead according to Ben Goertzel, the researcher who popularised the acronym Agi (Artificial General Intelligence) at the turn of the century, now used as a synonym for superhuman AI, i.e. capable of doing everything a human being can do but at a higher level. This is a prerequisite towards that evolutionary stage of AI that scientists call "singularity" and which in films and science fiction books is referred to as "machine self-awareness". According to Kurzweil, the moment is when we will merge our minds with the cloud. When will it happen? Around 2050 according to him, while Goertzel has not set precise dates, although the time horizon he suggests is similar. Amodei, on the other hand, does not commit himself but speaks generically of a super (powerful) AI in 2026 ("even if there is the possibility that it could take much longer", he points out) and of its possible development in the next five to ten years.

Despite the doubts, the research continues

Let's clarify: not everyone in the industry agrees that we will eventually reach the singularity, whether near or far. Achieving this milestone would require a significant technological leap, and since the necessary knowledge to attain it does not currently exist, some experts believe we are still a long way from Artificial General Intelligence (AGI) and even further from the singularity. In fact, some of them assert that we may never reach that point at all. Nonetheless, research continues, so - in theory - we are constantly approaching the development of AGI and the singularity: systems like Gpt-4 or Claude Sonnet 3.5 have shown remarkable linguistic abilities, self-driving cars are getting better and better, systems like AlphaZero not only systematically beat all human champions in complex games like chess and Go, but also learn by themselves how to do it from just the basic rules of the game. Is this good or bad? Opinions are also divided here and fluctuate. Similar to another technology guru, Bill Gates, who expressed concerns in 2015 but has since aligned himself with the optimists, the future remains uncertain for everyone. However, one thing is clear: artificial intelligence—whether it eventually leads to AGI or Singularity—is already bringing and will continue to bring significant changes to our society.

Between optimism and pessimism, ethics remains the beacon in uncertainty

Beyond the prevailing catastrophist strand of today's science fiction (from the Hal 9000 computer in the film "2001: A Space Odyssey" to the sentient machines of "Terminator" or "Matrix" the examples are endless), the scientific community oscillates between the optimists who speak of a future golden age, a society with almost no human labour, a universal basic income and an ever longer life span, and the pessimists who warn of possible social upheavals, with ever greater disparities between those who will have access to advanced technologies and those who will not. Therefore, all agree on the need to develop AI in an ethical and controlled manner to maximise the likely benefits and minimise the possible risks. This is one of the most complex (the cultural, social and ethical diversity of humanity makes it very complicated to define a universal set of values and goals) and crucial challenges (AI can develop unpredictable behaviour patterns even for programmers and trainers) that scientists and researchers face today. Perhaps, however, this is a false problem: as recently stated by Paolo Benanti, a scholar of ethics, bioethics and technology ethics, what we really have to fear "is not artificial intelligence, but the natural stupidity" of humanity, which has given endless proof of itself throughout history and will probably never cease to amaze us. In this sense, perhaps, if one day, near or far, a machine will ever equal or surpass us, it might not be such a bad thing.

The interview - Luca Gambardella: "Coexistence probably manageable"

Probably few people know this, but when talking about artificial intelligence (AI), one must remember that one of the first research centres in the field in Switzerland and Europe was established in Lugano in 1988. It was at the request of Mr Angelo Dalle Molle, who found in Canton Ticino a fertile environment for his projects and established the institute that bears his name: Dalle Molle Institute for Artificial Intelligence (IDSIA USI-SUPSI). Luca Gambardella, now Pro-Rector for Innovation and Business Relations at USI, Professor of Artificial Intelligence at the Faculty of Informatics and Director of the Master's degree programme of the same name, was one of the first employees of IDSIA, which he directed from 1995 to 2020 and to which he still belongs. Thirty-six years ago, however, "a "toy" artificial intelligence was being designed, certainly not addressing today's problems". Even then, however, there was a "visionary side" precisely because Dalle Molle "was interested in the quality of life of human beings in relation to technological evolution". And on this, Gambardella has no doubts: "Today, AI is a tool that helps us live better, even if we are not working less, on the contrary. We are, however, the generation experiencing an epochal technological change".

The hybrid world

In fact, we now live "in a hybrid world", where "the real challenge is to equip ourselves with the tools to deal with it, to have the ability to interrogate AI and interact with it, to look at it very much as a tool to support humans", certainly not as a threat. It is not for nothing that "last year, an amount of money more or less equal to the GDP of the United States was invested in AI worldwide". An enormous sum which would be meaningless if there was no use for it, present and future, in numerous fields. As proof of this, two recent Nobel prizes, one for chemistry and the other for physics, have been awarded to AI scholars. "There is recognition from science, not just from business. This technology is revolutionising many sectors; we are in the midst of it, and perhaps we don't fully understand where we have ended up". And where we are going.

To date, human consciousness remains unique

And here, on developments towards a proto-human AI, if not even the singularity, Gambardella puts the brakes on. "The idea of understanding whether a machine has human intelligence," he explains, "goes back as far as the origins of artificial intelligence, to the 1950s and Alan Turing". He invented a test, which bears his name, precisely "to understand whether a machine is intelligent or not". The results are unequivocal: "To date, no device has passed it". Hence Gambardella's doubts. "We informatics, somewhat down-to-earth people, if they explain to us how something works, we can write a computer programme that replicates those mechanisms; but if I went to a neuroscientist and asked him to explain to me how exactly the human mind works, he would not be able to answer me. Not to mention consciousness, a mystery in both computational and philosophical terms". To sum up, we have progressed to supercomputing and notionism, but advancing to the next level is certainly not immediate. At the moment, therefore, we have to be content with the fact that we have an AI "that works very well, even if it is a parrot (Noam Chomsky said, ed.): it is a probabilistic machine, which means that by dint of reading books, it gives the most likely correct answer". This implies that "according to our criteria of evaluation, machines today are not intelligent; they have very advanced capabilities", but equally high limits: "If it looks me in the eye, it does not understand what I am thinking, even though its eyes see much better than mine; if it talks to me about the sea, it does not know what the sea is, even though its answers will all be very sensible and correct". Logical consequence, if he has doubts about the possible advent of the singularity, let alone a potential date. "Those who do so speak without presenting a scientifically sound way forward. A machine that understands, evaluates, is aware of itself, of what it says and understands the context is something I don't see today". Which is not to say that it can never happen. "I leave a door open, but I don't see a way forward today", even though AI has made a remarkable leap forward in recent years.

Catch 22 or a third way?

And yet, there are machines that learn by themselves, from their own mistakes or from their peers (a bit like humans): these are the latest generations of software for complex games, such as AlphaGo and AlphaZero, which have become unbeatable just by knowing the basic rules of the game. How was this possible? By making them play against themselves. "It's a bit psychopathic, but the opponent inside a computer can be the programme itself". The machine thus learns both how to win and lose and gets better at each game. And it also devises unconventional moves and strategies. "This is where the singularity might become not an absurdity: not in a machine that reads so much, but in a machine that evolves in an almost human-like way". The consequences so far are disturbing, to say the least. "Some time ago, two machines generated a language that only they understood to exchange information faster. A new, unknown form of digital language. Can I transfer it to another machine? I don't know. But this point of discontinuity, therefore, the idea that there could be a new kind of consciousness, that ours could be just one of many possible consciousnesses, fascinates me. There really are endless possibilities here". Positive or negative? Star Trek or Terminator? Gambardella is positive; it does not frighten him that machines have and acquire consciousness because "We are all accustomed to working with people who have consciousness and can read the emotions in our eyes.", so "a world where humans and machines interact with the same rules and logic should be manageable". Maybe better.