The Gospel according to a robot

She could. You could do that. It could be that. It could be a word. Another word with the highest probability. While talking to you, language bots like Google's Gemini, Chat GPT-ja from OpenAI, or Antropic's nifty Claude platform, form sentences by making a prediction about what is the best next word to say in a sentence. It could be the word that, with the ever-faster development of artificial intelligence (AI), will change the world forever.

It may seem quite impossible to you that machines that talk to you so smoothly (and there are also applications on the horizon that offer psychological advice, support, and even AI partners) use such a banal method. Or, on the other hand, you don't think it's God knows what. But this technique works so well that, based on powerful mathematical models and trained on a huge amount of data, the robots have unexpectedly skillfully mastered human language and not only learned to speak, putting words together by probability from the so-called nested vector space, but actually do it in such a way that in the last two years they have become our everyday life.

All this was not a particularly important topic during the elections in the United States of America. Or at least she wasn't until they ended. You may not have noticed, but in the sea of ​​news about the future members of Donald Trump's administration, the relationship with Russia and China, the Western media report that there are smaller VI companies, but also technological giants that develop artificially intelligent software for conversation, programming, surveillance, translation, photo production and video content, medical applications, and even for all kinds of military needs, extremely happy about Trump's victory. And that the world, currently worried about the terrible ongoing wars, the nuclear threat and climate change, may soon be faced with a challenge that is probably bigger and certainly more intelligent than all the previous ones - before the emergence of the so-called general artificial intelligence.

ROBOTS IN THE WHITE HOUSE

Machines that are trained on huge amounts of data often violate the copyrights of the people who produced certain contents used in the training, are very often "infected" with the prejudices that people also have, so they show racist characteristics when making decisions (when choosing a person, for example, which allegedly pose a threat at the airport) or can be used for various media manipulations. Because of all this, along with the development of AI, certain bans have been introduced in many countries, including the USA, while AI companies believe that this is hindering their development and that the Western world will lose this race with China.

That's why American AI companies are particularly happy about the victory of the Republicans, their shares are rising, along with the expectation that various bans and regulations will be lifted after Trump is enthroned in the White House. It was part of Trump's package of promises, which was particularly highlighted by vice-presidential candidate J. D. Vance. Trump's important ally, Elon Musk, has repeatedly spoken out in public against the unrestrained development of AI, which raises the question of whether Trump will stick to his line, but it doesn't necessarily mean anything. After all, Musk also owns a part of the AI ​​business, and his company xAI has also developed a language model - their robot Grok is rated as a very successful chatbot that gives very cynical and darkly humorous answers. It is not yet available in Serbia, but it is certainly part of Musk's portfolio.

Sometimes in history things develop in such an unexpected way. After Lyndon Johnson more than smoothly defeated Barry Goldwater by an overwhelming majority in 1964, few could have thought that his popularity would melt away so quickly and that his term would eventually, and subsequently in history, be marked by the Vietnam War . Although the situation in Indochina was still very complex in Kennedy's time, for most Americans it was a marginal problem, something that exists but is not the backbone of American history, only to turn into a general trauma just a few years later. What if, again, among all the other topics discussed after the US election, a marginal thing like AI ends up being the crucial legacy of Donald Trump? Indeed, despite so many other difficult questions, is the development of AI actually the greatest calamity of our age?

The famous author, evolutionist Yuval Harari last year compared the uncontrolled development of AI to the invasion of aliens, pointing out that language is "with human culture", and that these machines have just usurped it. Whether you agree with him or not, the mastery of human language that AI machines have achieved, so that they communicate in all natural languages ​​without much effort, represents the main breakthrough and the central place of the AI ​​revolution from which the next achievement will grow - general artificial intelligence.

ARTIFICIAL NEANDERTHALS

The ability of such models to speak or perform various other tasks does not make them equal to humans because each of these programs still only has a so-called special artificial intelligence - to imitate one of the human cognitive abilities, but not all. That's why the conquest in general considers intelligence to be a true revolution, and some authors traditionally call it the "singularity", which is the moment when man will overcome his physical and intellectual limits with the help of general artificial intelligence.

Singularity futurologists long ago predicted as an inevitable outcome that will happen sooner or later if man continues to develop artificial intelligence. There are plenty of scenarios that have been offered that lead to it, be it the development of deep learning or quantum computers. In each of them there is that moment when somewhere some machine becomes aware of itself and acquires general intelligence. When he starts to really think and - he exists. What happens after that is also the subject of a series of scenarios, but it is certainly the moment when man on the planet will get another intelligent competitor. It is very likely that the new entity (or more) will be of help and support to man, but it may not be.

It is impossible to predict the future, but human history does not suggest that the peaceful coexistence of two masters, and especially two intelligences, is the most likely outcome - the fossil remains of the Neanderthal, who disappeared 40.000 years ago, perhaps best demonstrate this. The trouble with general AI is that it has far greater resources than the biologically limited human brain, primarily because the brain is an incomparably slower system than one developed in a substrate like silicon.

Time for general AI, which will work incomparably faster than biological machines, must run faster and the whole thing for a human can be over in a fraction of a second - immediately after becoming aware, the AI ​​can conclude that the human is the real and only threat to both it and the world, and then (using nuclear bombs or some other means) destroy man. It doesn't have to go that way at all, but it's not completely ruled out either. Man adapts to all challenges, but the biggest of them is that general AI will also be able to do so, which, if it is more intelligent than man, will be a complete unknown to us.

A THOUSAND DAYS

Many pediatricians, as well as cognitive experts who study the human brain, will often say that the first three years, or the first thousand days, are crucial for its development. This is based on numerous observations about how the brain develops, how it learns and learns about the world - unlike other animals whose brains arrive "in a package", the human brain develops more slowly, and this is actually the reason why it masters cognitive skills that others animals do not. The first thousand days are certainly crucial - during them, structures will be formed in the superpowerful network of nerve cells in the baby's head that ensure that someone will be a top pianist, someone capable of controlling the deepest emotions, and someone will be an AI expert.

The CEO of OpenAI, the company that created ChatGPT, Sam Altman, did not choose this number by chance - he, in fact, stated that general AI will be achieved in 1000 days. According to the latest projection, it is already close and his estimate is that general AI will emerge as early as 2025. Most futurologists, however, think that this is too ambitious and that current models cannot pass the famous Turing test.

In the classic Turing test, a human and a robot in closed rooms answer the same questions posed by an examiner from a third room. If the examiner cannot distinguish which answers are human and which are not, we can talk about the artificial intelligence of the machine. This idea served the American writer Philip K. Dick in the story Do androids dream of electric sheep, after which first Ridley Scott, and recently Denis Villeneuve made films under the name Bladerunner. The conversation with ChatGPT is sometimes really difficult to distinguish from a human one, his sentences are quite meaningful, connected to the questions, and the reason seems really convincing, but perhaps what betrays him the most is excessive perfection - it is relatively difficult to believe that a human creature will always give the "right" ones. answers.

The European pioneer in artificial intelligence research, the Spanish scientist Dr. Ramon Lopes de Mantaras, who recently stayed in Belgrade, still thinks that we will have to wait a long time for the general intelligence of machines. Namely, Chat GPT it cannot solve some very simple dilemmas - you will very easily see for yourself, through a simple examination, that it basically does not distinguish between good and evil. Or at least does not distinguish them in human categories. Even without that, he is full of limitations - the data he gives in conversations are completely unreliable, you can get an infinite number of different answers to the same question if you ask them in different contexts, and so-called "hallucinations" have been recorded - when he starts giving crazy, mutually contradictory answers .

It is particularly interesting that if you tell an AI language model that you pierced an apple with a needle, the machine does not know whether the hole is in the needle or the apple. AI can speak, but has no experience with the actual meaning of words because it has not experienced them. Many AI experts, like Ramon, do not believe in so-called hard intelligence because machines do not have a body. And they could not, like us, experience the world.

photo: promo

NEURONAL NETWORKS

In addition to these very optimistic and extremely pessimistic ones, there are also those that give AI room to develop in the next few decades, which is actually not that far away. Twenty years ago, American innovator and futurist Ray Kurzweil in a notable book Singularity is near he opened this topic by analyzing the laws according to which technology develops at an exponential speed and by giving several predictions, among which is the estimate that by 2045 the singularity will appear. This assessment is based on the development of artificial neural networks.

The history of such models began in 1954, when, at the height of the Cold War in the USA, machine translation from the Russian language was developed as part of the Georgetown-IBM experiment. When a few years later, in 1956, John McCarthy coined the term "artificial intelligence", defining it as "the science and technique of making intelligent machines", the expectations from robots seemed much faster to achieve. In the XNUMXs, it became clear to many that mechanical problems seem unsolvable - robots are not able to walk, but neither are they able to "think". This period of almost complete shutdown of robotics is therefore often called the "AI winter", after the "nuclear winter".

However, along with digitization, the development of the Internet and classic computers, there are completely new breakthroughs. With Marvin Minsky's breakthrough in electrical neural networks, the problem of "learning" machines is being considered. It is a model of a network that simulates the workings of the brain and that can be learned by changing the weights that connect the neurons in the network through trial and error. Successes in the training of artificial neural networks with propagation and feedback, as well as the development of stages of logic and expert systems, are reviving the field of artificial intelligence step by step. Neural networks are proving to be a very successful field and with an abundance of possibilities, they attract a huge number of researchers, and so-called recurrent neural networks are used for language modeling.

Interestingly, this year the Nobel Prize in Physics was awarded to John J. Hopfield and Jeffrey E. Hinton for their pioneering work on the development of artificial neural networks. Hopfield's work on networks that can learn and remember, and Hinton's contribution to effectively training those networks laid the groundwork for today's advanced AI systems. By the way, this year's Nobel was all about artificial intelligence because the prize for chemistry is also related to AI. Namely, David Baker, Demis Hassabis and John Jumper were awarded for their revolutionary work on proteins. Baker developed methods to design entirely new proteins, while Hassabis and Jumper, using artificial intelligence (Alpha Fold), were able to predict the structure of proteins based on their amino acid sequence.

Otherwise, there were still various problems in working with neural networks and their training, until in 1997, Hochreiter and Schmidhuber proposed the architecture of the so-called long short-term memory (LSTM). This architecture was replaced in 2017 when Vaswani et al introduced the Transformer architecture. It was developed inside Google, but it caused explosions and many new companies started using Transformer as a basis for new language models. ChatGPT is based on this architecture, along with all other talking robots.

Altman's company, which created it, states that its mission is to develop general AI for the benefit of all humanity. The only question is whether he will be around to see these benefits. In this light, the thought of the famous British science fiction author Ian Macdonald, who wrote that "a machine that is intelligent enough to pass the Turing test, is intelligent enough not to pass it" sounds eerie.