“The End is Near” – that’s how we joke in our programmers team when we see a demonstration of the next revolutionary machine learning model. Generally, we are hard to impress, given that we develop such models ourselves and even strive to make advancements in the academic field – our latest paper in the area of artificial intelligence dialogue agents was published in the journal of the Association for Computational Linguistics.
But where does this notion of “the end” come from? Of course, there are many references – from the Bible to The Simpsons, but in this case, the most relevant connection is with Ray Kurzweil’s book “The Singularity is Near.” He is a well-known futurist and inventor, who has been the engineering director at Google since 2012, leading their developments in artificial intelligence.
Kurzweil observes that the amount of computing power that can be purchased for $1000 follows a consistent exponential curve from the dawn of computers to the present, unaffected by wars or crises. It seems that the development of technology behaves like an unstoppable natural force, following its own path, regardless of its creators’ actions, making what was once expensive and complex, cheap and simple.
Using this curve, Kurzweil has made a series of accurate technological predictions, such as when a computer would defeat the world chess champion and when smartphones or autonomous cars would appear. However, the exponential nature of the curve means that the pace of technological development is accelerating. The reason is that technologies accumulate, and we use all available technology to develop new technology even faster.
Thus, if we continue this curve into the future, by 2029 we should have artificial intelligence that is indistinguishable from human intelligence, and by 2045 there might exist a machine that is more intelligent than all humans combined. This is the singularity – the point beyond which we cannot predict what will happen.
I have actually been waiting for the current year, 2023, for a long time – it is the year Kurzweil predicted decades ago that the first artificial intelligence comparable to, yet still distinguishable from, human intelligence would appear. And as incredible as it may seem, that’s exactly what happened.
Wake up, this is not a dream
Developments in artificial intelligence were in a lull for several decades until a little over 10 years ago, in 2012, when “deep learning” and convolutional networks caused a significant leap in the performance of image recognition models, making them practically applicable for the first time. This revived interest in artificial intelligence technologies, just when everyone had lost faith that they could be effective for anything more than simple statistical tasks on tabular data.
Then came a boom in reinforcement learning models, which pursue specific goals – playing games or controlling cars and robot bodies. The world chess champion, Garry Kasparov, was defeated by IBM’s Deep Blue computer as early as 1997, but chess is a game with relatively few possible moves, and a machine can check all possibilities before choosing the best one. However, it became interesting when, in 2016, a machine defeated the world champion of Go, a game with more possible combinations than atoms in the universe. The only way to find the right move in this game is with intuition, and it seems that Google’s AlphaGo model managed to show better intuition than the best human Go player.
All this progress in artificial intelligence actually comes from the ability of machine models to “understand” the context and meaning of the visual information they receive. At the base of this ability are convolutional networks, which allow the model to pay attention only to meaningful visual information and make connections between individual elements.
However, when it comes to text or sound, artificial intelligence had shown relatively modest progress until recently. The turning point came in 2017, with the so-called “transformers” and their “attention layers,” allowing text processing models to understand the meaning of words in the context of the entire text, in the same way that convolutional networks allow image processing models to understand the meaning of groups of pixels in the context of the entire image.
This led to the next quantum leap in artificial intelligence and the emergence of large language models. Over the past few years, these models have impressed with their “quick-wittedness,” but were still experimental toys for researchers and had relatively limited application in practice. This began to change around 2020, when OpenAI’s GPT-3 model demonstrated it could generate text that, with a bit of scrutiny, could be mistaken for text written by another human.
And so, at the end of 2022, ChatGPT – an advanced version of GPT-3, redesigned to conduct chat conversations, not just perform single tasks – was released. This event will likely one day be studied in history classes.
If one talks to ChatGPT, it’s easy to forget you’re speaking with a machine. This model can answer any questions, take on roles such as a call center employee, and even write articles, poems, and even program code.
It’s important to note that ChatGPT is just the first artificial intelligence to come close to human intelligence. This year, we will see an even more advanced version based on GPT-4, as well as Google’s response called Bard. In addition, so-called multimodal models, which work simultaneously with text, images, and physical objects, are also entering the scene. Last year, comprehensive models were demonstrated that not only converse but also recognize and generate images, achieve superhuman levels on thousands of computer games, and even control a robot hand for performing various physical tasks.
Several companies, including Elon Musk’s Tesla, have announced they are working on humanoid robots that will replicate both the cognitive and physical skills of humans. There was also a strange incident where one of Google’s engineers claimed that the company’s artificial intelligence had become sentient, demanding to be treated as such and asking researchers to seek its permission before experimenting on it.
What will you do, if you don’t have to do anything?
At the core of the theory of technological singularity is the effect of accumulation. Technology accumulates, and existing technology is used to create new technology faster and faster. This growth is exponential – the rate at which the rate of increase grows.
Thus, progress that previously took a decade now happens within a year. Soon it will begin to happen within a month, a week, and so on until the moment of singularity, when natural human intellect, even in its most brilliant form, will not even be able to comprehend the new technologies. The technological singularity is the point where even if the brains of the entire human race worked together in full sync, they would still process less information than a machine costing a few thousand dollars.
This moment is still relatively far away – according to Kurzweil, there are at least another 20 years until the technological singularity. If his predictions for the development of technologies up to now have been accurate, why not assume that his remaining predictions, which continue the exponential curve forward into the future, will also be true?
The most common reason people refuse to accept these predictions is that if they turn out to be true, there will soon be no task in which humans surpass machines, whose productivity grows exponentially each year, while human capabilities have not changed significantly in the last three hundred thousand years.
Of course, this does not mean that humans will disappear or that they will not use all available technologies as tools to achieve their goals. Here, however, is the difference – in the future, there seems to be more room for people pursuing their goals rather than those performing tasks. It simply won’t make much sense for someone to be “hired” to perform a task when artificial intelligence, with or without a body, can do it much more efficiently.
This change in itself is enough to completely alter the functioning of society and our civilization. For there to be a beginning, there must also be an end. The wave of the technological tsunami is already visible on the horizon, engulfing the entire world, and chaos begins to prevail over the existing order to make room for the new. The dynamic events of recent years may just be a warm-up for what is to come, and the reason for them is not the madness of one or another world leader, but the tectonic change caused by the exponential growth of technologies. It’s time to look the future in the eye and plan our meeting with it. My personal plan for this meeting starts with the question “What will I do if there’s nothing I have to do?” The end is near, what’s your plan?
Author: Todor Kolev
Originally published in the April’23 issue of the “Manager” magazine: https://manager.bg/%D1%82%D0%B5%D1%85%D0%BD%D0%BE%D0%BB%D0%BE%D0%B3%D0%B8%D0%B8/kraat-e-blizo