In the 5th year of Anyong (1776), the Japanese Edo period painter Toriyama Ishiyan painted a work that influenced Japanese yokai culture for hundreds of years, called "Hundred Ghosts Nocturnal Chart". It is a fairly well-known work that includes some yokai art inspired by paranormal phenomena, folklore, and more.
But what is less known is that 3 years later, he painted a sequel called "Drawing Pictures of the Past and Continuing the Hundred Ghosts". In this sequel, there is a yokai called "Jue" (さとり), and this sequel differs from the previous game in that each yokai has a description in the sequel.
This ape-like youkai sense, which had an ability that only existed in legends at the time - mind reading. In the upper left corner of the painting is written: "Jue, born in the deep mountains of Mino (Oda Shinsen's house?) ), known as a mountain man, with a long black hair, good at human speech, often imitating human behavior for fun, and good at understanding people's hearts, can perceive malice from humans." The biggest feature of this monster is that it reads minds, which is an ability that human beings have been fantasizing about for a long time, and people are not good at reading through people's hearts like reading books, so they are easy to be deceived and misled.
But Toriyama Shiyan could never have imagined that just 247 years later, humans would have created a similar product, a human-like intelligence, good at human language, and insight into people's hearts.
In the last two weeks, when I discovered a paper in Nature Neuroscience describing how to use GPT to passively read human thoughts, I gradually realized that this ability seemed to be very close to us.
And artificial intelligence has also begun to develop in a more exaggerated direction. In November '22, a paper called "Seeing beyond the brain" proposed a way to convert brain signals into images. (Paper link: https://arxiv.org/abs/2211.06956)
To put it simply, when we look at a picture, we form a visual stimulus that produces complex brain responses, and it stands to reason that humans will never be able to observe these brain activities because we don't have eyes that can see. But humans are good at creating and inventing, and over the years, the maturity of nuclear magnetic resonance technology has given rise to another subdivision called functional magnetic resonance imaging (FMRI).
And this technology happens to specialize in one area, the brain. Its power is not only that it can observe people's thinking activities, but most importantly, it can track and record changes in brain signals in real time. FMRI quietly opened a window for mind reading, but it's not so easy to do it. To be precise, there are two big problems to overcome.
1: How to decode the recorded brain signals.
2: How to generate images.
At that time, the whole concept was not mature enough, and the deeper theoretical level was discussed, and the image effect generated after decoding was difficult to say, although occasionally right, but more mistakes such as identifying elephants as dogs and dogs as bears.
But soon after the paper was published, the time came to the magical year 2023. In this year, the explosive explosion of AI technology has radiated to all walks of life, in the blink of an eye, 2 months have passed since the last open letter calling off AI, in these two months, from the intervention of the national government to the open and dark battles between industries, to the mass layoffs of major companies, unemployment, anxiety, and various emotions fill our lives, and our society is being reshaped rapidly.
The iteration rate of the entire industry is so fast that it forms a certain degree of visual persistence, (just like when the electric fan speed is too fast, we think it is slow) so that we ignore its speed, which is a kind of speed that even if we react fast, we can't keep up with its own speed of change.
The blowout of AI generative applications also triggered the switch in FMRI technology iteration, and less than a year later, the team published a paper called Reconstructing Video from Brain Activity, accompanied by a landing website and ultra-high-accuracy video imaging after their experiment.
This time, the subject is no longer looking at a static picture, but a dynamic video, AI can through the test, generate the picture presented in the minds of others, from the test sample results, this time the accuracy rate is much higher than the picture test a few months ago. In just a few months, from flawed brain imaging to high-precision intracranial image mapping, for the first time, humans have truly achieved the long-standing mythical ability to glimpse the thoughts of others.
These once-distant technologies make people wonder if humans have the ability to reproduce concepts that were once considered legends in these 200 years, then at what level is our technology, and in which direction AI will continue to develop. It is definitely not a simple matter to define the entire level of human technology and AI in a macro way, and I had no clue when I made this video.
But I remembered another similar concept, the Kardashov index. This is a paper called "Information Transmission of Extraterrestrial Civilizations" produced by the Soviet astrophysics Kardashov at a conference in 1964, and it is also the first time that human beings examine the development of the level of science and technology in terms of planetary civilizations, simply put, if there are other advanced civilizations in the universe, then the existing civilizations on Earth are at what stage they are compared with. This theory is based on the amount of energy a civilization can use to measure the level of technology.
At that time, Kardashov set the 3rd level civilization of the universe (page 219 of the paper), while the human civilization only reached the level 0.728, which is still 1-200 years away from the first level civilization, and it will take thousands to tens of thousands of years to reach the stage of using all the energy of the entire solar system.
The Kardashov index estimates the definition of energy consumption of three types of civilization
This theory seemed so distant in the context of 1964 that it had nothing to do with humans, but less than 100 years after this theory was proposed, we seem to have bypassed the method of forcibly piling energy, found a shortcut, and relied on another technology that is smarter than humans to curve overtaking, and the technology I am talking about is AI.
When I was looking for similarities between the three-level civilization and AI development path in the Kardashev index, I found a paper from 6 years ago, 2017, which was an article that compared the Kardashev level and AI development path at that time, just as Kardashev proposed his theory 59 years ago, this article also examines the development of AI from a macro perspective.
If the development of AI had to go through several stages, what would it be, what conditions would be required to achieve those stages, and what stage we were at right now, these questions quickly squeezed into my head.
After a quick pass, I found some answers...
In the paper, AI is divided into 4 levels.
Level 0 is called weak artificial intelligence, also called narrow artificial intelligence, this stage of artificial intelligence includes speech recognition in our mobile phones, search engines, machine-written news and other more monotonous intelligence. This stage is also the closest to us, and the purpose of level 0 exists to measure the technology we have experienced and developed, based on the overall industry status quo in 2017.
The next stage is Level 1 AI, called Enhanced Narrow AI, which is characterized by the ability exhibited by AI beginning to approach humans. To reach this layer, it is necessary to break through several important technical bottlenecks, and there are 3 important indicators to judge whether AI has reached this stage.
For example, we begin to be unable to distinguish whether we are talking to a real person or an AI-powered chatbot; There is also L4 level of self-driving cars, automatic driving is divided into 6 levels, L0-L5, that is, to achieve level 1 artificial intelligence, you need to reach L4 level of automatic driving, commonly known as highly automated driving, this stage of the automatic driving function has been able to completely take over all driving operations, drivers do not need to focus on driving operations, and can even do other behaviors unrelated to driving, at present we can only achieve L4 level automatic driving in the established road section, it will take a while to popularize; In addition to the two indicators just mentioned, there is also an important indicator of first-level artificial intelligence, that is, bionic robots, robots with the ability to imitate biological movement and simple behavior, such as the cyber product "robot dog" in recent years and the simulation robot that Tesla is developing.
Looking back from today, it is clear that we have been close to level 1 artificial intelligence in 6 years, but it is too early to say that it has fully reached level 1, and many technologies are not yet mature. The prediction in the chart that it will start to move towards level 1 artificial intelligence after 18 years seems to correspond to the start of OpenAI, but the timeline after level 1 is broken, and the author does not give the time range required from level 1 to level 2, leaving only 3 question marks.
And in this timeline, level 1 is not directly to level 2, and before reaching level 2 artificial intelligence, there are two nodes that are not in the hierarchy. These two nodes are called narrow AI clusters and narrow AI super clusters, respectively. As their name suggests, the concept of clusters is to bring together narrow AI with a single function, to tie these seemingly ordinary AIs together, into stronger programming clusters. This stage may produce unexpected results, which may help humans break through the existing scientific bottlenecks, such as completing the leap of automatic driving level 4-5, but at the same time face great uncontrollable risks, such as the loss of control of AI. When you enter the hyper-cluster stage, you can see the door of level 2 artificial intelligence.
The reason why it is said that the gate can be seen, rather than directly entering level 2, is because there is an important problem inside, that is, in view of the current technology, it is still impossible to confirm whether it can truly enter level 2 artificial intelligence. Because one of the important indicators of this level is "awareness".
Level 2 AI has its own name, General Artificial Intelligence (AGI).
This stage of AI can fully mimic all human abilities, including planning, memory, analysis, and five senses, and may even extend more sensory systems.
But there's a bug in this layer, which is consciousness. Some people feel that only when AI has self-awareness can it truly be called AGI (General Artificial Intelligence), but the definition of consciousness is very vague. In nature, other creatures are also an important part of the composition of the planet, so do they really have self-awareness? In addition to their basic instinctive reactions to the weather and seasons, do they recognize the passage of time? Do you recognize yourself as an organism capable of acting independently?
People question these questions day in and day out, but the answer ultimately does not depend on us, but on the species itself. One thing is certain, however, that whether they are self-aware or not, they are core players in their respective ecosystems. So no matter what kind of consciousness AI eventually produces, even if the consciousness it produces cannot be understood, it will eventually play its role in the entire human society and even the entire natural world. It is also on this premise that scholars finally believe that no matter what kind of machine consciousness appears, the general artificial intelligence AGI will eventually be realized.
2017, 6 years ago. Let's go back to today, after the emergence of GPT 4, Microsoft published this paper called "The Spark of General Artificial Intelligence" overnight, which I also mentioned in the previous video, because some abnormalities in the neural network layer of large language models have caused many AI scholars to begin to worry that general artificial intelligence may be inadvertently produced. In just 6 years, 2,190 days, and more than 50,000 hours, mankind saw the looming AGI gate. The scary thing is not only the speed of scientific and technological progress, but also the uncertainty, because no one can judge when and how AGI will appear in the true sense.
Although academics can't tell when it will come, it is clear what will happen after it is born. Because level 2 is not yet the final destination of artificial intelligence, there is also the last level of artificial intelligence, level 3 artificial intelligence, Artificial super intelligent (ASI) super intelligence.
At this stage, the AI ability is far beyond humans, and it can constantly improve itself, self-perceive and self-reprogram, and its IQ will be immeasurably high, and it can create things far beyond the scope of human understanding.
Unlike the Kardashov index, it may take thousands or even tens of thousands of years to cross between each level, and it may only take an instant for AI to go from level 2 to level 3.
In the chart, the AGI to ASI is marked to indicate that this stage is likely to accelerate implementation. The leap from level 2 to level 3 can be faster than any stage. Given that Google OpenAI, Github and other giants are training AI to program themselves, and with the launch of "AutoML" by Google, machine programming has been performing better and better in recent years. After the birth of AGI, it is destined to identify its own weaknesses, quickly cycle and iterate, and finally form a genius artificial intelligence far superior to humans in intelligence and ability, overturning our existing cognition and entering the next stage of civilization.
In early June, I went to various places to participate in some AI activities and lectures to strengthen myself, and on the way to board the plane at Shanghai Airport, I found that 1/2 of the pages of the newspaper were dominated by AI-related news, and I can't remember the last time a newspaper was dominated by a certain type of news. In 2023, I have long been accustomed to a fast-paced life, but it is clear that not only I have adapted to the fast, looked at my phone, listened to Lao Huang's speech, I found that Moore's Law has also accelerated this month to break the previous law, and the speed of hardware evolution is also unprecedentedly fast. （https://www.youtube.com/watch?v=L6rJA0z2Kag&list=WL&index=21&t=329s）0:15
The speed of human scientific and technological development is destined to gradually accelerate, intelligence will become more and more human-like, and even give me the illusion that every day after that will be like humans witnessing the evolution of silicon-based life.