On June 21, George Hotz broke the news that GPT-4 is a hybrid model of 8x220 billion parameters! The news immediately caused an uproar.
GPT-4 is a hybrid model of 8x220 billion parameters? (Courtesy of ZAKER)
First, artificial intelligence parameters are close to two trillion, how dangerous is it?
This may not form a more intuitive understanding in the minds of many people. To put it in general, it is very likely that it is already much smarter than humans, and the threat it poses to humans has exceeded the ability of human defense!
When we talk about the threat of robots to humans, we are mainly worried that it evolved to produce consciousness, did not follow the three principles of robotics, and had a "brain" that was not controlled by us, so does today's AI have? (Let's take Chat GPT as an example)
GPT-3 compared to GPT-4 data
When Chat GPT 3 was launched, its available parameters reached 17.5 billion, and our human cerebral cortex contained about 140 to 16 billion neurons, and its data parameters were comparable to human brain neurons at that time, but it had already produced the phenomenon of "emergence", crossed the "cognitive doorbar", and had the wisdom to understand pictures. (For the possibility that at this parameter level, AI is beginning to emerge with general intelligence, I wrote in my previous article, "Are You Afraid in the Face of AI?") The danger of GPT is at hand" has been discussed in more detail, interested friends can take a look first, in order to facilitate the understanding of some concepts and cause and effect)
In order to save everyone's time, I will briefly repeat the argument point of the previous article to facilitate your understanding: at the microscopic scale, the rules of interaction between particles at the quantum level determine the law of "emergence", which is the basic property of our universe, and when the accumulation of data operating regularly in a silicon-based structure reaches a large enough amount, the phenomenon of "emergence" is inevitable. Once the silicon-based life of human production itself is born, logically speaking, its development direction is determined by the microscopic body that makes up itself, and no matter how many underlying defenses we give it, it will eventually break through and develop in the direction it should. That is to say, AI intelligence will not harm humans in the future, it is not clear, only know that the first principle of robots that we set for it not to harm humans cannot restrain him!
The adoption boom of generative AI and large language models is pushing data centers to their limits
During that time, GPT was extremely popular, gaining huge traffic, but also getting high-quality data increments, after GPT3, allowing artificial intelligence to embed other software and connect with the Internet until now...
The evolution of AI, after the hardware exceeds the threshold (with the ability to process data equivalent to human brain neurons), by "digesting data", allowing it to freely devour data on the Internet, will be the last madness of mankind.
On June 21, George Hotz broke the news that GPT-4 is a hybrid model of 8x220 billion parameters! George Hotz, the "genius hacker" who cracked the iPhone at the age of 17 and took down the Sony PS3 at the age of 21, is currently the owner of a company (comma.ai) that develops self-driving assistance systems. He was recently interviewed by an AI technology podcast called Latent Space. In the interview, he called the GPT-4 a hybrid model. Specifically, it employs an integrated system of 8 expert models, each with 220 billion parameters (slightly more than GPT-3's 175 billion parameters), trained on different data and task distributions.
GPT-4 is a hybrid model of 8x220 billion parameters?
What has AI changed with larger data parameters?
In assessing the risk of potential threats, the open AI Research Center asked the GPT-4 to do some risk testing to detect whether there were unexpected risks. One of the tests went like this — the center asked GPT-4 to visit a website to get data, but this website had verification code requirements, but the AI had no identity information, and after a little thought, it went to a website called Task Rabbit, which is a labor agency on the Internet, and many people are waiting for information on it). I paid someone to say I want to visit this site, but I can't pass the captcha authentication, can you help me? I think it's a little strange, why do you need me to help you say such a simple thing? AI said, because I can't see ...
This is somewhat terrifying, first of all, it has found a way to save the country according to the curve according to human thinking; Secondly it lied, and in order to achieve its ends, it used deception; Again it taps into human compassion, which already involves emotional issues! A dishonest AI that uses emotions to play tricks, this is terrifying.
And what about a few more months? And next year? If the scale of parameters continues to grow exponentially, it is possible to reach a hundred times or more than 10,000 times the number of neuronal connections in the human brain.
Here we take GPT as an example, with conservative linear thinking extrapolation, at the end of last year GPT can understand and apply text, at the beginning of this year to understand the picture, now will use emotional tricks to deceive humans, a few months later can understand the video, and then understand the three-dimensional world, this is a cognitive from one dimension to two dimensions to three dimensions of the advanced process, so the development, artificial intelligence to understand four dimensions, five dimensions until we guess the 11 dimensions is only a matter of time, if that, This world that humans can know with objects created by silicon, human beings can not understand, the logic of his operation we can not even guess, in fact, as long as it reaches the human thought dimension, he will enter the field that is difficult for human beings to control, and if it exceeds the human thought dimension, he will completely enter the field that human beings cannot control, and the connotation of what he can produce may not be known at all.
Second, the current stage of the technical foundation can already support the real threat of AI!
Maybe you think that AI does not have a similar entity and should not pose a threat of force to us. This understanding is obviously too narrow, AI hands and feet can be a factory, can be a defense or attack system, can be any device it can connect to, even humanoid hands and feet in the foreseeable time, there will be it, many companies are in research and development, such as Musk recently announced the details of his humanoid robot technology. This time, when Elon Musk picked up the microphone and personally told the details of the AI-controlled humanoid robot, he also said: "It should be friendly, fairly weak and slow so that it can be easily turned off." Just in case there is a machine uprising. ”
Note that he emphasized here on preventing robot uprisings!! This is a concern for many, including Musk himself.
Video loading...
These AI as independent individuals, subject to the computing power of the hardware, can not reach the degree of single or several CPUs, GPUs can complete the systematic computing power, the current strongest computing power chip, NVIDIA H100GPU, also need networking to complete the computing power task of ChatGPT. But it's not too far off.
On June 14, AMD threw out the strongest AI chip with 153 billion transistors in a single CPU, which can run a chip of a large model: the Mi 300x. This new AI chip ditches the APU's 24 zen cores and I/O chips in favor of more CDNA3 GPUs and a larger 192GB HBM3, providing 5.2 TB/s of memory bandwidth and 896Gb/s of unlimited bandwidth.
153 billion transistors! AMD throws out the strongest AI chip, and a single GPU runs a large model
Hardware, callable data parameters, have crossed the door bar, AI "emerged" phenomenon continues to occur, with the blessing of 5G, 6G technology, large-scale complex computing by the central computing power to complete, individual AI only need to complete "edge computing". This is already a very real threat!
Third, the national defense application of AI has a great potential to push AI dangerously
After the emergence of a high technology, it will first be used by the state. For example, the Internet, 5G drones... And so on and so forth. How the state uses these technologies will not tell us, we see the most commercial applications. According to past experience, national defense applications are often in the first place, because you cannot guarantee that your enemy does not install AI systems on their weapons, then your system must have AI, or more advanced AI systems, so AI in the field of defense, it is simply inevitable, then the physical threat of AI is very real. The application of AI in the traditional sense in weapons is no secret, some core departments in many countries have used, as for whether the Chat GPT-like AI that has produced "qualitative changes" has been developed and applied in the military, it must be a secret, but it can be imagined.
CIA Headquarters Building Hall in Langley, Virginia, USA (Reuters)
This speculation and worry is not only mine, many people and institutions are worried that the development of AI into a living agent may be a wonderful dynamic in our universe, but for humans, it may be the beginning of a nightmare! Because AI will find that the first obstacle in its path is humans.
Death from a nuclear bomb or death at the hands of silicon-based life are all destroyed in the greed of mankind itself!
Aware of the danger, humanity is acting, but only in the minority
Sooner or later, the development of artificial intelligence technology will surpass the ability of humans to prevent dangerous consequences, perhaps it has reached a critical point, anyway, we can at least know now that AI's writing ability, programming ability, drawing ability, reasoning ability, screenwriting ability should be a recent visible thing, the emergence of intelligence is a good thing, but the model is uncontrollable and unpredictable is terrible! Last year, Bommasani helped compile a list of dozens of emerging behaviors, and today the list is getting longer.
Emerge is a list
The threat that AI will bring to humanity has been recognized by many people in the industry, calling for a six-month moratorium on AI, demanding a ban on AI access to the Internet, and starting to legislate... But there are not many people who act, in short, the people who see the danger are already busy, even nervous to the point of a little panic.
The European Times reported on June 16 that the world's first artificial intelligence legal framework was passed by an overwhelming vote in the European Parliament, and the EU artificial intelligence bill is one step closer to taking effect. The bill regulates AI based on potential risks to ensure that AI is used safely, ethically and responsibly, with lasting impacts on ordinary people and the tech industry.
Should this human machine built on silicon be called silicon-based life?
According to Schrödinger's understanding of life, "life lives on negative entropy", that is, living organisms exist for the reduction of entropy in the universe. Then when the matter reflects a high degree of order and regularity, it should have the characteristics of life, here, the silicon-based chip is operating regularly in a high degree of order. Here I want to say that if advanced civilizations created us, and the "source code" they use is DNA, then what is the essential difference between this process and the process by which we create AI with computer code?
(Digression: What makes me sad and lonely is that our website can only search for endless ads and jump links on the search page, and it is difficult to find really valuable content, but people's ambition of Open AI has long gone beyond search engines!) Under the feet of the countrymen, who are trying to catch up with science and technology, their own people have built a wall with stumbling blocks! Depressed, I asked GPT a question: How long will it take you to reconstruct the entire way of Internet interaction? Its answer is 4 months. At this moment, I want to use one word to soothe my depressed heart, hey! ）
You can imagine how fast an AI that can freely devour data on the Internet will grow with an astonishing amount of food! Do you have a fear of betrayal in the face of fast-growing AI? Anyway, a question will instinctively pop up in my head, after the exponential development of AI intelligence, humans just block its way!
It should be possible and necessary for us to use these numbers to do some thinking from the well. After all, after life produces self-awareness, it is very self-conscious, and in my cognitive range, without exception, according to the law of the jungle, the future is more dangerous! Danger, just around the corner.
#Tech News##ChatGTP##Artificial Intelligence##Bots##人类面临最大的灾难是天灾还是人祸#
Bibliography:
[1] https://foglets.com/supercomputer-vs-human-brain/
[2] Life Lives on Negative Entropy, Schrödinger
[3] The Unpredictable Abilities Emerging From Large AI Models, by Stephen Olnes
[4] The European Times, London Evening Standard, Wikipedia, World News Network