Human beings are only a transitional stage in the evolution of intelligence.
-
Jeffrey Hinton
”
After the college entrance examination scores, both adults and children are busy with a pot of porridge, because no one dares to say that they are nine out of ten when it comes to filling in the volunteers.
Especially in today's social environment where AI is popular, what major can be chosen to avoid being "eliminated" by artificial intelligence? Before making up your mind, you need to understand not only your interests and abilities, but also how far AI has come. In the series "It's Not Like a Human, It's Called a Robot" in the "Juvenile China Science and Technology, Future Science +" series, Chen Runsheng, an academician of the Chinese Academy of Sciences, shared the future of AI large models with us from the perspective of biomedicine.
Chen Runsheng / Huang Tiejun / Shan Shiguang et al
The impact of large models on biomedicineYou must know a very hot topic recently, that is, artificial intelligence has entered the era of large models. In other words, artificial intelligence has evolved from being able to do one thing to being able to handle multiple things at the same time, and we call this new development of artificial intelligence a "big model" of artificial intelligence. What is a large model? A large model is a combination of broad, disparate data for analysis. Let's briefly discuss the progress of large models and their impact on the development of biomedicine as a whole. In fact, whether it is the current artificial intelligence or large models, its basic principle is to imitate the human brain. The human brain has many neurons, and when many neurons are connected, they form neural circuits that can do a lot of things. Now, we're just making things bigger, making things more complicated, and the fundamentals haven't changed.
In what areas has the current large model progressed? Primarily language, to be exact, natural language. For example, all kinds of papers, books, and medical records can now be read by computers in this way, greatly upgrading the capabilities of large models. After the advent of large language models, it immediately moved to the field of biomedicine, because biomedicine is relevant to everyone. For example, when OpenAI launched ChatGPT, Microsoft launched BioGPT, which directly applied large models to biomedicine; Google has also made a large model for biomedicine, Med-PaLM - we can see from this that the entire large model development industry attaches great importance to biomedicine. So what about the level of these large models? For example, the passing score on the licensing test for occupational doctors in the United States is 60 points, and BioGPT and Med-PaLM both score above 80 points. I think that if a "white ding" like me wants to take the professional doctor's license, he will definitely be able to pass the test with this software. So, now artificial intelligence has been learned very well in some basic areas of biomedicine. Continental has also launched a large model for biomedicine. In my opinion, Baidu should be the one that has attracted the most attention and spent more energy and time on large models. Baidu has put a lot of effort into building China's large models, and its subsidiary, Baitu Biotech, was established specifically to apply large models to biomedicine, and has done a lot of groundwork. The New England Journal of Medicine (NEJM), the world's top medical academic journal, will launch a special issue, NEJM-AI, in 2024. According to one of its reviews, AI can actually be applied to many aspects of medicine. Therefore, the application of artificial intelligence, including large models, in biomedical medicine is a long-term development, not a sudden rise.
Will the large model be able to reach the level of strong artificial intelligence in the future? The reason why the large model has attracted wide attention around the world is mainly because some new phenomena have been discovered after the emergence of the large model. If the model size is large enough and the parameters learned are large enough, two things can happen. One phenomenon is called emergence, and the other is called epiphany. This is not a phenomenon that occurs with regular computer programs. What is emergence? I teach you so much that after learning a certain level you will have new ideas that I did not have in the process of teaching you. This phenomenon has caused people in the industry to worry deeply and think: will what emerge surpass the knowledge of human beings? Epiphany is such a phenomenon: the big model, like the human brain, did not learn the first time, did not learn the second time, did not learn the third time, but the fourth time it opened its mind and learned it completely. We know that this doesn't happen with a computer, it can only run one program, you input it and you always get the output. The emergence and epiphany of large models show that they are different from the artificial intelligence we used to play Go and predict the structure of biological macromolecules in the past, which makes the current large model have more room for deeper and broader discussion in addition to being a technology.
After OpenAI just released GPT-4, on May 14, 2023, Claude, known as ChatGPT's "strongest competitor", released an update. Essentially, Claude is the same as ChatGPT, but Claude can process nearly 100,000 words a minute, which means that its text processing power surpasses ChatGPT in a short period of time. These two models are developing too fast, you are catching up. In addition, there are already attempts to use large models to drive third-party systems. In other words, if a large model can drive a surgical robot, it can perform surgery in place of a surgeon. Such developments raise concerns about how far AI will go, as well as about how quickly it will develop and the downstream equipment it can mobilize. Therefore, the large model is particularly noticeable. At present, the focus of debate is whether large models can reach the level of strong artificial intelligence in the future. Strong AI is a technical term in the field of artificial intelligence. If artificial intelligence reaches the level of the human brain, it can think, think, and make decisions, it is a strong artificial intelligence. Of course, the current artificial intelligence has not yet reached this level, and whether it can reach it has become a point of great concern to people. For this reason, many people have deep concerns. Here, I will only mention one representative figure - Jeffrey Hinton, who is one of the earliest founders of artificial intelligence. He has expressed concern about artificial intelligence, especially this sentence: "Humans are only a transitional stage in the evolution of intelligence." The implication is that he believes that artificial intelligence may surpass human intelligence in the future. It's a question worth pondering, and it's because of this concern that he resigned from Google. In addition, on March 22, 2023, the Future of Life Institute released an open letter to the whole society on the "suspension of large-scale artificial intelligence research", which was signed by experts such as Musk and Turing Award winner Joshua Bengio. They want to pause the development of models for artificial neural networks that are more capable than GPT-4 for six months. They believe that humans are not ready, and if AI develops too quickly, maybe humans will not be able to grasp it well.
Open letter on "Moratorium on Large-Scale AI Research".
In addition, the European Union proposed the Artificial Intelligence Act, which means that the entire European Union has developed a legal regulatory scheme for artificial intelligence. On May 13, 2023, the U.S. government established an artificial intelligence working group, one of the two leaders of this working group is the famous Chinese scientist and mathematician Tao Zhexuan, and the Chinese female scientist Li Feifei is also a member of the advisory group. The above examples show that the current development of large models is not only a scientific problem, but also raises many problems, including the philosophy of science, which arouses people's deeper thinking and concern about large models. I would like to express some of my own opinions, in the future, of course, artificial intelligence will continue to develop, and it will be difficult to stop any force, because it is an advanced technology, and the development of cutting-edge technology is difficult to stop. But is the improvement of AI capabilities continuous or barriers? If it is continuous, it will continue to develop; If there are barriers, that is, it needs to overcome certain barriers in order to reach a new stage of development, we may delay its development on certain barriers. At present, there is no evidence to prove that there are barriers to the development of artificial intelligence, so it is still developing at a high speed. Therefore, if there are barriers to improving the capabilities of artificial intelligence, I believe that the complexity of the current structure of artificial intelligence neural networks is far from comparable to that of the human brain. In other words, AI will continue to evolve, and new emerging phenomena that people are worried about may still emerge, but it will take a very large time scale for AI to surpass human intelligence.
The picture above is a human neural network, and you can compare it to an artificial neural network. Physicists always ask the question: What if the laws of physics changed from three-dimensional to two-dimensional space? I believe that many physical laws are not valid. We can compare that the current AI model is still not as complex as the real human brain. I believe that artificial intelligence will continue to develop, and will definitely provide a great boost to the development of productivity, social production and life, and the development of medicine, but we must also pay attention to the various impacts it may bring. But at least in the short term, it is still a long way from reaching the level of strong artificial intelligence.
Article source: Unread