Since the explosion of big models, global interest in so-called "big language models" has never been higher. It is said that a hundred model war, a thousand model war can be. Major companies, such as Google's BERT, OpenAI's GPT, etc., are doing their best to make their big models more usable, better, and more powerful. These companies not only invested a lot of manpower and material resources in research and development, but even consumed most of the world's GPU resources in computing power, almost buying out the chips of the old Huang family. However, no matter how hard you make a miracle, you just put the world's data set there, and the big model is not much smarter, nor is it strong. It feels like a lot of big models are off track now.
Yesterday I saw a paper called "Language Models are Few-Shot Learners," which says that to improve large models, you have to give them variety. What is diversity? To put it bluntly, it is to provide all the world's languages, all the fields of text to this big model, AI can learn to understand the world through big data. This actually involves transfer learning and meta learning, which allow models to flexibly switch between different tasks and domains. But even if you get it, what about it? These big models can speak well, but they have limited capabilities. They can't see, touch, smell, or hear. Just a brain with a mouth. The brain in the tank still lacks clear eyes, ears ....
The big language model does not have five senses and five senses, which is the key to restricting the development of the large model. It's just that there is no future in developing a single facial features, even if someone now says multimodal, just say that let these large language models generate videos and generate pictures, and this generation ability is limited even if it is well developed.
Some time ago, someone put an eye on the large language model. But this is an ordinary B station blogger do, but also involves computer vision (Computer Vision) technology, not complicated. Large language models that can see objects with their eyes and answer questions should be smart. The next challenge is whether it can be further equipped with "ears" and even "smell" to make large models more like humans can perceive and understand the world.
This blogger is also a very ordinary blogger is not the world's top artificial intelligence bull, his technology is not advanced, not so profound, he can have such a thought, is it our big bulls, our so-called far ahead experts are doing?
Are the big cows of those big factories sleeping? Probably no either. It's just that these technologies are not taken seriously, and they can't use this technology, they can't use it, and they can't use it. Moreover, all large companies are thinking about how to make money, a little function, quickly brag a big bull, to let the big model quickly create economic benefits for themselves, as for technological innovation? As for technological transformation, standing aside, these are actually short-sighted.
There are limits to how big models can do wonders. Simulating the five senses and six senses of human beings is the breakthrough to become smart. When these functions are complete, strong artificial intelligence will soon come. Some people should not be afraid. When the steam engine appeared, some people said that it destroyed feng shui. When the loom appeared, someone took tools to smash it. There are many people crying and shouting, the suppression of various restrictions on artificial intelligence, including the recent domestic AI painting can not be used, various painting manufacturers AiGC painting websites, many have reached the edge of profit and charging, directly destroyed, the reason for the policy we may know a little, but how to meet the huge market demand? Do those big manufacturers provide the necessary technology and the necessary capabilities to meet the huge market of China?
For AI painting, generative artificial intelligence, AI videos and the like, can it only be banned?
When you are suppressing artificial intelligence technology, maybe other countries have applied these technologies to the military, so what can be done then? People's robots are getting smarter and smarter, can it be said that when other people's strong artificial intelligence appears, we will send Huaqiangbei? Ma Jiajun? And then far ahead? The problem is that sometimes, Hua Qiangbei and Ma Jiajun are helpless. What used to be useful may not be useful in the future.
Today's large language models have begun to experiment with multimodal combinations, which can talk, hear sounds in two days, and have vision in two days. This actually involves research in areas such as Joint Models and Cross-Modal Learning. Our big models are still selling services to various manufacturers, what services? It is to take your company's documents, make all your company's data into a local database, and then chat and play. Then when the customer asks questions, all kinds of tai chi, all kinds of recommended goods purchase, all kinds of enterprises are playing well. All kinds of plug-in platforms are actually a routine, that is, chat.
There is also a famous platform to do the platform to start, now directly open source their model, meaning let everyone open source, continue to do the platform dream, but the world has undergone a lot of changes, do the platform developed, can not always do the platform developed, my elementary school teacher once told me, can not go all the way to the dark, doesn't the boss of this platform know? Success cannot be copied, once started as a platform, just 10,000 years to make a platform? The resources are wasted, and in fact, there is not much splash.
To reiterate, there are limits to how powerful a big model can do wonders. The real breakthrough lies in imitating the five senses and six senses of humans, so that the model can perceive the environment and actually act. Only in this way can strong artificial intelligence truly move towards our lives and open a new scientific and technological revolution!