As long as the capabilities of this kind of AI intelligence that interacts with people at high intensity are improved a little, it will be difficult for people to continue to go back to the worse performing version.

Just like the resolution, after everyone is used to watching HD, they can't accept the blurry 480p.

Once you lose the arms race of the big model, the vast majority of users will be siphoned away by the giants at the top.

And if you want to achieve such success as ChatGPT, in fact, the proportion of the technical level is not particularly large.

GPT and GPT2 are both open source technologies, but at that time, the BERT-like technology of dividing language tasks into several branches was more popular.

Therefore, GPT series of technologies that focus on language generation have been neglected.

GPT3 and GPT3.5 continue to iterate on new technologies in obscurity, which is very difficult in the Internet atmosphere of China.

Meng Fanqi, who has traveled to 2-3 major factories in Yanjing and Shanghai, knows the urine nature of these Internet giants very well.

As long as we see a little bit of technological breakthroughs and achievements, all of them are asking the question, "How can I turn it into money?" ”

Excessive meaningless involution and competition make it difficult for 996 manufacturers to accept a big breakthrough that needs to be iterated for several years in order to achieve a blockbuster.

And the technology in this direction is precisely what colleges and universities basically cannot afford.

Meng Fanqi has never been a person who admires foreign flattery and is the same as foreigners. But looking at the history of AI, he has to admit that AlphaGO of Go, AlphaFold of protein, and even the breakthrough large model, ChatGPT.

But none of them are produced by Huaguo.

Could it be that the Chinese are not smart enough? Of course not.

If you flip through the top AI papers, you can easily find many names that are obvious at a glance.

But the more groundbreaking and groundbreaking work becomes, these Chinese institutions are always foreign universities and research institutes.

This is the reason why Meng Fanqi had to choose to go to Silicon Valley for development even with the advantage of rebirth.

If you want to lay a solid research foundation in a large domestic factory, the effort and pressure you need to bear are too great.

The design of the model, the concept of the proposal, can only have such a significant effect in the early days.

Further afield, high-quality data plays a more critical role.

"So you're basically completely denying the value of small models?" Robin Li disagrees with this view. "In my opinion, small models and mobile models and intelligence have a very broad market and prospects."

"It depends on the specific application scenario, what is the purpose of the terminal's intelligence, and whether it can be stably networked. If it is only the simplest image classification and detection, and in the case of no stable networking, of course, it still has some value. ”

"For example, autonomous driving scenarios." Meng added, "How can I deny the value of small models, as a programmer who studies algorithms, I am quite concerned about the accuracy and runtime of specific models and technologies. ”

Meng Fanqi took out a few experimental record forms of detection algorithms from his bag, "Mr. Li, you can take a look at the results of my experiments in the past few months. I made my judgment with a good understanding of the acceleration optimization of small models, and when I talk to you about cooperation, I am not just relying on the champion of a competition. ”

NOT LONG AFTER THE IMAGENET COMPETITION ENDED, THE IDEA OF SUCH AN EXPLOSION OF GENERATIVE ADVERSARIAL NETWORKS, AND THE PAPER HAD JUST BEEN RELEASED, THIS KID STILL HAD THE TIME AND ENERGY TO DO OTHER RESEARCH? What is the detection algorithm for the image?

I have to admit that Li Yanhong was shocked, but soon he found out that he was still shocked.

What kind of horrible result is this? While the detection accuracy has made a qualitative leap, the inference speed of the model is more than 100 times that of the current top algorithms.

You're not kidding me, are you? There is definitely something wrong with this result.

Robin Li is a person who is proficient in computer technology, and because of this, he can't think of a good way, or even a good possibility, to explain the principle or feasibility of this exaggerated result for a while.

Speech and image detection are the technologies that Baidu is now focusing on, and he even set up the first research institute in China to specialize in this area.

Many of the talents in it are like today, he personally went to talk to communicate, or sent very close subordinates to recruit.

Li Yanhong didn't want to deny Meng Fanqi's talent, but this gap was too exaggerated, so he couldn't explain it or accept it for a while.

"Do you want me to admit that I, a large number of researchers, with dozens of people up and down, are not as good as one percent of the undergraduate students in front of me?" Li Yanhong, who is at the peak of his life, can't accept it psychologically.

Even if the inference speed of detection is worse by more than 150 times, the other party has not even sacrificed performance in exchange for detection speed.

What the hell is going on here?

Robin Li has read the table three times, and it is full of easy-to-understand numbers and a few visualizations to make the content easier to understand.

It is clear which is better and which is better. I asked junior high school students to see it, and they could understand it in a minute.

But Li Yanhong has been watching it repeatedly for three minutes, he didn't say a word, flipped around, and didn't know what he was confirming.

It seems that I want to find something unreasonable, but when I look for it, I realize that something is wrong.

None of the numbers in this experiment record are correct.

All of them are levels that I have never thought of at this stage.

"It's not. Forehead.. You this.. Uh.." Li Yanhong opened his mouth to ask something, but he was speechless for a while, which is a very rare situation in his life.

"You promise, uh. You're sure that the data and the results are fine, right? ”

This is actually a nonsense, Meng Fanqi smiled, "If there is a problem, how dare I show it to Mr. Li?" ”

Li Yanhong put the thin pieces of paper back on the table, leaned back in the chair and closed his eyes for a long time, and finally asked, "How do you want to cooperate?" ”

Meng Fanqi smiled secretly in his heart, it seems that Li Yanhong was also shocked by the technological breakthrough two years ahead. Maybe there is nothing new in other industries in two years, but in this decade of AI, especially in the first five years of 13-18 years.

There are many techniques to ascend in place.

"Mr. Li, for some personal and other considerations, my focus in the next few years may not be in China. But if we are not in China and do not join Baidu, it does not mean that we cannot cooperate technically. ”

Meng Fanqi knew that with Li Yanhong's rhetorical question, the initiative of communication had gradually come to his side.

"Our country has always had its own management and restrictions on the Internet and some technologies. Many foreign companies do not have the technology and services that cannot be provided directly and smoothly in China. Mr. Li should have felt this deeply. ”

"The future development of AI technology will rely heavily on massive data and enterprise control, both of which China will not completely open up to foreign companies no matter what."

Speaking of which, Meng Fanqi's meaning is quite clear.