This kind of AI intelligence that interacts intensively with people, as long as its capabilities are improved a little, it will be difficult for people to continue to accept a version with worse performance.
Just like the resolution, after everyone is used to watching high-definition, they cannot accept the blurry 480p.
Once the arms race of large models is lost, most users will be siphoned away by the giants at the top.
To achieve such success as ChatGPT, the technical aspect is actually not particularly large.
GPT and GPT2 are both technologies whose paper codes are open source, but at that time, BERT-type technologies that specifically divided language tasks into several branches were more popular.
Therefore, technologies such as the GPT series that focus on language generation have been ignored.
GPT3 and GPT3.5 continue to iterate new technologies in obscurity, which is very difficult in China’s Internet atmosphere.
Meng Fanqi, who has worked for 2-3 large factories in Yanjing Shanghai, is well aware of the urinal nature of these large Internet companies.
As soon as they see some technological breakthroughs and achievements, everyone is asking one question, "How can we turn it into money?"
Excessive and meaningless involution and competition have made it difficult for major 996 manufacturers to accept a big breakthrough that requires several years of iteration in order to become a blockbuster.
And the technology in this direction is precisely what universities basically cannot afford.
Meng Fanqi has never been a person who worships foreigners, is harsh on foreigners, and rejects others. But looking at the history of AI development, he has to admit that Go’s AlphaGO, protein’s AlphaFold, and even the breakthrough large model, ChatGPT.
But none of them are produced in China.
Are the Chinese not smart enough? of course not.
If you look through the top AI papers, you can easily find many names that are obviously Chinese at a glance.
But when it comes to pioneering and ground-breaking work, these Chinese institutions are always foreign universities and research institutes.
This is the reason why Meng Fanqi had to choose to develop in Silicon Valley even if he took advantage of his rebirth.
If you want to lay a solid research foundation in a large domestic factory, the effort and pressure you have to bear are too great.
The design of the model and the proposing of the concept can only have such a significant effect in the early stage.
Later on, high-quality data will actually play a more critical role.
"So you basically completely deny the value of small models?" Li Yanhong disagreed 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 the terminal's intelligence is used for, and whether it can stably connect to the Internet. If it is just the simplest image classification and detection, and it cannot stably connect to the Internet, of course it still has a certain value."
"For example, in the autonomous driving scenario." Meng Fanqi added, "How could I deny the value of small models? As a programmer who studies algorithms, I pay close attention to the accuracy and running time of specific models and technologies."
Meng Fanqi took out several experimental record forms of detection algorithms from his bag, "Mr. Li, you can take a look at my experimental results over the past few months. I made this judgment based on a good understanding of the accelerated optimization of small models. And when talking about cooperation with you, I don’t just rely on a competition champion.”
The IMAGENET competition has just ended. Generative adversarial networks are such an explosive idea, and the paper has just been released. This guy actually still has time and energy to do other research? Image detection algorithm?
I had to admit that Li Yanhong was shocked, but he soon discovered that he was shocked too early.
What kind of horrific result is this? While there is a qualitative leap in detection accuracy, the model's running inference speed is more than a hundred times that of today's top algorithms? ?
Are you kidding me? There is definitely something wrong with this result.
Robin Li is a person who is proficient in computer technology. Because of this, he could not think of a good way or even a good possibility to explain the principle or feasibility of this exaggerated result.
Voice and image detection are the technologies that Baidu is currently working on. He even established China's first research institute specializing in this area.
Many of the talents there are like today. He went to talk and communicate with them personally, or sent out very trusted subordinates to recruit them.
Li Yanhong didn't want to deny Meng Fanqi's talent, but the gap was so exaggerated that he couldn't explain or accept it for a while.
"Do you want me to admit that I, a researcher in a big family, with dozens of people from top to bottom, are not as good as one percent of the undergraduate student in front of me?" Robin Li, who was at the peak of his life, really couldn't accept it psychologically.
Even though the detection inference speed is more than 150 times worse, the other party has not sacrificed performance in exchange for detection speed.
What the hell is going on?
Robin Li has read this table three times. The table is full of concise and easy-to-understand figures, plus several visual charts that make the content easier to understand.
Which one is better and which one is worse is clear at a glance. Even if I let junior high school students watch it, they can understand it in just one minute.
But Li Yanhong had been looking at it over and over for three minutes. He turned over and over without saying a word, not knowing what he was confirming.
It seemed that he wanted to find some unreasonable places, but as he searched, he found something was wrong.
Of all the numbers in this experiment record, not one of them is right.
All of them are at a level that I have never thought of at this stage.
"No... uh... you... uh..." Li Yanhong opened his mouth to ask something, but was speechless for a moment. This was a very rare situation in his life.
"You promise, um... you are sure that there is no problem with these data and results, right?"
This is actually nonsense. Meng Fanqi smiled and said, "If there is a problem, how dare I show it to you, Mr. Li?"
Li Yanhong put a few thin pieces of paper back on the table, leaned back in his chair and closed his eyes for a while, and finally asked, "How do you want to cooperate?"
Meng Fanqi smiled secretly in his heart. It seemed that Li Yanhong was also shocked by the technological breakthrough two years ahead. In other industries, there may not be anything new in two years, but in the past ten years of AI, especially the first five years from 2013 to 2018.
There are many technologies that allow people to soar in place.
"Mr. Li, due to some personal and other considerations, my focus in the next few years may not be in China. But not being in China and not joining Baidu does not mean that we cannot cooperate technically."
Meng Fanqi knew that with Robin Li's rhetorical question, the initiative in communication had gradually come to his side.
"Our country has always had its own management and restrictions on the Internet and some technologies. The technologies and services of many foreign companies cannot be directly and smoothly provided domestically. Mr. Li should be deeply aware of this."
"The future development of AI technology will rely heavily on massive amounts of data and corporate control, both of which China will never completely open to foreign companies no matter what."
At this point, Meng Fanqi's meaning was already quite clear.
