The AI model represented by ChatGPT is the C position of technology in 2023.
On March 27, 2023, the "Catching Horse Series" triggered by ChatGPT will be updated with a new drama.
Recap:
Wang Huiwen, the former co-founder and senior vice president of Meituan, announced his entry into the AI field, saying that he will build China's OpenAI.
News headline 1: "Wang Huiwen said: I will invest 300 million first"
Of course, he is now a non-executive director of Meituan, a shareholder and executive director of Beijing Light Year Beyond Technology Co., Ltd.
Later, Wang Xing said in the circle of friends: "...Since he is determined to embrace this big wave, I must support it."
News headline 2: "Wang Xing will personally participate in the A-round investment of Wang Huiwen's "Light Years Away" and serve as a director"
According to Wang Huiwen, 75% of the company's shares are used to invite top R&D personnel.
The phrase "invite top R&D personnel" is bright, which is worth paying attention to.
Back and forth, Wang Huiwen said in the post: "At present, three co-founders have been confirmed, one with an Infra background, one with an algorithm background, and me."
The word "joint creation of an Infra (infrastructure) background" is on again, and the joint creation of this background should be within my sight.
Finally, on March 27, 2023, "Dear Data" exclusively learned that the former Meituan co-founder Wang Huiwen's team "is" acquiring the domestic AI framework OneFlow.
There will be an important meeting to discuss "acquisition matters", it is not clear when the meeting will be held.
At this time, there is no need to contact anyone who participated in the meeting. According to the requirements, no one can talk about the terms of the transaction.
Since the press time is in the progress of this incident, it is impossible to know the acquisition amount of OneFlow.
News headline 3: "Exclusive 丨 Former Meituan co-founder Wang Huiwen is "acquiring" the domestic AI framework OneFlow, adding new generals from light years away"
Pictured: Yuan Jinhui, the founder of First-class Technology
Articles related to "Dear Data" in the past:
Movie "Runaway Gamer": Software 2.0, "awakened" the game characters?
(Click on the title of the article to read it)
The short version of the introduction is:
Yuan Jinhui (Teacher Mu), received his Ph.D. in Engineering from the Department of Computer Science, Tsinghua University in July 2008, and won the Excellent Doctoral Dissertation Award of Tsinghua University. In 2013, he joined Microsoft Research Asia to engage in the research and development of large-scale machine learning platforms.
From 2015 to the end of 2016, he focused on building a deep learning platform based on heterogeneous clusters. The project won the Microsoft Research Asia Dean's Special Award (top 1%). Founded Beijing First-Class Technology Co., Ltd. in 2017, dedicated to building a distributed deep learning platform.
In February 2021, First-Rate Technology was exclusively led by Hillhouse Venture Capital in the A round, with a total financing of 50 million yuan.
For the story of who Dr. Yuan Jinhui is and what kind of company OneFlow is, please read this article.
Those who engage in deep learning frameworks are either crazy or liars (clickable)
Although the article was completed in the Spring Festival of 2021, the stories in the article are not outdated.
End.
One More Thing
My fate with the AI model starts with a 14,000-word manuscript published on the Lantern Festival in 2021.
After the publication of the manuscript "Those who engage in deep learning frameworks are either lunatics or liars" , I encountered a double sky of ice and fire. Teacher Tan, who claims to have little experience and little knowledge, was dumbfounded.
On the one hand, the article spread wildly in the circle of friends of more than a hundred experts in the field of AI and computer, and my WeChat was in a state of explosion in those few days.
There are a lot of people I know who send me screenshots like crazy. Tell me how this article went viral in their circle of friends.
For a while, I became a screenshot collector of Moments, wow so happy.
I saw a lot of retweet records of famous people, and group chat records of famous people. Although we have not met each other, we have admired his name for a long time.
Soon afterwards, an expert named Xie Yutao called me. He told me that Mr. Shen Xiangyang wanted him to convey it on his behalf, and asked me if I would like to join IDEA, and the main job was to write manuscripts.
IDEA Research Institute is the well-known Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute.
Teacher Tan is really limited.
Although this article was listed on Zhihu's weekly hot spot, the comment area was overturned.
There was a lot of sarcasm, sarcasm and sarcasm. I thought at the time that a person must have done something utterly devoid of conscience to be worthy of this "honor".
Some people even said that you are not worthy to write about technology, why don't you, a woman, write an emotional column.
Wow, what a great suggestion I hadn't thought of.
Closer to home, what is the difference between the rare reposting by more than a hundred experts and the ugly comments from Zhihu netizens?
It is likely to show that people who eat melons have a greater degree of division in their perception of artificial intelligence systems than real experts than the Great Rift Valley in East Africa.
Having said that, the world's top large-scale computer system is a headache for top architects and developers. Don't the people who write the story of this thing have a headache?
Hard things pay off, and the headaches are clearly worth it.
Because of this article, many well-known experts came to add WeChat and chat with me. Some expert resources accumulated from previous writings are also willing to recommend me to "top experts".
Couldn't ask for more, thank you so much.
One month later, in March 2021, I met a top artificial intelligence expert named Yang Hongxia at Alibaba in Hangzhou.
Don't be suspicious when you see the word "red" in the name, that's right, female scientist.
It was a sunny day in spring, and the lake-blue glass was set against Ali Baba's unique orange decoration, like beating piano keys.
From a distance, I saw the oversized "Tao Doll" standing on the outer wall of the building and grinning.
Dr. Yang Hongxia waved to me, and I walked up the semicircular slope and waved to her in response. The spring grass is green, and the wheels of the suitcase are rattling.
That time, we were talking about an AI model called M6.
丨The author of this article was photographed by Dr. Yang Hongxia, hahaha...
On that day, we also talked about GPT-3, the hottest AI model at that time.
GPT-3 was released on June 10, 2020, which touched Chinese players a lot. More bluntly, Chinese AI scientists are under pressure.
Ali's conference room is as numerous as trees in a forest, and finally sat down, a table, two cups of coffee.
Dr. Yang Hongxia has a neat ponytail, clean forehead, no bangs, and a small amount of broken hair.
When she speaks, she is straightforward, and her Mandarin is extremely fluent and standard.
She told me: "Let me tell you the truth, why did you go back to China?"
Her answers are so good, even the punctuation reveals gentle sincerity and strong technical vision.
"My favorite technology is to abstract from complex business and use technology to solve practical problems. I don't like to go the other way. If I go the other way, I can choose to go to a research-oriented university."
"Everyone must have some patience. Without patience, there will always be only a single point of result. It is difficult to realize the line and the surface."
The main contents of the exclusive interview on the day are as follows:
First, today's deep learning models can no longer meet our more requirements. Because to put it bluntly, this algorithm model may only meet some special low (low) requirements when it first appeared. Gradually, we will get closer and closer to general artificial intelligence, and the requirements must be higher and higher.
Second, Alibaba hopes to be a world leader in some very important core directions, especially in technologies and products that support the next stage of artificial intelligence.
Third, other technical details.
I won't write the technical details. If I write it, according to the rules of Teacher Tan's work, it must be submitted to the students in charge of this department for review before the manuscript can be published. If I don't write it, it will be avoided.
It can be seen that Alibaba recognized the value of the next generation of artificial intelligence very early, and they also set off early.
At the end of the conversation, Dr. Yang Hongxia invited me to taste the duck blood vermicelli soup from Alibaba Canteen, which was delicious.
As night falls on Taobao City, I understand that Chinese AI scientists are thinking: In the wave of next-generation artificial intelligence, which technologies can China be the world's number one?
After the interview with Dr. Yang Hongxia, I chatted with several technical leaders.
Jia Yangqing, Lin Wei (former researcher at Microsoft Asia Research Institute, former researcher at Microsoft Silicon Valley Research Institute, person in charge of Alibaba Cloud Machine Learning PAI platform), Cao Zheng.
They represent three different layers: framework, platform, and cloud infrastructure. Then count Yang Hongxia's model algorithm team.
An AI big model brings together so many big guys.
It can be seen that for the large AI model, Ali invested in a whole-system team, mobilizing not only the strength of Dharma Academy.
Therefore, there will be the advent of the M6 large model.
This is basically the standard way for Internet companies to make large models, and multiple teams "co-build" a model.
After many long chats with many experts, I clearly understood that to make a super AI model (ChatGPT was not available at that time), it was not enough to rely on an algorithm team alone.
The large model team is composed of several technical teams who are good at different fields (AI computing power, AI framework, AI platform).
Due to some external forces, the story between me and Dr. Yang Hongxia came to an abrupt end here, and the interview was not published.
Around the Spring Festival in 2023, I learned that Dr. Yang Hongxia had gone to Toutiao. Since then, when everyone asked me about her resignation, I kept silent.
Let's talk back to AI large models such as ChatGPT. For such an important mega-project, can the bottom layer be unimportant?
It needs to be explained that in major Internet companies, the bottom layer is shared. This is true for Alibaba, Huawei, and Tencent.
What is unique about OneFlow is that this team chose a track with a large number of manufacturers. It seems that they are not going to fight, but more like "sending death".
This is a story of "put it to death so that you can have a new life". I am thrilled to have witnessed this story "told" to this day.
At the time when ChatGPT exploded, talents were particularly valuable. "Buying" an independent team is easier to operate than "digging" a team organized by a major Internet company. This is one of the reasons why the acquisition may happen. It depends on the price negotiated by Dr. Yuan Jinhui and Light Years Beyond up.
Back to ChatGTP, AI computing power, AI framework, AI platform, and AI model work together to create the success of the AI model.
Of course, ChatGPT has an "exclusive recipe" at the model level.
Why have I been repeatedly emphasizing the joint force?
Everyone has seen that the success of ChatGPT and GPT-4 is "other wise".
ChatGPT is not the success of a lone hero, but a collaboration between technology giants and general artificial intelligence madmen, one huge and one tiny.
To put it bluntly, no matter how successful the giant Microsoft is, no matter how powerful it is, it has not managed ChatGPT by itself.
Fortunately, Microsoft has vision. OpenAI has an amazing idea of ownership structure and corporate governance.
Thus, the strong alliance succeeded.
Thus, the new model was successful.
At the same time, "Dear Data" learned that a technology giant that is "top-ranked" in China has met with multiple non-Internet AI large-scale model teams.
Its negotiation strategy is: "I" is the US Microsoft, and "you" is the US OpenAI. This strategy is really good, understand the applause.
Just to add a little bit, some major manufacturers have laid out more than one large model of the GPT type. Of course, that's another story.
All in all, China's large model teams are divided into two types: Internet and non-Internet teams.
Non-Internet teams include:
The team of Wang Huiwen, the former co-founder of Meituan, the Zidong Taichu team of the Institute of Automation, Chinese Academy of Sciences, the Zhipu Huazhang team of Tsinghua University, the West Lake Xinchen team in Hangzhou, the Zhiyuan Research Institute team in Beijing, and the Project AI 2.0 team of Li Kaifu.
The length of this list will grow, because I see some investors are already "impatient".
An AI boss once said:
"Observing the hardware and computing power platform, if there is no trouble from a certain country, on the whole, our hardware computing power, computing power, and some of our underlying core technologies, such as high-performance networks, high-performance storage, heterogeneous Institutional computing, etc. Internationally, everyone’s capabilities are leveled, similar with minor differences, and more emphasis is placed on how to make good use of these resources.”
Now that the infrastructure is solidly supported, the "behemoth" ChatGPT was born.
Now OneFlow has also been acquired by a well-known large-scale model team, which shows its significance.
The scope of new AI infrastructure is expanding.
In the beginning, the GPT series large models were not infrastructure.
In 2023, the GPT-4 large model is now part of the new infrastructure.
Under the product architecture led by ChatGPT, the linkage between the terminal application and the underlying large model is very close.
From today's point of view, the big model of AI is electricity, and we must have our own power plants.
Fang Lei, founder of Jiuzhang Yunji, said: "The big model of AI is electricity, but the power plant is not a monopoly. Everyone can build some power plants."
Next, whether it is a big thing or a small thing, we play with the big model and develop it at the application level.
I did not mention this sentence. Coincidentally, many AI bosses have said the same sentence as me. Even if it's not a super app, even if it's just a useful little thing, it's important to use it, and it's king to be popular.
Next, it depends on how the masters from all walks of life break the ceiling of ChatGPT first, and then how to make super applications.
Long-winded this a lot. Finally, I just want to say, congratulations to Dr. Yuan Jinhui and the OneFlow team.
I especially want to say to Mr. Yuan: "Please forgive me for publishing this paper before the shareholder meeting is over and the agreement has not been signed."
(This is also the origin of the word " is" in the title of the article "Former Meituan executive Wang Huiwen is "acquiring" the domestic AI framework OneFlow, and the new Lianchuang Yuan Jinhui is light years away". )
It's not easy, mutual understanding, long live understanding.
Someone asked me in the reader group: "Who will be the four little dragons of the Chinese AI model?"
The thinking of the friend who raised this question is still stuck in the last round of AI boom.
In a sense, the singularity of the big AI model has passed.
My answer is: It is possible that there will be no four tigers in the Chinese AI large model, but there will be multiple teams competing for the high ground of the large model.
What if I can't catch up? It is very possible that one family dominates and everyone dies.
In other words, my point is, if OpenAI and Microsoft win all, and there is only one winner (a couple), what will happen?
In this case, a large number of industries may at most become service providers attached to giant APIs.
This is too bad.
The good news is that the window in which advanced technologies lead is very short.
The bad news is that the "winner" layout is super snappy.
This also reflects that they know that the leading time is not so much, and they also have a strong sense of crisis.
Entrepreneurs of China's AI large model are facing mechanism changes amidst anxiety.
The "liar" I wrote in the past has disappeared, and the "madman" is in the place where the sun rises, bathed in the golden and distant morning light, and set off again, leaving a new view of the horizon.
What a thrill to have the opportunity to witness a moment in history...
(over)
read more