Datawhale dry goods
Latest: The status quo of the AI large-scale model industry
The large-scale melee caused by Chat GPT is very similar to the Internet 20 years ago.
In the past two months, there have been more than five well-known domestic Internet companies, including Baidu, Alibaba, and ByteDance, who have announced that they will enter the large-scale model. In addition, game companies, cloud computing companies, software companies... as long as they release an action related to AI or the Chinese version of Chat GPT, it can trigger heated discussions, drive a round of expansion of the capital market, and trigger a carnival of followers.
Will ChatGPT become a fleeting meteor like the metaverse concept, or will it break the old order of the Internet, build a new order, and trigger a new change in production relations and productivity?
No one can define it at this stage, and Zero State LT tries to sort out the basic context and direction.
01
gods melee
There are three major patterns
Let’s take a look at the status quo of the large-scale model melee. The current players can be divided into three categories: one is the Internet giants, the other is the small technology giants, and the other is the individual entrepreneurial team.
Internet giants include: Baidu Wenxin Yiyan, Alibaba Tongyi Qianwen, Shangtang Rixin SenseNova system, Huawei Cloud Pangu, Zhihu "Zhihaitu AI", 360GPT, Kunlun Wanwei "Tiangong" large model, Jingdong Yanxi and so on. Xiaomi CEO Lei Jun posted on Weibo on April 14, stating that Xiaomi is going all out, resolutely embracing large models and generative artificial intelligence (AIGC), and will launch related technologies and products.
Under the leadership of the giants, some small giants in the technology field have also come to an end.
Since they do not have the data volume and capital scale of the giants, their general direction is to choose to access a large model and develop more adaptable applications on this basis. For example, JD Cloud’s Yanxi artificial intelligence application platform will launch ChatJD; Dingding announced that it will officially connect to Alibaba’s “Tongyi Qianwen” model; ByteDance’s Volcano Engine will release a self-developed DPU (data processing unit) chip, and Released a large model training cloud platform, etc.; Kingsoft Office announced that it will launch WPS AI, which is similar to Microsoft Copilot products based on ChatGPT technology. Kingsoft Office CEO Zhang Qingyuan revealed that this technology is provided by the domestic company MiniMax; APUS launched 100 billion parameters, self-developed multi-mode AiLMe, the state-of-the-art Tianyan model, and 7 products based on its technology; Banma Zhixing announced the access to the "Tongyi Qianwen" model, and Zhiji Automobile became the first car brand...
There are also some "veterans" in the Internet entrepreneurial track, such as Wang Huiwen, Li Kaifu, and Wang Xiaochuan, who use their personal influence and IP to announce their entrepreneurial team to build large-scale generative models.
On March 19, Kai-fu Lee, chairman and CEO of Sinovation Ventures, announced the establishment of Project AI 2.0, positioning itself as a world-class company with a new AI 2.0 platform and AI-first productivity applications.
Wang Huiwen's post on the social platform to "form a team to embrace the new era and build China's Open AI" has also been screened. He said he invested $50 million in the company (light years away). His old partner Meituan CEO Wang Xing then announced in Moments that he would personally participate in the A-round investment of Wang Huiwen’s start-up company “Light Years Beyond” and serve as a director.
Former Google scientist Li Zhifei said in a media interview at the end of February that the short-term goal is to make a 60-point general-purpose model in June next year.
Former Sogou CEO Wang Xiaochuan also said that he will enter the battlefield of the Chinese version of OpenAI, and admitted to the media that the project is "in preparation."
Zhou Bowen, the former head of JD.com AI, also announced that he would like to recruit partners without setting an upper limit to create the Chinese version of ChatGPT together. Vertical integration from its own basic large model to application and user full-scenario closed-loop, realizes the "dual landing" of generative artificial intelligence technology and commercial value.
However, in this race against time in the official announcement, many people joked that the update speed of the PPT of domestic GPT start-up companies cannot keep up with the version iteration speed of ChatGPT.
In addition, some experts who have been deeply involved in the field of science and technology for a long time also take advantage of the east wind of AI to harvest, for fear of being left behind by the times. For example, Wu Jun, a computer scientist and natural language model expert, publicly announced on April 3 that "Chat GPT is not a new technological revolution, and it will not bring any new opportunities." The sale time is set on April 26, and the price is 69.9 yuan.
In addition to fierce competition, the "mutual confrontation" under different structures is not absent. Wang Xiaochuan said that Robin Li is a person living in a parallel universe. Although Li Yanhong did not respond, Baidu cannot be bullied. His executives responded: Wang Xiaochuan has been away from the front-line business for too long, and by the way, he also satirized the old story of Sogou not doing Baidu.
This state not only makes people a little excited, but it has been a long time since such a grand occasion has been seen by major domestic Internet companies.
With the explosion of Chat GPT, chips, artificial intelligence, and even the stock prices of companies that release a concept similar to Chat GPT have received positive feedback.
In addition to Baidu and Ali, thanks to the popularity of the Chat GPT concept, all listed companies that have entered the market have tasted the sweetness in the stock market. 360 announced the GPT large model - "360 Smart Brain" landing search scene, and released the large model the next day The stock price opened 5.53% higher, and then rose by the limit again on April 3.
CMB International released a research report stating that due to the pull of ChatGPT on the stock prices of related concept stocks in the past three months, chips, optical modules, PCBs, and servers have each risen by 65%, 146%, 45%, and 62% year-to-date.
Can the big model of giant melee start the next wave of technological revolution? Who can stand out?
02
Don't underestimate the threshold of large models
After the launch of Chat GPT-4 in February this year, the market has once again perceived the power of large models. Compared with the Chat GPT-3.5 released at the end of last year, the new generation of GPT has stronger computing power and logic capabilities, and its commercial value has also become prominent. Industry experts generally believe that this will be the high ground of technological competition that will open the next era, just like the Internet in the millennium.
But it is not easy to really do a good job in large models.
First of all, Chat GPT threshold is very high. Guosheng Securities computer analysts Liu Gaochang and Yang Ran estimated in the report "How Much Computing Power Needed for Chat GPT" that the cost of GPT-3 training is about 1.4 million US dollars, and for some larger LLM models, the training cost is between 200 between $10,000 and $12 million.
In addition to the demand for large amounts of data, training large models also requires support from chips, servers, computing power, and even electricity bills.
Some organizations estimate that if an average of about 13 million users visit Chat GPT every day, more than 30,000 NVIDIA A100GPU chips will be needed, with an initial investment cost of about US$800 million and a daily electricity bill of about US$50,000.
From this perspective, with such a large investment, the most likely to win must be the Internet giants. They are well-funded and can withstand the wear and tear of a protracted war. Compared with large manufacturers, companies such as WPS and Turing have chosen a more flexible model, accessing existing large models, and developing more practical applications on this basis.
For example, the generative artificial intelligence application "WPS AI" released by Kingsoft Office is provided by MiniMax, a domestic partner; Banma Zhixing, a subsidiary of Alibaba Group, is also connected to the large model of "Tongyi Qianwen"; Baidu's internal communication product "Ruliu" is naturally It is also connected to the "Wen Xin Yi Yan" model.
As for individual entrepreneurial teams, unless there is a continuous influx of funds, it will be difficult to form. Although Wang Huiwen, the co-founder of Meituan, declared that "50 million US dollars will be included in the group", many people exclaimed that it was a lot of money, but experts said that 50 million US dollars may not last long!
Secondly, even if you have money, it is not easy to train a large model that can withstand the test of the market when it is actually implemented.
In the eyes of many industry professionals and observers, the current shortage of domestic computing power, insufficient data model training, and even copyright disputes are difficult to ignore. For example, after the release of Wenxin Yiyan, due to frequent mistakes in AI drawings, the appearance of seemingly incongruous pictures such as "Confident", "Husband and wife lung slices", and "Fish-flavored pork shreds" made users feel both interesting and full of slots. As a stalk of "being played badly", even Li Yanhong came forward to laugh at himself and said that "Wen Xin Yi Yan is a typical product that needs feedback to continuously improve."
Ali was also overturned in the live demonstration. DingTalk’s AI function is supported by the large model of “Tongyi Qianwen”. However, in the Wenshengtu demonstration, similar problems to Wenxin Yiyan appeared, such as inputting the expression of “going through fire and water”, But there was a pot of soup with a fire-like anime character resembling glutinous rice balls.
Some people in the industry also said, "The current big language model still has the problem of outputting wrong information in nonsense. ChatGPT pearls are ahead, and it seems that there is not much time left for the big manufacturers."
From the overall situation, these large models are indeed launched in a hurry, but this is not an essential problem, because large models require long-term training and accumulation, as well as the collection and understanding of Chinese corpus data sets. So from this aspect, big factories still have many advantages. Many start-up companies do not have enough data and training trial and error tolerance opportunities.
Thirdly, with the application of large models and the popularization of AI, what is essentially deconstructed is the production relationship. In the past, people produced content, but now people train machines, and machines produce content. With the rise of AIGC (AI Generated Content), the use of AI to generate content is considered to be a new content creation method after PGC and UGC. Companies of any size will struggle to adapt to this new era without transformative restructuring.
From this perspective, the adjustment of internal personnel and production relations may pose a greater challenge to large factories, and it is difficult to turn around when the ship is too big. However, for small and medium-sized technology companies and start-up companies, it can be solved in the early stage. these questions. You can meet the new technology revolution with a lighter attitude.
03
miniature melee
Can the next "20 years of the Internet" be opened?
Based on the current situation, AI technology is generally considered to be the basis for a new round of industrial revolution. Just like the era of the Internet explosion 20 years ago, participants hope to seize this technological innovation and take advantage of the dividends of the times.
Why do you say that the large-scale melee caused by Chat GPT is similar to the Internet era 20 years ago? In essence, the emergence of Open AI, like the beginning of the Internet in 1998, is considered by many to be a revolutionary beginning.
From an objective point of view, because they all have similar backgrounds, the Internet 20 years ago benefited from the rapid development of communication and data transmission networks. The golden decade of the mobile Internet. At the current stage, thanks to the popularization of cloud computing, the improvement of infrastructure such as communication networks and base stations, domestic AI technology has been bred for many years, so taking advantage of the popularity of Chat GPT, the explosion of large models also conforms to the trend of the times. Rhythm.
From a subjective point of view, in the past two years, as the development of the Internet has entered a bottleneck period, the industry has fallen into a downturn, so whether it is the call of the market or the sentiment of users, we all expect something innovative.
But the situation today is different from the Internet era 20 years ago.
The Internet era 20 years ago has experienced different subdivisions such as the portal era, e-commerce era, mobile Internet era, and online games. In each stage, there are leading companies to promote industrial progress. In addition, in the process of rapid development in the Internet era, whether it is a portal, e-commerce, service or game, the logic is based on the construction of infrastructure, technical support, and application scenarios for APP development.
The state of the large model is still in the exploratory period, whether it is security or application scenarios, it is in the stage of confusion. Even OpenAI CEO Sam Altman (Sam Altman) said: "After we completed the training of GPT-4, we spent more than 6 months studying the safety of large models..." He also said, "Now Didn't train GPT-5, and won't be training anytime soon."
Everyone must admit that because the technology of Open AI is too new, the application scenarios are not yet clear. At present, it only stays in AI-generated content and chatting. In addition, there is no successful experience in the United States that we can completely copy, so in this competition On the road, it seems that everyone has reached a similar starting point and is crossing the river by feeling the stones.
As Li Tao, the founder of APUS, said, the current large-scale model players, many technical routes are not clear, and the industry is still in an early stage, which can even be called "the BC of AI". Under this situation, for many innovative companies , but the greatest opportunity. "New scenarios and new species. Every time a new wave of technology comes, new companies will emerge. New species will eliminate old species. This is almost impossible to be a process of inheritance and inheritance."
But who can really get the ticket to the next twenty years? No one can give a conclusion yet.
Like " Share " is a kind of support↓