(Image source: Stockcake)
Microsoft's Chinese AI team was packaged abroad, and the incident continued to ferment.
Recently, some employees of the American technology giant Microsoft in China received emails asking if they would like to move to other regions to work, asking them to choose countries including the United States, Australia, Ireland and other countries. The employees involved include teams working on machine learning and other cloud computing-related issues, the company will be responsible for family visa issues, or 700-800 employees are involved.
In this regard, Microsoft told Titanium Media App on May 16: "A small number of employees have the opportunity to choose international rotation, and employees can choose to accept rotation or continue to work in their current positions." Microsoft stressed that the company remains committed to growing in the Asian region and will continue to operate in China.
Titanium Media App learned that this time it mainly involved some employees of Microsoft's Asia-Pacific R&D Group in Shanghai, Beijing and other places.
It's not just Microsoft. Domestically, Huawei's large model is suspected of being manually manipulated when displaying Wensheng drawings on site; At the same time, ByteDance's Volcano Engine announced that the price of its main bean package model token is 0.0008 yuan/1000 Tokens, which is 99.3% cheaper than the industry average price; In addition, Tencent Cloud announced that the hybrid Wensheng diagram model has been fully open-sourced, and the Tencent hybrid MoE model of various sizes will also be open-sourced soon.
Judging from the reduction of token prices to the full open-source of Ali and Tencent models, China's AI large model market is facing an accelerated reshuffle of "bad money driving out good money": as soon as foreign countries open source, domestic countries will catch up with GPT-4, and then domestic models will be open source, and the token price will drop to 0, and commercialization is impossible. At the same time, in the context of Sino-US competition and the scarcity of AI talents, American technology giants, including Microsoft, are opening up the "Chu River Han Boundary" of AI talents with China, and the two sides are "strong and clear".
At present, the AI industry in China and the United States is ushering in new nodes and trends.
As Bloomberg analysts Robert Lea and Jasmine Lyu write in their latest report: "China will have a long road ahead of AI profitability, and an industry reshuffle could drive the sector to profitability, albeit in an industry with excess capital, this (industry profitability) scenario seems unlikely to happen anytime soon." ”
Selling bodies, laying off employees, and relocating AI talents, American AI has entered the "second half"
Overnight, a number of American generative AI startups were exposed to a crisis of funding shortages.
In the early morning of May 17, Amjad Masad, CEO of Replit, an American AI programming startup, posted an email on the social platform X, announcing that the company had laid off 30 employees, accounting for about 20% of its total workforce.
At the same time, it was reported that Reka AI, a large language model startup, was revealed to be likely to be acquired by data storage and analytics company Snowflake for $1 billion.
In addition, according to The information, Stability AI, an AI unicorn company that was in danger of selling itself or even going bankrupt, is fighting for a "life-saving money", with investors including Sean Parker, the first president of Facebook, and Prem Akkaraju, the former CEO of visual effects company Weta Digital.
Stability AI spokesperson Ben Ullmann confirmed the news, saying in a statement: "Stability AI is working exclusively with a consortium of world-renowned technology investors. Stability AI leadership is working closely with the consortium to make a significant equity investment in the company. He called the consortium "completely aligned" with the company's leadership.
Not just Replit, a number of generative AI startups have started layoffs in recent months. Including Tome, Jasper AI, Deepgram, etc., they have also shifted their business focus to enterprise users.
In July last year, one of the early winners of the generative AI boom, American AI unicorn Jasper AI, laid off an unspecified number of employees nine months after it raised $125 million at a $1.5 billion valuation. Its CEO, Dave Rogenmoser, said the layoffs were to "focus on gathering resources and becoming the best AI copilot for the marketing team."
The above news of generative AI startups highlights the important pressure faced by AI companies in the United States to develop AI models and products and explore commercial profits.
(Image source: edited and photographed by Titanium Media App)
In contrast to AI startups, tech giant Microsoft has begun to establish a "Chu River Han Circle" with Chinese AI talents.
On May 15, a number of people broke the news on social platforms that some Microsoft employees in China collectively received company emails asking if they were willing to relocate to other regions to work, including the United States, Australia, Ireland and other countries. The employees involved are mainly engaged in AI research at Microsoft. Microsoft's "packaging of China's AI team abroad" has become a hot topic of public opinion.
Microsoft's China cloud computing and AI team is involved in the global development of the company's core products. One person familiar with the matter said that it was some of the thousands of engineers in Microsoft's cloud computing division who were asked to consider a transfer. The people also said that these employees can continue to work in China if they choose not to transfer out of China.
Another person familiar with the matter said Microsoft made the redeployment proposal earlier this week.
In fact, Titanium Media App has previously reported that compared with Microsoft's investment of more than 3.9 billion US dollars (about 28.237 billion yuan) in Southeast Asian countries including Malaysia, Thailand, Indonesia, etc., its attitude towards China is "intriguing".
As tensions between China and the United States rise over the issue of leading the world's technological future, there have been discussions within Microsoft, including Nadella and President Brad Smith, about how to handle Microsoft's AI research in China. The New York Times bluntly said that Microsoft faces "a tricky balancing act."
In a previous subcommittee hearing on AI, Smith responded to U.S. senators that China accounted for only 1.5 percent of Microsoft's sales, compared with total sales of $212 billion last fiscal year.
As a result, the 1.5% share, coupled with the inability of the new Bing AI search to land in China, makes Microsoft's "isolation" of Chinese AI more and more obvious.
Since the establishment of a representative office in Beijing in 1992, Microsoft has been present in the Chinese market for 32 years. At the 2022 Microsoft China 30th Anniversary Event, as one of the first foreign-funded companies to enter China, Microsoft announced that it will further increase investment in four aspects: talent absorption, campus expansion, education investment and local ecology, and on the basis of expanding recruitment, Microsoft also plans to upgrade and expand its campuses in Beijing, Shanghai and Suzhou in the next three to five years.
The announcement of the withdrawal of Microsoft's China AI team is mainly from Microsoft's Asia-Pacific R&D Group, which has about 7,000 engineers, most of whom are based in China.
According to public information, Microsoft Asia-Pacific R&D Group was established on January 18, 2006, is Microsoft's largest and most complete R&D base outside the United States, with a complete innovation chain, covering basic research, technology incubation, product development and strategic cooperation, and is committed to creating an "open, innovative, and win-win" industrial innovation ecological environment, playing an active role in the scientific and technological innovation and economic development of academia and industry in Greater China and even the entire Asia-Pacific region, and continuously promoting the development of cutting-edge technologies in the field of computer scienceIt has quickly transformed the latest research results into Microsoft's core products and services globally and locally in China, and has played an active role in empowering the digital transformation of various industries.
Microsoft Asia Pacific R&D Group focuses on the research of next-generation revolutionary technologies, including natural user interfaces, intelligent multimedia, artificial intelligence, cloud and edge computing, big data and knowledge mining, computer science fundamentals, etc., and is committed to rapidly translating the latest research results into core products and services for global and local users and developers in China, including Microsoft Azure, Microsoft 365, Microsoft Dynamics 365, Bing Search, online advertising platforms, big data platforms, Visual Studio, Xbox, Surface, HoloLens, as well as Microsoft Translator, Microsoft Cognitive Services, etc.
At present, in terms of AI research, Microsoft has imposed a mandatory "quarantine" in China.
According to reports, Microsoft Research, located in China, does not allow Chinese researchers to use the beta version and core technology of GPT-4 in advance. At the same time, the company has also limited the research work of the institute in quantum computing, facial recognition and synthetic media. But at the Vancouver branch, researchers have free access to key technologies, including the computing power and OpenAI systems needed to conduct cutting-edge research.
In a previous offline exchange, Microsoft employees admitted to the Titanium Media App that the enterprise Copilot version of Microsoft's Greater China version hopes to cover those Chinese companies that are committed to globalization and going overseas, rather than competing with local Ali Tongyi, Baidu Wenxin, etc., rather than insisting on approval.
"The lesson of history is that only countries that learn from the world can succeed," Smith said in a statement. "Safety barriers and controls are crucial, but participation is still essential."
The "manual manipulation" of Huawei's large model was denied
Huawei's large-scale model "manual control" incident has recently appeared on the Zhihu hot list.
On May 16, in response to the news that Huawei's large-scale model Wensheng diagram was displayed on the spot and suspected to be manually manipulated, Huawei's Ascend community responded: "Instead of retrieving preset images, all the real code displayed this time will also be available on the Ascend community." ”
The reason for this is that six days ago, at a technical seminar at the Kunpeng Ascend Developer Conference held in Haidian, Beijing, Huawei demonstrated the function of the MxRAG SDK, which can be used to develop RAG applications with more than a dozen lines of code.
According to the screenshots of the video and chat uploaded on the Internet, when Huawei was demonstrating the Wensheng diagram function, it pressed Crtl-C to interrupt, and the corresponding code was displayed as time.sleep(6), and some netizens explained that the code meant: pause for 6 seconds, and then read a local picture to display.
As a result, Huawei has been questioned about its large-scale model technical capabilities.
In response, the Ascend community said that the on-site images are generated in real time, and the open source model is called. There are expressions such as time.sleep(6) in the code, which are commands to wait to read the images generated by the external open-source model in real time, rather than calling up the preset images. The real code will be available on the Ascend community, and developers are welcome to use it and provide valuable suggestions.
It is reported that the so-called mxRAG is simply understood as the retrieval enhancement generation function - retrieval, augmentation and generation. This capability is one of the most important capabilities needed to develop large AI models today.
Some analysts believe that although the Ascend community has not yet released the source code, and developers can only make empirical judgments and discussions through online code screenshots, many developer users of Zhihu said that the Wensheng diagram process is theoretically "completely unnecessary to sleep(6)". And according to the normal understanding, this does not need this string of code, if it is generated with a delay of 6 seconds, the feedback in the live demonstration is generated in 0 seconds, even if the on-site network is particularly good, the performance exceeds the computing power of NVIDIA A100, there will be no 0 seconds to generate AI pictures.
At a recent public event, Zhang Ping'an, executive director of Huawei and CEO of Huawei Cloud Computing Technology Co., Ltd., revealed that in AI large model scenarios, Huawei's Ascend server performance can be 1.1 times that of NVIDIA A100 and 1.2 times that of NVIDIA A800.
The revenue growth rate is 1%, the token is reduced to 0, and the AI large model is not profitable
Compared with the new situation of AI in the United States, at present, China's AI large model industry is forming a "volume" price war.
On May 15, ByteDance's Volcano Engine announced the price of its main bean package model token (a string of characters or symbols), saying that the price in the enterprise market is 0.0008 yuan/1,000 tokens, that is, the price of 0.8 centimeters can process more than 1,500 Chinese characters, which is 99.3% cheaper than the industry average.
Earlier, the Zhipu Large Model Open Platform (bigmodel.cn) launched a new price system, and the price of the entry-level product GLM-3 Turbo model was reduced by 80%, from 5 yuan/million tokens to 1 yuan/million tokens, and 1 yuan can buy 1 million tokens. In response to the price reduction trend of large models, the face wall intelligence said that its own product small steel cannon MiniCPM has been purchased at 0 yuan, and there is no way to reduce it.
Tan Cheng, president of Volcano Engine, said that the main reason for the price reduction is that the industry's large-scale model capacity has been greatly improved this year, and it has become very important to do the application, that is, the ecology must be prosperous. Tan said that many of the customers he is currently in contact with are trying to make large models, but the risk of innovation is very high, especially in the field of AI, so it is necessary to reduce the cost and promote more widespread use. From this point of view, both large enterprises and individuals need large models with lower cost and higher quality.
However, the decline in the price of this large model token has directly led to a slowdown in the revenue growth of AI companies, and the AI large model has become "worthless".
On May 16, the latest financial report released by Baidu Group showed that in the first quarter of 2024, Baidu's revenue will be 31.513 billion yuan, a year-on-year increase slowing down to 1%, which is far less than the 10% growth rate in the same period last year, and the growth rate of 6% in the fourth quarter of last year has also slowed down. Among them, generative AI contributed 6.9% of cloud revenue.
In fact, a few paragraphs recently posted by Li Zhifei, the founder and chairman of Mobvoi, in the circle of friends, I think it is worth pondering.
Li Zhifei said some time ago that cutting-edge technology companies have been doing it for several years and losing a lot of money, and they are still far from making money, while the CEOs of many non-frontier technology companies are full of "making money".
The brutal fact is that computing power requires money, it takes money to find AI talents, it takes money to train and inference models, and it also takes money to catch up with GPT-4. However, at the same time, after the free application of the C-end large model, the commercialization is far away, and even the landing of the B-end is also facing a market of "more wolves and less meat" and fierce competition.
So why does this phenomenon exist? Titanium media AGI believes that there are three main points:
1. The "three highs" of large models: high cost of computing power, high R&D, and high cost of talents. At present, the application of AI models emphasizes the attributes of productivity tools, and the commercialization of such software tools is single and the "hematopoietic" ability is too poor.
2. Everyone frequently "rolls" the price of tokens, whose price is lower and the service experience is better, downstream scene application companies will choose it, and can even "package" procurement with cloud services, instead of choosing non-cost-effective solutions, the concentration of the industry will be higher and higher, and the situation of "the boss eats meat, the second child drinks soup, and the third child has no water to drink".
3. Cutting-edge technology is an industry that is constantly running, whether it is NVIDIA, AMD, or OpenAI, Microsoft, Google, everyone is constantly chasing technological innovation, so the cost of research and development will continue to increase, and China's model is "you open source, I change and innovate", the basic technology research is poor, and the industry has not formed a production, learning, and research system.
"Poverty limits imagination", Zhang Jianzhong, founder of Moore Threads, said that AI startups have finally raised billions of billions, or 10 billion, which is already a lot, but in fact, most companies have more than enough intentions and insufficient strength, plus you can't buy a computing power card, it is even more difficult to make AI large models.
If we add the cost of AI talent in China: the gap is large and the price is "expensive", the annual salary of a master's degree can reach 700,000 yuan.
"At present, high-end talents are mainly concentrated abroad, and there are relatively few talents in the field of large models in China, especially those with practical experience are particularly difficult to find, and the annual salary of top talents can reach millions of yuan." Luo Yi, vice president of Yuntian Lifei, an AI listed company, once said.
As the cost becomes higher and higher, the large model is gradually becoming a game for Internet manufacturers.
Robin Li, founder and CEO of Baidu, once said that the so-called "two-wheel drive" talked about by some startups that make models in the outside world is not a good model. Doing both models and applications is bound to be distracting. Startups have limited energy and resources, and it's self-evident which has a higher success rate, doing two things at the same time or doing just one. In any case, we are very focused on concentration, "force out of a hole", when resources are limited, we should be more focused, rather than to engage in the so-called "two-wheel drive".
"Of course, the more fundamental thing is that Wenxin's function is the strongest, the cost performance is the best, and we will continue to invest, with this, all other things can be established." Robin Li still believes that Baidu is the No.1 in the field of large models.
Regarding the price reduction, Tan Bei emphasized that it is unsustainable for the To B business to exchange income through losses, so Volcano does not take this path, and more consideration is to let more people use it, "A large amount of use can polish a good model, and it can also greatly reduce the unit cost of model inference." Tan said that after the price is brought down by technical means, the industry will enthusiastically come in and try this.
Baidu later said that the use of large models should not only look at the price, but also look at the comprehensive effect.
Li Di, CEO of Xiaoice, once said that the current AI technology innovation is in a period of shock, some large model products are in a state of unfalsifiability, the industry's suitable business model has not yet been established, and commercialization is facing challenges such as low-price vicious competition, reducing costs but unable to generate high profit value, and falling into a "strange circle".
Li Di emphasized to the Titanium Media App, "We are anti-'involution' and anti-'open source', blind open source will make the AI field very messy, and it is easy for the (AI) industry to 'drive out good money', which is not a benign format." In the past, I have always believed that in the first quarter of 2024, basically the industry pattern (large models) will be settled, and more than 100 large models will become undifferentiated. ”
(This article was first published on Titanium Media App, author | Lin Zhijia, editor | Hu Runfeng)