Per reporter: Ye Xiaodan Per editor: Dong Xingsheng
On June 1, Alibaba Cloud disclosed the progress of the Tongyi big model and launched the AI product "Tongyi Listening Understanding" focusing on audio and video, which became the first large model application product open for public testing in China.
Alibaba Cloud revealed that Tongyi Listening Access has access to the understanding and summarization capabilities of the Tongyi Qianwen model, which can efficiently complete the transcription, retrieval, abstraction and organization of audio and video content anytime, anywhere, and can become a powerful AI assistant in users' work and study.
Previously, on May 6, iFLYTEK officially released the "iFLYTEK Spark Cognitive Model". It is reported that on June 9, iFLYTEK will also release new progress in this large model, including breakthroughs in open-ended question and answer, multiple rounds of dialogue and mathematical ability upgrades, and continuous improvement of text generation, language understanding, and logical reasoning capabilities.
Zhou Jingren, CTO of Alibaba Cloud Intelligence, revealed in an interview with media including the "Daily Economic News" reporter that Tongyi Listening combines the voice capabilities previously developed by Alibaba Cloud, and will launch enhanced versions and commercialize them in the next stage.
Alibaba Cloud Intelligent CTO Zhou Jingren (left) Image source: Photo by Ye Xiaodan, reporter
On April 26 this year, Alibaba Cloud proposed to move towards "product integration" and form a "1+3+1" product integration structure: the top layer is MaaS model as a service, and Alibaba Cloud opens the big model capabilities and training base to the ecosystem. It is worth noting that Alibaba Cloud's Tongyi Qianwu released this time can be used as both a SaaS application layer and can be integrated as a MaaS layer.
It can be said that Alibaba Cloud has raised the big model to a new level. Daniel Zhang, chairman and CEO of Alibaba Group, said that Alibaba Cloud wants to fully move towards "model as a service". On the one hand, Alibaba Cloud hopes that the application of large-model technology will open up a larger entry point. On the other hand, Alibaba Cloud took the lead in announcing a large price reduction for public cloud at the end of April, which triggered the price adjustment of many cloud service vendors in the public cloud market.
Whether technology iteration and price decline can release new incremental space and form a positive cycle remains to be verified by the new quarterly financial report. For many cloud computing vendors, large-model applications have brought the "iPhone moment" of the cloud computing market, and a new pattern is opening.
Public cloud "price war" chain reaction
Following Alibaba Cloud's announcement on April 26 that the price of public cloud has been significantly reduced, leading cloud vendors such as Tencent Cloud, Mobile Cloud, and JD Cloud have also reduced prices.
On May 16, Tencent Cloud announced price reductions for a number of core cloud products, with some product lines reduced by up to 40%; On May 23, JD Cloud said that the official website list prices of all core products are lower than the price of the target product list on the official website of specific cloud vendors, and the actual transaction unit price will be discounted by nine on the basis of the actual lowest unit price of specific cloud vendors.
It is worth noting that the public cloud price war has also stirred up the private cloud market. In early May, an IT service provider listed company told reporters that the price war in the public cloud market will further cause the price of the private cloud market to fall.
Image source: Photo by Ye Xiaodan, reporter
Recently, a third-party service provider of a leading cloud service vendor also revealed in an exchange with the "Daily Economic News" reporter that the decline in the price of public cloud will inevitably bring certain volatility to private cloud. "But this is related to the scale and industry of the company, such as large central enterprises, state-owned enterprises, and the financial industry pay more attention to security and confidentiality, and choose privatization more, while ordinary companies will choose according to price and demand."
Gan Yunfeng, chairman of Shulan Technology, said in an exclusive interview with reporters at the "Digital Survival Conference" at the end of May that the large price reduction of the public cloud market has an impact on the construction cost of private cloud, but because the price cost of private cloud is relatively fixed, the price fluctuation space of private cloud will not be very large.
As for the public cloud market price war, Gan Yunfeng said that this can have a better promotion effect on downstream data application service providers. "In the past, customer digital services often accounted for most of the funds invested in hardware, but with the decline of public cloud market prices, customers can invest more expenses in business strategy and application level, which is also more conducive to the development of the industry, especially the development of data application fields, allowing end customers to invest more funds in business scenarios and mine more value in industrial data."
As a data application infrastructure provider, Dataland Technology has built a one-stop data middle office infrastructure "data habitat platform", providing product matrices such as data middle office solutions, data visualization services, and data intelligence solutions. However, Gan Yunfeng also mentioned that the customer's choice of cloud computing vendors is generally not adjusted according to the price, but first based on the customer's choice of matching servers. However, based on cost considerations, more cost-effective servers will be recommended so that customers can have more funds to invest in real business scenarios.
And big model technology has brought new changes to the cloud computing market. According to the analysis of China Merchants Securities Research Report, on the whole, the positioning of large models has shifted from simple technology empowerment to platform ecological entrance card slots, and the emergence of ChatGPT plugins will accelerate the clarity of ecology and the prosperity of the application layer in the "AI+" era, and the synergy between the two will bring full-dimensional subversive changes to modern human life.
Zhou Jingren revealed that there are still many new needs based on large models among various enterprises and industries, and Alibaba Cloud is still continuing to explore, and will release a series of new products next. For Alibaba Cloud, which proposed "model as application", while the price of public cloud is declining, the effectiveness of creating a new ecological entrance through large models and large-scale effects remains to be verified by the market.
Is it inevitable to move towards open source?
However, from the current point of view, large model technology is still some way from the prosperity of the entire ecology.
As for the AGI industry structure represented by the large model, at the previous 2023 Alibaba Cloud Summit, Wang Jing of China Merchants Venture Capital said in her speech that the overall structure is divided into four layers, namely the computing layer, the model layer, the middle layer and the application layer. Among them, such as how fast and how much the current computing power and training costs will fall, will affect the entire industry chain.
China Merchants Venture Capital Wang Jing Image source: Photo by Ye Xiaodan, reporter
AI large models have spawned huge computing power requirements, and at present, A100 is the "main chip" for large model applications. According to media reports, in the current AI era, "computing power is power" seems to have become the main theme of the industry, and the demand for computing power facilities continues to rise. OpenAI pointed out that in order for AI large models to continue to make breakthroughs, the computing resources required to be consumed must double every 3~4 months, and funds also need to be matched by exponential growth.
The A100 costs about $10,000 and is currently the "workhorse chip" for AI applications. It is also understood that some retailers previously quoted another chip H100 for about $36,000, and the price has risen significantly recently.
NVIDIA's research shows that the largest GPT3 model requires 175 Billions, 512 V100 graphics cards for 7 months or 1024 A100 chips for a month. Large model training costs millions of dollars per month.
The development of large models is expensive and has become an expensive attempt by domestic and foreign giants.
Image source: Visual China - VCG41N1462568496
In the eyes of the industry, the public cloud is the best carrier for the new AI infrastructure. The head cloud vendors have built ultra-large-scale, high-performance, low-cost computing power public facilities, and do not need to "reinvent the wheel".
Wang Jing said that in the basic general model layer of heavy capital and talent investment, domestic large manufacturers will be important participants in the general model. At the same time, a very small number of startups have the opportunity to make a general-purpose model that is useful in China with their talent and technology accumulation, as well as innovation and flexibility.
In the field of application layer and middleware, Wang Jing said that many fields have not yet developed in China, and with the improvement of the ability of subsequent domestic basic models and the decline of industrial costs, it is expected that a hundred flowers will bloom in the future.
Fang Yi, chairman of Daily Interactive, revealed in an exclusive interview with the "Daily Economic News" reporter that the large model algorithm will slowly move towards open source, and it is estimated that in the next three years will usher in the "Android era", that is, the era of open source large models or reaching a considerable level, there will be better and more lush ecological development.
In response to the open source question, Zhou Jingren said: "We actually support open source very much, and through a series of open source models, we can reduce the cost of learning and try some of our own breakthroughs." Each enterprise has a specific core business environment, so it needs to make more refined choices about what kind of model to use. ”
On May 30, the China Academy of Information and Communications Technology and all parties in the industry jointly compiled the "Paper Kite" open artificial intelligence model license, aiming to lay a solid foundation for jointly building a transformative large-model open source project.
Guo Xue, director of the open source and software security department of the Cloud Institute of the China Academy of Information and Communications Technology, said that the nature of AI big models determines that open source is the only way to go. The compilation of the "Paper Kite" open artificial intelligence model license officially in order to give full play to the versatility advantages of large-scale models and promote the real landing of large-scale model technology in the industry. In the next step, the Paper Kite Open Artificial Intelligence Model License (Draft for Comments) will be released.
Looking for high-quality "calculations"
Daily Interaction (300766. SZ, stock price 18.63 yuan, market value 7.454 billion yuan) Chairman Fang Yi believes that big model computing requires big data training, so high-quality data is very important. Algorithms and computing power are very important, and the "calculation material" of data may be more important, and "calculation material" will become a very important engine.
Gan Yunfeng, chairman of Shulan Technology, pointed out that large model technology has brought two main impacts to the big data industry. First, basic capacity building can be completed through large model technology, which greatly improves efficiency. On the other hand, there is currently a contradiction between the rapid growth of data and the growth of computing power, the progress of technology can no longer keep up with the growth rate of data, and the application of large model technology has very high requirements for GPUs.
Since there is no solution to the high cost of computing power, when hardware development cannot keep up with the speed of software development, can we start with data and make up for the lack of computing power by mining the organizational capabilities of the data itself? Gan Yunfeng believes that it is very possible, and the soil of AI is data.
Gan Yunfeng, chairman of Shulan Technology Image source: Photo provided by interviewee
In the context of the "hundred flowers" of large models, Minsheng Securities pointed out in the research report that "the threshold for simply releasing a large model is not as high as the market imagines", "it is not difficult to have a large model, the difficult thing is to have a high-quality large model that can continue to iterate and improve performance". It is also proposed that high-quality data is a scarce element in the development of large models.
Data elements have also become a segment of the capital market that has attracted much attention. On June 6, the data element concept stock rebounded, Golden Bridge Information (603918. SH, stock price 34.08 yuan, market value 12.533 billion yuan), China Science Chuan (601858. SH, stock price 48.14 yuan, market value 38.055 billion yuan), Cape Cloud (688228.SH, stock price 61.84 yuan, market value 4.152 billion yuan), COSCO Haike (002401. SZ, stock price 28.11 yuan, market value 10.454 billion yuan), Meiyapaike (300188. SZ, the stock price is 25.18 yuan, the market value is 21.642 billion yuan) and so on.
Recently, the latest research reports of many securities companies have also intensively covered the data elements.
Deppon Securities introduced that the upstream basic layer of the digital economy lies in the value of data, and the upstream includes three subdivisions: data elements, information creation, and digital infrastructure, of which data elements are the basis of the entire digital economy industry chain cycle.
TF Securities analysis believes that algorithm, computing power and data are the three core elements of AI, the current training of AI large models, the algorithm end converges to the neural network Transformer model, the computing power side relies on AI server clusters with large-scale parallel computing capabilities, and the data side needs a large-scale dataset feeding with a huge amount of data. Among the three elements of AI, data is the key to directly affecting the landing effect of AI large models in vertical industries, while vertical data is usually mastered by governments and industry institutions, and data scarcity is obvious compared with models and computing power.
At present, the training data set of general large models is mostly from Internet text data, for example, the ChatGPT training dataset comes from Wikipedia, Internet news, social media, e-books, etc. Bloomberg released a large-scale language model BloombergGPT for the financial field, training using 51.27% financial information data and 48.73% public data, under the scale of 50 billion parameters, BloombergGPT is better than the general large model for financial tasks.
So, will the game against "calculation materials" become more heated in the future?
Fang Yi believes that the data element market will eventually form a certain degree of open cooperation. The first-level core data developers may be mainly central enterprises and state-owned enterprises, and the second-level developers may have more private enterprises and market-oriented forces involved
Daily economic news