07/01 2024 461
Does China's large models usher in a "splendid fortune"?
After the "palace struggle," OpenAI has once again created "trouble."
Starting from July 9, 2024, OpenAI will take additional measures to strictly control API calls. If usage is detected in unsupported regions such as mainland China, Hong Kong, and Macao, it will cut off traffic or ban accounts, depending on the severity.
This means that related domestic development will face a "supply cutoff."
What about the development needs of AI applications and industry models? Fortunately, China's large models have long since improved and are vying to become "alternatives," eagerly preparing to embrace this wave of "splendid fortune."
After all, opportunities are always reserved for those who are prepared.
Replacing OpenAI, China's large models are on the move.
OpenAI is caught in a conflict between technological ideals and real-world interests.
In March 2023, OpenAI allowed third-party developers to call APIs and borrow ChatGPT's capabilities, thereby stimulating domestic enterprises to engage in related AI development.
It is not difficult to see that OpenAI's move this time targets B-end enterprises rather than C-end users.
A senior programmer told ZnKedu: "There have always been relevant restrictions on OpenAI's API calls. Before, they turned a blind eye to making money, but now they are just seriously enforcing it."
The senior programmer further stated that restrictions can be bypassed through third-party platforms, but using any method to "borrow" counts as indirect calling, which is not as convenient as direct calling. This is also why Microsoft's "Azure OpenAI Service" is difficult to become a mainstream choice in China, despite being compliant.
In fact, China's large models are flourishing, and while there are overall differences, they have advantages in specific capabilities, especially in Chinese understanding.
Coupled with the increasingly favorable prices of domestic large models, many enterprises hold an open attitude towards "moving."
In this context, internet giants such as Tencent, Baidu, and Alibaba quickly took action, launching migration services and welcoming OpenAI users to "move" to their own large models.
For example, Tencent Cloud announced that from June 25, 2024, to July 31, 2024, it will give away 100 million Tencent Hunyuan large model Tokens for free to newly migrated enterprise users and provide free dedicated migration tools and services. Currently, Tencent Cloud offers multiple versions and sizes of Hunyuan Pro, Standard, Lite, and more, allowing users to choose freely.
An industry insider said: "Domestic top model service providers offer higher cost-effectiveness, but previously many developers were dependent on GPT-4 and were unwilling to give domestic large models a chance. In the future, if they spend more time adapting to domestic large models, it's highly likely they won't go back."
Therefore, the importance of this opportunity speaks for itself.
What qualifies the Hunyuan large model to be on the table?
Amidst the competition, Tencent's Hunyuan large model has garnered significant attention due to its practicality.
Internally, the Hunyuan large model has been integrated into Tencent's product matrix and has been applied in over 600 internal business scenarios, serving a vast number of users through Tencent Cloud.
For example, WeChat's input method has introduced a "one-click AI Q&A" function based on Tencent's Hunyuan large model. Simply enter content in the chat box within WeChat and add an equal sign "=" to obtain an AI response.
Externally, leveraging the approach of "industry large models + vector databases + real-time data extraction," Tencent Cloud has provided over 50 solutions for more than 20 industries such as finance, cultural tourism, pan-internet, energy, and healthcare, helping clients build exclusive large models and intelligent applications.
This performance is closely related to Tencent's deep involvement in the AI field.
From 2018 to 2023, Tencent invested a cumulative 269.654 billion yuan in research and development expenses over six years, with 2023's R&D expenses alone reaching 64.078 billion yuan. A significant amount of resources was invested in research in key areas such as deep learning and natural language processing, ranking first in the "China AI Invention Patent Enterprise Rankings" with 15,626 patents.
Supported by high investment, Tencent's Hunyuan large model has reached the first tier in the industry, being the first to adopt a mixed expert model structure, which improves the model's overall performance by 50% compared to the dense version...
According to the "2024 China Large Model Capability Evaluation" report released by Frost & Sullivan, a global growth consulting firm, Tencent Hunyuan leads other large models in China in terms of general basic capabilities and other aspects.
It is worth noting that it is also the industry's first Chinese-native DiT architecture text-to-image open-source model, available for free commercial use by enterprises and individual developers.
To improve the efficiency of large model operations, Tencent Hunyuan's text-to-image large model has launched a dedicated acceleration library, reducing inference time by 75%. The Hunyuan DiT model has been deployed to the Hugging Face official model library Diffusers, allowing it to be called with just three lines of code without the need to download the original code library, significantly lowering the usage threshold. Within just one month of open sourcing, Tencent Hunyuan reached 2,400 stars (followers) on the open-source community Github, ranking high among popular DiT models in the community.
As Jiang Jie, vice president of Tencent, said: "In the highly competitive field of large models, Tencent Hunyuan has always adhered to its technological roots and application orientation. We will continue to learn and iterate long-term, hoping to create more valuable models, better applications, and better services for the industry."
AI search, a crucial battlefield for large models.
Apart from the B-end market, Tencent's Hunyuan large model is also aimed at the C-end market, creating the Tencent Yuanbao APP based on its model capabilities.
Nowadays, search is considered a key lever for creating killer AI applications, and AI search has become a crucial battlefield for large models. Compared to traditional search, AI search has become a standard feature for internet giants. Instead of providing a bunch of relevant links for users to judge on their own, AI search deeply understands keywords and proactively provides refined answers in the form of text, images, charts, etc., meeting users' personalized search needs and integrating into daily life scenarios, thereby solving problems that traditional search cannot address.
The issue is that the threshold for AI search is not particularly high, with many players entering the market, leading to severe homogenization.
An internet observer told ZnKedu: "The underlying logic of AI search is a competition of ecosystems. Whoever has an irreplaceable ecosystem will have differentiated competitiveness and is more likely to stand out."
And AI search capabilities are a major highlight of Tencent Yuanbao.
Backed by Tencent's ecological content system and other authoritative internet sources, Tencent Yuanbao can provide more high-quality and precise answers. Combining core functions such as AI summarization, AI writing, and multiple featured AI applications, as well as personalized intelligent agents, Tencent Yuanbao has become a practical AI tool for use in various scenarios such as work and life.
Overall, the actual impact of OpenAI's "supply cutoff" is limited. The competition among domestic large models has just begun.