From July 6th to 8th, the 2023 World Artificial Intelligence Conference (WAIC 2023) was held in Shanghai, and the theme of the conference "Intelligent Connected World, Generating the Future" directly gave the hottest topic of this year - generative artificial intelligence. And its technical foundation, the big model, has also become the hottest technology topic this year, and there may not be one. According to the official data of the conference, more than 400 companies participated in the exhibition this year, nearly double that of last year and 1/3 more than the year before.
To say who is the most beautiful in this artificial intelligence conference, if Huawei says second, I am afraid that no one dares to recognize the first. How to intuitively feel Huawei's influence at WAIC 2023, just look at its booth: the area is about equal to Baidu booth + Tencent booth + Alibaba booth.
On July 6, the opening day of the World Artificial Intelligence Conference, the news that "HUAWEI CLOUD Pangu was featured in the official issue of Nature" was on the hot search. To sum up its significance in one sentence: this is the first paper published by a Chinese technology company in the past ten years as the only signatory of the "Nature" official paper.
On July 7, Huawei released Pangu Model 3.0, which was featured in Nature the day before.
Although the outside world has not been able to wait for "Huawei Chat" and "Pangu Chat", Huawei has provided a new paradigm for the real landing of large models.
What is Pangu Grand Model 3.0?
According to incomplete statistics, up to now, at least 20 domestic Internet giants have announced or released their own large models, including but not limited to Baidu, Tencent, Alibaba, Huawei, Xiaomi, 360, etc. But it seems that just as building a tram cannot show its technology if it does not benchmark Tesla, in most of the large models that have been announced, they like to use ChatGPT as a benchmark, such as Baidu's "Wen Xin Yiyan" and Alibaba's "Tongyi Qianqian". Although most of them will emphasize their B-side capabilities when promoting their products, the real commercialization has not yet arrived. Except, of course, Huawei.
Completely different from ChatGPT, Huawei Pangu 3.0 has been aimed at the industrial and scientific fields since the beginning of the project.
Specifically, in 2021, Huawei began to establish a project to make the Pangu Big Model, and in April of that year, it released the Pangu NLP (Natural Language) Big Model, Pangu CV (Visual) Big Model, and Pangu Scientific Computing Big Model. In September 2021, a large model for drug development subdivision scenarios was launched. This is Pangu Grand Model 1.0.
In 2022, Pangu Grand Model 2.0 was released. Huawei and Energy Group released the Pangu Mine Model, Pangu Meteorological Model, Pangu Wave Model, and Pangu Financial OCR Model, and began to apply the model to enterprises, industries, and scientific research.
The Pangu Grand Model 3.0, released on July 7, goes a step further, not only with larger model parameters, but also targeting more industries and more practicality. Zhang Pingan, CEO of HUAWEI CLOUD, said that Pangu Model 3.0 is an industry-oriented model series, including a three-layer architecture of "5+N+X".
Among them, the L0 layer includes five basic large models of natural language, vision, multimodality, prediction, and scientific computing, providing a variety of skills to meet the needs of industry scenarios.
The L1 layer is N industry big models, and HUAWEI CLOUD can provide industry-wide big models trained using industry open data, including large models such as government affairs, finance, manufacturing, mining, and meteorology. It can also be based on the industry customer's own data, on the L0 and L1 layers of the Pangu large model, to train its own proprietary large model for customers;
The L2 layer provides customers with more detailed scenario models, mainly focusing on specific industry applications or specific business scenarios such as government hotlines, network assistants, lead drug screening, conveyor belt foreign body detection, typhoon path prediction, etc., and provides customers with "out-of-the-box" model services.
What is the strength of Huawei's large model?
It can also be seen from the development process of Pangu large model that Huawei's commercial exploration of large models has been based on the B-end market since the beginning of development.
For this wave of artificial intelligence, Huawei founder Ren Zhengfei also made his own judgment, "In the future, there will be storms in AI large models, not only Microsoft." The direct contribution of artificial intelligence software platform companies to human society may be less than 2%, and 98% of it is the promotion of industrial society and agricultural society. ”
This paragraph can also explain why Huawei is not obsessed with "Huawei Chat", but firmly follows its own road of industrial big model.
And unlike other solutions that use external computing power providers such as NVIDIA, the computing power of Huawei's large model comes from its own Ascend computing platform, the Ascend 910 and Ascend 310 processors. Among them, the Ascend 910 was released in August 2019, and Huawei said that its computing power is twice that of the NVIDIA V100 chip (the previous generation of the NVIDIA A100) under the same power consumption. Moreover, Ascend computing is not only the basis of Huawei's Pangu model, but also can be supplied externally.
Ken Hu, Huawei's rotating chairman, said at the 2023 World Artificial Intelligence Conference on July 6 that Huawei has incubated more than 20 basic large models and adapted to more than 10 mainstream large models in the industry, and "half of the current large models in China are supported by AI ascend computing power." For example, iFLYTEK's Spark model.
Zhou Bin, CTO of Huawei's Ascend computing business, also said not long ago that the Ascend AI basic software and hardware platform can carry the computing power requirements of ChatGPT or GPT-4, achieving 20 times model compression, 1% accuracy loss, and inference latency less than 50ms.
Ren Zhengfei once said that the application of the model is sometimes more promising than the model itself, Huawei will be the underlying computing power platform of AI, but the application platform is not Huawei's option, "In the 2% of the platform contribution, we can account for a little bit." What are the opportunities for ChatGPT for us? It will stretch up the calculation, stretch the pipeline flow, so that our products have market demand. ”
Both developing large models and mastering the underlying computing power platform means that Huawei wants to do both NVIDIA and OpenAI - of course, mainly in the industrial and scientific fields. In the context of the increasingly stringent US chip policy towards China, Huawei's approach is obviously safer and more controllable, and profitable.
Zhang Dixuan, president of Huawei's Ascend computing business, also regarded the ban on the sale of GPUs such as the NVIDIA A100 in the United States as an opportunity for Huawei Ascend computing, "Now the domestic demand for computing power is strong, many large manufacturers can get computing power, but many small enterprises can't." The implication is that many small businesses need to ascend, but many large manufacturers have avoided Huawei. For example, in June this year, it was reported that ByteDance ordered nearly $1 billion worth of GPUs from Nvidia.
Although ByteDance may have chosen NVIDIA for performance considerations, for those big models that are powerful and may compete with Huawei, Huawei, which is both a referee and an athlete, may trigger the fear of other big manufacturers.
However, in any case, compared with other large models stuck in "plans", "PPT" and "demonstrations", Huawei has taken a big step forward in commercializing large models and verified its feasibility. This is a valuable experience for ChatGPT, which is still in the commercial exploration period, as well as many similar large models in China.
AI big models are overwhelming
Like the Internet "black talk" that once "deeply empowered traditional industries", the big model has become a prominent science. Although the Internet has mixed reviews, no one doubts that the Internet has transformed all walks of life, and countless enterprises have undoubtedly proved its influence. Now, the big model is generally regarded as the next Internet, and countless bigwigs are cheering for it.
Li Yanhong said that the big model will penetrate more fields and reconstruct the global digital industry; Zhou Hongyi said that the big model is not a vent and bubble, and will lead the new industrial revolution; Lei Jun said that the revolution brought by AI big models is coming...
In addition to the industry, governments around the world are also actively legislating, while being vigilant against the social and legal impact brought by AI, and actively maintaining and promoting the development status of AI large models.
At this World Artificial Intelligence Conference, the National Artificial Intelligence Standardization Group guided by the National Standards Commission announced that the leader of the first large model standardization task force in mainland China will be jointly served by Shanghai Artificial Intelligence Laboratory and Baidu, Huawei, Alibaba, 360 Group, iFLYTEK, China Mobile Research Institute and other enterprises, and officially start the formulation of national standards for large model testing.
In 1889, Paris hosted the World Fair. The most striking exhibit at the fair is the 320-meter-high, 9,000-tonne Eiffel Tower, assembled from more than 18,000 steel components and millions of rivets. Later the history books wrote, "The Eiffel Tower became a symbol of the second industrial revolution that swept the world".
Perhaps many years later, when we look at today's World Artificial Intelligence Conference, we will also find that it has also become an imprint of an era.