Huawei's Pangu large model will be upgraded in July, where is the domestic large model arms race going

The wave set off by the AI big model is one wave after another. Not long ago, NVIDIA's total market capitalization exceeded trillion dollars for the first time, becoming the fifth largest company in the US stock market by value after Apple, Microsoft, Google and Amazon. This is a landmark event, with the rapid development of large models and generative AI, the strong demand for AI computing power from the global market has sent NVIDIA to the top of the wave.

"NVIDIA's market value skyrocketed by $200 billion a day, equal to one AMB and two Intels, largely stemming from the demand for computing power for AI training." On June 17, Zhang Ke, Executive Chairman of Beijing Frontier Financial Regulatory Technology Research Institute and Chairman and CEO of Baker Capital, sighed at the first stop of Huawei Developer Co-creation Day in Shenzhen.

Industry experts pointed out that the parameter scale of large models is getting bigger and bigger, ChatGPT-3 has 175 billion parameters, GPT4 is not disclosed, but it is expected to exceed the order of trillions, which is similar to the number of human neuron connections, which also makes the reasoning decision-making ability of large models reach the point of panic for scientists.

Huawei Developer Co-creation Day was the first stop in Shenzhen

At the event, the participants exchanged views on the trend of AI technology, industry applications, and developer growth. Xu Jinsong, Director of Developer Relations Dept of Huawei, said that the digital economy has become one of the main engines to promote economic growth, artificial intelligence is pushing the industry into the digital era, AI is moving from perception and understanding of the world to creating generative worlds, generative AI is very popular, and promotes the upgrading of the entire industry.

Big models are leading a new round of human-machine revolution

Looking back at the history of the development of artificial intelligence, from the 1960s to the 1990s, the first generation of artificial intelligence was knowledge-driven artificial intelligence, mainly represented by expert systems; From 1990 to 2020, the second generation of artificial intelligence was connectionism represented by data-driven deep learning, which was characterized by data + computing power + algorithms; Since 2020, the third generation of artificial intelligence has been driven by "knowledge + data", characterized by the ability to think like humans such as common sense, experience, and reasoning.

Since the advent of ChatGPT in December 2022, AI big models have accelerated the development of a new generation of artificial intelligence. When the technology of large models and generative AI continues to make breakthroughs, many people in the industry have begun to think about what kind of utility large model technology will have in commercial operation.

At the meeting, Wang Chen, an expert in Huawei's AI technology planning, said that large models are leading a new round of human-machine revolution and bringing users a subversive user experience. He introduced that in the PC era, "mouse + graphical user interface" opened the first interactive revolution; "Touch screen + gesture" opened the era of mobile Internet; "Dialogue + 5G" has opened the era of intelligent interconnection, marked by the large model technology represented by ChatGPT.

Wang Chen said that big model technology is also being used in enterprise services, through public cloud, private cloud, Finetune (proprietary training + inference resources) and other deployment models to help enterprise applications upgrade. As AI capabilities continue to grow, what Huawei and many industry partners need to consider in the future is how to use AI responsibly so that it can truly become a driving force for the progress of human society.

Wang Chen believes that the current deep learning of artificial intelligence still faces three major problems: one is poor generalization, the other is difficult to integrate with domain knowledge, and the third is safe and credible. The US Natural Science Foundation pointed out that the key to the entry of AI systems into thousands of industries is to integrate more key knowledge areas, such as the combination of AI and technologies in agriculture, physics, and chemistry; At this year's AI conference, a charter was issued on adhering to the responsible use of AI, and put forward the principles adhered to in the era of large models, such as human auditability, open source accessibility and other standards.

Wang Chen said that it is currently in the stage of alternating from the second generation of artificial intelligence to the third generation of artificial intelligence, and the "knowledge + data" drive makes the AI system have the ability to think about human beings such as common sense, experience, reasoning, etc.

He believes that there are three directions for future AI technology breakthroughs: first, based on large models, "language" drives the "AI codec" model architecture to become unified, and the future model evolves to "modal spatiotemporal dynamics"; The second is the general solver, from games to AI4Science, combined with expert knowledge to break through the NP-hard problem, and the future "prediction and control" will be deeply applied; The third is EmobodiedAI, integrated AI, AI interaction with the physical world, accumulate more experience and common sense, improve reasoning ability, and achieve symbiosis with humans.

Wang Chen also reminded that in the future we must use AI responsibly, in the past people discussed what AI can do, now talk about what AI cannot do, and in the future we must discuss what AI cannot do.

AIGC exploded, and by 2025 10% of data will be created by AI

Artificial intelligence has entered a new stage, and AIGC continues to be popular. Al Generated Content (AlGC), generative Al refers to the computer through machine learning to learn the elements of an object (item, product or task) from existing data, and then generate a new, original, real, similar to the original content of the object.

Xia Fei, an AI ecosystem technology expert at HUAWEI CLOUD, believes that the popularity of AIGC indicates the trend of artificial intelligence development, from perceiving the world to understanding the world, and now it has begun to create the world.

Gartner's 2021 book The Impact of Artificial Intelligence on Humans and Society predicted that 20% of content will be created by generative Al by 2023, and that by 2025, generative AI will generate 10% of all data, up from less than 1% today. Current generative AI is still in its infancy and is expected to be applied at scale within 2-5 years.

According to the data cited by Xia Fei, it is expected that in 2030, China's AIGC market (including only content creation, excluding Al code generation) will reach a trillion yuan. Similar to the proportion of Al market space, the global market is about 5~7 trillion. In the past 1-2 years, the market space has mainly focused on AI-generated content as a business monetization point. In the next 3~5 years, the larger market space will extend to marketing promotion, data synthesis (as a way to complete data), virtual companionship, game strategy generation, game character generation, etc.

Xia Fei also pointed out that many companies have entered the track of AIGC, but AIGC not only needs technological breakthroughs, but also a systematic project, from the underlying computing power resource requirements, to the above integration framework layer and AI platform layer need to have a certain accumulation, can not blindly follow the trend. For example, the cost of model training is very high, and it is estimated that the cost of training a large model ranges from millions of dollars to tens of millions of dollars.

The large-model arms race is a war between giants, and Baidu Wenxin, Aliyun Tongyi Qianwen, Tencent Mixed Yuan Big Model, Huawei Pangu Big Model and so on have been born successively in China. Industry insiders believe that there are only 5 companies in China that can train 20,000 GPU cards, and only 3 to 4 can make general-purpose large models.

And Huawei is an important seed player in this competition. Previously, a reporter from Nanduwan Finance Agency learned from a close person within Huawei that Huawei's Pangu Chat will be released in July this year. At the event site that day, Xia Fei also mentioned that the Pangu model will be upgraded in July.

Xia Fei introduced that Huawei Pangu Big Model has released CV large models, NLP large models, scientific computing large models, etc., Huawei has been researching in the field of large models for many years, and has a multi-modal Chinese database, combining discriminant models and generative models, which can flexibly support downstream tasks, such as text and life diagrams, text-oriented image completion, and image editing.

At the event, Xia Fei also showed that with the support of Huawei's Pangu model, AIGC has completed internal applications such as text generation, image generation, and video editing, and AIGC cooperates with mobile phone manufacturers to automatically generate mobile phone screensavers, and cooperates with brand retailers to generate posters and advertising images for different products.

Written by: Cheng Yang, reporter of Nanduwan Finance Agency