Are you still chatting with ChatGPT?
From the whole people keen to try early, to only a small number of people are still using, the first half of the ChatGPT hustle into the end, and from another dimension observation, the vitality of new technology in the market, only customers take real money to pay for the large model, is the positive cycle of technology-business, in the second half of the year, the industry large model rushed to open a new competition.
At the Huawei Developer Conference 2023 (Cloud) held on July 7, HUAWEI CLOUD officially released Pangu Model 3.0. Pangu Model 3.0 is a large model system designed based on industry needs and provides services for the industry, including a 5+N+X three-layer architecture.
Without haste, Huawei's Pangu large model unveiled a new veil, and it was also the first time that Huawei systematically talked about large models after the ChatGPT boom.
Titanium Media App learned that Huawei is not keen on the "100-model war", although as early as April 2021, HUAWEI CLOUD released the Pangu large model, including the NLP large model and the CV large model, and since then Huawei has successively released the scientific computing large model, the drug molecule large model, the Pangu mine large model and the meteorological large model.
"People who are familiar with Huawei think about it, 'Pangu chat' does not conform to Huawei's main channel, Huawei's strategy is to win the B-end market, the basic model has long been released, although the domestic and foreign C-end large models are loud, but Huawei does not want to mix a foot, or firmly do what they are good at, to the whole market attaches importance to the landing of large models, talking about the industry big models, Huawei must stand up," a person close to Huawei said.
Huawei is one of the earliest manufacturers in China to release large models, and the hype of the concept of the capital market has been round after round, and when the industry tide rushes to the industry large model, Huawei still can't hold back, and will put its own big model strategy and inventory.
Ken Hu, Huawei's rotating chairman, said at WAIC that the key to the development of Huawei's AI is to "go deeper and get to the ground", focusing on making AI serve the production activities of thousands of industries and serve scientific research and innovation.
At this stage, Huawei has two focus points in the development of artificial intelligence: First, build a strong computing power foundation to support the development of the Chinese industry of industrial intelligence. Second, from general large models to industry large models, let artificial intelligence serve thousands of industries and serve scientific research and innovation well.
Large model "volume" landed
A lesson from the past, a lesson for the future. In the past few decades, the curve of the development of artificial intelligence technology has ebbed and flowed, and "difficulty in landing" has always been a barrier in the industrial reality.
Before the ChatGPT boom, artificial intelligence faced the problem of scene fragmentation, and artificial intelligence did not enter the core scenario of enterprises, technology and business are not closely coupled, so it is difficult to form scale effects.
According to monitoring data from third-party website SimilarWeb, global traffic (PV) to ChatGPT's website and mobile client fell 9.7% month-on-month in June, and traffic in the United States fell 10.3% month-on-month. At the same time, the number of unique visitors (UVs) to ChatGPT fell by 5.7%, and the time spent by visitors on the site also decreased by 8.5%. This is the first time since its launch on November 30, 2022, that ChatGPT has experienced negative traffic growth.
The arrival of the inflection point was unexpected by some, but in the sense of others.
Zhang Pingan, Executive Director of Huawei and CEO of HUAWEI CLOUD, said, "At present, most applications of large models are concentrated in the 2C field, and due to the difficulty of obtaining industry data and combining technology and industry know-how, the implementation of large models in the industry is slow. ”
While the general public is still immersed in the amazing performance of ChatGPT chat, artificial intelligence manufacturers are already envisioning the commercialization of large models, and internationally, Microsoft, Amazon and other large manufacturers seek commercialization paths from enterprise-level services and explore multiple industries; In China, large and small manufacturers such as Huawei, Baidu, Alibaba, and Tencent are accelerating the investment of large models in the industry.
Huawei has seen this direction very early, and it is reported that in 2020, Huawei judged that artificial intelligence has two development directions, one is the trend of small models to large models; The second, the combination of artificial intelligence and industry, is AI for Industries, Huawei believes that AI has great imagination space in thousands of industries.
Zhang Pingan introduced that Pangu 300 can provide customers with a series of 10 billion parameters, 38 billion parameters, 71 billion parameters and 100 billion parameters of the basic large model, matching the diversified needs of customers in different scenarios, different delays, and different response speeds.
The latter, before GPT was hot, Pangu large model has been deeply cultivated in the industry, creating large models and capability sets in mining, meteorology, pharmaceutical molecules, railways and other fields, combining industry knowledge know-how with large model capabilities, reshaping thousands of industries, providing expert assistants for every enterprise and everyone, making work easier.
If Huawei's strategic prediction was still a little abrupt at that time, and there were not many references, then today's large model is enough to prove that Huawei's technology and business route are both correct.
Since the beginning of this year, Huawei has been slow to "rub" the wind of large models, but has done some basic work below the surface of the water. Since the release of the Pangu model, Huawei has been thinking about the concerns of customers in industries such as customer operations, product development, software engineering, production and supply, and marketing, adhering to its own technical propositions and R&D rhythm, not rushing to achieve results, and always pursuing technological breakthroughs and technological leadership to ensure product quality and delivery quality.
"Huawei firmly chose the big model route as early as 2020, when the popularity in the market was not as high as it is today, and there were many skeptical voices, we still persevered, regardless of the hype or not, the heat is high or low in the future, we will try our best not to be disturbed by the outside world and insist on doing the right thing." Tian Qi, chief scientist of HUAWEI CLOUD artificial intelligence, told Titanium Media App.
Talking about the overheated state of the industry, Tian Qi said, "For the top technology such as large models, the heat of the market reflects the capital's expectation of the profitability of large models on the one hand, and the public's expectations for the application ability of large models on the other hand." ”
The market is the biggest driving force, the biggest change of the large model, is to create a scale effect of the export, the upper application can be based on the large model to develop, the fragmented scene, unified, to form a set of large model solutions, Pangu model 3.0 upgrade also follows a similar logic.
In the 5+N+X three-layer architecture of Pangu 3.0 big model system, the basic big models of the five L0 layers, including natural language big model, visual big model, multimodal big model, prediction big model, and scientific computing big model, can provide various general skills and support various applications of enterprises.
N L1 layer industry big models, such as government affairs big model, financial big model, mine big model, etc., can help enterprises build their own big models based on a variety of capability combinations of basic big models, through secondary training of industry data and enterprise-owned data.
X represents a large number of L2 layer scene models, compared with the basic large model and the industry large model, the scene model is more focused on a specific application scenario or specific business, providing customers with out-of-the-box model services, for example, in the medical field, for small molecule screening, small molecule optimization, etc.
From "Nobody Believes" to Nature
On the eve of the Huawei Developer Conference 2023 (Cloud) conference, the research results of the high-resolution global AI weather prediction system developed by the HUAWEI CLOUD Pangu model team were officially published in the official journal Nature, and the accuracy of the weather prediction system based on 3D neural networks exceeded that of traditional numerical forecasting methods, and the speed increased by more than 10,000 times.
What few people know is that as recently as December last year, experts and professors in the field of international meteorology generally believed that AI is a very distant thing to achieve the accuracy of traditional numerical methods.
“There are a lot of comments I could make indicating that this is perhaps not yet quite the triumph of AI over physical modelling. despite the claims in the paper. Never the less it is a big step forward compared to other efforts. The paper has also been causing a degree of existential angst at ECMWF. However, the paper caused a degree of anxiety at the ECMWF. ）
The European Center for Medium-Range Weather Forecasting (ECMWF), the world's authoritative international weather forecasting research and business institution, first produced real-time medium-term weather forecasts in June 1979, and now Huawei Pangea Meteorological Big Model shows the world another possibility.
The core members of the Pangu Meteorological Large Model R&D team told Titanium Media App that everyone did not believe that AI methods could achieve higher accuracy and better efficiency, and ECMWF was also exploring the use of AI to predict weather, but the planned timeline was calculated in ten years, and they believed that AI methods had many problems that were difficult to break through at this stage.
For example, the resolution is not enough, provincial and district-level weather forecasts, the amount of data varies greatly, if you want to achieve higher resolution, the amount of data to reach thousands of terabytes, which is much larger than other AI application data, big data means consuming a lot of computing power, this part of the problem can be solved by heap hardware, engineering.
For example, most of the accuracy of existing AI forecasting methods is significantly lower than numerical forecasting methods, which is the main reason why many people do not believe that AI can surpass numerical prediction methods, the existing AI weather forecasting models are based on 2D neural networks, which cannot handle uneven 3D meteorological data well, and AI methods lack mathematical and physical mechanism constraints, so it will continue to accumulate iterative errors in the iterative process.
HUAWEI CLOUD proposes the 3D Earth-Specific Transformer method, which introduces absolute position coding related to latitude and height in each visual transformer module to better handle complex 3D weather data, and disperse training of models in different time periods to reduce the number of iterations of a single model, thereby reducing iteration errors.
"Not only did we make a model that was more accurate than the European Meteorological Center's data forecast, but we quickly put this model on the ground, overcoming many problems and allowing meteorological experts to verify the model results with actual measurements, and they have no reason to deny the advanced nature of AI methods." As the person above said.
The construction of a large weather model has become an empirical testimony, and HUAWEI CLOUD not only has the willingness to build a large model of the industry, but also has the tools and capabilities to put it into practice. Corresponding to Huawei's Pangu model, L0 is the basic model of scientific computing, L1 is the large model of the meteorological industry, and L2 is the application of weather forecasting.
In order to accelerate and simplify the development and implementation of large industry models, HUAWEI CLOUD provides the Pangu large model engineering platform, covering three major links: data processing, model training, and application development.
In terms of data platform, compared with traditional annotation platforms (such as automatic data cleaning), HUAWEI CLOUD Data Engineering Platform provides template-based Prompt online auxiliary compose function for SFT training, and multi-person Rank online annotation and task allocation functions for RLHF training. Compared with offline tasks for these two tasks, the measured efficiency can be increased by 3 times.
With high-quality data how to produce high-quality models, it is also necessary to ensure that the process of model development is accurate, in terms of model training, the large model development suite provides self-supervised pre-training, supervised SFT training, reinforcement learning training 3 kinds of workflows, covering the whole process from dataset creation, hyperparameter configuration to model training, evaluation, deployment, condensing the practical experience of large model experts, the complex large model development, process, standardization, simplification, to help industry users one-click start, one-stop development.
In terms of model development, HUAWEI CLOUD provides the Pangu Application Development Kit, which combines traditional software engineering with large models and provides a variety of APIs and tools to call to support enterprises in building native applications of large models in minutes.
For example, based on the basic capabilities of Pangu language large model and visual large model, as well as the Pangu large model engineering platform, after learning more than 200,000 pieces of government affairs data, including public government affairs knowledge such as policy documents and government affairs encyclopedias, and proprietary government affairs knowledge such as 12345 hotline scenes, Shenzhen Futian District Government has created a Futian government affairs big model with rich industry knowledge such as laws and regulations, office procedures and so on.
It is understood that with reference to GPT-3, the end-to-end development of a 100 billion industry model has been completed, and based on the Pangu large model engineering platform, the development of large models has been shortened from 5 months in the past to 1 month now, and the overall speed has increased by 5 times.
The other pole of the AI world
Artificial intelligence has become the focus of national strategic competition, AGI (general artificial intelligence) may change or even subvert the original logic of the world's operation, the national level emphasized: "artificial intelligence is a strategic technology leading this round of scientific and technological revolution and industrial transformation, with a strong spillover driving 'head geese' effect." ”
The combination of AI and the real economy plays an important role in the industry model, and industry reshaping, technology rooting, and openness are the differentiated advantages of HUAWEI CLOUD.
The advantage of HUAWEI CLOUD AI is that it has hundreds of projects in various industries, and based on in-depth understanding of the industry and precipitation of the core know-how of the industry, HUAWEI CLOUD Pangu model can better implement in the main business scenarios of industry customers.
Pangu has learned public data from more than 10 industries, covering finance, government affairs, meteorology, medical care, health, Internet, education, automobile, retail, etc. HUAWEI CLOUD and its partners have jointly built aPaaS in seven industries: industry, heat supply, government affairs, coal mining, education, electric power, and highways, providing the most profound industry accumulation for the Pangu model.
Zhang Pingan mentioned that others can rely on the industry's most mature AI computing power and AI ecosystem, but Huawei can only rely on its own AI root technology.
Zheng Weimin, an academician of the Chinese Academy of Engineering, previously said that the large model is one of the foundations of the new key infrastructure, and the competition of the large model is also the competition of the national science and technology strategy, and China must lay out large-model products with full-stack independent innovation, while building localized computing power, and also solving the balance between computing power energy consumption and the national "dual carbon" strategy.
To this end, Huawei has built the deepest AI stack root technology, built an AI computing power cloud platform based on Kunpeng and Ascend at the lowest level, built Ascend's computing engine CANN, AI framework MindSpore, and AI development platform ModelArts, providing key capabilities such as distributed parallel acceleration, operator and compilation optimization, and cluster-level communication optimization for large model development and operation.
"Now, based on Huawei's AI stack, our training performance for large models is not only not lagging behind, but our training performance in the big model scenario is 1.1 times that of mainstream GPUs in the industry," he said.
In addition, HUAWEI CLOUD provides an easy-to-use and reliable large model tool suite, Kaitian aPaaS that gathers APIs for massive multi-industry scenarios, and an exclusive large model community with rich high-quality courses and technical certifications, helping developers get started as experts in one stop.
Huawei has also accumulated a high density of large model talents: more than 50% of the Pangu team are doctors, and there are many "genius teenagers", the core members of the above-mentioned meteorological big model are one of them, the large model will encounter various difficulties and challenges in the training process, a team with excellent technology and daring to innovate is the core guarantee for the large model to be trained, and it is also the support for Huawei's ability to export large models to the outside world.
In terms of security, HUAWEI CLOUD provides three modes: public cloud, hybrid cloud, and large model zone, to ensure secure deployment. Establish a long-term mechanism to ensure the security and compliance of large models: including dataset source and usage compliance, data lifecycle security, building a complete data annotation and review mechanism, building model compliance policies, and ensuring model usage boundaries.
In the era of AI big models, facing the grand proposition of bottom-up independent innovation, Huawei is building another pole of AI in the world.