Speaking of which, although Huawei has suffered unimaginable setbacks, its core business HiSilicon chips have almost been pumped from the bottom of the cauldron, and have been blocked in various links such as chip computing architecture, EDA, and wafer foundry, so that the technology company with the highest level of technology in the field of 5G communication technology cannot apply 5G to its own products, but Huawei itself is more optimistic than anyone.
This optimism is not only reflected in the signals released by Ren Zhengfei on various occasions, but also in the fact that Huawei's products and services can still maintain a high level under sanctions, only from the results of the past few years of sanctions, Huawei is worthy of being surrounded by the United States at all costs of technology enterprises, under such high pressure, Huawei actually survived tenaciously, not only its revenue scale, research and development investment, etc. basically did not cut, and even achieved positive growth in difficulties.
What makes the Chinese people's blood boil is that in the field of artificial intelligence, which is related to the next generation of technological revolution, Huawei has unknowingly been at the forefront of the world, and the strength it has shown this time has broken through US sanctions and constraints, and truly achieved independent and controllable.
Why do you say that?
This has to mention Huawei's three major advantages in the field of artificial intelligence.
The first advantage: artificial intelligence chip layout
As early as when Huawei was not sanctioned by the United States, Huawei has applied its own artificial intelligence chips in the mobile phone business, and it is the world's first smartphone that integrates artificial intelligence chips, which has become one of the main reasons why Huawei HiSilicon is focused on "taking care".
Later, the United States launched extreme sanctions against Huawei's chip business, from EDA software (chip design tools), basic hardware architecture (chip design framework), to chip foundry (chip mass production), setting barriers in almost every link, and also had a huge impact on the research and development and manufacturing of Huawei's six series of chips HiSilicon, Kunpeng, Ascend, Baron, Tiangang, and Lingxiao.
However, the sanctions did not prevent Huawei HiSilicon from continuing to develop on the chip side, nor did it stop Huawei's rapid progress, all of which stemmed from Huawei's foresight in the chip field - as early as many years ago, in order to avoid the future sanctions that are likely to occur in emerging fields, Huawei launched a research on basic technologies in the field of artificial intelligence in the field of three major elements.
The achievement in the field of computing power is today's famous Ascend AI chip. Ascend is mainly NVIDIA's GPU in the world, including Cambrian and other manufacturers of a series of AI chip products, in the past two years, in the field of domestic artificial intelligence hardware construction, Ascend's market share has risen rapidly, unique in the field of AI training.
There are two main types of Ascend chips, namely the training-oriented Ascend 910 and the inference-oriented Ascend 310, the latter was integrated into the Kirin chip as early as 2018, and the former has been extremely popular in the past two years, especially this year.
Huawei's advantage in the field of AI chips is not only the launch of chip products, but more importantly, it has established a completely independent and controllable AI chip ecosystem: it has created the da Vinci architecture for AI chips, and created a large number of development tools and software for developers to support developers' various needs for AI chips. This feature has made Ascend chips and Ascend servers with them as the core gain more and more attention and applications in the field of artificial intelligence.
Today, under the ban on the export of NVIDIA high-end GPU products to China in the United States, domestic substitution has become an inevitable road, and the field of artificial intelligence is also the next round of "hot competition" in Sino-US technology, in this context, the development of Ascend is not only a commercial "siege", but is likely to become an important turning point for the further export of Chinese technology products.
The second advantage: the layout of large models in the industrial field
Not long ago, People's Daily inadvertently mentioned Huawei's Pangu series of large models in a report, according to the content of the article, the Pangu series of large models were applied to the process of developing new drugs by a leading pharmaceutical institution, shortening the time to develop new drugs by 70% and reducing the cost by 30%.
Huawei's Pangu series of large models include NLP large models, CV large models, multimodal large models, and scientific computing large models, among which NLP large models correspond to content production and content understanding. CV large models for classification, segmentation and detection; The multimodal large model has four capabilities: cross-modal retrieval, cross-modal generation, and image speaking. The big model of scientific computing is divided into molecular large model, financial large model and meteorological large model. The People's Daily mentioned the scientific computing model in the Pangu series of large models, and according to the official website of HUAWEI CLOUD, the above four large models are "coming soon".
It is worth mentioning that the much-watched Baidu Wenxin is based on multi-modal large models, and Baidu is also the only technology company in China that has opened internal testing of large model applications. Huawei is obviously much more low-key than Baidu, but its application in industries such as industry has already begun to promote.
It can be said that if Baidu is likely to be ahead of Huawei in the field of multimodal large models, then in other fields, especially industrial applications, and basic layout, Huawei has been far ahead of domestic technology companies, and even more advanced than its American counterparts in some fields.
This leadership is not only based on the computing power provided by Huawei's own Ascend servers, but also the fundamental leadership lies in the algorithms.
Huawei and OpenAI have an essential difference in the starting point of large-scale model development: the former mainly serves specific objects in the industrial and commercial fields, while the latter serves general-purpose objects. This difference determines that the tendencies of the two algorithms are essentially different: the former will be more inclined to optimize specific objects, the optimization direction is more accurate, and it pays more attention to "accurate and correct"; The latter, because the service object is not specific, must tend to be "vaguely correct" in parameter settings.
This essential difference will cause a huge difference in the order of magnitude of parameters between the two, the general type requires larger parameters to cope with a variety of more types of needs, and the large model of a specific object class will reduce a lot of unnecessary work because the service object is very accurate.
In the vernacular, the general-purpose model is like a department store, while the object-specific model is more like a category store, such as a hardware store, shoe store, clothing store and other stores that provide specialized product services.
The strength of Huawei's Pangu model is that it has also reached the level of hundreds of billions in the parameter settings of the large model of specific objects, using the metaphor above, that is, to open a hardware store as big as a large department store - then it is conceivable to provide people with products and services to what level.
And Huawei not only competes with Baidu and OpenAI in the field of multimodal large models, it is deployed in multiple fields, and the architecture used in its large models also comes from the self-developed Encoder-Decoder architecture.
In other words, in the field of artificial intelligence software, Huawei has also created an "ecosystem", and the depth and breadth of the layout exceed the imagination of most people, and it is obviously deeper than the large model application of its counterparts in specific object scenarios in the United States.
Meng Wanzhou returned to China
The third advantage: market prestige
Some people may feel that this advantage does not seem to be directly related to the rise of Huawei's artificial intelligence, but it actually exists.
For Huawei, among the three elements of artificial intelligence, it is not worried about the constraints of computing power and algorithms, but data, which is a natural shortcoming for a manufacturing technology company.
Why was ChatGPT born out of Internet companies? Why are Baidu, Tencent, and Ali making rapid progress on big models? The fundamental reason for this is that they have the advantages of algorithms and data in the three elements, especially data, and computing power can be smashed out with money - without sanctions. Internet platform companies naturally have a large amount of data, and this data is also where their greatest wealth lies.
If Huawei wants to make a difference in the field of artificial intelligence, it must solve the data problem. Some people may say that spending money can buy data, which is a solution, but this solution is too inefficient for Huawei.
The data used in the training of large models is not only "orders of magnitude large", but also "rich" and "high-quality". These two requirements actually put forward high requirements for the authenticity, diversity and validity of data, but this requirement is difficult to quantify for the data package provided by the data vendor, in other words, even if the data provider or platform hands over a data package that "looks standard but is actually not so useful" to Huawei, it is difficult to distinguish, and can only passively accept "running and seeing", but the cost of AI training is quite high, and it may cost tens of millions of yuan after training.
How does Huawei solve the data problem? In my opinion, it's about using its market prestige to get a quality supply of data in a particular area.
For example, the pharmaceutical institutions mentioned by People's Daily, Huawei has used its market reputation to obtain a large amount of data support in the pharmaceutical field, so as to complete the formation of scientific computing large models and feed back to the pharmaceutical field.
This huge market reputation enables Huawei to solve the problem of data credibility more quickly, and at the same time establish a more long-term "mutually beneficial relationship" with related industries through technology feeding, so as to realize the continuous supply of industry data sources and provide "nourishment" for the iteration of Pangu-related large models.
Write at the end
In addition to strengths, we should also face up to shortcomings.
First of all, in terms of consciousness, we should be clear - only one Huawei is not enough. At the same time, we should not pin all the breakthroughs of high-end technology, especially in the field of chips, electronic information and even artificial intelligence, on Huawei, a private enterprise, because China needs a technological leap at the overall level, and it is impossible to rely on only one company to support this technological leap.
Secondly, for Huawei and domestic artificial intelligence-related enterprises, algorithms and data may not be the most important, we for a long time, its main energy should still be focused on the core technology of chip manufacturing, do not hope for the recovery of the semiconductor international supply chain, but to be determined to create a lithography machine that can support advanced processes.
Finally, I would like to say,. I believe that even if OpenAI has a first-mover advantage, we still have a high probability of being able to "come last" after making up for the shortcomings, and then it is really only a matter of time.
Hopefully, that day will come sooner.