30 companies contend for WAIC: large models enter the high-dimensional war

Source: Zero State LT, Author: Zhuo Xinyue, Editor: Hu Zhanjia

How fast is the large model from concept to implementation?

The 2023 World Artificial Intelligence Conference (WAIC 2023), the most eye-catching event in the technology circle recently, gave an answer that shocked everyone: half a year! Half a year is not too long for any technological track, and even some tracks are only in the "infancy" stage, but the large-scale model track has completed the process from concept to implementation in such a short period of time.

In the first quarter of this year, the industry was still discussing the question of "whether or not". In the second quarter, the trend of various major manufacturers getting together to make large models is that players are answering "how to do it".

This answer, in essence, has been focused by more than 30 large-scale model manufacturers participating in the WAIC conference-first solve technical problems, followed by scene implementation, and finally commercial and scale.

"Previously, the difficulties in the implementation of artificial intelligence were more at the technical level, such as the field of autonomous driving. It has been developed for many years, but the commercialization still prevents players from seeing the dawn." An industry insider mentioned after visiting the WAIC conference: "Relative However, it should be noted that the market competition for large models will be more intense.”

In this WAIC conference, more than 30 companies focused on "showing off their muscles", including giants such as Baidu, Tencent, Ali, Huawei, and JD.com, as well as players in vertical fields such as SenseTime and NetEase.

At this time, news came out that the leader of my country's first large-scale model standardization task force was jointly held by Shanghai Artificial Intelligence Laboratory and Baidu, Huawei, Ali and other companies. These first batch of shortlisted companies were also given the "National Team" title. ** There are various signs that a higher-dimensional large-scale model PK battle led by the "national team" has officially started. **

From "group model dancing" to "national team leading"

Since the beginning of this year, large models have undoubtedly become the hottest point in technology, and soon became the "main theme" of various companies. It seems that no major manufacturer dares to publicly say that it will not be involved in this. According to incomplete statistics, in the first half of this year, there were nearly a hundred companies that officially announced to build large-scale models.

Major Internet companies such as Ali, Baidu, Tencent, ByteDance, and JD.com, AI companies represented by iFlytek and SenseTime, and various industries represented by educational companies such as Kidswant and financial companies such as Huashun Companies have entered the game one after another.

In addition, there is a rapid rise in the country to start a business of AI large-scale models. Many technology giants, executives and former executives of major manufacturers have devoted themselves to large-scale models to start a new business. For example, Wang Huiwen, the co-founder of the original Meituan, entered the AI big model with a high profile (currently, "Light Years Beyond" established by the company has been acquired by Meituan); Li Kaifu, CEO of Innovation Works, Wang Xiaochuan, founder of Sohu, and former JD.com AI business pioneer Zhou Bowen and others have joined the entrepreneurial wave of AI large models. Although this scene is not as spectacular as the "Thousand Regiments War" back then, it is enough to surprise the outside world for a large-scale model track that requires extremely high technical thresholds of artificial intelligence + solutions.

In recent years, the to B track has been extremely hot, and the demand for digital transformation and upgrading of enterprises has become increasingly strong, and cost reduction and efficiency enhancement are also the core demands of enterprises. Even many people in the industry believe that large-scale models have become the most promising field for incremental growth in the technology circle. And this has become one of the reasons for the rapid popularity of large models.

**At the WAIC conference held in Shanghai from July 6th to 8th, more than 30 large-scale model companies competed to display the layout and achievements at the large-scale model level, which also became a landmark event in the development of the industry. **

Baidu, as the first manufacturer in China to announce All In artificial intelligence, demonstrated its Wenxinyi style, one of the "treasures of the town hall", at the conference, allowing everyone who entered the exhibition hall to realize that everyone can "P-picture". Huawei moved its Atlas 900 PoD A2 to the scene to show off its basic computing power on the hardware side; at the Alibaba Cloud sub-forum, the "Tongyi Family" added an AI painting model "Tongyi Wanxiang", which is Ali The third large model product announced in three months time.

The industry's enthusiasm for large models has accelerated the establishment of the "national team".

On July 7, at the WACI 2023 conference, the national artificial intelligence standardization under the guidance of the National Standards Committee announced that the leader of my country's first large-scale model standardization task force will be jointly held by Shanghai Artificial Intelligence Laboratory and Baidu, Huawei, Ali and other enterprises. Since then, the "national team" of large models has assembled.

In fact, as early as May of this year, the country started the work related to the standardization of large models. The newly formed special group will undertake the standardization of large models, with the purpose of promoting the combination of large models and standardization practices, and promoting the development of the artificial intelligence industry. healthy growth. Since then, the large-scale model track, which has been noisy for half a year, has officially announced that it will move towards standardization.

The giants "compete", the track competition enters the high-dimensional battle

At the 2023 World Artificial Intelligence Conference, the large model deserves to be the "top class".

Baidu Wenxin, Ali Tongyi, Huawei Pangu, Xunfei Xinghuo, Shangtang Ririxin, Netease Fuxi and more than 30 large-scale models were all unveiled. At the conference site, the giants abandoned the concept one after another, took a step closer, and showed off the achievements of each company.

On the afternoon of July 7, at the Huawei Developer Conference 2023 (Cloud), Zhang Pingan, Executive Director of Huawei and CEO of Huawei Cloud, announced that Huawei Cloud Pangu Model 3.0 was officially released. He also said that Huawei's "Pangu model is very busy, busy with things, and has no time to write poems." This move is also considered to imply that the previously released model likes to write poems and prose at the press conference. Because Huawei hopes to use the Pangu large model to help various industries, such as finance, government affairs, mining, meteorology, etc., instead of focusing on the voice large model level.

It is said that up to now, the Pangu large model has been implemented in meteorology, medical research and development, electric power, language and other fields, and has delivered multiple large models with hundreds of billions of parameters.

Zhou Jingren, CTO of Alibaba Cloud, mentioned that "the primary goal will be to promote the prosperity of China's large-scale model ecology, and provide all-round services to large-scale start-up companies." Obviously, this continues the MaaS (Model as a Service) concept proposed by Alibaba Cloud.

Baidu is an early player, and its Wenxin large-scale model has always attracted the attention of the industry. At this conference, Baidu Chief Technology Officer Wang Haifeng said that Baidu has now upgraded to Wenxin Model 3.5. The effect is increased by 50%, the training speed is increased by 2 times, and the reasoning speed is increased by 30 times.

In addition to the "national team" news, major Internet companies such as Tencent are also accelerating the process of large models from concept to implementation.

In the past 20 days, Tencent's heavy news on the large model has spread frequently. On June 19, he publicly revealed his thoughts on large models for the first time. On June 26, he disclosed the self-developed Xingmai high-performance computing network for the first time. On July 7, Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, introduced that in terms of application innovation , Tencent Cloud's large-scale industry model capabilities are applied to scenarios such as financial risk control, interactive translation, and digital smart customer service, which improves the efficiency of intelligent applications.

** It can be seen that with the rapid influx of manufacturers, the domestic large-scale models have quickly passed the concept stage, and now each company is focusing on how to implement and commercialize. **

In the war of ascension that has already started, the threshold has been raised and the difficulty has increased. This is no small challenge for any manufacturer.

Open the "volume" in all directions to seize the correct posture of the "high ground" of the large model

In fact, although the large model is very popular, it is quite difficult from entering the game to actually landing and pushing it to the market, and many difficulties have already emerged. Funding, talents, infrastructure, scenarios, and commercialization have become an "obstacle race" that every player must overcome.

In the early stage of the development of large models, some people in the industry said that "large models are the game of big manufacturers", implying that "large models are very expensive", and only big manufacturers can afford it. According to incomplete statistics from Titanium Media, in 2022, Huawei will invest 161.5 billion yuan in R&D expenses, becoming the company with the most R&D investment; followed by Tencent with 61.4 billion yuan and Alibaba with 55.5 billion yuan. In the past ten years, Baidu has invested more than 100 billion yuan in the field of AI. The annual investment in R&D by major Internet manufacturers has enabled them to have a strong R&D team and become a well-deserved "first echelon" on the large-scale model track.

But they can't blindly put in without expecting anything in return. Judging from the current dynamics, they are all speeding up the implementation of the industry. Strong funds are invested in research and development, commercialized as soon as possible, and then invested in the development and training of AI models-this is a cycle that supports large-scale model players.

Just as Li Qiang, vice president of Tencent and president of Tencent's government and enterprise business, said: "In the era of large models, data, network, and computing power constitute the 'iron triangle' of the underlying infrastructure." But at the same time, he also said, "Models for vertical industries, It will be the tipping point of the value of large-scale models.” The implication is that capital and technology are only the necessary conditions and momentum for entering the large-scale model industry, and the real highlight is still at the landing level.

Talent is an extremely important part of the large-scale model landing competition.

In the first quarter of this year, various companies started a "war for talent". At that time, Wang Huiwen expressed his willingness to take out 75% of the shares to invite top R&D talents, and Li Kaifu called for recruiting world-class talents around the world. Baidu is willing to recruit AI large-scale model algorithm engineers with a monthly salary of 25-40k. 40-70k monthly salary to recruit large-scale model training and algorithm engineers. At the same time, on a recruitment website, the salary of large-scale model products and operation positions has reached a monthly salary level of 35-60k.

Secondly, "infrastructure" such as algorithms, computing power, and data are still the top priority of the large model. According to the evaluation of 10 large AI models at home and abroad by relevant institutions, overall, domestic large models surpass foreign models in terms of word comprehension and knowledge questions, that is, domestic AI large models [gf] 2f42[/gf]Basic cognition and learning ability of characters [gf] 2f12[/gf] is stronger. But at the same time, it should also be noted that at the data level, the development of large models requires high-quality training data sets.

In terms of computing power, although some leading technology companies such as Ali, Baidu, Tencent, and Huawei have completed the construction of data centers in my country, and players in vertical fields such as SenseTime and Megvii have invested heavily, there is still a lot of capacity. Little room for improvement.

Finally, there are landing scenarios and commercialization.

The high cost of commercialization in different subdivision scenarios and in different industries is also a common issue faced by the industry. It is estimated that the cost of training a large model is between US$2 million and US$12 million. On the whole, it will take time for the commercial realization of large AI models.

More pragmatic manufacturers choose to focus on some of their strengths. For example, Tencent took the lead in landing in the fields of finance and education, and the one-stop MaaS service reduced the burden on enterprises; after the upgrade of Baidu Wenxin's large model, the cost was reduced to 10% of the past. Up to now, Baidu Smart Cloud has achieved good test results in more than 400 scenarios with more than 300 ecological partners. Huawei has begun to work hard in its own government and enterprise fields...

It can be seen that this big model battle about the future is in full swing.

Write at the end

Under the "group model war", any player must seize the opportunity, and many people regard it as a bonus of the times. This is understandable. In this increasingly competitive track, even though there are many difficulties, domestic large-scale models are still moving towards a more complete and pragmatic technical direction and commercialization. ** This kind of industry competition situation that is high-spirited and practical is bound to accelerate the pace of development of my country's large-scale model technology and promote the overall technological upgrading of China's AI industry. **