From dancing in groups to leading by the national team
After half a year of blowing the wind of the big model, it finally has a new direction.
On July 7, at the 2023 World Artificial Intelligence Conference (WAIC 2023), the Artificial Intelligence Standardization Overall Group, guided by the National Standards Commission, announced that the leader of the mainland's first large model standardization task force was jointly served by Shanghai Artificial Intelligence Laboratory and Baidu, Huawei, Ali and other enterprises.
For the first batch of selected "national team" lineup, the outside world is not surprised, after all, the development of large models needs to be led by players with particularly strong technical strength. After the state clarified their status and tasks, the wind direction of the domestic large model market began to have new changes.
01 Half a year, the industry ushered in the "national team"
Since the beginning of this year, the large model has been soaring, and the speed has exceeded the development process of any previous technology. If in the first quarter of this year, players flocked to the entrance of the large model and were still discussing the question of "whether to do it", by the second quarter, each company had evolved to the specific "how" problem.
And such a lively scene ushered in a climax at WAIC 2023. More than 400 enterprises participated in the exhibition, more than 30 large models concentrated on the highlights, and the exhibition area reached 50,000 square meters, setting a record for the previous session.
In this lively conference, many people could not enter the conference site because they did not make an appointment in advance. At the conference, which was called "high specifications" by industry insiders, Internet celebrity Musk, Yang Likun, one of the three giants of the Turing Award, Hu Houkun, rotating chairman of Huawei, Tang Xiaoge, professor of Hong Kong Chinese University, and bigwigs from academia and entrepreneurship all attended.
In the exhibition hall, the era of large models, generative AI, and general artificial intelligence, which were unfamiliar to them half a year ago, have now become symbols everywhere in the exhibition hall.
Of course, the players of more than 30 large models at the conference site did not disappoint the outside world and gave their own answers to large models. In particular, the actions of the members of the "national team" have attracted the attention of the outside world.
For example, Baidu, as the first manufacturer in China to announce All in artificial intelligence, its exhibition hall at the conference site is particularly attractive. Of course, on this important occasion, Baidu will naturally exhibit the "treasure of the town hall" that more people can experience, a product known as Wen Xin Yige, which allows the audience entering the exhibition hall to achieve P map freedom.
Huawei moved its "world's fastest AI training cluster" Atlas 900 PoD A2 to the site. Ken Hu, vice chairman of Huawei, said that with Atlas 900, people can complete the training of a typical neural network ResNet-50 on the ImageNet dataset in only 59.8 seconds, which is 15% faster than the second place with the same accuracy. "It's the equivalent of hitting the line in the first place in the sprint and then drinking a bottle of water to see the second place run to the finish." Undoubtedly, Huawei's show of basic computing power on the hardware side has made industry insiders and audiences shift their attention from the complexity of large models to the competition on the hardware side.
At the Alibaba Cloud Forum, Alibaba Cloud's "Tongyi Family" added an AI painting model "Tongyi Wanxiang", which is said to be able to assist humans in graphic creation, and can be applied to application scenarios such as art design, e-commerce, games, and cultural creativity in the future. Zhou Jingren, CTO of Alibaba Cloud Intelligent Group, said at the scene that this is a key step for Alibaba Cloud Big Model to fully grasp the multi-modal capability, and this capability will be gradually opened to industry customers in the future.
"There is money, people, technology, and scenes", this is the innate advantage of large factories to make large models, but also a gap that is difficult for many start-ups to fill, and even some people directly pointed out on the scene that China's large model landing, will only be in the five large factories, namely BAT + Huawei + Tencent.
But what are the giants going to do? Where the next step will go is a big proposition.
02 Abandoning concepts and feelings, giants have focused on scene landing
At this year's artificial intelligence conference, large models have become well-deserved top.
National teams such as Ali Tongyi, Baidu Wenxin, and Huawei Pangu have shown their hard power, and at the same time, more than 30 large models of pendants, such as iFLYTEK Xinghuo, SenseTime Ririxin, and NetEase Fuxi, have not lost their momentum and are working hard in their respective fields.
But judging from the situation on the ground, it seems that they have abandoned the practice of big and empty, storytelling, and emotional telling, and have begun to focus on telling landing scenes and cases. This is the only way forward for big models, and it is also very likely to be the highlight of the next stage.
At the conference, HUAWEI CLOUD Pangu Model 3.0 was officially released, attracting many industry insiders. What impresses the industry even more is that Zhang Pingan, executive director of Huawei and CEO of HUAWEI CLOUD, said that the Pangu model is very busy, busy with things, and has no time to write poetry. And composing poetry is exactly what players who released large models in the previous six months loved to do.
In Zhang Pingan's view, Huawei hopes that the Pangu model can help all walks of life, such as finance, government affairs, mining, meteorology, etc., rather than focusing on the language model. According to it, up to now, the Pangu large model has been implemented in the fields of meteorology, medical research and development, and electric power, and has delivered a number of 100 billion parameter large models.
Also putting the scene into practice is Baidu. Baidu, as an early player, released the Wenxin model four years ago, but the industry did not pay enough attention to the large model at that time, so it did not stir up too much splash. But for Baidu, the Wenxin big model is a leading layout that is one step ahead of the industry. Now, this forward-looking product is also very rewarding.
At WAIC 2023, Baidu CTO Wang Haifeng said that Baidu has now upgraded to version 3.5 of Wenxin model, which has increased the effect by 50%, the training speed has been increased by 2 times, and the inference speed has been increased by 30 times compared with the previous version. In terms of cost, it has dropped to 10% of the past.
"Take promoting the prosperity of China's big model ecology as the primary goal, and provide a full range of services to large model startups." Alibaba Cloud CTO Zhou Jingren said. Obviously, this continues the MaaS (Model as a Service) concept proposed by Alibaba Cloud.
Tencent, which is the latest to enter the market in the field of large models, has been moving continuously in the past 20 days. On June 19, Tencent publicly revealed its thinking on the big model for the first time; On June 26, the self-developed Xingmai high-performance computing network was disclosed for the first time; At WAIC 2023 on July 7, Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, disclosed Tencent's achievements in large-model application innovation, and said that Tencent Cloud's large-model capabilities have been applied to scenarios such as financial risk control, interactive translation, and digital sapiens customer service, improving the efficiency of intelligent applications.
Of course, the large model of the subdivision field also shows strong vitality. Tang Wenbin, co-founder and CTO of Megvii, said in an interview with the media: "Application landing is the only standard to measure the value of large models, and Megvii Technology will march from visual large models to general multimodal large models. ”
Focus on the landing of scenarios, effectively provide enterprise users with solutions to reduce costs and increase efficiency, and become the focus of current large model players. In the future, the big model has already moved from "do or don't do" to the question of "how to do". And that's the next step in the big model battle.
03 Participate in the battle for the future and answer these four questions first
Although the big model is very popular, there is still a long way to go from the beginning to the market. In the process, many difficulties have been exposed.
However, in the view of Yidu Finance, the future battle for large models will probably be carried out in four latitudes. Namely: technology, talents, capital and commercialization.
Let's start with the technical aspects. There is no doubt that artificial intelligence is one of the most advanced technologies at present, and at the technical level, the accumulation required for it cannot be completed in a short time. "Large" computing power, "big" data, "large" model is the basic characteristics of the current large model, but also the challenge of the industrialization of the large model. Secondly, the size of the model is large, the training difficulty is higher, and the third is the large scale of computing power, which will have higher requirements for hardware performance.
This also means that without sufficient funds to support it, it is difficult to build such a super team. A marketing cloud founder once mentioned when communicating with YiDU Finance: "Since the investment in the industry model in March, the overall capital investment has been very large, even exceeding the total before the establishment of the company to the big model." However, he also mentioned that if it is done, it will definitely be a reassurance for the company's development in the next ten years.
Prior to this, many people in the industry had proposed that "big models are big factories burning money". This statement is not without reason.
Although the big model is hot, on a global scale, capital has not kept pace with the technological recovery. Global venture capital funding almost halved in the first six months of the year, falling 48 percent to $173.9 billion, and the number of deals fell 19 percent, according to research firm PitchBook.
In China, as of the end of June this year, more than a dozen large-model startups have received financing, and among the companies that have announced the amount of financing, the largest is MiniMax, which received more than $250 million in Series A financing from Tencent in June this year; Lightyear also received $230 million in angel+ round financing before being acquired by Meituan.
Looking at the investment of large manufacturers, previously, the statistics of titanium media are quite illustrative, in 2022, Huawei will invest 161.5 billion yuan in R&D expenses, becoming the largest domestic R&D investment enterprise; It is followed by Tencent, although much lower, but also at the level of 61.4 billion yuan, Ali ranked third, with research and development expenses of 55.5 billion yuan. According to public information, Baidu, as an early player in artificial intelligence, has invested more than 100 billion yuan in the field of AI in the past decade. Such input specifications are obviously not comparable to ordinary enterprises.
With technology and funds, large factories are relatively more attractive to talents. At the beginning of this year, the companies began a crazy war for talent. Baidu recruits AI large model algorithm engineers with a monthly salary of 25-40K, and Ali recruits large model training and algorithm engineers with a monthly salary of 40-70K.
After searching for the keyword "big model" on a recruitment platform, you will find that some companies are willing to give graduates in 2023 a monthly salary of 15-25K. At the same time, some vertical track companies have also participated in this round of grabbing wars. For example, a medical large model product manager recruited by a trading company has a salary range of 25-50K, and a game company recruiting algorithm engineers with a large language model also gives a salary of up to 50K. Even the annual salary of product managers of large-model platforms recruited by China Telecom can reach the level of 840,000.
The rising talents, technology and funds all urge the players of the big model to land and commercialize as soon as possible, after all, according to the laws of business, in the end, these inputs need to be output returns to be valuable.
However, the landing cost of large models is also the threshold that major players need to cross. Some industry insiders have estimated that the cost of training a large model is extremely high, reaching $2-120 million. This also means that the commercialization of AI large models may have to return to cost accounting.
epilogue
Looking at the big model from the present, the whole is very similar to the Internet in 1998, which was in its infancy, with a big bubble and a lot of opportunities. In this case, a really strong good company will have better growth and value in the future.