After the melee, it began to differentiate. Where did the wind of the AI ​ ​ big model blow？

Source: Securities Star Official Weibo

The tide of AI big model has risen, and the rapid development of AIGC technology represented by ChatGPT has provided the possibility of subversive innovation in content production for all walks of life. Therefore, the wave of AI big model has swept the world and continued to be hot. In China, technology giants, small and medium-sized enterprises, and scientific research institutions have entered the game one after another, starting at the track they are good at, and general big models, industry big models, and vertical big models are constantly emerging. According to statistics, China has released more than 200 large models, making it the country with the largest number of large models in the world, which shows the fierce competition.

Recently, the AI ​ ​ big model has sent big news one after another.

Apple gave up its dream of building a car for ten years and turned to a generative AI project; OpenAI released the Wensheng video model Sora；; Kimi intelligent assistant, a large model application that can support 2 million words of long lossless context input, brings the fire of long text; On the end side, notebook manufacturers such as Hewlett-Packard, Dell, Lenovo and Asus have launched new product release plans for AI PC. Samsung, Xiaomi, Glory, vivo When mobile phone manufacturers announced their efforts to deploy AI technology, and planned to launch related products … Companies in the fields of finance, construction, electricity, data, medical care, etc., also launched AI big models through self-research or cooperation, and the big models began to develop in the direction of verticalization and industrialization.

The application of AI big model has entered an explosive period. Some insiders believe that big model is helpful for enterprises to improve efficiency, but it also faces problems such as mismatch between technology and actual business and difficulty in adapting to large-scale production.

In the capital market, the AI ​ ​ big model is not as sought after as it was in the early days, and it is rare that it will rise with AI, but the determination of A-share listed companies to embrace the big model has not changed. According to IFinD data, up to now, the total market value of 152 AIGC concept stocks in A shares is 1,524.485 billion yuan. More companies are officially announcing their own big models, and the wind of AI big models is still blowing.

The head factory competes in a general big model, and its profitability appears.

In February last year, chatGPT4 came out, which once again made the market feel the power of the big model with its superior computing power and logical ability. Domestic head manufacturers ran into the market and launched a general big model.

The research and development of super-trillion-parameter large model is an all-round competition between AI and cloud computing, including algorithms, huge computing power at the bottom, networks, big data, machine learning and many other fields, which often means high costs. Therefore, the technological innovation of the basic big model is more like the arena of big companies. Baidu’s ERNIE Bot, Alibaba Cloud’s Tongyi Qianwen, Kunlun Wanwei’s Tiangong, Iflytek’s Spark and Shang Tang’s Day-by-Day Innovation appeared.

Recently, AI has made significant changes in large-scale models and applications, with GPT-4 as the symbol, and exerted its strength in the multi-modal direction. There is a breakthrough in language processing, which enriches the possibility of multimodal dialogue through visual ability, making AI more widely used in various fields.

For example, Huawei released the innovative service of Pangu big model, and released Pangu big car model and medical big model, which went deep into mining, government affairs, automobile, meteorology, medical care, live broadcast, research and development and other industries; Baidu released the first "industrial" medical model in China; Iflytek released iFiyCode1.0, an intelligent programming assistant, focusing on the code level …

From the research process, compared with openAI, the domestic general big model has changed its previous catching-up state and the gap is getting closer and closer.

Iflytek announced in its 2023 performance forecast that the Spark Model V3.5 has greatly improved its capabilities in logical reasoning, language understanding, text generation, mathematical answering, code and multi-mode, further approaching the latest level of GPT-4 Turbo; The company said that the Spark Voice Model will be released for the first time, which has surpassed Whisperlarge-v3 launched by OpenAI in the first batch of 37 mainstream languages.

Shangtang Technology said that SenseChat V4, a large language model for discussion, has a comprehensive evaluation score comparable to GPT-4. Beyond GPT-3.5.

Baidu Robin Li publicly stated that "Wen Xin Da Model 4.0 obviously surpasses GPT-4 in Chinese, and" ERNIE Bot has less English training data than ChatGPT, so there is still a gap in English effect ".

In the third quarterly report in 2023, Kunlun Wanwei said that in terms of logical reasoning, "Tiangong" was significantly ahead of GPT-3.5 and LLaMA2-70B.

Not only that, domestic large-scale model applications have entered the stage of accelerating the landing of commercial applications and rapid accumulation of effective users, and some companies have even reached the harvest period. On March 26th, Shangtang Technology announced its annual results in 2023. The company’s total revenue in Shang Tang was 3.4 billion yuan, gross profit was 1.5 billion yuan, and gross profit rate was 44.1%. Among them, the most striking thing is that Shang Tang’s revenue of generative AI exceeded 1,183.7 billion yuan, achieving a rapid growth of 200%, taking the lead in completing the "closed loop of big model business".

The 2023 annual report released by Baidu can intuitively see the real situation of Baidu AI. In 2023, Baidu’s total revenue was 134.6 billion yuan, and its non-GAAP was 28.7 billion yuan, a year-on-year growth rate of 39%. Exceeding the expected revenue and profit, among which, the growth of Baidu advertising, intelligent cloud and autonomous driving business all benefited from the help of AI.

Morgan Stanley predicts that Baidu’s advertising revenue is expected to achieve a year-on-year growth of 7% in 2024, which is mainly driven by AI, which will effectively promote the improvement of advertising conversion rate. It is estimated that by 2024, AI technology will bring about 3 billion yuan of advertising incremental revenue to Baidu.

There is also 360. In terms of revenue of AI big model, according to the first half of 2023, 360 Zhinao big model has achieved nearly 20 million yuan of related business income, mainly including membership fee income of soda office and SaaS service income of enterprise security cloud.

Generative AI, known as the "king of future realization", is hot, which reflects the value of big models. It has flooded into more than 200 enterprises in a short period of time. Industry insiders predict that more than 200 big models across the country will soon enter the convergence period, and the final market needs only one or two big models. Under the big waves, many enterprises began to look for the application opportunities of large models in the industry.

The application gradually landed, and the vertical big model "Cambrian" life explosion came?

"In the era of AI big model, the application of every industry is worth doing again."

This sentence has become the consensus of many internet giants. For some start-up AI companies, the vertical field is also a very good opportunity. The AI big model is indeed developing in the direction of verticalization and industrialization, and the wind of the AI big model has blown into the vertical fields of various industries.

At this time, the application of the big model becomes very important. No matter who can stand out from the crowd in the end, the whole industry will not be affected, and the scenes that are more likely to be realized in finance, medical care and office will perform better.

In December last year, the National Development and Reform Commission issued an announcement on publicly soliciting the opinions of the "Three-year Action Plan for Data Elements ×" (2024-2026) (draft for comments). Among them, it is proposed to support the development of large models with scientific data, deeply mine all kinds of scientific data including scientific documents, build a scientific knowledge resource base through fine-grained knowledge extraction, build a high-quality corpus and basic scientific data sets, and support the training of general artificial intelligence large models and vertical artificial intelligence large models.

Technically, the industry big model uses industry knowledge to fine-tune the big model to meet the needs in different fields such as finance, automobile and media. In this context, in addition to the vertical models introduced by Huawei, JD.COM, Baidu and other technology giants, A-share listed companies are enthusiastic about developing vertical models, and companies such as Kingdee, UFIDA, Tomcat, Dongfang Guoxin, Insai Group, Zero Count and Alte are actively deploying them.

Judging from the application, the financial industry seems to be commercialized faster. For example, Hang Seng Electronics released LightGPT, a large model of the financial industry, in June last year. In September, LightGPT and four large model products were tested by 20 financial institutions. Flush AI open platform provides customers with a variety of services, and the company’s self-developed financial model "Asking for Finance HithinkGPT" has entered the memory and filing stage …

On the end side, institutional analysis believes that AIGC brings new applications, increased functions and improved intelligence level, which is expected to increase the value of stand-alone software, and the layout of AI mobile phones is just the right time. Recently, OPPO, vivo, Huawei, Xiaomi, Glory, etc. have announced their own mobile phone models, and vivo, the first mobile phone in the industry to launch AI models, has also reaped the "dividend" under the blessing of this new technology-breaking the sales records of all new machines in vivo history on the first day of the X100 series.

In the era of automobile intelligence, the large model is expected to improve the generalization ability of autonomous driving. At the level of autonomous driving cognitive decision-making, there have been related AI models in China. Last April, Mimo Zhixing released the industry’s first self-driving generative model, DriveGPT, with the Chinese name "Xuehu Hai Ruo". DriveGPT established RLHF (Human Feedback Reinforcement Learning) technology by introducing driving data, and continuously optimized the cognitive decision-making model of autonomous driving.

Mimo Zhixing has also released three second-generation intelligent assisted driving products to get rid of the dependence on high-precision maps, and at the same time, the expensive laser radar has become an optional option.

At the same time, the application of AI large model plus cockpit is also frequently landing. In June 2023, Ideal’s voice assistant "Ideal Classmate" joined its self-developed cognitive MindGPT, which can support voiceprint recognition, content recognition, AI painting, AI calculation and other functions. In October, with the listing of Zhiji LS6, its self-developed GPT model was officially launched.

In addition, the large model injects the brain into the humanoid robot and opens up the realization of the universal robot. Humanoid robots need to have more powerful intelligence and autonomous decision-making ability, while large models have huge prior knowledge base and strong general understanding ability, which can meet the requirements of humanoid robots for universal scenes and skills, and realize complex motion control, multi-modal perception and decision-making, autonomous learning and optimization. Many domestic science and technology enterprises and cross-border leaders have laid out one after another, such as Xiaomi Cyberone and Ubisoft WalkerX.

After the two sessions, the "big model" became a hot word, which was strongly supported by local policies

In the 2024 Government Work Report, it was mentioned that "vigorously promote the construction of modern industrial system and accelerate the development of new quality productivity", and the artificial intelligence industry is undoubtedly a "new quality productivity". The report mentioned "artificial intelligence+"many times, and even mentioned the "artificial intelligence+"action for the first time. Large models, intelligent manufacturing and autonomous driving … once again became hot topics in the market in 2024.

Among them, the big model is the most popular AI subdivision field in recent years, which is regarded as an important research direction to realize general artificial intelligence and one of the core forces leading the new generation of industrial transformation.

On July 13th last year, seven departments, including the Network Information Office, jointly issued the Interim Measures for the Management of Generative Artificial Intelligence Services, which provided a policy basis for the filing of large models. Since August, 2023, SenseChat, which was discussed by Baidu ERNIE Bot and Shang Tang, and Baichuan Big Model of Baichuan Intelligent have passed the Interim Measures for the Management of Generative Artificial Intelligence Services, and started to open services to the whole society, and the filing of big models has been officially started.

It is worth mentioning that with the rapid development of large-scale model technology in artificial intelligence, local governments have issued relevant support policies to accelerate the sustainable development of large-scale model industry. According to the report jointly issued by People’s Network Finance and Economics Research Institute and Zhiding Technology, at present, Beijing, Shanghai, Guangdong, Anhui, Fujian, Shenzhen, Hangzhou, Chengdu and other places have issued relevant industrial policies for AI big models.

However, the focus of each place is different. The report shows that Beijing is striving to promote technological innovation related to large models and build an efficient and coordinated industrial ecology of large model technologies. Shanghai emphasizes building a large model with international competitiveness; Shenzhen focuses on supporting the creation of an open source general model based on domestic and foreign chips and algorithms, and supporting key enterprises to continuously develop and iterate commercial general models; Anhui started to attract large-scale model enterprises from the aspect of resources; Hangzhou supports head enterprises to tackle key technical problems of multi-modal general large-scale models, and small and medium-sized enterprises to dig deep into vertical fields to make precise special models …

Summary: After years of development, artificial intelligence finally broke out. The large-scale commercialization of ChatGPT, especially the release of ChatGPT4, stimulated the nerves of major domestic companies and began to develop their own large models, which were in a state of blooming. In just over a year, the development of AI big model began to differentiate. Some large manufacturers continued to develop general big models, and at the same time, they made efforts in the vertical field, while others small and medium-sized companies polished the industry big models in the vertical field. It is foreseeable that in a short period of time, the AI ​ ​ big model can’t decide the outcome. If major enterprises want to share a piece of the market, make good use of the useful big model, find their own track and find their own way of survival.