Since the explosion of ChatGPT in March this year, there has been a trend of "preparation" for large models in China, and major manufacturers have made heavy bets: all in AI for many years Baidu launched "Wen Xin One Word" to fight with ChatGPT, Huawei's "Pangu" swept the industry, Ali's "Tongyi Qianqian" landed in the office field... The fiery degree of the big model seems to shine a ray of light into the haze of "the peak of the flow of the big factory" that has been singing for a long time.
Not long ago, on August 31, on the AppStore free list, "Wen Xin Yiyan" after opening to the public rushed to the top of the list, Baidu official showed that Wen Xin Yiyan exceeded the 1 million mark on the first day of its launch.
Prior to this, a number of large model products including Wen Xin Yiyan required internal testing applications. On March 16 this year, Wenxin Yiyan officially opened the closed beta, and the first users can log in through the invitation code and experience the product on the official website of Wenxin Yiyan. In early July, Baidu launched the Wen Xin Yiyan Apple App, but that is that users still need to obtain the internal test qualification to experience the corresponding functions.
This time, while Baidu's AI big model reached the top in one fell swoop, it seems to have unveiled the "100-model war" of domestic large models. Because, several other companies such as Baichuan Damodel, Zhipu Qingyan, and Discussion SenseChat were also opened for use on the same day.
Domestic large models have entered the era of blowouts!
The battle of big models, a war without gunsmoke
A large model, or a large language model, refers to a deep learning model trained with large amounts of text data to generate natural language text or understand the meaning of language text.
Simply put, large models that can simulate the process of human learning language, understand and generate text in a human-like way, is an important way to artificial intelligence.
Siri and Tmall Genie that we often use in our daily life are actually based on the basic application under the large model, but because the AI in this period is "not intelligent enough" and is nicknamed "artificial intellectual disability", but no one thought that in just a few years, AI will evolve from "understanding instructions" to "executing instructions", and even be able to learn by itself based on instructions, creating copywriting, scripts, drawings.
From the perspective of the research and development of large models, large models are mainly divided into three categories: one is self-developed by large manufacturers. For example, Baidu's Wen Xin Yiyan, Ali's Tongyi Qianwen, Huawei's Pangu, Tencent's Mixed Yuan, JD.com's Yanxi, Byte's Volcano Ark, Ant Group's Zhenyi, Xiaomi's MiLM-6B and so on. One category is independent entrepreneurial teams. For example, Wang Xiaochuan, the founder of Sogou, officially announced "Baichuan Intelligence" on the second floor of Sohu Building in Wudaokou. There is also a category of academics. Compared with the MOSS large model released by the team of the School of Computer Science of Fudan University at the beginning of this year, ChatGLM of Tsinghua, and Zidong Taichu of the Institute of Automation of the Chinese Academy of Sciences.
In the capital market, large model manufacturers at home and abroad are also favored by capital.
After the explosion of ChatGPT, in just six months, Anthropic, an artificial intelligence company founded by former OpenAI leaders, received three rounds of large financing, totaling more than $850 million.
The number of AI large models publicly released in China has already been hundreds, and the companies with the AIGC label alone have raised 5.89 billion yuan in financing transactions in the first half of 2023, with 42 events, far exceeding previous years.
The Chinese manufacturers with a keen sense of smell have even shown the courage to make a desperate bet!
The first shot was fired at Baidu's "Wen Xin Yiyan" - a iteration of Baidu's ERNIE 3.0 Titan iteration model launched in December 2021, which Baidu founder Robin Li said would be used to reconstruct all of Baidu's applications. Six years ago, Baidu launched the "all in AI" strategy to transform artificial intelligence, which now seems to be a very forward-looking strategy.
Subsequently, on April 11, Alibaba Cloud launched the Tongyi Qianwen language model to face Wen Xin's words. Daniel Zhang, Chairman and CEO of Alibaba, announced that all Alibaba products will be fully upgraded to access the large model.
If you often use DingTalk documents, you will find that the AI document function is quietly on the upper left corner, and AI has unknowingly entered your office field, which can easily help workers "move bricks" - make PPTs, forms, design posters, and write copywriting.
On April 24, the official website of iFLYTEK Spark Cognitive Big Model was officially launched. As an AI voice leader, iFLYTEK Xinghuo cognitive big model is arranged around scenarios such as "knowledge question and answer, code programming, mathematical calculation, creative association, and language translation", and tries to deeply integrate with educational vertical scenarios.
On June 28, Byte's Volcano Engine released the large model service platform "Volcano Ark", providing enterprises with a full range of platform services such as model fine-tuning, evaluation, and reasoning, and integrating multiple large models for customers to directly compare.
On July 13, JD.com released the Lingxi Big Model: The Lingxi Big Model integrates 70% of the general data and 30% of the native data of the digital intelligence supply chain, and penetrates into knowledge-intensive and task-oriented industrial scenarios such as retail, logistics, finance, health, and government affairs to solve real industry problems.
On July 18, Huawei and Shandong Energy Group held a press conference to announce the first commercial use of Huawei's Pangu model in the mining field. Help all walks of life, such as finance, government affairs, mining, meteorology and other industries, use the Pangu model to empower product development, production supply chain, marketing, and digital operations.
Tencent's hybrid model was announced on August 3 to enter the closed beta stage, mainly used in Tencent advertising, automatic generation of 3D virtual scenes, conversational intelligent assistants, etc. On September 7, Tencent announced that the hybrid model will be open to the public today, allowing users to experience it through Tencent Cloud, supporting direct calls to API interfaces, or using hybrid elements as a base model to fine-tune on the public cloud.
Since then, from BAT to Huawei Xiaomi, domestic and foreign giants have all come down.
What is the use of large models? Will it affect my monthly salary of 3,000 yuan?
Overall, cloud vendors use large models to consider two main levels:
The first is to improve productivity: including universal AI applications, that is, using large models to complete the rapid customization of small models, and the other is to quickly use large models into existing products, including intelligent Q&A, clothing design and other scenarios, and constantly using large models to explore the boundaries of applications.
The second is to increase influence: including relying on large models to promote a wave of their own frameworks (Baidu's PaddlePaddle, Huawei's MindSpore); Challenge the limits of human intelligence with the help of large models (CLUE in Chinese, GLUE in English, etc.); Develop all kinds of fun but not necessarily profitable apps (AI writing poetry, AI painting, especially good at exhibitions), maybe one day you can monetize.
Specifically, they include:
Natural language processing (NLP): In the field of natural language processing, large models are widely used for tasks such as language models, machine translation, and question answering systems. By using a large-scale corpus for training, large models can better understand the meaning and context of the language and produce more accurate text.
Pay attention to those who learn languages or do translation!
Computer vision (CV): In the field of computer vision, large models are used for tasks such as object detection, image classification, and image generation. By training on large-scale image datasets, large models can extract deeper and more advanced features, thereby improving the accuracy of image recognition and understanding.
Mainly used in digital marketing, AI can save manpower to do image recognition and label management, yellow inspectors may be unemployed!
Financial risk management: In the financial sector, large models are used for risk prediction, market prediction, and fraud detection. By processing large amounts of market data and transaction records, large models can analyze market trends and risks and provide strong decision support.
The arrogance of fraudsters in northern Myanmar will be stifled!
Medical diagnosis: In the medical field, large models are used in disease diagnosis, image interpretation, and drug development. By processing large amounts of patient data and medical images, large models can assist doctors in making accurate diagnoses and treatments.
Gospel for all earthlings!
Transportation and urban planning: In the field of transportation and urban planning, large models can help optimize transportation networks, improve traffic flow, and safety. By using large-scale traffic data and urban data, large models can simulate and predict the operation of urban transportation and provide decision support to optimize transportation planning.
Perhaps, the traffic jam problem will be solved?
For ordinary people, if you find that most of your daily work can be replaced by AI, you need to have multiple eyes.
Everyone should improve their understanding of the field of AI, get started with AI, learn to use AI, and control AI just as we use computers and mobile phones now.
We should proactively introduce AI tools into our work. After all, under the crush of AI, no job is absolutely safe. Instead of being passively optimized, it is recommended that everyone take a step first, take the initiative and thoroughly, integrate AI capabilities into their own workflows, and use AI to revolutionize their work.
Giant games, besides the big model, what are they still rolling?
"No matter what the track, as soon as the big factory does it, you know that it will start rolling" - this industry joke is now being fulfilled on the big model.
At the 2023 Baidu Cloud Intelligence Conference, Li Yanhong, founder, chairman and CEO of Baidu, also mentioned that the volume model is meaningless, and the opportunity for volume application is greater.
In China, the 100-model battle at the general model layer is far from being won or lost, and no one has a decisive advantage in terms of effect, but it is followed by a lot of innovative exploration of application forms.
AI seems to have more imagination space at the application layer. Just like when the computer was invented, no one could see the application form of the Internet now. The greatest value of AI, where new giants are born, is the future things it produces, which we don't see and can't see.
The application field of AI, from easy to difficult technical difficulty, can be divided into three levels: "helping decision-making, assisting creation and replacing execution".
1) Helping decision-making is AI forming knowledge on the basis of data and information, and then helping humans make decisions and complete specific tasks that do not require high accuracy. Mainly used in life and office and professional services. For example: intelligent assistant: daily life, office management, etc.; Professional services: advertising, education, finance, medical care, logistics, security, electricity, etc.;
2) Assisted creation is the ability of AI to form logical reasoning on the basis of knowledge, assist content creation, and achieve creative goals. Mainly used in information, text, images, film and television, games, etc.;
3) Alternative execution is the solution of high-precision execution capabilities formed by AI on the basis of logical reasoning, which is mainly used in the field of intelligent machines to replace human execution with high-precision requirements. Such as smart cars, smart robots, smart factories, etc.
At present, generative AI can be applied to B-side and C-side scenarios, typical application scenarios such as B-side office, marketing, finance, etc., C-side education, social networking, e-commerce.
From the perspective of application landing speed: the B side is faster than the C side. Since AI To B applications are mainly aimed at specific minority user groups such as internal enterprises, they focus on reducing production costs and increasing efficiency in specific fields, and have strong tool attributes, so they are easier to use.
The implementation speed of AI To C applications is slower than that of To B products, mainly because it is aimed at using as many users as possible, and has higher requirements for the standardization of products, and also needs to be combined with specific scenarios, such as education products should be clearly oriented to teachers or students, e-commerce products to buyers or sellers, etc.; Moreover, educational and social products also have certain media attributes and need to be supervised by relevant departments.
From the perspective of data quality: the B side is higher than the C side. Since AI To B applications are mainly oriented to enterprise customers, they have higher requirements for the professionalism and accuracy of the output content, and the error tolerance rate is low, so the data used by the underlying model for training must also have high quality. AI To C is mainly for mass users and has certain social entertainment attributes, so the professionalism and requirements for output content are weak.
Written at the end: Evolve with AI
This is an era in which silicon-based intelligence comprehensively crushes our IQ. The AI revolution brought about by ChatGPT will profoundly change the direction of society, and now it is a chaotic chaotic situation, and everyone is trying to do something. But models are not something that ordinary people can do, so there are more products based on scene applications, and it also provides entrepreneurs and developers with many ideas for AI applications.
In addition, localization is another important topic, domestic manufacturers are obviously lagging behind in progress, but the application layer has taken the lead and has a lot of useful products, which will also be the most competitive and the most opportunities in the future. For ordinary people, the entire industry is still evolving, and there are opportunities in the future.
*Commercial soft implants are not included in this article