"In the future, all applications will be developed based on large models. Every industry should have its own large models, and the large models will be deeply integrated into the real economy."
Author Sihang Doudou
Editor | Pi Ye
Produced | Industrialist
"Large models are about to change the world." On May 26, Robin Li said at the Zhongguancun Forum.
Regarding the sentence at the beginning of the article, the head of Baidu, who has always adhered to the AI line, made a more complete statement, " We are now at a new starting point. This is a new era of artificial intelligence centered on large models. Large models Changing artificial intelligence, large models will change the world."
Large models, a concept that has exploded rapidly in the past six months, are becoming the core topic of communication among countless investors, media professionals, and market practitioners. While people are talking about it, it refreshes its capabilities and boundaries at a rapid pace.
And Baidu is the first practitioner. With regard to Wenxin large model, if it is said that people’s first focus on this first general-purpose large model in China lies in its release itself, then in more and more enterprise and industrial scenarios, it is increasingly breaking people’s attention to large models. cognition and definition.
According to a set of incomplete data statistics, in the past two months or so, Wenxin Yiyan has made rapid progress, and has carried out four iterations. The QPS query and reasoning response speed per second has increased by 10 times, which means that the reasoning cost has been reduced. One-tenth of the original. Up to now, 150,000 companies have applied for Wenxin Qianfan’s internal test, and more than 300 ecological partners have signed contracts with Baidu.
In fact, these changes from the product and market side are just confirming the views expressed by Robin Li on the forum. "In the future, all applications will be developed based on large models. Every industry should have its own large models, and the large models will be deeply integrated into the real economy."
At the moment when large models are becoming new productive forces, Robin Li and Baidu are being revalued by the market.
1. Big models are about to change the world
"Emergence", along with the explosion of large models, is becoming one of the labels of large models.
"The artificial intelligence in the past is that I want the machine to learn whatever skills I want to teach it. Those who have been taught may be able to, and those who have not been taught will not. After the emergence of intelligence in large models, skills that have not been taught before, it I will too." Robin Li said.
Objectively speaking, "emergence" corresponds to the way information appears. If we say that in the past few years, the first three industrial revolutions have completed more or less new arrangements and combinations of the world's inherent elements, such as tool-based production. The improvement of efficiency, such as the divergence of information-based transmission, etc., among them, itself corresponds to the more efficient transmission of information.
But "emergence" is not. Its greater value lies in information generation, that is, in the training line of deep learning, large models can realize the generation and creation of new information that has never been seen in the world. For people, in the future, in addition to information acquisition, how to obtain accurate information will become a new difficulty.
In Li Yanhong's words, "In 10 years, 50% of the world's jobs will be prompt word projects. Asking questions is more important than solving them. Our education should teach children to ask questions, not just solve them."
The changes don't stop there. If the label "emergence" corresponds to the unique value created by the large model itself for information circulation, then on the industry side, the large model is also reconstructing the inherent enterprise business and industrial business with a sufficiently stable and irreversible attitude.
Taking marketing and customer service as an example, as the front-end tentacles of enterprise business activities, marketing and customer service often need a sufficiently personalized and suitable contact and communication method, but before, due to the standardization of SaaS products, personalized equipment difficult to realize.
But large models can. Based on the information creation of the large model, after fine-tuning on the enterprise side, the enterprise can express and serve thousands of people in marketing and customer service, directly improve the efficiency of enterprise customer acquisition, and realize precise marketing.
In addition to business, products have also undergone essential changes under the catalysis of large models.
"Future applications will be realized by mobilizing native AI applications through natural language prompt words. This means that your salary level in the future will depend on how well your prompt words are written, not on your code. Is it good or not?" Li Yanhong said.
If it is said that before the research and development of the enterprise's products, or the work process, it was produced through market research, or the process of standard service software development, then after the large model, the enterprise's products, service processes, and personal ability training All links in which can be designed and optimized based on the prompt words with the support of AI.
Li Yanhong’s observation is, “For example, DoNotPay is an application that uses AI to help people file lawsuits and write legal documents. AI helps you get back the money you shouldn’t have paid. Jasper is an application that uses AI to help companies and individuals write marketing promotion documents. App. Speak is an app for learning foreign languages in Korea. The big model becomes a one-on-one teacher, providing personalized education for each child.”
In fact, this is exactly what Baidu and Robin Li have done themselves. "Baidu is also using AI native thinking to reconstruct all our products, services and work processes. For example, our Ruliu intelligent work platform allows every employee to have a work assistant with rich professional knowledge and real-time response. Through dialogue Comprehension ability to achieve intelligent summary of chat records.”
It can be said that, subtly, the big model is gradually building new C-side and B-side logic. For individuals, this change lies in the attitude and method of obtaining information, while for enterprises, it represents a new concept of work and products, as well as a completely different IT design from the past.
AI has really come to people.
2. Large models, giving birth to AI native applications
In this wave of trends, especially for enterprises, how should they obtain their own "AI power"? In other words, how to get a ticket for a large model ship? In fact, before answering this question, it may be possible to first examine the new industrial digital boundaries that have been created by the big model.
In the past, with the popularization of cloud computing, cloud-native technology has gradually emerged as a new architectural approach in the cloud era. It puts the process of application development, delivery, and deployment, as well as the data generated in it, in the cloud to achieve efficient, reliable, and flexible operation and maintenance management. Cloud native has become the first choice for digital transformation of many enterprises.
However, after the data is uploaded to the cloud, there are still many problems in the middle. For example, for many enterprises, the data is more visible and unavailable, that is, "going to the cloud" is more of an action and has little value to the actual business.
The large model brings a new way of solving problems - AI native.
That is, it can be understood that AI-native is different from the previous enterprises that need to develop software based on cloud-native development, delivery, and deployment. Instead, it is based on the generation of upper-layer application software natively on AI, and can even directly help enterprises make intelligent decisions.
"The rules of the cloud computing game have been completely changed. When customers choose cloud vendors, they mainly look at whether the performance model and framework are good, rather than traditional capabilities such as computing power and storage." Robin Li said.
This means that, among them, computing power, algorithms, and development models can all become "black boxes", and what enterprises ultimately accomplish is to pay for the effects. Corresponding to the service model of cloud vendors, no matter IaaS, PaaS or SaaS, enterprises in the future will examine the MaaS (Model as a Service) capabilities of service providers, and carry out native development based on the AI capabilities of the MaaS layer to meet business needs more intelligently .
Objectively speaking, this change in the underlying model is also changing the inherent digital transformation path of the enterprise.
Specifically, due to the late start of digitalization of Chinese enterprises and the low degree of standardization, domestic digital service providers often have insufficient service capabilities. For large customers, there is a lack of necessary industry mapping capabilities, and the ability to customize services is insufficient; while for small enterprises, it is difficult to enjoy personalized development without sufficient IT budget.
With the development of general large models, these needs will be gradually met. Take the Wenxin large model as an example. For example, the heterogeneous biological supercomputing platform jointly built by it and Baitu Biological Sciences supports ultra-large-scale GPU parallel computing to train multi-modal large models exceeding 100 billion levels. Tushengke conducts research and development of innovative drugs; helps Changan Automobile build an artificial intelligence infrastructure platform, and the two parties are also jointly developing a new artificial intelligence product based on the Wenxin Yiyan model to empower Changan Automobile's mass-produced models, etc. wait.
For small enterprises, based on large models in the future, enterprises can access general-purpose large models such as the Wenxin large model, and conduct text, image, audio, video, digital artificial intelligence, and even data processing based on a small amount of training or no training. This is a simple software. Generation, and self-individualized satisfaction of one's own digital needs.
"In the past, whether in the PC or mobile era, there were three layers of IT technology stacks, chip layer, operating system layer, and application layer. In the era of artificial intelligence, the IT technology stack is becoming four layers." Robin Li said .
"The bottom layer is still the chip layer, but the mainstream chip has changed from CPU to GPU. Baidu's layout at the chip layer is Kunlun Core; the chip is called the framework layer, which is the deep learning framework. Baidu's Paddle, Meta's PyTorch, Google's TensorFlow They are all at the framework layer. Above the framework is the model layer, while ChatGPT and Wenxinyiyan belong to the model layer. The topmost layer is the application layer, which is the native AI applications we mentioned earlier.”
It can be understood that the AI large model redefines cloud native, and cloud native is moving towards a more practical AI cloud native; while AI cloud native redefines the application layer and is heading towards AI native applications.
3. Baidu enters the competition table first
In fact, as the world's first major Internet company to launch a general-purpose model, Baidu has been under the spotlight of the market. That is to say, with the rapid development of large-scale models, people want to measure the development of large-scale models in China through Wenxin large-scale models and Baidu.
As the frontrunners, while opening the Wenxin model to each industry partner, Li Yanhong and Baidu have become more and more creative about their own path choices.
"There is a saying in the industry that the era of large models has come, and every product is worth redoing. But who will really do it again? Baidu will be the first company to redo all products, not integration, not succession Enter, it is redo, refactoring!"
For Baidu, this is a difficult but inevitable thing to do. That is, with the deepening of the large model at the enterprise and industry ends, it will grow into a new enterprise and industry bottom layer, and build a new digital logic for enterprise growth and industrial evolution. Based on its own accumulation of AI capabilities, Baidu is just becoming the first demonstrator and practitioner.
In fact, in China, Baidu is already on the AI table.
As early as 2013, Baidu has already started the layout of "All in AI". In order to build its own AI capabilities, it has invested more than 100 billion in 10 years, and now the Wenxin large model that leads the country is also as early as 2019 Wenxin large model version 1.0 was released. Now the kilocalorie parallel linear acceleration ratio of Wenxin large model AI base can reach more than 90%, the utilization rate of training resources exceeds 70%, and the iteration efficiency of model development is increased by 100%.
From the perspective of the technology stack, Baidu’s four-layer technical architecture of the large model is self-developed. Based on its Kunlun chip, Feishui deep learning platform, and Wenxin large model, it can achieve controllable data, controllable framework, and controllable models. control.
This is the prerequisite for Baidu to take the lead in the AI poker table, or in other words, the trend of large models, and it is also the confidence that Robin Li dared to say "reconstruct Baidu products".
It can also be said that behind the big model, Baidu is being revalued.
In the past many years, the market has more or less doubted and puzzled Baidu's heavy investment and direction in the field of AI. However, this kind of doubt and confusion has been broken one by one with the release of Wenxin Yiyan and the implementation of Wenxin's large-scale models in different industries.
It is also based on Baidu's accumulation on the AI side for more than 10 years in the past years that the speed and confidence of Robin Li and Baidu in today's large-scale model era have been formed. Reasons for choice and expectations. And this centralized release of AI capabilities is bound to bring new value to Baidu's TO B business, including Baidu Smart Cloud.
The value of this business has actually been transformed into a specific industrial landing value.
According to Baidu's 2023 Q1 financial report, Baidu's core operating profit increased by 45% year-on-year to 5.36 billion yuan. Among them, Baidu Smart Cloud achieved profitability for the first time.
Technology Baidu, "craftsman" Robin Li, is now ushering in the best era after the heavy snow on Changpo.