The technological world we take for granted is being changed or even reconstructed by large models at a speed visible to the naked eye. Technology companies that don't want to miss out on opportunities all have the dream of building AI heavyweights, and various large-scale models come in one after another.
The release of the large-scale model is just the beginning. Like a rocket launch, ascension is the first step. Whether it can successfully rush to space and enter the industrial orbit is a road full of uncertainties.
Some companies spend a lot of money to develop large models, but the practical effect is not good. After the release, they are silent and uninterested, which increases the burden; Continue to promote the iterative upgrade of large models; some lose their way and cannot find the exact angle for large models to enter the industry...
The big language model that successfully rushed to the sky and the industrial track, Wenxin said, is the one that leads.
Data is the most powerful evidence. On May 16, Baidu released its unaudited financial report for the first quarter of 2023. Both revenue and profit exceeded market expectations, of which revenue was 31.1 billion yuan, a year-on-year increase of 10%, and net profit (non-GAAP) was 57. billion, a substantial increase of 48% year-on-year.
Baidu attributed this solid performance answer sheet to the revolutionary potential brought by generative artificial intelligence and large language models to various industries, as well as Baidu's timely layout of large models, Wenxin started internal testing, and plans to gradually introduce Wenxin Xinyiyan is integrated into all businesses of Baidu, and a new ecology is built around Wenxinyiyan in the new era to achieve long-term and sustainable growth.
With the help of Wenxin Yiyan's "AI rocket", Baidu's performance is rushing to the top.
We can disassemble the "power device" of Wenxin Yiyan to understand the driving force and industrial value of the large model.
Flying Paddle + Wenxin Yiyan, full speed ahead
Regarding Baidu's performance, some brokerages gave such an evaluation- Baidu is a first choice with multiple catalysts, such as generative artificial intelligence, cloud and ADAS, which can promote multiple expansions.
Multiple motivations, this feature is also reflected in the Baidu big language model Wenxin Yiyan.
For a large model to fly to the industrial orbit, it is not enough to have a front-end "spacecraft". Multiple powers such as computing power, framework, and ecology are also needed to provide sufficient technical momentum to support the continuous acceleration of the entire journey.
Take Wenxin as an example. After the release, industry partners generally gave good feedback. One of the main reasons is that it can iteratively strengthen model capabilities, improve model reasoning efficiency, and optimize computing power utilization in a timely manner based on user feedback, thereby supporting large-scale, High-growth, high-demand industrial demand.
Wenxinyiyan's full-speed advancement is inseparable from the joint optimization with Baidu's self-developed industrial-level deep learning framework Flying Paddle.
The collaboration of Wenxin Yiyan + Flying Paddle injects multiple powers into the large model:
First, the continuous improvement of reasoning ability, performance improvement of nearly 10 times.
The performance of large models on common tasks needs to be strengthened and improved in a timely manner based on user feedback.
Wenxin Yiyan and the large model behind it, supported by flying paddles, combined with Yiyan's model structure characteristics and quantification technology, completed the optimized version reserve of the inference engine. Since the release, it has maintained an average rate of completing one iteration per week . Four iterations have been completed. Compared with the first version of large model inference service, the QPS of a single machine has increased by nearly 10 times. This also means that the model reasoning cost of Wenxin Yiyan has been greatly reduced, and it can provide services for 10 times the number of users. The large model is expected to "fly into the homes of ordinary people" and land in various industries more inclusively.
To put it simply, Wenxinyiyan learns faster and becomes smarter.
In addition, the number of users of large language models has soared, and computing power has become a bottleneck that cannot be ignored, which has further pushed up the cost of computing power. Spending money to buy computing power is a "bottomless pit" that requires long-term high investment. Improving the utilization rate of model computing power is the way to "fix the root cause". Based on the optimization of the flying paddle distributed parallel strategy and the adjustment of the training accuracy strategy, the computing power utilization rate of the Baidu model will be greatly improved.
Flying Paddle supports the implementation of the whole process of AI model from training to reasoning. The collaboration of AI large model + deep learning framework is the key driving force for Wenxinyiyan to soar into the sky and accelerate to the industrial track.
Multi-stage power plant for large models
In 1909, Robert Goddard read science fiction novels such as "The First Men on the Moon", and began to conduct extensive theoretical research on rocket dynamics, and proposed a blueprint for a multi-stage rocket. According to the principle of conservation of momentum, when a rocket enters space orbit, it needs multi-stage propellers to continuously supply fuel at each stage of the voyage in order to break through the earth's gravitational circle.
In order to break through the limitations of technology landing and move towards the industrial track, the large model also needs a multi-stage power device that advances layer by layer.
Baidu's full-stack AI technology constitutes such a "multi-level power device". In the era of large-scale models, Wenxin Yiyan can show differentiated capabilities:
One is rapid growth.
As mentioned earlier, the joint optimization, mutual promotion, and mutual achievement of the Wenxin Yiyan at the model layer and the flying paddle at the framework layer are the most fundamental reasons for the improvement of the efficiency and effect of large models. Flying paddle integrates the core framework of deep learning, basic model library, end-to-end development kit, and rich tool components. Baidu's full-stack self-developed products are distributed in the four-layer architecture of Baidu's artificial intelligence. They are more compatible with each other and can be efficiently Collaboration to achieve end-to-end optimization, the speed of model iterative development is naturally faster.
Large model service providers without self-developed deep learning frameworks either start from scratch or redevelop based on deep learning frameworks such as TensorFlow and PyTorch. It is difficult to fully meet the training needs of large models.
Second, it is comprehensive and reliable.
The large model is a well-deserved "AI heavyweight", and the game around the large model may become a headache that plagues a large number of industry users in the near future.
The large-scale model of question-and-answer with countless Chinese people every day is built on the deep learning framework of other countries, and the risk factor can be imagined. Therefore, for the self-reliance, safety and reliability of large models, the core technology must be mastered by Chinese corporate audiences.
Wenxin series of large models are built on Baidu's long-term deployment of AI infrastructure and underlying self-developed technology, which can eliminate the worries of applying large models in thousands of industries and accelerate the transformation and upgrading of industrial intelligence.
The third is sustainable development.
As a new technology and product, the large model has no mature path to learn from. It requires long-term investment, exploration of no-man's land, and trial and error iterations. The process of exploration also means a lot of innovation investment. Falling behind, keeping up with industry trends, and supplementing the infrastructure and supporting technologies required by large models will bring huge R&D workload and cost pressure.
Baidu's artificial intelligence four-layer framework "chip layer-framework layer-model layer-application layer" has a complete layout, leading technology, and application implementation, which has built a sustainable competitiveness for Wenxin Yiyan.
For example, it is still unknown in many industries how large-scale models will be implemented in business scenarios. Thanks to the industrial practice and accumulation of Baidu's AI ecology, large-scale models in the Wenxin industry have been used in electric power, gas, finance, aerospace, media, cities, Applications have been launched in fields such as film and television, manufacturing, and social sciences.
IDC released the "2022 White Paper on the Development of China's Large-scale Models", which shows that Baidu Wenxin's large-scale models are in the first echelon of the industry, with comprehensive leading product capabilities, application capabilities, and ecological capabilities.
Just as the breakthrough of OpenAI was not accomplished overnight, Wenxin's "flying into the sky" did not rely on the explosion of a single technology, but the long-term accumulation of Baidu AI, which has no shortcomings, and finally allowed Wenxin's large model to break through the limit. Fly to the sky.
Baidu AI, jumping towards infinite possibilities,
The final stage of the rocket's voyage is getting the spacecraft safely into planetary-synchronous orbit. The end point of the large model is to be in line with the industry, and Baidu has already reached this stage.
The wave of large models will continue. How to embed large models into the industry and combine them with industrial applications? I am afraid that it will become an important topic, or challenge, of domestic large-scale models in the next few years.
This is why we pay attention to the meaning of Wen Xin Yi Yan.
A brokerage report shows that since the announcement of ERNIE Bot, enterprise customers have a strong demand for cloud-based ERNIE Bot services, and many of them are new cloud customers. In addition, ERNIE Bot has obtained a large number of high-quality sales leads since its launch, many of which have been transformed into joint development projects exploring AIGC applications.
The direct driving effect of AI infrastructure and SaaS services on cloud services is also clearly reflected in Baidu's financial report. Baidu Smart Cloud's revenue in the first quarter of 2023 will be 4.2 billion yuan, an increase of 8% year-on-year, and it will be profitable in this quarter. Achieving profitability in the highly competitive cloud market is a difficult problem faced by many cloud giants. Baidu Smart Cloud has found its own rhythm by relying on the core advantages of AI. With Wenxinyiyan's continuous efforts, Baidu's performance in the cloud market has more possibilities.
These solid industry transcripts are proof of the integration of large models with the industry and the successful transformation of Baidu's AI technology. The success path of dismantling Baidu AI may be far more important than the financial report itself for the subsequent impact of the big model and the industry.
Wenxin completes the "last meter" connection with the industrial track, and we can see a path: first, the self-certification and other certification of technical capabilities, and then the platform-based power supply and complete tools. The combination and landing of Baidu Heart and Baidu business and industry applications.
Wenxin Yiyan and Flying Paddle co-evolve to form a new AI infrastructure. This idea also paved the way for the later innovation of large models.
Baidu AI, which dares to be the first in the world, is also driving out of the Chinese speed and power we expect in the roar of the "flying paddle + Wenxin" power engine, and is heading towards the infinite future of industrial intelligence.
