“More than two months and four iterations, the core scene reasoning performance has increased by 50 times.”
This is the report card that Baidu handed over for Wenxin Yiyan in the past 70 days. At the technical exchange meeting just two weeks ago, the performance of Wenxin Yiyan was only ten times higher.
It was also at that exchange meeting that Baidu demonstrated the fine-tuning process of the large model for the first time on the spot. The absolute hero behind Wen Xin’s words, “Wen Xin Qian Fan” began to quietly enter the public eye.
Previously, the public had always been more familiar with Baidu’s conversational AI product “Wenxin Yiyan”, and few people have seen the Wenxin Qianfan large-scale model platform that provides support for it.
And now, in good time, it’s finally starting to surface, which is good news for businesses keen to embrace big models.
1. Supporters behind Wenxin Yiyan
The iteration speed of Wenxin Yiyan is amazing. At the technical exchange meeting on May 23, Zhu Yong, vice president of Baidu Smart Cloud, said that since the internal test, through the continuous optimization of algorithms and models, Wenxin Yiyan’s reasoning The performance has been greatly improved by 50 times.
Just over two weeks ago, the reasoning performance of Wenxinyiyan was still at the level of 10 times improvement. In the past 20 days, based on the Wenxin Qianfan large-scale model platform, Baidu has further improved the reasoning performance in high-frequency scenarios through the end-to-end optimization of the four-layer architecture (chip-framework-model-application). 5 times. Previously, the reasoning cost of Wenxin Yiyan has been reduced to one-tenth of the original, and the improvement of performance also means that the related costs have been further reduced.
Many people may still be curious about the difference and relationship between Wenxin Qianfan and Wenxin Yiyan. In the most popular terms, Wenxin Yiyan is more like a To C product, providing users with the ability to communicate with large models, while Wenxin Qianfan is a platform for developing, training and applying large models.
More specifically, the services provided by Wenxin Qianfan can cover the entire life cycle of artificial intelligence research and development, including: data management, model training, model evaluation, predictive deployment and plug-in services, etc., and can help customers develop and deploy large-scale AI end-to-end. Model application.
At the technical exchange meeting on May 9, Baidu Smart Cloud showed the “Wenxin Qianfan Large-scale Model Platform” to the outside world, saying that it is the world’s first one-stop enterprise-level large-scale model platform.
In the exchange meeting on May 23, Xin Zhou, general manager of Baidu Smart Cloud AI and Big Data Platform, further elaborated on this definition. There are two sets of keywords in the aforementioned concept: one is “one-stop”, which means that Wenxin Qianfan will provide a full closed-loop process from model development, application, reasoning, to data return and development, covering the entire development of large models Life cycle; the second is “enterprise level”, that is, Wenxin Qianfan will provide many enterprise-related services, such as intelligent management and control, data security, account management and so on.
Over the past period of time, the iteration speed of Wenxin Yiyan has proved Wenxin Qianfan’s ability in training large models. We have seen that Baidu has become the first company in China to release a large language model, and we are also pleased to see the rocket-like upgrade speed of Wenxin Yiyan in the past few days.
Based on such changes, we can foresee that in the future, with the support of Wenxin Qianfan, the upgrade speed of Wenxin Yiyan may exceed our imagination.
At the same time, the rapid iteration of Wenxinyiyan has fully proved the capabilities of Wenxin Qianfan platform, and it has been firmly embraced by more enterprises that want to use large-scale model capabilities.
From the perspective of the enterprise, the four technical exchange meetings held by Wenxin Qianfan recently were full, and almost all the participants were corporate customers. It is reported that as of now, 150,000 companies have applied for the internal test of Wenxin Qianfan. More than 300 ecological partners have signed contracts with Baidu to explore more than 400 scenarios.
It is worth noting that Wenxin Qianfan not only supports Wenxin Yiyan’s large-scale model service, but also supports third-party large-scale models. In other words, customers can also deploy and train on Wenxin Qianfan using third-party large models.
From the perspective of enterprise demand, today’s enterprises embrace large-scale models, and the three focuses they pay most attention to are model effects, costs, and safety. That is, how much business improvement the large model can bring to your own company, whether you can afford it, and whether the data can be guaranteed safe.
Let’s look at the effect first. Wenxin Yiyan’s training results have already reflected from the side, Wenxin Qianfan’s ability as a large-scale model platform. Previously, the live demonstration on May 9 showed that after only about ten minutes of fine-tuning, the effect of the model has been significantly improved. It is understood that such a fine-tuning process only needs to label about 100 pieces of data each time, and the model can generate a certain generalization ability in similar problems. According to Baidu Smart Cloud, the improvement of Wenxin Qianfan’s development effect mainly relies on the ability of the AI base. The kilocalorie parallel linear acceleration ratio of the AI base can reach more than 90%, the utilization rate of training resources exceeds 70%, and the iteration efficiency of model development Increased by 100%. Feedback from customers also proves this point. The end-to-end optimization capability provided by the AI base greatly improves the model iteration speed, which impresses Jinshan Office, which has been established for 35 years. implementation.
As far as the cost side is concerned, Wang Yanpeng, Baidu’s outstanding system architect, said that the main focus of enterprises is on development costs, that is, computing power. But in fact, it may not account for a high proportion of the total cost, and this cost can be reduced through many methods, such as end-to-end optimization of the four-tier architecture. Algorithms, frameworks, systems, and chips are integrated. Zhu Yong, vice president of Baidu Smart Cloud, said that in addition to the consumption of the entire underlying resources such as computing power, the development of the entire intelligent application requires a lot of labor costs. The emergence of large models makes the threshold of these tasks low enough. It’s a good choice.
In terms of security, Xin Zhou said frankly that the public cloud requires a very high security mechanism, with complete command monitoring, log monitoring, auditing, and resource isolation mechanisms to ensure data security. In this way, the data security of the user’s private domain is guaranteed to a certain extent. In addition, customers can also choose the way of private deployment. At the same time, in order to protect user data privacy, Baidu will not use user data for model training and iterative optimization. And when policy requirements are met, these numbers will be deleted.
In the communication after the meeting, Xin Zhou further explained: “We need to make every angle of this function good enough, without making it too complicated. In fact, it is more flexible, so that our customers and users Based on sufficient flexibility, you can develop applications for your own scenarios according to your own business needs, which requires careful design from a product perspective.”
2. From a large model to a thousand industries
The landing scene of the large model is far more than what Wen Xin said. At a more specific industry level, the application scenarios of large models have more room for imagination.
In the past quarter, Baidu Smart Cloud achieved profitability for the first time. Zhu Yong, vice president of Baidu Smart Cloud, told Leifeng.com that Baidu Cloud’s profitability lies in the continuous improvement of product standardization on the one hand, which is product-driven; End-to-end optimization keeps costs down.
Of course, Zhu Yong also said, “After the big model appeared, we also saw the enthusiasm of the industry, customers, and partners, and they are actively jointly testing and developing. The release cadence is very helpful and may generate additional revenue streams in the future.”
The six major intelligent products of Baidu Smart Cloud are being upgraded based on Wenxin Yiyan, and will be launched after the safety assessment is completed. Including finance, government affairs, customer service, writing and other industries and fields.
Taking intelligent creation as an example, the platform will fully empower content producers to create efficiently, and high-quality content will be in place in one step. From topic selection planning, text creation to picture and video production, it covers all levels.
In terms of topic selection, the intelligent creation platform will provide a series of functions such as news clues, hotspot discovery, and event context; the auxiliary creation based on large models will be upgraded to AI automatic creation, and the creation time of a single article will be reduced from hours to minutes. The whole process of article writing is automated and interactive content revision is supported; 16 types of scene pictures are provided, and functions such as picture editing and picture generation are provided; one-stop generation and processing of pictures, texts and videos is realized.
According to Liu Qian, general manager of AI application product department of Baidu Smart Cloud, after the upgrade of smart customer service response assistance, conversation summary, smart work order, customer service knowledge base and other functions based on the large model, knowledge production efficiency increased by 9 times, and multiple rounds of dialogue Construction costs are reduced by 65%, enabling a smarter and more human-like experience for end users.
As another example, in the government affairs industry, Baidu created Yiwang Office | Government Service Assistant, Yiwentong | Community Consulting Service Assistant, Yiwentong | Community Consulting Service Assistant, Yiwangtong | A series of products such as intelligent analysis assistants help government agencies achieve double growth in digital intelligence and government affairs experience.
It can be seen that the large model is no longer an unreachable cutting-edge technology, and its application scenarios are gradually penetrating into the production chain of enterprises and people’s daily life.
In addition to a series of intelligent applications, it is expected that more and more products will be upgraded based on large models in the future. The first to benefit from this are the seven enterprise-level products such as Baidu Netdisk. It is reported that the Baidu Netdisk Enterprise Edition based on the large model will provide enterprises with a digital intelligent management platform. The upgraded Baidu Netdisk can help enterprise version users to summarize, translate, ask and answer the contents of files through conversational interaction. According to Zhu Yong, vice president of Baidu Smart Cloud, Baidu will integrate the ability of Wenxinyiyan in all businesses, and has made very detailed internal planning.
3. Where does the enterprise go from here?
Previously, in the article “From the first profit of the smart cloud, see how Baidu is running in the era of large models”, Leifeng.com(Public number: Leifeng.com)It has been mentioned that with the arrival of large-scale models, the customers of Baidu Smart Cloud have exceeded expectations, and many organizations that were reluctant to use it in the past have begun to actively communicate with Baidu. In other words, more and more businesses are choosing to embrace big models.
At the press conference on May 9, Baidu engineers demonstrated the fine-tuning process of the large model for the first time through “Wen Xin Qian Sail”. At that time, Xin Zhou also said, “Fast and convergent is a very important indicator in the application training of large models. If it is fast but not convergent, the training of large models is useless. In terms of multi-machine and multi-card training performance, Wenxin Qianfan has It can reach a state of convergence faster, and ranks first in the world in the MLPerf list of the global authoritative AI benchmark evaluation.” In Xin Zhou’s view, if an enterprise wants to make good use of a large model, only the large model is not enough, and a mature and complete large model is also needed. Model production platform.
In fact, it is true, the positive communication and hugs from customers say it all. In addition to Kingsoft Office mentioned above, Kingdee, iSoftStone and many other companies have already negotiated or are in the process of signing contracts with Wenxin Qianfan.
On May 18, at the Baidu Smart Cloud Partner Conference, Kingdee signed a contract with Wenxin Qianfan on the spot and officially became a partner of Wenxin Qianfan. Han Geying, Assistant President of Kingdee China, said in her speech that Kingdee has now connected to the Baidu Wenxin Yiyan model, and has been exploring and verifying in various ways and has produced good results. In the future, Kingdee will combine large models such as Baidu Wenxin Yiyan, use Kingdee’s industry knowledge and proprietary data to build industry models for specific tasks, continuously upgrade Kingdee’s original NLP, OCR and other applications, and explore more Many business scenarios.
Zhu Yong said frankly that Wenxin Qianfan large-scale model platform will provide the best environment for enterprises to develop and apply large-scale models, and it is the best path for customers to embrace AI. We hope that Wenxin Qianfan large-scale model platform can enable our customers and partners to use large-scale models in the simplest way and make good use of large-scale models.
As mentioned above, Wenxin Qianfan not only provides large-scale model services including Wenxin Yiyan, but also provides customers with a complete tool chain and development and training environment, which can fully meet customer needs.
In the future, Wenxin Qianfan will mainly provide two services:
First, with Wenxin Yiyan as the core, it provides large-scale model services to help customers transform products and production processes. Zhu Yong said that this is a bit of a “gold panning”, the real gold digging, empowering customers by providing model services.
Second, as a large-scale model production platform, enterprises can develop their own exclusive large-scale models based on any open-source or closed-source large-scale models on Wenxin Qianfan. Zhu Yong explained, “From Baidu’s point of view, as the leading AI company in the industry, Baidu still hopes to be more comprehensive and provide customers with various services. As long as it is related to large models, it is ours.” vegetable’.”
In addition, according to different customer needs, Wenxin Qianfan will also provide a variety of service models: at the public cloud level, Wenxin Qianfan provides services such as reasoning, fine-tuning, and hosting; in terms of private deployment, Wenxin Qianfan uses software authorization , software and hardware integration, and rental services are delivered to customers.
Although Wenxin Qianfan provides customers with a fully closed-loop process from development, training, fine-tuning, and data return of large models, covering the entire life cycle, not all companies are suitable for developing and training basic models from scratch.
Zhu Yong said frankly that in fact, the cost of training a basic model is very high. If you really want to make a large model with a scale of 100 billion, you need a computing power of more than 10,000 cards in a single cluster. Utilizing cluster resources, this is often not something that many companies can do.
In his view, in the future, there may be only a few basic models (large underlying models), but based on this, there will be many large models of different professions and industry types. These models in different fields will support the very prosperous upper-level field applications in the future. .
According to Xin Zhou, from the perspective of industry, the fields where large models are more popular are mainly industries with high penetration rate of informatization and technology, such as finance, energy, and pan-Internet industries.
Postscript: The future of generative AI can be expected
In the past two months, large models have blossomed everywhere. The focus of enterprises has shifted from the product function of the large model to how to make good use of the large model at a faster speed, complete the innovation and reengineering of the industrial chain, and realize overtaking in curves.
According to Baidu Smart Cloud, 150,000 enterprises have applied for internal testing of Wenxin Yiyan, among which more than 300 ecological partners have achieved test results in more than 400 specific scenarios, including office efficiency improvement, knowledge management, intelligent customer service, and intelligent marketing. and other representative scenes.
As the first company in China to release large-scale model products, Baidu’s exploration in the field of large-scale models has been on the way: from the release of Wenxin Yiyan, to the display of a series of application scenarios, and now Wenxin Qianfan has entered the internal testing stage, for Enterprises provide various tool chains and complete environments for developing AI applications, trying to cover the entire life cycle of artificial intelligence research and development.
As Xin Zhou, head of Wenxin Qianfan, said: “Generative AI is a technological change that is hard to see in a century. Although we are in the initial stage and the technology is still developing, I believe that its future will definitely bring us many applications of our own. Innovation and efficiency improvement. We have the most people, enterprises, and scenarios, and Baidu is working hard on the road of generative AI large models.”
Leifeng.com original articles, unauthorized reprinting is prohibited. See Reprint Instructions for details.