Introduction
A developer platform in the era of large models must not only provide native AI capabilities, but also have ecosystem creation and commercialization closed-loop capabilities.
The full text is 6115 words, and the estimated reading time is 16 minutes.
When people mention "spiritual realms", they often picture the scenic spots they have traveled to in their minds. However, in the context of the "intelligence emergence" of general large models, what will the "spiritual realm" platform look like that can provide opportunities for developers who are ready to explore the AI world here?
This question is difficult to answer. After all, operating a developer platform involves multiple dimensions such as product polishing, rule establishment, and ongoing operations. The beginning of a new era is bound to bring more unpredictable variables.
But at this time, we also see some pioneers trying to give their own answers. For example, Baidu, which was the first to launch "Wen Xin Yi Yan" in China, also quickly launched its own "Lingjing Matrix" developer platform (hereinafter referred to as "Lingjing") in September this year.
We learned that since the public beta launch of Spirit Realm Matrix in September, the overall number of applications received is about 27,000. Among them, the proportion of individual developers has been showing an upward trend and is now around 30%. What's different about Lingjing Matrix is that it is not only a platform that focuses on providing development capabilities, but also designs a set of traffic monetization channels for developers, hoping that developers can actually gain revenue through the platform.
This is a quite cutting-edge exploration mechanism, which makes people curious. What kind of thinking and exploration went into the launch and design of the spiritual realm matrix?
On the afternoon of November 15th, during the latest "Baidu Search Innovation Competition", we came to Baidu Science and Technology Park with various questions about "spiritual realm" and had a 2-hour discussion with Zhang Ruixing, the person in charge of Baidu Spiritual Realm Matrix Platform. access.
1. The birth of the spiritual realm: making large models successful is the key to the battle situation
"Should we build a low-cost platform for developing AI applications to empower developers, or should we help developers directly access large models and then empower various enterprises through large models? Which direction should we choose? We have experienced a lot of A long time of thinking." Faced with the question of why the spiritual realm was born, Zhang Ruixing responded frankly.
At the beginning of this year, in the new large model battlefield, most people chose to make applications, such as domestic Miaoya Camera, foreign Midjourney, etc. But in addition to rebuilding its own products across the board, Baidu is also thinking about how to help developers reduce costs.
This is a very important judgment. A large model is a powerful brain, but it requires more hands and feet. Therefore, when everyone focused on large models at the beginning of this year, Baidu has begun to think about how to provide more capability support for large models.
Developer platforms became an option. If you make an application like Midjourney, it is the individual who succeeds. As a developer platform, what we achieve is a large model and ecosystem. If more developers come here and develop more capabilities, the big models will benefit. Baidu’s Wenxin large model will further win the AI battle in the new era.
This judgment was also made through "chatting" with many customers who use large models.
"We also communicated with many developers and companies at that time and found that they still had some pain points in specific scenarios when using large models, and they needed products such as plug-ins to help solve them," Zhang Ruixing said.
Therefore, the positioning of Lingjing is to help the brain of Wenxin’s large model have more hands and feet, so that Wenxin’s service capabilities are stronger.
2. Throw away concepts, plug-ins and agents are both spiritual realms
In the large model industry, iteration is very fast, and many concepts have emerged on the market: such as Bot, plug-in, Agent, etc. However, entangled with these concepts is of little significance in the era of large models that are changing with each passing day.
"We need a mechanism that allows developers to connect to the base of large models at a low enough cost. Is this mechanism called a plug-in, an agent, some kind of API, or something else? Everything is just a concept. With the evolution of technology and the development of the market, they will eventually be gradually standardized and unified into a paradigm." Zhang Ruixing believes.
But another question is, since Agent is more like a future paradigm, are plug-ins unimportant? What is the relationship with the Agent?
In Zhang Ruixing's eyes, Agent may be the mainstream access method in the future, and it itself involves several major capabilities.
"The agent first needs to understand the natural language that describes the requirements. The second is to make effective plans for the user's actions and tasks based on understanding. At the same time, it needs long-term memory for user interaction information." Zhang Ruixing further explained , in this process, in order to meet user needs, the Agent actually needs to call a variety of tools.
Therefore, plug-ins may be the basis of Agent and the basis of a large number of tools that Agent needs to call.
For example, a travel itinerary planning agent allows users to quickly develop a personal assistant based on personal habits to plan travel itineraries.
However, the Agent itself needs to call a large number of tools. That is, when it needs to query the weather, call the flight or train timetable, flight information, and some personal preferences such as whether you like to sit by the window, it also needs to query the interesting attractions in the destination city, whether to travel with children and other personalized attributes. When using it, the Agent will call many plug-in tools to meet the user's personal needs in a combination.
From this perspective, various plug-ins are the base that supports the Agent.
"What plug-ins do now is a large number of tools. For example, the travel intelligent itinerary planning agent may directly help you book a ticket after you interact with it. In order to complete this action, it needs to adjust a ticket booking tool, that is, What the plug-in is doing now. The plug-in will become an important support for Agent in the future." Zhang Ruixing said.
He also made an interesting metaphor to describe the relationship between large models, Agents and plug-ins in the future: "The Agent may be a cerebellum, and the large model is a brain. In the future, each company's big and small brains may have similar capacities. But if I have 10 If you only have the same tool in your hand, then my ability to meet customer needs will definitely be stronger.”
3. Why does it have to be the spiritual realm?
Although many large model companies are currently launching developer platforms, to be a developer platform in the large model era requires not only technical support, but also an understanding of user needs, as well as the ability to build an ecosystem and commercialize closed-loop capabilities.
The birth of Lingjing began with Baidu's forward-looking research and judgment on the large-scale model pattern, and it is also inseparable from the accumulation of the entire technology, product, ecology, and commercialization system in the past.
The actual project establishment time for the Spiritual Realm Matrix Platform is July. From project establishment to online testing to external release in September, Lingjing only has about 2 months.
Zhang Ruixing recalled the intense pace of the team at that time, as every second counted: "In less than 30 days, the internal beta (invitation beta) version was released in mid-to-late August. In early September, we officially released the public beta version at the Alliance Summit."
The launch of the "Lingjing" developer platform at Lightspeed, in addition to the team's full efforts, also has a great relationship with Baidu's long-term accumulation in the past. This is something that many large-scale model companies in China do not possess.
First of all, Baidu has accumulated a strong developer ecosystem - whether it is the original webmaster or the later mini-programs, these millions of webmasters and 300,000 to 400,000 small developers are Baidu's natural advantages.
Many companies on the market want to build a developer ecosystem, but they have not accumulated experience at this level. If you build a platform without developers or a B-side ecosystem, unless the company has strong layout capabilities, even if the platform is built, the revenue will be very low.
Furthermore, compared to large model players with development ecosystems, Baidu also has rich “traffic scenarios”.
This also makes Lingjing's positioning in the market more differentiated - it is not just a pure development platform, but an integrated empowerment platform covering "development + operation + commercialization".
In contrast, many plug-in developers at home and abroad are faced with the dilemma of “no ecosystem and no traffic scenarios”. For example, in the OpenAI ecosystem, a plug-in developer in the financial field said that the plug-in was completed, but few people asked financial questions on ChatGPT, which resulted in few plug-in calls and very small profits. A developer platform without traffic distribution is likely to be ineffective.
Zhang Ruixing explained, "Why don't we just write the word 'development' in the positioning? Because Lingjing includes development, operation and commercialization, but in the early stage, development is definitely the most important."
He revealed to us that Baidu has partially opened up the distribution scenarios before, and expects to completely open up the two major distribution scenarios, Baidu Search and Wenxinyiyan, in the fourth quarter. “Like Maps, Wenku and other Baidu’s other major distribution scenarios, Lingjing is all in Access. Therefore, it will soon be seen that developers can not only develop on Lingjing, but also have an influx of traffic that is more suitable for the scene. It is expected that in Q4 this year or Q1 next year, Lingjing will lay out commercialization capabilities and allow some Plug-in developers monetize traffic in corresponding scenarios.”
So, if the opponent also has these two advantages, what is the support of the spiritual realm?
The answer lies in the capabilities of the large model itself. In China, Baidu's investment in large models is very long-term and huge, and the results are also catching up with international standards.
4. The devilish details of looking forward
The completeness and ease of use of functions can often determine the success or failure of a product in the early stages.
So, in terms of the initial development model, what are the differences between Lingjing and the ChatGPT plug-in? Through Zhang Ruixing's narration, we feel that the devil is often hidden in the details.
First, Lingjing adopts the latest visual development and orchestration model. Through drag and drop, it can eliminate intermediate platforms like Langchain to a certain extent and reduce the development cost of large model applications. It is understood that Baidu Lingjing Matrix actually adopted this relatively advanced method in July and August, even earlier than OpenAI.
Second, at the access level, the spiritual realm supports a variety of methods. Like ChatGPT, it supports more API access, but Lingjing supports more types of data access, such as local upload, database, API, etc., covering all capability-based access. In addition, application access is also a major feature of the spiritual realm. Some developers with strong capabilities, such as Ctrip, are willing to come to Lingjing to make a similar smart application. In the future, this smart application can also be connected to the large model. "We support access to all three modes, so the access methods are richer than Plugin." Zhang Ruixing said.
Third, in the top-level design of Lingjing, the distribution of plug-ins has been designed into an integrated and multi-terminal model. Moreover, this model will be fully implemented in Q4.
"The so-called one-in-one multi-terminal design is based on the characteristics of Baidu's traffic scenarios, and is designed to be compatible with various mobile APPs and PC browsers for web search." Zhang Ruixing said that with the design of one-in-one multi-terminal distribution capabilities, in the future, spiritual realm The plug-in can be natively implanted into the original App, even into its own mini-program. This is a good foundation and something that ChatGPT-Plugin cannot achieve.
With many differentiated features, Lingjing currently covers a relatively complete range of fields. In addition to law, emotion, office creation, product evaluation, etc., there are also many plug-ins that are being developed and launched. Nowadays, office creation plug-ins are mostly used in Lingjing, because this is still the most important scenario for large model applications.
We also noticed that the two previous Lingjing applications, Legal Bot and Travel Bot, have been well received by developers for a long time.
From the creation methods of these two types of Bots, we can also see the details of the operation of the spiritual matrix.
First, Baidu established special teams for these two industries.
Such as legal bots. Baidu's legal department took advantage of its busy schedule to provide many professional lawyers to correct the Bot and evaluate the accuracy of the answers verbatim. "It seems that we don't invest much in the product, but the entire team adds up, and the investment in each small direction is basically around a few dozen people, which includes strategy, evaluation teams, and related professional teams." Zhang Rising introduction.
The Travel Bot is not only the contribution of the Lingjing team, but also the result of the strong collaboration between the Ctrip team and Baidu. "It is equivalent to a very close team of brothers who have also made a lot of investment and collaborated with us to do this." Zhang Ruixing Replenish.
In addition, domain alignment is also a hot topic for large models.
Zhang Ruixing shared his experience in the process of training legal bots, "In the process of pre-processing and pre-training of large model training, repeated strategic adjustments are required. For example, if you answer the same question this way or that way, which one will matter? Which one is right or good requires positive feedback. In this process, Lingjing has a large number of evaluation teams and sample generation teams to assist in the training of large models."
There are two points here. First, in the process of the entire team going from not understanding to understanding, they actually need to communicate frequently with internal and external professionals. "We have talked about everything that can be discussed on the market, and the legal team will also provide training to the product team. The team has also introduced professionals, and the data annotation process has hired people with legal backgrounds." He told us.
Furthermore, cooperation with professional organizations can also significantly improve product effectiveness. We learned that the legal content on the Lingjing platform has also received support from the China Judicial Big Data Research Institute.
"When you enter the deep water area, when you enter this vertical field, the effectiveness of your strategies and models, and your understanding of the entire 'French Language' will reach a big level." Zhang Ruixing concluded. The feeling of being deeply involved in the vertical field.
5. The spiritual realm of the future
"AI can be used by everyone" is the concept that Lingjing adheres to. This is a long-term goal that requires continuous hard work to achieve.
We understand that in the next year, Lingjing will continue to focus on three key directions.
First of all, the Lingjing platform will continue to reduce costs. This is the core of Lingjing and also the trend of the industry. Developers in the AI era are not just practitioners with professional R&D capabilities in the traditional sense, but more and more ordinary people who are also eager to try it. To make this happen, reducing developers' costs will be the top priority.
The emergence of the Agent development paradigm has given this matter a turn for the better.
In the Agent development model, creating functions through code, development APIs, and prompts will bring great room for cost reduction.
"To what extent will the cost be reduced? It is possible that in the future the entire Lingjing platform will be an Agent. Users only need to type: 'I want to create something', or 'I want to plug in a database', and only need to interact through natural language This is a clear path for us." Zhang Ruixing said.
And this is also a new flywheel. This model will lower the development threshold, attract more developers, and allow people with development capabilities to develop better things.
Zhang Ruixing also detailed some plans for Lingjing to continue to reduce costs.
The most direct way is to compress the developer's development process and let the large model directly help the developer convert the process into prompts and prepare the corresponding Agent.
In addition, the spirit realm will also provide developers with many resources. "Lingjing will empower everyone with more computing power and more free tokens. In addition, Lingjing will also have some economic empowerment or support. For example, at the server level, software and hardware conditions, such as cloud Servers, etc., we can open them for free for many individual developers or enterprise developers." Zhang Ruixing added.
The second key approach involves closed-loop business. In other words, Lingjing will also empower developers with sufficient business capabilities . "We believe that this line is still the core advantage of Baidu's differentiation. We can really bring benefits to developers instead of just talking empty words." Zhang Ruixing emphasized, "The developer has made something good, and Lingjing hopes to help him launch it."
Among them, Lingjing will help developers truly obtain traffic and users on the one hand, and will also provide various commercialization empowerment at the same time. On the other hand, for developers who lack C-side demands, Lingjing will also create "D2C" and "D2D" models. "Other developers can buy the good things you develop or use them directly." Zhang Ruixing said that even A "D2P" business model may be launched, in which developers can make premium products for a fee for the platform.
Of course, in order to ensure user experience, Lingjing has strict requirements for the online review of plug-ins. The launch of a certain plug-in means that it must interact with users through products such as Wenxinyiyan and Baidu Search. In this case, we have higher requirements for its accuracy, triggering accuracy and answer quality. To this end, the Lingjing team has strict online evaluation standards to avoid giving users a bad experience.
Of course, the road to continuous polishing of the spiritual realm also presents various challenges. Most directly, it is the tuning of the model and Query, and finally the execution of the action.
For the development paradigm of Agent, the original logic of web page distribution is no longer applicable - any web page or website may be defined as a small Agent in the future, and future searches will become calls to various types of agents. Agent model.
For example, website Agent, App Agent, and Data Agent are all different types of Agents. So, how to intelligentize and automate this process to achieve the effect of accurately identifying, calling and distributing agents requires not only optimizing the product, but also forming a strategic flywheel and data flywheel integrating B and C in the process.
This is only possible if the large model has strong judgment and recognition capabilities, as well as the ability to interact with the Agent. This is also a core direction of Lingjing’s follow-up: creating a search AGI + a universal Agent.
This is obviously not an easy task, but for Spirit Realm Matrix, a bigger vision is still on the way.
In Baidu’s vision, Lingjing will also build a store similar to GPTs in the future, “but unlike the App Store in the mobile era, Lingjing can build a construction paradigm that is truly suitable for developers, thereby creating new things in similar fields. "In Zhang Ruixing's view, building such a platform is the only way to stimulate more imagination among creators and make the entire Agent ecosystem more prosperous.
——END——
Recommended reading
Hundreds of thousands of QPS, Baidu's stability guarantee practice for hot event search
Baidu search trillion-scale feature calculation system practice
Baidu search intelligent computing power control and allocation method
UBC SDK log level repetition rate optimization practice
Spring Boot 3.2.0 is officially released. The most serious service failure in Didi’s history. Is the culprit the underlying software or “reducing costs and increasing laughter”? Programmers tampered with ETC balances and embezzled more than 2.6 million yuan a year. Google employees criticized the big boss after leaving their jobs. They were deeply involved in the Flutter project and formulated HTML-related standards. Microsoft Copilot Web AI will be officially launched on December 1, supporting Chinese PHP 8.3 GA Firefox in 2023 Rust Web framework Rocket has become faster and released v0.5: supports asynchronous, SSE, WebSockets, etc. Loongson 3A6000 desktop processor is officially released, the light of domestic production! Broadcom announces successful acquisition of VMware