Compared with the numbers in the financial report, the industry and outside the industry are more concerned about what Baidu AI has done and whether it can meet expectations or even exceed expectations.
On the evening of May 16, Baidu released its Q1 financial report for 2024, with a total revenue of 31.5 billion yuan in the first quarter. The net profit attributable to Baidu (Non-GAAP) was 7.011 billion yuan, a year-on-year increase of 22%, and the net profit attributable to Baidu's core (Non-GAAP) was 6.628 billion yuan, a year-on-year increase of 26%, both exceeding market expectations.
During the earnings call, Baidu founder Robin Li emphasized that Baidu will transform from Internet Centric Business to AI First Business. AI is undergoing a comprehensive restructuring of Baidu, which is not only a simple launch of a product or model, but a complete change in product logic.
如何理解从Internet Centric Business到AI First Business？
If an Internet company emphasizes AI investment, but the core pillar is only Internet business, it is still Internet Centric Business. Baidu is confident that it will shout AI First Business, which means that Baidu's AI reform has entered the deep water area.
The article "Crossing the S Curve" writes that the so-called evergreen foundation is not to stretch it very long and steep in a curve, but to be able to cross the second curve again and again, and to open the second curve as soon as possible when the enterprise does not have a stall point, which may be the biggest secret for enterprises to pursue the success of the S curve.
Good companies have their own S-shaped growth curves, and most of them have suffered losses without a second growth curve, such as Microsoft, the overlord of the PC Internet era, in the mobile Internet era, the strategic shift to "cloud first, mobile first" and then smoothly transformed.
Baidu has long been looking for its second growth curve, or even the third growth curve, but between the ups and downs of industry and technology, the strategy to the landing is not always smooth, but this time the arrival of the big model, Baidu is a little different.
The search for Day1 in the era of large models
Most of the products in the Internet era are divided into three steps, the first is to build high-frequency use scenarios and acquire users through the free model, the second step is to precipitate business scenarios to improve the retention rate, and the third step is to complete the business closed loop to make sustainable profits.
In recent years, Baidu has mainly balanced user experience and commercialization in the search business, but can you think of a Baidu search that has no commercialization at all and only considers user experience?
This means that Baidu has brought search back to Day1 in the era of large models.
Titanium Media App observed that Baidu is gradually improving AI-generated search results, and adding an "interactive" column, completely changing to a question-and-answer search, all kinds of actions show that the speed of AI reconstruction search is accelerating, and it is also more free. Currently, more and more search results are generated by AI in more and more diverse formats, such as text, images, third-party links, points of interest, etc.
After the reconstruction of the Wenxin model, Baidu's new search has three characteristics: extreme satisfaction, recommendation stimulation and multiple rounds of interaction. At present, the proportion of Q&A requirements existing in Baidu Search every day has reached 70%, and the average daily new Q&A demand for Baidu Search exceeds 50 million.
Internet-centric search assesses revenue and profits, while AI-first search considers user experience and scale.
The financial report also mentioned that in March 2024, the monthly active users of Baidu App will reach 676 million, a year-on-year increase of 3%, and the quarterly online marketing revenue will be 17 billion yuan. Baidu continues to promote AI to reconstruct search, and the agents built based on Wenxin's intelligent twins platform AgentBuilder can be distributed in Baidu search and other ecosystems to enrich the content ecosystem and improve customer efficiency.
In the earnings call, Robin Li gave the latest judgment that search is the most likely application to become a KillerApp in the AI era, and AI can generate more accurate and direct results, so that users can be satisfied in one step - this is the most obvious.
At present, Baidu11% of the search results are generated by AI, which is just the beginning, search is Baidu's basic disk, can Baidu really gain a foothold in the AI era, how to combine AI with search, Baidu thinks clearly enough, the most difficult part may be to return to Day1's entrepreneurial mentality, and for the path of new search after that, there may be no company more familiar than Baidu.
Digging down the new search, you can also see two trends, one is to distribute agents on Baidu search, which may be the "applet" in the AI era, which can improve the user experience and improve the efficiency of merchants in the short term. Second, there is no advertising generated by AI, which will affect advertising revenue in the short term, but in the long run, it can make up for Baidu's content ecology and increase user stickiness.
Previously, Baidu has launched the agent development tool AgentBuilder, allowing everyone and every organization to become an agent developer. With AgentBuilder's "zero-code mode", novices can create an agent in a few sentences using natural language. Now, Baidu search, Xiaodu, map, post bar, car machine and other Baidu ecological products have been connected to the relevant capabilities of the agent, to solve the worries of traffic distribution for developers, and to obtain real benefits.
For example, when a user searches for "Singapore Tourism", they can interact with the Singapore Tourism Board agent. Ask questions such as "When are the least number of people going to Singapore", "What should I pay attention to when I travel to Singapore", "Top 3 hotels in Singapore", etc., and the intelligence will synthesize information from multiple sources to generate an answer. In addition, users can also book tickets directly to Universal Studios Singapore, a one-stop solution to their needs, which greatly saves time.
Up to now, more than 50,000 developers and tens of thousands of enterprises have settled in the Wenxin intelligent twin platform, and the intelligent twin is distributed millions of times per day.
AI prices and thresholds have been lowered, and intelligent cloud revenues have been raised
Intelligent cloud is Baidu's second curve, Robin Li revealed on the earnings call that Baidu's intelligent cloud revenue reached 4.7 billion yuan in the quarter, a year-on-year increase of 12%, and continued to achieve profitability (Non-GAAP). At the same time, generative AI contributed 6.9% of revenue, and the proportion is gradually expanding.
Wenxin model is becoming an important source of growth for the cloud, as Robin Li said in the conference call, "model reasoning is an important long-term opportunity", the current Wenxin model daily average processing of Token text reached 250 billion, the daily call volume of 200 million times, in the previously disclosed China Samsung, Honor cooperation on the basis of Xiaomi, vivo, OPPO, Lenovo and Weilai have also become Wenxin model customers, through smart devices, further drive the growth of model calls.
From the perspective of strategic layout, the direction of AI in the field of intelligent cloud is becoming clearer and more clear, and it is implemented in place, that is, the price and threshold of AI applications are lowered, industry customers are aggregated, and the ecological cycle of cloud and AI is truly running.
In terms of price, Baidu has done two things, one is to reduce the cost of Wenxin large model flagship version to 1% of the original, and the other is to launch 5 lightweight/scene-specific models, combined with flagship models such as 3.5/4.0, to form the use of MoE (large and small model collaboration).
Flagship models like Wenxin Model 4.0 have strong performance, but the call price is very high, and developing all applications based on 4.0 is also a waste of resources, so Baidu provides a rich variety of choices, including the flagship version of ERNIE3.5, ERNIE4.0, and the lightweight version of ERNIE Speed, Lite, Tiny, etc., for customers to choose in terms of performance and cost.
In terms of lowering the barrier to entry, Baidu provides a variety of application tools required for large-scale model development, including agent development tool AgentBuilder, AI native application development tool AppBuilder, and model customization tool ModelBuilder of various sizes.
The three major tools are Baidu's ability to transform AI into advanced productivity, such as Xiaodu's use of ModelBuilder, which mixes these five models with Wenxin3.5/4.0The flagship version, ordinary Q&A calls small models, and math problems are only used4.0, and finally the response speed has been increased by 2 times, and the cost has been reduced by 99%.
In the first step, Baidu laid down the price of the large model to reduce the cost of enterprises and developers to call the model. In the second step, Baidu has lowered the development threshold, allowing the public to use natural language to develop intelligent twins and AI-native applications, and the income of Baidu Intelligent Cloud will naturally be guaranteed.
Relevant data statistics show that from January to April 2024, nearly 200 domestic large-scale model-related projects have been announced, and the number of winning bids has far exceeded the total number in 2023; In terms of the amount of winning bids, the disclosed amount of large-scale model winning projects has reached about 77% of the whole year of 2023.
Among them, Baidu won three firsts among mainstream large-scale model manufacturers: the largest number of winning bids (7), covering the most industries (6), and the highest total number of winning bids (more than 56 million). The corresponding project procurement units include the All-China Federation of Trade Unions, China Southern Power Grid, China Unicom, China Tower Group and other leading customers in the industry, which also reflects the stronger growth trend of large-scale model technology compared with traditional innovative technologies driven by policies and markets.
The demand for AI will undoubtedly increase the use of intelligent cloud, and the intelligent cloud system can provide more possibilities for AI implementation, and the value of AI and cloud can be multiplied.
Autonomous driving, on the eve of scale effect
After ten years of autonomous driving, Baidu is ushering in a new dawn of autonomous driving. At the Apollo Day held recently, Baidu Apollo released the Apollo ADFM (Autonomous Driving Foundation Model), a large model that supports L4 autonomous driving, and at the same time launched a new sixth-generation unmanned car equipped with Baidu's sixth-generation intelligent system solution and priced at only 200,000 yuan.
Autonomous driving is Baidu's third-level curve, and now the technology and industrial environment are more mature, especially large-scale model-driven autonomous driving, which is expected to achieve a technological leap. Baidu Apollo ADFM reconstructs autonomous driving based on large model technology, which can take into account the safety and generalization of technology, achieve safety more than 10 times higher than that of human drivers, and achieve city-level global complex scene coverage.
It is also relying on the application practice of the autonomous driving model that Baidu Radish Express has overcome the complex road scene in Wuhan, realized the coverage of the whole city of Wuhan and the whole time and space scene, and provided convenient unmanned travel services for nearly half of the Wuhan citizens. At the same time, in the field of L2+ intelligent driving, ANP3, a pure visual urban pilot assisted driving product, will also fully apply the autonomous driving model Apollo ADFM, which will be upgraded to ASD (Apollo Self-Driving), which will be mass-produced and launched in all models of Jiyue, so that intelligent driving can be opened all over the country and everywhere there is a Baidu map.
The industry is paying attention to the news of Tesla's FSD and Robotaxi entering China, Tesla FSD and Baidu Maps are partners, Baidu Maps real lane-level navigation is about to make its global debut in Tesla, but Baidu Apollo and Robotaxi are competing products, the technology that supports Tesla's unmanned driving is FSD, and the core technology of Baidu Carrot Express is ASD, Baidu has invested in unmanned driving for 10 years, and is not too worried about "wolf".
Baidu intelligent driving said that for Tesla FSD and Robotaxi to enter China, Baidu is particularly looking forward to this, and it is very welcome, but at the same time, it is not easy to do Robotaxi in China's complex road scenarios, such as in Wuhan, Tesla wants to achieve the current level of Baidu's fully unmanned operation, it still takes a long time, less than 3 years, more than 5 years, if you don't believe it, you can come to Wuhan to run first.
Radish Express is doing two things, one is to make Wuhan a real benchmark, double the operating area, achieve 7X24 hours of operation, and cover 7 million people in Wuhan, so that radish in Wuhan from "experimental" to "practical", and truly become a part of Wuhan citizens' travel.
The second is to actively promote technology, launch sixth-generation vehicles, release autonomous driving models, further improve safety and risk prevention capabilities, and at the same time reduce costs, so that the business model can truly run through.
Combined with Baidu Radish Run's actions in Wuhan, perhaps the large-scale application of autonomous driving is not far away.
In a recent internal speech at Baidu, Robin Li said, "We are at the forefront of large models, and we need to brave the no-man's land, and we need to take risks that no one has taken before." ”
When Robin Li said AI First Business, what Baidu is doing is to bring itself back to the first day of entrepreneurship, to do what no one has ever done, from search to cloud to autonomous driving, and use large models to reshape its three-level growth curve. In the end, Baidu will be able to succeed again.
(This article was first published in Titanium Media App, by |.) Zhang Shuai)