news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

after more than a year of rolling, is the big model really popular?

there seems to be an endless debate about big models: should we pursue "large parameters" or consider "scaling law"? which framework is better, "dense" or "sparse"? which is the real future of big models, "language" or "multimodality"? players have clear-cut positions and opinions, but this enthusiasm seems to have never been transmitted to the more practical market.

among all the large companies, baidu is a rare one in the field of large-scale model engineering. according to baidu's second quarter financial report for 2024 released today, the total quarterly revenue was 33.9 billion yuan, baidu's core revenue was 26.7 billion yuan, and baidu's core operating profit was 5.6 billion yuan, a year-on-year increase of 23%.

the daily call volume of wenxin big model exceeded 600 million, and ai drove cloud revenue to 5.1 billion yuan, a year-on-year increase of 14%, which is a gratifying growth.

what did baidu do right to achieve such great results in bringing the "floating" ai back to earth?

the call volume continues to grow, relying on price cuts and strength

let’s first talk about the most eye-catching data in baidu’s q2 financial report - the sharp increase in the number of calls to the wenxin model.

the daily call volume exceeds 600 million, and the average number of tokens processed per day exceeds 1 trillion. the surge in call volume always reminds people of the "big model price war" in may this year, which attracted the attention of the entire industry.

in this battle, baidu's price-cutting efforts can be described as "ferocious":

soon after the battle started, baidu directly announced that it would make the pre-setting service of ernie-speed, ernie-lite, and ernie-tiny series models free;

in july this year, during waic, baidu smart cloud announced further price cuts, significantly lowering the prices of the two flagship models, ernie 4.0 and ernie 3.5, with input and output prices as low as 0.03 yuan/thousand tokens and 0.06 yuan/thousand tokens respectively.

by removing the price threshold, many users will naturally flock in. in particular, many large, medium and small companies and institutions, without the cost concerns, are enthusiastic about trying large models.

the price reduction does reduce the “push” for customers to call;the competition in the large model market is so fierce. even if the price is lower, especially when the price is already as low as it can be, why can wenxin stand out?

an article has already analyzed that baidu's ability to make such an almost decisive concession is essentially due to baidu's more advanced full-stack ai technology, which gives it the confidence to offer services for free or at a lower price.

since its birth, the wenxin big model has been keeping up with the trend of the times. the average weekly training efficiency of the wenxin big model has reached 98.8%. compared with the release of wenxin yiyan, the training efficiency has increased to 5.1 times, the reasoning has increased to 105 times, and the reasoning cost has dropped to 1% of the original. in june, baidu also launched paddlepaddle 3.0. this upgrade significantly improved the compatibility of the paddlepaddle framework with baidu's ai infrastructure and the wenxin big model, and is expected to further reduce the cost of model reasoning in the future.

baidu has conducted a very in-depth analysis of the pain points of current large model users. the first type of pain point for users is that the threshold for application implementation is relatively high, and how to quickly combine scenarios to develop suitable applications. the second type of pain point is how to use dedicated models to meet their own industry needs for specific scenarios.

in essence, for a technological product to gain users’ support, the first and foremost consideration is still whether it is “usable” and “easy to use.”

just one day before the release of the financial report, baidu announced that it would support fine-tuning of the flagship large model ernie 4.0 turbo. previously, baidu qianfan large model platform has supported fine-tuning of ernie 3.5, ernie speed, ernie lite, ernie tiny, and ernie character. as of now, a total of 6 wenxin large models can be fine-tuned on the qianfan platform, with a total of 21,000 models fine-tuned, serving more than a thousand core business scenarios of enterprises, with many successful cases.

in the field of government affairs, the all-china federation of trade unions and baidu smart cloud have applied ai capabilities to legal consultation, allowing workers to understand legal issues more accurately, calculate labor compensation amounts more quickly, and complete case assessments more personally through big models.

compared with traditional development methods, this system, which requires extremely high accuracy and feedback speed, was delivered in just one month, which is a significant improvement in efficiency compared to the traditional 3-6 month delivery cycle.

with such an improvement in productivity, more users will see its value, and it will be difficult for the number of calls to the wenxin model not to increase.

cloud business is growing rapidly, relying on maas and tools

the rapid development of ai has directly brought baidu the continued growth of the company's cloud business.

the financial report disclosed that in the second quarter, baidu smart cloud's revenue was 5.1 billion yuan, a year-on-year increase of 14%, with ai revenue accounting for 9%, higher than 6.9% in the previous quarter.

the excellent large model is regarded as a business card, and the cloud business closely integrated with it will naturally be driven along with it. the boom in ai applications has gradually made gpu cloud a "standard configuration" for corporate procurement.

as a result, the model as a service (maas) model is gradually coming onto the stage and becoming a new growth driver for baidu cloud.

according to the latest idc report, baidu smart cloud will rank first in china's large model platform market share in 2023, reaching 19.9%. similarly, among the nearly 260 large model projects in china that have been announced as winning bids this year, baidu leads the ranking of winning bidders. among the mainstream large model manufacturers, baidu has the largest number of winning projects, covers the most industries, and has won more than 64 million bids, ranking first in three categories.

during the period when the large model was implemented, qianfan platform has "made a fortune in silence" and has served a total of 150,000 customers. it has made efforts in many industries such as government affairs, electricity, and education. not only has its technology products been recognized by the market, but it has also brought feedback in the form of productivity upgrades to society.

in villages in henan, chongqing and other places, qianfan is promoting the "big model going to the countryside", allowing villagers to safely hand over the most trivial and time-consuming issues such as medical insurance payment and household registration to big models for assistance.

when encountering problems that are difficult to accurately cover in the local government knowledge base, the big model can also provide answers in combination with baidu search, and automatically @reply to the person asking the question, bringing round-the-clock convenient services to the countryside.

at present, resident assistants have entered more than 6,000 villages, allowing ordinary people to enjoy a smarter life.

in order to make customers feel that the product is "usable" and "easy to use", baidu relies more on upgrading the big model tool chain in the commercialization of big models.

deepening tool chain upgrades is one of the main themes of baidu's technological development this year.

from the perspective of model fine-tuning, modelbuilder has launched the high-quality data function for hybrid wenxin large models.users can integrate general mixed corpora and vertical field corpora with business data to fine-tune more stable and effective industry-specific large models.

from the perspective of ai native application development, appbuilder solves many problems that customers face when developing ai native applications. it enhances aspects such as massive knowledge retrieval, custom strategies, and enterprise-level security, and takes the usability of large models to a new level. currently, more than hundreds of thousands of applications have been created on the platform, covering industries such as online education, e-commerce, and government affairs.

baidu's big model system has actually improved customer productivity, which is obvious to the industry.

looking back, baidu itself also has a huge technology product ecosystem. if the capabilities of the big model are really so powerful, can it allow baidu, a 24-year-old "old tree", to sprout new shoots?

reconstructing the product model, seeing the world and seeing yourself

there is a widely circulated story on the internet: in the middle of world war ii, when the war was at its most intense, the quality of the us military's parachutes was worrying. in order to make the pass rate of parachutes reach 100%, the military came up with a solution: let the parachute sales staff test them personally, and they would not pass unless they jumped. in this way, the pass rate of parachutes finally approached 100%.

"if you don't use a product yourself, how can you let others trust it?"

the same principle applies to baidu as well. as early as last year, robin li announced that he would use the wenxin model to reconstruct all of baidu's products. the first to bear the brunt of this was baidu's headquarters - search.

currently, 18% of search results are generated by ai, which can provide users with more accurate and direct results. this just confirms the concept of "new search" proposed by baidu a year ago, with the three characteristics of "extreme satisfaction", "recommendation stimulation" and "multi-round interaction", making users' search results more accurate, organized and intuitive.

intelligent agents are a key part of this. baidu is accelerating the distribution of intelligent agents in search results. currently, baidu's average daily distribution of intelligent agents has exceeded 800 million, which is twice as much as in may.

robin li repeatedly emphasized that intelligent entities are like ai websites, with low barriers to entry but high ceilings.

from a development perspective, developing intelligent agents is even simpler than developing websites. li yanhong said, "how were websites made at that time? it was very simple to browse the source code through a browser. with a little modification, i could also make it. making intelligent agents today is very similar to this... just give it a name, tell it what to answer and what not to answer, and an intelligent agent is done."

as one of the earliest large companies in the industry, baidu has built an ecosystem of initial scale in the field of intelligent agents. on baidu's wenxin intelligent agent platform agentbuilder, 200,000 developers and 63,000 companies have settled in. when developers create intelligent agents on the wenxin intelligent agent platform, they can flexibly choose the wenxin model 3.5 or 4.0 version. both versions of the model can be used free of charge, which can be said to have "lowered the threshold for use".

with the opening of the paris olympics this summer, many fans of athletes took action and spontaneously gathered on the baidu wenxin agent platform to develop a number of fan support agents. fans of chinese female table tennis player sun yingsha developed "sun yingsha's little fangirl" for her; fans supporting female tennis player wang manyu developed "manyu's little taro balls"; and the agent of female diving athlete quan hongchan was named "chanchan's little schoolbag" by fans because of her popular schoolbag pendant. the communication was unique and very interesting.

in the field of agriculture, academician zhu youyong of the chinese academy of engineering also assisted baidu in creating the "farmer academician intelligent agent". this intelligent agent, equipped with the research results of zhu youyong and his team, can answer a variety of questions for farmers, and can easily grow crops such as dryland high-quality rice and winter potatoes, bringing the benefits of technology to farmers to a new level.

with low barriers to entry and good results, how can intelligent agents not be popular? it is not difficult to see that the scale of intelligent agents in the future can be compared to the numerous websites today, forming a huge ecosystem.

baidu's ambition is to use search as the largest entrance for the distribution of intelligent entities and to stand at the forefront of the great prosperity of intelligent entities.

among baidu’s traditional products, baidu wenku is famous for its “second success”.

reconstructed by a big model into a "one-stop ai content acquisition and creation platform", baidu wenku is moving further forward on the road of ai, with hundreds of multimodal ai functions such as smart ppt, smart documents, smart mind maps, smart research reports, etc., which greatly enhance users' document experience and usage efficiency.

as of now, the total number of ai users of baidu wenku has exceeded 180 million, and the number of times the ai ​​function has been used has exceeded 2.2 billion, allowing more people to enjoy the productivity dividends brought by the big model and giving this 15-year-old product a "second spring".

conclusion

this year, robin li once pointed out that the focus of the big model is still "volume application": "without application, the basic model alone is worthless, whether it is open source or closed source."

but what we can actually see is that baidu is trying to do both basic models and applications well.

over the past decade, baidu's hard work in big models and basic ai research has finally produced an acre of fertile soil. but soil alone is not enough. baidu has entered the second stage of their big model development, focusing on creating ai native applications - growing rich "crops" on this "fertile soil."

deeply integrating generative ai into its business and resolutely using large models to reconstruct search have once again boosted the confidence of a number of securities firms in baidu.

“from a business perspective, such a change will expand baidu’s role in the value chain from pure traffic generation to pre-sales consulting,” jpmorgan said. “we believe the latter will increase the flow of deal conversations and, if executed reasonably successfully, will have a positive impact on earnings in the medium to long term.”

if other large companies want to truly implement big models, they may really need to learn from baidu: reduce the usage and cost thresholds to a minimum, rely on cloud strength and tool chains to establish a well-functioning ecosystem, and rely on big models and intelligent agents to reconstruct existing businesses.

these three steps are the key operations that drive large models to truly “enter the homes of ordinary people.”