"Today's shocking release, tomorrow's epic update, but I have to ask, where is the application? Who has benefited from it?"
On July 4, Robin Lee, chairman of Baidu, strongly criticized the competition of big models at the main forum of the World Artificial Intelligence Conference (WAIC) on industrial development, saying that many companies focus on the basic models, running scores and ranking all day long, and who has surpassed GPT4? OpenAI has launched sora, GPT4o, etc.
Big model ranking refers to companies and research institutions competing to achieve the highest scores on various benchmark tests, in order to demonstrate the performance superiority of their models. In the pursuit of high scores, some models may be trained on samples containing benchmark test datasets, resulting in unfair advantages.
In addition, some benchmark tests may be too simple or not comprehensive enough to fully reflect the true capabilities of the model in complex scenarios, and may also lead to imbalanced resource allocation.
In Robin Lee's view, the most important thing at present is applications. "Without applications, only basic models, whether open source or closed source, are worthless." He called, "Don't roll models, roll applications."
In fact, applications based on basic models have gradually penetrated into various industries. Taking the express delivery industry as an example, a large model can shorten order processing time from over 3 minutes to 19 seconds, solving over 90% of after-sales problems with efficiency. In the field of code generation, about 30% of Baidu's internal code has already been generated using AI, with a adoption rate of over 44%.
According to Robin Lee, the daily adjustment volume of Wenxin Model exceeded 200 million more than two months ago, and more than 500 million recently. "The changes in the number of calls reflect the real needs, and some people really benefit from the big model."
But he also reminded that to avoid falling into the "super application trap", "DAU's 1 billion APP is called success" has been the thinking logic of the mobile Internet era. In the AI era, the "super capable" application is more important than the "super application" of DAU only. As long as it can generate great benefits for the industry and application scenarios, the overall value is greater than that of mobile Internet.
The promising AI application development direction inside Baidu is agent, and the produced agent is simpler than making web pages in the Internet era. In the future, in fields such as healthcare, finance, education, manufacturing, transportation, and agriculture, various intelligent agents will be created based on their own scenarios, unique experiences, rules, data, and so on.
"There will be millions of agents in the future, forming a huge ecosystem of agents," Robin Lee said.
He once again expressed his views on the heated debate between open source and closed source in the industry, stating that some laymen have confused model open source and code open source. In his view, under the same parameter scale, the ability of closed source models is better than that of open source models. If open source wants to catch up with the ability of closed source models, it needs to have larger parameters, which means that the inference cost will be higher and the reaction speed will be slower.
"Open source models do not allow you to iterate and develop on the shoulders of giants," he said. Some companies have modified their models through open source models, creating an isolated model that cannot benefit from continuous upgrades of the basic model or share computing power. Open source models are only valuable in a few fields such as academic research and teaching, and are not suitable for most application scenarios, especially in a fiercely competitive market environment. To make a company's business efficiency and cost better than its peers, commercial closed source models are the most competitive.
However, some people in the industry expressed different opinions to the interface news, believing that Robin Lee's comments on the closed source model are only based on Baidu's interests, and the open source model has its commercial value, especially in the long run, the ability of the open source model will continue to catch up with the closed source model.
The debate over open and closed source models is just a corner of the domestic "hundred model war". According to Robin Lee, the war caused a huge waste of social resources, especially computing power. On the other hand, China began to catch up with the most advanced basic models in the world.
Behind the above statements, it also reflects to some extent Baidu's urgent desire to change the current situation.
In the last round of mobile Internet competition, Baidu kept "falling behind" and chose to bet on AI instead. In this round of AI competition, there are not only old Internet giants, but also big model startup star companies, with more intense competition.
Even in Baidu's most powerful search field, it has to face the "erosion" of competitors - not only occupying the content search of WeChat and Xiaohongshu, but also major model companies have launched AI search, such as Kimi, Doubao, Conch AI, Zhipu Qingyan and other products. Search has become a must-have for large-scale consumer applications.
Image source: CICC
At least in terms of volume, Baidu has fallen behind. According to the latest research report released by CICC, Kimi has always been ahead of ERNIE Bot of Baidu and won the first place in Chinese AI website traffic since its explosion in March.
Taking the week from June 19 to 25 as an example, in the weekly visits to Chinese AI websites, Kimi chat under the Dark Side of Moon ranked first with 571.83 points, and ERNIE Bot ranked second with 412.97 points; In the list of China's Al app downloads and average daily active users, the Doubao under ByteDance ranks first, and Baidu always ranks second.
Robin Lee and Baidu need to face the issue of how to maintain their advantages in the subsequent competition and achieve breakthrough.