Robin Li: The combined usage of over 200 domestic large models is still less than that of Wenxin alone
-
"Continuously developing various foundational large models is, in my opinion, a tremendous waste of social resources."
"To this day, whether in China or the U.S., I believe the best AI-native applications have yet to emerge."
On November 15th, at the Xili Lake Forum held in Shenzhen, Robin Li, founder, chairman, and CEO of Baidu, shared his two cold reflections and three hot drivers regarding large models. He has previously expressed similar views on multiple public occasions, stating that entrepreneurs competing in large models is meaningless, while competing in applications offers greater opportunities. At the Xili Lake Forum, Robin Li reiterated this perspective, saying, "I believe the hallmark of humanity entering the AI era is not the creation of numerous large models, but the emergence of many AI-native applications."
"Reports indicate that by October, China had released 238 large models, up from around 79 in June—a threefold increase in just four months. Yet, how many native AI applications can anyone here name? While dozens of foundational large models have been released globally, there are already thousands of native AI applications abroad, something notably absent in the Chinese market," said Robin Li.
Drawing parallels to the PC and mobile internet eras, Li noted that the Windows OS spawned countless software, while Android and iOS now host 8 million apps. "Large models are foundational platforms, much like operating systems. Ultimately, developers will rely on a select few large models to create diverse native applications. What we need is millions of AI-native apps, not just another hundred so-called large models."
The above is Robin Li's first cold reflection. Speaking of the second cold reflection, Robin Li mentioned that he has observed a phenomenon where many enterprises and even cities are still hoarding chips and building AI computing centers, attempting to train their own dedicated large models from scratch. Due to the lack of emergent intelligence capabilities, the value of these dedicated large models is very limited.
It is worth noting that Baidu itself is a provider of foundational large models. On October 17, Baidu launched the Wenxin Large Model 4.0, claiming it to be the most powerful Wenxin model to date, with comprehensive capabilities comparable to GPT-4. During his hour-long speech, Robin Li also demonstrated to the public how Baidu's various businesses have been reconstructed based on the Wenxin Large Model. Earlier, at the end of August this year, a batch of large model applications, including Baidu's Wenxin Yiyan, were made available to the public.
At the Xili Lake Forum, Robin Li stated that since its opening on August 31, the API calls for Wenxin Model have shown exponential growth. "There are over 200 large models domestically, appearing on this list or that ranking, but their actual call volumes are quite small. The call volume of Wenxin Model alone exceeds the combined total of these 200-plus large models."
Robin Li also discussed three 'hot' drivers in the AI-native era. The first is powerful foundational models, which will drive the explosion of AI-native applications. Just as the mobile era gave birth to 'mobile-native' applications like WeChat, TikTok, and Uber, the AI-native era will also see outstanding AI-native applications developed based on large models.
Secondly, embracing the AI era requires leadership from CEOs and top executives. Robin Li noted that many companies lack a deep understanding of the essence of the issue. CEOs often delegate tasks to IT leaders, who, along with engineers, are swayed by buzzwords like 'groundbreaking releases,' 'epoch-making updates,' 'iPhone moments,' and 'explosive innovations.' They either attempt to build foundational models themselves or select high-scoring large models based on online evaluations, mistakenly believing this constitutes embracing the AI era. 'Only CEOs truly care whether new technologies positively impact their business's key metrics.'
Thirdly, a thriving ecosystem of AI-native applications will drive economic growth. Li believes that high-quality applications can stimulate the market and force changes. Using new energy vehicles as an example, he pointed out that China holds 65% of the global market share, primarily due to government policies that boosted demand and application development. Similarly, the AI industry is demand-driven. Efforts should focus on the demand side and application layer, such as incentivizing businesses to leverage large models for developing AI-native applications, thereby using market mechanisms to propel industry growth.
According to The Paper reporters, compared to its initial hype, large language models have seen diminishing public attention. A key concern is whether Baidu's early enterprise clients have continued using the ERNIE model and whether it has generated revenue for the company.
In response, a Baidu executive mentioned during media discussions that all technological developments follow a hype cycle. While current enthusiasm has moderated, industry focus and expertise in large models continue to deepen. The technology remains in its early stages, with significant evolution expected. Current applications don't represent the full future potential.
Another Baidu representative noted that from a commercial perspective, if large models fail to find practical applications, sustaining the technology will be challenging, requiring ongoing investment.
It is reported that in the field of foundational large models, most companies that have released such models believe that many key industry processes are worth transforming with large models, indicating a vast market. However, compared to the enthusiasm of releasing companies, the general public has yet to perceive the 'charm' of large models. Some To C (user-end) applications only generated short-lived popularity after their release.
At the Xili Lake Forum, Robin Li also stated, 'I've talked so much about large models and AI-native applications because I hope everyone will take action—use them, understand them, experience them, and engage in the innovation of AI-native applications.'
Baidu has not yet released its Q3 financial report. Its Q2 report did not disclose the revenue contribution from Ernie Bot. When discussing the impact of AI technology on cloud business during the Q2 earnings call, Baidu Cloud executives mentioned that it is still too early to talk about operational models.
It is reported that Baidu will release its third-quarter earnings report next week. Investors remain particularly focused on the company's R&D investments in artificial intelligence and whether its Ernie Bot has attracted new demand.