As we approach two years since the inception of ChatGPT, the highly anticipated era of AI-generated content (AIGC) seems to be cooling rather than ignitingOne of the hottest discussions in the AI community recently has been about an AI startup that announced a shift away from investing heavily in large models to focusing on profitable AI applications.

The current exploration of AI can roughly be categorized into three classes: universal foundational large models, industry-specific large models, and native AI applications developed based on the first two categoriesThe first two serve as the backbone of infrastructure, while the latter relates to products that users can engage with directlyFor instance, Baidu's AI product Wen Xiaoyan is powered by the Wenxin large model, whereas ByteDance’s Doubao operates on its proprietary large model.

All major tech companies have embarked on creating universal large models, subsequently developing native AI applications from there, while also opening their models for developers to access via APIs

This approach is akin to establishing a shopping mall—providing a space with basic amenities so that vendors (developers and businesses) can set up shops, with the mall operator charging for services while also participating in commerce themselves.

This year, voices indicating that the competition surrounding large models is wavering have gained traction, with reports of layoffs in some international firmsIn this context, actions taken by well-funded and technically equipped giants serve as significant indicators of market trends.

The experienced consultancy, “Ding Jiao One,” has endeavored to decipher the AI landscape shaped by five major domestic players: Baidu, ByteDance, Alibaba, Tencent, and Kuaishou, by probing their strategies and the significant AI products that have emerged over the past couple of yearsThe pressing question being whether any breakout AI applications are on the horizon.

Last year marked a peak interest in large models, with nearly all competent companies—Baidu, ByteDance, Alibaba, Tencent—launching foundational large models like Wenxin's universal model, the Doubao large model, the Tongyi universal model, and the Hunyuan large language model

Kuaishou took a different approach with a model centered on video generation.

Moreover, Baidu, ByteDance, and Alibaba have also invested in niche industry models closely tied to their core businessesThe diversity of AI products they have developed further illustrates a broadening scope in this field.

Recent public information compiled by “Ding Jiao One” reveals notable consumer-facing (to C) and business-oriented (to B) AI application products from these five companies over the last two yearsAmong them, Baidu has shown a commendable commitment to AIIts chairman and CEO, Robin Li, has often emphasized the necessity of practical AI applications, stating that “without application, foundational models are worthless.”

Much of the consumer engagement with Baidu's AI products notably leans into search and language processing, with Wen Xiaoyan receiving the highest mentions among end-users

Additionally, it ranks among the top three across various domestic AI application lists based on monthly active user statistics.

Wen Xiaoyan primarily serves as a multifunctional AI application, but it particularly shines in search capabilitiesBaidu promotes it as "new search," allowing users to engage in conversations with Wen Xiaoyan rather than merely retrieving keyword-driven data from web pages.

The application also features diverse interaction modalities such as writing, job searching, entertainment, and office-related tasksIn a workplace context, users can access tailored conversations for activities like crafting PPT presentations, simulated interviews, and formal document writing, simply by selecting the corresponding function.

Beyond this, Baidu’s offerings have also extended into image generation, digital human avatars, and smart customer service solutions.

In contrast, ByteDance has emerged with a broad range of AI applications, with its Doubao app currently leading in user engagement

alefox

According to QuestMobile data, as of July, Doubao ranks first among comprehensive AI applications in terms of monthly active user size in China.

The appeal of Doubao stems from its extensive functionality coupled with effective results, catering to various tasks including image generation, essay grading, and work summary assistance, as well as more whimsical features such as name scoring and MBTI personality testingThe app’s engaging persona is highlighted by a virtual Doubao character greeting users, with support for simultaneous text and voice outputs.

AI practitioner Li Jingjin has lauded the music generation feature of Doubao, asserting that it stands at a leading position domestically.

Alibaba’s Tongyi represents a holistic AI, integrating capabilities for generating text, images, and videos into one frameworkLi Jingjin notes that its responsiveness to societal trends is commendable, exemplified by its ability to facilitate popular online phenomena such as robots playing basketball or dancing cats swiftly

With the Nobel prizes recently awarded in AI-related fields, Tongyi launched a feature allowing users to create their personalized Nobel portrait almost as an immediate engagement.

Tencent's offering, the Tencent Yuanbao, also targets AI-based search and dialogue applications but is perceived to prioritize educational, occupational, and creative enhancement for users, distinguishing itself from the more entertainment-focused Doubao.

Kuaishou, on the other hand, has fewer AI products to showcase, with its most notable being the Keling AI, which emphasizes video generationCurrently, Keling AI is emerging as a leader in the text-to-video domain, with numerous users remarking that Keling excels in text understanding, generation speed, and video clarity, potentially reaching commercial-grade quality.

“For promotional videos and ads, I sometimes need precise control of certain visuals; specific elements need to function in designated trajectories, and Keling’s motion brush allows for this,” said one industry practitioner.

In summary, while the offerings of these major companies predominantly revolve around AI assistant products, they share similarities in functionality and often do not differ vastly in performance, especially as many are available for free

Other product types, such as the “Miao Duck Camera,” saw brief popular phases but lacked sustained engagement across ongoing monthly active users.

Some companies are steadfastly safeguarding their interests, while others are biding their time.

A detailed comparison of the AI strategies among these five giants reveals distinct approachesBased on insights from various industry specialists, “Ding Jiao One” has encapsulated these strategies effectivelyByteDance stands out, having launched the highest number of AI products, covering areas like assistants, social networking, imagery, video, and education.

Insiders have characterized ByteDance's approach as “human monitoring defense,” ensuring they are not missing out on any significant AI product domestically recognizedFor instance, their AI camera application Xinghui directly competes with Alibaba's Miao Duck Camera.

ByteDance is also pushing into new hardware territory, recently unveiling the AI headset called Ola Friend.

Renowned AI entrepreneur Lian Shilu analyzes ByteDance's strategy as mirroring their tactics in other sectors, which underscores a multi-pronged product strategy akin to a racing competition

Unlike other giants, ByteDance has notably focused on consumer-oriented AI applications.

Alibaba has been involved in substantial investments, having backed promising AI startups like Zhiyu AI, Ling Yi Wan Wu, Baichuan Intelligent, MiniMax, and Yue Zhi An Mian alongside developing its proprietary Tongyi large model.

Another marked strategy for Alibaba has been creating AI applications tightly woven into its existing ecosystemSeveral practitioners have noted that compared to its counterparts, Alibaba’s AI products have a closer synergy with e-commerce, introducing tools like Pic Copilot and Dui You explicitly geared towards aiding merchants and enhancing e-commerce marketing effectiveness.

During this year's Cloud Summit, Ant Group rolled out three AI products—Zhi Xiaobao (Alipay AI Assistant), Ma Xiao Cai (Financial Planning), and AI Health Manager (Healthcare)—all conceptually aligned with its foundational business focus.

Baidu, however, continues to double down on AI, with a pronounced pivot toward enterprise solutions

Many professionals believe that the success of Baidu in reclaiming its status among internet frontrunners pivots significantly on its AI ventures.

Baidu’s commitment to AI is substantiated in its quarterly reportsIn Q2 of this year, its smart cloud service recorded revenues of 5.1 billion yuan, showcasing a growth rate of 14%, while also highlighting an average usage of Wenxin's large model surpassing 600 million calls daily—demonstrating a tenfold increase over six months.

At a recent corporate-led directors meeting, Robin Li reiterated the commitment to investing in foundational models while referencing business avenues like search, digital humans, AI agents, large model calls, and applications like "Rutabaga Runs." He also dismissed venturing into spaces such as Sora’s text-to-video applications due to their extended timelines, suggesting that such endeavors could take 10 to 20 years before yielding financial returns.

Conversely, Tencent and Kuaishou present a calmer posture, with Tencent launching its Hunyuan large model officially just last September

Additionally, Tencent has invested in several notable AI firms like Zhiyu AI and MiniMax.

Tencent has focused on developing a robust portfolio of B2B database applications, while their B2C offerings, apart from the well-recognized Tencent Yuanbao, have garnered relatively less visibility.

Some industry insiders speculate that Tencent seems poised to unleash a major strategy“Currently, AI hasn’t birthed groundbreaking applications, and no primary application direction has emerged; hence, Tencent is perhaps waiting for others to pinpoint larger application potential.” Qin Yu, an AI data research expert, concurs that Tencent, which possesses a wide array of potential application scenarios, has the luxury of patience.

Lian Shilu adds that Tencent appears to be in a moment of waiting, “As soon as Tencent identifies another player discovering a precise AI product direction, they can move fast and potentially leapfrog them.”

As for Kuaishou, their current AI products are mostly limited in variety and remain primarily concentrated within niche applications in video-related arenas—text-to-image, text-to-video, and AI editing.

In conclusion, Baidu, Tencent, and Alibaba are emphasizing their B2B AI products, while ByteDance concentrates on consumer-oriented solutions, particularly in the social and entertainment spheres.

The usability feedback across various products indicates that AI solutions integrated within a company's existing framework often demonstrate superior market performance; notable examples include Baidu’s Wen Xiaoyan search assistant and Kuaishou's video generation tool Keling AI, alongside Alibaba's Tongyi Qianwen, which excels in e-commerce discount strategy modeling and marketing manuscript development.

In terms of emerging breakout AI applications, the industry is still navigating through exploration.

Lately, there have been numerous reports of startups contemplating a retreat from pre-trained models (the foundational training data feeding large models). Though some companies have publicly refuted these rumors, the sentiment indicates an overarching realization that the race surrounding large models is becoming less tenable, redirecting focus towards more tangible AI applications as a path forward.

Lian Shilu contends that the overarching strategy among the major players appears to be a temporary pause on foundational large models while pivoting towards development of AI applications.

This shift is primarily influenced by two aspects

First, there is a relative stabilization within the domestic foundational large model arena.

AI practitioner Li Siyu explains that last year’s intensive competition stemmed from the belief that substantial growth potential exists for large models, prompting all companies to initiate groundwork, which has now largely been accomplished—focusing more on optimizing rather than groundbreaking advancements.

Large model projects incur hefty expenses; as per data from Sequoia Capital, a staggering $50 billion was spent on NVIDIA chips by the AI industry alone in 2023, with total revenues merely reaching $3 billion.

Qin Yu remarks that uncertainty prevails regarding the exact extent of training investment required for pivotal models—making it challenging for companies to ascertain the necessary capital for pre-training regimens.

In other words, continuing to saturate the market with large model investments appears economically impractical at this juncture, as distinctions in technology among industry leaders have significantly diminished.

Lian Shilu elaborates on the underlying infrastructure's architecture, hardware, and workforce, noting that foundational models among the five leading firms share a similar backbone, alongside a calculative power factor hovering around ten thousand accelerator cards, with Alibaba slightly ahead

Furthermore, the robustness of their core AI scientist teams remains stable; accordingly, they differ predominantly in the volume of training data collated.

On the note of evolving from the aggressive growth last year, players are gravitating back towards rationality.

Experts echo that this year's AI industry has effectively weeded out much of the superfluous influx of speculative participants and funding, leaving a core of earnest contributors who understand that building large models ultimately serves the purpose of creating products that genuinely address pressing challenges for enterprises and consumers.

However, traversing the product application road also poses challenges.

Foremost among these is the lack of breakthrough scenarios to discover, an assertion supported by various AI usage metricsQin Yu has noted that currently, AI product types are predominantly confined to searches, role-playing, and chatbots with little in terms of novel utility scenarios

Over the last couple of months, monthly active users continue to be concentrated among established products such as Doubao, Wen Xiaoyan, and Kimi, with scant fresh solutions emerging.

Additionally, the engagement levels of AI products are closely correlated with marketing expenditures.

Qin Yu divulges that, despite stability in the rankings among top AI applications, fluctuations in monthly active users often reflect advertising investments—“Whichever product sees a surge in advertising for the month experiences rising active user stats, with Douyin emerging as the primary ad placement vehicle.” In another vein, media insights reveal that acquiring users remains heavily tied to marketing endeavors.

Nevertheless, whether with large models or functional AI applications, the competitive drive among major firms fuels their relentless ambition to seize a slice of the lucrative AI landscape.