In recent months, the developments in the AI landscape have taken a fascinating turn, particularly as we approach the end of the yearNotably, major players in the large model sector are racing to meet their KPIs, flooding the market with positive announcementsFor instance, one of the six major players in the large model arena, Zhipu AI, has recently secured a significant financing round amounting to 3 billion yuanIn the meantime, ByteDance has unveiled its Doubao visual understanding model, while Kuaishou's Keling 1.6 has officially launched, marking a significant milestone for AI applications.
As we look back, 2023 has undoubtedly been viewed as the inaugural year for large AI models, while 2024 is anticipated to mark the dawn of application-centric AIThe rapid evolution of this field, characterized by advancements in deep learning, the rise of AI agents, and the competition between multi-modal models, reveals an astonishing pace of innovation that often overshadows the fact that AI in China has only been burgeoning for less than two years.
The phrase "AI for a year is like ten years for mankind" rings true as this revolutionary product awaits its moment of explosive growth
Despite being heralded as a revolution in the internet landscape, we have yet to witness a significant breakthrough in the AI application ecosystem.
The fundamental driver in discussions surrounding the development of large models in China has long focused on computational powerIn the intensely competitive environment of AI, many leading companies once believed that increasing the scale of their models would help them rapidly gain market shareHowever, a shift has occurred this year, amplifying discussions about the stagnation of progress in the global large model ecosystemRecent months have seen fewer announcements from major companies regarding new model releases—not as an indication that the fierce battle for supremacy is ending, but rather a reflection that these firms are increasingly aware of the challenges posed by relying solely on computational power and parameter scaling to surpass their competitors.
The reality is that training and operationalizing large models demands substantial and consistent investments in computational resources—a daunting challenge for companies still struggling to establish viable business models
- Can the U.S. Stock Market Continue to Rise?
- Uncertainty in Japan's Central Bank Interest Rate Trends
- How to Weigh Economic Prospects?
- This year, agencies have started leveraging PR for profit.
- Quantum Computing Breakthrough Ignites the Market!
ByteDance, for example, plans to invest an astonishing 80 billion yuan in AI during 2024, which is almost equivalent to the combined capital expenditures of the three giants: Baidu, Alibaba, and TencentEstimates suggest their capital outlay might increase to 160 billion yuan by 2025, with 90 billion yuan allocated specifically for procuring AI computational resources.
Even renowned models like ChatGPT grapple with the scarcity of computational powerMicrosoft, for instance, has spent hundreds of millions of dollars and deployed thousands of NVIDIA A100 chips to create supercomputing platforms to enhance the performance of ChatGPT and its revamped Bing search engineFurthermore, Microsoft has outfitted over 60 data centers with hundreds of thousands of GPUs to facilitate the inference processes involved with ChatGPT.
From a long-term perspective, however, relying on computational power alone does not constitute a sustainable competitive moat; rather, it highlights core competencies at the hardware level
It is well-established that the quality of a model improves with stronger computational capabilities and more extensive training datasetsYet, at the end of the day, both factors largely derive from economic strength.
Discussing competitions from last year, engineers within Google recognized the challenge of keeping up with competitors like OpenAI, stating, "We have no moat, and neither does OpenAI." This sentiment reflects an apprehension that even with full commitment from Google, victory in the AI race is far from guaranteedThe rapid advancement of open-source AI models has further eroded any potential technological barriers, with the current superiority in data quality swiftly narrowing.
As it stands, gaining an edge in the AI sector and monopolizing profits is exceptionally challengingA multitude of companies in the large model space possesses models of varying strengths, yet their core functionalities tend to be homogenous, resulting in plenty of substitutes in the marketplace
Features introduced by one company are often quickly replicated by competitorsFor example, Kimi, initially leveraging external data sources, successfully entered the ranks of top AI models thanks to its long text capabilities, which have since become standard functionalities across most AI search products.
Moreover, the training materials utilized often hail from the same datasets, allowing multiple companies to capitalize on similar resources, effectively diminishing the competitive advantage that might have derived from computational strength or training datasets.
According to a recent research report titled "The Deployment of Large Models and Frontier Trends," it has been noted that while various key elements in the internet age can create competitive moats—such as data flywheels, network effects, switching costs, economies of scale, and user mindset—these constructs fall short when applied to the business models associated with large models
Thus far, a clear competitive barrier remains elusive in this emerging business landscape.
In the latter half of the year, the discourse within the AI sector has noticeably shifted from focusing on the model layer toward the application layer.
Wall Street star fund manager Cathie Wood has aptly pointed out that during the infrastructure building phase, hardware manufacturers often exhibit higher growth trajectoriesHowever, once this phase concludes, the market shifts its focus toward software and applicationsOver the past year, many large model manufacturers in China have displayed a lack of distinction in their capabilities across general scenariosAlthough there exists a variety of technological niches, such as Kimi's long text functionality and the deep reasoning capabilities championed by AI search products, not to mention the recent emergence of multi-modal models and popular visual AI models, the pressing issue of excessive homogeneity looms large over the AI product landscape.
On December 18, the Doubao visual understanding model was officially released, placing a significant emphasis on content recognition, understanding, reasoning capabilities, and visual description and creative abilities
Despite substantial media attention, just two days earlier, Kimi had launched its own visual cognition version, targeting foundational scientific imagery interpretation and reasoningThis highlights the pressing question posed earlier: In a landscape devoid of competitive moats, can an attacker find a distinguishing gap through application scenarios?
Large models fundamentally serve as a baseline, and to genuinely witness an explosive breakthrough, AI must establish a strong correlation with everyday usersIndustry insiders have previously highlighted that the emergence of killer applications could ultimately determine which firms rise victoriousThe assertion made by Li Yanhong resonates: “Without applications, foundational models are worthless.”
The advent of such killer applications could catalyze the adoption of technology across various fields, creating new scenarios and business models
For instance, in the evolution of 3G networks, the introduction of the iPhone served as a killer application that not only propelled the proliferation of smartphones but also paved the way for a multitude of mobile internet solutions.
However, whether AI can ignite another iPhone-like revolution remains uncertain at this juncture—at least as we look ahead to 2024, no sparks of innovation are visibleEven today, numerous reviews assessing user experiences across various AI applications, ranging from chatbots to image and video generation, reveal significant difficulty in identifying a single AI product that stands preeminent amidst the evolving landscape.
While ordinary users acknowledge the transformative impact of AI search products on existing search systems, no single company has thus far demonstrated a commanding lead in market perception or search volume, akin to the unequivocal dominance of Baidu or Xiaohongshu.
Although AIGC technologies have registered advancements in certain domains—evidenced by Kuaishou's collaboration with notable filmmakers like Jia Zhangke and Li Shaohong or the widespread adoption of AIGC in independent gaming—it is clear that the technology still languishes in its developmental stage
Obstacles involving content generation quality, stability, and controllability, coupled with limited application scenarios primarily focused on auxiliary roles, hinder widespread adoption fulfilling large-scale user demands.
Even in the realm of hardware, the homogeneity issue surfaces concerning mobile AI modelsPresently, AI smartphones primarily feature enhancements around voice interaction, image processing, and call enhancementTo realize an AI smartphone, leading manufacturers across the board have focused on upgrading voice interaction—promising to transform existing voice assistants into more intelligent AI agents, with functionalities that have now become typical for AI assistants like Doubao or Kimi gaining traction.
However, as Wang Hua, co-CEO of Innovation Works, aptly put it, the AI technology field has barely crossed its year-and-a-half markExpectations for rapid advancements in product forms seem overly ambitious
Genuine applications are likely to emerge only in early next year, provided a solid foundation is established.
A pivotal reason fueling the urgency surrounding AI applications is the market's quest to see AI technology transition into tangible business modelsAs significant capital influxes rush into the realm of AI, the demand for increasingly impressive returns prompts investors to expect that these projects deliver swift commercial application insights reflecting their investments.
In September of this year, Sequoia Capital partner Pat Grady stated that the firm's focus within the AI sector is shifting toward application developmentGrady indicated that Sequoia anticipates that the majority of groundbreaking AI companies in the future will emerge from the application tier rather than traditional foundational model developmentThis shift implies we will witness a greater symbiosis between innovation and practical application in AI.
Nonetheless, reports suggest Doubao's user metrics have increased over recent months, yet comparable to all AI conversational products within the industry, the duration of use, interaction tendencies, and commercialization potential remain suboptimal
The ByteDance management perceives AI dialogue products as a possible "intermediate state" and believes that a more visually engaging user interface could deliver a long-term ideal product formConsequently, they are elevating the priority of the Dream product, actively seeking creative routes to forge a “Douyin” for the AI era.
This information has not yet been officially validated by ByteDanceWhat remains evident, however, is that ByteDance lacks a first-mover advantage in AI product developmentThe debut of Doubao on the market in August 2023 came six months after Baidu launched its own generative AI solution in March.
Despite this, Doubao’s monthly active user count soared to nearly 60 million by November, achieving a growth rate exceeding 10%. Within the realm of native AI applications, Doubao has emerged as the undeniable leader, boasting over 108 million downloads—the highest in the sector—while its closest competitors, Wen Xiaoyan and Kimi, trail with user counts of 12.99 million and 12.82 million, respectively.
By October of this year, the cumulative downloads of Doubao ranked first with 1.08 billion, leaving Wen Xiaoyan (with 226 million downloads) in the dust
Interestingly, back in February, Doubao recorded merely 1.73 million visits, meaning its monthly active count skyrocketed by nearly 35 times over three quarters, indicative of significant user growth potential that contradicts claims of underperformance.
In terms of profitability models for AI search products, they circle around subscription fees and advertising revenue streamsHistorically, the domain of subscription-based services has presented considerable challenges in China, and overloading ads could dramatically hinder user experienceThis poses a palpable challenge when relying on conversational AI products for revenue generation.
Reports indicate that Doubao’s dialogue rounds and duration metrics, among other critical indicators, are not yet optimalHowever, whether considering AI search or its advanced form in AI agents, such productivity-enhancing tools are not necessarily seen as urgent consumer needs.
From a certain perspective, AI search products, being among the most immediately compelling outputs of generative large models in consumer-facing markets, primarily serve as environments for collecting sample data rather than generating revenue.
For ByteDance, emphasizing the strengthening of its training capabilities and ecosystem within conversational AI serves to pave broader commercial avenues in AI, reinforcing competitive experiences
As long as active users and sustained utility exist, the importance of AI dialogue products remains intactObserving contemporary dialogue AI products in the market reveals that ByteDance's competitors seem less eager to position conversational AI as a vital component of commercialization.
Former Baidu vice president and head of the mobile ecosystem commercial framework Chen Yifan once remarked on generative AI's commercial landscape that Baidu does not feel hurried to unleash all available traffic simultaneously; the priority is ensuring true satisfaction of user needsThe objective is to enhance the quality of responses and dialogue interactions as key performance indicators, asserting that fulfilling user demands precedes addressing commercial aspirations.
Reflecting on the innovations that shaped the previous wave of the mobile internet: the launch of the App Store in 2008 provided platforms for mobile applications, setting the stage for the profound evolution of smartphone ecosystems culminating in widespread adoption by 2012. This pathway unfolded over five years, while now the AI landscape has made remarkable strides within less than two years since its emergence