Will 2024 Be the Year of Multi-Track AI Application Wars?

[baoshi.rao]

"In just one year, 11 versions have been updated, from simply adding design templates to the recently launched animation generation and website building features. An AI-powered PPT tool has undergone hundreds of modifications and upgrades." Last week, when an investor who had originally decided not to focus on AI projects noticed that the momentum of consumer AI applications—decline, growth, and iteration—still showed no signs of slowing down this year, he clearly became hesitant.

This year, content editing tools have become the second most popular category of AI products among users on web platforms, accounting for 22%, nearly catching up with content generation applications like ChatGPT and Gemini, which are completely "monopolized" by major tech companies. This is a key finding highlighted in a16z's Top 100 Generative AI Consumer Applications report released this month. The product that made this investor hesitate is called Gama AI, whose user base experienced explosive growth last August. However, Gama AI's performance has not been stable. By February of this year, the number of queries on the website had dropped by more than half from its peak of 25 million last year. In contrast, Prezi, a Hungarian company founded in 2009 and known as the disruptor of slides, has shown more stable growth. Prezi has achieved nearly 20% new user growth in recent months.

These are just the tip of the iceberg in the current AI application landscape. According to website traffic statistics from Toolify.ai, among Chinese companies, ByteDance's AI creative platform CapCut became one of the fastest-growing applications globally in February. Additionally, Moon Dark Side's Kimi Chat quickly entered the global top 15 in growth last month.

Even with an active market, this investor still has some concerns. The situation is completely different from when he first entered the industry a decade ago, during the explosion of mobile internet applications. Nowadays, newly emerging software tools like Gama AI can become excessively bloated in a much shorter time, facing severe homogenization. Many developers focus far more on features than on addressing user pain points. In fact, 90% of current AI applications or plugins still lack practical value.

He told Forbes China, "There are still too many shell applications and conceptual products in this market. But this year, it's time to shift attention from the computational power anxiety surrounding domestic large models back to focusing on product strength." For example, after Kimi Chat increased its context length to 2 million words, the product's user experience improved exponentially, and its scenarios became more diverse—something GPT-4 has indeed not achieved. "To some extent, Chinese companies still have strong capabilities and instincts in application development."

During the Jobs era, the product principle of pursuing ultimate user experience was to make "simplicity the ultimate sophistication." However, after the advent of ChatGPT, everyone began to believe in brute-force miracles, eagerly integrating Open AI's APIs without hesitation, as the AI label on a product now equates to traffic. This methodology is unlikely to work effectively this year. For instance, Microsoft, the biggest beneficiary behind OpenAI, failed to replicate last year's overwhelming praise when integrating GPT-4 into its Office suite as it did with the Bing+ GPT smart chat feature. Some users even commented, "It's hard to believe this is software powered by GPT-4." Tech influencer Luke Barousse also expressed frustration after using the "Preview" version of Copilot. During his tests, he found that Copilot's Excel functionality currently only handles up to 500 rows of data (basically, tasks that normal users could complete with a few clicks). Additionally, the data visualization lacks personalization (far from the dazzling variety showcased in Microsoft's demos).

The most critical issue, however, is user privacy and security. Excel users must upload all their data to OneDrive to access AI features. While Microsoft has promised not to retain commercial user data or use it for training, its stance on personal users remains ambiguous.

In contrast, WPS AI, Microsoft Copilot's competitor in China, appears more stable. It hasn't overpromised to users and offers a more user-friendly interaction experience, with document functionalities resembling the popular editing tool Notion AI. Yet, WPS AI has similarly failed to earn revolutionary acclaim from users. There's a common saying in Silicon Valley: never touch Microsoft's first-generation products. Columnist Gene Marks noted, "Copilot's reputation decline closely resembles Microsoft's 1985 Windows debut." However, the good news is that Copilot or WPS AI will continue to improve.

This year, the adoption threshold for AI technology will further decrease. With GPT-5's release in the second half of the year, Sam Altman stated: "Most simple AI wrapper applications will be phased out by the market because GPT-5's powerful capabilities will render them obsolete." Content editing tools like Gama AI, which currently have weak brand recognition and unclear product advantages, may face significant threats.

Under the influence of mobile internet's user habit cultivation, all traffic and information gateways are already guarded by established companies. Meanwhile, new entry points like VR/XR hardware continue to show sluggish shipment volumes, making them unlikely to present structural terminal disruption opportunities in the short term. Therefore, the winner-takes-all consumer market rule from the internet era may still hold true in the AI age. For instance, ChatGPT and NewBing accounted for 60% of user engagement time among the top 100 AI applications in February this year. In mainland China, AI products from major tech companies captured about 50% of user traffic. Consequently, focusing on vertical niche applications will be a key theme for startups in 2024.

Beyond this, the greater challenge for innovation comes from user ecosystems. Since Apple launched iTunes+App Store, the most significant trend over the past decade has been the shift from open to semi-closed platforms. This has benefited mobile app developers by making it easier to enhance user stickiness and loyalty amid reduced information flow efficiency—achieved through cross-functional product expansions. Thus, platforms like WeChat, Alipay, and Douyin have evolved into comprehensive service platforms integrating social networking, news, short videos, payments, and transportation.

Zhang Xiaolong once sarcastically remarked on Fanfou: "How many features must a product add before it becomes trash?" By 2021, at WeChat's 10th anniversary, China's most outstanding product manager revised his statement: "How many features must a product add to avoid becoming trash?" For major tech companies' AI products and AI-native startups this year, moving beyond the simple additive thinking of the internet era is becoming increasingly crucial, and some companies' user declines this year already hint at this argument.

Character.ai is an AI character generation product developer that went viral last year, founded by former Google engineers Noam Shazeer and Daniel De Freitas. This "product" allows users to create AI versions of real and fictional celebrities, generating viral content through conversations with them, while customized bots can also meet users' various emotional needs.

After the mobile version of Character.ai launched in May 2023, it achieved over 1.7 million installations in its first week. Within less than half a year, monthly active users on mobile devices exceeded 5 million. More importantly, the nearly 10,000 bots on Character.ai attract a younger audience, with 60% being aged 18 to 24, compared to less than 30% for ChatGPT. To some extent, Character.ai is not a chatbot company but rather a content company cloaked in AI. Last year, its content created a trend popular among young people, which also brought traffic and revenue.

However, this year, the hyper-saturation of AI bot creation and viral strategies has begun to raise alarms. In fact, user numbers on top-tier character-generation platforms have been declining for some time. For example, in February, Character.ai saw a 1.48% drop in visitors, Janitor AI experienced a 13.19% decline, and SpicyChat AI's traffic fell by about 8.23%. (Data source: aicpb.com)

In contrast, the fastest-growing platforms in terms of visitor numbers during the same period were AI products with specific use cases and clearly defined service boundaries. For instance, in the education sector, Q-Chat and CheggMate achieved visitor growth rates of 26.74% and 59.40%, respectively. Another example of PMF (Product Market Fit) involves search tools.

Over the past few years, as algorithmic recommendations have gained popularity, the importance of internet search functionality has declined. We can clearly observe Baidu falling behind and Google's diminishing presence. However, Perplexity represents another AI product category that has achieved stable growth this year. This search engine reached 50 million visits in February, marking an 8% increase from the previous month. Founded in 2022, Perplexity differs from traditional search engines by providing direct answers to search queries along with citation links for verification, rather than merely offering a list of links. This approach goes a step beyond querying ChatGPT and also supports real-time information retrieval.

Market feedback from the first quarter of this year shows that OpenAI has only demonstrated the problem-solving capabilities of large models in terms of performance. However, practical application scenarios require addressing engineering and interaction challenges, ultimately shaping a form of user education with clear values. As Steve Jobs once said, "People don't know what they want." The global high-intensity battle for large model positioning concluded in just one year. From OpenAI opening its API interface in March last year to Grok's complete open-source in March this year, large model companies hoped to build a massive application ecosystem once again. A story seemingly similar to the iOS vs. Android competition appears to have restarted.

However, contrary to the expectations of many large model companies, the prelude to this year's flourishing AI innovation began with a collective disillusionment with large model firms.

On March 10 this year, Jensen Huang emphasized in a public speech, "We should not focus solely on computational power but consider the comprehensive issue of energy consumption. The future of AI lies in photovoltaics and energy storage. If we only think about computing power, we would need to burn the energy equivalent of 14 Earths." If the core competitiveness of large model companies ultimately depends only on chip assets and energy, then it will increasingly resemble an "engineering management" business, much like cloud computing. Additionally, OpenAI's ecosystem vision has not progressed as smoothly as expected. OpenAI, which launched the GPT Store in January, originally hoped to make it easier for users to utilize and build GPT tools while providing developers with new revenue opportunities. However, this initiative did not receive the same positive feedback as the Apple Store's launch in 2010. The number of users engaging with these AI tools was significantly lower than expected, with many of the top 5% bots failing to reach 1,000 daily active users. Amid controversies over product reviews, copyright issues, and inappropriate content, the GPT Store rapidly lost market attention within just two months.

Learning from OpenAI's experience, Google's Gemini Pro and Anthropic's Claude 3, released this year, have similarly shifted their focus from consumer-facing applications to enterprise solutions. In other words, if large model companies cannot establish themselves as payment gateways for more users, they will lose their power in revenue distribution and are unlikely to become groundbreaking application platforms. Compared to North America, China's AI consumer market appears far more intriguing. In an environment where no absolute giant has yet emerged, the battle of large models has just begun, with greater emphasis on innovation in usage methods.

On March 18th, Chinese company Moonshot AI increased lossless context length tenfold to reach 2 million characters. In comparison, GPT-4 Turbo-128k's announced text range is only 100,000 Chinese characters, while Claude3-200k context handles 160,000 characters. User feedback indicates this upgrade significantly enhances Moonshot AI's capabilities in long-text analysis, web search, document processing, and particularly in deep analysis of voluminous book content. This also means Moonshot AI has taken a step closer to users who fear massive, obscure prompts.

Four days later, Alibaba's Tongyi Qianwen and 360 respectively launched long-text processing capabilities of 10 million and 5 million characters - 5 times and 2.5 times that of Moonshot AI's Kimi. China's innovation ecosystem has its unique narrative logic, from the battles of messaging apps, payment dominance, group buying wars to today's large models. The wheel of history is spinning at a familiar pace once again. Will the biggest battlefield for AI application innovation be in China, just like the mobile internet era?

PwC predicts that the largest economic gains from AI in five years will be seen in China (26% GDP growth by 2030) and North America (14.5% growth), totaling $10.7 trillion, accounting for nearly 70% of the global economic impact.

We have provided a selection of AI products that have achieved significant user volume and rapid growth over the past three months, based on traffic data from Toolify.ai, aicpb.com, and Data.ai.