Source | Bohu Finance (bohuFN)

Author | All too well

The AI sector has initiated yet another phase of intense competition.

In early September, Meta unveiled its personal agent, Muse. Within 22 days of its release, cumulative downloads surpassed 5 million. On September 21, Meta's stock price surged by 11.43%, adding over $190 billion to its market capitalization in a single day.

Meta also managed to turn the tables, transitioning from an under-the-radar participant to a central figure in the market.

Subsequently, a familiar scenario unfolded.

Within three weeks, OpenAI launched its resident agent, Dots, at DevDay; Alibaba introduced its personal agent, "Qianwen Go!"; ByteDance was rumored to be developing a native mobile personal assistant; Tencent was also reported to be deploying multiple product lines simultaneously.

From the "lobster farming" craze earlier in the year to the current resurgence of personal agents, the AI industry appears to be caught in a vicious cycle: whenever one product gains momentum, other major players find it difficult to remain passive.

01 Why Did Muse Suddenly Gain Popularity?

Muse is positioned as a digital assistant capable of comprehending user objectives and executing tasks around the clock.

Doesn't that sound reminiscent of lobster farming? In fact, Muse drew inspiration from Open Claw and can be seen as an enhanced iteration of lobster farming.

However, unlike lobster farming, which necessitates self-deployment, environmental configuration, and learning how to use it, Muse presents virtually no barriers to entry. More significantly, lobster farming requires users to spend money on tokens, sometimes incurring exorbitant costs for minor tasks. Muse, on the other hand, currently employs a hybrid model of subscription plus transaction commissions, with the free tier offering approximately 100 million tokens per week.

Tencent Technology's 2026 Q3 AI Trends Report also underscores three technological advancements in Muse, summarized as follows:

Enhanced memory capabilities enhance its reliability. It continuously records and updates information. For instance, if a school announces a postponement of a sports event, it can synchronously update the old date in the user's memos.

It exhibits spontaneous initiative, making it more user-friendly. For example, Muse's design lead once entrusted his son's pre-school preparations to Muse, instructing it to check school emails, organize dates, purchase stationery, and find hoodies. During this process, Muse identified an easily overlooked task: the deadline for registering for his son's sports program was approaching. The lead was boarding a plane at the time but managed to submit the registration form before the deadline.

More granular permission management bolsters security. For sensitive operations like booking flights, users can restrict the AI's permissions and execution processes to prevent it from gaining unnecessary access.

These factors also contribute to the rapid ascent of AI agents.

The popularity of Muse highlights the significant real-world demand for personal agents. In the narratives of AI vendors, agents could emerge as the next major traffic entry point following search engines and app stores. Users would no longer need to open numerous apps; a single conversation with an agent would suffice.

Notably, Muse is integrated into apps like WhatsApp and Instagram without incurring substantial additional customer acquisition costs. Apptopia reveals that 95% of Muse users are also Facebook users, and 63% use Instagram. This suggests that existing social products can serve as vital entry points for agents to reach users.

This naturally excites domestic big tech firms with robust ecosystems.

However, making personal agents viable seems as challenging as conquering Mount Everest.

Conflicts of Interest Between Platforms

The value of agents lies in cross-platform integration, which requires them to access different platforms on behalf of users, search for products, and even execute purchases. For retail platforms like Amazon, if users increasingly shop via agents, the platforms' control over product displays, bid rankings, and ad placements could be undermined. Consequently, some platforms have begun identifying and blocking automated traffic from agents.

For instance, Amazon and other retail and SaaS platforms swiftly identified and blocked automated traffic from Muse.

The situation is even more severe domestically. While agents derive value from cross-platform integration, core services in China are confined within the ecosystems of super apps.

To grasp the intensity of permission issues, consider Doubao Mobile. After the 30-day public comment period for the Screen Automation Operation Declaration Protocol (SAEP), Doubao Mobile Assistant will gradually grant screen operation permissions for third-party apps. However, apps that have explicitly declined permission will not be operable, and developers can request to disable permissions at any time.

Currently, the latest list of rejected apps includes 151 applications, such as WeChat, QQ, Taobao, Alipay, Meituan, DeepSeek, Kimi, Xiaohongshu, Pinduoduo, and various banking and telecom apps. One can't help but chuckle at this list.

But this isn't even the most significant issue; commercialization seems like a distant dream.

Morgan Stanley's research indicates that even if users only consume 25% of their weekly token quota, maintaining an always-on, continuously thinking agent like Muse would cost $37 per month in server expenses, or $445 annually. This far exceeds Meta's global average annual revenue per user of approximately $70.

This implies that relying solely on subscription fees to cover costs for personal agents makes profitability a distant prospect.

Despite this, big tech firms remain uniformly committed to following the trend.

02 If You Can't Lead, At Least Don't Fall Behind

Logically, internet giants venturing into the AI space should possess overwhelming advantages.

Cutting-edge foundational models require top talent, massive training data, and substantial financial support. At the application layer, big tech firms have existing user ecosystems and business scenarios to integrate AI into existing products and explore commercialization opportunities.

However, resource advantages do not equate to product superiority, and the strong do not always dominate.

Over the past two years, two breakthroughs in Chinese large models have sparked global discussions.

In early 2025, DeepSeek-R1 gained global attention for its reasoning capabilities, open weights, and relatively low usage costs. It demonstrated that Chinese teams could not only create competitive models but also reduce the cost of high-performance AI through architectural design and training/inference efficiency optimizations. DeepSeek briefly topped the U.S. App Store's free apps chart.

In July of the same year, Yuezhi Aimi released Kimi K3, a model with 2.8 trillion parameters and a 1 million-token context window, designed for programming, knowledge work, and reasoning tasks. Its launch once again drew global attention and showcased China's ability to advance open-source models.

Of course, this isn't to say that DeepSeek or Yuezhi Aimi's models are necessarily leading; it merely illustrates the unpredictability of competition in the AI industry.

Even companies without significant financial or resource advantages can become market focal points through technological breakthroughs or product innovations.

This is what makes big tech firms feel insecure. In such an environment, product timelines and strategies become secondary; the priority is not falling behind.

Thus, we witness a peculiar phenomenon: AI vendors are simultaneously cutting AI products en masse (e.g., Tencent announced that QClaw, its "nationwide lobster farming" product, would shut down on December 24) while aggressively updating AI models and chasing trends.

Because the benefits of following are tangible.

To exaggerate a bit, Tencent was arguably saved by WorkBuddy this year. In May, Ma Huateng admitted at a shareholder meeting that Tencent lagged behind competitors: "A year ago, I thought we were on board, but then I realized the ship was leaking. Now, it feels like we've boarded but can't sit comfortably."

Today, Tencent has secured the top position in China's desktop office agent market. According to Analysys' Q2 2026 China Office AI Market Report, total AI-native office agent visits on PCs exceeded 60 million in June, with WorkBuddy accounting for 20.97 million visits—more than the second- and third-place competitors combined.

The overseas market is no different. Codex, positioned as a rival to Claude Code, achieved remarkable growth. In August, OpenAI Codex lead Tibo shared on X that Codex had surpassed 20 million active users.

Big tech firms aspire to shape industry directions, but when leading becomes increasingly uncertain, quickly following the pack becomes a prudent choice.

03 Final Thoughts

Warren Buffett has always maintained an extremely cautious stance toward tech stocks.

His concerns are twofold: First, moats are unstable. The tech industry is constantly disrupted; today's leaders may be replaced by new technologies within a few years, making it hard to predict the industry landscape a decade later. For example, Nokia's mobile empire was suddenly destroyed by the iPhone. Second, tech companies burn through cash on R&D in their early stages, making future cash flows extremely unpredictable.

These concerns perfectly describe today's AI vendors.

However, while Buffett can keep his distance from tech stocks, big tech firms have no choice but to confront these challenges head-on.

Reference Sources:

1. Tencent Technology: The Ambition of Agents: Boundary Expansion and Power Grabs | 2026 Q3 AI Trends Report

2. China Economic Net: Doubao Mobile Rejected by Over 150 Apps!

3. Qingcheng Finance: Domestic and Foreign Big Tech Companies Scramble for Muse Access, but Personal Agents Remain a False Proposition

4. David Senra: Creating Muse: The Fastest-Growing AI Product Since ChatGPT

The cover image and illustrations are owned by their respective copyright holders. If the copyright owners believe their works are unsuitable for public browsing or should not be used free of charge, please contact us promptly, and our platform will immediately make corrections.