110: Ming Ventures' Xia Ling on the Agent Gateway Battle | Agent #2
Summary
Xia Ling believes DeepSeek showed how an L2 reasoner can crush an L1 Chatbot from a higher dimension, so every portfolio company must take L3 Agent seriously and make it an explicit R&D priority. Manus still completes only just over 50% of GAIA’s hardest Level 3 tasks, but model companies and practitioners believe that figure could exceed 80% within a year, perhaps even by year-end; the blunter interpretation is that “80% of jobs” or “80% of the tasks within a job” could be covered. If a startup is still selling an old product form by then, “there’s a very good chance it will be crushed by someone else’s L3 Agent product within a year.”
The Agent race is not determined by the foundation model alone; three capability curves—models, orchestration and the tool ecosystem—must rise together. The orchestration layer covers goal instructions, short- and long-term memory, planning and reasoning; the tool layer expands the system’s reach through Function Calling, Computer Use and MCP. OpenAI’s global memory, MiniMax’s exploration of 4M tokens through Linear Attention, and multi-model, multi-Agent Workflows all show that model companies’ product boundaries have moved far beyond “just providing a model.”
The battle for the general-purpose gateway could begin in the second half of this year, with the most intense front on smartphones. OpenAI, Anthropic, Google and Meta overseas, ByteDance, Alibaba and Tencent in China, as well as Li Auto and Meituan, are all pursuing the opportunity; platforms such as Meituan and Didi have no desire to become resources dispatched through someone else’s gateway again. Xia Ling calls this competition “somewhat close to the endgame,” with the outcome hinging on “a first-tier model plus a first-tier position.” Even before models fully break through, the appeal of a To C general-purpose Agent may push major companies into the fight early.
If the next generation of hardware is glasses, AI Agent could become their operating system, further narrowing the choice of gateways. Xia Ling estimates that this shift is “very likely to happen within two or three years”: constrained by space, battery life and performance, a device may keep only one 3B or 7B model resident in memory. His observations on Meta are especially worth noting—its social assets, investment in small models, customized perception and control chips, and possibly a standalone box that bypasses the phone all point to an ambition to make glasses an independent AI terminal.
General-purpose Agents may erode traditional advertising, but subscriptions have not yet proved an efficient enough replacement business model. In Save Time scenarios, people may stop browsing search results, reducing both auction-based ranking and ad exposure; overseas products rely mainly on subscriptions, while domestic general-purpose products are “more or less free.” By contrast, ads in Kill Time content can be embedded in generative context; Xia Ling has seen early conversion efficiency “many times higher than traditional advertising,” though everything still depends on whether the placement feels natural.
A startup’s safe zone is not the model companies’ main line of advance, but professional users, vertical data, proprietary Workflows and the last mile. GPT-4o’s text-to-image capability initially threatens broad consumer tools built around one-shot output, not professional production that requires iterative control; Xia Ling even believes it could accelerate Liblib AI’s evolution into an Agent, using its LoRA aesthetic ecosystem, ControlNet and human-machine feedback to rebuild the Workflow and ultimately take share from Photoshop. Products closer to the general-purpose main line, such as Manus, Perplexity and Kimi, must find differentiation and move toward more precise users and deeper products.
For Chinese vertical Agents, the more viable business model is not selling SaaS seats but taking over the service layer, delivering the outcome and charging a 10% to 30% Take Rate. Aiyu Intelligence is itself a law firm handling bulk litigation for financial institutions, filing 15,000 cases in a single month and taking 30% of recovered funds; hotel Douyin operations and online product-sales Agents could turn roughly 20% of traditional sales costs into about 10% of sales paid to Agent operators. Xia Ling believes companies that can structure non-standard services, reshape business flows and even redesign organizations could still grow into “$100B-class companies.”
The real risk in embodied intelligence is not merely the lack of PMF today, but the absence of a generalizable “GPT-3 moment” governed by a Scaling Law. Hardware engineering, imitation learning and reinforcement learning are advancing, but the industry has yet to converge on the sources and processing of pretraining data or the algorithmic framework; it is currently closer to Technology-Problem Fit than Product-Market Fit. Xia Ling is betting on a breakthrough within two or three years and accepts that commercialization may take three to five years or longer, while explicitly acknowledging: “If we can’t find it, embodied intelligence really will become a huge bubble.”
Deep dive
1. Agent Is a Task System Spanning Virtual and Physical Worlds
Xia Ling places autonomous driving, embodied intelligence and cloud-based software Agents in the same framework: their environments and hardware-software stacks differ, but their core is the same—calling tools, interacting with the outside world and executing a sequence of actions to complete a task specified by a person.
This perspective comes from Ming Ventures’ observation across software and hardware investments. Xia Ling believes that understanding autonomous driving, embodied intelligence and AI applications through the evolution of model capabilities yields insights unavailable from looking at any single field.
2. DeepSeek Turned L3 Agent from a Distant Vision into Every Company’s R&D Agenda
In OpenAI’s five-level framework, L1 is the Chatbot, represented by ChatGPT, Doubao and Kimi; L2 is AI with reasoning capabilities, represented by DeepSeek in Xia Ling’s framing; L3 goes beyond answering questions to complete an entire task by solving a chain of complex problems.
Xia Ling had already invested in Agents in the second half of last year, but DeepSeek was what made every portfolio company take them seriously: “When an L2 product comes out, it can crush an L1 product from a higher dimension.” The natural follow-up is whether a mature L3 will again crush existing L1 and L2 products in the same way.
Agent was therefore not a concept that appeared only after Manus took off in March; Manus merely brought it to the mainstream. The harsher implication for startups is that businesses not designed for a post-L3 environment could lose their competitive dimension within a year.
3. GAIA’s 80% Is a Time Anchor, Not a Capability Already Delivered
Xia Ling uses GAIA to break down current maturity: Level 1 typically uses no tools or one tool and involves no more than roughly five reasoning steps; Level 2 calls several tools in fewer than ten steps; Level 3 coordinates a large number of tools to complete a complex task system. Manus still completes only just over 50% of the hardest tier.
Based on conversations with people at model companies, his preliminary judgment is that leading Agent systems could push GAIA Level 3 above 80% within a year, perhaps even within this calendar year. “80% of human work” is simply the blunt version; it may actually mean 80% of jobs or 80% of the tasks within each job, not a validated employment forecast.
4. Google’s Three-Layer Framework Shows Model Companies Have Moved Beyond the Foundation Model
Xia Ling adopts the structure from Google’s white paper earlier this year: an Agent consists of Model, orchestration and Tools. The model supplies the capability base, orchestration handles goals, memory, planning and reasoning, and tools let the system reach the external world and act.
Pretraining, post-training and capability fusion at the Model layer have not slowed, and multimodal progress is not limited to images. Xia Ling has heard production-grade, rather than demo-grade, voice that approaches human quality, with emotion, accents and even dialects at very low latency.
This also changes the answer to how far model companies will go: Memory, Computer Use, MCP, Google’s extensions and early Function Calling are all parts of model companies’ effort to build L3 products, not natural white space reserved for application companies.
5. Memory Is Both a Technical Bottleneck and a Potential Relationship Moat
For most general-purpose or vertical Agents, native Long Context is still insufficient. The episode mentions OpenAI context windows of roughly 128K and 64K, along with similar lengths for DeepSeek; MiniMax is experimenting with Linear Attention to support 4M tokens while maintaining a very high hit rate on “needle in a haystack” tests.
After OpenAI launched global memory on April 11, Sam Altman even pointed to Memory as a possible next “Scaling Law.” The industry is still guessing whether it relies on Context, traditional RAG or Agentic RAG; Xia Ling believes that if it is merely traditional RAG, the hit rate will certainly be inadequate.
After a founder Xia Ling knew was put through a blind test, the founder immediately asked how the system understood his career and personality. The AI identified him as a CEO and summarized his management style and positive traits, observations that matched Xia Ling’s own view as an investor and produced a genuine Wow Moment.
Manqi therefore points to a potential flywheel: the more an AI remembers about a user, the better it understands the user’s work, habits and relationships, and the harder it may become to leave. Xia Ling agrees that long-term relationships could become a moat, but believes the technical requirements for moving from accurate memory to a stable emotional relationship remain far higher than those for completing general tasks.
6. Workflow Is an Engineering Compensation for Current Model Gaps
When a task is sufficiently well-defined, an Agent may not need the strongest reasoning model; a team can decompose a human SOP into finer-grained serial and parallel reasoning steps, explicitly filling in the tacit knowledge and judgment that humans invoke automatically but models lack.
Multi-Agent systems divide the work further: one Agent tracks overall progress, another handles emotional expression, and a third performs numerical calculations. A real system may also combine DeepSeek or Claude for reasoning with Doubao or OpenAI for emotional output. As models improve, these steps and roles could theoretically be recombined.
Xia Ling warns that discovering a Workflow is initially a product-definition or engineering moat, but it will inevitably weaken as the Agent barrier falls. More durable moats should migrate to platform effects, new business flows or new organizations rather than a reproducible process arrangement.
7. MCP Expands the Tool Ecosystem, but Model Companies May Still Capture the Biggest Upside
MCP was initially introduced by Anthropic to help its Agent use external tools more effectively, then developed into an ecosystem because of its industry value and was gradually adopted by OpenAI. Application companies benefit, but the companies that need the most cross-industry APIs are precisely the model companies building general-purpose Agents.
Manus reportedly integrated 29 tools. During development, MCP was not yet mature, so it relied mainly on engineering methods such as Function Calling. MCP makes it easier for a company like Manus to call more tools, but it does not solve the problem of choosing the right tool from a large number of Servers.
Manqi adds that once the number of available tools exceeds 50, task quality drops noticeably. Xia Ling acknowledges that the problem remains unsolved but expects model selection and invocation to improve gradually; when that happens, the richer the tool ecosystem, the more unavoidable the fight for the general-purpose gateway becomes.
8. The General-Purpose Gateway Is Both a Growth Ambition and a Defensive War for Platforms
Anthropic, OpenAI and Google all want to become general-purpose gateways through their tool ecosystems; in China, the list includes Alibaba, ByteDance, Tencent, Li Auto and Meituan, with Xia Ling also leaving room for Xiaohongshu and Didi. “General-purpose” here does not mean adding an assistant to an existing core business, but answering questions, writing emails, generating images, recognizing images and executing everyday tasks at the same time.
Manqi compares the idea with WeChat: if WeChat itself were an Agent, Mini Programs would resemble callable MCP Servers. The problem is that Meituan, Pinduoduo and Didi worked hard to escape WeChat’s nine-grid menu and have little interest in becoming supply-chain resources allocated and dispatched by someone else again.
“Not opening an API” may not hold forever. Manqi suggests that if Meituan refuses to connect while Ele.me opens up, incremental orders could eventually force Meituan to adjust; Xia Ling agrees this is possible but believes each company will still prioritize control of its own traffic and gateway.
In Xia Ling’s view, the winning formula is “a first-tier model plus a first-tier position.” Google, Meta, ByteDance, Alibaba and Tencent already have ecosystems, while OpenAI, Anthropic and DeepSeek are trying to use leading models to launch era-defining products; lacking either side leaves a company at a disadvantage.
9. The Smartphone Agent War May Start Before Models Fully Mature
The aggressive timing call is the second half of this year, broadly consistent with major companies’ public product schedules and GAIA capability expectations. The first form is likely to be a standalone general-purpose Agent App on smartphones, because the users and scenarios served by Meituan and Didi are naturally mobile and broader than the cloud-based production scenarios covered by Manus, Deep Research and Computer Use.
Manqi asks whether winning the first PMF could become a “blessing in disguise” or a “sweet curse”: the Chatbot may be only a transitional product. Xia Ling believes that whoever truly captures the Agent gateway will occupy “a position somewhat close to the endgame”; what may appear in the second half of this year is not the endgame itself but the opening act of the endgame competition.
The biggest variable remains model capability and who controls it, but the competition may begin even earlier. Xia Ling’s view is that “the temptation of a To C general-purpose Agent is simply too great for major companies and model companies”; positioning may come first, before the capabilities fully break through.
10. If Glasses Break Free from the Phone, Agent Could Become Their Only Resident OS
Xia Ling estimates that the next generation of AI hardware is “very likely to happen within two or three years.” Constrained by space, battery life and performance, glasses may allow only one 3B or 7B model to remain resident in memory; the Agent driven by that model would effectively be the operating system, with fewer gateway choices than on a phone.
Alibaba and ByteDance are also investing in models and glasses domestically, with Alibaba’s related products placed within its consumer-facing information-intelligence business group. Manqi remains skeptical: current experiences are far from replacing the phone, but a truly independent new terminal could open up far more room for startups.
The speculation about Meta is explicit. Although Llama 4 has drawn criticism, Xia Ling recalls that when Meta trained the 7B version of Llama 3, it used roughly 15T tokens of data, while the 70B model used a smaller data volume; he suspects this preference for small models may be related to glasses.
Ray-Ban Meta did not become popular directly because of AI, but it has already demonstrated glasses’ potential as a terminal. Meta is also customizing cameras, perception chips and control chips, and hopes to bypass the phone with a standalone box; Xia Ling therefore believes it has made an unusually early investment on both the software and hardware sides.
11. Save Time Weakens Advertising, While Kill Time May Multiply Ad Efficiency
When an Agent replaces people in searching, comparing and executing, users may no longer browse the original information, and the auction rankings returned by Google or Baidu APIs will not be displayed as-is. Perplexity- or Doubao-style answers could therefore reduce the reach of traditional search advertising, while recommendation systems designed for humans cannot simply be transferred to Agents.
Overseas general-purpose Save Time products currently rely mainly on subscriptions, while domestic products are almost “more or less free.” Xia Ling questions whether subscriptions are truly an efficient monetization model in the AI era: they offer fewer commercial levers than advertising, and domestic products will ultimately need other sources of revenue.
The mechanism may be reversed for Kill Time content. If an ad is naturally embedded in the context of a virtual character or generative chat, Xia Ling has seen conversion efficiency reach many times that of traditional advertising; the trade-off remains between user experience and commercial value, with the quality of integration determining whether users feel the ad has been forced on them.
12. Model Companies Deliver a Newly Minted PhD; Startups Must Deliver a Professional
Xia Ling characterizes a model company’s L3 Agent as “a newly graduated PhD”: broad knowledge, strong reasoning and the ability to call general tools for common-sense tasks, but no particular industry’s methods, workflows, historical experience, data or proprietary tools.
If capability is plotted on two axes—generalization and precision—model companies will continue to absorb general-purpose tasks and may cover 80% of a job. The remaining 20% becomes more specialized and long-tail, with higher customization costs and less fit with model companies’ logic of lowering marginal costs through generalization.
Professional tools and vertical Agents are therefore not impossible for model companies to build; they simply will not usually be their priority. Model companies will sharply lower the development barrier for the last mile, but are unlikely to personally become lawyers, designers and service providers for every industry at the same time.
13. GPT-4o Will First Pressure Broad Consumer Apps, Not High-End Professional Production
GPT-4o’s text-to-image capability is impressive on text consistency and other dimensions, and OpenAI did not initially open an API. Manqi believes the restriction may be intended to make the feature relatively exclusive in the market and use it to drive consumer growth first; she also expects the API may open later.
Broad consumer users want a one-shot result and generally do not care about extensive local detail; professionals know that high-quality output requires repeated control and iteration. On that dividing line, Manqi argues that GPT-4o could hit mass-market text-to-image products such as Jimeng harder than professional platforms such as Liblib AI, a view Xia Ling shares.
The boundary applies to language, images and video alike: as model companies move from L1 to L2 to L3, they will inevitably advance along the general-purpose main line. Startups serving the same broad consumer users with the same one-shot experience face the greatest risk of being crushed from a higher dimension.
14. Adobe’s Selloff Priced a Workflow Rewrite, Not Photoshop’s Immediate Replacement by 4o
Manqi asks why Adobe’s stock fell sharply after GPT-4o was released. Xia Ling believes the market was pricing both the long-term disruption that generative images could bring to Adobe’s existing software and Workflows, and the fact that Adobe’s actual products have underperformed despite its aggressive “AI” positioning.
GPT-4o’s more immediate impact may not be on Photoshop’s core professional users, but on lower-quality-demand segments in Meitu Xiuxiu and Canva. What the market is trading is a higher probability of a credible replacement path, not the disappearance of professional image editing on the spot.
Photoshop is built on layers, filters and a standardized editing paradigm around layers; text-to-image generation does not naturally have the same structure. Xia Ling believes the real long-term threat is a reconfigured production process, not merely a generation button added to legacy software.
15. The More Professional the Image Market, the Smaller the User Base—but the Higher the Price and Paid Conversion
Xia Ling estimates that Photoshop and the Adobe Creative Suite have more than 100M monthly active users globally, with a legitimate subscription costing just over $50 per month; Canva has more than 200M users at roughly $10 to $15 per month; Meitu’s full product line has more than 200M users at only about $2 to $3 per month.
The three products trace a diffusion path from professional to mass market: Photoshop serves professional creators; Canva serves marketers who came of age with social media but never learned Photoshop; Meitu Xiuxiu serves mobile users seeking beautification and light editing.
The broader the audience, the lower the price and paid conversion, and the stronger the expectation of free, one-shot output. Meitu had already raised paid conversion through AI before GPT-4o, but that does not change Xia Ling’s core segmentation: going deep enough on a specific need can open a market, while broad consumer products are easiest for general-purpose models to absorb.
Manqi’s question preserves an important tension: if generative models eventually achieve continuous fine-grained control, the professional boundary will not be permanent. Xia Ling’s defense is not that models can never do it, but that aesthetic ecosystems, professional Workflows and tool combinations will not automatically arrive with the foundation model.
16. Liblib AI’s Path Is to Attack Photoshop with an Agent, Not Defend Text-to-Image
More than half of Liblib AI’s users are professionals, and its primary asset is the aesthetic, style and high-quality reference-image ecosystem built around LoRA. Xia Ling says AI can learn from and borrow aesthetics and styles, and may also create new ones that humans do not understand; aesthetics ultimately still require humans to understand, choose and guide them.
A stronger foundation model is not Liblib AI’s endpoint but its base layer. The platform must add Prompt interaction, ControlNet, human-machine feedback and other tools, potentially even incorporating Photoshop itself as a Tool, and rebuild the full Workflow from style selection to professional delivery.
This Agent form could help users who previously reached only “five or six points” with Photoshop reach “eight or nine points” faster, while also increasing professional designers’ output. Xia Ling therefore sees GPT-4o as positive for Liblib AI: it could accelerate the company’s encroachment into Photoshop’s professional market rather than simply weaken its value.
17. Manus and Perplexity’s Problem Is Not Lack of Growth, but Proximity to the Main Line
Perplexity, Manus and domestic products such as Kimi all sit close to the general-purpose capabilities that model companies are fighting to own. Even at the height of DeepSeek’s shock in February, Kimi and Perplexity continued to grow, though marketing spend may have contributed; the biggest domestic beneficiaries remained DeepSeek and the Yuanbao and Nami products that absorbed its traffic.
Xia Ling watches these figures to test two questions: whether companies close to the main line can continue to thrive, and how they will adjust. Current growth does not eliminate the long-term risk, because model companies themselves are incorporating search, research, tool use and general-purpose execution into their products.
Product companies find it difficult to build lasting moats through technology alone; Manus, Cursor and Perplexity’s early advantages came more from product ideas, precise positioning and engineering lead time before models were Ready. Xia Ling’s direction is explicit: “Depth matters more than breadth”; the target user must be more professional and more precisely defined.
18. China’s Vertical Agent Opportunity Is Not the RPA Boom or a Patch on Legacy SaaS
After DeepSeek became popular, many executives pushed automation out of technology anxiety. The most immediate beneficiary was often RPA, because executives may not distinguish AI from RPA as long as process automation alleviates the anxiety.
Growth-stage SaaS companies such as Kujiale, Sensors Data and Yikuai can all build AI teams and layer model capabilities into existing products. But if the business logic and Workflow remain unchanged, this is merely “adding a patch”—an upgrade for an incumbent company, not an opening for an AI-native startup.
A genuine startup opportunity must first reshape the original business process; otherwise the legacy software vendor can simply add AI. It should ideally enter an industry that is long-standing, fragmented and difficult to structure, using AI to standardize and scale non-standard work. Finally, it must start with a “low-hanging fruit” that current technology can actually reach.
Xia Ling divides vertical opportunities into two types: one is Save Time, serving only professionals or vertical scenarios; the other is Kill Time, using new content and new experiences to compete for consumer time. Both must avoid the direct advance path of general-purpose models.
19. The US-China Difference Is Not Only How Companies Charge, but Whether the Agent Assists People or Replaces a Business Layer
US companies are willing to pay for efficiency. Harvey handles contract review, while sales Agents act as SDRs; they typically charge by seat or usage. An AI SDR seat costs roughly $2,500 per month, while Harvey charges by usage and is not directly tied to how many contract-review staff a law firm employs. These products serve mainstream businesses and assist professionals; they do not directly take responsibility for the final result.
Chinese companies pay in smaller numbers and have less capacity to pay, making them more willing to pay for outcomes than efficiency. Acceptable delivery generally takes only two forms: providing information or leads that can be processed further, or taking over the entire business and delivering the result directly.
Chinese vertical Agents are therefore better suited to full-service outsourcing and a Take Rate of roughly 10% to 30%. They are not making professionals more efficient; they are “replacing or directing people,” giving non-experts the ability to execute professional tasks.
This explains why directly copying the US SaaS model into China often goes wrong: for the same legal or sales use case, overseas companies sell tools and seats, while a domestic startup may become the law firm, operations agency or channel intermediary itself and realize value through revenue sharing.
20. Manbang Shows That the Lowest-Hanging Fruit Can Be Narrow and Still Become Huge
Xia Ling sees Manbang as a standard example of the previous generation of industrial internet: it initially digitized the small blackboard in an information office, creating a truck-and-freight matching platform that delivered one useful piece of information to trucks temporarily seeking cargo and cargo temporarily seeking trucks.
The entry point did not attempt to rebuild the entire transportation chain; the information itself was the customer’s desired outcome. The company encountered many challenges as it expanded into other businesses, but the core matching need was sufficient to build a large-scale platform.
The lesson for Agent founders is not that fewer features are always better, but to first find one step that current technology can deliver reliably, with a clear enough value loop. Companies that ultimately emerge from each wave often begin with a narrow, hard-demand entry point like this.
21. SHEIN’s Moat Came from Reshaping the Organization, Not Selling Digitalization to Factories
The apparel industry has seen many informationization, industrial internet and SaaS projects, but getting factories with 100 workers and low-education owners to proactively purchase systems is a “particularly painful, particularly bad” business. The company that actually completed supply-chain digitalization was SHEIN.
SHEIN was not selling software. It wanted small-batch, fast-turn production and lower inventory, using real orders to persuade upstream factories to accept information-system upgrades, then connecting front-end promotion data with production progress.
Xia Ling emphasizes that SHEIN integrated marketing and supply through digital means, reshaped the existing business flow and created a new organization. A standalone factory SaaS or traditional informationization company cannot replicate this structure; that is a higher-order moat than an engineering Workflow.
For a vertical Agent company, the best outcome is likewise not to add an assistant to an old link, but to reorganize a service that previously could not be standardized. If surrounding competitors remain small, fragmented service providers, the new company can use technical density to achieve an order-of-magnitude advantage in labor productivity.
22. Aiyu Intelligence Is Not Selling an Agent to Law Firms; It Is an AI Law Firm
While Harvey serves roughly 700 law firms with contract review, Ming Ventures-backed Aiyu Intelligence directly established a law firm serving financial institutions in bulk civil and commercial litigation, focusing on debt collection, breaches and cases that traditional firms avoid because the amounts are small or the matters are old.
The company filed 15,000 cases in March alone, while even China’s largest traditional law firms may handle only several hundred lawsuits per month. Xia Ling estimates that its efficiency is 100x or more than that of a traditional small law firm.
The business model is not a seat fee but a Take Rate of roughly 30% after recovery. Xia Ling emphasizes that AI can guide non-experts, even people without legal training, through large parts of the process; this is not “making lawyers more efficient,” but rebuilding the production structure of a law firm.
23. A Sales Agent Can Charge on Closed-Won Revenue Only If It Takes the Entire Channel
Overseas SDR Agents identify opportunities and send customers into the downstream sales process, so they cannot take a cut of final sales and can charge only by seat or usage. They improve front-end efficiency but do not bear responsibility for closing.
China already has companies running full-Agent Douyin operations for hotels and completing the end-to-end sale of online courses and digital products. They look more like channel agents, taking over the entire sales layer: if traditional sales costs account for roughly 20%, an Agent operator might charge about 10% of sales.
The model has an upper bound: the inventory and traffic belong to someone else, so platforms or suppliers may constrain the company as it grows. Teams therefore hope to keep improving through Post-Training, so that when general-purpose Agents all reach a 10% conversion rate, they are already at 13% or 15%; Xia Ling does not treat this as a permanent moat already established.
24. “Rules of the Human World” Can Make Apparently Standard Agent Scenarios Unworkable
The US already has Agents that automate bookkeeping and closing the books, but Chinese companies may operate with “two sets of books.” Xia Ling calls these informal rules, subjective judgments and non-technical constraints “the rules of the human world”: whether AI should participate in the second set of books is not a question technology can decide by itself.
Manqi summarizes the boundary with an old joke: “AI cannot replace lawyers, AI cannot replace accountants, because AI cannot replace people in going to prison.” Xia Ling’s conclusion is that technology should first enter problems that can be standardized and structured, rather than assuming every manual process is worth automating.
Traditional law firms and operations agencies could theoretically use Agents, but shifting from organizations driven by professional knowledge and sales relationships to ones driven by technology is difficult. Small, fragmented competitors usually lack sustained engineering, R&D and product capabilities, leaving an opening for AI-native companies.
The right founder needs to understand both the industry and AI. Aiyu Intelligence’s founder comes from finance, understands customer scenarios, reads papers personally, has a core technical team and works closely with model companies such as Qwen; Xia Ling believes opportunities like this will appear in batches, but the window for low-hanging fruit is time-sensitive.
25. Kill Time Agents Start with New Content, Not Mature Companion Relationships
In virtual-character platforms such as Xingye and Zaomeng, every character can be viewed as an Agent, but Xia Ling is reluctant to call it companionship. Current technology can provide instant emotional value, often through role-play, but still cannot reliably provide a relationship requiring long-term understanding, memory and trust.
The real novelty is high participation: “Creators are also consumers, and consumers are also content creators.” Every user input changes the content. Compared with the choice branches in games such as “Love Is All Around” or “Detroit: Become Human,” the degree of freedom is far higher.
The weakness also comes from participation: text chat is less immersive than video, while typing and imagining consume more effort than scrolling Douyin. Retention is therefore currently stronger among certain core and niche users, but this has not yet become a mainstream mode of content consumption.
Multimodal characters and interactions, more plot-driven structures and better choice design can raise immersion and lower the input threshold. Balancing freedom with low friction, however, requires stronger reasoning. Xia Ling believes model progress will continue to raise penetration, giving this category a chance to become a genuine new content platform.
26. New Content Platforms and Single-Relationship Agents Are Two Paths That May Eventually Converge
One class of company accepts current technical limits and uses many different Agents to supply content, with creators and consumers jointly forming a platform; Xingye and Zaomeng fit this model. Only by creating a new content form can a company escape existing distribution channels such as Douyin, Fanqie and Qidian.
Another class offers only one or a small number of Agents, hoping to serve everyone through highly personalized memory. For an individual user, the time, mood and emotion invested ultimately produce not a work, but “the relationship itself.”
Manqi compares the long-term form with Her: if a user trusts a long-term relationship Agent, the Agent could also shop, handle tasks and act as a secretary, merging the Kill Time and Save Time gateways. Xia Ling agrees with the ultimate direction but believes building deep emotion and trust is much harder than delivering a clearly defined general-purpose task.
Manqi proposes another evolution: a smartphone general-purpose Agent first connects with users as a “heartless worker,” then deepens the relationship through models, memory and interaction. Xia Ling agrees this is possible and emphasizes that the product experience must keep pace with the latest models; otherwise, the arrival of the next generation will create an obvious gap between the product experience and the new model.
27. Embodied Intelligence Has Yet to Reach a GPT-3 Moment for Scalable Generalization
Xia Ling agrees with parts of Zhu Xiaohu’s description of embodied commercialization but rejects the conclusion that the sector is not worth backing. Over the past few years, the engineering of bipedal and quadrupedal hardware has matured visibly, dexterous hands may continue to improve, and combining imitation learning with reinforcement learning has raised intelligence levels.
The unresolved issue is pretraining: the industry still does not know where to obtain massive datasets, how to process them or which algorithms can make capability generalize across scenarios through a Scaling Law. In other words, embodied intelligence has not even reached the GPT-3 Moment at which large language models demonstrated a path to generalization.
The greater danger is therefore not that the market lacks demand today, but that the industry ultimately becomes another generation of vertically integrated robotics vendors. Integration businesses in China are difficult; without large-scale generalization, the commercial value technology can deliver will diverge fundamentally from expectations.
28. Embodied Investment Bets on TPF, and Open-Source Analysis Must Find the Underwater Flywheel
Xia Ling divides technology commercialization into four stages: scientific discovery, technological breakthrough, engineering replication and product deployment. Embodied intelligence is currently closer to the transition from technology to engineering, or Technology-Problem Fit, than to Product-Market Fit; niche PMF may exist, but broad commercialization could still take three to five years or longer.
He invested in LimX Dynamics in August 2022 by extrapolating from the results of MiniMax’s second-generation end-to-end, data-driven model in text. The instinct among first-line practitioners is that embodied intelligence may find its GPT-3 Moment within two or three years. Xia Ling admits that if it cannot, embodied intelligence will become a huge bubble.
In identifying early opportunities, he cares more about whether a founder can grow faster than the AI industry: “Three months later, you talk to him and there has been no significant change” often indicates insufficient growth speed. Revisiting “Fifteen Years of Boiling,” “Nine Losses, One Win,” “The Android Story” and “同行开源” is an attempt to identify, through history, the organizations and flywheels that truly win in a technology wave.
Open source alone does not guarantee victory. Open-source databases rely on cloudification, upstream and downstream adaptation and deployment costs to form a product flywheel; Android also depended on Google’s core applications and advertising incentives to attract developers. DeepSeek’s open source can generate technical influence, recruiting advantages and compatibility with China’s heterogeneous infrastructure, but community adoption will not automatically train the next-generation model better—“the reasons for winning are actually all underwater, beneath open source.”