Vol.59 2025: The Year of AI Agents—Where Is the Industry Headed? — Kuangxi Podcast Festival: Er Kuang

Summary

• The short-term capabilities of Agents are overestimated, while their long-term value in reshaping software architecture may be underestimated. For now, they are still automation layered with model-based understanding, action, and tool use, while ordinary users have not even moved beyond the Chat interface; but 小狼 has revised his view from a year ago that Agents were “merely an extension of automation.” Enterprise IT may eventually stop stacking layers of systems horizontally and instead consist of a group of vertically integrated Agents, “really just a useful group of engineers.” 小狼 places the possible inflection point around 2030, while explicitly acknowledging that “it’s hard to say exactly when this can happen.”

• Open source is shifting from a technical-route debate into an industry accelerator, while China is leaning more heavily on open source and ecosystems. 庄明浩 believes the capability gap between open and closed models has become “negligible”; 小狼 summarizes the China-US difference this way: US vendors sell services to the Fortune 500 through large BD teams—“they’re already making money, and we’re behind them flipping the table.” DeepSeek, 千问, and their ecosystems mean application-layer vendors no longer need to bet on a single model and could reproduce the internet era’s “overtaking on the bend,” but Chinese companies are more likely to monetize through private deployment, appliances, and project procurement than pure SaaS.

• MCP is not primarily about technical difficulty; its significance is building the connectivity, transaction, and settlement infrastructure for a Multi-Agent world. 朱峰 uses hotel booking to explain the value: a front-end Agent checks availability and calls the payment service, subsequent Agents communicate among themselves, and the user receives only the result; 小狼 warns that China’s mobile internet super-apps have created natural “reproductive isolation,” so an MCP network will remain fragmented unless platforms open up. The real variable is therefore how platforms divide traffic and revenue—and whether a new entry point could emerge that bypasses today’s incumbents.

• General-purpose Agents currently resemble interns who cut corners and run over budget; genuinely investable products still require vertical workflows, trustworthy data, and human review. 庄明浩 asked Manus to analyze how many of the 50 websites and 50 Apps on A16Z’s list were Chinese companies: it initially copied the official article and returned 7 or 9; after being challenged, it raised the number to the teens; when asked to verify each company, website, news report, and business registration, it eventually said, “context limit exceeded.” 朱峰 therefore rejects treating general-purpose Agents as a “silver bullet”—the more complete a search report looks, the easier it is for errors to be packaged as answers, so users must retain “the ability to be a contrarian.”

• AI Coding may be the first application in this cycle to deliver meaningful interim results, but its value depends heavily on code quality and task complexity. Cursor represents Human-in-the-loop, while Devin aims to become “the developer” end to end; at large companies with mature engineering standards, it can already deliver consistent productivity gains, while generating more code in systems buried under technical debt only produces “noodles stacked on noodles, turning into silver-thread rolls.” For individual developers and one-person companies, landing pages, websites, front-end design, and decomposed microservices are already highly usable; asking it to build a large full-stack project in one shot can still mean “two days of writing, two weeks of debugging.”

• The startup narrative is shifting from “make one big bet” toward one-to-three-person cash-flow workshops, with AI as a capability multiplier rather than the moat itself. The lesson 小狼 took from Hangzhou was that “there’s nothing shameful about doing business,” while 朱峰’s venture studio even prefers projects with “no more than three people”; both place more weight on clear demand, fast cash collection, and founders’ existing industry expertise. A thin layer of AI features will soon be absorbed by models or Big Tech; the stronger model is to turn domain experience into a solution and use AI to scale it.

• Autonomous driving shows the hardest part of moving from “language to behavior”: once autonomous execution enters the physical world, responsibility, regulation, and scenario boundaries become part of the problem. 于欣烈 notes that OpenAI’s L3 Agent, mapped onto SAE levels, only makes sense starting at autonomous-driving L3, while autonomous-driving L4 still belongs within the Agent layer; human-vehicle co-driving is Co-pilot, not Autopilot. Chinese manufacturers are betting on VLA, putting Vision, Language, and Action into one model, but “nobody knows” whether this route can reach L4; the US may bypass L3 and directly validate unsupervised FSD and Robotaxi.

• Nvidia’s core question is not whether compute demand will grow, but whether it will continue to accrue exclusively to Nvidia. Agent workloads are described as consuming thousands of times the compute of Chat products, and falling costs could stimulate total demand; but slower training growth, competition in inference chips, and automakers developing their own silicon all make the binary narrative of “taking off from here or collapsing completely” misleading. 于欣烈 mentions that NIO ET9 has launched, XPeng plans to put its Turing chip on the road in Q2, and Li Auto plans to release its “Schumacher” chip the following year; however, intelligent-driving chips account for a small share of Nvidia’s revenue, so a falling automotive share does not imply a reversal of the overall AI-chip thesis.

Deep dive

1. Agents Are Not New Magic; Automation Is Starting to Understand and Act

• 小狼 places Agents in the broader history of technology: large amounts of automation engineering already existed during the digitalization and informationization phases. What is new today is that models can “infer what you need” and then use accounts, tools, and Actions to do some things on their own; the underlying technology has not suddenly become completely unfamiliar.

• 于欣烈 offers a stricter product test: the capability must rely on a large model, execution must be Autonomous, and the system must be able to continuously call external tools. Tools that merely assist users and wait for instructions at every step do not qualify as Agents under his definition.

• What this cycle has truly expanded is participation. 小狼 says AI has moved “one layer up” from infrastructure for Infra and engineers into applications and business, expanding the conversation from technical teams to every organization and ordinary users.

2. The Current Gap Explains “Overestimated in the Short Term, Underestimated in the Long Term”

• 朱峰’s field survey is telling: many people use AI at work, but once those who only chat in 豆包 or DeepSeek are excluded, the number of users who can build or deeply use Agents drops sharply. “In the short term, I think its capabilities are overestimated; in the long term, they’re underestimated.”

• Kuangxi Podcast Festival’s own Agent also exposed the boundary: it can check the course schedule, but depends heavily on external data quality, the knowledge base, and the underlying model. The product world’s daily pursuit of buzzwords such as Agents and MCP does not mean ordinary users have crossed the gap of configuration, judgment, and verification.

3. 小狼 Revises His View: Agents May Rewrite the Shape of Enterprise IT

• A year ago, 小狼 saw Agents as an extension of automation, mainly handling efficiency problems for engineers; a year later, he voluntarily “slapped his own conclusion in the face,” arguing that what had been underestimated was the system depth Agents might ultimately take on, not just their ability to perform individual tasks.

• Traditional IT architecture “grows horizontally”: middleware, real-time audio and video, and various business layers are stacked on top of the base layer. 小狼 imagines a future architecture that may “grow vertically”—a single UI at the front, with a group of Agents underneath, each running a closed-loop business and able to understand requirements, all built on the cloud.

• This is not a near-term promise. 小狼 places a rough timeline for a more mature form around 2030 and repeatedly stresses that “it’s impossible to say how far away it is”; but if the paradigm holds, building Agents could become an engineering activity that almost anyone can participate in.

4. Open Source Wins on Collaboration Speed, Not Just Model Weights

• 庄明浩 recalls the industry’s reversal in perception: a year and a half ago, leading companies broadly believed that heavily funded closed models were stronger; after DeepSeek’s rise, he considers the efficiency and score gap between open and closed models “negligible,” and open source has gradually become the consensus.

• 朱峰 emphasizes that open source is not merely a narrow software-licensing question, but a community mechanism for building together. Vendors can combine multiple models such as DeepSeek and 千问 on the back end instead of betting on one supplier; model matching, engineering integration, and ongoing community contributions will jointly accelerate the formation of Agent networks.

• 小狼 describes engineers’ motivation bluntly: on one hand, there is the ideal of “breaking monopolies”; on the other, there are Reputation, higher pay, and entrepreneurial opportunities. Open collaboration is not one discrete technology, but a mode of production that allows demand, dependencies, and surrounding projects to keep growing.

• 小狼 remains speculative in his assessment of DeepSeek: its goal may not be to become a large-scale commercial model company immediately, but to first secure an ecosystem position; once it is established, “everyone will know how to make money with it,” while the entire domestic application industry is lifted along with it.

5. China and the US Are Commercializing AI on Two Different Infrastructure Paths

• 小狼’s summary is: “They’re already making money, and we’re behind them flipping the table.” The US has a large number of Fortune 500 headquarters, allowing companies such as OpenAI to deploy large BD teams to sell and implement services inside enterprises; Chinese vendors are more likely to train their own models and stockpile chips, then build another ecosystem through open source.

• 朱峰 points out that Chinese companies prefer “自主可控一体机” when buying IT: GPUs, models, and applications are installed inside the enterprise. Purchased equipment becomes an asset and data remains under the company’s control; cloud and SaaS are often booked as expenses and also raise concerns about data ownership and security.

• As a result, revenue from Agent services in China may come more from equipment purchases, private deployment, and project implementation. Open-source models are also often installed in appliances to match Chinese enterprises’ delivery preferences.

6. The Failure of the Previous SaaS Cycle Sets the Valuation Backdrop for Enterprise AI

• 庄明浩 recalls that after 2010, the main TMT narratives in China and the US began to diverge: China grew through mobile internet and algorithms, while the US also built a To B ecosystem around Cloud, PaaS, and SaaS. The To C group at his dollar fund “made a fortune,” while the To B group was virtually “wiped out.”

• Beisen was the example he remembered most vividly: the HR SaaS company may have raised tens of billions of yuan in the primary market and may once have been valued at several billion dollars; its financing terms specified a qualified IPO price of about HKD30B, below which it could not list. 庄明浩 says the stock fell about 50% on its first trading day, another roughly 50% the next day, and was left with only about HKD3B a week later.

• This does not prove that enterprise software will never have opportunities. The guests expect AI, Agents, and delivery models better suited to local procurement habits to potentially create a different opening for China’s To B market than traditional SaaS did.

7. MCP Builds the Agent Network; the Commercial Moat Is Harder Than the Interface

• 朱峰 uses hotel booking to illustrate MCP: a user asks for a hotel near a cinema, the model calls a booking interface to obtain availability, then calls a payment interface to complete the transaction; the Agents communicate among themselves, while the front end gives the user only the options and final result. “AI can now call AI.”

• MCP itself is not particularly deep. Its real value is connecting multiple Agents and eventually carrying clearing and settlement. It serves the next stage of Multi-Agent systems, rather than adding a polished button to an individual chat window.

• 小狼’s reservation is that domestic Apps, official accounts, and mini-programs have long tried to do everything in one place, creating natural “reproductive isolation” between platforms. Opening up would mean redistributing traffic and revenue; if platforms cannot agree, the tool-calling chain will remain fragmented, just as Apps are today.

• 庄明浩 suggests that a new giant or entry point could emerge, using a new standard to bypass existing internet players; 小狼 responds with the example of WeChat displacing Baidu, but this remains a possibility, not a firm prediction.

8. The Technology Narrative Can Be Reduced to “From Language to Behavior”

• 庄明浩 compresses more than two years of evolution into one phrase: “from language to behavior.” Large Language Models first solved language and conversation; Agents, Multi-Agent systems, tool use, and MCP are now trying to turn understanding into executable Action.

• He cites a whiteboard shown in a news segment on “the development of next-generation AI agents”: longer-context memory, Multi-Agent communication, communication protocols, and expanding the task horizon from one person’s work over several hours to the workload of several people over a week.

• In the OpenAI grading framework he cites, L1 is Chatbot and Agents sit at L3; the intermediate stages correspond to the continued evolution of model capabilities and task forms. The labels are numerous, but the shared goal is for models not merely to answer “how to do it,” but to keep doing the work until it is finished.

9. The Physical World Is Evolving in Parallel with LLMs Along the Vision—Action Route

• 于欣烈 cautions that Agents do not exist only in the LLM world. Cars and robots initially relied on visual perception, moving from Vision to Action; this line has evolved in parallel with language models moving from Language to Action.

• He says models in automotive and robotics are moving from Vision-Action toward end-to-end systems. The next step, VLA, adds Language between Vision and Action, putting all three capabilities into one training framework.

• A traditional Vision Language Model first has to identify a road sign and then translate it into natural language for the driving model, which is poor for both latency and accuracy. VLA aims to exchange information internally, avoiding repeated translation between two models.

10. “General-Purpose” Belongs to the Whole; Each Agent Still Needs to Be Specialized

• 于欣烈 questions “general-purpose Agent” as a product name: generality already comes from the foundation model, while an Agent is closer to a productized specialized workflow. Deep Research focuses on research, Coding on generation, and even Manus combining the two does not mean it truly covers every task.

• 小狼 believes current general-purpose products primarily serve individual users’ learning, search, and personal growth. As foundation models improve, these generalized capabilities will increasingly be absorbed by the models themselves, and standalone general-purpose Agents may ultimately look more like integration tools.

• 于欣烈 uses an analogy from the old world: the large model is the operating system, and the Agent is the application. One vertical Agent solves one type of need; enough specialized Agents combined together create generality in the aggregate.

• This is also where the Agent’s tension lies: it needs the foundation model to become stronger to improve execution quality, yet fears that a sufficiently strong model will absorb its function outright. Product value therefore cannot rest solely on a one-time capability gap in the model.

11. A Polished Report Is Not a Reliable Answer; Users Must Learn to Push Back on Models

• 朱峰 uses general-purpose Agents less precisely because the results “look very good” but are often wrong. AI search can generate a complete report, but users cannot open every citation one by one; when he needed to verify the facts, he ultimately returned to traditional search and checked multiple sources.

• The same problem appears in Coding: a developer may finish in two days what previously took a week, then spend two weeks debugging it. Generation speed rises sharply, but strange logic and uncontrollable behavior created by hallucinations can swallow the time saved.

• 小狼’s advice to current users is to develop “the ability to be a contrarian”: keep asking follow-up questions, criticizing the model, and even PUAing it into checking whether it is talking nonsense. The exercise improves the output while reminding users every day that the answer is not inherently correct.

12. The Manus Test Shows That an Agent Most Resembles an Intern Who Cuts Corners

• 庄明浩 chose a task that could not be solved by a simple search but was not completely unbounded: analyze the 50 AI websites and 50 AI Apps on A16Z’s latest list and determine how many belonged to Chinese companies. Because the list had just been released, no in-depth external analysis yet existed.

• Manus initially extracted only the companies named in A16Z’s article and returned 7 or 9; after being told that this could not possibly be so few, it relied on news searches and produced a number in the teens. Each time, it chose the least effortful path closest to an existing answer.

• 庄明浩 ultimately broke down the process explicitly: list every website name and App icon, identify the company behind each one, and verify its country using the official website, news reports, and business registrations. Manus immediately returned “context limit exceeded”—exactly like training an intern, except interns usually do not use that excuse to stop working.

• 朱峰’s takeaway is that users must “learn to act like managers”: the task objective, delivery standard, and path must all be clear. If users continue communicating like passive executors, “the two of you are both just wage slaves,” and the result will not improve automatically.

13. AI Coding Has Moved from Autocomplete to End-to-End Workflows, but Engineering Quality Sets the Ceiling

• The products were framed as several stages: GitHub Copilot is closer to a half-finished component on a platform; Cursor emphasizes Human-in-the-loop; Devin’s vision is to complete workflows such as building and deploying websites end to end, positioning itself as “the developer.”

• 小狼 does not see the vision as new: when he worked on a code-hosting platform in 2013, engineers were already imagining another “wage slave” writing code on their behalf. The real change is that today’s products are smoother than early copy-and-paste tools and can now enter mature engineering systems.

• At large companies with strong code conventions, review rules, and system stability, AI Coding is relatively mature. Models can improve productivity on top of an established engineering taste, standards, and quality foundation, making the efficiency gains easier to realize.

• If a team is already buried in technical debt, AI will only keep producing technical garbage. 小狼’s metaphor is: “stack noodles on noodles, and you get silver-thread rolls.” Whether to adopt the tool depends first on whether the CTO, architecture, and codebase are clean enough.

14. Small Teams Can Already Capture the AI Coding Dividend, but Complex Projects Still Need to Be Decomposed

• 于欣烈 sees individual developers and entrepreneurs as the most direct beneficiaries today. It is unrealistic for a small company to employ a full-time designer, while AI can already produce increasingly modern front-end pages, fit them into frameworks, and connect to the back end when the documentation is clear.

• The best tasks are landing pages, temporary websites, domain-registration pages, and clearly bounded services; complexity remains the main constraint. Trying to have AI build a complete large-scale project in one shot quickly sends the front end, back end, and state management out of control.

• 于欣烈’s practical rule is decomposition and modularization: have AI write individual microservices, then let humans connect them. The model is not a replacement for architecture; it fills gaps for independent developers in design, front-end work, or a particular programming language.

15. Coding Tools Will Also Move into Writing, but the IDE May Not Exist Forever

• 于欣烈 notes that code is also a language, which means an IDE can become a writing tool. He mentions Deep Cloud: first use DeepSeek’s CoT to break down the problem and writing requirements, then hand the chain of thought to Cloud, which is better at writing or Coding, to generate the article.

• 朱峰 expects all kinds of IDEs to integrate AI in the short term, with products such as Cursor—built by modifying existing editors—becoming the entry point.

• The long-term question is this: if a user only needs to say, “Change this graphic from green to blue,” and AI makes the edit directly, why write code at all, and why would a PRD still be needed? The guests immediately added that this stage is “still a long way off.”

16. “Wrapping” Is a Process of Discovering Demand Together; There Is No Need to Start with a Super-App

• 小狼 compares today with his tool startup in 2015: the technology, scenarios, and dependencies are all still unsettled, so entrepreneurs can “keep wrapping” indefinitely, move on to the next company after failure, and accumulate professional judgment about scenarios, ecosystems, and business models through repeated attempts.

• 小狼 does not see wrapping as derogatory, because “everything is wrapping.” Early product experiments reveal what is missing underneath, what users actually want, and which niche scenario is sufficient to form a closed-loop business.

• 小狼 also warns that while everyone is competing around large models, some are genuinely competing on capability and others may simply be competing on concepts. The application layer is closer to users, and continuous experimentation makes it easier to discover demand.

• 小狼 has consciously abandoned the super-entry narrative of building “the next WeChat” or “the next Toutiao.” The question she leaves entrepreneurs with is more basic: “Ask yourself every day: what problem are you solving?”

17. AI Is Pushing Entrepreneurship from “Make One Big Bet” toward Small Businesses That Can Collect Cash

• 庄明浩 summarizes the shared narrative of the previous mobile-internet and VC cycle as “make one big bet.” In this AI cycle, personal capabilities are amplified, significantly increasing the viability and odds of success for one-person companies, small but excellent products, and low-headcount teams.

• After studying e-commerce, 小狼 found that having “a product in hand” and understanding distribution channels can quickly produce feedback on sales, profit, and cash flow. Software has a shorter industry chain and thinner business models, while Big Tech controls too many links, leaving small companies “dirt poor.”

• 小狼 hopes that as China improves its industrial software, infrastructure, and position in the industrial chain, it will produce more value chains long enough to resemble e-commerce, allowing experts to build small businesses in familiar industries rather than fight for a single platform entry point.

• After moving from Shanghai and Beijing to Hangzhou, 小狼 learned the phrase: “There’s nothing shameful about doing business.” People there care more about costs, sales, profit, and supporting a family than about first packaging everything as a grand entrepreneurial narrative.

18. A One-to-Three-Person Workshop Can Work, but a Thin Feature Has No Moat

• The projects 朱峰 and the 科技乱炖 team invest in are often “workshops”: “Ideally, it’s one person; don’t exceed three.” The first metric is whether the business can generate cash flow, not whether it should immediately expand the team, raise financing, or pursue a higher valuation.

• His challenge still comes from traditional investing: if a company merely uses AI to build a very thin Feature, a foundation model or Big Tech will quickly embed that capability, and the original product will immediately lose its value.

• The better target is a founder who already has experience, capability, and accumulated knowledge in an industry, with AI then used to improve efficiency or scale the operation. The moat comes from past capability and accumulated knowledge; AI amplifies them rather than replacing them.

19. SAE Levels Remind the Market That an Agent’s “Autonomy” Cannot Be Smuggled Across Definitions

• 于欣烈 says intelligent-driving professionals generally dislike SAE’s mechanical grading system, but cannot do without it, because L1–L5 is the easiest language for compressing consensus when communicating with markets and users.

• Under a strict definition of “autonomous execution,” driver assistance does not qualify as OpenAI’s L3 Agent. Only from autonomous-driving L3 onward does the system begin to complete driving independently; autonomous-driving L4 still belongs to the OpenAI Agent layer, not a higher intelligence tier.

• The distinction can be compressed to: “When humans and cars drive together, it is Co-pilot; it is not Autopilot.” This is the responsibility boundary that a language-world Agent must confront as soon as it enters the physical world.

• The slow rollout of L3 is not only a technical issue, but also a matter of regulation and accident liability. Whether responsibility for a collision falls on the driver, automaker, or system determines whether the product can formally cross from “assistance” into “autonomy.”

20. China Is Betting on VLA; the US May Directly Validate L4

• 于欣烈’s view at the time was that Huawei could be the first in China to achieve “conditional autonomous-driving L3” this year in limited scenarios such as highways. Regulatory language was also tightening, making it unacceptable to use “intelligent driving” or “autonomous driving” casually; Tesla FSD in China was likewise renamed within the driver-assistance framework, while Huawei could use terms such as “intelligent driver assistance” and “conditional autonomous driving.”

• The US route has no obvious L3 transition. Tesla may move directly toward unsupervised FSD and planned to launch Robotaxi in Austin in May or June that year, turning L4 directly from a passenger-car function into a mobility service; this was also one of the important reasons behind the change in its market value.

• China had long followed Tesla’s publicly disclosed roadmap, but Tesla no longer holds public AI Day events on its technical route, leaving domestic companies to make their own judgments. Li Auto, XPeng, and others have chosen VLA, and the combination of technical sophistication and product experience may develop differently.

• This is not universally regarded as the optimal route. 于欣烈 relays 李想’s view that VLA may not be efficient, but is an acceptable practical solution for “not standing still”; whether it can ultimately achieve L4, “nobody knows.” Most hotel robots and delivery vehicles still rely on path planning, rules, and path following, which cannot be used to claim a leap between levels.

21. Nvidia Remains the Biggest Player, but the Inference Era Will Weaken Expectations of Total-Market Dominance

• 庄明浩 summarizes the simple strategy from the end of 2022 through the end of 2024 as “buy Nvidia.” After DeepSeek emerged, the debate became whether greater training efficiency would weaken chip demand or, in the manner of Jevons paradox, stimulate more usage by lowering costs.

• Agent workloads were described at the event as consuming thousands of times the compute of Chat products, so the logic of total demand growth has not disappeared. 于欣烈’s key question is: “Can that demand actually land on NVIDIA?” Future incremental demand will come mainly from inference, and inference chips may no longer be monopolized by one company.

• He rejects the binary view that Nvidia will either “collapse completely or continue taking off from here.” Nvidia may remain the largest player without maintaining its previous growth rate; valuation must account simultaneously for market share, growth expectations, and whether the existing PE is too high.

• Alternatives are also emerging in automotive: issues related to the Blackwell architecture have repeatedly delayed the launch and vehicle deployment of Thor, while the Thor U solution, with a peak of about 1,000T, still faces cost and validation problems. NIO ET9 has launched, XPeng plans to roll out its Turing chip in Q2, and Li Auto plans to release its “Schumacher” chip the following year. 于欣烈 emphasizes, however, that intelligent-driving chips account for a small share of Nvidia’s revenue, so automotive share cannot be used to extrapolate an overall conclusion.

22. AI Deployment Is First Held Up by “Dirty Data Designed for Human Eyes”

• The Agent 朱峰 built for Kuangxi Podcast Festival was essentially a knowledge base made with 扣子; the hard part was not building it, but reorganizing course schedules—easy for humans to read but difficult for machines to understand—into CSV or JSON.

• Large amounts of guest information and event details existed only in images, with no Word versions, so the Agent naturally could not find them. 庄明浩 then said AI also needs to grow up and gradually recognize that “data in human society is simply this dirty.”

• This makes data structuring, knowledge bases, databases, and internal enterprise governance foundational work for Agent deployment. No matter how strong the model is, if its inputs remain scattered images and unverifiable documents, the product will remain stuck at the demo stage.

23. The China-US Gap Is a Full-Industry-Chain Issue; China’s Opportunity Comes from Different Soil

• 庄明浩 believes the “China-US AI capability gap” is not just a gap in AI technology itself, but more a gap across the entire industry chain: the US has mature software, cloud, and SaaS business models, while China has open source, industrial restructuring, and the conditions for “overtaking on the bend.” The outcome depends on the full upstream and downstream chain.

• 庄明浩 sees clear tailwinds in policy and, based on past experience, expresses an overall optimistic but non-precise forecast: “Putting serious force behind it produces extraordinary results.”

• If AI truly enters the business layer, Chinese companies may also enter new markets through projects such as the Belt and Road Initiative. He uses Saudi Arabia’s need to rebuild its IT infrastructure as an example, imagining systems rebuilt directly on an AI architecture. “Flowers of the next generation grow more easily in barren soil,” while Chinese users are also more flexible in adopting and adapting to new technologies.

24. “Agent” Works for Industry Discussion, but Not as a Product Pitch

• Asked whether marketing and fundraising should use the word Agent, 庄明浩’s answer is no. 智谱’s official description—“AutoGLM 沉思版, an AI Agent product similar to Deep Research”—was almost impossible for anyone at the event to understand on the first hearing, demonstrating how far technical terms, product language, and user perception have diverged.

• 朱峰 adds that PR can exploit market trends, but a hot word is not the same as product positioning. 庄明浩’s final advice is to leave the complexity of the “Year of Agents” in the conference room: when speaking to users, the product should first make clear what problem it solves.