和蚂蚁、千问、吉利、OPPO、vivo 的五位总裁，聊 Agent 的商业元年、漏斗与放大器

Jane's talk

AI-assisted summary of 和蚂蚁、千问、吉利、OPPO、vivo 的五位总裁，聊 Agent 的商业元年、漏斗与放大器 — original source. Check the original for full context.

Source published: 2026-10-26

Key ideas

2026年被视为Agent商业化元年，行业痛点已从前端获客与内容生产转向转化效率瓶颈，智能体正成为跨越信息呈现与交易履约鸿沟的核心工具。 (0:01)
OPPO与支付宝合作落地AHA与A2A协议，主动摒弃强制跳转App等移动互联网流量旧逻辑，不仅打破数据壁垒，还带动支付宝在OPPO终端上半年DAU净增超200万。 (17:27)
vivo指出PC端Agent成熟度背后包含逾3万次系统级适配，而手机移动端追平该体验至少需要一年以上的工程基建补课期。 (28:30)
蚂蚁集团CEO韩歆毅预测Agent生态商业闭环时间表：前3个月先锋商家见效，6个月（春节后）生态广泛入场，明年下半年将迎来全行业规模化商业拐点。 (30:35)

Detailed notes

## Background 本内容由 YouTube 频道「Jane's talk」于 2026 年 10 月 26 日发布，记录了 2026 支付宝年度生态大会上的一场闭门圆桌。对话由主持人 Jane（魏世杰）主持，汇聚了蚂蚁集团 CEO 韩歆毅、千问事业部总裁吴嘉、吉利汽车任向飞，以及手机双雄 OPPO 唐凯与 vivo 周围五位产业核心掌舵人，深入探讨 2026 Agent 商业化元年的流量重构、端侧生态博弈与商业化拐点。

## TL;DR 圆桌围绕 Agent 如何重构商业规则与产业生态展开：蚂蚁集团韩歆毅指出当前数字经济的根本摩擦在于流量昂贵稀缺，Agent 的核心价值在于提升转化效率，而最大挑战则是建立多方共赢的利益分配机制；OPPO 唐凯与 vivo 周围从手机终端视角剖析了意图入口、全栈感知与分布式执行器的演进，并揭示了模型飞速发展与产业链组织力滞后、移动端面临三万项系统适配技术债之间的巨大鸿沟；吉利任向飞则阐述了智能汽车向主动式具身智能跃迁的闭环场景与整车兜底责任。与会者一致认为必须摒弃移动互联网的流量劫持思维，以真实闭环与价值贡献为基准，预判行业将在经历数月商户验证后于次年下半年迎来商业化爆发拐点。

## Atmosphere & style 整场对话风格极其务实、克制且充满工程落地的紧迫感，完全脱离了概念宣讲的虚热。几位身处风暴眼的企业高管毫无保留地倒出“如履薄冰”的焦虑与睡眠被严重压缩的交付压力，在台上直面端侧厂商、汽车主机厂与平台巨头之间的商业防备、数据壁垒与信任重建。主持人的追问层层剥开商业算盘与现实瓶颈，在严苛的技术债清偿与“将时间还给人类”的终极愿景交织碰撞中，让观众深切感受到大变革前夕产业巨头们在博弈中试探同盟的真实张力。

## Bulletin - [00:01](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1s) 2026年被视为Agent商业化元年，行业痛点已从前端获客与内容生产转向转化效率瓶颈，智能体正成为跨越信息呈现与交易履约鸿沟的核心工具。 - [17:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1047s) OPPO与支付宝合作落地AHA与A2A协议，主动摒弃强制跳转App等移动互联网流量旧逻辑，不仅打破数据壁垒，还带动支付宝在OPPO终端上半年DAU净增超200万。 - [28:30](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1710s) vivo指出PC端Agent成熟度背后包含逾3万次系统级适配，而手机移动端追平该体验至少需要一年以上的工程基建补课期。 - [30:35](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1835s) 蚂蚁集团CEO韩歆毅预测Agent生态商业闭环时间表：前3个月先锋商家见效，6个月（春节后）生态广泛入场，明年下半年将迎来全行业规模化商业拐点。

## Key takeaways - [00:01](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1s) 流量稀缺且昂贵背景下，智能体的本质是解决信息转化为行动的转化效率瓶颈，完成端到端履约闭环。 - [01:28](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=88s) 汽车正从被动的信息检索工具跃迁为主动提供服务的“智慧生命体”，但面临传统电子电气架构重构的巨大压力。 - [04:00](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=240s) 基础大模型能力已显著超越产业链组织与协同能力，生态割裂与数据孤岛成为Agent商业化落地的核心阻碍。 - [05:12](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=312s) 端侧AI技术演进在2026年3月出现关键提速，Op-loop自主循环架构使行业在5个月内经历了过去七八年未有的体验跨越。 - [08:31](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=511s) 硬件终端（手机与智能汽车）正在取代传统单一App，成为统领用户意图并分发子智能体的“超级入口”。 - [11:46](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=706s) 硬件终端在Agent时代将承担三重角色演进：从应用启动器转向意图入口、全栈感知与记忆体，以及分布式执行器。 - [11:46](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=706s) 终端厂商的核心技术重心聚焦于边缘算力优化，目标是在1至2年内推动DeepSeek等高参数MoE模型在端侧原生运行。 - [15:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=927s) Agent生态落地的深层挑战不在于技术本身，而在于能否打破流量抽成逻辑，建立基于多方共赢的新型利益分配机制。 - [15:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=927s) 支付宝将自身定位下沉为Agent生态的基础设施底座，专注于提供身份核验、互信体系、支付路由与系统安全能力。 - [21:48](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1308s) 车端与手机端存在权责机制差异：车内智能体涉及人身安全与整车控制，车企必须对第三方的端到端交付结果承担完全系统级责任。 - [23:46](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1426s) 商业价值的衡量从单一应用时长转向全流程解决方案交付，利益分配必须依据核心数据、算法决策与最终交付的实际贡献值动态结算。 - [25:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1527s) 未来千万级Agent生态将呈现海量原子化微服务形态，智能座舱需部署专门的AIOS以协同主Agent与各领域外围智能体。 - [25:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1527s) 商业底层逻辑从抢夺眼球的“注意力经济”逆转为替用户解决琐碎事务、解放生命时间的“时间解放经济”。 - [30:35](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1835s) 由于AI技术以指数级跨越式发展，产业组织形式可能永远处于被动追赶模型能力的非对称状态。 - [30:35](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1835s) 生态繁荣的前提是普通商家与开发者能否快速实现商业闭环，平台方必须通过开放共建降低参与门槛，驱动正向飞轮。

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1s) 2026: The Commercial First Year of Agents and the Traffic Efficiency Bottleneck

The session opens with moderator Wei Shijie welcoming the audience to what is framed as a pivotal discussion on the business evolution of artificial intelligence. Setting the stage for 2026 as the recognized inaugural year of Agent commercialization, Wei immediately poses the central, unvarnished questions to the panel: What is the most critical issue keeping industry leaders up at night, and what is the fundamental essence of the commercial shift toward Agents? Han Xinyi, CEO of Ant Group and host of the forum, sets a calm, pragmatic tone by dismissing the notion of sleepless anxiety. Instead, he diagnoses the core friction in the digital economy today as an economic bottleneck: traffic has become both scarce and prohibitively expensive.

To Han, the true promise of artificial intelligence lies in resolving this fundamental contradiction by driving conversion efficiency. Over the past two to three years, the industry has aggressively optimized other phases of the digital business lifecycle—specifically user reach, content production, and downstream fulfillment. However, the conversion of information into action has remained an inefficient bottleneck precisely because it lacked true cognitive intelligence. Agentic intelligence, Han argues, serves as the decisive tool that bridges this final gap between information presentation and transactional conversion.

> “To solve this, we must improve efficiency, and I believe intelligence is the primary means to achieve that.” — Han Xinyi

[01:28](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=88s) From Transport Tool to Intelligent Living Entity: Geely's Vision and Anxieties

Geely Automobile Group’s Chief Engineering and Technical Scientist, Ren Xiangfei, addresses the shift from an industrial standpoint. For Ren, the symbolic value of an automotive executive sharing the stage with mobile handset giants and platform leaders illustrates the tectonic change already underway: vehicles are transitioning from passive machines requiring manual inputs and static information retrieval into proactive, service-delivering entities. Drawing an evocative parallel to the 1980s television show *Knight Rider*, Ren reflects on how childhood visions of an autonomous, highly intelligent companion vehicle are finally becoming viable through the convergence of foundation models and AGI.

Beneath the aspirational vision, however, Ren acknowledges deep anxiety about navigating an unprecedented epochal transition. Geely has positioned itself as an industry pioneer by committing to reframe cars from daily transportation tools into "intelligent living entities" and deploying AI across its entire corporate and engineering footprint. Yet transforming this vision into a sustainable product ecosystem under tight technical and competitive pressures creates immense friction. The stakes are extraordinarily high as traditional automotive architectures must be radically re-engineered to accommodate proactive, agentic workflows.

> “Perhaps the technological empowerment of agents is a crucial path that helps transform cars from mere daily transport tools into a form of intelligent living entity.” — Ren Xiangfei

[04:00](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=240s) The Structural Mismatch: Model Capabilities vs. Industrial Chain Organization

Tang Kai, Vice President of OPPO and President of Software Engineering, shifts the dialogue to the harsh realities of execution within consumer electronics. Challenging the idea that no one is anxious, Tang points directly to an emerging structural bottleneck: the raw capabilities of contemporary foundation models have outpaced the organizational maturity and cooperative capacity of the broader industrial ecosystem. While assembling a functional, impressive-feeling Agent prototype using current frontier models is relatively trivial, building an Agent that achieves repeatable commercial success remains an uphill battle.

According to Tang, an effective Agent cannot operate in a vacuum. It requires immense volumes of deeply contextual user data to correctly decipher complex human intents, combined with unfettered access to a wide array of downstream third-party services to deliver meaningful economic value. However, the industry remains heavily balkanized, fragmented across proprietary walled gardens and isolated data silos. Tang warns that aligning the entire supply and service chain—requiring cross-company consensus, technical standard-setting, and inevitable commercial turf wars—will take considerable time before the industry's organizational fabric catches up with model capabilities.

```mermaid flowchart LR accTitle: The Model vs Organization Capability Gap accDescr: Shows how raw model intelligence meets bottlenecks in data silos and fragmented service chains. subgraph S1["Upstream Capabilities"] A["Frontier Foundation Models"] --> B["Agent Prototypes & Intent Understanding"] end subgraph S2["Structural Bottlenecks"] B -.-> C["Data Silos & Privacy Walls"] B -.-> D["Fragmented Service Ecosystems"] end subgraph S3["Commercial Outcome"] C --> E["Stalled Commercialization"] D --> E E --> F["Need for Industry-Wide Alignment"] end ```

> “Industrial organizational capacity can no longer keep up with model capacity.” — Tang Kai

- **Model vs. Industry Gap**: Basic Agent user experiences are easy to build with current models, but sustainable commercialization is blocked by ecosystem friction. - **Data and Service Requirements**: True intent understanding requires contextual user data, while end-to-end execution demands deep integration into third-party commercial services. - **Ecosystem Fragmentation**: Siloed platforms and proprietary boundaries prevent Agents from completing multi-step tasks autonomously.

[05:12](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=312s) The Evolution of the AI Stack and the Six-Month Delivery Crunch

Zhou Wei, Vice President and Head of the AI Global Research Institute at vivo, frames his anxiety not as sleeplessness, but as having literally "hardly any time to sleep." Tracing the seven-to-eight-year arc of mobile AI, Zhou outlines the rapid technological shifts leading up to 2026. The journey began in 2018 with computer vision and early voice assistants that proved largely rigid and dim-witted. When large language models arrived in 2023, the industry spent 2024 attempting to unlock capability through prompt engineering, only to realize its severe limitations. By 2025, the paradigm shifted toward context engineering, which dominated developer forums like Google I/O.

Everything accelerated abruptly around March 11, 2026, when breakthroughs around operator loops (op-loops) and agentic frameworks took hold, aiming to string together perception, planning, and execution in an autonomous loop. Zhou stresses that the past five months have delivered more radical shifts in user experience than the previous seven or eight years combined. For a smartphone manufacturer whose foundational mission has been to serve as an intelligent, personalized bridge between everyday users and digital services, the compressed timeframe creates intense operational pressure. In the hardware sector, annual software deliverables must be locked in by September; teams had a mere six-month window to take agentic breakthroughs from research labs into commercial operating systems, leaving barely a month before final delivery.

> “I feel these five months and a few days have seen the greatest changes and shifts in experience over the last seven or eight years.” — Zhou Wei

- **2018**: Computer vision breakthroughs spark the initial wave of native mobile voice assistants. - **2023–2024**: Emergence of LLMs dominated by prompt engineering; user experiences remain comparatively crude. - **2025**: Industry-wide shift toward context engineering to supply rich background data to models. - **March 11, 2026**: Emergence of op-loop architectures enabling autonomous multi-step execution. - **Delivery Timeline**: Hardware release cycles dictate a strict September deadline, forcing companies to condense five to six months of agentic R&D into immediate production software.

[08:31](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=511s) Super Entrances and Ecosystem Silos: The Friction in End-to-End Delivery

Moderator Wei Shijie highlights a structural power transition: in the mobile internet era, individual apps served as the primary traffic entry points, but in the Agent era, physical hardware terminals—smartphones and connected vehicles—are asserting themselves as the definitive super entrances. Wei asks how the strategic role of devices is evolving. Ren Xiangfei of Geely elaborates on the friction this transition generates. In the traditional app paradigm, software on vehicle head units operated in closed, point-to-point silos where users manually searched for information or triggered specific commands.

In contrast, an Agent-centric architecture requires transforming passive information retrieval into proactive, end-to-end result delivery. If an automotive OEM attempts to build a master vehicle Agent capable of dispatching specialized third-party sub-agents, it immediately runs into commercial walled gardens. Major tech ecosystems—such as Ant Group's closed service networks or rival platforms—operate their own proprietary stacks. Because these commercial entities have conflicting business models and data governance boundaries, integrating them into a seamless, uninterrupted workflow poses a profound challenge. Ren points out the central irony: while AI architectures inherently require openness, interoperability, and collective data sharing to fulfill user intentions, the commercial entities controlling the services remain fiercely territorial. Tang Kai of OPPO prepares to expand on how mobile terminals navigate these exact proprietary tensions.

> “AI itself needs to be more open, more inclusive, with the whole society, the whole ecosystem, and all parties collaborating.” — Ren Xiangfei

- **Shift in Entry Points**: Hardware devices (smartphones, connected vehicles) are superseding standalone apps to become master orchestrators of user intent. - **Retrieval vs. Fulfillment**: The fundamental paradigm shifts from static information retrieval to end-to-end task completion and proactive service. - **The Walled Garden Clash**: Master Agents face structural integration roadblocks because major digital service ecosystems operate as closed, competing commercial silos.

[11:46](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=706s) The Terminal Evolution: Intent Entry, Full-Stack Sensing, and Distributed Actuation

周围 (Zhou Wei, Vice President and Head of the AI Global Research Institute at vivo) outlines how the fundamental role of hardware terminals will undergo a three-fold architectural shift in the emerging Agent era. First, smartphones and other terminal platforms are transitioning from being mere application launchers into entry points for user intent. Users care about obtaining the final service outcome rather than the underlying mechanism—whether delivered via APIs, APKs, or Agent Skills. As intent comprehension advances, on-device AI will orchestrate tasks and make autonomous decisions. Second, the terminal will evolve into an immensely powerful sensor; because frontier AI models simultaneously demand vast real-world data and excel at contextual utilization, hardware must embody the principle that the “full stack must perceive and the full stack must remember.” Third, terminals will act as distributed actuators coordinating an ecosystem of peripheral smart hardware—such as AI smart glasses, specialized audio and visual gear, and biometric health monitors—effectively evolving into a system of embodied intelligence.

```mermaid flowchart TD accTitle: Terminal Role Transformation in the Agent Era accDescr: Diagram illustrating how hardware terminals shift from app launchers to intent entries, sensors, and actuators.

subgraph Inputs["Perception & Sensing"] S1["Ambient Sensors & Biometrics"] S2["Peripheral Hardware (Glasses/Audio/Health)"] end

subgraph Core["Terminal Intelligence Engine"] T1["Full-Stack Perception & Memory"] T2["Intent Parsing & Agent Decision-Making"] T3["On-Device MoE Models (e.g., DeepSeek)"] end

subgraph Execution["Actuation & Delivery"] A1["A2A Protocols / Shared Intent Standards"] A2["Service Aggregators (Alipay Ecosystem)"] A3["End-to-End Task Closure"] end

Inputs --> Core Core --> Execution ```

Addressing the industry-wide preoccupation with immediate commercial monetization in Agent year one, Zhou notes that terminal manufacturers view their priorities differently. Rather than rushing to extract commercial tolls, terminal makers are prioritizing the heavy lifting required at the intersection of computing power, silicon architecture, and model deployment. With massive Mixture-of-Experts (MoE) architectures such as DeepSeek needing to be optimized to run natively on mobile chipsets within a tight one-to-two-year horizon, handset makers must concentrate on device-level readiness and building point-to-point Agent-to-Agent (A2A) protocols and shared intent standards in close partnership with service platforms like Alipay.

- “I believe that as the capability for intent grows stronger, AI will help us perform more orchestration and even decision-making,” asserts 周围 (Zhou Wei). - Mobile manufacturers anticipate deploying high-parameter MoE architectures natively on edge devices within a 1- to 2-year cycle. - The terminal's primary technological mandate is transforming from APK management to distributed hardware orchestration across vision, audio, and wellness sensors.

---

[15:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=927s) The Core Challenge of the Agent Ecosystem: Benefit-Sharing Mechanisms

Responding to the host’s inquiry regarding ecosystem bottlenecks and Alipay's infrastructure ambitions, 韩歆毅 (Han Xinyi, CEO of Ant Group) argues that the defining bottleneck of the Agent era is neither algorithmic power nor technical plumbing, but rather the creation of an equitable commercial distribution model. In an AI landscape characterized by fragmented user behaviors across diverse environments, no single hardware device or standalone super-app can realistically monopolize user needs. A functioning ecosystem requires open collaboration and distinct division of labor. Alipay aims to anchor the foundational substrate—taking responsibility for the non-negotiable primitives of digital interactions: identity verification, mutual trust, payment routing, and systemic security.

Han points out that getting multiple independent entities to invest in coordinated Agent workflows requires replacing the legacy web’s traffic-centric monetization models with a new system of benefit sharing. This mechanism cannot be artificially mandated from a top-down, “god’s-eye view”; it must emerge iteratively through commercial experimentation and mutual compromise among terminal vendors, platform hosts, and model builders. While organic evolution introduces friction and uncertainty, it represents the only viable path to a sustainable multi-party ecosystem.

- “So, to summarize, I think the real challenge isn't about capabilities or the technology itself, but rather the underlying system of benefit sharing,” states 韩歆毅 (Han Xinyi). - The emergent Agent infrastructure moves away from legacy traffic arbitrage and page-view economics toward outcome-based, multi-party value distribution. - Alipay’s self-defined role in the Agent architecture focuses on capability sinking—providing trust frameworks, financial clearance, and security backbones.

---

[17:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1047s) Building A2A Protocols, Mutual Trust, and Abandoning the Traffic Trap

唐凯 (Tang Kai, Vice President and President of Software Engineering at OPPO) illustrates how cross-platform Agent workflows function in practice through OPPO’s collaboration with Alipay on the AHA and A2A protocols. Establishing seamless interoperability between OPPO's Breeno (小布助手) and Alipay required clearing psychological hurdles before technical ones could be dismantled. Once early technical bridges were established, the teams confronted complex questions surrounding edge privacy, transaction authority, and data governance, prompting them to engineer an underlying risk-free verification layer. As a result, Breeno can invoke Alipay’s transaction and service capabilities natively without forcing users to pass through independent login prompts, creating frictionless cross-agent handoffs.

Tang dispels the notion that ecosystem participants should be haggling over revenue splits today, arguing that parties must first focus on expanding the total market size. He shares an instructive debate: early in the Breeno-Alipay integration, the teams clashed over whether invoking a service had to force-open the Alipay app. Traditional mobile-internet metrics incentivized app-jumping to preserve daily active users (DAU) and capture ad traffic. However, both teams agreed to abandon legacy metrics and prioritize seamless user experience. By surrendering app-jump demands and letting user data guide the interaction, Alipay’s daily active users on OPPO devices actually surged by more than 2 million in the first half of the year.

周围 (Zhou Wei, vivo) reinforces Tang’s points, emphasizing that durable partnerships rest on three pillars: trust, security, and openness. Zhou shares that within vivo’s Blue Heart (蓝心) models and Xiao V (小V) assistant, when a user asks a medical or health-related question, vivo immediately routes the request to Alipay’s healthcare agents unconditionally, without conditioning the integration on commercial rev-share agreements. This institutional confidence, Zhou adds, stems from years of reliable, transparent engagement between the executive teams.

- “Whether it's better to jump to the app or not, we let the technology, the data, and the user experience decide,” explains 唐凯 (Tang Kai). - “When any user asks a health-related question in our Blue Heart or Xiao V assistant, we provide services unconditionally,” remarks 周围 (Zhou Wei). - OPPO’s native A2A integration with Alipay generated an increase of over 2 million daily active users for Alipay across OPPO devices in H1.

---

[21:48](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1308s) The Automotive Paradigm Shift: From Passive Traffic to End-to-End Service Delivery

任向飞 (Ren Xiangfei, Chief Engineering Scientist of Geely Auto Group) details how the automotive industry’s position is being redefined as interaction shifts from screen-based traffic to proactive Agent intelligence. Historically, the automotive cabin lagged behind the consumer internet; vehicle digital services were largely offered for free, giving automakers minimal economic leverage or monetization power compared to smartphone platforms. In the Agent era, however, vehicle intelligence must transition away from passive information delivery toward an economic model predicated on end-to-end task completion and proactive physical service delivery. The priority is not dividing an existing revenue pie, but proving that an autonomous cockpit can deliver enough tangible utility that drivers willingly pay for executed tasks.

Ren highlights a structural accountability divide between smartphones and connected vehicles: on a handset, a failed task is attributed to the third-party developer—if an Alipay mini-program glitches, users fault Alipay rather than OPPO or vivo. In stark contrast, any systemic, mechanical, or safety failure within a vehicle cabin is attributed directly to the automaker. Because a modern car functions as a complex, highly embodied AI system, it must orchestrate dozens of concurrent tasks while safeguarding human life. For example, rather than waiting for a manual voice command, a vehicle's multi-modal sensors will continuously assess driver and passenger fatigue, biometrics, and physical discomfort, autonomously intervening to adjust environmental conditions, suggest rest stops, or summon medical resources.

- “In the past, there was no solution, but in the future, in the Agent era, in the AI era, the vehicle will be able to actively detect the driver's health or physical state,” explains 任向飞 (Ren Xiangfei). - Automotive economic models are moving from zero-cost infotainment toward paid end-to-end service delivery. - Unlike smartphone OEMs, automakers assume total system-level liability for the outcomes and safety implications of third-party Agent orchestrations inside the cabin.

[23:46](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1426s) End-to-End Mobility Scenarios and Value-Based Profit Sharing

Addressing how multi-agent collaboration creates tangible user value, the discussion explores an integrated in-cabin scenario where an automobile transitions from a passive transportation machine into an intelligent, proactive caregiver. If a driver suddenly falls ill behind the wheel, the vehicle leverages embedded perception and planning capabilities alongside deep ecosystem integration to immediately intervene. Rather than burdening an incapacitated user with manual navigation or searches, the vehicle automatically engages autonomous driving features to steer toward the nearest hospital. Simultaneously, an onboard agent powered by a large foundation model diagnoses the driver's real-time physical condition, schedules an urgent hospital appointment, and pre-selects the appropriate medical specialist.

```mermaid flowchart TD accTitle: End-to-End In-Cabin Agent Orchestration accDescr: Flowchart illustrating how vehicle perception and a master agent coordinate peripheral medical and driving agents to deliver closed-loop user outcomes.

subgraph CabinPerception["Vehicle Perception Layer"] A["Driver Vitals & Health State Detection"] end

subgraph MasterCoordination["Master Agent / AIOS"] B["Vehicle Master Agent Orchestration"] end

subgraph PeripheralExecution["Peripheral Agent Ecosystem"] C["Autonomous Driving to Nearest Hospital"] D["AI Medical Assessment & Triage"] E["Hospital Registration & Specialist Booking"] end

subgraph DeliveryValue["Closed-Loop Outcome"] F["End-to-End Problem Resolution"] G["Value-Based Profit Distribution"] end

A --> B B --> C B --> D B --> E C --> F D --> F E --> F F --> G ```

This integrated workflow demonstrates that consumer perception in the agent era will no longer revolve around isolated information lookups or single-point mobile apps, but around end-to-end task completion. When users demonstrate a willingness to pay for comprehensive problem resolution, sustainable commercial value is generated. Regarding revenue sharing among cross-industry participants, the consensus emphasizes establishing market-driven incentives tied directly to functional contributions rather than legacy gatekeeping. As the speaker emphatically argues: “It must be a true distribution based on value contribution: who provided the core data during this process? Who provided the primary core decisions, and who provided the final delivery to the user?” The prevailing philosophy is pragmatic: build the integrated capability first, establish customer value, and let market mechanisms dictate fair commercial yields.

- In-cabin agents synthesize health diagnostics, autonomous vehicle dispatch, and live clinical scheduling into a unified safety intervention. - User valuation shifts from fragmented software interactions to complete, closed-loop solutions. - Commercial profit-sharing models must dynamically compensate core data providers, algorithmic decision engines, and terminal delivery interfaces.

---

[25:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1527s) The Tens-of-Millions Agent Era: AIOS and the Shift from Attention to Time

Jane steers the panel toward the overarching commercial horizon of autonomous agents, asking the leaders to forecast the definitive trends of the next one to two years, envision the economic landscape once agent deployments reach tens of millions, and pinpoint potential structural bottlenecks. Geely's Ren Xiangfei responds by highlighting a profound structural divergence from the mobile internet era. While smartphone ecosystems have been dominated by a handful of centralized, monolithic super-apps, an ecosystem supporting tens of millions of agents will be powered by vast fleets of atomic-level micro-services developed by nimble teams and independent creators. This impending "Agent Internet era" dramatically elevates the role of hardware terminals, placing a premium on their aggregation capacity and their overarching "Master Agent" orchestration frameworks.

```mermaid flowchart LR accTitle: Mobile Internet vs Agent Internet Paradigm accDescr: Comparison of the mobile internet app-centric attention economy versus the agent internet AIOS-driven time-liberation economy.

subgraph MobileEra["Mobile Internet Era"] M1["Monolithic Super-Apps"] --> M2["Attention Capture & Ad Revenue"] M2 --> M3["User Trapped in Tedious Screen Tasks"] end

subgraph AgentEra["Agent Internet Era"] A1["Tens of Millions of Atomic Agents"] --> A2["Master Agent & Dedicated AIOS"] A2 --> A3["End-to-End Task Liberation"] A3 --> A4["User Reclaims Time for Life & Health"] end ```

To accommodate this artificial general intelligence (AGI) evolution, Ren explains that vehicle technical architectures must undergo structural reinvention. Automobiles will require a specialized AI operating system (AIOS) designed to host a master agent that coordinates specialized peripheral sub-agents to execute multifaceted missions. Expanding on this paradigm shift, another executive points out that while the mobile internet economy fundamentally operated as an attention-extraction engine that commodified human screen time, the agent maturity curve will invert this logic. Autonomous agents handle tedious bureaucratic chores and daily operational friction, returning precious hours to human beings. As the panelist observes: “As the Agent ecosystem becomes more mature, people's attention should shift; people should be engaging in exercise and health, rather than continuing to be stuck in tedious work.” In this light, projects like Ant Group's "A-Fu" healthcare initiative reflect a forward-looking pivot toward human time liberation over passive ad consumption.

- The proliferation of tens of millions of agents marks the transition from monolithic software applications to distributed atomic services built by individual developers and small studios. - Smart vehicles will require dedicated AIOS architectures capable of coordinating master and peripheral agents for real-time task routing. - The macroeconomic model shifts from an "attention economy" competing for eyeball duration to a "liberation economy" engineered to save time.

---

[28:30](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1710s) Walking on Thin Ice: System Adaptations and Mobile Infrastructure Debts

Offering a sobering counterweight to visionary optimism, vivo’s Zhou Wei confesses that he approaches the hype cycle feeling as though he is "walking on thin ice." He emphasizes that beneath the shiny promises of next-generation AI, underlying consumer requirements remain fundamentally conservative: whether seeking clinical triage or scheduling a vehicle wash, user expectations for reliability have not radically altered over the past decade. The superficial novelty stems not from an intrinsic shift in human demand, but from the arrival of more agentic foundation models and harnesses. However, delivering reliable consumer experiences on edge devices requires an immense engineering foundation that is far from complete.

Zhou explains that the sudden productivity leaps observed in desktop and PC agent harnesses did not emerge out of thin air; they were forged through more than 30,000 comprehensive, full-scale system adaptations and deep software-hardware alignments. On mobile handsets, this rigorous integration work is only in its infancy. For mobile agents to deliver true parity with desktop agent workflows, smartphone manufacturers cannot take shortcuts or rely solely on cloud model prompts. Expressing the sheer scale of technical debt facing device manufacturers, Zhou Wei warns: “If we want to have that same great experience next year, I believe not a single one of those 30,000 adaptations can be missed.” When pressed by Jane on how rapidly terminal makers can conquer this mountain of system-level skills, Zhou provides a blunt assessment: “At least a year is a very optimistic figure.”

- Core consumer demands for physical and digital services remain stable; the primary disruption lies in the systemic foundation supporting delivery. - Desktop agent maturity relies on over 30,000 full-scale operating system adaptations—a technical benchmark smartphone makers are only beginning to tackle. - Device vendors face massive engineering hurdles and technical debt before mobile AI agents can achieve parity with desktop counterparts, requiring upwards of a year under optimistic projections.

---

[30:35](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1835s) Commercial Closed Loops, Ecosystem Co-Creation, and the 2027 Inflection Point

Jane directs the conversation to Ant Group CEO Han Xinyi, picking up on a concern raised earlier by OPPO's Tang Kai regarding the widening chasm between hyper-accelerating AI model capabilities and lagging industrial coordination and organizational readiness. When asked how many years it will take for organizational competence to catch up with raw model performance, Han offers a startlingly candid paradox: “To be honest, given the current development of technology, I think we might never catch up.” Because AI capability advances at an unprecedented, exponential clip—progressing from open-source baselines to agentic execution harnesses in mere months, and traversing the entire trajectory from GPT-3 to contemporary multi-modal models in roughly twenty months—traditional corporate structures may remain in perpetual pursuit.

```mermaid flowchart TD accTitle: Ecosystem Commercialization and Adoption Timeline accDescr: Timeline mapping Ant Group's projected milestones for agent commercialization from immediate merchant response to late-2027 adoption inflection.

T1["Months 0–3: Initial Merchant Pilots"] -->|Early revenue proofs & closed-loop tests| T2["Month 6: Post-Spring Festival Influx"] T2 -->|Broader merchant & developer onboarding| T3["H2 Next Year: Commercial Inflection Point"] T3 -->|Scalable ecosystem flywheel across platforms & terminals| T4["Mass Consumer & Industry Realization"] ```

Nevertheless, Han maintains that ecosystem maturity can materialize far faster than pessimists anticipate, provided the industry successfully establishes self-sustaining commercial closed loops. While platform holders and handset makers possess the balance sheets to invest unilaterally for long-term user scale and experience, ordinary software developers and commercial merchants require immediate fiscal viability. As Han Xinyi underlines: “Only with such a closed-loop business model is it possible to get more people involved and form a positive cycle.” For Alipay, this mandate requires stripping away proprietary silos, opening up foundational infrastructure, and co-creating operating environments alongside competitors and hardware vendors rather than attempting centralized platform control.

Looking forward to concrete deployment timelines, Han sketches a clear roadmap for merchant adoption and user impact: - Within the initial three months, pioneering merchants will begin registering measurable commercial benefits from early agent integration. - Around the six-month mark—falling directly after the forthcoming Spring Festival—a much broader wave of enterprise merchants and ecosystem developers will actively onboard. - By the second half of next year, the market is poised to encounter a genuine commercial turning point and widespread adoption inflection.

## Notable quotes > [04:00](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=240s) Industrial organizational capacity can no longer keep up with model capacity. — Tang Kai

> [15:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=927s) I think the real challenge isn't about capabilities or the technology itself, but rather the underlying system of benefit sharing — 韩歆毅

> [17:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1047s) Whether it's better to jump to the app or not, we let the technology, the data, and the user experience decide — 唐凯

> [01:28](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=88s) Perhaps the technological empowerment of agents is a crucial path that helps transform cars from mere daily transport tools into a form of intelligent living entity. — Ren Xiangfei

> [25:27](https://www.youtube.com/watch?v=tB6TX1rXxfs&t=1527s) As the Agent ecosystem becomes more mature, people's attention should shift; people should be engaging in exercise and health, rather than continuing to be stuck in tedious work.

## Mentioned 韩歆毅 (Han Xinyi), 吴嘉 (Wu Jia), 任向飞 (Ren Xiangfei), 唐凯 (Tang Kai), 周围 (Zhou Wei), 魏世杰 (Wei Shijie / Jane), 蚂蚁集团 (Ant Group), 阿里巴巴 (Alibaba), 千问 (Tongyi Qianwen), 吉利汽车 (Geely Automobile), OPPO, vivo, 支付宝 (Alipay), 小布助手 (Breeno), 蓝心大模型 (Blue Heart), 小V (Xiao V), 安福 (A-Fu), DeepSeek, Openclaw, Harness, Knight Rider

## Video description 过去两三年，我们都在为大模型的「智力涌现」而兴奋。但今天，整个商业界正在经历一场更深彻的质变——

2026，被视作 Agent 商业化的关键元年。

年初，从 Openclaw 的爆火到 Harness 的广泛应用，一幅新的商业地图正在展开—— 在移动互联网时代，商业的核心是 APP，是流量和注意力的争夺；但在 Agent 时代，从意图识别到端到端的交付闭环，当 Agent 可以为我们办事，智能正在绕过移动互联网时代筑成的流量墙，构建全新、不同以往的 AI 原生交互体验。

人机交互的入口在变化，这意味着，原有的商业规则、流量格局和生态天花板，都面临着重构。

这场重构会如何发生，价值的核心会是什么，生态会怎样建成——带着这些关键问题，在2026支付宝年度生态大会上，我主持了一场信息密度极大的闭门圆桌。

圆桌难得地召集了几位重量级角色：

坐镇东道主、思考 Agent 生态底座与多方受益共享新机制的——蚂蚁集团 CEO 韩歆毅；阿里巴巴集团副总裁、千问事业部总裁吴嘉；探索汽车从“出行工具”向“智慧生命体”跃迁的——吉利汽车集团首席工程技术科学家任向飞；以及坐拥数亿用户的手机终端双雄——OPPO 副总裁、软件工程系统总裁唐凯与 vivo 副总裁、AI 全球研究院院长周围。

五位嘉宾分别代表平台生态方、大模型入口，以及汽车与手机这两大超级硬件终端。身处 Agent 风暴眼核心的掌舵人们坐在一起，分享了各自真正的焦虑、跨端协同背后的博弈与信任，以及对 Agent 到来后商业拐点的预判。

[Source](https://www.youtube.com/watch?v=tB6TX1rXxfs)

中文简报

对话蚂蚁、千问、吉利、OPPO、vivo五位总裁：Agent商业元年、漏斗与放大器圆桌对话围绕Agent如何重构商业规则与产业生态展开：蚂蚁集团韩歆毅指出当前数字经济的根本摩擦在于流量昂贵稀缺，Agent的核心价值在于提升转化效率，而最大挑战则是建立多方共赢的利益分配机制；OPPO唐凯与vivo周围从手机终端视角剖析了意图入口、全栈感知与分布式执行器的演进，并揭示了模型飞速发展与产业链组织力滞后、移动端面临三万项系统适配技术债之间的巨大鸿沟；吉利任向飞则阐述了智能汽车向主动式具身智能跃迁的闭环场景与整车兜底责任。与会者一致认为必须摒弃移动互联网的流量劫持思维，以真实闭环与价值贡献为基准，预判行业在经历数月商户验证后，将于次年下半年迎来商业化爆发拐点。 2026年被视为Agent商业化元年，行业痛点已从前端获客与内容生产转向转化效率瓶颈，智能体正成为跨越信息呈现与交易履约鸿沟的核心工具。 (0:01) OPPO与支付宝合作落地AHA与A2A协议，主动摒弃强制跳转App等移动互联网流量旧逻辑，不仅打破数据壁垒，还带动支付宝在OPPO终端上半年DAU净增超200万。 (17:27) vivo指出PC端Agent成熟度背后包含逾3万次系统级适配，而手机移动端追平该体验至少需要一年以上的工程基建补课期。 (28:30) 蚂蚁集团CEO韩歆毅预测Agent生态商业闭环时间表：前3个月先锋商家见效，6个月（春节后）生态广泛入场，明年下半年将迎来全行业规模化商业拐点。 (30:35)

Reading summary

在被定义为商业化元年的背景下，蚂蚁集团、吉利汽车、OPPO 与 vivo 的多位产业高管指出，智能体（Agent）的本质是通过认知智能突破传统流量转化效率瓶颈，但其落地正受制于模型能力超速发展与产业组织协调、底层系统工程深度适配之间的严重脱节。流量瓶颈与商业本质：从掠夺注意力转向交付闭环 (0:01) 蚂蚁集团董事长兼首席执行官韩歆毅指出，当前数字经济的核心矛盾在于流量稀缺且成本昂贵。过去数年行业在用户触达、内容生成和履约交付环节进行了大量优化，但因缺乏真正的认知智能，信息向交易行为的转化始终存在效率漏斗。智能体的核心商业价值正是填补这一转化缺口。与此同时，智能体经济学打破了移动互联网依赖时长与流量套利的旧模式；随着千万级微型服务取代单一超级应用，商业重心由争夺用户眼球的“注意力经济”转向代劳繁琐事务的“时间解放经济”，用户的付费意愿将直接建立在端到端问题解决的完整闭环上。硬件终端角色重塑：意图入口、全栈感知与分布式执行 (11:46) 智能终端正逐步取代独立App成为主导性的超级意图入口。vivo副总裁周围阐述了终端架构的三重蜕变：首先是从单一的“应用启动器”升级为由AI统一调度决策的“意图入口”，用户不再关注交付介质是API还是APK；其次演进为“全栈感知与全栈记忆”的超级传感器；最后则充当连接音频、智能眼镜与健康配件的分布式执行中枢。吉利汽车总工程师任向飞亦指出，汽车正从被动代步工具向主动服务的“智能生命体”与车载AIOS演进。然而，车端智能体需要承载多并发任务调度与人身安全底线，汽车主机厂对第三方协同故障承担全系统责任，这一风险机制与手机端存在显著差异。打破围墙花园：构建A2A互信协议与动态价值分配机制 (15:27) 多位高管一致认为，构建智能体生态的最大阻碍并非算法能力，而是跨平台的商业信任与利益分配机制。在既有格局下，大型平台与各硬件厂商受制于数据孤岛与商业壁垒，若各方过早争夺分成比例，将抑制整体市场拓展。OPPO副总裁唐凯以OPPO小布助手与支付宝合作AHA和A2A协议为例说明，双方放弃了以强行“跳转App”来维持日活和广告曝光的传统指标，转而以体验为核心实现原生无感调用，这一让步反而在上半年为支付宝在OPPO设备上净增超200万日活。周围也提到vivo蓝心大模型与小V助手在健康场景无条件接入支付宝医疗智能体。韩歆毅强调，未来的分润体系无法自上而下预设，必须由核心数据提供方、主要决策引擎以及终端交付方依据各自的价值贡献由市场动态协商确立。工程债务与组织滞后：移动端底座适配与推进时间表 (28:30) 针对模型能力的狂飙突进，业内呈现出冷静与危机感。周围坦言团队处于“如履薄冰”的交付高压下，他指出桌面端Agent的突破依赖于超过30,000项全方位的软硬件系统级适配，而移动手机终端目前才刚刚起步，乐观估计至少需要一年时间才能补齐这一工程技术债务。此外，韩歆毅直言，考虑到大模型每隔数月即发生底层跃迁的指数级速度，传统产业的组织能力“可能永远追不上技术能力的演化”。面对中小型开发者和商家对现实盈利周期的硬性要求，韩歆毅给出落地预期：前3个月首批商户跑通闭环收益，6个月（春节后）迎来更广泛开发者入驻，预计次年下半年整个生态将迎来真正的商业拐点与规模化爆发。智能体商业化面临的核心未解难题在于：在大模型技术以数月为周期加速迭代的背景下，产业链上下游能否在资金消耗和技术债务压顶之前，建立起兼顾数据隐私、跨平台互通与合理收益分摊的可持续商业循环。

精炼阅读摘要

蚂蚁集团、OPPO、vivo 与吉利等产业高管认为，2026 年开启的 Agent 商业化本质上是从争夺眼球的“流量经济”向以促成交易为核心的“交付与时间解放”转型，但其落地正受制于模型演进速度与产业链组织协同、底层系统工程之间的巨大断层。商业本质：从流量争夺转向确定性闭环与价值分配 (0:01) 蚂蚁集团 CEO 韩歆毅指出，当前数字经济的根本摩擦在于流量日益稀缺且昂贵，过去行业虽然优化了触达、内容与履约，但信息转化仍缺乏认知智能。Agent 的真正价值在于打破转化效率瓶颈，填补信息呈现与交易履约之间的鸿沟。这一范式转移促使商业逻辑从依赖展示和应用跳转的旧有流量经济，走向注重最终结果的收益分配机制。韩歆毅强调，Agent 时代的真正挑战不是技术本身，而是如何建立兼顾终端、平台与模型厂商的利益分配体系，该机制无法自上而下预设，只能通过商业实验动态演进。OPPO 软件工程事业部总裁唐凯指出，OPPO 与支付宝基于 AHA 和 A2A 协议的合作中，双方放弃了以应用跳转保日活（DAU）的旧指标，最终反而促成支付宝在 OPPO 终端上半年 DAU 增长超 200 万，印证了摒弃流量防线、以用户体验为先的可行性。终端演进：从单点 App 走向超级意图入口与具身执行器 (11:46) 智能手机与智能汽车正取代传统独立 App 成为调度意图的超级入口。vivo 副总裁周围将终端变革概括为三重角色：从应用启动器转变为意图入口、全栈感知与记忆的传感器，以及调度智能眼镜、健康设备等外设的分布式执行器。vivo 规划在一至两年内将类似 DeepSeek 的混合专家（MoE）大模型原生部署于端侧芯片，并推动 A2A 协议以直接闭环调度外部服务。在智能汽车领域，吉利汽车集团任向飞指出，汽车座舱正从被动信息查询转变为主动交付服务的主体，未来需要专用的 AI 操作系统（AIOS）协调主 Agent 与外围子 Agent。例如在驾驶员突发疾病场景中，系统需自主接管行驶路线、联动大模型诊断并预约号源。但与手机故障多由第三方应用承担责任不同，座舱内的执行失效与行车安全直接由主机厂背书，因而车端对跨生态调度的系统确定性有着严苛要求。落地瓶颈：模型爆发与产业组织滞后、生态壁垒的冲突 (4:00) 尽管基础模型能力突飞猛进，但产业上下游在数据整合、商业边界与系统适配上的摩擦严重制约了规模化落地。唐凯提出，当前的结构性矛盾在于“产业组织能力跟不上模型能力”，构建一个惊艳的 Agent 原型门槛极低，但要实现商业成功，必须融合深度上下文数据并打通碎片化的第三方服务链条，而各平台的商业壁垒和数据孤岛仍十分顽固。任向飞亦指出，汽车主 Agent 在跨平台调度服务时，屡屡遭遇大厂自建生态的封闭边界，AI 架构所要求的开放协同与商业机构的领地意识形成鲜明讽刺。韩歆毅甚至提出一个悖论：在技术以月为周期指数级演进的背景下，传统企业组织的协同与制度演化可能永远难以赶上模型能力的发展速度。工程债务与商业化节拍：如履薄冰的基础设施适配 (28:30) 面对 Agent 元年的行业兴奋，vivo 周围直言自己“如履薄冰”。他指出，消费者对医疗分诊或洗车等实际服务的可靠性要求十年来未曾改变，桌面端 Agent 带来的效率飞跃背后依赖于超过 3 万次全系统深度适配与软硬件协同，而移动终端在此类基础工程上才刚刚起步。端侧设备商若想完全达到同等体验，无法仅靠云端 Prompt 走捷径，必须补齐底层工程欠账，其推进周期在乐观预估下也至少需要一年。对于整体商业化的时间表，韩歆毅预测，商业闭环是撬动中小商户与开发者的核心动力，最初 3 个月内标杆商户将验证早期收益，6 个月后（春节前后）会迎来更广泛的商户接入，而整个生态的真正商业拐点预计将在下半年的次年阶段全面爆发。五位高管形成的共识表明，Agent 商业化的核心障碍已从算法原型转向底层系统工程的补课与跨平台利益机制的重构。如何在各大科技巨头的商业护城河之间建立开放、免密且权责清晰的 A2A 协议，并在端侧完成极其繁重的系统级适配，仍是该体系迈向数千万级规模前悬而未决的现实挑战。

Source quotes

[4:00] Tang Kai: Industrial organizational capacity can no longer keep up with model capacity. [15:27] 韩歆毅: I think the real challenge isn't about capabilities or the technology itself, but rather the underlying system of benefit sharing [17:27] 唐凯: Whether it's better to jump to the app or not, we let the technology, the data, and the user experience decide [1:28] Ren Xiangfei: Perhaps the technological empowerment of agents is a crucial path that helps transform cars from mere daily transport tools into a form of intelligent living entity. [25:27] As the Agent ecosystem becomes more mature, people's attention should shift; people should be engaging in exercise and health, rather than continuing to be stuck in tedious work.

Original transcript

[0:01] Good afternoon, everyone. I am Wei [0:03] Shijie, founder of the video podcast " [0:05] Man Tan Star." I am very pleased to [0:08] moderate this high-profile roundtable, [0:10] and I believe everyone here can feel [0:12] the immense value of our guests on [0:13] stage. So, in the limited time we have, [0:17] I will ask some pointed questions on [0:19] your behalf and invite our guests to [0:21] speak freely. Let’s start with the [0:24] most important question: 2026 is seen [0:26] as the first year of Agent commerce. [0:29] What is the most important issue you [0:30] are thinking about this year? Is there [0:32] anything that keeps you up at night? [0:34] How do you understand the essence of [0:36] the Agent business transformation? [0:38] Let’s start with our host, Mr. Han. [0:40] Thank you. I don't really have issues [0:42] that keep me up at night. If we look at [0:46] the essence of AI, we see it as [0:48] intelligence, which today resolves our [0:50] biggest contradiction: traffic is [0:51] insufficient and too expensive. To [0:56] solve this, we must improve efficiency, [0:58] and I believe intelligence is the [0:59] primary means to achieve that. What [1:03] does intelligence solve in this context [1:05] ? We believe it addresses the [1:08] efficiency of information conversion. [1:11] Various business stages, from reach and [1:13] production to fulfillment, have seen [1:15] significant improvements over the past [1:17] two or three years. However, the [1:20] efficiency of information conversion is [1:22] still insufficient because it lacks [1:24] this intelligence, so I believe [1:26] intelligence can fill that final gap. [1:28] >> Next, I’d like to invite Mr. Ren from [1:30] Geely to share his thoughts. Do you [1:32] have any sources of anxiety? [1:34] >> I will talk about my anxieties in a [1:36] moment. First, let me answer your [1:40] second question: what is the [1:41] fundamental change in this industry? I [1:45] believe my sitting here participating [1:47] in this event is itself a manifestation [1:50] of that fundamental change. As an [1:55] automotive company, collaborating with [1:57] top-tier terminals and large models, we [2:00] are exploring a new form of vehicle. [2:08] Moving from the passive information [2:10] retrieval or use of cars in the past, [2:12] towards a future of proactive service [2:14] —I believe this is a truly [2:16] significant turning point. I remember [2:25] watching a movie as a child in the 80s [2:28] called "Knight Rider"—some of the [2:30] slightly older audience might have seen [2:32] it—which envisioned an intelligent, [2:34] AI-driven type of car. It gave people [2:40] endless imagination. I feel the timing [2:43] for that seems to be here now. It’s [2:46] the era of an AGI-powered car. Perhaps [2:54] the technological empowerment of agents [2:56] is a crucial path that helps transform [2:58] cars from mere daily transport tools [3:00] into a form of intelligent living [3:02] entity. So, returning to your first [3:09] question: what am I anxious about? [3:15] Indeed, as a veteran in the automotive [3:17] industry and its ecosystem, I am truly [3:19] filled with anxiety at this specific [3:22] moment of epochal change. From [3:27] technical forms to ecosystems, [3:29] including the expansion of this [3:31] ecosystem, is an extremely important [3:33] environment. Because Geely was the [3:37] first in the automotive industry to [3:39] propose turning cars from transport [3:41] tools into intelligent living entities, [3:43] and the first to deploy AI across the [3:45] entire organization. We also hope to [3:51] bring some new changes and ideas to the [3:53] industry in this field, which is also a [3:55] rather anxiety-inducing thing. Thank [4:00] you, Mr. Ren. We will focus on the [4:02] challenges involved in a moment. Next, [4:04] we invite Mr. Tang from OPPO to tell us [4:06] what the mobile phone industry thinks. [4:08] Everyone just said they aren't losing [4:10] sleep over anxiety, so let me mention [4:12] something that still makes me feel a [4:14] bit anxious. Uh, I’ve noticed a [4:16] phenomenon: our current model [4:18] capabilities may have already surpassed [4:20] the organizational capacity of the [4:22] entire industrial chain. Today, I [4:25] believe everyone feels that building an [4:27] agent with a decent experience based on [4:29] these models is not a very difficult [4:31] task. However, if we want to achieve [4:34] commercial success, it is not that easy [4:37] . This agent needs a lot of data to [4:40] become more sensible, meaning it can [4:42] better understand the intentions [4:43] everyone keeps mentioning. At the same [4:46] time, it also needs a lot of services. [4:49] Only then can it generate greater value [4:51] . But objectively speaking, whether [4:54] it’s services or data, there are [4:56] still many silos. I think it will take [4:59] some time for the entire industry chain [5:01] to reach a consensus, collaborate, and [5:03] sometimes even fight for it. Industrial [5:06] organizational capacity can no longer [5:08] keep up with model capacity. A very [5:10] good point, let's discuss it in depth [5:12] later, Mr. Zhou. [5:16] >> Regarding whether this makes people too [5:18] anxious to sleep, my feeling is that [5:20] there is hardly any time to sleep. AI [5:25] has been developing for seven or eight [5:28] years now. In 2018, the trend really [5:31] picked up, and when image recognition [5:33] became possible, everyone was very [5:36] excited. As a result, everyone rushed [5:39] to launch their own voice assistants. [5:42] It still seemed pretty dim-witted, [5:43] though. By 2023, when large models [5:46] arrived, everyone felt that AGI was [5:48] coming this time, only to realize that [5:50] large models have actually gone through [5:53] three or four long years from the [5:55] beginning to now. So back in 2024, we [5:57] saw the entire industry talking about [6:00] prompt engineering, focusing on how to [6:02] use prompts to get the best out of [6:04] large models. The experience in 2024 [6:07] didn't seem that great, but when I [6:09] attended Google I/O in 2025, I found [6:11] the whole of Google talking about [6:13] context engineering, feeling that [6:15] prompts weren't really that smart and [6:18] were a bit silly, so we had to shift to [6:20] context engineering. Good heavens, 2025 [6:23] was a busy year with everyone launching [6:26] voice assistants based on large models [6:28] and constantly refining them, until [6:31] March 11, 2026, when I remember [6:33] Anthropic came along, and this op-loop [6:35] idea emerged, aiming to string [6:37] everything together within this loop. [6:40] For the past three years, we've been [6:42] doing one thing: using the emergence of [6:45] large models, combined with prompts and [6:47] context, to move better and further [6:49] along the path toward AGI. But this [6:52] time, from March 2026 until now, it's [6:55] actually been just five months and a [6:57] few days. I feel these five months and [7:00] a few days have seen the greatest [7:02] changes and shifts in experience over [7:04] the last seven or eight years. [7:08] Returning to myself, I'm in charge of [7:10] mobile software and AI at F, so my [7:12] anxiety stems from how we define our [7:14] own position. Looking back over the [7:20] years since AI emerged, how should we [7:22] actually position our company's [7:23] products? What exactly do our mobile [7:27] phone software products bring to the [7:29] users? Actually, our answer has never [7:33] changed: we hope to use the phone as a [7:35] bridge between people and the vast [7:37] number of services in the digital world [7:39] . And what do we use AI for in this [7:45] system? We hope to build a personalized [7:52] assistant for people through continuous [7:54] personalization efforts; this original [7:56] intention hasn't changed in eight years [7:58] , and we've been digging deeper, so [8:00] returning to the question, what are we [8:02] anxious about? The anxiety is about how [8:08] to turn this positioning and vision [8:10] into reality. Why do I say there's no [8:12] time to sleep? Well, from March to [8:14] these five-plus months, we in the [8:16] mobile industry usually call it " [8:18] submitting our homework," and the [8:20] deadline is in September, right? So [8:22] there is just over a month left. The [8:24] whole process has been six months to [8:26] submit this work; let's wait and see, [8:29] it's still a very tight schedule. Can [8:31] you feel the pressure in the hardware [8:33] industry? Let me lead you to focus on [8:35] the trends next. I've discovered a [8:37] particularly interesting phenomenon: [8:39] when we say the agent era has arrived, [8:41] distinct from the previous mobile [8:43] internet era, the entrance is [8:44] essentially changing, and I think power [8:46] is shifting as well. Previously, we [8:48] said apps were the entrance for traffic [8:50] , but now hardware terminals have [8:52] become a very important entrance, and [8:54] everyone wants to be this super [8:56] entrance. So, I can tell that today, [8:58] except for Mr. Han, all the guests on [9:00] stage represent the super entry points [9:02] for hardware. So my next question is, [9:06] what changes do you all think will [9:08] happen to the role of the device in [9:09] this Agent revolution? What is your [9:13] strategic layout? Let’s start with [9:16] Mr. Ren from Geely. [9:20] >> Uh, over the past few years, Geely has [9:22] actually been experimenting and working [9:25] hard on AI in vehicles and the [9:26] agentification of these areas; we were [9:29] also the earliest to start in the [9:31] automotive industry. And in this [9:35] process, we have also discovered that [9:37] there are quite a lot of problems and [9:39] challenges. Beyond the model maturity [9:46] and iteration speed we usually talk [9:48] about, in terms of the ecosystem—the [9:50] ecosystem—the future for mobile [9:52] phones and cars is the same; they are [9:55] super terminals, so they naturally need [9:57] to introduce many ecological partners. [10:06] In the past, for a car in the app era, [10:08] when we introduced certain apps, they [10:11] were relatively closed, and services [10:13] were point-to-point. But in the agent [10:21] era, you will find that the previous [10:23] forms of information retrieval are [10:25] actually unable to adapt to future [10:27] development; we need to turn past [10:29] information retrieval into services, [10:31] into result delivery. In this way, it [10:39] requires a series of APPs. If we [10:44] transition from original APPs to AGENTS [10:46] to form end-to-end services, and if we [10:48] car manufacturers want to build a [10:50] master AGENT that calls various AGENTS, [10:53] you will find a problem: the terminals [10:55] themselves are also relatively closed [10:57] silos. For example, Mr. Han’s [11:01] ecosystem and someone else's ecosystem [11:04] —they also have their own ecosystems. [11:06] So, this makes it difficult for future [11:08] agents to truly achieve end-to-end [11:10] delivery and end-to-end service. It has [11:13] natural obstacles. AI itself needs to [11:15] be more open, more inclusive, with the [11:18] whole society, the whole ecosystem, and [11:21] all parties collaborating. However, the [11:23] interest groups that AI belongs to are [11:25] relatively fragmented. This [11:28] contradiction brings challenges, and I [11:30] think that in the short term, for [11:32] deeper agents on the terminal side, it [11:34] poses a significant challenge. [11:37] >> Understood, Mr. Tang. [11:42] >> I think that as a terminal company, [11:44] there should be some changes to our [11:46] terminal platforms moving forward. [11:49] There might be a few points. The first [11:52] point is that today, the terminal is an [11:54] entry point for applications. But [11:56] subsequently, it may transform into an [11:58] entry point for intent. Because what [12:00] the user needs is the final result of [12:01] the service. As for whether we use APIs [12:05] , APKs, or Skills to serve them, these [12:07] are all processes we have developed [12:09] over the years. I believe that as the [12:13] capability for intent grows stronger, [12:15] AI will help us perform more [12:17] orchestration and even decision-making. [12:21] Secondly, I think the terminal will [12:22] become a much more powerful sensor. [12:26] Because the model actually has a great [12:28] need for data, yet it is also very [12:29] adept at utilizing that data. And this [12:33] sensing capability actually provides [12:34] the possibility of offering much more [12:36] and larger amounts of data. So, we [12:40] often say now that the full stack must [12:42] perceive and the full stack must [12:43] remember, so you can see that the role [12:45] of the terminal as this sensor will [12:47] become increasingly important. Then, [12:53] the third point is that I think the [12:54] entire terminal will become an actuator [12:56] for multiple devices. Because we may [13:03] soon see many new types of hardware [13:05] devices appearing in the industry, such [13:07] as the glasses just mentioned, or other [13:09] audio devices, visual devices, and even [13:11] health-related devices. And these [13:16] intelligent hardware components will [13:18] all form a distributed intelligent [13:19] agent system. Actually, I am also [13:23] thinking that when these intelligent [13:25] hardware pieces function like an [13:26] ecosystem, it is essentially like [13:28] embodied intelligence evolving. [13:30] >> Understood, Zhou. Because. For a [13:37] terminal device to better serve our [13:39] users, we only provide the carrier, but [13:41] what users are truly enjoying is the [13:43] service of the entire digital world, [13:45] such as various intelligent agents [13:47] aggregated by platforms like Alipay. [13:52] Today, I see many questions asking [13:54] about commercial monetization and how [13:56] we view this inaugural year; actually, [13:59] as a terminal smartphone manufacturer, [14:01] we have considered this issue [14:03] relatively less. Perhaps it is more [14:05] about how, in this massive process of [14:07] transformation, we focus on the [14:09] integration of computing power chips, [14:11] as well as the underlying large-scale [14:14] models. On the path toward on-device [14:19] model deployment, everyone has seen [14:21] that super-large MoE models like [14:23] DeepSeek will need to run on phones [14:25] within just a one-to-two-year cycle; [14:27] this is what we should be preparing for [14:29] . Beyond that, we may collaborate with [14:38] Alipay on how to build point-to-point [14:40] A2A protocols and construct shared [14:42] standards for intent, which I think is [14:44] where we should invest more effort. [14:49] Thank you, Mr. Zhou. Han, I’d like to [14:54] ask, since you’re hosting today’s [14:56] event and clearly aiming to build a [14:58] robust infrastructure and foundation [15:00] for the Agent era, while also serving [15:02] as a key leader in organizing and [15:04] coordinating the ecosystem, what do you [15:07] see as the actual bottlenecks and [15:08] challenges to making Agents a reality? [15:14] What kind of ecosystem are you hoping [15:16] to build? Regarding the client side and [15:18] our industry partners, what do you hope [15:20] everyone will contribute? What role [15:22] should they play? [15:27] >> Yes, I personally feel the real [15:28] challenge lies in how we build a new [15:30] mechanism for sharing benefits in a [15:32] more open ecosystem. In the AI era, [15:39] it’s hard to imagine that in the [15:41] foreseeable future, a single device or [15:43] super-app will be able to encompass [15:44] everything, right? Because user habits [15:48] are, in fact, quite diverse. With [15:52] different users having different habits [15:54] , it requires a division of labor among [15:56] various participants, which really [15:57] demands that all of us become more open [15:59] . In this regard, the second thing [16:04] Alipay wants to do is to sink deeper [16:06] into the capability side. We are [16:10] building out the most essential [16:12] elements needed in this ecosystem, such [16:14] as trust, payment, and even parts of [16:16] security—we are taking charge of [16:18] building those. And finally, for [16:23] everyone to participate, there must be [16:25] a benefit-sharing mechanism, which [16:26] might not be exactly the same as the [16:28] traffic-based sharing models of the [16:30] past. But what is this new mechanism? [16:36] Will it be designed from a "god's-eye [16:37] view," or will it emerge gradually as [16:39] all participants work through it [16:41] together? I believe it will be the [16:46] latter, but we all know that this [16:48] approach is extremely challenging and [16:49] fraught with uncertainty. So, to [16:55] summarize, I think the real challenge [16:57] isn't about capabilities or the [16:59] technology itself, but rather the [17:01] underlying system of benefit sharing. [17:05] Understood. Tang, since Han just [17:10] mentioned building this mechanism [17:12] together, in an ecosystem where an [17:14] Agent performs a closed-loop task [17:15] across multiple ends through joint [17:17] efforts, what kind of rights and [17:19] responsibilities distribution mechanism [17:21] do you hope for? [17:27] >> Well, I think I can start by sharing [17:29] some stories from our collaboration [17:31] with Alipay on the AHA protocol. [17:36] Because for this, to connect these [17:37] agents together really well. First, [17:40] everyone needs to have full trust. Yes, [17:42] so there must be psychological trust. [17:45] Then, we also need to break down the [17:47] technical walls. So, in our [17:51] collaboration process, I think we [17:53] quickly built that psychological trust [17:55] and reached a consensus on cooperation. [17:58] Technically, it was also opened up very [18:00] quickly. This way, the A2A connection [18:04] looked like it was made, but then we [18:06] discovered that after connecting Breeno [18:08] and Alipay, there were still many [18:10] issues regarding security, privacy, and [18:12] even mutual trust. So we sat down again [18:19] and started discussing how to build an [18:21] underlying mechanism to make A2A even [18:23] more risk-free. This was the entire [18:29] process of A2A development. Of course, [18:34] with this A2A in place, it will solve [18:36] many underlying security issues. One [18:41] simplest scenario is that today, when [18:43] we request services from Alipay via [18:45] Breeno, there is no need to log in to [18:47] Alipay separately. I think this will [18:53] provide users with a very smooth [18:55] experience, and at the same time, we [18:57] both believe that the large-scale [18:59] deployment of A2A will greatly improve [19:01] efficiency. As for business profits and [19:07] how to split them, I don't think it's [19:09] time for that yet. At this stage, we [19:13] are probably more focused on how to [19:15] grow the cake together. Just in our [19:20] collaboration with Alipay in the first [19:22] half of the year, I recall that on all [19:24] OPPO brand phones, Alipay’s daily [19:26] active users increased by over 2 [19:27] million. I think this is a great [19:33] synergy process, and that’s how the [19:34] cake gets bigger. [19:37] >> Let me ask a crucial question: how do [19:39] you build mutual trust? [19:40] >> Let me give a very small example. When [19:43] we first started collaborating, we were [19:45] stuck on one thing: when Breeno invokes [19:48] an Alipay service, should it [19:49] necessarily jump to the app? At first, [19:53] we insisted on jumping to the app. [19:56] Because, after all, it’s about DAU, [19:58] it’s about traffic. But later, we [20:00] quickly reached a consensus that we [20:02] must prioritize user needs and [20:04] experience. Whether it's better to jump [20:07] to the app or not, we let the [20:08] technology, the data, and the user [20:10] experience decide. We gave up the DAU, [20:13] right? But the real users started using [20:17] it, and our overall scale actually grew [20:19] instead. I’ve gathered one thing: [20:22] setting aside what we have already [20:24] owned and achieved to face the future [20:26] is the first step toward trust. Well, [20:29] for cooperation, there are three [20:31] pillars: trust, security, and openness. [20:35] I deeply agree and feel strongly about [20:37] them. Let me give an example to show [20:41] the level of partnership between Weibo [20:43] and Alipay. When any user asks a [20:47] health-related question in our Blue [20:49] Heart or Xiao Wei assistant, we provide [20:51] services unconditionally. That is what [20:55] we call trust. Throughout the process, [20:57] we didn't discuss any commercial terms [20:59] —nothing like that. We didn't talk [21:00] about it at all; it was just trust. As [21:02] Tan Kai mentioned earlier, it might be [21:04] because we’ve shared many meals, but [21:06] the truth is we trust him because he is [21:08] a trustworthy person. Because over [21:11] these many years, he has never done [21:12] anything untrustworthy, and that is [21:14] very important. Mr. Ren, before we [21:20] started, we chatted briefly backstage, [21:22] and one thing you said left a deep [21:24] impression on me: car shipment volumes [21:26] cannot be compared to phone shipment [21:28] volumes, so in the previous era, cars [21:30] were quite passive in the traffic era [21:32] and lacked much say. But it seems [21:36] different now. Facing this new [21:38] ecosystem, what kind of rights and [21:40] responsibilities system do you hope to [21:42] build? [21:48] >> In the past, during the era of [21:49] information payment, the automotive [21:51] sector was actually quite lagging [21:53] behind in terms of the payment [21:55] ecosystems involved. Service fees have [22:02] never really been charged to users; [22:04] they have generally been free. Cars are [22:12] relatively unique, but in the next [22:14] phase, I believe they must [22:16] fundamentally evolve from the [22:17] information and traffic era toward a [22:19] service and delivery-based economic [22:21] model—a different way of paying. We [22:31] don't focus on how to distribute value [22:33] first, but rather on whether we have [22:35] the opportunity to create more value [22:37] for users under this model, making them [22:39] willing to pay for the end-to-end [22:41] delivery and services we provide. That [22:45] is the first step. The difference with [22:47] cars is that with a phone, you might [22:49] only have it do one thing. But in a car [22:52] , you are asking it to do many things. [22:55] At the same time, the one ultimately [22:57] responsible for the outcome is the car [22:59] itself. For phones, if a service isn't [23:03] done well, it's the app developer's [23:05] responsibility. If Alipay has an issue, [23:09] you wouldn't blame OPPO; you would go [23:11] to Alipay. But cars are different—all [23:14] problems that arise with the car are [23:16] the responsibility of the automaker. To [23:18] give an example, in this new era—our [23:21] new era of intelligence—consider a [23:23] new scenario: what happens if a driver [23:25] or passenger is suffering from [23:27] prolonged fatigue or physical [23:29] discomfort? In the past, there was no [23:32] solution, but in the future, in the [23:34] Agent era, in the AI era, the vehicle [23:37] will be able to actively detect the [23:39] driver's health or physical state. [23:42] Because a car is a highly embodied form [23:44] of artificial intelligence; it [23:46] possesses perception, planning [23:48] capabilities, and powerful ecosystem [23:50] integration. At that point, when I [23:55] detect that they are feeling unwell, I [23:57] will directly activate my intelligent [23:59] driving function to take them to the [24:01] nearest hospital. This is where an [24:05] Agent's large model might come into [24:07] play, serving as just one component [24:09] within my ecosystem. It will [24:13] proactively diagnose the driver's [24:15] condition based on their actual state, [24:18] book a hospital appointment, and locate [24:20] the relevant specialist—all tasks the [24:22] car can complete automatically. In this [24:29] context, it is about ecological synergy [24:31] along a chain, not just a single [24:33] service provider in a niche. So, what [24:40] the customer perceives is not just the [24:43] value of a single service or piece of [24:45] information, but the end-to-end [24:46] resolution of a series of problems. If [24:54] more customers are willing to pay for [24:56] this value in the future, then true [24:58] value is being created. As for how to [25:02] divide the gains, how should it be done [25:04] ? I believe it should be distributed [25:06] based on value contribution. It must be [25:09] a true distribution based on value [25:11] contribution: who provided the core [25:13] data during this process? Who provided [25:15] the primary core decisions, and who [25:17] provided the final delivery to the user [25:20] ? The market should provide the best [25:23] pricing mechanism for this. Build first [25:25] , then talk about profit sharing. Given [25:27] our limited time, I’d like to ask [25:29] everyone one final question. Today we [25:32] have a very grand topic, which is that [25:34] Agents will usher in a new era of [25:36] commerce. What is the one trend you are [25:38] certain will happen in the next one to [25:40] two years? What kind of new economy do [25:43] you think will emerge when Agents reach [25:45] the tens of millions level? Assuming [25:48] this new economy doesn't appear, where [25:50] might the bottleneck be? Let's start [25:53] with Geely's Ren. I think, for example, [26:04] for Agents to go from the current state [26:06] to the tens of millions level, the [26:08] difference might be that a large number [26:10] of Agents providing atomic-level [26:12] services will emerge, which is [26:14] different from the past. In the past, [26:25] services were mostly provided by a few [26:27] large app providers, but if we reach [26:29] the tens of millions level, it must [26:31] mean that Agents representing services [26:34] developed by many individuals or small [26:36] teams will appear. So at this point, [26:43] wouldn't a form called the "Agent [26:45] Internet era" emerge? At this time, [26:50] what the terminal actually needs to do [26:52] becomes even more important; its [26:54] integration capabilities and its master [26:56] Agent capabilities will highlight its [26:58] significance. Take cars for example: if [27:03] they are to adopt future AGI forms, [27:05] their technical structure will [27:07] certainly have to change. They might [27:14] need to develop a dedicated AIOS to [27:16] support these tasks, and the master [27:18] Agent will focus on task delivery, [27:20] calling upon peripheral Agents to [27:22] provide users with better services. [27:30] During this period, I believe the [27:32] vitality and creativity brought by this [27:35] intelligence will surge, and the user [27:37] experience brought by the terminal will [27:39] also be greatly improved; I think new [27:41] economic models centered on this will [27:43] definitely bring about some new changes [27:45] . [27:49] >> Understood, thank you. Ren. My [27:54] perspective is a bit unique. I think [27:58] the business model of today's mobile [28:00] internet is essentially an attention [28:02] economy. As the Agent ecosystem becomes [28:07] more mature, people's attention should [28:09] shift; people should be engaging in [28:11] exercise and health, rather than [28:13] continuing to be stuck in tedious work. [28:18] So in this regard, I think the arrival [28:20] of this new era will give us more time, [28:22] in a certain sense. That is why I [28:25] believe Ant's work on "A-Fu" is very [28:27] forward-looking. [28:30] >> Zhou. [28:33] >> I look at this matter from another [28:35] angle, and I still feel like I'm [28:37] walking on thin ice. It's the same [28:41] medical consultation, the same car. If [28:45] it's still the same thing, the user's [28:47] basic needs for it next year won't be [28:49] any different from the last ten years. [28:54] But why are we still so expectant of [28:56] this brand-new thing happening? It's [28:59] the same car wash, the same request, so [29:02] why would it be different? The [29:05] fundamental difference isn't the [29:06] service itself, but that the foundation [29:08] supporting it has changed. This change [29:12] is precisely the large model, let's [29:14] call it "Harness"—it’s more [29:16] AGI-like and offers a better experience [29:19] . But will these services really be [29:23] different when they reach us next year [29:24] compared to last year or the year [29:26] before? Well, what I see is that the [29:30] foundation supporting this is actually [29:32] still on thin ice. For one, the massive [29:37] shift in "Harness" on PCs today and the [29:39] explosive growth in productivity don't [29:41] actually come from the business tasks [29:44] themselves, but from a revolution in [29:46] productivity. I can explain this by [29:50] mentioning just a few metrics. The [29:53] reason "Harness" on PCs today is [29:55] different from last year's is that it [29:58] has over 30,000 full-scale system [30:00] adaptations, whereas that work has only [30:02] just begun on mobile phones. If we want [30:06] to have that same great experience next [30:08] year, I believe not a single one of [30:10] those 30,000 adaptations can be missed. [30:14] This is what I see, so I see more of [30:16] the time we need to race against, the [30:18] debts we need to repay, and the work we [30:20] still have to do. I spend relatively [30:23] little time thinking about the [30:24] commercial side of things. And for [30:26] these 30,000 skills that cannot be [30:28] missed, how long do you expect it to [30:30] take to finish them? [30:32] >> At least a year is a very optimistic [30:34] figure. Moving to Mr. Han, Mr. Tang [30:36] raised a very important issue: he said [30:39] that current industry coordination and [30:41] organizational capabilities aren't [30:43] keeping up with model capabilities. [30:46] Give me a number—how long do you [30:47] think it will take to catch up? [30:50] >> To be honest, given the current [30:52] development of technology, I think we [30:54] might never catch up. But the upside is [30:58] that I believe it could happen much [31:00] faster than any of us imagines. As [31:03] mentioned earlier, if you look at the [31:05] progression from this year’s "Llama" [31:07] to "Harness," it’s only been a few [31:10] months. If you look further back, from [31:13] GPT-3 to now, it’s only been about 20 [31:16] months. The speed of development is [31:18] extremely fast. I think that is the [31:20] first point. It’s hard to have [31:23] certainty with technology, but the only [31:24] thing I’m sure of is that it might be [31:26] faster than any of us expects. Secondly [31:31] , I’m still thinking that another way [31:33] to accelerate this is to find a way to [31:35] build a commercial closed loop, so that [31:38] all participating parties can see a [31:39] commercial incentive throughout this [31:41] process. For example, as both Mr. Tang [31:47] and Mr. Zhou mentioned earlier, we can [31:49] invest for the sake of users, scale, [31:51] and experience, but quite a few [31:53] merchants and developers actually hope [31:55] that their offerings can be utilized [31:58] and generate revenue for them. Only [32:01] with such a closed-loop business model [32:04] is it possible to get more people [32:05] involved and form a positive cycle. [32:09] That is also why I think from Alipay's [32:11] perspective, what we need to do is [32:13] become more open and do a good job with [32:15] our underlying capabilities. [32:18] Furthermore, this ecosystem likely [32:20] cannot be designed by us or any single [32:22] company, but rather must be co-created [32:24] by all participating parties. I hope [32:28] that in the foreseeable future, this [32:30] ecosystem will grow faster and better, [32:32] and be able to attract more people to [32:34] join in. Now, let me ask one last quick [32:38] question that I believe everyone is [32:40] concerned about: when do you expect [32:42] merchants and the platform side to [32:44] perceive the actual value of the policy [32:46] implementation, and how soon will users [32:48] be able to feel it? [32:52] >> Uh, I’ll be bold and say that I think [32:54] some merchants might be able to feel [32:55] the initial impact within three months. [33:01] Then, perhaps around the six-month mark [33:03] , which is roughly after the Spring [33:04] Festival next year, it’s possible [33:06] that more and more merchants will start [33:08] to join in. If that’s the case, we [33:13] believe a turning point, or a real [33:15] inflection point, could possibly emerge [33:17] in the second half of next year. Great, [33:20] that gives everyone a lot of confidence [33:22] . Thank you. It’s an honor to be here [33:24] today to record these thoughts at such [33:26] a critical moment. I look forward to [33:28] meeting again in the future. Thank you [33:29] all. Yeah.