雷递网 乐天 5月19日5月19日在上海举办的 AMD AI 开发者日 2026AMD AI DevDay 2026上零一万物CEO李开复博士与AMD董事会主席及首席执行官苏姿丰博士展开了一场主题为“AI智能体新范式”的巅峰炉边对话。两位长期站在 AI 技术与产业基础设施前沿的行业领军者围绕智能体、多智能体协作、开源生态、企业 AI 转型、开发者角色演进与下一代算力基础设施等议题展开了深入交流也对与会的技术人提出呼吁。这场对谈把一个十分迫切的具体问题摆到了开发者和企业面前当智能体真正进入核心业务一线技术架构、组织责任、算力部署和商业回报将如何被重新定义从单个 Agent 的能力边界到企业核心职能的重构从开发者个人能力的放大到本地智算基础设施的重要性这场对话试图回答的正是 AI 从“可用”走向“可交付结果”之后的新命题。如果说过去两年行业关注的是大模型能否完成任务那么接下来更值得关注的是AI 能否稳定承担目标、协同执行结果并真正影响企业经营。李开复博士与苏姿丰博士的讨论不仅呈现了 AI 技术路线正在发生的根本变化也揭示了下一阶段产业AI竞争真正的分水岭。以下为访谈全文2026年AI 的核心问题能否替代一个部门苏姿丰博士过去一段时间你一直在谈生成式 AI 正在迈向智能体时代。如今越来越多人开始认为2026 年可能会成为这一转变真正落地的一年。那么站在现在的时间点你观察到了哪些变化从而让你认为智能体与此前的生成式 AI 浪潮已经有了本质区别李开复博士非常高兴来到 AMD 上海开发者大会也很高兴能和真正“在创造未来”的开发者们坐在一起。我认为有两件关键的事情发生了变化其中第二件尤为重要。首先AI 编程能力跨过了临界点。一年前AI 还只能辅助编写代码、函数等等而现在它已经可以端到端地交付一整套功能。这听起来像是一个渐进式的进步但其实不然。在座各位都知道智能体在数字世界中的所有行为本质上最终都会落到代码层面。一旦 AI 的编码能力跨过那个门槛自主智能体就真正具备了成为现实的可能。其次更重要的变化在于我们开始意识到单一智能体的能力是有上限的。无论模型参数规模有多大只依赖单个 Agent 的推理能力在面对真实复杂问题时终究会碰到瓶颈。而多智能体架构第一次打破了这个上限。负责规划、评估、研究和执行的不同智能体开始彼此协作、相互辩论并在彼此结果之上继续迭代。这其实非常接近“美第奇效应Medici Effect”当不同领域的专家被放进同一个房间时最终产生的成果会远远超过任何单一个体能力的简单叠加。五百年前在文艺复兴时期人类已经发现了这一规律。直到21世纪的今天我们第一次把这种机制带到了 AI 世界。从技术路径上看这意味着我们正在逐渐摆脱过去那种“试图用一个模型完成所有事情”的模式。未来的 AI不会是一个“超级大脑”的独角戏而会更像一个由不同智能系统协同运作的交响乐。正是基于这一趋势我们着手部署专业化的多智能体系统并逐渐走向“异构智能Heterogeneous Intelligence”阶段。不同类型的模型与算法会被组合在一起用群体智能去解决更加复杂的问题。2024 年最关受关注的问题是“AI 能否完成一个任务”2025 年这一问题转变为“AI 能不能完成一整条工作流”在2026年这个核心问题已经进阶为“AI 能否替代一个企业的职能部门”以现代HR人力资源部门为例。当招聘 Agent 与绩效 Agent 实现联动后系统就能够根据员工入职后的真实绩效数据自动调整前端的人才筛选标准。从简历筛选、面试到新员工入职再到月度和季度绩效自动化跟踪这些多智能体系统会围绕统一的人力资源数据持续运转升级。随着这种能力不断扩展它最终会演变成一个彼此互联的企业多智能体协作网络覆盖 HR、研发、产品、销售和市场等不同部门。这种架构也正在推动“One-Person Company一人公司”趋势的出现。借助模块化的多智能体框架单个开发者或领域专家如今已经有能力像“总架构师”一样快速启动一家高度自动化运转的公司。在由智能体驱动的新范式下我们实际上已经跨过了“自主执行”的门槛。AI 正在从过去被动的“Prompt-and-Response从提示词到响应”模式转向主动的“Goal-and-Execution从目标到执行”模式。未来你不再是给 AI 一个 Prompt而是直接给它一个组织目标。随后智能体们会自行完成协同、执行、评估、优化并形成完整闭环。而这一新范式也正在催生当前 AI 领域最巨大的商业机会产业级 AI 转型。新时代真正的经济价值不会来自只会“回答问题”的 AI 系统而会来自能够真正执行企业目标的自主多智能体基础设施。这也是零一万物所关注的核心方向。过去一段时间里我一直在与全球各地的 CEO 和企业高管交流以便更深入地理解AI 将如何重塑生产力、组织结构以及未来的领导力。同时我想这也会影响今天在场每一位开发者驱使大家重新思考自己未来在 AI 时代会扮演怎样的角色。AI 转型为什么不能只靠 CIO苏姿丰博士在你与 CEO 们的交谈中他们是如何对待 AI 转型的这对开发者社区有什么影响李开复博士我看到了许多明显的问题。几乎每个企业目前都选择在不出错却价值很低的场景部署 AI。比如会议纪要、人力资源员工答疑聊天机器人、企业内部搜索等等。这些都只是表面文章。我很直白地告诉各位 CEO不要只听你们的 CIO 。典型CIO们关注的是系统稳定运行、软件运行安全不出错在这一轮深入企业核心业务命脉的 AI 变革中反而可能成为阻碍进化的旧势力。因为 CIO 的职责本质上是管理软件运营而不是重新定义公司。CIO 擅长安全地部署 AI但并不擅长推动组织层面的真正变革。多数由 IT 部门自下而上推动的 AI 转型最终都会失败。传统 CIO 这个角色不会消失但它的重要性会被大幅削弱。因为 AI 并不只是一个新的软件工具企业AI转型绝对是是一把手工程需要企业领袖根本性的思维转变。真正能够改变公司经营结果的往往是那些直接影响损益表PL的核心业务环节。而这些领域恰恰也是很多高管最不愿意让 AI 介入的运营职能部门收入、利润、防欺诈、动态定价、供应链、产品上市速度以及核心创新能力。具有前瞻性的 CEO 们正在重新校准他们公司的运营方式、组织应如何改变以及领导方式应该如何调整。我也经常对 CEO 们说如果你的 AI 部署最终没有改变任何一个会出现在季度财报电话会议上的数字那么你公司做的就不是真正意义的 AI 转型只是浪费钱打造了一个 AI 实验室。同样的话我也想送给今天的开发者。任何参与商业研发的人都应该用同样的方式思考问题。停止浮于表面的表演式 AI 开始构建能真正深入业务实质的结构性引擎。中国的开源生态将催生产业AI的“安卓系统”苏姿丰博士现在最让我兴奋的事情之一就是开源 AI 社区正在涌现出大量创新而且这个生态已经越来越全球化。你与中国的开源社区一直保持着非常密切的联系。那么在这个生态中开发者和贡献者们最近最让你感到兴奋的变化是什么李开复博士开源的趋势势不可挡它从根本上重写了全球 AI 的游戏规则其发展格局与经典的智能手机大战如出一辙。闭源模型类似于苹果的 iOS追求高利润并保持着对生态系统的强硬控制。而开源社区则越来越像 AI 世界里的 Android。它拥有更广泛的全球覆盖以及更大的用户规模。正如 Lisa 刚才提到的中国开源生态之所以表现得如此出色背后存在着很深层的结构性原因。因为硬件资源有限中国开发者和创业公司并没有条件依赖“大力出奇迹”的算力堆叠。在这种约束下整个生态反而开始把重点转向极致的工程效率更加关注算法优化、架构创新以及如何把底层基础设施做得更精简、更强大。它就像一个充满活力的、去中心化的学习小组大家齐心协力为了在考试中取得好成绩每个人都在其他人公开发布的成果之上进行创造整个群体的能力也因此呈指数级增长。我确信这种机制将为未来带来更多的进步与创新。DRI模式重新定义技术人人人都要把自己当CEO苏姿丰博士在 AMD我们自己的工程师也在使用 AI 智能体加速产品设计和验证流程。我们越来越明显地看到今天一个人如果拥有合适的工具和足够的算力已经能够完成几年前需要整个团队才能完成的工作。你就“人与智能体协作”的趋势写了很多文章那么在那些真正以这种方式创业和开发的人身上你最近观察到了哪些变化李开复博士受限于此前的训练大多数开发者都习惯于在代码层面思考所有权Ownership问题。比如由一个人负责 GitHub 上的代码仓库和 PRPull request另一个人负责值班轮换另一个人负责某一个具体服务。这种责任边界其实是有边界的。它本质上只覆盖你能够通过键盘直接控制的部分而现在越来越多编码工作已经开始被 AI 智能体接管。我想和大家谈谈一种打破这个边界的运营架构DRIDirectly Responsible Individual直接责任人概念。在软件工程中交付产品的主要瓶颈很少是代码本身。而是所有权的模糊不清。责任分散、停滞的拉取请求以及偏离的路线图通常都源于很多人只是负责项目管理大表上的某一个环节却没有人真正对最终结果负责。DRI 模型改变了这一点。我预测 DRI 模型会成为 AI 原生公司最核心的组织架构。所谓 DRI就是由一个人对某个跨职能结果承担端到端责任。这不是一个职位头衔而是一种非常明确的责任机制。就像系统运行手册里唯一指定的值班工程师最终结果如何、业务影响如何都由 DRI 责任人负责。在这个模式下一个人类 DRI 会处于整个智能体系统的中心。围绕他协同工作的是由研究、执行、合规和监控等不同 Agent 组成的专业化集群。DRI 不把时间精力花在具体执行上而是负责整体编排、关键决策并对最终的输出契约负责。与此同时实时数据流会逐渐取代传统的汇报体系业务运转也会越来越围绕具体、可量化的结果展开。我认为一个优秀工程师所具备的很多能力和成为优秀 DRI 所需要的能力几乎是高度一致的。当你编写技术规范时你其实是在尝试定义可量化的商业成果当你给系统做监控、配置自动告警时你其实是在建立衡量结果的机制。当你凌晨两点主动去 de-bug排查故障而不是等别人通知你时你展现出来的其实正是 DRI 模式最核心的主人翁意识。选择 DRI 模式也意味着你必须重新定义“什么是个人成功”。在智能体时代一个优秀工程师的价值不再只是由“写了多少代码”来衡量。这也意味着今天很多工程师的工作方式都会发生变化。你不再只是关注系统而是要对结果负责。优秀工程师通常都会非常重视监控系统使服务具备极强的可观测性。DRI 则是把这种技术严谨性延伸到了他们所拥有的业务结果上。如果你是一个负责增长的 DRI你不仅仅监控 API 延迟。你还要监控用户激活率、转化漏斗以及对收入的影响。你端到端地对完整结果负责。你拥有决策权而不仅仅是建议权。 工程师通常很擅长分析但他们常常把产出物交给产品经理或高管去做选择。DRI 则需要自己完成闭环。你进行分析你做决定并对接下来发生的任何事情负责。刚开始会有些不习惯但很快你就会进入状态。你会有规划性地去配置你的智能体集群。 大多数非技术的 DRI 会把 AI 智能体当作黑盒来对待。但工程师不一样。工程师们发挥技术能力去监控智能体、评估它的输出、识别它的故障模式并懂得如何围绕智能体集群建立更可靠的验证机制。工程师的优势会在这个时代被无限放大。你们拥有的不仅仅是给产品或业务的建议权而是直接拥有决策能力。AI 正成为赋能技术人的新形态超能力。AI 正成为赋能技术人的新形态超能力。智能体经济爆发前夜推理算力走向前台苏姿丰博士你刚才谈到的这一切背后都需要非常庞大的算力支撑。而且需要的还不只是单一算力而是一整套能够协同工作的全栈算力体系。那么当开发者和企业真正开始大规模运行智能体时底层算力基础设施究竟需要具备哪些能力才能支撑这一切真正运转起来李开复博士Lisa这已经完全进入你的专业领域了。智能体 AI 趋势底层的计算模式正在从根本上改变底层的计算模式。传统 AI 系统所要求的更多是稳定、持续的计算负载而智能体系统则完全不同它具有高度突发性而且会产生大量并行计算。一个用户请求可能会被拆分成20个或更多并行运行的智能体这些结果汇总之后又会再次触发下一轮 Agent 协同。从本质上看智能体经济是“推理驱动”的经济而推理与训练其实是两种完全不同的计算模式。要让多智能体协同真正具备现实可行性系统必须满足几个条件本地优先、端侧处理以及低于 100 毫秒的响应延迟。而这正是当前硬件竞争真正分出胜负的地方。我认为在这一趋势上AMD 比很多公司都看得更早、更清楚。随着 AI 开始走向多智能体架构我们也必须重新思考“算力”本身。未来极致的 token 效率以及本地化处理能力会是关键。自主企业将诞生数据主权与ROI成为产业AI新坐标苏姿丰博士我想用今天早上大会开场白环节的一句话作为结束AI 时代仍然处于非常早期的阶段真正精彩的部分其实还在后面。展望未来开发者接下来最有可能创造出的东西里什么最让你感到兴奋李开复博士未来真正意义上的“自主企业”会诞生。驱动它的将是跨部门、多层级协同运作的智能体网络。下一阶段的产业 AI 转型会同时围绕两个核心问题展开数据主权以及清晰可验证的 ROI投资回报率。类似 AMD 的头部合作伙伴正是构建“智算主权”的关键地基。对于今天的开发者来说最大的机会是去构建那些过去需要一个完整团队才能完成、如今却可以由 AI 独立交付商业结果的 AI。AI 的角色已经不再是“帮助营销人员提升效率的 AI工具”而是能够真正承担营销职能的 AI Agent不是“协助金融分析师的 AI工具”而是能提供自动化财务分析的 AI Agent。我曾与零一万物的工程师紧密合作构建了一个“开复 AI”作为我个人的决策智能体。我们发现在大型企业中推动 AI 落地最快的方法往往是由 CEO 或董事长自上而下推动。因为一旦 CEO 或 CFO 真正开始使用这些智能体他们很快就会离不开它。当管理层真正接受智能体之后AI转型自然会沿着企业组织结构不断向下推进。如果你正坐在这个会场里带着一台笔记本电脑对系统编排有所了解并有一个大胆的想法那么你现在会比世界上任何一家财富500强企业的战略部门都更有优势。这一代开发者正站在一个极其少见的时代窗口面前。在这样一个时代里最不应该做的就是把自己的创造力提前锁进一家大公司的组织体系里。去创造属于你自己的事业吧Cube01多智能体时代的智算基础设施李开复博士我们一直非常重视开发者的使用体验。这次在 AMD 强大的硬件能力基础上我们进一步整合了零一万物“万智企业大模型平台”的模型能力与工具体系。万智不仅内置了多种领先模型同时也能够直接打通企业知识库与核心业务工作流。这意味着开发者可以根据不同场景灵活选择最合适的模型并快速把 AI 能力真正接入自身企业研发体系。接下来我想重点介绍 Cube01 的三个核心能力。我们认为这也是它在智能体 AI 开发中的真正价值所在。以下三大核心特性我相信会让 Cube01 成为你开发智能体项目时独一无二的利器•多智能体编排层Multi-Agent Orchestration Layer 万智的多智能体框架实现了毫秒级的响应速度。这款“一体机”One-Box解决方案允许自主工作流持续运行并动态调用不同工具对多个 Agent 进行实时协同编排。你的 AI 智能体将保持“永远在线”always-on。•AI 员工实体映射AI Worker Entity Mapping 与传统的 AI 机器人不同系统会为每个 AI 智能体分配一个独立的“数字身份”。你可以为它配置真实的业务权限、数据权限以及组织角色。AI 智能体能够精准识别业务上下文并主动介入协作工作流。这样一来AI 智能体就能像人类一样在组织内部嵌入各自的虚拟角色。它们每一个都会成为独立的生产力节点并与其他智能体协同合作来完成任务。•算力主权与安全Compute Sovereignty and Security 通过与 AMD 的此次合作Cube01 将高昂的云计算算力转化为自主可控、安全可靠的本地基础设施从而为企业和组织确立了“智算主权”。过去大家可能一直在为云计算按量付费就像打网约车一样。而现在你拥有了一辆绝对安全的专属专车。最后我还是想再次强调我鼓励所有的工程师都要带着“结果导向”的思维去进行开发。我们非常期待看到大家借助 Cube01 蜕变成为一名能够独立负责结果在 AI 时代组织里不可或缺的 DRI。谢谢大家以下为英文访谈内容Dr. Lisa Su: You have been writing and talking about the shift from generative AI to agentic AI for a while now. And it feels like 2026 is the year it is actually happening.What are you seeing right now that tells you this moment is genuinely different from the generative AI wave we just lived through?Dr. Kai-Fu Lee: Thank you, Lisa. It’s an absolute pleasure to be here at the AMD Shanghai Developer Day. Its great to be in a room with the people who are actually building the future.Two things changed, and the second matters more.First, coding got past a threshold. A year ago, AI could write functions. Now it can ship features end-to-end. That sounds incremental, but it isnt. Everyone in this room knows that everything an agent does in the digital world is code underneath. When coding crossed the line, autonomous agents became possible.Second, and this is the bigger one: we figured out that a single agent has a ceiling. No matter how big the model, one agents reasoning alone hits a wall on real problems. Multi-agent architectures broke that ceiling. Specialized agents—a planner, a critic, a researcher, an executor—debating and building on each other. Its the Medici effect: put diverse specialists in one room, and the output exceeds any individual. The Renaissance figured this out for humans 500 years ago. We just figured it out for AI.Technically, we have achieved this by moving away from a single, fragile LLM trying to do everything. Instead, we deploy specialized multi-agent systems. We are also moving toward heterogeneous intelligence, which combines distinct types of AI models and algorithms to solve complex problems.So, if 2024 was Can AI do a task? and 2025 was Can AI complete a workflow? then 2026 is strictly Can AI run a corporate function?Take a modern HR department as an example. A recruitment agent that syncs with a performance agent enables the system to automatically adjust upfront talent-filtering criteria by analyzing post-hire employee performance data. From automated resume screening to interviewing, and from onboarding new employees to tracking performance monthly and quarterly. These multi-agents operate on an integrated HR data flywheel. Ultimately, these functional layers scale into an interconnected web of HR, RD, product, sales, and marketing agents.This architecture fuels an emerging One-Person-Company trend. By utilizing modular, multi-agent frameworks, a single developer or domain expert can now function as a macro architect to kickstart a highly functional company.In this new paradigm powered by agentic AI, we have crossed the threshold into autonomous execution. We are moving from a passive Prompt-and-Response model to an active Goal-and-Execution model. You do not give an AI agent a prompt; you give it an organizational objective. The agents then coordinate, excute, measure, optimize, and close the loop.This new paradigm is empowering the largest commercial opportunity for AI today: Industry AI Transformation. The next era of economic value wont be driven by systems that answer your questions. It will be driven by an autonomous, multi-agent infrastructure that executes your corporate goals.At 01.AI, this is the exact core of our focus. Ive been engaging with CEOs and senior executives around the world to gain deeper insights into how AI will shape new productivity, new organizations, and new leadership. It will also direct how developers here today to think about their role in building for the future AI economy.Dr. Lisa Su: In your conversations with CEOs, how are they approaching AI transformation? And what are the implications for the developer community?Dr. Kai-Fu Lee: I see many obvious mistakes. Almost every enterprise is deploying AI in safe and trivial places right now. Meeting summaries. Chatbots for HR FAQs. Internal search. All of it sub-scale. All of it cosmetic.I tell CEOs bluntly: Dont Listen to Your CIO.Because CIOs are here to protect the software environment, not to re-invent the company. CIOs are good at deploying AI safely, not in transforming their organizations. The bottom-up approach from the IT department often fails. The traditional CIO role won’t disappear, but it will be vastly diminished in importance.AI is not just another software addition. The scale of AI transformation demands a fundamental leadership shift.The functions that actually move the dial are the ones that fundamentally change a companys PL. They are often the exact operational functions executives are afraid to touch with AI: revenue, profits, fraud, dynamic pricing, supply chain, time to market, and core innovation. Forward-thinking CEOs are recalibrating how their companies operate, how their organizations should change, and how they should lead differently.I also tell CEOs: if your AI deployment doesn’t change a number that shows up on your quarterly earnings call, you’re running an AI Lab, not AI Transformation.My call to action for developers is this: anyone involved in RD for commercial organizations should think exactly the same way. Stop building cosmetic wrappers. Start building structural engines that move the needle for the bottom line.Dr. Lisa Su: One of the things that excites me most right now is how much innovation is happening in the open-source AI community... and how global it has become.You are very close to the open-source community here. What are you seeing from developers and contributors in this ecosystem that excites you most right now?Dr. Kai-Fu Lee: The open-source landscape is completely unstoppable, fundamentally rewriting the global AI playbook in a dynamic mirroring the classic smartphone wars. Proprietary models resemble Apple’s iOS, which pursues high margins and tightly controlled ecosystems. The open-source community has become the Android of AI, capturing a massive global footprint and adoption.There is a structural reason why the Chinese ecosystem has excelled so fiercely here, as Lisa pointed out earlier. Due to limited hardware resources, Chinese developers and startups couldnt afford brute-force compute. Out of absolute necessity, the community leaned into extreme engineering efficiency. The focus shifted toward algorithmic optimization, architectural innovation, and making foundational infrastructure incredibly lean and powerful. It operates like a dynamic, decentralized study group working together to ace a test, where everyone builds on top of each others public releases. Their collective capability rises. I am sure this will lead to more improvement and innovation down the road.Dr. Lisa Su: At AMD, our own engineers are using AI agents to accelerate how they design and verify our products. And we are seeing what one person with the right tools and the right compute can do today that would have taken a whole team a few years ago. You have written a lot about this idea... one person plus a team of agents.What are you seeing from founders and developers who are actually building this way right now?Dr. Kai-Fu Lee: Most developers are trained to think about ownership at the code level. You own the repo and the PR on GitHub. You own the on-call rotation. You own a specific service. But theres a ceiling on that kind of ownership—its strictly scoped to what you can directly control through a keyboard, and coding agents can increasingly replace that.I want to talk to you about an operational architecture that shatters that ceiling: the concept of the DRI—the Directly Responsible Individual.In software engineering, the primary bottleneck to shipping products is rarely the code itself. It is ownership ambiguity. Diffuse accountability, stalled pull requests, and drifting roadmaps often stem from the fact that individuals only own items on a project management chart. The DRI model changes that.I predict that the DRI Model will become the definitive organizational architecture for AI-Native companies. A DRI is a single human who explicitly owns one cross-functional outcome end-to-end. This is not a title; it is a clear operational contract. Acting like a single designated on-call engineer in a service runbook, the DRI is entirely accountable for the final result and the ultimate business impact.In this model, a single human DRI sits at the center of a specialized swarm of research, execution, compliance, and monitoring agents. The DRI does not spend time on manual execution. They orchestrate, decide, and own the final output contract. Real-time data pipelines replace traditional reporting latency. Business activities are executed in concrete, quantifiable outcomes.The skillset that makes a great engineer maps almost perfectly to what makes a great DRI. When you write a technical spec, youre defining outcomes. When you instrument a service and set up automated alerts, youre building measurements into an outcome you own. When you debug a production incident at 2 AM without waiting to be told, youre demonstrating exactly the ownership instinct the DRI model is built on.Stepping into a DRI role means shifting how you define your personal success. In the agentic AI era, great engineering isnt measured by how many lines of code you write. This means a few things have changed in practice for many of you here today:You instrument the outcome, not just the system. Great engineers already instrument their services obsessively. DRIs extend that same technical rigor to the business outcome they own. If youre a Growth DRI, you dont just monitor API latency. You monitor activation rates, conversion funnels, and revenue impact. You own the full outcome end-to-end.You own the decision, not just the recommendation. Engineers are often great at analysis, but they frequently hand the output to a product manager or executive to make the choice. DRIs close that loop themselves. You do the analysis, you make the call, and you own whatever happens next. This can feel uncomfortable at first, but it becomes natural fast.You configure your agent swarm deliberately. Most non-technical DRIs will treat AI agents as black boxes. Engineers, however, will instrument them, evaluate their outputs, identify failure modes, and build better validation pipelines around them.This is where your engineering background becomes a genuine superpower.Dr. Lisa Su: Everything you are describing requires real compute... and a lot of it. Not just one type of compute... the full stack, working together.When you think about what developers and enterprises actually need to run agents at scale... what has to be true about the compute infrastructure to make that work?Dr. Kai-Fu Lee: Lisa, this is where it lands directly in your world. The compute pattern underneath the agentic AI trend is fundamentally changing. Agents are bursty and highly parallel, not steady-state. One user query fans out to twenty parallel agent calls, collapses, and fans out again. The agent economy is fundamentally an inference economy, and inference looks nothing like training.It requires local-first, on-device processing and latency under 100 milliseconds for multi-agent orchestration to feel real. That’s where the hardware question is being decided right now—and I think your team has read that landscape better than anyone.The shift to multi-agent architectures requires us to look at computing through the lens of extreme token efficiency and localized processing.Dr. Lisa Su: I want to close where I started this morning. The AI era is still in its early stages. The best is ahead of us.Looking to the future…what excites you most about what developers are building next?Dr. Kai-Fu Lee: What lies ahead is the birth of the truly Autonomous Enterprise, driven by multi-layered, cross-departmental agent networks. The future of industry AI transformation will be balanced by two uncompromising corporate imperatives: Data Sovereignty and Visible ROI. Partners like AMD are building Compute Sovereignty as the critical backbone to make this safe.The biggest opportunity open to you right now is to build AI that delivers a business outcome that used to require an entire team. We arent talking about AI that helps a marketer do their job, but an AI agent that is the marketing function. Not AI that assists a financial analyst, but an AI agent that delivers automated financial analysis.Ive worked closely with our engineers at 01.AI to build a Kai-Fu AI that acts as my personal decision-making agent. We are currently rolling out pilots with more enterprise leaders. We found that the fastest way to deploy AI in a large enterprise is through a top-down mandate from the CEO or Chairman. If you build the tools for CEO or a CFO, they get addicted. Once they are hooked, they drive the agentic transformation downward through their own corporate hierarchies.If you are sitting in this room with a laptop, an understanding of system orchestration, and a bold idea, you are in a better strategic position right now than the strategy department of any Fortune 500 company in the world.Don’t trade this historic position for a conventional job at one of them. Just build.Dr. Lisa Su: Thank you, Kai-Fu.Dr. Kai-Fu Lee: We value the ease of use for developers. On top of this powerful AMD foundation, weve integrated the brain and toolkit from the 01.AI Worldwise Enterprise LLM Platform. Worldwise not only embeds a wide range of leading models, but also connects the enterprise knowledge base directly to core business workflows. This allows developers to choose the right models and adapt them to a specific RD use case.Three key features that make Cube01 unique and powerful for your agentic AI projects:•Multi-Agent Orchestration Layer: Worldwises Multi-Agent framework enables millisecond-level response times. This One-Box solution allows autonomous processes to run persistently, and dynamic tools to orchestrate agents seamlessly. Your AI Agents will be always-on.•AI Worker Entity Mapping: Unlike traditional AI bots, the system assigns each AI Agent a distinct digital identity. You can configure them to map with real business permissions and assets. AI Agents can precisely identify business context and proactively step into collaborative workflows. This way, AI agents can plug into virtual roles within the organization just like humans. Each of them becomes an independent productivity node and collaborates with the other Agents to get things done.•Compute Sovereignty and Security: Through this collaboration with AMD, Cube01 transforms high-cost cloud computing into controllable and secure local assets, establishing Compute Sovereignty for companies and organizations. You have probably been paying as you go for cloud computing, like riding a Didi. Now you have a safe personal limousine.Again, I encourage all engineers to develop with an Outcome mindset. We would love to hear how you all are evolving into a DRI with Cube01.———————————————雷递由媒体人雷建平创办若转载请写明来源。