提问方：某世界500强企业
回答方：快思慢想研究院院长 田丰
中文版
问题一：你如何看待未来五年人工智能技术的发展趋势？
田丰：
趋势1.微型AI设备普及化
2纳米芯片将催生微型、可穿戴、无屏幕的人工智能终端，例如OpenAI的耳后音频设备SweatPea、智能笔；带屏幕的终端也将变得更轻便，例如电池续航更长、算力更强、嵌入了空间智能体的增强现实眼镜。这些新的人工智能终端能自主感知环境：智能笔可感知并记录手写文字，而增强现实智能眼镜能主动检测二维码并识别交互对象。到2030年，全球领先国家90%的物联网终端将嵌入人工智能芯片和智能体。
趋势2.智能体形成"价值网络"
互联网巨头正在构建统一的通用智能体门户，而初创团队则开发用于会议/办公、编程、医疗、教育、人力资源、金融、法律、广告/营销等领域的专用智能体。通用智能体根据任务特性调用和组合专用智能体，并按照Token使用量共享收益。到2030年，跨产业链的智能体之间将出现"价值互联网"，Token将作为智能价值的载体实现跨组织流动。例如，公司A的智能体消耗库存，可以触发公司B的供应链智能体自动补货，最终提升品牌C的市场满意度。Token流通率（即Token在跨系统协作中创造附加值的速率）将成为衡量生态系统健康程度的关键指标。目前爆火的OpenClaw智能体间正在形成初步的自主流转价值网络。
趋势3.模型即服务（MaaS）平台涌现"模型工厂"与"智能体工厂"
云服务商成为“模型工厂”。云端模型即服务平台的应用开发模式，正从"API调用"演进为"工作流编排"和"自主模型网络"。2030年的变革在于：用户不再直接调用模型，而是将目标任务（如"优化供应链成本"）提交给智能云平台。平台自动分解任务、调度多模态模型、工具链及数据资源，生成定制化智能体并交付结果。模型如同"芯片中的晶体管"，被封装在智能体内，用户焦点从"如何调用"转向"效果与价值"。客服智能体按"客户满意度提升百分比"计费，营销智能体按"订单转化率提升"分享收益。借助模型即服务平台上的低代码工具，医疗专家、工程师、设计师等领域的专家，可以利用私有数据、私有技能(Skills)轻松微调出高度定制化的"微模型"，比如"某医院专用的胰腺癌影像筛查模型"。这些"模型微工厂"可通过模型市场进行交易或联合优化，同时确保数据隐私。
趋势4.AI芯片将变得非常廉价
随着摩尔定律、香农定律和冯·诺依曼架构遇到瓶颈，芯片和计算机架构将发生颠覆性创新，同时推动人工智能芯片成本持续下降。人工智能应用创造的价值将是计算成本的100倍，从而迎来多样化、创新型人工智能应用的爆发。
趋势5.AI生成视频互联网
2025年，在当前视频生成场景中，电子商务产品展示和个性化广告的AI生成视频成本比传统方法低100倍以上，社交媒体营销视频成本低50倍以上，品牌宣传视频成本低30倍以上。预计到2030年，得益于多模态大模型和世界模型的快速进步，全球互联网上70%的视频将由人工智能生成。AI视频在长达一小时的时间连贯性和情感表达方面将接近人类制作水平。AI视频将在漫剧、短视频、广告、游戏、电影等领域跨越“图灵测试”里程碑。
趋势6.自然语言正在成为新一代编程语言
当前的AI编程工具主要提供代码补全功能。到2030年，自然语言将取代传统编程语法，成为主要的交互方式。开发者只需用自然语言描述需求，代码智能体就能生成完整、可部署的代码。此外，人工智能将理解复杂的项目上下文和业务逻辑，独立处理代码生成、测试和部署的全过程。人工智能将自主满足80%的开发需求，不仅能生成代码，还能深度集成图像、音频、视频的多模态生成式开发。新一代的开发者——具备行业知识的专家——将取代专业程序员，而传统程序员将进阶为软件架构师和产品经理。全球AI编程工具市场规模将从2024年的大约62.1亿美元增长到2030年的近1000亿美元，增长16倍。
问题二：人工智能进一步发展的动力是什么？瓶颈又是什么？
田丰：
AI发展瓶颈有三个：
1.算力瓶颈
当前GPU算力年增长约60%，但内存带宽年增长仅10%。"内存墙"导致计算单元60%的时间处于闲置状态。高端人工智能芯片制造成本每代增长30%，而性能提升从7纳米时的40%下降到2纳米时的15%，边际收益递减明显。
2.能源供应瓶颈
按当前趋势持续，到2030年人工智能将消耗全球10%的电力，但可再生能源仅能供应35%，存在65%的绿色能源缺口。数据中心电能使用效率（PUE）的优化已进入平台期，三年间仅从1.5降至1.4，冷却仍占能耗的40%。
3.训练数据瓶颈
高质量训练数据需求年增200%，但全球合规高质量文本数据供应年增仅15%，形成13:1且不断扩大的供需缺口。合成数据目前占训练集比例已超35%，但在关键任务上表现比真实数据差42%，存在显著的"真实性鸿沟"。
AI发展动力也有三个：
1.算力：突破在于异构架构创新
芯粒（Chiplet）技术通过堆叠提升集成度，而存内计算架构则突破"内存墙"，将数据搬运能耗降低90%以上。更根本的解决方案是将计算基础设施迁移至太空：利用宇宙低温环境进行自然冷却，并获取天基太阳能（效率是地球的7-10倍），可使计算单元成本降低100倍。马斯克指出，若星舰发射成本降至100美元/公斤，轨道数据中心将成为扩展人工智能算力的最优解。
2.能源供应：充分利用太阳能
短期解决方案依赖电网级储能（如特斯拉Megapack）进行"削峰填谷"，使现有电网吞吐量翻倍；长期则需建立天基太阳能卫星网络（目标：年部署100吉瓦），利用外太空24小时不间断且强度比地球高30%的日照。马斯克强调，中国凭借其年部署1500吉瓦太阳能的能力（2025年其新增电力的70%来自太阳能），预计到2026年发电量将是美国的3倍，从而在人工智能算力基础上获得领先优势。
3.训练数据：多模态实时数据闭环
马斯克认为基于互联网文本训练人工智能是低效的"认知捷径"。xAI的突破在于构建多模态实时数据闭环：X平台每日数万亿帖子提供了人类认知动态，特斯拉车队通过每秒数百万视频帧捕捉物理规律，Neuralink则探索神经信号等生物原始数据。下一代"世界模型"将直接学习重力、摩擦力等物理规律，使人工智能能够模拟宇宙的基本运行机制，而非从文本中推断现实。例如，Grok已能分析电路图错误并递归优化AI芯片设计。
问题三：在你的视频中，曾经提到独特的企业数据成为竞争优势。企业有效利用其数据应迈出的第一个实际步骤是什么？
田丰：按照Palantir的企业级“数据本体方法论”，第一步是将组织的所有数据连接起来，形成一个统一的、可操作的业务模型。这超越了技术集成——关键在于清晰定义客户、订单、资产等核心实体及其关系，从而有效构建企业的"数字孪生"。最终将结果直接转化为自动化工作流、警报和行动指令。这就在企业每个决策点创建了从洞察到执行的秒级闭环。
该图片可能由AI生成
问题四：随着我们从“数字AI”转向“物理AI”，你预计人们在社会或日常生活中首先注意到哪些日常变化？
田丰：发生的深远影响很多，至少从以下三个方面会有重大变革。
1) 劳动力结构调整
物理AI将首先取代重复性体力劳动、危险性体力劳动，机器人同事在服务业和制造业将变得普遍，使人类能够转向更具创造性和决策密集型的工作。
2) 成本与效率革命
从根本上说，随着机器人的规模化，商品和服务的边际成本将趋近于能源和原材料成本，从而大幅降低物流、清洁、护理等日常开支。
3) 人机协作新常态
像特斯拉Optimus这样的机器人将成为家庭和公共场所的"可编程伙伴"——从辅助老年人到在危险环境中作业——将生产力的概念从"人类劳动"转变为"人机协同"。
该图片可能由AI生成
英文版
Question1: How do you foresee the trend of AI technology development in next 5 years?
Tian Feng:
1.Proliferation of Miniature AI Devices
2nm chips enable tiny, wearable, screenless AI terminals, such as OpenAI's behind-the-ear audio device SweatPea and smart pens; screen-equipped terminals are also becoming lighter, exemplified by XR glasses with longer battery life, stronger computing power, and embedded spatial agents. These new AI terminals autonomously perceive their environment: smart pens sense and record handwritten text, while XR smart glasses actively detect QR codes and identify the next object for interaction. By 2030, 90% of IoT terminals in leading global nations will be embedded with AI chips and agents.
2.Agents Form a "Value Network"
Internet giants are building unified, general-purpose agent portals, while startup teams develop specialized agents for meetings/office work, programming, healthcare, education, HR, finance, legal affairs, advertising/marketing, etc. General agents invoke and combine specialized agents based on task characteristics, sharing commissions according to Token usage. By 2030, a "Value Internet" will emerge among agents across industry chains, with Tokens flowing cross-organizationally as carriers of intelligent value. For instance, Tokens consumed by Company A's R&D agent can trigger automatic replenishment by Company B's supply chain agent, ultimately enhancing Brand C's market satisfaction. The Token circulation rate (i.e., the rate at which Tokens create added value in cross-system collaboration) becomes a key metric for ecosystem health.
3."Model Factories" and "Agent Factories" Emerge on MaaS Platforms
Cloud providers become model foundries. Cloud-side Model-as-a-Service application development platforms evolve from "API calls" to "workflow orchestration" and "autonomous model networks." The 2030 transformation: users no longer directly invoke models; instead, they submit target tasks (e.g., "optimize supply chain costs") to the AI cloud platform. The platform automatically decomposes tasks, dispatches multimodal models, toolchains, and data resources, generating customized agents and delivering results. Models, like "transistors in a chip," are encapsulated within agents, shifting user focus from "how to call" to "effectiveness and value." Customer service agents are billed based on "percentage increase in customer satisfaction," marketing agents share revenue based on "conversion rate lift." Using low-code tools on MaaS platforms, domain experts like medical professionals, engineers, and designers can easily fine-tune highly customized "micro-models" (e.g., "a specialized model for pancreatic cancer image screening at a specific hospital") with private data. These "model micro-factories" can be traded or jointly optimized through model markets while ensuring data privacy.
4.AI Chips Will Become Very Cheap
As Moore's Law, Shannon's Law, and the von Neumann architecture encounter bottlenecks, chip and computer architectures will undergo disruptive innovation, simultaneously driving down AI chip costs. The value created by AI applications will be 100 times the computing cost, ushering in an explosion of diverse, innovative AI applications.
5.The AI-Generated Video Internet
In 2025, for current video generation scenarios, the cost of AI videos for e-commerce product displays and personalized ads is over 100 times lower than traditional methods, over 50 times lower for social media marketing videos, and over 30 times lower for brand promotion videos. It is projected that by 2030, 70% of videos on the global internet will be AI-generated, thanks to rapid advancements in multimodal large models and world models. AI videos will approach human production levels in hour-long temporal coherence and emotional expression.
6.Natural Language Will Become the New Programming Language
Current AI programming tools primarily offer code completion. By 2030, natural language will replace traditional programming syntax as the primary mode of interaction. Developers will only need to describe requirements in natural language, and an AI agent can generate complete, deployable code. Furthermore, AI will comprehend complex project contexts and business logic, independently handling the entire process of code generation, testing, and deployment. AI will autonomously fulfill 80% of development needs, not only generating code but also deeply integrating multimodal generative development with images, audio, and video. A new generation of developers—experts with industry knowledge—will replace specialized programmers, while traditional programmers will advance to become software architects and product managers. The global market size for AI programming tools will grow from approximately $6.21 billion in 2024 to nearly $100 billion by 2030, representing a 16-fold increase.
该图片可能由AI生成
Question2: What is the driving force for AI's further development and what is the bottleneck for that?
Tian Feng:
1.Computing power Bottleneck
Current GPU computing power grows ~60% annually, but memory bandwidth increases only 10% yearly. The "memory wall" leaves computing units idle 60% of the time.
The manufacturing cost of high-end AI chips rises 30% per generation, while performance gains drop from 40% at 7nm to 15% at 2nm, showing clear diminishing marginal returns. 4.Computing power：The breakthrough lies in heterogeneous architecture innovations: Chiplet technology boosts integration via stacking, while computing-in-memory architectures break the "memory wall," cutting data movement energy by over 90%. A more fundamental solution is migrating computing infrastructure to space: leveraging the cosmic 3K background for natural cooling and capturing space-based solar energy (7-10x Earth's efficiency) could reduce computing unit cost 100-fold . Musk notes that if Starship launch costs drop to $100/kg, orbital data centers become the optimal solution for AI compute expansion.
2.Energy Supply Bottleneck
Continuing current trends, AI will consume 10% of global electricity by 2030, but renewables will supply only 35%, leaving a 65% green energy gap.
Data center PUE optimization has plateaued, dropping only from 1.5 to 1.4 in 3 years, with cooling still accounting for 40% of energy use. 5.Energy Supply: Short-term solutions rely on grid-scale storage (e.g., Tesla Megapack) for "peak shaving and valley filling," doubling existing grid throughput; long-term requires a space-based solar satellite network (target: 100 GW annual deployment), utilizing 24/7 uninterrupted sunlight in outer space (30% more intense than on Earth). Musk emphasizes that China, leveraging its annual 1,500 GW solar deployment capacity (70% of its 2025 new electricity came from solar), is set to triple U.S. power generation by 2026, thereby gaining a foundational lead in AI computing power.
3.Training Data Bottleneck
Demand for high-quality training data grows 200% yearly, but the global supply of compliant, high-quality text data increases only 15% annually, creating a widening 13:1 supply-demand gap.
Synthetic data now exceeds 35% of training sets but performs 42% worse than real data on critical tasks, showing a significant "authenticity gap." 6.Training Data: Elon Musk views training AI on internet text as an inefficient "cognitive shortcut" . xAI's breakthrough lies in building a multimodal real-time data closed loop: X platform's trillions of daily posts provide human cognitive dynamics, Tesla's fleet captures physical laws via millions of video frames per second, and Neuralink explores biological raw data like neural signals . The next-generation "world model" will directly learn physics like gravity and friction, enabling AI to simulate the universe's underlying mechanics rather than infer reality from text. For instance, Grok can already analyze circuit diagram errors and recursively optimize AI chip design.
该图片可能由AI生成
Question3: In your preliminary video, you mentioned that unique enterprise data becomes a competitive advantage. What is the first practical step a company should take to start using its data effectively?
Tian Feng: The first step is to connect all your organization’s data into a single, actionable model of your business (Data). This goes beyond technical integration—it's about clearly defining core entities like customers, orders, and assets along with their relationships, effectively building a "Digital Twin" of your enterprise (Biz Logic). Then ultimately translate outcomes directly into automated workflows, alerts, and action commands (Action). This creates a second-by-second closed loop from insight to execution at every decision point across the enterprise.
Question4: As we move from digital AI to physical AI, what everyday changes do you expect people will notice first in society or daily life
Tian Feng:
1) Restructuring of Labor
Physical AI will first replace repetitive manual labor, with robotic colleagues becoming commonplace in service and manufacturing industries, allowing humans to shift to more creative and decision-intensive roles.
2) Revolution in Cost and Efficiency
From first principles, as robots scale, the marginal cost of goods and services will approach the cost of energy and raw materials, drastically reducing daily expenses in logistics, cleaning, and care.
3) The New Normal of Human-Machine Collaboration
Robots like Tesla Optimus will become "programmable partners" in homes and public spaces—from assisting the elderly to operating in hazardous environments—transforming the concept of productivity from "human labor" to "human-machine synergy".
该图片可能由AI生成
书名：《AI商业进化论：“人工智能+”赋能新质生产力发展》
出版社：人民邮电出版社
作者：田丰
帮助你定位AI当下发展坐标的指南针
帮助你洞察AI未来演进趋势的航海图
通俗化解读AI的原理、特性和四大发展规律、提供AI赋能商业、引发新质生产力变革的一手案例分析。既有宏观视角的全局观照，又有各行业应用层面的下探记录，聚焦AI的原理与实践、现在与未来，是当下AI应用的全景图、更是身处AI技术浪潮之中的探路书。