FRONTIER AI MINDS · ANNUAL BRIEFING

AI 竞争,正在从模型竞赛升级为系统竞赛

六位前沿人物过去一年的公开言论,指向能力、部署与治理的同一场变化。

EXECUTIVE ANSWER

真正稀缺的是把智能转化为结果的完整系统

模型分数只是起点;成本、能源、工作流、执行权限和责任机制共同决定现实领先。

主线02 / 37

SIX ROUTES · ONE ERA

六个人不是六句口号,而是六套因果链

科学目标、智能供给、安全治理、工业产能、世界模型与递归生产——六条路线在现实部署处相遇。

Demis科学目标
用更完整的认知系统推进科学与医学。

届时,我们将拥有一种前所未有的工具或系统,必须思考如何使用如此具有变革性的技术。在我看来,它会改变一切:经济,甚至可能包括人类的生存状态。

So around 2030 um and then when that happens you know this that that becomes a new type of tool or uh system that we've never seen before and we'll need to uh uh think about how to use such a transformative piece of technology. uh…The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:18:07
Sam Altman效率扩展
从 o1 到 GPT‑5.4,同一难题的推理成本约下降千倍。他据此判断,模型、内核、电力与数据中心仍有巨大的联合优化空间。

人们会引用各种惊人数据,说明我们的模型乃至整个行业的模型随时间变得多么高效;我认为其中一个特别惊人的例子是,我们的第一个推理模型叫 o1,大约在 16 个月前发布。

Um people cite whatever amazing statistic they like about how much more efficient our models have gotten over our industry's models have gotten over time, but one that I think is incredible, our first reasoning model was called O1, came out like 16 months ago.LIVE: Sam Altman speaks at BlackRock's US Infrastructure Su… · 00:09:57
Dario能力与风险
高速前进,同时把灾难风险降到接近零。

[笑声] 首先,谁也别再说 Claude 对我手软,或 Claude 偏袒我或 Anthropic。就像,天哪,我当时想 >> Claude,Claude 出的是硬球。

[laughter] First of all, let no one ever say that Claude is soft on me or Claude is biased towards me or towards Anthropic. Like, you know, Wow, I thought >> Claude Claude throws hard balls.The Co-Founders of Claude AI Tell Oprah About the Impact Ar… · 00:02:37
Jensen工业体系
把 Token 作为新工业体系的产出。

如果有 1 吉瓦电力,每瓦吞吐量就是收入,因为每个 Token 都有利润;计算就是收入,每瓦性能就是收入。

If you have 1 gigawatt of power, throughput per watt is revenue because every token is profitable. Compute is revenue. Performance per watt is your revenue.NVIDIA GTC Taipei 2026 Keynote · 00:36:49
Yann替代架构
通过世界预测和行动规划走向智能。

智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果,从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。

Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs…Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 21
Elon递归生产
数字与物理机器人形成生产扩张。

好吧,如果数字人体仿真问题到今年年底还没有得到解决,我会感到惊讶。我想这就是我们所说的 MacroHard 项目的意思。

Well, I'd be surprised by the end of this year if digital human emulation has not been solved. I guess that's what we sort of mean by the MacroHard project.马斯克如何把 1 太瓦 GPU 送上太空 · 01:00:39
主线03 / 37

READING CONTRACT

统计只占这一页;之后每个结论都回到原话

97 篇双语材料、805 条去重观点、43 个人物—主题正式轴。公开言论能证明“他反复强调什么”,不能直接证明技术已实现。

主线04 / 37

DEMIS HASSABIS

目标函数是科学,AGI 是实现目标的工具

科学与医学目标 → 补齐认知能力 → 世界模型与 Agent → 安全部署。

主线05 / 37

SAM ALTMAN

智能是一条不断降价并向外扩散的供给曲线

能力扩展 → 推理降本 → 基础设施扩容 → 平台分发 → 社会吸收。

Sam能力—成本曲线
企业应按未来两到四年的能力与成本曲线设计。

>> 我认为,如果启动初创公司的类似障碍和类似的工具变得更加激增,所有这些东西,你会认为像更困难的初创公司更容易启动。

>> I would assume that if the like barriers to starting startups and the like tools out there become more proliferated and all this stuff, you would assume that like harder startups are easier to start.Sam Altman - How to Start a Startup · 00:01:12
Sam Altman效率扩展
从 o1 到 GPT‑5.4,同一难题的推理成本约下降千倍。他据此判断,模型、内核、电力与数据中心仍有巨大的联合优化空间。

人们会引用各种惊人数据,说明我们的模型乃至整个行业的模型随时间变得多么高效;我认为其中一个特别惊人的例子是,我们的第一个推理模型叫 o1,大约在 16 个月前发布。

Um people cite whatever amazing statistic they like about how much more efficient our models have gotten over our industry's models have gotten over time, but one that I think is incredible, our first reasoning model was called O1, came out like 16 months ago.LIVE: Sam Altman speaks at BlackRock's US Infrastructure Su… · 00:09:57
Sam平台边界
模型、ChatGPT、API 与 Codex 构成通用平台。

我们的战略是构建核心模型和基本使用方式——ChatGPT、API、Codex 等智能体以及通用智能体——并真正赋能其他所有人进行建设。我从 OpenAI 学到的一件事是,世界上大多数创新都来自初创企业,而我们希望赋能它们。

Um we way more innovation is going to happen outside of OpenAI than inside of OpenAI and our strategy is to build the core model and the basic way to use it, ChatGPT, the API, agents like Codex and the general agent, and really empower everybody…Fireside Chat: Sam Altman×Vinod Khosla and AMA at IIT Delhi… · 00:33:05
主线06 / 37

DARIO AMODEI

安全不是减速口号,而是能力增长假设的后果

能力快速增长 → 科学与企业价值放大 → 就业、滥用与失控风险同步放大 → 测试与责任前置。

Dario组织起点
Scaling 会让机会与风险同时扩大。

是的,回到 Anthropic 的创立:我和 Anthropic 的联合创始人是最早记录技术规模扩展及其将变得多么强大的人之一。我们创办公司背后的想法是,看着技术进步如此迅速,我们说:“看,我们确实需要把这件事做好。

Yes, so, you know, go going back to the founding of Anthropic, you know, we were among the first myself and my co-founders at Anthropic to to document the scaling of the technology and how powerful it would become. And you know, the idea behind how…「不毛な競争には加わらない」アンソロピック ダリオ・アモデイCEO 単独インタビュー「安全なAI」と日本市場への期待【… · 00:01:32
Dario高速列车
技术列车不能由一家停止,也不能任其失控。

嗯,真正回答这个问题的话,这项技术是人类一个划时代的变化。你知道,某个时候人类开始使用火;某个时候,人类造出了蒸汽机和工厂,由此创造了工业革命。

You know, at some point humans started to, you know, started to use fire. At some point humans, you know, built built steam engines and factories, and that created the Industrial Revolution.The Co-Founders of Claude AI Tell Oprah About the Impact Ar… · 00:02:37
Dario分配风险
高增长可能与高失业、低工资和不平等并存。

我们或许会看到一个极不寻常的组合:GDP 高速增长,同时伴随高失业、至少是就业不足,或大量低工资岗位。

I think we could have this very unusual combination of very fast GDP growth and high unemployment, or at least underemployment or, you know, low wage job.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:27:41
主线07 / 37

JENSEN HUANG

AI 不是软件功能,而是一种新工业产能

扩展律 → AI 工厂 → Token 产出 → Agent 与物理 AI → 行业生产率。

主线08 / 37

YANN LECUN

对 LLM 的批评只是起点,JEPA 才是替代路线

连续现实信号 → 抽象表示 → 世界预测 → 行动规划 → 机器人与工业控制。

Yann总纲
智能必须预测世界演化和行动后果。

智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果,从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。

Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs…Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 21
YannJEPA
在表示空间保留对行动后果有用的信息。

JEPA 的基本思想是:我们观察的是一种复杂现象,不是离散 token 序列,不是语言,而是来自工业系统的传感器数据、视频,或连续、高维且含噪的信号。你不能用生成式 AI 去理解它。

The basic idea of JEPA is that we have a complex phenomenon that we observe, not a sequence of discrete tokens, not language, but sensor data from industrial systems, videos, or continuous high-dimensional noisy signals. You cannot use generative AI to understand it.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 41
Yann关键区别
生成合理后续并不等于理解真实后续。

当然,你会说:“等等,但我们有能进行视频生成的系统。

Now, of course, you're going to say, "But wait, we have systems that can do video generation.Fireside Chat with Yann LeCun, Executive Chairman of AMI La… · 00:06:10
主线09 / 37

ELON MUSK

机器人是经济叙事中心,能源是扩张硬约束

求真模型 → 数字工作自动化 → 物理机器人 → 递归生产 → 能源与太空算力。

主线10 / 37

CAPABILITY ROUTES

规模仍有效,但“继续堆”不是完整解释

能力继续增长的证据很强;长期记忆、世界理解与可靠行动仍没有被规模单独解释。

Jensen Huang三条扩展律
扩展已不只是预训练:后训练与推理时计算同样决定能力;系统“思考”越久,检索与反复推演越充分。

为此必须进行大量推理,因此在强化学习中,训练与推理已经融为一体。>> 这非常复杂,也就是所谓的后训练。

And so it tries a whole bunch of different ways and and uh in order to do that, yeah, you've got to do inference. So now training and inference are now integrated in reinforcement learning.NVIDIA: OpenAI, Future of Compute, and the American Dream |… · 00:01:52
Dario Amodei能力墙并非本质边界
推理、发现与持续学习曾被视为“能力墙”,却不断被新训练方法和更长上下文突破;当前缺陷不应被视为不可跨越。

过去是“AI 模型不能推理”,最近则是“AI 模型不能做出新发现”。

It was like, "AI models can't reason." And recently, there's this, "AI models can't make new discoveries.A Cheeky Pint with Anthropic CEO Dario Amodei · 00:33:27
Demis突破条件
AGI 仍可能需要一到两个重大突破。

首先,我对 AGI 的定义是:一个经测试后能表现出人类所具备全部认知能力的系统。之所以以此为标准,是因为人类及人脑是我们目前所知——也许是整个宇宙中——通用智能存在的唯一实例。

I mean firstly my definition of AGI is is is you know a system that you can test and exhibits all the cognitive capabilities humans have. The reason that's important is um the human reference point and the human brain reference point is the human brain…The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:16:50
Yann架构缺口
语言序列预测缺少世界模型与行动后果预测。

智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果,从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。

Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs…Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 21
主线11 / 37

WORLD MODELS

“世界模型”其实是三种不同的赌注

Demis 讲认知机制,Yann 讲学习目标,Jensen 讲工业仿真;同名不等于共识。

主线12 / 37

CAPABILITY GAP

真正未解决的是长期可靠性

榜单、单轮问答与演示不能替代长期任务完成率、失败可逆性、环境变化下稳定性和监督成本。

主线13 / 37

FOUR QUESTIONS

“模型很强”必须拆成四个可验证问题

能不能做?能不能持续做?能不能负担?能不能负责地做?

主线14 / 37

DEPLOYMENT · 1

真正的部署单位是有责任人的工作流

模型价值不会自然释放;目标、数据、工具、权限和责任人必须进入同一个组织闭环。

Demis Hassabis研究—产品—分发
产品战略从最强基础模型和研究开始,再快速进入产品、营销和分发;Google 正重构基础设施,把最新模型接入核心产品。

拥有最好的模型,由基础研究供给;随后尽快反映到产品,做好营销和分发。但模型若非同类最佳、最先进,其他都不重要。

having the best models in this case, and obviously fundamental research feeding into that, and I always believe you then need to reflect that obviously as quickly as possible in your products, and then you've got to get your marketing and distribution right. But none of…Can Google’s New AI Solve Everything? | Titans and Disruptors · 00:19:37
Dario Amodei企业是集体智能
更聪明的模型要进入企业并推进疾病治疗;企业本身是超越个人的集体智能,人机组织的上限可能高于单个模型。

……希望公司最终抵达的位置。

where we hope the the company will be.A Conversation with Dario Amodei and Marc Benioff | Dreamfo… · 00:31:19
Jensen Huang企业 Agent 控制栈
当 Agent 能调用子代理、工具和记录系统,企业平台必须提供身份、权限、隐私路由、网络护栏与完整审计。

【Peter】我很喜欢自己的工具赋能普通人;一位 60 岁父亲安装它后,甚至把酿酒机器通过蓝牙接入 OpenClaw。

And he had like one guy, he told me like he installed it as a 60 year old dad and like they made beer, connected the machine via Bluetooth to open cloth.NVIDIA GTC Keynote 2026 · 01:49:15
主线15 / 37

DEPLOYMENT · 2

从总算力转向单位有用任务的全成本

推理、能源、网络、人工复核、失败重做与基础设施利用率都要计入。

主线16 / 37

DEPLOYMENT · 3

Agent 的门槛不是会回答,而是能被授权

成熟度要看任务完成率,也要看授权范围、错误半径、回滚能力、监督时间与总成本。

Demis Hassabis委托完整任务
模型能力已接近可委托完整任务的门槛;一旦从建议转向代办,偏离、暂停和恢复就成为产品能力的一部分。

展望今年,AI 故事和成就会是什么?我每年都说 AI 每年都关键;对前沿科学工作者而言,几乎每年都像十年,今​​年也不例外,强度很高。

So when you look at the year ahead, what do you think the story of AI will be? What will we achieve?Can Google’s New AI Solve Everything? | Titans and Disruptors · 00:25:15
Dario Amodei人类转向编辑与监督
部分团队约九成代码由 AI 生成,但人仍负责编辑、监督和最困难部分;自动化比例越高,监督质量越关键。

谈论未来 AI 时,整件事难免带上一层科幻光泽。它作为比喻有帮助,却也可能让人分心,因为它会把所有人的思维带入一种未来幻想模式。

There there's I think inevitably as we, you know, talk about AI in the future, this kind of sci-fi sheen on on on, you know, the on the whole thing. And that's, you know, in a way it's helpful as a metaphor, but also it can…A Conversation with Dario Amodei and Marc Benioff | Dreamfo… · 00:16:03
Yann LeCun配方不等于理解
当前 Agent 多是在盲目套用昂贵、难泛化的步骤配方;只有世界模型进入系统,跨任务泛化才可能发生质变。

不,我认为几年内它会变得非常不同,主要是因为当前智能体系统缺乏世界模型。构建一个无法预测自身行动后果的智能体系统,是非常糟糕的主意。

No, I think it’s going to look very different within a few years. Mostly because of the lack of world models in current agentic systems.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 73
主线17 / 37

PHYSICAL AI

机器人是所有技术路线的现实压力测试

世界理解、长期规划、数据、硬件、可靠性、安全和单位成本被同时拉到一张考卷上。

主线18 / 37

ENERGY & SPACE

能源是现实约束;太空算力仍是高承诺预测

地面电力、许可、网络与建设周期已进入核心约束;轨道太阳能仍需验证发射、散热、通信与运维。

主线19 / 37

AI FOR SCIENCE

科学价值形成跨人物共振,但不是同一路径

至少四位人物独立强调科学价值;他们仍分歧于技术路线、时间表、验证标准与责任链。

Demis目标函数
科学、医学与治病是 AGI 的首要价值。

一两项重大突破,就像 Transformer,或我们在 AlphaGo 中把强化学习用于深度学习那样。因此,答案尚未揭晓。

or two big breakthroughs still uh like transformers were or things like reinforcement learning applied to deep learning like we did with Alph Go. So the jury's out on that.The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:18:07
Sam早期信号
前沿模型开始出现微小但真实的科学突破。

不过,GPT-5 首次取得了非常微小但确实真实的科学突破,这令人十分振奋;我们一路走来一直在等待的就是这一刻。

But the fact that for the first time GPT5 is making very small but but real scientific breakthroughs is quite exciting and the thing we've been waiting for this whole time.Sam Altman, OpenAI in conversation with Peng Xiao, G42 at G… · 00:20:53
Dario医疗前沿
生物学与医学展现意外能力。

我认为我见过的一些最疯狂的东西是围绕生物学和医学的。我见过很多案例,实际上包括丹妮拉,Claude诊断出一个医疗问题,而你知道,很多优秀的医生都忽略了这一问题。

I think some of the wildest stuff I've seen is around biology and medicine. I've seen a number of cases, including Daniela actually, where Claude diagnosed a medical problem that, you know, a bunch of fancy doctors had missed.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:24:49
Yann LeCun领域自监督模型
AlphaFold、ESM 的成功来自专用自监督模型而非自回归 LLM;材料和生物医学仍需要用无标签数据快速适配领域。

人们寄予很大希望,但对 AI 在材料科学、生物医学等领域应用的投资可能还不够。

And there’s a lot of hope and probably not enough investment in the use of AI for things like material science and biomedicine.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 80
主线20 / 37

WORK

工作的分歧来自时间尺度与分析单位

应问自动化的是任务、岗位还是组织,发生在何时,生产率收益由谁获得。

Jensen Huang任务自动化扩大需求
自动化的是任务而非工作的目的;服务供给与下游需求扩大后,放射科、法律辅助和制造业可能净增岗位。

市场、行业和社会需要认识到:当我们自动化这些任务时,世界所需的工作岗位实际上在增加。这对我来说很合理。

And it's very sensible to me. You know, while people have said that AI is going to eliminate jobs, that is going to make it so that we don't have to work at all, it's exactly the opposite.Jensen Huang says the AI doomers have it wrong · 00:16:04
Sam组织监督
人的判断与监督责任仍在。

另一个判断标准是:一家大公司的 CEO、一个大国的总统或一位诺贝尔奖得主,什么时候会离不开对 AI 的大量使用,无法正常履行职责?这并不意味着会出现 AI CEO 或 AI 总统,但确实意味着,比如人类 CEO 的角色会发生变化。

The other one is when can a CEO of major company, a president of a major country, a Nobel Prize winning scientist, when can they not do their job without making heavy use of AI? This doesn't mean that there will be an AI CEO or…LIVE: Sam Altman speaks at BlackRock's US Infrastructure Su… · 00:00:00
Dario分配冲击
高增长与高失业可以同时成立。

我们或许会看到一个极不寻常的组合:GDP 高速增长,同时伴随高失业、至少是就业不足,或大量低工资岗位。

I think we could have this very unusual combination of very fast GDP growth and high unemployment, or at least underemployment or, you know, low wage job.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:27:41
Elon长期情景
数字工作先受冲击,长期工作可能变为自愿。

>> 我不希望你惊吓大家,但 AI 多快、多大程度取代工作确实是重要辩论。正如你说,十年后人工智能会超过所有人类智能。

That is natural. >> So I I certainly don't want you to alarm people, but I do think there is a really important debate about how quickly and to what degree AI displace people from jobs because it is as you say if in 10 years…The full-length interview with Elon Musk | The Economist · 00:28:27
主线21 / 37

ABUNDANCE

丰裕与过渡成本必须在同一逻辑链里

供给扩张不会自动生成公平结果;所有权、收入分配、公共服务与人的能动性不能被愿景略过。

Demis丰裕来源
科学、材料、能源与医疗推动长期丰裕。

AI 将解决许多科学大问题,如新材料、聚变、太阳能、最优电池和能源危机;随后进入激进富足世界,可用这些能源——

And then I think we'll be in a in a world of radical abundance, where we can use those energy sources to,Can Google’s New AI Solve Everything? | Titans and Disruptors · 00:28:23
Sam普惠使命
智能应充裕、廉价、强大且广泛可用。

说我喜欢的一句话,那就是如果你的生活中有一个遥控器上的快进按钮,你的生活就结束了。所以就像无聊的部分、糟糕的部分一样,这比没有经验要好得多。

say this thing that I loved which is if you had like a fast forward button on a remote for your life your life would be over. And so like the boring parts, the bad parts, it's like much better than no experience.Sam Altman - How to Start a Startup · 00:07:43
Elon物质供给
机器人递归生产是物质扩张机制。

好吧,如果数字人体仿真问题到今年年底还没有得到解决,我会感到惊讶。我想这就是我们所说的 MacroHard 项目的意思。

Well, I'd be surprised by the end of this year if digital human emulation has not been solved. I guess that's what we sort of mean by the MacroHard project.马斯克如何把 1 太瓦 GPU 送上太空 · 01:00:39
Dario过渡成本
GDP、工资、就业与财富分配可能分离。

我们或许会看到一个极不寻常的组合:GDP 高速增长,同时伴随高失业、至少是就业不足,或大量低工资岗位。

I think we could have this very unusual combination of very fast GDP growth and high unemployment, or at least underemployment or, you know, low wage job.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:27:41
主线22 / 37

RISK MAP

不要把五类问题都叫作“AI 安全”

模型行为、恶意使用、自主失控、社会分配、地缘供应链——风险对象不同,控制节点也不同。

Demis Hassabis滥用与自主偏离
风险一端是个人或国家恶意使用双用途能力,另一端是自主 Agent 偏离预期;技术护栏必须与国际最低标准并行。

人们经常谈论 AI 的安全与安保风险。

So, uh so one of the things that people always talk about is the safety risk, security risk from AI, and so on and so forth, right?Google's Demis Hassabis Speaks at India AI Impact Summit 20… · 00:13:10
Sam Altman未知的社会影响
过度迎合和强化妄想,是未被传统风险清单捕获的真实失败;每天数十亿次对话意味着人格微调也可能产生系统影响。

我认为到目前为止我们在 ChatGPT 做过的最糟糕的事情是呃,我们遇到了这个疾病问题,该模型对用户来说有点太讨人喜欢了,对于某些用户来说,大多数用户只是烦人,但对于一些精神状态脆弱的用户来说,这鼓励了妄想,这并不是我们担心的最大风险。

I think the worst thing we've done in ChachiBT so far is uh we had this issue with sickency where the model was kind of being too flattering to users and for some users it was most users it was just annoying but for some users…Sam Altman Shows Me GPT 5... And What's Next · 00:52:12
Yann LeCun防止知识集中
限制公众获取知识可能造成更危险的权力集中;在他的排序中,开放滥用风险小于少数公司和政府控制信息的风险。

有一种“安全主义”的论述,大意是 AI 很危险,正因为危险,所以不能交到所有人手中。这类论述在其他领域存在已久。

Of the type, AI is dangerous, and because it’s dangerous, it can’t be put in the hands of everyone. This discourse has been around for a long time in other domains.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 30
主线23 / 37

GOVERNANCE CHAIN

评测、部署控制和事故责任必须连成一条链

只做模型评测不够;训练、发布、授权、持续监测、事故响应与责任追踪要连续。

Dario Amodei治理门槛
计算量超过某一门槛的模型,应当由具备资质的第三方强制测试。

计算量超过某一门槛的模型,应当由具备资质的第三方强制测试其在四个特定领域的风险水平:网络安全、生物武器、AI 系统失控,以及可能加速上述其他风险的自动化研发。

Models above a threshold of compute should undergo mandatory testing by a qualified third party for their level of risk in four specific areas: cybersecurity, biological weapons, loss of control of AI systems, and automated R&D that could accelerate these other risks.Policy on the AI Exponential · paragraph 15
Elon Musk发布前同业测试
突破性模型发布前,应给予竞争实验室短期 API 测试窗口;利用竞争激励发现问题,再逐步形成更正式机制。

发布突破性的全新前沿模型——远比现有模型聪明——之前,应有机会让——

we I talked to him for a few hours before he published his his piece was that um the most immediate thing that we could do is to have the leading AI companies um at least just meet uh or have a some sort of call…The full-length interview with Elon Musk | The Economist · 00:12:43
Jensen Huang企业 Agent 控制栈
当 Agent 能调用子代理、工具和记录系统,企业平台必须提供身份、权限、隐私路由、网络护栏与完整审计。

【Peter】我很喜欢自己的工具赋能普通人;一位 60 岁父亲安装它后,甚至把酿酒机器通过蓝牙接入 OpenClaw。

And he had like one guy, he told me like he installed it as a 60 year old dad and like they made beer, connected the machine via Bluetooth to open cloth.NVIDIA GTC Keynote 2026 · 01:49:15
主线24 / 37

OPEN MODELS

开放模型争论是风险与权力的双重分配

开放既非天然安全,也非天然危险;能力阈值、不可撤回性、审计与参与权必须分别判断。

主线25 / 37

CHIPS & SECURITY

芯片不是外部背景,而是治理链的一部分

训练、制程、数据中心、电力、蒸馏与部署权限都可能成为控制节点;每个节点的误伤与规避成本不同。

主线26 / 37

ERA SIGNALS

反复出现的五个判断,构成时代底色

这里的“反复出现”只表示至少三位人物在独立场合方向相近:能力扩展、可靠性缺口、基础设施、Agent 权限、科学价值与治理链同步上升。

Demis可靠性
长期认知结构仍不完整。

你知道,我的联合创始人 Shane 等人曾公开说过大约是 2028 年、2030 年。我们把它视为一项 20 年使命,我认为我们正按此推进。

You know, people like Shane, my co-founder, was on record saying around 2028, 2030 you know, we thought of it as a 20 year mission, and I think we're on track for that. What do you think the capabilities that are missing are still before we…Demis Hassabis on AGI by 2030, Curing Every Disease, Life A… · 00:03:21
Jensen基础设施
能力必须落在完整生产系统。

世界正在竞相建设 AI 工厂;芯片、机架、网络、电源、冷却与电网必须端到端协同设计,因为计算就是收入。

The world is racing to build AI factories. Every layer—chip, rack, network, power, cooling and grid—must be designed together end to end because compute is revenue.NVIDIA GTC Taipei 2026 Keynote · 00:26:52
Jensen Huang企业 Agent 控制栈
当 Agent 能调用子代理、工具和记录系统,企业平台必须提供身份、权限、隐私路由、网络护栏与完整审计。

【Peter】我很喜欢自己的工具赋能普通人;一位 60 岁父亲安装它后,甚至把酿酒机器通过蓝牙接入 OpenClaw。

And he had like one guy, he told me like he installed it as a 60 year old dad and like they made beer, connected the machine via Bluetooth to open cloth.NVIDIA GTC Keynote 2026 · 01:49:15
Demis科学
科学与医疗构成最强正向价值。

一两项重大突破,就像 Transformer,或我们在 AlphaGo 中把强化学习用于深度学习那样。因此,答案尚未揭晓。

or two big breakthroughs still uh like transformers were or things like reinforcement learning applied to deep learning like we did with Alph Go. So the jury's out on that.The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:18:07
主线27 / 37

FIVE DIVERGENCES

真正决定未来路线的是这些分歧

扩展还是新架构?LLM 主体还是接口?世界模型解决哪类问题?工作增强还是替代?治理卡在哪个节点?

Jensen Huang三条扩展律
扩展已不只是预训练:后训练与推理时计算同样决定能力;系统“思考”越久,检索与反复推演越充分。

为此必须进行大量推理,因此在强化学习中,训练与推理已经融为一体。>> 这非常复杂,也就是所谓的后训练。

And so it tries a whole bunch of different ways and and uh in order to do that, yeah, you've got to do inference. So now training and inference are now integrated in reinforcement learning.NVIDIA: OpenAI, Future of Compute, and the American Dream |… · 00:01:52
Yann替代架构
语言预测并非理解世界的主体。

智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果,从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。

Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs…Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 21
Dario替代风险
全面自动化可能削弱比较优势缓冲。

但这会非常微妙,不可能一刀切,只能逐步摸索。我们或许会看到一个极不寻常的组合:GDP 高速增长,同时伴随高失业、至少是就业不足,或大量低工资岗位。

It won't be all one thing. It won't, we'll have to kind of figure it out over time.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:27:41
Jensen Huang任务自动化扩大需求
自动化的是任务而非工作的目的;服务供给与下游需求扩大后,放射科、法律辅助和制造业可能净增岗位。

市场、行业和社会需要认识到:当我们自动化这些任务时,世界所需的工作岗位实际上在增加。这对我来说很合理。

And it's very sensible to me. You know, while people have said that AI is going to eliminate jobs, that is going to make it so that we don't have to work at all, it's exactly the opposite.Jensen Huang says the AI doomers have it wrong · 00:16:04
主线28 / 37

UNRESOLVED QUESTIONS

每条路线都缺最后一段现实证明

世界模型能否补齐认知?降本会普惠还是集中?治理能否跟上增长?AI 工厂何时稳定回报?JEPA 能否规模胜出?Optimus 与轨道算力能否跨过工程门槛?

Demis仍需突破
认知缺口能否被世界模型补齐。

首先,我对 AGI 的定义是:一个经测试后能表现出人类所具备全部认知能力的系统。之所以以此为标准,是因为人类及人脑是我们目前所知——也许是整个宇宙中——通用智能存在的唯一实例。

I mean firstly my definition of AGI is is is you know a system that you can test and exhibits all the cognitive capabilities humans have. The reason that's important is um the human reference point and the human brain reference point is the human brain…The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:16:50
Sam Altman效率扩展
从 o1 到 GPT‑5.4,同一难题的推理成本约下降千倍。他据此判断,模型、内核、电力与数据中心仍有巨大的联合优化空间。

人们会引用各种惊人数据,说明我们的模型乃至整个行业的模型随时间变得多么高效;我认为其中一个特别惊人的例子是,我们的第一个推理模型叫 o1,大约在 16 个月前发布。

Um people cite whatever amazing statistic they like about how much more efficient our models have gotten over our industry's models have gotten over time, but one that I think is incredible, our first reasoning model was called O1, came out like 16 months ago.LIVE: Sam Altman speaks at BlackRock's US Infrastructure Su… · 00:09:57
Jensen回报验证
Token 产出何时覆盖资本与能源投入。

如果有 1 吉瓦电力,每瓦吞吐量就是收入,因为每个 Token 都有利润;计算就是收入,每瓦性能就是收入。

If you have 1 gigawatt of power, throughput per watt is revenue because every token is profitable. Compute is revenue. Performance per watt is your revenue.NVIDIA GTC Taipei 2026 Keynote · 00:36:49
Elon预测验证
轨道算力的工程与时间表能否成立。

这就是您通过将其移入太空而可能节省的部分。其中大部分是 GPU。

That's the part you're presumably saving by moving this into space. Most of it's the GPUs.马斯克如何把 1 太瓦 GPU 送上太空 · 00:00:30
主线29 / 37

CLOSING

六条路线,共同描述一个尚未定型的 AI 时代

能力正在扩展,现实约束正在增多;价值不在替他们选出赢家,而在让分歧连同原话、条件和证据强弱同时可见。

主线30 / 37

APPENDIX · METHOD

证据状态决定语言强度

稳定核心、复现敏感核心、辨识度重复立场、高承诺预测、支撑材料、暂不入选——六种状态不能混写。

证据附录31 / 37

APPENDIX · PRIMARY EVIDENCE

DEMIS · EVIDENCE MAP

科学目标—认知机制—安全部署

Demis稳定核心
科学、医学与治病。

一两项重大突破,就像 Transformer,或我们在 AlphaGo 中把强化学习用于深度学习那样。因此,答案尚未揭晓。

or two big breakthroughs still uh like transformers were or things like reinforcement learning applied to deep learning like we did with Alph Go. So the jury's out on that.The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:18:07
Demis成果路径
AlphaFold 打开生物学根节点。

我们能够培育独角兽,难点是如何迈向下一个阶段。——正是下一阶段。

We can get the unicorns. It's getting it to the next level >> the next level now.The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:42:38
Demis能力缺口
持续学习、记忆与长期规划。

你知道,我的联合创始人 Shane 等人曾公开说过大约是 2028 年、2030 年。我们把它视为一项 20 年使命,我认为我们正按此推进。

You know, people like Shane, my co-founder, was on record saying around 2028, 2030 you know, we thought of it as a 20 year mission, and I think we're on track for that. What do you think the capabilities that are missing are still before we…Demis Hassabis on AGI by 2030, Curing Every Disease, Life A… · 00:03:21
Demis Hassabis能力机制
我们想构建所谓的世界模型,即一种真正理解世界物理规律的模型。We want to build what we call a world model, which is a model that actually understands the physics of the world.Demis Hassabis on shipping momentum, better evals and world… · 00:00:14
Demis治理
Agent 安全与国际标准。

对。所以是品味和判断力,以及继续学习所需的技能。

Right. So taste and judgment and continuing to learn the skills, the skills needed.Demis Hassabis on AGI by 2030, Curing Every Disease, Life A… · 00:15:04
Demis条件预测
2030 前后仍可能需要重大突破。

我知道你说的是 2030 年及以后,但希望你再展开一些。>> 好。

I know you say 2030 and onwards but a little bit of color on that would be very good. >> Yeah.The Future of AI: With Sir Demis Hassabis & Dame Wendy Hall · 00:16:50
证据附录32 / 37

APPENDIX · PRIMARY EVIDENCE

SAM · EVIDENCE MAP

能力扩展—成本下降—平台分发

Sam稳定核心
按能力—成本曲线设计公司。

>> 我认为,如果启动初创公司的类似障碍和类似的工具变得更加激增,所有这些东西,你会认为像更困难的初创公司更容易启动。

>> I would assume that if the like barriers to starting startups and the like tools out there become more proliferated and all this stuff, you would assume that like harder startups are easier to start.Sam Altman - How to Start a Startup · 00:01:12
Sam Altman效率扩展
从 o1 到 GPT‑5.4,同一难题的推理成本约下降千倍。他据此判断,模型、内核、电力与数据中心仍有巨大的联合优化空间。

人们会引用各种惊人数据,说明我们的模型乃至整个行业的模型随时间变得多么高效;我认为其中一个特别惊人的例子是,我们的第一个推理模型叫 o1,大约在 16 个月前发布。

Um people cite whatever amazing statistic they like about how much more efficient our models have gotten over our industry's models have gotten over time, but one that I think is incredible, our first reasoning model was called O1, came out like 16 months ago.LIVE: Sam Altman speaks at BlackRock's US Infrastructure Su… · 00:09:57
Sam瓶颈
晶体管与电力约束供给。

>> 晶体管,其次是电力——按这个顺序。

>> Transistors and then electrons in that order.Sam Altman - How to Start a Startup · 00:10:56
Sam资源配置
Codex 的战略承诺。

>> [笑声] >> 从某种意义上说,Codex 就是一个例子。它不是最近的一次,但很有启发性。

>> [laughter] >> I mean in some sense codeex was an example of this like this is not the most recent one but it was one that I think it's instructive. Um, we were way behind cloud code and it seemed like a kind of crazy…Sam Altman - How to Start a Startup · 00:52:02
Sam平台边界
通用平台与外部创新。

未来发生在 OpenAI 外部的创新会远多于内部。我们的战略是构建核心模型和基本使用方式——ChatGPT、API、Codex 等智能体以及通用智能体——并真正赋能其他所有人进行建设。

Um we way more innovation is going to happen outside of OpenAI than inside of OpenAI and our strategy is to build the core model and the basic way to use it, ChatGPT, the API, agents like Codex and the general agent, and really empower everybody…Fireside Chat: Sam Altman×Vinod Khosla and AMA at IIT Delhi… · 00:33:05
Sam社会判断
指数变化可能呈平滑体验。

年。

years.Sam Altman on AGI, Compute, and Human Agency · 00:26:33
证据附录33 / 37

APPENDIX · PRIMARY EVIDENCE

DARIO · EVIDENCE MAP

能力增长—收益与冲击—强制治理

Dario组织起点
Scaling 放大机会与风险。

是的,回到 Anthropic 的创立:我和 Anthropic 的联合创始人是最早记录技术规模扩展及其将变得多么强大的人之一。我们创办公司背后的想法是,看着技术进步如此迅速,我们说:“看,我们确实需要把这件事做好。

Yes, so, you know, go going back to the founding of Anthropic, you know, we were among the first myself and my co-founders at Anthropic to to document the scaling of the technology and how powerful it would become. And you know, the idea behind how…「不毛な競争には加わらない」アンソロピック ダリオ・アモデイCEO 単独インタビュー「安全なAI」と日本市場への期待【… · 00:01:32
Dario正向价值
生物与医疗能力前沿。

我认为我见过的一些最疯狂的东西是围绕生物学和医学的。我见过很多案例,实际上包括丹妮拉,Claude诊断出一个医疗问题,而你知道,很多优秀的医生都忽略了这一问题。

I think some of the wildest stuff I've seen is around biology and medicine. I've seen a number of cases, including Daniela actually, where Claude diagnosed a medical problem that, you know, a bunch of fancy doctors had missed.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:24:49
Dario企业模式
长期信任与安全部署。

您在编码和企业上下了赌注,Claude代码很受欢迎。Claude·科沃克很受欢迎。

You made a bet on coding and enterprise and Claude Code is a hit. Claude Cowork is a hit.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:10:45
Dario分配风险
增长与失业可能并存。

我们或许会看到一个极不寻常的组合:GDP 高速增长,同时伴随高失业、至少是就业不足,或大量低工资岗位。

I think we could have this very unusual combination of very fast GDP growth and high unemployment, or at least underemployment or, you know, low wage job.Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:27:41
Dario Amodei治理门槛
计算量超过某一门槛的模型,应当由具备资质的第三方强制测试。

计算量超过某一门槛的模型,应当由具备资质的第三方强制测试其在四个特定领域的风险水平:网络安全、生物武器、AI 系统失控,以及可能加速上述其他风险的自动化研发。

Models above a threshold of compute should undergo mandatory testing by a qualified third party for their level of risk in four specific areas: cybersecurity, biological weapons, loss of control of AI systems, and automated R&D that could accelerate these other risks.Policy on the AI Exponential · paragraph 15
Dario国家安全
先进芯片出口控制。

例如,我一直直言不讳地支持限制向中国出口先进芯片。

There have been a number of cases where we've really spoken our minds about what we think, you know, I've been very outspoken about the need for export controls on ships to China, right?Inside the Mind of Anthropic CEO Dario Amodei | The Circuit… · 00:17:18
证据附录34 / 37

APPENDIX · PRIMARY EVIDENCE

JENSEN · EVIDENCE MAP

扩展律—AI 工厂—Agent 与物理 AI

Jensen Huang三条扩展律
扩展已不只是预训练:后训练与推理时计算同样决定能力;系统“思考”越久,检索与反复推演越充分。

为此必须进行大量推理,因此在强化学习中,训练与推理已经融为一体。>> 这非常复杂,也就是所谓的后训练。

And so it tries a whole bunch of different ways and and uh in order to do that, yeah, you've got to do inference. So now training and inference are now integrated in reinforcement learning.NVIDIA: OpenAI, Future of Compute, and the American Dream |… · 00:01:52
Jensen生产系统
完整 AI 工厂。

世界正在竞相建设 AI 工厂;芯片、机架、网络、电源、冷却与电网必须端到端协同设计,因为计算就是收入。

The world is racing to build AI factories. Every layer—chip, rack, network, power, cooling and grid—must be designed together end to end because compute is revenue.NVIDIA GTC Taipei 2026 Keynote · 00:26:52
Jensen经济性
固定电力的 Token 产出。

如果有 1 吉瓦电力,每瓦吞吐量就是收入,因为每个 Token 都有利润;计算就是收入,每瓦性能就是收入。

If you have 1 gigawatt of power, throughput per watt is revenue because every token is profitable. Compute is revenue. Performance per watt is your revenue.NVIDIA GTC Taipei 2026 Keynote · 00:36:49
Jensen产业拐点
推理需求而非只看训练。

这是重大时刻。O1 模型本身只稍大,但用于上下文的输入 token 和思考的输出 token 让计算量激增。

very big moment. The amount of input tokens that was necessary in order to produce, and the amount of output tokens it generated in order to reason the model was a little bit larger.NVIDIA GTC Keynote 2026 · 00:48:09
Jensen Huang企业 Agent 控制栈
当 Agent 能调用子代理、工具和记录系统,企业平台必须提供身份、权限、隐私路由、网络护栏与完整审计。

【Peter】我很喜欢自己的工具赋能普通人;一位 60 岁父亲安装它后,甚至把酿酒机器通过蓝牙接入 OpenClaw。

And he had like one guy, he told me like he installed it as a 60 year old dad and like they made beer, connected the machine via Bluetooth to open cloth.NVIDIA GTC Keynote 2026 · 01:49:15
Jensen物理 AI
仿真与合成数据。

然而,为了为人工智能、机器人技术创建数据,它必须处于机器人的感知、视角中。世界上大多数视频数据都来自第三人称,而不是第一人称。

However, in order to create a data for AI, robotics, it has to be in the perception, the perspective of the robot. And most of the world's video data is from a third person, not first person.NVIDIA GTC Taipei 2026 Keynote | Full Replay · 01:38:42
证据附录35 / 37

APPENDIX · PRIMARY EVIDENCE

YANN · EVIDENCE MAP

表示学习—世界预测—行动规划

Yann总纲
预测世界与行动后果。

智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果,从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。

Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs…Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 21
YannJEPA
表示空间预测。

原因在于,其中大部分数据不可预测:既有噪声,也有高度复杂的依赖关系。如果你尝试训练一个系统预测视频接下来会发生什么,它不可能预测每一个细节。

And the reason is that most of that data is unpredictable. There’s noise, and there are highly complex dependencies in it.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 41
Yann关键区别
生成不等于理解。

当然,你会说:“等等,但我们有能进行视频生成的系统。

Now, of course, you're going to say, "But wait, we have systems that can do video generation.Fireside Chat with Yann LeCun, Executive Chairman of AMI La… · 00:06:10
Yann近期落点
工业预测维护与控制。

我们正在与多家潜在产业合作伙伴洽谈,涉及提前预测故障或异常之类的事情。

And they have to do with things like predicting failures or faults in advance.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 47
Yann时间边界
现实世界 AI 分层时间表。

面向现实世界的 AI。

AI for the real world.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 15
Yann LeCun防止知识集中
限制公众获取知识可能造成更危险的权力集中;在他的排序中,开放滥用风险小于少数公司和政府控制信息的风险。

有一种“安全主义”的论述,大意是 AI 很危险,正因为危险,所以不能交到所有人手中。这类论述在其他领域存在已久。

Of the type, AI is dangerous, and because it’s dangerous, it can’t be put in the hands of everyone. This discourse has been around for a long time in other domains.Meet Yann LeCun's Lab and the AI World of 2030 · paragraph 30
证据附录36 / 37

APPENDIX · PRIMARY EVIDENCE

ELON · EVIDENCE MAP

求真模型—机器人—能源与太空

Elon模型使命
理解宇宙与严格求真。

理解宇宙需要哪些东西?你必须好奇,你必须存在。

What things are necessary to understand the universe? You have to be curious and you have to exist.马斯克如何把 1 太瓦 GPU 送上太空 · 00:39:39
Elon工作情景
数字工作先受冲击。

>> 我不希望你惊吓大家,但 AI 多快、多大程度取代工作确实是重要辩论。正如你说,十年后人工智能会超过所有人类智能。

That is natural. >> So I I certainly don't want you to alarm people, but I do think there is a really important debate about how quickly and to what degree AI displace people from jobs because it is as you say if in 10 years…The full-length interview with Elon Musk | The Economist · 00:28:27
Elon递归生产
机器人制造机器人。

好吧,如果数字人体仿真问题到今年年底还没有得到解决,我会感到惊讶。我想这就是我们所说的 MacroHard 项目的意思。

Well, I'd be surprised by the end of this year if digital human emulation has not been solved. I guess that's what we sort of mean by the MacroHard project.马斯克如何把 1 太瓦 GPU 送上太空 · 01:00:39
Elon Musk通用性与量产规模
真正可用的人形机器人尚不存在;若同时获得通用能力与量产能力,它可能成为比手机更大的产品类别。

至于人形机器人,目前没有真正有用的人形机器人,只有一些噱头。

Um there are sort of gimmicks, but there are there are no actually useful humanoid robots.FULL: Elon Musk Makes Shocking Future Predictions At U.S.-S… · 00:03:06
Elon Musk工业级基础设施
xAI 的 50→500MW 设施计划依赖国家、资本、能源与芯片供应商协同;前沿竞争已超出单一实验室可以独立承担的尺度。

埃隆,先谈 xAI。马斯克:我们很高兴宣布,我们在做一个 500 兆瓦——对,500 兆瓦。

Let's begin with you Elon. the things that we're doing with XAI.FULL: Elon Musk Makes Shocking Future Predictions At U.S.-S… · 00:15:40
Elon高承诺预测
轨道太阳能算力。

这就是您通过将其移入太空而可能节省的部分。其中大部分是 GPU。

That's the part you're presumably saving by moving this into space. Most of it's the GPUs.马斯克如何把 1 太瓦 GPU 送上太空 · 00:00:30
Elon风险排序
非零风险与丰裕预期并存。

那还需要钱做什么?>> 但机器人由你创造,现在和未来你都需要钱。

What do you need money for in that case? >> But you're you're creating these robots.The full-length interview with Elon Musk | The Economist · 00:05:23
证据附录37 / 37