AI Interview Wiki
主题综述

分歧图 · Disagreement Map

全库级的正面对撞清单：同一命题、双方都有真信念（conviction）的分歧，跨越主题边界收集。与各主题页里的「分歧在哪」小节互补——那里按主题内聚，这里按命题对撞。

更新日志

如何读

模型与研究前沿

缩放定律正在触顶 · Scaling laws are plateauing

正方 3 · 反方 6 · 持中 2

正方

Andrew Ng——规模这颗柠檬还有一点汁可榨，但已极难；「规模第一」的叙事是少数公司 PR 机器带出来的偏见。

"There's probably a little bit more juice out of the scalability lemon to be squeezed, so hopefully, we can still make progress there, but it's getting really, really difficult."
「可扩展性这颗柠檬可能还有一点汁水可以榨取，所以希望我们仍然可以在这方面取得进展，但这变得真的非常困难。」
Andrew Ng · How Agentic AI is Transforming the Startup Landscape

Rob Toews——GPT-4 以来增量递减，持续学习与样本效率是根本性缺口。

"Things like continual learning or sample efficiency, those are fundamental problems that we're not seeing improvements happening in the current set of models. So I think there's no question that the models are plateauing."
「像持续学习或样本效率这类根本问题，在当前这批模型里我们没有看到改进。所以我认为毫无疑问，模型正在进入平台期。」
Rob Toews · AI Vibe Check: The Actual Bottleneck in Research

反方

Sam Altman——OpenAI 的创立信条当年被当成异端，如今被验证；规模仍被低估。

"When we started, yeah, the core beliefs were deep learning works and it gets better with scale. And I think those were both somewhat heretical beliefs."
「我们创立时的核心信念就是：深度学习有效，而且随规模变好。当时这两条都多少算是异端。」
Sam Altman · How To Build The Future: Sam Altman

Jakub Pachocki & Szymon Sidor (OpenAI)——预训练范式没有消失，会与推理范式复合。

"One thing we always try to not underestimate is the importance of scaling. Even as we look at these reasoning models, it's not like the previous scaling paradigm of pre-training has vanished. I think we will see these things compound."
「有一件事我们始终提醒自己不要低估：规模的重要性。即便在看这些推理模型时，之前的预训练缩放范式也并没有消失。我认为这些东西会复合叠加。」
Jakub Pachocki & Szymon Sidor · Episode 5 - Defining AGI and the road ahead

Joelle Pineau (Cohere)——历史上押注缩放定律失败的人都输了，但它不单独起效。

"The scaling laws have been remarkably robust. They don't play exactly as we expect, but still, they've been remarkably robust. Lots of people have bet against scaling laws in the past, and I would say overall, you know, we've seen a pretty robust effect. They don't work alone. We also need these algorithmic innovations."
「缩放定律一直异常稳健。它们并不完全按我们预期的方式展开，但仍然异常稳健。过去很多人押注缩放定律会失效，总体来说我们看到的效应相当稳健。它们不单独起效，还需要算法创新配合。」
Joelle Pineau · 20VC: Cohere's Chief Scientist on Why Scaling Laws Will Continue

Nathan Labenz——IMO 金牌是 GPT-4 做不到的跃迁；「变慢」的观感多来自 GPT-5 发布事故。

"We had an IMO gold medal with pure reasoning models with no access to tools from multiple companies. And, you know, that is night and day compared to what GPT-4 could do with math, right?"
「多家公司用不带工具的纯推理模型拿到了 IMO 金牌。这和 GPT-4 在数学上能做到的相比，是天壤之别。」
Nathan Labenz · Is AI Slowing Down? Nathan Labenz

Mark Chen (OpenAI 首席研究官)——「预训练已死」的叙事在 LLM 史上每遇瓶颈就出现一次，但更细致的研究/数据工程总能解锁下一步；预训练如今反而是被低估的方向。

"Like more careful research engineering, more careful data engineering, more careful scaling. And it always unlocks that next ability to scale further. So I mean, it's held for almost 10 orders of magnitude. There's no reason it should not keep holding."
「更加细致的研究工程，更加细致的数据工程，更加细致的规模。它总是解锁更进一步的规模能力。所以我的观点是，这已经持续了近 10 个数量级。没有理由它不会继续保持下去。」
Mark Chen · Cooking with OpenAI's Research Chief: AGI, o1, Evals, and Scaling Laws

Igor Babushkin (xAI 联合创始人)——大规模集中预训练已收益递减，数据与算力的继续扩张都愈发艰难。（罕见的实验室联创正方声音）

"But eventually something will have to give. So at some point, you can't cover the entire earth with GPUs."
「但最终总会有些事情必须让步。因此在某些时候，你无法用 GPU 覆盖整个地球。」
Igor Babushkin · Ep 92: xAI Co-Founder Unpacks the Future of Model Development

持中/条件立场

现有范式（Transformer+规模+RL）足以达到 AGI · The current paradigm suffices to reach AGI

正方 2 · 反方 7 · 持中 3

正方

Sholto Douglas (Anthropic)——趋势线尚未弯曲，他坚信算法层不存在根本局限。（该篇逐字稿为中文版）

「话虽如此，我认为我看到的每条趋势线都表明，我们有望在大多数我们训练的任务上获得专家级超人的可靠性。」
— Sholto Douglas · Ep 66: Sholto Douglas (Anthropic)

Mark Chen (OpenAI 首席研究官)——不接受「还需 2-3 个突破（如持续学习）才到 AGI」的框架：持续学习未必是必须解锁的基本原语，现有多条路线（shots on goal）中会有奏效者。与同在 OpenAI 的 Noam Brown（下方反方）形成同实验室内部张力。

"I don't know if that same framing like continue learning is a basic primitive that you have to unlock. There's so many different techniques. ... I don't know what would consider as a breakthrough versus not, but I think there are clearly many shots on goal and I am pretty sure they'll work."
「我不知道是否那种框架，比如持续学习，是你必须解锁的基本原语。有很多不同的技术。……我不知道什么算是突破与否，但我认为明显还有很多机会，而且我很确定它们会奏效。」
Mark Chen · Cooking with OpenAI's Research Chief: AGI, o1, Evals, and Scaling Laws

反方

Noam Brown (OpenAI)——只堆预训练到不了超级智能，会先触到经济可行性的天花板。（该篇逐字稿为中文版）

「LLM 领域的一些人确信，只要我们不断扩大预训练的规模，我们就能达到超级智能。我对这种观点持怀疑态度。」
— Noam Brown · Scaling Test Time Compute to Multi-Agent Civilizations

Joelle Pineau (Cohere)——奖励训练的理念根本且不会消失，但 RL 开箱即用到不了 AGI。

"Where we're maybe getting a little bit ahead is thinking that just RL out of the box is going to give us AGI. That part, a lot less so. You know, if you look at the curve of progress, RL is terribly inefficient."
「我们可能有点想过头的地方，是以为开箱即用的 RL 就能带来 AGI。那部分我远没那么信。看看进展曲线，RL 的效率低得可怕。」
Joelle Pineau · 20VC: Cohere's Chief Scientist on Why Scaling Laws Will Continue

Edwin Chen (Surge AI)——需要能模仿人类上百万种学习方式的新算法与新数据。

"I'm in a camp where I do believe that something new will be needed. ... I believe that in the same way that there's a million different ways that humans learn, we need to build models that can mimic all those ways as well."
「我属于相信需要新东西的阵营。……就像人类有上百万种不同的学习方式一样，我们需要造出能模仿所有这些方式的模型。」
Edwin Chen · The 100-person AI lab that became Anthropic and Google's secret weapon

Fei-Fei Li (World Labs)——语言之后的下一个前沿是空间智能：AI 必须能生成、理解并作用于三维世界，纯语言路线到不了。

"We are building the next frontier of AI, which is what we call spatial intelligence. And spatial intelligence is about Creating AI that has the ability to generate, understand, reason with, and interact with spaces, whether it's physical or virtual."
「我们在构建 AI 的下一个前沿，也就是我们所说的空间智能。空间智能是要创造出有能力生成、理解、推理并与空间交互的 AI——无论是物理空间还是虚拟空间。」
Fei-Fei Li · Fei-Fei Li on Spatial Intelligence and Robotics

持中/条件立场

基准测试与竞技场排行榜能衡量真实的 AI 进展 · Benchmarks and arenas measure real progress

正方 2 · 反方 5 · 持中 1

正方

Nathan Lambert (AI2)——Arena 争议缠身，但人类偏好数据的价值可能被低估。

"It's like people are down on chatbot arena, but it might be that the human data helps boost retention time and general preference a lot."
「人们现在看衰 Chatbot Arena，但很可能正是这种人类数据大幅提升了留存时长和整体偏好。」
Nathan Lambert · The RLVR Revolution

反方

Edwin Chen (Surge AI)——对着 Arena 优化等于把模型训练成点击诱饵。

"One of the things I think is a giant plague on AI is LLMSys, LLM Arena. ... if you don't do this, you're basically just training your models on the analog of clickbait."
「我认为 AI 领域的一大瘟疫就是 LMSYS、LLM Arena。……如果你不做（耗时的专业人工评估），你基本上就是在用点击诱饵的等价物训练模型。」
Edwin Chen · The Power of Quality Human Data

Brendan Foody (Mercor)——学术评估与买家关心的结果脱节，评估正转向真实工作能力。

"Historically, everyone's been pointing to these academic evals of PhD-level reasoning with GPQA, Humanities Last Exam, or Olympiad Math, but now it's moving towards the capabilities that people practically care about, of how do we get models to automate the way that we build a software platform or automate the way that we do an investment banking analysis."
「过去大家都盯着 GPQA、Humanity's Last Exam、奥数这类博士级推理的学术评估，但现在正转向人们实际关心的能力：如何让模型自动化我们构建软件平台的方式、自动化投行分析的做法。」
Brendan Foody · Why experts writing AI evals is creating the fastest-growing companies

Michelle Pokrass (OpenAI)——为基准优化的模型「看起来很棒」但不好用。（该篇逐字稿为中文版）

「有时你会为了基准而优化模型，这样看起来很棒。然后你真正尝试使用它时，会遇到一些基本问题，比如，它不听从我的指示，或者格式很奇怪。」
— Michelle Pokrass · Ep 64: GPT-4.1 Lead Michelle Pokrass

Mia Glaese & Olivia Watkins (OpenAI Frontier Evals)——SWE-Bench Verified 已饱和且被污染，近期「停滞」其实是基准失效。

"We realize that this is because the eval is effectively saturated and also highly contaminated. So at this point, we think that it's not really measuring coding performance improvements well anymore."
「我们意识到这是因为这个评估实际上已经饱和、且被高度污染。所以到这个时点，我们认为它已经无法很好地衡量编码性能的提升了。」
Mia Glaese & Olivia Watkins · SWE-Bench-Dead: The End of SWE-Bench Verified

持中/条件立场

编程已基本被解决、即将全自动化 · Coding is largely solved

正方 2 · 反方 3

正方

Boris Cherny (Anthropic, Claude Code 负责人)——对他做的这类编程，编码已是被解决的问题。

"I think at this point it's safe to say that coding is largely solved. At least for the kind of programming that I do, it's just a solved problem because quad [Claude] can do it."
「我认为到现在可以放心地说，编码已经基本被解决了。至少对我做的这类编程而言，它就是个已解决的问题，因为 Claude 能做。」
Boris Cherny · Head of Claude Code: What happens after coding is solved

Fiona Fung (Anthropic, Claude Code & Cowork 团队负责人)——与同队 Boris Cherny 互为印证但换了度量：编码不再是瓶颈（工程师人均季度交付 8x、设计师 PM 都在提交代码），未解决的部分已转移到验证与质量。

"And that's a shift that you know, like we're seeing with With Claude Code and co-work is coding is no longer the bottleneck. ... Now, not only engineers, we also have designers, PMs, everybody on the Claude Code team checks in code. When not only more people checking in code, but different disciplines, but also the throughput is so high, how do we think about verification? That's this Other shift that I'm seeing."
「这是一个转变，你知道的，现在我们和 Claude Code 和团队合作时，编码不再是瓶颈。……现在，不仅仅是工程师，我们还有设计师、PM，Claude Code 团队中的每个人都在提交代码。不仅更多的人在提交代码，还有不同的学科，而且产出率如此高，我们如何考虑验证？这是我看到的另一个转变。」
Fiona Fung · Building the most AI-pilled engineering team in the world

反方

Michael Truell (Anysphere/Cursor)——存在「漫长而混乱的中间阶段」，人类必须留在驾驶座上。（该篇逐字稿为中文版）

「我们希望人类掌握主导权。而且，我们认为即使在最终状态下，让人们掌控一切也非常重要。你需要专业人士来完成这些，并决定软件的样子。所以我认为，是的，绝对需要工程师。」
— Michael Truell · The rise of Cursor

Tony Fadell——无人类架构监督的 AI 代码是快时尚软件，积累技术债。

"There is this dichotomy of fast and throwaway. It's called fast fashion. We got fast software. But software, if you're going to build a real company, can't be throwaway."
「存在这种『快与即抛』的二元对立，时尚业叫快时尚。我们现在有了快软件。但如果你要建一家真正的公司，软件不能是即抛的。」
Tony Fadell · Father of the iPod and iPhone on building taste and judgment

Andrew Ng——拒绝 vibe coding 一词：AI 辅助编码是深度智力活动。

"Vibe coding leads people to think, you know, like, I'm just going to go to the vibes and accept all the changes ... So when I'm coding for a day or for an afternoon, I'm not like going with the vibes."
「vibe coding 这个词让人以为，我只要凭感觉走、接受所有改动就行……但当我编码一整天或一下午时，我根本不是在凭感觉。」
Andrew Ng · How Agentic AI is Transforming the Startup Landscape

单一通用模型胜过专用/垂直模型 · One general model beats specialized models

正方 10 · 反方 8 · 持中 1

正方

Jeff Dean (Google)——垂直模型只是过渡形态（AlphaProof 被通用 Gemini 取代）。

"I mean, I think general models will win out over specialized ones in most cases."
「我认为在大多数情况下，通用模型会胜过专用模型。」
Jeff Dean · The AI Frontier from Gemini 3 (Jeff Dean)

Boris Cherny (Anthropic)——The Bitter Lesson 的最大推论：永远押注更通用的模型。

"Rich Sutton had this blog post maybe 10 years ago called The Bitter Lesson. ... His idea was that the more general model will always outperform the more specific model. ... For me, the biggest one is just always bet on the more general model."
「Rich Sutton 大约十年前写过那篇《The Bitter Lesson》。……他的观点是：更通用的模型总会胜过更专用的模型。……对我来说最大的推论就是：永远押注更通用的模型。」
Boris Cherny · Head of Claude Code: What happens after coding is solved

Jakub Pachocki & Szymon Sidor (OpenAI)——刻意不做点状部署，重大发现来自 generality。

"We don't really think of it as let's take these specific domains and let's deploy this technology there. I think that is a way to make faster pointwise progress, but I think the potential for the really big discoveries and the most meaningful technology advancement comes from this generality."
「我们并不把它想成『挑几个特定领域、把技术部署进去』。那样能更快取得点状进展，但真正重大的发现、最有意义的技术进步，潜力来自通用性。」
Jakub Pachocki & Szymon Sidor · Episode 5 - Defining AGI and the road ahead

Mark Chen (OpenAI 首席研究官)——模态维度上的同一立场：文本/图像/音频/视频应统一在单一模型与基础设施栈下，架构种类越少越好。

"For a research lab, I think there are a lot of advantages for it to being under one. So you just have to maintain one infrastructure stack, for instance. ... I think there's a strong bias for us to keep it in as few different architectures as possible."
「我认为对于一个研究实验室来说，将所有东西整合在一个模型下有很多优势。比如，你只需维护一个基础设施堆栈。……我认为我们有很强的倾向，希望将其保持在尽可能少的不同架构中。」
Mark Chen · Cooking with OpenAI's Research Chief: AGI, o1, Evals, and Scaling Laws

Josh & Matt (Chai Discovery)——把 bitter lesson 推进到生物学：拒绝为生物手写领域架构，扩数据/算力/模型胜过编码生物规则，糖基化位点这类隐藏特征模型自己会学出来。

"It's this bitter lesson pillness of scaling compute and data and models. It's being really rigorous. That's something that's so important in this space. You can fool yourself so easily in biology."
「就是这种『苦涩教训』式的信念：扩算力、扩数据、扩模型。同时要极其严谨——这在这个领域太重要了。在生物学里你太容易骗过自己。」
Chai Discovery 联合创始人 · Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem

反方

Edwin Chen (Surge AI)——世界太丰富，永远不会有 one-size-fits-all。

"Over the past year, I've started to realize that it's almost like every company should have a thesis on, like, the world is just so rich. There's never going to be a one-size-fits-all solution."
「过去一年我开始意识到：每家公司都该有自己的论点——世界实在太丰富了，永远不会有一个放之四海而皆准的方案。」
Edwin Chen · Ep 80: Edwin Chen on Why Frontier Labs Need Their Own Thesis

Wade Foster (Zapier)——最好的 agent 很小且专注。

"I think the real challenge that folks run into is the best agents are small. And you really focus them on a particular task that you want to go achieve."
「我认为大家真正遇到的难题在于：最好的 agent 是小的。你要把它聚焦在你想完成的一个特定任务上。」
Wade Foster · AI Agents, Clearly Explained in 40 Minutes

Jessy Lin & Dan Biderman (Engram)——直接对撞 Frontier Lab 世界观：大量需求是私有、专有甚至相互冲突的（永远进不了后训练数据集），未来是每个人/每个团队拥有自己持续训练的模型。

"The worldview that we have is a bit different from the Frontier Lab worldview, where it's like we want one model that's bigger and bigger, that's more and more intelligent across a variety of domains. Instead, how we see it, like we kind of imagine this world where everybody has their own model. A lot of the things that people want to learn are either private, like things that will never see the light of day in a post-training data set, or even conflicting."
「我们所持的世界观与前沿实验室的世界观有点不同，他们想要一个越来越大、越来越强的模型，在各种领域中变得越来越智能。相反，我们的观点是，我们想象的世界是每个人拥有自己的模型。很多人想要学习的东西要么是私密的，比如在后训练数据集中永远不会曝光的内容，要么是相互矛盾的。」
Jessy Lin · Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin

Scott Wu (Cognition)——Devin 刻意做成复合模型系统，按子任务在 Anthropic/OpenAI/Google/自研开源模型间动态路由，拒绝绑定单一模型。

"So Devin, rather than being pegged to one model, And saying, oh, we're only going to serve you GPT for this, or we're only going to serve you Opus for this. Devin can use any of the different models it has in its arsenal, which include all of these models from Anthropic, OpenAI, Google, etc. But also our own models, right, or open source models out there. And it will, you know, dynamically go and choose these models for these tasks."
「因此，Devin 不再限制于一个模型，而是说，我们只会为您提供 GPT。或者，我们只会为您提供 Opus。Devin 可以使用其武器库中不同的模型，包括来自 Anthropic 的所有模型，OpenAI、Google 等等。还有我们自己的模型，当然，或是其他开源模型。它会动态选择这些模型来完成这些任务。」
Scott Wu · Scott Wu, Cognition

Lin Qiao (Fireworks)——把 Engram 的「每人一个模型」推到企业侧的极端版本：未来不是少数 AGI 主导，而是数百万个专用模型，一个应用一个用例一个。

"And I really believe the future will not be a few small number of AGI models, dominant world. I really believe the future will be, it may be scary, but I think that's true. It will be millions of specialized model, one per application per use case."
「我真的相信未来不会是少数几个主导世界的 AGI 模型。我真的相信未来可能会很可怕，但我认为这是真实的。它将是数百万个专业化模型，每个应用和用例一个。」
Lin Qiao · 20VC: Are OpenAI and Anthropic Overvalued? The Open-Weight Case

持中/条件立场

记忆与知识应内化进模型权重 · Memory & knowledge belong in the weights

正方 1 · 反方 3

新对撞：agent 的长期记忆（以及世界知识）应该训练进参数，还是外置到文件系统/检索层？双方都承认另一侧机制会存在，分歧在于哪一侧是主干、哪一侧是补丁。

正方

Jessy Lin & Dan Biderman (Engram)——不内化就无法组合出复杂抽象思维；「事实 vs 技能」是伪二分法，检索的根本瓶颈是寻址（模型不知道该查什么）。

"I think if you look at like how the models think, if you need to recall basic facts in order to like take the next step in your thinking, you can't get very far. ... In order to like, think more and more complex and deep thoughts about things, you kind of need to internalize some things so that you can compose them into more abstract concepts."
「我觉得如果你看看模型是如何思考的，如果你需要回忆基本事实才能进行下一步思考，你就无法走得很远。……为了能够更复杂、更深刻地思考事物，你需要内化一些东西，以便将它们组合成更抽象的概念。」
Jessy Lin · Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin

反方

Oriol Vinyals (Google DeepMind, Gemini 共同负责人)——情景记忆的解在非参数的文件系统式存储；把不同用户的记忆编进权重在大规模 serving 下不现实。

"But I think the mechanism I want it to work kind of, or I mean, it's going to clearly work better and better is this kind of File system style like nonparametric. It's a bit more convenient than integrating those back into the weights because even from a practical point of view, we try to serve one model at scale. So what it would be really like painful to have to serve one model with different memories to users."
「我认为我希望其运作的机制，或者说，显然会越来越好的是这种文件系统风格的非参数化机制。这比将那些重新整合到权重中更方便，因为即使从实际角度来看，我们也试图以规模服务一个模型。因此，服务一个具有不同记忆的模型对用户来说会非常痛苦。」
Oriol Vinyals · Ep 87: Gemini Co-Lead on World Models, RL's Next Domains & Continual Learning

Will Bryk (Exa)——权重里记世界知识是浪费：知识外置给检索、智能留在参数里，「从未见过世界的爱因斯坦」才是省算力的正确架构——恰是 Engram 反对的「剥掉事实的纯核心模型」。

"If you're wasting those weights on like all sorts of information about the world, like the capital of France or, you know, this random blog that you read, like you're just wasting tokens. Sorry, you're wasting weights. Those weights should be focused only on like intelligent processing. And you could probably get to models that are like one billion, even less than a billion parameters that are extremely hyper intelligent and completely Unknowledgeable. It's like Einstein who never saw the world."
「如果那些权重在关于世界的各种信息上浪费，比如法国的首都，或者你阅读的某个随机博客，那么你就是在浪费代币。对不起，你在浪费权重。这些权重应该只集中在智能处理上。你可能会得到像十亿，甚至少于十亿参数的模型，它们是极其高度智能且完全无知的。就像是爱因斯坦从未见过世界。」
Will Bryk · Building Search for AI Agents with Exa CEO Will Bryk

Ali Taha (Baseten)——第三条路线：不进权重也不靠外部检索，而是把 KV-cache 压缩到近乎无限——他坦承自己在内部争论中被说服、认输。

"If you're able to sort of make your KV almost infinite and you're able to compact in such a way that you don't lose any of the knowledge. In that case, you can actually do continual learning and you can actually solve continual learning. And it's a result of this argument that Charlie and I had, that I do concede that his point was correct. And I do see that KV-cache is the way forward."
「如果你能让 KV 近乎无限，并且压缩得不丢失任何知识——那种情况下你就真的能做持续学习、真的能解决持续学习。这是我和 Charlie 那场争论的结果：我承认他的观点是对的，我确实认为 KV-cache 才是前路。」
Ali Taha · The Inference Engineering Masterclass (Baseten)

（正方 Engram 一侧新增独立佐证：Dan Biderman 单独受访时把外置路线拆成两层反驳——上下文越长模型越混乱（context rot），千万 token 也无法整体推理；RAG 擅长取数但替代不了内化知识，面对万亿 token 专有数据反复重读语料不可持续。（逐字稿未含此句，结构层转述）The AI Memory Problem: Why Long Context Isn't Enough）

递归自我改进将带来快速起飞 · Recursive self-improvement leads to fast takeoff

正方 1 · 反方 3

新对撞：分歧不在「AI 会不会加速 AI 研发」（双方都承认会），而在这个回路是否在数年内压缩成一夜式的智能爆炸——Ben Mann 押 2028 年 50 百分位，Noam Brown 认为测试时算力的物理时间本身就是起飞的稳定器。

正方

Ben Mann (Anthropic 联合创始人)——AI 2027 式的递归自我改进循环（Claude 5 造 Claude 6）到 2028 年导致多数领域超人 AI，他认为「很有可能」。（该篇逐字稿为中文版）

「我读过 AI 2027，它基本上就是你刚才描述的故事。它预测在 2028 年……这是我们拥有这种递归的自我改进循环的第 50 个百分位预测。带领我们走向在大多数领域看起来像超人 AI 的东西。我认为这对我们来说非常重要。」
— Ben Mann · Will we have Superintelligence by 2028? (Ben Mann)

反方

Noam Brown (OpenAI)——解锁模型最高能力要跑数周乃至数月的大规模测试时计算，时间本身成为物理瓶颈和起飞的稳定器。

"And you have basically, in an instance, the models just You know, becoming very superhuman across the board in moments. And I don't think we're headed to that world, largely because of the fact that the models rely so much on large-scale test-time compute in order to achieve their greatest intelligence."
「然后，你基本上在某一刻，模型就 你知道，瞬间变得超级人类。而且我认为我们并不是走向那样的世界，主要是因为模型在实现最大智能时，依赖于大规模的测试计算。」
Noam Brown · Why Traditional Benchmarks Fail Modern AI Models

Jonathan Siddharth (Turing)——现在才第一局，把所有领域数据灌进模型需要很长时间；缓慢稳定起飞。

"It's innings one and I believe in slow takeoff. I'm sorry to pour cold water on all the AI doomers that might be listening to this, but we are not in a rapid takeoff scenario. I believe in slow steady takeoff for AGI and eventually superintelligence."
「现在是早期阶段，我信奉缓慢起飞。我很抱歉要给那些可能正在收听此节目的 AI 末日论者泼冷水，但我们并没有处于快速起飞的情形中。我相信 AGI，以及最终的超级智能，将会缓慢而稳定地起飞。」
Jonathan Siddharth · 20VC: Scale, Surge, Turing, Mercor — Who Wins & Who Loses

Scott Wu (Cognition)——承认「事件视界」存在（且「Devin 的大部分代码由 Devin 写」），但六个月太短：大量问题非纯智能可解，组织变革、流程和硬件都是硬约束。

"I agree there's some point at which there's almost like this event horizon in AI. There's some point in which the AI gets so smart that it's hard for us to even reason about what that world will look like. ... So the event horizon does exist, and I think past that point it's hard to predict, but it's just like, you just gotta stay sane, you know?"
「我同意 AI 里存在某个近乎『事件视界』的点——某个时刻 AI 聪明到我们连那个世界长什么样都难以推理。……所以事件视界确实存在，越过那一点确实难以预测，但你就是得保持清醒，你懂吗？」
Scott Wu · Inside the Fastest-Growing Category in AI

资本与市场

我们正处于 AI 泡沫之中 · We are in an AI bubble

正方 3 · 反方 10 · 持中 3

正方

David Cahn (Sequoia)——泡沫论一年前是逆共识，如今连大多头都承认；脆弱性人人可见。

"I do think we're in an AI bubble. You can see the fragility. Everybody can see the fragility. The thing that I think is more interesting is who's going to survive the bubble."
「我确实认为我们在 AI 泡沫里。你能看到脆弱性，每个人都能看到。我觉得更有意思的问题是：谁能活过这场泡沫。」
David Cahn · 20VC: Sequoia's David Cahn on The Winners and Losers

Shensi Ding & Gil Feig (Merge)——3,000 倍收入倍数的融资令人沮丧，这些公司多数会失败。

"It was really frustrating to see some companies where we knew what their revenue was and we knew they were raising it like a 3,000x multiple on that revenue."
「看到一些公司——我们知道他们的真实收入——却在以大约 3,000 倍收入的估值融资，真的很让人沮丧。」
Shensi Ding · The Token-Maxxing Bill That Shocks Every CFO (Merge)

Trae Stephens & Delian Asparouhov (Founders Fund)——内部硬数据版的泡沫论：Delian 披露 FF 最新风险基金平均入场估值约 $700M（此前五期约 $80-100M），Trae 直言系统里的钱太多了；但两人的解法是坚守微观选股而非宏观择时退场。

"I think the thing that the founders are missing in this equation is that like the economics At these different prices at the early stages shift significantly. And so like there's so much money in venture that there are all these new funds that are setting up. It's like way too aspirational. There's way too much money in the system."
「我认为创始人在这个方程中缺失的东西是，像这种不同价格的经济要素在早期阶段有显著变化。所以，风险投资中有如此多的资金，这些新基金正在设立。这实在是太过于追求理想了。系统中有太多资金。」
Trae Stephens · Uncapped #53 | Trae and Delian from Founders Fund

反方

Ben Horowitz (a16z)——所有人都喊泡沫时就不是泡沫，泡沫需要「投降」。

"The one thing about bubbles is anytime everybody thinks it's a bubble, it's not a bubble. Because in order for it to bubble, you need capitulation."
「关于泡沫有一条：只要所有人都认为是泡沫，它就不是泡沫。因为要形成泡沫，你需要『投降』——需要人们普遍相信它不是泡沫。」
Ben Horowitz · Ben Horowitz: Why Hesitation is a CEO's Worst Enemy

Jensen Huang (NVIDIA)——数万亿美元存量通用计算必然更新为加速计算，泡沫论误解了基本盘。

"General purpose computing is over and the future is accelerated computing and AI computing. That's the first point. And so the way to think about that is there's how many trillions of dollars of computing infrastructures in the world that has to be refreshed."
「通用计算已经终结，未来是加速计算和 AI 计算。这是第一点。理解它的方式是：全世界有多少万亿美元的计算基础设施必须更新换代。」
Jensen Huang · NVIDIA: OpenAI, Future of Compute, and the American Dream (BG2)

Sam Altman——足够低价位下对智能的需求实际无上限。

"In some sense, I think demand for intelligence at a low enough price is effectively uncapped."
「某种意义上，我认为在足够低的价格下，对智能的需求实际上是没有上限的。」
Sam Altman · Sam Altman in conversation with Patrick Collison

Doug Leone (Sequoia)——泡沫意味着投入即亏损；资本过剩不等于泡沫。

"The word bubble implies you invest money in and you lose money because either due to lack of supply of companies or abundance of capital. There's certainly an abundance of capital."
「『泡沫』这个词意味着你投进钱就会亏掉——要么因为好公司供给不足，要么因为资本过剩。资本确实过剩，但（这一轮是真实周期的前端）。」
Doug Leone · Why AI Will Transform Customer Experience (Cresta)

David George (a16z Growth)——capex 由现金流支撑，占收入比远低于 dot-com。

"So relative to the dot-com, CapEx is actually supported by cash flows and CapEx as a percentage of revenue is considerably lower."
「相对于互联网泡沫时期，这轮 capex 实际由现金流支撑，而且 capex 占收入之比要低得多。」
David George · The State of Markets

Jonathan Siddharth (Turing)——看不到泡沫：我们只是习惯了魔法。

"I don't see an AI bubble. These models are incredibly powerful today."
「我看不到 AI 泡沫。今天这些模型已经强大得不可思议。」
Jonathan Siddharth · 20VC: Scale, Surge, Turing, Mercor — Who Wins & Who Loses

Sarah Friar (OpenAI)——需求唯一的限制是算力供给而非市场饱和。

"Today, we feel absolutely constrained on compute. There are many more products that we could launch. Many more models that we would train, many more multimodality things we would explore if we had more compute today."
「今天我们在算力上是绝对受限的。如果现在有更多算力，我们可以发布多得多的产品、训练多得多的模型、探索多得多的多模态方向。」
Sarah Friar · Episode 12 - State of the AI Industry (Friar & Khosla)

Gavin Baker (Atreides)——本节最细的反方论证：建设主体靠经营现金流而非信贷；但他不掩盖唯一真实的软肋（信用利差走阔、CDS 放大）。

"So real yields go up and spreads widen. Meta priced a bond last week and it did not price where you would think a meta bond would price and this just shows that the credit market is."
「所以实际收益率上升、利差走阔。Meta 上周发债，定价并不在你以为 Meta 债券该在的位置——这恰恰说明信用市场（在紧张）。」
Gavin Baker · Gavin Baker - AI Market Jitters

持中/条件立场

AI 市场赢家通吃，将收敛为一至三家 · AI markets are winner-take-all

正方 3 · 反方 3

正方

Alfred Lin (Sequoia)——品牌网络效应+企业需要制衡选项，自然收敛到两三家。

"I think in technology, there is an element of network effects. There is a brand network effect and a lot of businesses I want to be able to trade off between one or two alternatives. ... And so there's a natural network effect in those businesses where it consolidates to number one, number two, and maybe a number three."
「科技行业存在网络效应的成分——品牌网络效应，而且很多企业希望能在一两个备选之间制衡。……所以这些生意里有一种自然的网络效应，会收敛到第一名、第二名，也许再加一个第三名。」
Alfred Lin · Alfred Lin on Airbnb, DoorDash, OpenAI & Citadel

David George (a16z Growth)——绝大多数市值创造归市场领导者，这一点被低估。

"I happen to think strongly, and my experience has been, the vast majority of market cap creation is going to go to the market leader. And this is probably underappreciated."
「我强烈认为——我的经验也是如此——绝大多数市值创造将归于市场领导者。而这一点很可能被低估了。」
David George · David George - Building a16z Growth

反方

Danny Rimer (Index)——历史表明会有数百家受益者。

"History would serve us well to assume that there's not going to be one beneficiary or even five beneficiaries of a specific sector, but hundreds of beneficiaries."
「历史经验告诉我们：一个行业不会只有一家受益者，甚至不止五家，而是数百家。」
Danny Rimer · Index Ventures' Danny Rimer on Figma's IPO

Martin Casado & Sarah Wang (a16z)——「不存在 AI」，只有一堆各需策略的子空间。

"We've kind of come to the opinion that there is no AI. There's like a bunch of subspaces that are totally different that all require their own strategy."
「我们逐渐形成一个观点：不存在『AI』这个东西。只有一堆完全不同的子空间，每个都需要自己的策略。」
Martin Casado · The State of AI: Growth, Fragmentation, and the Next Wave

Vinod Khosla——下一个数万亿美元公司不是另一个 AGI 实验室。

"Right now everybody wants to invest in the next open AI [OpenAI] and probably the next multi-trillion dollar company will not be another AGI research lab. It will probably be the thing that got built because AGI now existed."
「现在所有人都想投下一个 OpenAI，但下一个数万亿美元公司大概率不是另一个 AGI 研究实验室，而是『因为 AGI 存在了才建得起来的东西』。」
Vinod Khosla · Where is AI Taking Us (Altman & Khosla)

推理成本下降将缩减总算力需求（反杰文斯悖论） · Cheaper inference shrinks total compute demand

正方 1 · 反方 3

正方

Sebastian Siemiatkowski (Klarna)——AI 的核心价值是压缩企业冗余数据，照此推演算力需求会显著更少（他本人也承认娱乐侧生成是反向力量，「两边都能论证」）。

"If you look at any large enterprise company, they will have the same information over and over and over again. But if you look at Wikipedia, how many articles is there about Klarna? One. Why aren't there 15?"
「看任何一家大企业，同样的信息一遍又一遍地重复存储。但看 Wikipedia：关于 Klarna 的条目有几个？一个。为什么不是十五个？」
Sebastian Siemiatkowski · 20VC: SaaS is Dead — Why Systems of Record Will Die (Klarna)

反方

Tuhin Srivastava (Baseten)——降价的实证结果是塞入更多智能、跑更长的 agent。

"From the developer's perspective, they would insert more intelligence if you make it cheaper. They will insert more intelligence anyway, but if you make it more cheaper, they'll insert a hell of a lot more intelligence. You see this with agents. Agents are just longer running."
「从开发者的角度：你把价格降下来，他们就塞进更多智能。他们本来就会塞，但你降得越多，他们塞得越狠。agent 上就能看到——agent 只会越跑越长。」
Tuhin Srivastava · Baseten CEO Tuhin Srivastava on the AI Inference Cloud

Philippe Laffont (Coatue)——更省电省芯片的模型只会加速采用。

"If AI becomes materially cheaper, there's the whole Jevons paradox argument, which is like, yeah, the cheaper it becomes, the more people will find creative ways to use it and do other things."
「如果 AI 显著变便宜，就有整套杰文斯悖论的逻辑：它越便宜，人们越会找出各种创造性的用法去用它、去做别的事情。」
Philippe Laffont · Coatue CIO on AI's Biggest Winners

Roman Chernin (Nebius)——DeepSeek 时刻是公司史上最好的商业周。

"That was the best commercial week in the history of the company because so many people figured out that they can run inference in their production workloads with DeepSeek and economics will work. ... Every time we got intelligence cheaper. Same unit of intelligence cheaper, we are not reducing the consumption but we're increasing the [consumption]."
「那是公司历史上商业表现最好的一周——因为那么多人发现可以用 DeepSeek 在生产负载里跑推理、账算得过来。……每次智能变便宜——同一单位的智能更便宜——我们的消耗不是在减少，而是在增加。」
Roman Chernin · 20VC: Nebius Co-Founder on AI Infrastructure Bubbles

能源/电力是 AI 扩建的决定性瓶颈 · Energy is the binding constraint on the AI buildout

正方 2 · 反方 2 · 持中 1

正方

Jared Kushner——头号任务是加速国内能源建设与审批改革。

"We need to do two main things, which is one is make sure we have all the tools to compete, which is really accelerate our ability to build energy in the U.S. I think that'll be really, really bad if we're constrained on that."
「我们要做两件主要的事，第一是确保我们拥有竞争所需的全部工具——也就是切实加速美国国内的能源建设。如果我们在这上面被卡住，那会非常非常糟糕。」
Jared Kushner · Jared Kushner: BrainCo, Affinity Partners, and the Middle East

Shaun Maguire (Sequoia)——推理的瓶颈将是电力，因此「原子世界」最强的人被低估。

"We all know what's happening in AI. We're limited by power, limited by regulation, which is probably even more restrictive."
「我们都知道 AI 正在发生什么。我们被电力限制着，也被监管限制着——后者可能还更严。」
Shaun Maguire · How Elon Builds Trillion-Dollar Companies

反方

Sam Altman——他「当然希望」AI 消耗巨量能源：能源丰裕度与生活质量最相关。

"I mean. I sure hope so. I think that the thing that has most correlated with improvements in quality of life over history is increasing abundance of energy. I have no reason to believe that's going to stop."
「（被问到 AI 是否会消耗巨量能源）我当然希望如此。纵观历史，与生活质量改善相关性最强的，就是能源丰裕度的提升。我没有理由认为这会停止。」
Sam Altman · Sam Altman: The Future of AI

David Owen & Yafah Edelman (Epoch AI)——贵但可行的方案（光伏+储能）意味着能源并非真瓶颈。

"There are expensive technologies that exist right now, you could pay for solar power plus batteries. This is fairly small lead times. It might cost twice as much as normal power, but that's still way less than your GPUs, so you're going to do it if you have to."
「现在就存在贵一点的技术：你可以花钱上光伏加电池，交付周期相当短。可能比常规电力贵一倍，但仍远低于你的 GPU 开销——所以真到了那一步你就会去做。」
David Owen & Yafah Edelman · The 2045 Superintelligence Timeline (Epoch AI)

持中/条件立场

企业 AI 落地迅速且已产生正回报 · Enterprise AI adoption is fast and ROI-positive

正方 1 · 反方 5 · 持中 1

正方

Sarah Guo & Elad Gil——历来最慢采用技术的职业正在最快拥抱 AI。

"The people who have tended to be the slowest adopters of technology love AI. That's physicians. That's lawyers. That's certain accounting types. It's actually kind of fascinating. It's compliance. It's all the people who always never adopt technology are now adopting this stuff fast."
「一直以来采用技术最慢的那群人爱上了 AI：医生、律师、某些会计岗位、合规岗。那些从来不采用新技术的人，这次都在快速上手。这挺耐人寻味的。」
Sarah Guo & Elad Gil · The 2026 AI Forecast

反方

Sherwin Wu (OpenAI)——「很多 AI 部署实际是负 ROI，我不会惊讶。」

"It's actually really hard to measure these things. But especially from observing some companies trying to do AI, I would not be surprised if a lot of AI deployments are actually negative ROI."
「这些东西其实很难量化。但尤其是观察一些公司做 AI 的方式之后，如果很多 AI 部署实际上是负 ROI，我一点也不会惊讶。」
Sherwin Wu · "Engineers are becoming sorcerers" (Sherwin Wu)

Jason Droege (Scale AI)——演示误导人，真实部署要挖开每一条路。

"These things take six to 12 months to get them truly robust enough where an important process can be automated. Like with any of these major tech revolutions, headlines tell one story and then on the ground, laying broadband means you need to dig up every single road in America to lay it."
「要让一个重要流程真正自动化，这些东西需要六到十二个月才够健壮。就像每次重大技术革命：头条讲一个故事，而落到地面上——铺宽带意味着你得把全美国的每一条路都挖开。」
Jason Droege · First interview with Scale AI's CEO

Aaron Levie (Box)——财富 500 强接入 agent 是多年旅程。

"The context engineering is an incredibly hard problem because, again, you have access control challenges, you have different data formats ... That's where the Fortune 500 is. And so we ... have to be prepared as an industry where we're going to be on a multi-year march to be able to bring agents to the enterprise for these workflows."
「上下文工程是极难的问题：访问控制、异构数据格式……财富 500 强就处在这个阶段。整个行业要有准备：把 agent 带进企业工作流将是一场多年行军。」
Aaron Levie · Every Agent Needs a Box (Aaron Levie)

Patrick Collison (Stripe)——经济数据中几乎看不到生产力提升。

"There was a new paper published on this very recently ... Its claim is that one does not, in fact, observe productivity improvements stemming from use of language models."
「最近刚有一篇新论文……它的结论是：事实上并没有观察到语言模型的使用带来生产力提升。」
Patrick Collison · Patrick Collison on Stripe's Early Choices

持中/条件立场

收入规模本身能给公司去风险 · Revenue scale de-risks a company

正方 1 · 反方 3

新对撞：AI 时代的收入还等不等于安全边际？正方视极速收入爬坡为 PMF 的最佳信号；反方认为「越大越安全」的旧链条（PMF→单位经济→TAM 依次去风险）已断，减值风险与规模甚至呈怪异正相关。注意正方 Cahn 与反方 Lin 同属 Sequoia——同门对撞。

正方

David Cahn (Sequoia)——「$0-$100 俱乐部」：极速冲到 1 亿美元收入，是「你造出了真正有用的东西」的最佳可得指标。

"I think of it as the $0-$100 club. So I think it's a variation on this, which is the best AI companies right now are going $0-$100M of revenue very quickly. ... the best companies, it is the best indicator we have that you built something really useful."
「我把它看作是 “0 到 1 亿美元俱乐部”。所以我认为这是其中的一种变体，那就是现在最好的 AI 公司正在非常迅速地从 0 增长到 1 亿美元的收入。……而对于最好的公司来说，这是我们拥有的最佳指标，表明你构建了一些真正有用的东西。」
David Cahn · 20VC: Sequoia's David Cahn on The Winners and Losers

反方

Ev Randle (Benchmark)——电子表格投资时代终结：十亿美元收入的公司仍未证明单位经济与持久差异化。他同时警告：AI 应用公司毛利太高反而是红旗——推理本来就贵，高毛利说明没人在用你的 AI 功能。

"And I think that the thing that's changed massively in the new AI paradigm is that you can have businesses that are well over a billion dollars in revenue that haven't proven out their unit economics. They haven't proven out durable product differentiation. ... maybe there's like a weird positive correlated relationship with scale and risk."
「我认为在新的人工智能范式中，发生了巨大的变化，企业的年收入超过十亿美元，但尚未证明他们的单位经济学。他们尚未证明可持续的产品差异化。……甚至可能与规模和风险之间存在一种奇怪的正相关关系。」
Ev Randle · Benchmark's AI Bets: Cerebras, Sierra, Legora, Fireworks..

Alfred Lin (Sequoia)——收入是产出指标，「冲到 1 亿」的口号掩盖收入质量；健康度要看参与度、留存等输入指标与吞吐效率。

"This idea that we can just get to $100 million in revenue belies the fact that that's the goal, and yet I think quality of the revenue matters."
「我们认为能达到 1 亿美元营收的想法掩盖了一个事实，那就是目标，但我认为营收质量很重要。」
Alfred Lin · Inside Sequoia's Strategy

消耗制/用量计费是 AI 产品的正确定价模式 · Usage-based pricing is the right model for AI

正方 1 · 反方 2

新对撞：token/用量计费正在成为行业默认，但它是终点还是过渡态？正方视其为对齐成本与价值、防止被高频低价值任务「吃破产」的唯一现实解；反方认为 token 账单会复刻计时收费的全部复杂性、最终回摆订阅制，或该直接跳到按结果计费。

正方

Sarah Sachs (Notion)——不按用量计费就无法上线 agentic 功能：让想多用多付的客户找到出口，而不必把成本摊到曲线低端；且「并非所有知识工作都相同」。

"And if we wanted to ship this functionality, we couldn't afford it. It would bankrupt the company if we let, for instance, like autofill or the database autofill feature will soon be agentic. That will be associated with usage-based pricing because if every Single autofill action was an agent running on OBIS. On every single database cell, it would be billions of dollars, right?"
「如果我们想发布这个功能，我们就负担不起了。如果我们允许，例如，自动填充或数据库自动填充功能很快将变成代理式的，那将会使公司破产。这将与基于使用量的定价相关联，因为如果每一个自动填充操作都是一个在 OBIS 上运行的代理。在每一个数据库单元格上，那将是数十亿美元，对吧？」
Sarah Sachs · Notion's Token Town

反方

Gabe Pereyra (Harvey)——「转向消耗计费就解决一切」是当前最大误解：千万美元 agent 账单难以归因，token 计费将复刻计时收费的逐项审计生态，且他预判「某个时点你可能会回到类似订阅的模式，因为很难预测这些东西」。

"Like, I think when I talk with most people, they think, oh, just move to consumption pricing and that will solve all the problems. But I think there's going to be this very interesting dynamic where there's kind of a couple ways to price these things."
「我觉得，当我与大多数人交谈时，他们认为，哦，只需转向消费定价，这将解决所有问题。但我认为将会出现一个非常有趣的动态，因为定价这些东西的方式有几种。」
Gabe Pereyra · Harvey Co-Founder Gabe Pereyra on the Token Pricing Reckoning

Bret Taylor (Sierra)——token 用量与商业价值相关性弱，应像展示广告→CPC 那样把计费直接推进到结果层。

"We are trying to. So we do outcomes-based pricing. For a customer service context, that means if the AI agent resolves the case, no human intervention, there's a pre-negotiated rate for that. ... I'm a huge believer in this. I think the analogy of going from impression-based ads to CPC ads is apt."
「我们正在努力。所以我们做基于结果的定价。对于客户服务环境，这意味着如果 AI 代理解决了问题，没有人为干预，那么有一个预先商定的价格。……我非常相信这一点。我认为从基于展示次数的广告到 CPC 广告的类比是恰当的。」
Bret Taylor · Bret Taylor of Sierra on AI agents

就业与人

AI 将造成快速、破坏性的净失业 · AI causes fast, destabilizing net job destruction

正方 3 · 反方 12 · 持中 3

正方

Brendan Foody (Mercor)——替代会快速、痛苦，并成为重大政治问题。（Sarah Guo & Elad Gil 在年度回顾中特别标记了这一立场 · The Best of 2025 (So Far)）

"I think displacement in a lot of roles is going to happen very quickly, and it's going to be very painful and a large political problem. I think we're going to have a big populist movement around this."
「我认为很多岗位的替代会来得非常快，而且会非常痛苦，成为重大政治问题。我认为围绕这件事会出现一场大规模民粹运动。」
Brendan Foody · From Job Displacement to AI Trainers (Brendan Foody)

Nathan Labenz——Klarna、Salesforce 已公开因 AI 削减人力，冲击已真实发生。

"We are definitely seeing no less than like Marc Benioff has said that they've been able to cut a bunch of headcount because they've got AI agents now that are responding to every lead. Klarna, of course, has said very similar things for a while now."
「我们确实看到了：连 Marc Benioff 都说他们因为有了响应每条销售线索的 AI agent 而裁掉了一批人。Klarna 当然也早就说过非常类似的话。」
Nathan Labenz · Is AI Slowing Down? Nathan Labenz

Felix Rieseberg (Anthropic)——「我们自动化掉的烦人工作」正是初级员工的入场任务。

"At Anthropic, as a group of people, we're deeply worried about the impact that the tools are going to have on the labor market, especially for junior employees. I think it's only honest to say that when we talk about automating away a lot of the work that we personally find annoying ... In a lot of industries, that kind of work would have been given to a junior entry-level employee, right?"
「在 Anthropic，作为一个群体，我们对这些工具将给劳动力市场——尤其是初级员工——带来的冲击深感忧虑。诚实地说：我们嘴上说要自动化掉的那些『烦人工作』，在很多行业里恰恰是交给初级入门员工做的活。」
Felix Rieseberg · Why Anthropic Thinks AI Should Have Its Own Computer

反方

Marc Andreessen——人口崩塌+技术停滞 50 年，我们恰好在最需要 AI 时得到它。

"If we didn't have AI, we'd be in a panic right now about what's going to happen to the economy. ... We're going to have AI and robots precisely when we actually need them. The remaining human workers are going to be at a premium, not at a discount."
「如果没有 AI，我们现在应该正为经济的前景恐慌。……我们将恰好在真正需要的时候得到 AI 和机器人。剩下的人类工人将是溢价品，而不是折价品。」
Marc Andreessen · Marc Andreessen: The real AI boom hasn't even started

Alex Schultz (Meta CMO)——劳动总量谬误。

"One of the biggest problems here is what's called a lump of labor fallacy, which you assume that there's a certain fixed amount of work in the world. And it's not true for manual labor. It's not even true for, it's certainly not true for knowledgeable labor."
「这里最大的问题之一是所谓『劳动总量谬误』——假设世界上的工作总量是固定的。这对体力劳动不成立，对知识劳动更是绝对不成立。」
Alex Schultz · Meta CMO: How Advertising Works Under the Hood

Tyler Cowen & Alex Tabarrok——「50% 失业」与「工作周减半」几乎是同一件事。

"Suppose I tell you that AI is going to create 50% unemployment. Half of the people in the workforce will lose their jobs. That sounds terrible. ... Suppose, however, that I tell you that the work week will be cut in half. People will do half as much work. That actually sounds glorious. ... And yet, these are almost the same thing."
「假设我告诉你 AI 将造成 50% 的失业——一半劳动力失去工作。听起来很可怕。……但假设我告诉你：工作周将缩短一半，人们只需做一半的工作。这听起来简直美好。……然而这两者几乎是同一件事。」
Tyler Cowen & Alex Tabarrok · Tyler Cowen & Alex Tabarrok on AI, Jobs, and Economic Growth

Jensen Huang (NVIDIA)——radiology 12 年反例：被 AI 全面渗透后，放射科医生反而更多了。

"Radiology was completely penetrated by computer vision. ... However, the interesting thing is this, radiology demand went up. The number of radiologists in the world went up."
「放射科被计算机视觉全面渗透。……然而有趣的是：放射诊断的需求上升了，全世界放射科医生的数量也上升了。」
Jensen Huang · LIVE: Jensen Huang on Building the Dynamo of the Intelligence Age

Ryan Roslansky (LinkedIn)——LinkedIn 实时数据里看不到 AI 导致的大规模裁员。

"So the data we see so far, at least that I see real time right now, not seeing the former, not seeing companies reduce their workforce because of AI, definitely seeing a ton of job descriptions that now include some form of AI in them."
「就目前我们看到的数据——至少是我现在实时看到的——没有看到公司因 AI 裁减员工；确实看到大量职位描述里加入了某种形式的 AI 要求。」
Ryan Roslansky · LinkedIn's CEO on How AI Changes Your Job

David Friedberg——把举证责任放在正方身上：从 1963 年 Newsweek 的大型机恐慌到 80 年代 PC，每一轮技术恐慌都被后来的就业数据证伪。

"As soon as AI starts deleting jobs from the economy, I will change my position on this. But I need to see it actually happen first. There was a Newsweek article, you can pull it up. 1963, all about how computers are going to take all jobs and America's going to run out of jobs."
「一旦 AI 真开始从经济中抹掉岗位，我就会改变立场。但我需要先看到它真的发生。有篇 Newsweek 的文章你可以调出来——1963 年，通篇讲计算机将拿走所有工作、美国会没有工作可干。」
David Friedberg · David Friedberg on the Economic Crisis Nobody Wants to Talk About

持中/条件立场

AI 将全面替代（而非仅增强）专业工作者 · AI fully replaces professionals, not just augments

正方 3 · 反方 8 · 持中 2

正方

Victor Lazarte (Benchmark)——「增强论」是大公司的场面话。（该篇逐字稿为中文版）

「大公司 CEO 说，哦，不，AI 不会取代人类。AI 实际上是在增强人们的能力。这纯粹是胡说八道。它完全是在取代人类。」
— Victor Lazarte · 20VC: Benchmark's Victor Lazarte

Brendan Foody (Mercor)——在多数评估上模型评估人才已胜过人类招聘经理。

"We're already seeing on most of our evals that models are better than human hiring managers assessing talent, and it's still like the very early innings."
「我们已经在大多数评估里看到：模型评估人才的水平超过了人类招聘经理，而这还只是极早期。」
Brendan Foody · From Job Displacement to AI Trainers (Brendan Foody)

反方

Josh Tobin (OpenAI Deep Research)——「我根本不认为这是劳动力替代。」（该篇逐字稿为中文版）

「我认为没有任何工作是有风险的。我根本不认为这是一种劳动力替代。但对于这些知识型工作，你需要花费大量时间浏览信息并得出结论，我认为它会赋予人们超能力。」
— Josh Tobin · OpenAI's Deep Research Team

Winston Weinberg (Harvey)——律所拿到的总工作量会增加。

"I don't think so. I think we'll just get more work. There are certain things I don't want to pay for anymore, like marking up NDAs. But there's all these new things that I'm paying law firms for: AI risk, should you buy this company, is there a problem in X, Y, Z country with an act that is going to change it."
「我不这么认为。我认为我们只会得到更多工作。有些东西我不想再付钱了，比如 NDA 批注。但现在有一堆新事项要付钱给律所：AI 风险、该不该收购这家公司、某国某项法案会不会出问题。」
Winston Weinberg · 20VC: How Model Performance is Plateauing

Dylan Field (Figma)——AI 的价值是帮设计师探索更大的选项空间，距替代很远。

"In the engineering context, productivity is building something well. In the design context, it might be that it is just further exploration of the option space sometimes. Because I think right now, oftentimes, you're constrained by timeline, you can only explore so much, you might not be getting to the best solution."
「在工程语境里，生产力是把东西做好；在设计语境里，生产力有时就是对选项空间的进一步探索。现在你常被时间线限制，只能探索这么多，未必到得了最优解。」
Dylan Field · Uncapped #31: Dylan Field

Brian Halligan (HubSpot)——两个碳基生命之间基于信任的企业销售最晚被替代。

"I think ye olde enterprise sales where there's actual trust built up between two carbon-based life forms I think will be very, very, very late to go in the white-collar world."
「那种老派的企业销售——两个碳基生命体之间建立起真实信任的那种——我认为会是白领世界里非常、非常、非常晚才被替代的。」
Brian Halligan · Sequoia CEO coach: Why it's never been easier to start a company

持中/条件立场

AI 将缩减软件工程师岗位 · AI shrinks software-engineering headcount

正方 2 · 反方 6 · 持中 1

正方

Nathan Labenz——被迫二选一时，他已经选模型而不是初级工程师。

"I tend to think less. Already, if I just think about my own life and work, I'm like, would I rather have a model or would I rather have a junior marketer? I'm pretty sure I'd rather have the model. Would I rather have the models or a junior engineer? I think I'd probably rather have the models in a lot of cases."
「我倾向于认为（五年后工程师）更少。就我自己的工作而言：要模型还是要初级市场人员？我很确定选模型。要模型还是要初级工程师？很多情况下我大概也会选模型。」
Nathan Labenz · Is AI Slowing Down? Nathan Labenz

Michael Barton (Coatue)——今天的形态不是裁员而是招聘放缓。

"It's not that people are getting fired today or their jobs are being automated today. It's that the hiring is slowing."
「今天并不是人们在被解雇、岗位在被自动化，而是招聘正在放缓。」
Michael Barton · Inside Coatue: $70B Hedge Fund's AI & Retail Strategy

反方

Mike Cannon-Brookes (Atlassian)——与 AWS 的 Matt 同台「强烈一致」：五年后工程师更多。

"I was speaking right before I should say Matt from AWS and we both violently agreed that five years from now we'll have more engineers working for our company than we do today. More software developers working for our companies. We will create far more. They will be more efficient."
「我上台前正好是 AWS 的 Matt，我们俩『强烈一致』：五年后我们两家公司的工程师都会比今天更多，软件开发者更多。我们会创造多得多的东西，他们的效率也会更高。」
Mike Cannon-Brookes · 20VC: Atlassian CEO on Why Everything is Overvalued

Aaron Levie (Box)——软件工程是「受保护」的职业类别。

"This is why I always laugh when people say, you don't need to be an engineer, don't do computer science. I actually think like, that is like still one of the most protected job categories, because things are only getting more technical."
「所以每当有人说『不用当工程师了、别学计算机了』我都会笑。我其实认为那仍是最受保护的职业类别之一，因为一切只会变得更技术化。」
Aaron Levie · Every Agent Needs a Box (Aaron Levie)

Andrej Karpathy——软件的稀缺与昂贵才是需求受限的原因，成本下降触发杰文斯悖论。

"So if the barrier comes down, then actually you have the Jevons paradox, which is like, you know, actually the demand for software actually goes up. It's cheaper and there's more powerful."
「所以如果门槛降下来，你实际上会遇到杰文斯悖论：对软件的需求反而上升——它更便宜也更强大了。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch

Steve Huffman (Reddit)——「我们是建设公司」，瓶颈已转移到代码审查与部署。

"So let's say AI makes our engineers 50, 100% or even 10x more productive. We'll just build more stuff, like do more with more. It's not do the same amount with less."
「假设 AI 让我们的工程师效率提升 50%、100% 甚至 10 倍——我们只会造更多东西：用更多做更多，而不是用更少做同样多。」
Steve Huffman · How Reddit Went From $12M to $2.2B

Brad Lightcap (OpenAI)——Codex 的实证：成本趋零时需求大涨。

"The thing we're seeing in reality with tools like Codex and other things is actually when you reduce the cost of something to zero, the demand for it goes up significantly."
「我们在 Codex 这类工具上实际看到的是：当你把某样东西的成本降到接近零，对它的需求会显著上升。」
Brad Lightcap · Uncapped #46: Brad Lightcap

持中/条件立场

应用层与创业

前沿实验室将吞噬应用层 · Frontier labs will eat the application layer

正方 3 · 反方 11 · 持中 3

正方

Brendan Foody (Mercor)——「模型即产品」，软件层防御性极难建立。

"The application layer companies' businesses are not far removed from the foundation model companies' businesses. It is not a far leap for Claude Cowork to add capabilities across medical and legal. Obviously, they did it with software engineering and can do that across finance. And so I feel like building defensibility in the software layer on top of the models is going to be incredibly difficult."
「应用层公司的生意离基础模型公司的生意并不远。Claude Cowork 补上医疗和法律能力并不是多大的跨越——他们在软件工程上已经做到了，金融也一样能做。所以我觉得，在模型之上的软件层建立防御性将会极其困难。」
Brendan Foody · 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility

反方

Sam Altman——OpenAI 宁做永远低毛利的基础设施。

"Like I would like us to be an infrastructure provider. I'd be happy for us to be like a forever low margin as long as we can be huge and growing fast business."
「我希望我们做基础设施提供方。只要能做成巨大且快速增长的生意，我很乐意让它永远是低毛利业务。」
Sam Altman · Sam Altman in conversation with Patrick Collison

Sherwin Wu (OpenAI)——死掉的创业公司死于产品无共鸣，而非被实验室碾压。

"Every startup that I've seen that has kind of fizzled out is not because OpenAI or Big Lab or Google or something has come to squash them. It's because they built something and it really didn't resonate with the customers. Whereas the ones that take off, even in very competitive spaces like coding, like Cursor is huge at this point. And it's because they built something that people really love."
「我见过的每一家没做起来的创业公司，都不是因为 OpenAI、大实验室或 Google 来碾压它，而是因为做的东西没有引起客户共鸣。而起飞的那些，即便在编码这种最卷的赛道——Cursor 现在已经很大了——是因为他们做出了人们真正喜爱的东西。」
Sherwin Wu · "Engineers are becoming sorcerers" (Sherwin Wu)

Tuhin Srivastava (Baseten)——独占用户信号+工作流深度集成是实验室无法复制的护城河。

"To the extent that that is encoded in a model, I think a lot of their business will be at risk, but to the extent that it is encoded in workflows, that is where they will be able to develop modes [moats]."
「如果那种价值被编码进模型里，他们的很多生意会有风险；但如果它被编码进工作流里，那正是他们能建立护城河的地方。」
Tuhin Srivastava · Baseten CEO Tuhin Srivastava on the AI Inference Cloud

Gabe Pereyra (Harvey)——垂直层的价值实验室结构性拿不到：最有价值的法律数据因保密义务永远进不了通用模型，律所因利益冲突必须多模型——通用法律智能既不是护城河也不是价值所在。

"So imagine you're using only Anthropics models as a law firm and you want to represent OpenAI. OpenAI is not going to let you send their sensitive legal data to Anthropx models. And so as a law firm, you need to support all the different models if you want to represent OpenAI and Google and Microsoft and all these providers."
「想象一下，你作为一家律师事务所，只使用 Anthropics 的模型，并希望代表 OpenAI。OpenAI 不会允许你将他们的敏感法律数据发送给 Anthropx 模型。因此，作为一家律师事务所，如果希望代表 OpenAI、Google 和 Microsoft 等所有这些供应商，你需要支持所有不同的模型。」
Gabe Pereyra · Harvey Co-Founder Gabe Pereyra on the Token Pricing Reckoning

Scott Wu (Cognition)——拒绝「为时已晚/终将被巨头收编」的虚无主义：初创靠在特定未来愿景上插旗+更在乎取胜，引 Daniel Ek 对 Apple Music 的回答自况。

"He was like, you know, I can give you all the other reasons, but the truth is we're just going to care way more about music than they are. And I think for us, it's true. We are just going to care so much more about what does it look like to build software end-to-end at Goldman Sachs or at Mercedes-Benz or something like that."
「他当时说，您知道，我可以给您所有其他原因，但事实是，我们对音乐的关心程度远高于他们。我觉得对我们来说，这是真的。我们将更加关注在高盛或梅赛德斯 - 奔驰等地构建端到端软件的样子。」
Scott Wu · Scott Wu, Cognition

持中/条件立场

AI 时代应用护城河已死 · AI app moats are dead — wrappers commoditized, only speed matters

正方 5 · 反方 7 · 持中 2

正方

Ben Horowitz (a16z)——世界是肥尾的：薄封装不够，做深任何领域最终都「只是必需」专有数据与专用模型。

"It turns out that the universe is long-tailed, is fat-tailed, and humans are very fat-tailed in terms of human behavior, human conversation, and so forth. So to get to the real meaning of it and to get to the kind of essence of the problem, you know, in any domain, it turns out to be, I think, more complex than we thought."
「事实证明宇宙是长尾的、肥尾的，人类行为和对话尤其肥尾。所以要触及问题的真正含义和本质，在任何领域，事情都比我们想的更复杂。」
Ben Horowitz · Ben Horowitz: Why Hesitation is a CEO's Worst Enemy

Evan Spiegel (Snap)——编码不再是瓶颈，护城河转向独有资产与构想速度。

"One of our team members was like, okay, let me just see what I can vibe code and I'll get back to you. And two hours later, we were using the service together."
「我们一个团队成员说：好，让我看看能 vibe code 出什么来，回头找你。两小时后，我们已经在一起用这个新服务了。」
Evan Spiegel · Augmenting Reality, Both Online and Off (Evan Spiegel)

反方

Martin Casado (a16z)——「GPT wrapper 根本不是一个东西」。

"I gotta say, GPT wrapper was this derogatory term. I think we've kind of come to the conclusion that's not even a thing. When someone writes software on whatever the cloud, you don't call it a cloud wrapper."
「我得说，『GPT wrapper』曾是个贬义词。我们现在的结论是：这根本不是一个东西。有人在云上写软件时，你不会管它叫 cloud wrapper。」
Martin Casado · The State of AI: Growth, Fragmentation, and the Next Wave

Garry Tan (YC)——除非你信 AGI 清零一切，否则标准规则照旧。

"I think short of believing that AGI will happen and or ASI will happen ... I think all the standard rules still apply. ... It's like, you know what? This is just really, really good software."
「除非你相信 AGI 或 ASI 会降临（一切清零），否则我认为所有标准规则照旧适用。……说白了：这只是非常、非常好的软件。」
Garry Tan · Running Y Combinator Like a Founder (Garry Tan)

Jacob Lauritzen (Legora CTO)——随手可抄的是表面的 10%，「另外 90%」才是壁垒。

"It's very quick to get to the 90% where it looks the same and in like 80% of the cases, it works similarly. It's the other 90% that are difficult. You know, it's like ensuring all the edge cases work and all the unhappy paths and all the audit locking and all the RBAC and all the weird scenarios that you end up at at a certain scale."
「很快就能做到看起来一样的那 90%，而且 80% 的情况下用起来也差不多。难的是『另外那 90%』：保证所有边缘情况、所有异常路径、所有审计锁定、所有 RBAC，以及到了一定规模后你会遇到的所有诡异场景都能工作。」
Jacob Lauritzen · 20Product: Inside Legora's Tech Stack

Gavin Baker (Atreides)——给出「wrapper 何时不再是 wrapper」的可操作判据：有了推理云+路由器，用自有领域数据 RL 微调开源模型之后，防御性大增。

"Well, now with open source, you've generated some data. It's unique to your use case, whatever your vertical you're going after as a wrapper is."
「而现在有了开源：你已经生成了一些数据，它是你的用例独有的——不管你作为 wrapper 主攻的是哪个垂类。」
Gavin Baker · Gavin Baker - AI Market Jitters

持中/条件立场

AI 将杀死 SaaS 在位者与记录系统 · AI kills SaaS incumbents / systems of record

正方 4 · 反方 3 · 持中 3

正方

Jonathan Siddharth (Turing)——已知形态的 SaaS 已终结。

"All knowledge work is going to be automated. It's only a matter of time. ... SaaS, as we know it, I think is over. I think it's completely over."
「所有知识工作都将被自动化，只是时间问题。……我们所熟知的那种 SaaS，我认为已经终结了。彻底终结了。」
Jonathan Siddharth · 20VC: Scale, Surge, Turing, Mercor — Who Wins & Who Loses

Sebastian Siemiatkowski (Klarna)——转换成本消失后，「办公室排球和免费午餐的日子」结束。

"You create this service and there's a huge switching cost so your customers can't really switch that easily and hence you create this money printing machine and then life is sweet. Then you build these campuses and you play volleyball in your office and you go and get free lunches and you live off the spoils of this money printing machine in your basement. And that's not how normal business work."
「你做出一个服务，转换成本巨大，客户没法轻易走掉，于是你有了一台印钞机，日子甜得很。然后你盖园区、在办公室打排球、吃免费午餐，靠地下室里这台印钞机的战利品过活。但正常生意不是这样的。」
Sebastian Siemiatkowski · 20VC: SaaS is Dead — Why Systems of Record Will Die (Klarna)

Gil Feig (Merge)——agent 直接走 API/CLI，UI 退居次要。

"You never need to go to Salesforce. You can have your agent. It can go sign up through the API or through a CLI tool, like a command line tool. It can create accounts."
「你永远不需要打开 Salesforce。你可以让你的 agent 去做：通过 API 或 CLI 命令行工具注册、创建账户。」
Gil Feig · The Token-Maxxing Bill That Shocks Every CFO (Merge)

反方

Alfred Lin (Sequoia)——他 1997 年就信过一次「Amazon 杀死 Walmart」。

"The simple narrative, as we just talked about, is AI is going to kill SaaS. And I just don't think that's the case. And the reason why I know that that's not going to be the case is because I lived it. ... The simple narrative I made was that e-commerce was going to destroy brick and mortar, that Amazon was going to kill Walmart. ... And that just didn't happen. Walmart is 20 times larger today than it was in 1997."
「那个简单叙事——AI 将杀死 SaaS——我就是不信。我知道它不会成真，因为我亲历过：（1997 年）我给自己讲的简单叙事是电商将摧毁实体零售、Amazon 将杀死 Walmart。……结果并没有发生。今天的 Walmart 比 1997 年大 20 倍。」
Alfred Lin · Alfred Lin - The Future According to Sequoia

Sarah Guo & Elad Gil——SaaSpocalypse 是把 5 人初创的行为投射到财富 100 强。

"Markets are melting down about the end of software. ... is SAS [SaaS] actually dying or people just projecting five person startup behavior onto the fortune 100."
「市场正为『软件的终结』恐慌。今天我和 Elad 要问的是：SaaS 真的在死，还是人们只是把五人初创公司的行为投射到了财富 100 强身上？」
Sarah Guo · From SaaS to AI-First (Elad Gil & Sarah Guo)

Anish Acharya (a16z)——在位者还在涨价，这是产品市场契合的证据。

"Price is a measure of product market fit, right? And if you have enormous competitive pressure, you are not raising prices, you're typically cutting prices."
「价格是产品市场契合度的度量。如果你面临巨大的竞争压力，你不会涨价，通常只会降价。（而在位 SaaS 正在涨价。）」
Anish Acharya · Anish Acharya: Is SaaS Dead in a World of AI

持中/条件立场

应用层公司应自研/微调自己的模型 · App companies should train their own models

正方 4 · 反方 5 · 持中 2

正方

Edwin Chen (Surge AI)——只依赖前沿实验室，你优化的就不是你想优化的东西。

"So I definitely think that eventually every company should be training their own models. And that's because these models will be so important to the world and you want to eventually deploy them to like 99.999% of use cases, right? And if you simply rely on models from the Frontier Labs, what you're optimizing for may not be what you're optimizing for."
「我确定地认为：最终每家公司都应该训练自己的模型。因为这些模型对世界太重要了，你最终要把它们部署到 99.999% 的用例上。如果你只依赖前沿实验室的模型，你在优化的东西可能根本不是你想优化的东西。」
Edwin Chen · Ep 80: Edwin Chen on Why Frontier Labs Need Their Own Thesis

Will Brown & Johannes Hagemann (Prime Intellect)——ChatGPT 与 Claude Code 的优势正是产品-模型闭环，创业公司必须拿到这个能力。

"The general idea is to give more companies the actual ability and advantage that's currently only the big labs have in a sense of like this product model optimization loop ... that's the kind of reason why like a ChatGPT was created by OpenAI or like a Claude Code was created by Anthropic. They actually have the capabilities to optimize models for their specific scaffolds."
「大方向是把目前只有大实验室才有的能力和优势——产品-模型优化闭环——交给更多公司。ChatGPT 之于 OpenAI、Claude Code 之于 Anthropic，正是因为他们有能力针对自己的脚手架去优化模型。」
Will Brown & Johannes Hagemann · Building the GitHub for RL Environments (Prime Intellect)

反方

Shensi Ding & Gil Feig (Merge)——自训模型是过度工程。

"One hot take that I have is that I've noticed a lot of companies over-engineering their ML usage. They'll build their own custom models. They'll try to train their own models when really you could probably just use the generic model and then focus more on making your product better."
「我的一个尖锐观点：我注意到很多公司在过度工程化他们的 ML。他们要造自定义模型、要自训模型——而实际上你多半用通用模型就够了，然后把精力放在把产品做得更好上。」
Shensi Ding & Gil Feig · The Token-Maxxing Bill That Shocks Every CFO (Merge)

持中/条件立场

自主智能体今天已可投产 · Autonomous agents are production-ready today

正方 2 · 反方 5 · 持中 2

正方

Carlos García Ottati (Kavak)——宁可忍受一年增长停滞也不搞混合过渡。

"When you're doing a transition of the shift, you need to make sure that everybody understands that we're gonna go streaming or die."
「做这种范式转移时，你必须确保所有人都明白：我们要么全面转向流媒体，要么死。」
Carlos García Ottati · From Copilots to Agents: Rebuilding the Company Around AI (Kavak)

Steijn Pelle & Frederik Branken (Lassie)——给出可投产的具体阈值：做到 95%+ 就该卖，现实业务完全接受人管尾部边角案例。

"We want to get to a sufficient level of automation across any product before we sell it. So for us, I think that's like 95 plus, say, but not necessarily 100."
「在卖任何一个产品之前，我们都想达到足够的自动化水平。对我们来说大概是 95% 以上，但不一定要 100%。」
Lassie 联合创始人 · AI for America's Small Businesses (Lassie)

反方

Ben Mann (Anthropic 联合创始人)——能力已在，但安全不够，所以 Anthropic 自己不产品化全自动模式。（该篇逐字稿为中文版）

「我们未能部署基于计算机使用的消费者或最终用户级别应用的主要原因是安全问题，我们觉得如果让 Claude 访问你的浏览器及其中的所有凭据，它可能会出错并采取一些不可逆转的操作……它有这个能力，但安全系数还不够高，不足以让我们自己投入生产。」
— Ben Mann · Will we have Superintelligence by 2028? (Ben Mann)

Harrison Chase (LangChain)——浏览器操作还不够好，短期关键能力是代码执行。

"Browser use, I think the models just aren't good enough right now from what we've seen."
「浏览器操作——就我们观察到的情况，模型现在还不够好。」
Harrison Chase · Context Engineering Our Way to Long-Horizon Agents

Walden Yan (Cognition)——swarm 实验一片混乱，单 agent 反而意外好用。

"We found a surprising success of like, don't do a swarm or anything. Just have one Devin, you know, does his own context management. Just let it keep running for a while and give us some crazy tasks."
「我们发现一个出人意料的成功配方：别搞什么 swarm。就用一个 Devin，让它自己管理上下文，让它一直跑下去，接一些疯狂的任务。」
Walden Yan · The Age of Async Agents (Cognition & OpenInspect)

持中/条件立场

MCP/开放协议是智能体集成的正确层 · MCP/open protocols are the right integration layer

正方 1 · 反方 3

正方

Sherwin Wu & Christina Cai (OpenAI Platform)——竞争对手事实上承认了开放标准的引力。

"MCP was kind of like the natural protocol that developers were already using to bring all the tools into their system."
「MCP 差不多是开发者已经在用的、把所有工具接进系统的自然协议。」
Sherwin Wu & Christina Cai · DevDay 2025 (OpenAI Platform)

反方

Bret Taylor (Sierra)——真代理需要远超 MCP 所能提供的上下文。

"I think these agents need a lot more context than what MCP affords."
「我认为这些 agent 需要的上下文，远多于 MCP 所能提供的。」
Bret Taylor · Bret Taylor of Sierra on AI agents

Simon Last & Sarah Sachs (Notion)——CLI 环境自带渐进披露且能「自举」，MCP 适合窄场景。

"Broadly speaking, I'm really bullish on CLIs. I'm still bullish on MCPs in a certain environment. I think MCPs are really great for when you want a narrow, lightweight agent."
「总体来说我非常看好 CLI。我在特定环境下仍看好 MCP——当你想要一个窄的、轻量的 agent 时，MCP 非常好用。」
Simon Last · Notion's Token Town

Eiso Kant (Poolside)——比 Notion 更彻底：不是「MCP 适合窄场景」，而是塞工具定义这条路本身是歧途——给模型一台装好二进制和代码库的虚拟机、让它自己写代码才是终局。

"They're increasingly no longer here. We're going to stuff 50 tools in the like system prompt to know here's a virtual machine with these binaries installed, this code base you can operate in here, a folder where you can write, you know, your memory if you want to. And the model is using code to do complex asks."
「它们越来越不再是『我们要往系统提示里塞 50 个工具』，而是『这里有一台虚拟机，装好了这些二进制、这个你可以操作的代码库、一个你想写记忆就能写的文件夹』。模型用代码去完成复杂请求。」
Eiso Kant · Inside the Model Factory (Poolside AI)

智能体开发平台将被商品化 · Agent-building platforms get commoditized

正方 1 · 反方 1

正方

Bret Taylor (Sierra)——通用 agent builder 会像 1995 年后的建站工具一样商品化。

"It just feels inevitably to be a commodity in my mind because you know, maybe making a website was hard in 1995, but today there's like a million ways to make a website. ... In practice, I think the same will happen with agent building. I think OpenAI will have a great tool."
「在我看来它不可避免会沦为商品：1995 年做网站也许很难，今天做网站有一百万种方法。……我认为 agent 搭建工具会重演同样的历史。OpenAI 会有很棒的工具，（所有基础模型公司都会有，开源也会有。）」
Bret Taylor · Uncapped #42: Bret Taylor

反方

Harrison Chase (LangChain)——harness 与模型家族绑定、自建比想象难得多，长期价值沉淀到 harness 提供方。

"The issue is the models weren't really good enough and the scaffolding and harnesses around them weren't really good enough. And I think the models got better. We learned more about what makes a good harness over the past few years. And now they start to like really, really work."
「问题在于过去模型不够好、外围的脚手架和 harness 也不够好。后来模型变好了，我们这几年也学到了什么才是好的 harness。现在它们真的、真的开始能用了。」
Harrison Chase · Context Engineering Our Way to Long-Horizon Agents

万能应用/聊天代理将取代专用应用 · Everything apps & agents replace purpose-built apps

正方 1 · 反方 1

新对撞：AI 通用界面之下，专用应用是待吸收的「孤岛」还是不可替代的富体验终点？分歧核心在 agent 的角色——吞噬者还是路由器。

正方

Ryo Lu (Cursor 设计负责人)——专用应用是「自私的」：按文件格式和工作流把人隔成孤岛；简单核心概念+AI 通用界面的万能应用会克服这种局限。

"The purposeful thing is no longer simple. And all of these purposeful apps, they're kind of selfish. They are siloing people, siloing workflows for our file formats, creating islands."
「有目的性的东西不再简单。所有这些有目的性的应用程序，它们有点自私。它们正在隔离用户，隔离我们的文件格式的工作流程，创建孤岛。」
Ryo Lu · Ryo Lu (Cursor): AI Turns Designers to Developers

反方

Josh Elman (a16z)——「应用之死」问错了方向：agent 的目标是更快把你送进配置好的富应用；可视化、探索、沉浸式体验都不会消失（Zork 不会取代巫师 3）。

"I mean, a lot of people have used this moment to go, is this the death of apps? Because you'll just get to everything through a command line. And I'm very much a believer that Apps and experiences and being able to visualize and explore information or spend time in an app consuming content or interact with things. None of that's going to go away."
「我是说，很多人利用这个时刻来问，这是不是应用的死亡？因为你只需通过命令行访问一切。我非常相信应用和体验，能够可视化和探索信息，或花时间在应用中消费内容或与事物互动。这些都不会消失。」
Josh Elman · What's Next for Consumer AI? | Josh Elman Joins a16z

政策与社会

前沿 AI 应先监管/暂停再部署 · Regulate/pause frontier AI before deployment

正方 2 · 反方 9 · 持中 1

正方

Sebastian Mallaby（转述并认同 Demis Hassabis 的转变）——单个实验室的安全不能让世界更安全。

"It's almost pointless for one lab to pursue safety by itself, because if one lab is safe and then the other ones aren't, it doesn't make the world safer. So he really has shifted to seeing this as a collective action problem that only a government can solve."
「单个实验室独自追求安全几乎没有意义——如果一家安全而其他家不安全，世界并不会更安全。所以他（Hassabis）已真正转向把这看作只有政府才能解决的集体行动问题。」
Sebastian Mallaby · Ep 88: Unpacking DeepMind's Quest for SuperIntelligence

Lukas Petersson & Axel Backlund (Andon Labs)——一旦意识到 AI 不只是聊天机器人，暂停就成为可行选项。

"If you think that AIs are just chatbots, then it sounds ridiculous to advocate for a pause of AI."
「如果你认为 AI 只是聊天机器人，那么呼吁暂停 AI 听起来就很荒谬。（但一旦看到模型可能接管并执行危险任务，暂停就是可行的必要选项。）」
Lukas Petersson · Reality: The Final Eval

反方

Marc Andreessen——过早监管 AI 会是「非常非常非常大的错误」。

"I think it would be a very, very, very big mistake to do that in AI."
「（为对冲风险而提前监管、切断收益——）我认为在 AI 上这么做将是一个非常、非常、非常大的错误。」
Marc Andreessen · Marc Andreessen: The Future of Venture Capital

Sam Altman——迭代部署+民主化访问是最有争议也最正确的决定。

"The most controversial decision we made in the history of OpenAI ... was what we now call iterative deployment. But a lot of the thinking at the time that we released ChatGPT was this was insanely dangerous to do. ... And I thought then and I believe now that it is extremely important that we avoid that kind of power concentration."
「OpenAI 历史上最有争议的决定……就是我们现在所说的迭代部署。发布 ChatGPT 时，很多人认为这样做危险到疯狂。……但我当时认为、现在仍然相信：避免那种权力集中极端重要。」
Sam Altman · Sam Altman in conversation with Patrick Collison

Trae Stephens (Founders Fund / Anduril)——政策本应滞后于技术；自主武器的法律正是在国防界内部演化出来的。

"A lot of the existing law in the books around autonomy and artificial intelligence actually started in the defense community."
「现行法律里关于自主系统和人工智能的很多条文，实际上就起源于国防界（先有数十年实践、再有成文规则）。」
Trae Stephens · Uncapped #35: Trae Stephens

Patrick & John Collison (Stripe 年度信)——应削减 EU AI Act 式负担、改革否决制监管。

"Entrepreneurs in Europe can boost tepid economies with new tools, but only if well-intentioned yet counterproductive burdens such as the EU AI Act are curtailed."
「欧洲的创业者可以用新工具提振疲软的经济，但前提是削减像 EU AI Act 这样善意却适得其反的负担。」
Patrick & John Collison · Stripe's 2025 annual letter

Mark Chen & Nick Turley (OpenAI)——「在自由一侧犯错」，把控制的苦功夫做足。

"But I've always felt we need to err on the side of freedom and we need to do the hard work."
「但我始终觉得，我们应该在自由的一侧犯错，然后把（控制与安全的）苦功夫做足。」
Mark Chen & Nick Turley · Inside ChatGPT (Mark Chen & Nick Turley)

David Sacks（All-In）——把「先监管」的诉求拆成动机链：virtue signaling、CYA 甩锅、Dario 要的「AI 版 FDA」、群体信仰，最后才是护城河。

"Number three is Red Capture. Dario wants an FDA for AI. He's not going to stop until he gets it. And in order to get it, you have to keep spiking the cortisol and panic people."
「第三条是监管俘获。Dario 想要一个 AI 版的 FDA。他不拿到手不会罢休。而要拿到它，你就得不断抬高皮质醇、让人们恐慌。」
David Sacks · Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs

Steven Sinofsky——顺序错了：在理解监管对象之前就立法是本末倒置，既有法律框架已覆盖绝大多数 AI 风险，责任应始终落在使用工具的持照专业人士身上。

"It's not the AI's medical license. Completely unambiguous. Now if she went and hurt the patient, it's super tragic. But she still hurt the patient and it's her accountability no matter who wrote the note."
「行医执照不是 AI 的。这毫不含糊。如果她伤害了病人，那非常不幸——但伤害病人的仍是她，无论那份病历是谁写的，问责都在她。」
Steven Sinofsky · Steven Sinofsky: AI Doesn't Need New Rules Yet

持中/条件立场

对华 AI 鹰派脱钩是正确的美国战略 · Hawkish decoupling is the right US strategy on Chinese AI

正方 4 · 反方 7 · 持中 1

正方

Trae Stephens (Founders Fund / Anduril)——直斥 Jensen Huang（本节反方）的「赢下中国市场」论：特斯拉/比亚迪是前车之鉴，根本不存在美国公司「拥有」中国市场这回事。

"I mean I think that's probably the thesis that Tesla had as well before they realized that BYD was just going to Copy all their best stuff and cut the price. And I think that's the much more likely outcome. It's like the super naive ramblings of a monopoly hungry American who ignores geopolitical reality."
「我的意思是，我认为这可能是特斯拉之前的论点，但后来他们意识到比亚迪会抄袭他们所有最好的东西并降价。我认为这更可能是结果。这就像一个贪图垄断的美国人的幼稚闲聊，他忽视了地缘政治现实。」
Trae Stephens · Uncapped #53 | Trae and Delian from Founders Fund

Péter Fankhauser (ANYbotics CEO)——把命题从模型层挪到硬件层：关键基础设施里的机器人满载摄像头传感器，供应链必须本国或盟国化。

"You don't want to have 15 cameras in your critical infrastructure that somebody else controls."
「你不会想让你的关键基础设施里有 15 个由别人控制的摄像头。」
Péter Fankhauser · The $1/Hour Worker: Four Robotics CEOs on Humanoids

反方

Bill Gurley / Brad Gerstner (BG2)——「如果这是对国家最重要的关系，为什么要选择了解得更少？」

"If it is the most consequential relationship for our country, why you'd want to know less, right?"
「如果这是对我们国家最重要的一段关系，你为什么会想要了解得更少呢？」
Bill Gurley / Brad Gerstner · China, China, China (BG2)

Jensen Huang (NVIDIA)——中国只落后「几纳秒」，正确策略是去竞争。

"Yeah, they're nanoseconds behind us. And so we've got to go compete. We've got to go compete."
「对，他们只落后我们几纳秒。所以我们必须去竞争。我们必须去竞争。」
Jensen Huang · NVIDIA: OpenAI, Future of Compute, and the American Dream (BG2)

Tuhin Srivastava (Baseten)——网络隔离后模型无法神奇外传数据，未见木马证据。

"If I network bound these models that they're not magically, you know, going to be able to cross those network boundaries and to data zitter. And, you know, I don't, and we, I've never seen any real evidence, except from some very early models that I think people picked up on very quickly, that there is some agenda or bias built into these."
「如果我限制了这些模型的网络访问，它们是不可能凭空突破这些网络限制去获取数据的。而且，我从未见过任何确凿的证据——除了早期一些模型被人们迅速察觉到的情况外——表明这些模型中内置了某种议程或偏见。」
Tuhin Srivastava · Baseten CEO Tuhin Srivastava on the AI Inference Cloud

Steven Sinofsky——给出历史对照：出口管制是创新领导权之战里政府仅有的钝器，重演 70-80 年代底特律指控日本车「倾销」的剧本，最终只逼出本地建厂与变通方案。

"And Detroit just convinced the administration that they were selling these cars below their cost, that they were dumping them. Does this sound familiar? ... So what did Japan do? They actually built all their factories. ... They still lost. Detroit doesn't even make cars anymore."
「底特律说服了政府：日本人在以低于成本的价格卖车，是在倾销。听着耳熟吗？……那日本怎么做的？他们干脆把工厂全建了过来。……底特律还是输了，如今它连车都不造了。」
Steven Sinofsky · Steven Sinofsky: AI Doesn't Need New Rules Yet

持中/条件立场

AI 伴侣有益社会、对人类是净收益 · AI companions are net good for humans

正方 3 · 反方 2

正方

Anish Acharya (a16z)——任何缓解孤独的进展都是人类进步。

"I think that there's a deep loneliness and any progress we make towards addressing the loneliness is human progress and is very pro-social."
「我认为（社会上）存在一种深层的孤独，而我们在缓解孤独上取得的任何进展，都是人类的进步，是非常有益社会的。」
Anish Acharya · Why Creativity Will Matter More Than Code (Rose & Acharya)

Eugenia Kuyda (Replika/Wabi)——AI 今天不会说，但会听。

"And I think that realization that maybe an AI can't talk well today, but it could listen, that that could be a groundbreaking thing for millions in the world."
「我意识到：也许今天的 AI 还说不好话，但它会倾听——对世界上数百万人来说，这可能就是开创性的。」
Eugenia Kuyda · Seeing The Future: from AI Companions to Personal Superintelligence

Dara Ladjevardian (Delphi)——数字分身提供原本不存在的访问权限。

"So the idea isn't replacing human connection, but providing access where previously there was none in a way that is more temporally convenient for the end user."
「所以本意不是替代人际连接，而是在原本完全没有（专家）接触机会的地方提供访问，而且在时间上对用户更便利。」
Dara Ladjevardian · Delphi's Dara Ladjevardian: How AI Digital Minds Can Scale You

反方

Trae Stephens (Founders Fund)——零摩擦认同机器会放大在线约会已经造成的社会问题。

"We've created this entire social demographic of incels that historically, if you look back over the course of human civilization, have been the powder keg for implosions of societies. You don't want to have a bunch of young 20-something single men that are underemployed running around in your economy. It doesn't bode well. And you add AI on top of that now."
「我们已经造出了 incel 这样一整个社会群体——纵观人类文明史，这类群体一直是社会内爆的火药桶。你不会想让一堆就业不足的二十多岁单身男性在经济体里游荡。这不是好兆头。而现在你还要在这之上再加上 AI。」
Trae Stephens · Uncapped #35: Trae Stephens

Reid Hoffman——把 AI 陪伴当朋友是危险的范畴错误。

"Friendship is a joint relationship. It's not a, oh, you're just loyal to me or you just do things for me. Oh, this person does things for me. Well, there's a lot of people who do things for you. Your bus driver does things for you ... but that doesn't mean that you're friends."
「友谊是一种双向的关系。不是『你对我忠诚』或『你为我做事』。为你做事的人多了——公交司机也为你做事——但那不意味着你们是朋友。」
Reid Hoffman · Reid Hoffman on AI, Consciousness, and the Future of Being Human

AI 生成内容注定沦为垃圾（slop） · AI-generated content is doomed to be slop

正方 2 · 反方 4

正方

Ethan Smith (Graphite)——严格对照研究显示无人参与的全自动 AI 内容排名无效；平台终将像 Google 打击 spam 一样打击它。

"Anything can be optimized, but if you're spamming it, they'll see that, and they'll have a whole team looking at that, and then they'll change your algorithm to prevent you from doing that."
「任何系统都可以被优化，但如果你在向它倾倒垃圾，他们会看见——会有一整个团队盯着，然后改算法阻止你。」
Ethan Smith · The ultimate guide to AEO (Ethan Smith)

反方

James Cadwallader (Profound)——「AI 写的就是垃圾」是转移视线。

"I think slop is a red herring that is going to be quite quickly disproven. ... I do think that this idea of if it's written by AI equals slop is a stupid one."
「我认为 slop 是个很快会被证伪的烟雾弹。……『AI 写的就等于垃圾』这个想法，我认为是愚蠢的。（NYT 盲测里 53% 的读者更喜欢 AI 写的那篇。）」
James Cadwallader · From SEO to Agent-Led Growth (Profound)

Ryan Roslansky (LinkedIn)——验证身份+给真人更多平台权限可以守住真实性。

"We introduced what's called verified identity, which is basically either through ... your work email address ... or through your driver's license or through your passport, you can actually verify who you are."
「我们推出了『验证身份』：通过你的工作邮箱、驾照或护照，你可以真正证明你是谁（验证过的真人在平台上获得更多权限）。」
Ryan Roslansky · LinkedIn's CEO on How AI Changes Your Job

Matt Tancik (Luma)——slop 与否的变量是背后指挥的人，不是模型：同一批工具既产出糟糕生成物、也产出惊艳作品。

"You've seen AI generations that are bad, right? I think we've all seen them that are bad. I can guarantee the same tools that made those bad generations, they've also made generations that you would think are amazing."
「你见过糟糕的 AI 生成的内容，对吧？我想我们都见过一些糟糕的内容。我可以保证，造成那些糟糕生成的同样工具，也创造出你认为非常惊艳的作品。」
Matt Tancik · Building AI for Creators | Luma & Phota Labs

语音是下一个主要的 AI 交互界面 · Voice is the next primary AI interface

正方 1 · 反方 1

正方

反方

Eugenia Kuyda (Replika/Wabi)——语音无法做发现与主动性，她永远不会做无屏设备。

"I love screens. I think there's no way with voice to solve for discovery, for proactivity."
「我爱屏幕。我认为语音没有任何办法解决发现（discovery）和主动性（proactivity）的问题。」
Eugenia Kuyda · Seeing The Future: from AI Companions to Personal Superintelligence

自我立场转变

Brendan Foody (Mercor)：从「替代快而痛」到「弹性无限、总就业上升」

"I think displacement in a lot of roles is going to happen very quickly, and it's going to be very painful and a large political problem."
「我认为很多岗位的替代会来得非常快，而且会非常痛苦，成为重大政治问题。」
Brendan Foody · From Job Displacement to AI Trainers (Brendan Foody)
"But I think that the key thing is that everyone underestimates the elasticity for demand and increased productivity in the economy. Ultimately, over the last 250 years, we've increased productivity by 25x, equivalent to automating about 96% of someone's job."
「但我认为关键在于：所有人都低估了经济中需求与生产力提升的弹性。过去 250 年我们把生产力提高了 25 倍——相当于自动化掉了一个人约 96% 的工作。」
Brendan Foody · 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility

构建方式：/lens 全库扫描 → 27 组候选命题 → 编辑闸门（同一命题、双方真信念）→ 25 组收录；引文逐条 grep 自 interviews/ 逐字稿。下次更新：新增访谈进入语料库后由 lens sweep 增量补充。

继续读
线索
阅读线索：agent 不到 9 nines 可靠，那应该卖给企业还是消费者
线索
阅读线索：如果 coding 是 commodity，AI startup 的护城河长在哪儿
线索
阅读线索：推理成本一年跌 10×，AI app 该按什么定价
线索
摘报
线索
摘报
线索
摘报