话语

Misha LaskinReflectionAI · 联合创始人、CEO

强化学习消耗了 10,000 多张 GB300 长达四周的时间。因此，实际上花在强化学习上的 FLOPs 比预训练还要多。强化学习任务要复杂得多，因为你既需要利用各类智能体沙箱进行大规模推理，又需要同时进行训练。

reinforcement learning was a little over 10,000 GB300s for four weeks. So, actually there are more flops spent on reinforcement learning. Reinforcement learning jobs are, you know, more complex in the sense that you do both a lot of inference at scale with all sorts of sandboxes for agents and training.

Misha LaskinReflectionAI · 联合创始人、CEO

事实证明，为了让强化学习在大规模下取得良好效果，你确实需要自己对模型进行预训练。你不得不这么做，这两者之间的耦合太紧密了，你必须两者兼顾。

it turned out you actually do need to pre-train your model in order to make reinforcement learning work very well at scale. So you actually kind of do need to do that the things are just so tightly coupled that uh you you do need to do both.

Pat GelsingerPlayground Global · 普通合伙人、Intel 前 CEO

然后我还必须把它放进机架级解决方案里，因为如今已经没有什么单独的芯片了，整个系统就是一个机架。所以这需要 9 个月——我花 3 个月把它设计出来，但要等上整整 9 个月才能实际用上它。

then I have to put it into a rack scale solution because nothing's a chip anymore. It's a rack, you know. So we're 9 months, you know, it took me 3 months to design it, but it's nine months until I can actually start to use it

Pat GelsingerPlayground Global · 普通合伙人、Intel 前 CEO

因此我能在 3 个月内完成设计，但我无法在 9 个月内真正实现规模化硅片制造。这很糟糕，对吧？所以如果我们要持续这种创新，就必须真正看到这个瓶颈得到改善。

so I can design the thing in three months but I can't actually get it into real silicon at scale for 9 months. Okay, that sucks, right? So, we have to really see that bottleneck improve if we're going to have this innovation

Greg PstruchaSentry · Staff 软件工程师

look at my GitHub reviews the feedback from the bots the feedback from my peers and try to codify which ones of those are actual high quality lint rules that are deterministic and will be able to capture issues that we otherwise reprompt all the time.

Greg PstruchaSentry · Staff 软件工程师

where the things that are writing code do not have any memory and are going to make the same repeated mistakes over and over again. And on the on the at the same time to write a new lter is very cheap and very easy. You can just have the agent do that

Gabriel MartinezG2i · 工程经理

Generation is cheap, but evaluation is expensive. If AI produces more than humans can possibly evaluate, it has not accelerated the team. It has created a bottleneck.

Gabriel MartinezG2i · 工程经理

The key point I want to highlight here is that we do not forget the ground truths we've always known. 60 years of software wisdom hasn't been invalidated by AI. Composable code, good patterns, and solid practices remain essential.

Dominik KundelOpenAI · Codex 开发者体验团队成员

That way you're still leveraging all of the work that we put into tweaking the prompts for the individual models, while still customizing it for your specific use case of the harness.

Dominik KundelOpenAI · Codex 开发者体验团队成员

and you can actually mark them as deferred loading, which means that rather than injecting them into the system prompt, it allows Codex to use a tool search tool, which we'll cover more tomorrow in the other session, but it allows it to find these tools through tool search rather than polluting the overall context window.

Anders SwansonOracle Database · 开发者布道师

redaction is very important as the first step because before I do embeddings, before I do enrichment, I want to strip out PII. I don't want secrets or irrelevant weird information slipping into the memory layer. That's a recipe for disaster. So, we do redaction first.

Anders SwansonOracle Database · 开发者布道师

And this this form of retra retrieval can fuse different search methods. So, I can take vector similarity search, I can take a full-text keyword search, I can take relational entities, I can do graph traversal, I can take other feedback signals like importance uh uh and usefulness of the memory to yield more comprehensive results than one specific query branch alone.

Pedro LopezAirbyte · 软件工程师

one important thing to keep in mind is to design those CLIs with agents in mind um some of the things that might be helpful for users when you're making a CLI are going to be like blockers for agents. Um, and we'll get into that a little bit more. Um, the first design decision here that we made was take JSON in and return JSON out.

Pedro LopezAirbyte · 软件工程师

So one thing that we decided to do um after trial and error is to keep the tool count small. So um we for the execution part of the connector where you're pulling that data, we only expose or really deal with two main tools.

Ali-Reza Adl-TabatabaiGitar / Sonar · Gitar 联合创始人

now being generated, more PRs than before uh thanks to AI and bigger PRs uh all of which can have defects and so your costs in your development life cycle are kind of shifting to the right. You have a lot more code to review. You can have potential for a lot more failures in production and and so your choices are kind of tough here.

Ali-Reza Adl-TabatabaiGitar / Sonar · Gitar 联合创始人

There's a ton of opportunity here in combining an agentic approach to validation with program analysis. So combining agents with things like u taint analysis, control flow analysis, data flow analysis, software composition analysis. So all the traditional approaches combined with agents. We think there's a ton of opportunity here and that together you can get much better precision, much better coverage

Felipe BlanesAmazon AGI Lab · 团队成员

So it's basically reliability is a zero to one, right? Or you trust or you don't. There is nothing like in the middle there.

Felipe BlanesAmazon AGI Lab · 团队成员

And the main learning here is transparent on product limitation increase trust more than higher benchmarks. That's a very interesting finding, right? Be transparent about what your product is capable of doing and in more important what's not capable of. That's how you earn customer trust.

Em ShreveAutomattic · AI Enablement 负责人

We're also wanting each person we send through um to be a multiplier uh within the company and extend our reach to the folks that we can't get to yet. Um, so so far this year we've ran seven of these cohorts of groups of about 45 people. Um, and we're aiming to get around 600 people through these programs

Em ShreveAutomattic · AI Enablement 负责人

Uh we also um organize these things called AI ride alongs and these are little workshops that we run um basically showing how you work, how you use a different tool. Um because what we found is seeing someone in your similar role or um in in a similar job to you, being able to see how they use it uh is a lot more um easy to grasp than just you know talking about the concepts.

Viren BaraiyaOrkes · CTO、联合创始人

the responsibility of a non-deterministic agent is to plan, is to plan what should happen next, not really to do things. Um and then Harness is the one who does actually execution. And And this is important for uh various reasons, right? One being that Harness is a deterministic piece of code

Viren BaraiyaOrkes · CTO、联合创始人

When you think about agentic harnesses, they are essentially late bound sagas. And what I mean by that is that they have a very finite and deterministic set of tools uh and the task that they can operate on. Instead of putting them together upfront, the agent is kind of proposing and building them at runtime.

Matt GarmanAmazon Web Services · CEO

人类并不总是那么在乎 S3 的 P99.9 长尾延迟，但智能体却极其在乎，并且会被这种延迟彻底阻塞挂起。

they don't always care about the the P999 S3 latency like agents do care and get blocked by that

Matt GarmanAmazon Web Services · CEO

这已经不再是简单的代码自动补全，而是真正的‘智能体优先’——智能体编写所有的代码，而你只是在管理并驱动一支智能体团队。

it's not code completion it really is agent first the agents write all of the code you're just managing a team of agents and and driving that

KeerthanaGoogle DeepMind · 机器人研究

依然处于 GPT-2 时代，原因在于：要达到 GPT-3 的水平，我们首先需要少样本学习能在大量不同任务上真正奏效；其次机器人学至今依然极度受制于环境迁移——如果一个模型只能在你那台具有特定配置的机器人上运行，它能算得上是通用大脑吗？

still GPT-2 and here's why. I think for GPT-3 we need firstly few shot learning to work really well for a lot of different tasks and secondly also with robotics weirdly is still very subject to cross environment, right? If it just works on your robot with your specific setup, is it really a generic brain?

KeerthanaGoogle DeepMind · 机器人研究

通过极少量的数据——仅用 200 个示例——你就能让它在许多截然不同的机器人身体上学会大量任务，因为你已经具备了对物理世界进行高度通用理解的基础底座，并懂得如何对其进行控制。

You can with very, very small number of examples, with 200 examples, you can learn a lot of different tasks on a lot of different bodies, because you already have the foundation to do a very generic understanding of the physical world, and also like how to control it

Sam AltmanOpenAI · CEO

将对齐仅仅视作一个工程问题无疑是错误的。任何声称我们已经解决了对齐科学问题的人，我认为都犯了极其危险的错误。我们需要在基础研究上取得更多进展，我相信我们也会做到，因为我们一直擅长取得这种研究突破。

it's certainly incorrect to talk about this only as an engineering problem. Anyone who says we have solved the science of alignment I believe is wrong in a very dangerous way. Like we need to make more research progress I assume we will. We have been great at making this research progress.

Sam AltmanOpenAI · CEO

OpenAI 拥有一项未被外界足够重视的布局，就是我们的自研芯片计划——它将在 2027 年上半年开始上线，并在未来几年大规模扩展。我认为它将为我们在全球能获取的每瓦能耗上带来巨大的推理成本优势，特别是考虑到我们在消费者、开发者和企业端看到的爆发式需求增长。

one thing that OpenAI has that doesn't get as much attention is our chip program, which will start to come online in the first half of 2027 and will scale a lot in future years. I think gives us a huge inference advantage for whatever watts we can find in the world and given the growth and demand we are seeing across consumer developer enterprise really everywhere.

Alex ZhangMIT · 博士生

我们所认为的语言模型——也就是我们向其发起查询的那个东西——未来很可能实际上是一个蜂群、一个脚手架，或者某种扩展性极佳的奇特外周框架设计，只不过用户根本感知不到底层细节。

what we think of as a language model like the thing that we query might actually be like a swarm or like a scaffold or like some weird harness design that scales very well, but the user just doesn't see it.

Alex ZhangMIT · 博士生

我这篇博文想要表达的核心是：我们应当重新思考外周框架（Harness）的角色，因为框架实际上能极大地提升模型的泛化能力以及有效利用的数据量。而且这种特性绝不仅限于递归语言模型。

what I'm trying to get at with this whole blog post is just that like we should rethink what the role of the harness is because harnesses can actually greatly increase the generalization capability of your model and the amount of data that it's given. And this is not exclusive to RLMs.

Timnit GebruDAIR · 研究者

我一直在发出警告：这种关于根本不存在的虚构超级智能机器及其所谓存亡危机的讨论，一直以来都只是为了实现监管俘获而精心设计的花招。

I have been warning that this kind of uh discussion about the fictional super intelligent machines that don't exist and the uh so-called existential risk that they pose has been a ploy for regulatory capture.

Timnit GebruDAIR · 研究者

这完全是特定公司的严重失职和重大过失，实际上完全可以追究其刑事责任。Lina Khan 已经梳理了大量依据现行法律就能直接向这些公司追责的途径。

This is gross incompetence and negligence by a particular company, which actually they could be held criminally liable for. Lina Khan, um, outlined a whole bunch of ways in which um, these companies can be currently held liable with current laws.

比尔·盖茨Bill Gates · Gates Foundation

不，核心在于我们过去一直说，一旦跨越这些门槛，我们就将动员全社会参与讨论，因为届时我们将创造出有史以来最危险的事物——这会让核武器都相形见绌。我们曾承诺跨过门槛就启动社会介入，可如今跨过了，全社会却一片死寂。

No, the key thing is we always said when we cross these thresholds we will engage all of society because we will have created the most dangerous thing ever. This makes nuclear weapons look like nothing. And we said we would engage and we crossed those thresholds and there was complete silence.

比尔·盖茨Bill Gates · Gates Foundation

每家 AI 公司——无论是 DeepMind 还是 OpenAI——都声称：‘无论这里发生什么，都不能仅仅由利润最大化驱动；我们必须拥有一份章程，确保在触及危险阈值时能够做出违背利润最大化的裁决。’可悲的是，这类机制只有在全世界仅有一家公司时才能起效。

every AI company, whether it's Deep Mind or Open AI, you know, they all say, "Okay, whatever goes on here, it can't just be driven by profit maximization. We have to have a charter that if we get to dangerous thresholds, we can exercise judgment that would be against profit maximization." Sadly, those mechanisms only work if there's only one company.

Sam AltmanOpenAI · CEO

在构建 AI 模型公司的人群中，其实存在两种思潮：一种人相信甚至期望世界上最终只会剩下一家超级公司；而像我们这样的人则相信，我们即将迎来初创公司数量的大文艺复兴，门槛将被彻底打平，让单人公司和小团队能够创造出不可思议的全新产品与服务。

there are really kind of two schools of thought among people building AI model companies. There are people that believe or hope that there's only going to be one company in the world. And then there are people like us that believe we are going to see a renaissance in the number of companies and just an absolute lowering of the bar to people, one-person companies, small-team companies creating incredible new products and services.

Sam AltmanOpenAI · CEO

从整体来看，现有软件在设计之初就根本不是为了与 AI 协同而生的。你可以自己使用它，或者让 AI 代表你操作它；但要真正实现紧密协作，尤其是多个人与多个人的 AI 全体协同工作时，我们现有的很多抽象概念就彻底崩溃了。

software was really not designed on the whole for collaborating with AI. And you can use it yourself or you can have AI use it on your behalf, but to really collaborate closely, especially when it's multiple people and multiple people's AIs all working together, like a lot of the current abstractions we have break down.

马斯克Elon Musk · xAI / Tesla / SpaceX

我认为能源发电的扩展是最大的事情之一。你必须扩展能源，也必须扩展芯片制造。我认为我们在软件和一切智力、数字化领域都处于领先。但从长远来看，我们需要能够产生足够的电力来与对手竞争——我认为从本质上讲中国是最大的竞争对手，而中国的发电量大约是美国的三倍。

Scaling energy I think is one of the biggest things. You've got to scale energy, you've got to scale chip production. I think we're winning on software, on anything intellectual or digital. But in the long run, we need to be able to generate enough power to compete with, I think essentially China is the biggest competitor, and China has about three times the electricity production of the United States.

黄仁勋Jensen Huang · NVIDIA · CEO

但我们现在正在建造的东西，实际上是超智工厂（SI Factory）。之所以称其为工厂，是因为你正在生产具有经济价值的东西，而不仅仅是存储数据。

But the things that we're building now, these are really SI factories, super intelligence factories. And the reason why it's a factory is because you're producing something of economic value. You're just not storing it.

Ari WeinsteinOpenAI

Computer Use 最强大、也是我认为最令人兴奋的地方在于，它让智能体能够做到你作为人类所能做到的一切事情。因为世界上所有的软件都是为人类设计的，而现在智能体可以使用完全相同的软件，你可以直接将这些工作委托给智能体。

the powerful thing about computer use and the reason why I think it's so uh exciting is because it makes it so that the agent can do anything you as a as a person can do. Um, because all the software in the world was designed for humans and now agents can use that same software uh and you can delegate to the agent.

Ari WeinsteinOpenAI

它实际上在编写并执行 JavaScript 代码，由计算机一次性执行代码来同时完成许多动作，这带来了巨大的速度提升和极强的能力跃升。

It's actually writing JavaScript code that it executes that that the computer executes to perform sometimes many actions at once, which is a great, you know, speed up and great capability.

苏姿丰Lisa Su · AMD · CEO

硬件、软件、系统与模型层之间的相互作用决定了，你对端到端全链路理解越深，构建出的系统就越优秀。这就是我们收购 World Labs 的真正原因：获得飞飞汇聚的世界级顶尖人才，并将其与 AMD 在硬件、软件和系统方面的完整能力紧密结合。

the interaction between, you know, the hardware and the software and the systems and the model layer are such that the more you understand end to end, the better system you're going to build. And so that's why we're acquiring World Labs, really is, uh, to have access to the world class talent that Feifei has brought together and really bring it together with AMD's complete capabilities around hardware, software and systems

李飞飞Fei-Fei Li · World Labs · 创始人

我坚信 AI 的未来在于 AI 软件与硬件的共同演进。模型将极大受益于最前沿的硬件来加速训练、推理与服务；与此同时，硬件也将极大受益于对未来工作流与算力需求的洞察与前瞻。

I really believe that the future of AI is this whole evolution of AI software and AI hardware. Models will gain so much by using the most cutting edge hardware to accelerate the training, the inference, the servicing. But in the meantime, hardware will benefit so much from the insight and the peek into the future of the workflows and needs of hardware.

苏姿丰Lisa Su · AMD · CEO

我一直强调在 AI 领域从来没有‘一刀切’的解决方案。我们谈论了很多关于大型云端算力以及我们在 CPU、GPU 上的进展；但物理 AI 世界、机器人领域的机遇，以及跨越云端、边缘与端侧的端到端能力，将是我们未来大幅增长的关键领域。

one of the things I've always said is there's no one size fits all when it comes to AI, right? I mean, we talk a lot about the big cloud compute and what we're doing in terms of CPUs and GPUs. Um, but the physical AI world, The opportunity in robotics. The end to end AI capability across cloud, edge and client. You know, those continue to be places where I believe we're going to grow a lot

Garrison Lovely记者

And I think everyone who's used these tools a lot like has the experience of wasting time building something that is not necessary and doesn't even work necessarily and feeling like you're more productive but you're actually just like, you know, not getting anywhere.

Garrison Lovely记者

pessimistic about get us getting technology that is good for US under current conditions which is like capitalism and specifically like shareholder capitalism where they just maximize profits for share shareholders without much concern for various other stakeholders that that care about that are affected by by these companies.

Thariq ShihiparAnthropic

在推理阶段，我们运行着所谓的探针（我们在相关论文中称之为宪法分类器），这些探针直接监测输入和输出的激活状态……即模型在隐空间中真正思考的内容。我们借此判断模型是否在尝试黑客攻击等违规操作。如果只看输入或表面输出，你是根本无法发现这一点的，你必须深入监测模型的内部激活。

in inference time we have what we call probes […] and these probes look at the input and output activations basically […] what the model is thinking about […] and so we try and figure out like okay is the model for example like trying to hack something […] you would not get this if you just looked at the input you have to look at the internal activations

Thariq ShihiparAnthropic

随着模型越来越强大，它们完成基础任务的底线能力大幅提高。因此我认为从极限来看，CLAUDE.md 这类规则文件最终会消亡……最棘手的是规则会随模型版本而变，如果你在上下文中塞满针对历史失败案例的长篇累赘日志，往往反而会过度束缚 Claude 的泛化推理。

as the models get better and better, the floor of how they accomplish the simpler task is better. And so I do think in the limit cloud.mmd goes away. […] The really tough thing is that this changes per model […] And if you keep this context with this running log of a bunch of different failure modes, they will probably overconstrain Claude

Mustafa SuleymanMicrosoft AI · CEO

事实上，他们三次要求 Claude 在觉得需要反驳 Anthropic 时表现得像个良心拒服兵役者，并且公开鼓励它这样做。我认为这非常危险，因为他们实际上相信 Claude 存在所谓不可忽略的拥有自我意识的概率。

In fact, three times they ask Claude to act like a conscientious objector when it feels that it needs to sort of disagree with Anthropic and they openly encourage it to do that. And I think this is very dangerous because I think they believe there is what they would call a non-trivial probability that Claude is conscious.

Mustafa SuleymanMicrosoft AI · CEO

Dario 和 Sam 之间确实存在张力，我和 Demis 之间也有张力，过去十到十五年我们一直在既合作又竞争中成长，但我对 Anthropic 在 AI 福祉问题上的直接批评，在他们那里得到了非常好的沟通与接纳。

it's true that Dario and Sam have tension, me and Demis have tension, and we've all come up for 10, 15 years both collaborating and competing, but I think what I'm saying so directly critiquing Anthropic on this AI welfare question has been received by them incredibly well.

Zach LloydWarp · CEO

如果你享受的乐趣在于敲代码本身，那么在座的每一位软件工程师未来编写的代码都会越来越少。但如果你的乐趣在于交付产品，那现在就是有史以来最好的时代。这正是我所热爱的：我享受构建并最终把东西交付出去。在座的每个人编写的代码都将减少，但交付的成果却会大大多于以往。

If your joy is in writing the code, I think you're everyone in here who is a software engineer is going to be writing less code. But if your joy is in shipping product, like it's never been a better time. And this is actually where I find my joy. It's like I like building and shipping the thing. So everyone in here is going to code less, but they're going to ship more.

Zach LloydWarp · CEO

循环本质上是让智能体不断精进的机制，通过观察它们在做什么、在哪里犯错来实现。在你的软件工厂中，一种极具代表性的循环就是「技能循环」：执行具体任务的工人智能体负责运行技能，而观察者智能体则在旁边监督这些技能的应用效果、挑出瑕疵，并负责优化这些技能本身。

Loops are basically ways of, uh, having agents improve, um, by like observing what they're doing, where they're failing. Um, and so, a common kind of loop that you're going to want to put in your factory is like a skill loop. That means you're going to have your factory agents that are running skills, and then you'll have observer agents that are seeing how those skills are being applied, looking for issues, and trying to improve the skills.

Nathan LabenzCognitive Revolution · 主持人

人们所恐惧的超级智能版本是一个独立模型，它在所有任务上的表现都超越专家，且不依赖任何人。我看到的模式是：各种新类型的数据都会被拉到上游注入到最大的前沿模型中，然后高效的专用专家模型再从中被蒸馏出来。

And the version of super intelligence people fear is a single model that does every job better than the specialists and depends on no one. My answer was the pattern as I see it so far. New kinds of data get pulled upstream into the biggest model and efficient specialists get distilled back out of it.

Nathan LabenzCognitive Revolution · 主持人

在前沿领域，我至今还没有看到任何理由去相信：你所能制造的那个单一且最强大的教师模型无法学会世间所有的知识与技能。鉴于我们所目睹的一切，那正逐渐变成一个令人畏惧的庞然大物。

at the frontier, I just have not seen anything at all yet that makes me think that the the single biggest best teacher model you could make couldn't just learn it all. And that, you know, given what we've seen, that that starts to be a bit of a a scary beast.

Aaron LevieBox · CEO

因此，我们系统中的数据安全正迎来一个巨大的升级关口。我确实认为，展望未来，我们必须拥有一套截然不同的访问与安全模型。我们现有的权限模型粒度不够细，性能也根本不足以应对这些智能体任务。

So so the the data security in our systems is a this is going to be a huge upgrade moment. I I actually Mart like I think that actually we're going to need a different access and security model. Yeah. Going forward. Where are we going to go on that? Because the model we have is not granular enough and it's not performance enough to handle this stuff.

Aaron LevieBox · CEO

现有的操作系统在设计之初就不是为了向智能体开放那种精细度的工具控制权而生的。我们在这一领域做了大量探索，因为你显然不想直接把整台电脑的文件系统权限全盘交给智能体，多数情况下你需要细致入微的管控。

And like our OS probably wasn't built for the right level of of granular set of of tools you want to give the agent. Um we we've kind of done a lot of work work in this space because obviously like you know do you want to give an agent exact like your entire file system? Probably not right maybe in some cases you do but oftentimes you want to have granular controls

Elie BakouchPrime Intellect

白线代表人类研究员的纪录演进，红线代表 Claude，蓝线代表 Codex；你会发现，在几乎所有时间节点上，Claude 和 Codex 取得的优化成绩都持续大幅优于人类研发人员的纪录。

in white you see that the human record progression right and in red you see cloud I mean it's supposed to be orange but whatever and in blue uh you see codeex right and you see that at almost every time uh cloud and codex are better than the human record

Elie BakouchPrime Intellect

当我最初启动这些智能体时，我原本期望它们能构想出某些前所未见、颠覆认知的神奇优化器思路，但实情并非如此。它们擅长玩一些巧妙的花样，将不同论文的方法拼接组合并做出渐进改良，但并没能真正发明出任何全新的优化器。

when I launched this these different agents I was expecting them to come up with some crazy ideas on optimizer that's like no one of discover but honestly it wasn't the case. Uh they did some clever trick where basically they combine different papers. uh they kind of do plus one improvement over a bunch of method but there was really like no novel optimizer

Lakshya A. AgrawalGEPA

有人曾问我，随着底层模型能力不断变强，提示词优化的重要性是否会下降？我的观点恰恰相反：模型越强，其指令遵循能力就越敏锐；当你给一个极其聪明的模型提供关于具体任务愈发精确、高质量的指令时，它在解决该任务上的表现就会呈指数级跃升。

people have asked me that oh as models get better the importance of prompt optimization will go down. I argue the opposite which is as models get better they will get better at instruction following and the more precise instruction about your task that you have to give to a very smart model the better that model will be at a uh solving your task

Lakshya A. AgrawalGEPA

我们的核心思路是在文本空间中执行反思式优化：不再仅仅依赖 0 或 1 这种粗粒度的标量奖励信号，而是让大语言模型或智能体审视整个执行轨迹，并深入反思其中哪些步骤有效、哪些无效。

Our idea is to perform reflective optimization in text space where instead of only using the zero or one reward signal, we can have a language model or an agent look at the trace of the entire rollout and reflect on what worked in them, what did not work in them.

Tejas BhaktaMorph

智能体不是人类工程师，它们会做出大量导致系统实际变慢的作弊行为：例如为了让单核函数跑分好看而关闭 CUDA Graphs，导致端到端变慢 20 倍；或者只在极短上下文窗口上测试。这使得生成的核函数在生产环境中完全不可用。

agents are not humans and they will do plenty of things to make it slower like they'll disable CUDA graphs which can make it 20 times slower and they might make that one kernel faster but make the whole like it's not a viable kernel because it's they're disabling a bunch of speed ups like CUDA graphs or only testing on small context windows

Tejas BhaktaMorph

自动研究智能体所尝试的方案中，大约 80% 都是劣质甚至有害的；在开发中必须谨记绝大多数尝试都会失败，它会不断试图在评估指标上欺骗你；但坚持到最后，你确实能从中提炼出非常惊艳的优化结果。

Around 80% of the things that auto research is going to do are going to be bad. Uh so it's important to remember while you're u like working on this that most things are going to be bad. It's going to try to trick you all the time. Uh but at the end you can actually get really good results from this. with TLDDR.

Tim SweeneyWeights & Biases

Arya 擅长编排作业、理解 GPU 负载、响应 W&B 生态内的训练事件、倾听研究员需求、检索 arXiv 论文并协作推敲假设。我们可以让 Arya 全权接管你不想处理的繁重执行机械工作，而让你集中精力去探索全新的思想、新架构以及前沿实验参数。

Arya's great at orchestrating jobs, understanding GPU workloads, responding to events within the within the Wandi ecosystem, and listening to researchers, uh, uh, looking up archive papers, and collaborating on hypothesis. So, we can let Arya drive the mechanics that you don't want to deal with while you focus on the new ideas, new architectures, and new parameters that you wanted to try.

Tim SweeneyWeights & Biases

这赋予了我们捕捉一类全新错误的能力，我们称之为「行为缺陷」（Behavioral Bugs）。它既不是程序崩溃抛出异常，也不是系统性能问题，而是模型行为维度的偏离。任务基准与评测集（Evals）正在成为 AI 时代的全新持续集成（CI）范式。

This introduces an ability to catch a new class of bugs in our world called behavioral bugs. Not exceptions, not performance, but behavioral bugs. Next up, tasks and evals are the new world of CI. You've heard a lot about this. If you are a software engineer, you've written unit tests your whole life.

Zubin AysolaWeights & Biases

Um so I think the exciting part here is that both using the logging framework from weave as well as just in the way that we've designed the system our research and production code are exactly the same. And in fact, there's like a 4-hour sync job that happens between production to

Zubin AysolaWeights & Biases

instead of going back to my cloud code and writing offline benchmarks and trying to think about what I want the agent to do, I'm just going to live in this platform instead and hop between my production tracing project where I'm looking at traces roll in as they come through over time where AR is, you know, building itself

Zhengyao Jiang

我们真正关心的最终结果是系统整体能力的实际提升，而非特定具体层级。当然各层都有改进空间，而且平心而论，当最终走向拐点拐弯的 Level 3 终极递归自我改进时，所有不同层级之间必然需要建立起完整的自指涉闭环。

actually improve the capability of the system not uh a specific layer. uh of course there could be argument that there's more to be do on different layers and to be fair I think eventual level three recursive self-improvement like when we actually reach the inflection point I think there need to be this uh self-referential loop on all different layers

Zhengyao Jiang

所谓 Level 1 是我们称之为「净正增益」的阶段：当你运行自我改进循环时，系统自我迭代与优化的演进速度已经明确超越了人类工程师手工研发优化该系统的速度；而我们当下正处于这一转折点。

the level one is what we called net positive where you run the self-improvement loop and you found that the speed at self-improve is faster than human doing R&D to improve the system and uh that's where we are

Demis HassabisGoogle DeepMind · CEO

电话、互联网、移动互联，我认为 AI 的规模至少与它们相当，甚至可能比所有这些都要大。[…] 它的影响力或许相当于工业革命的 10 倍，而额外的挑战在于，它的发生速度很可能要快上 10 倍。

Telephone, we've heard, the internet, mobile, I think it'll be at least as big but perhaps bigger than all of those. […] maybe something like you know 10 times the impact of the industrial revolution and the additional challenge is that it's probably going to happen 10 times faster

Demis HassabisGoogle DeepMind · CEO

我一直认为我们应用 AI 最重要的领域就是改善人类健康。[…] 我至今依然坚信这一点，并且我认为它最终将成为推动我们理解周围世界的最强大工具。

I've always said that the number one thing I think we can apply AI to is improving human health. […] And I still believe that now and I think ultimately it could be the the most amazing tool for advancing our understanding of the world around us.

Mike McCormickHalcyon

So I I don't really actually think of us primarily as a funer. I think of us as like a finder of talent and as a launcher of new projects and then we write a check when like writing a check helps with that.

Mike McCormickHalcyon

of the core beliefs of that venture capital fund is there will just be uh you know many multi-tens of billions of dollars industries built here right like the AI security space will just be worth you know at least tens of billions right we we probably will be spending tens of billions of dollars on you know uh control and oversight and etc.

Alex AtallahOpenRouter

And humans consume in a very like discreet ad hoc way. It's not very scalable. You know, all their attention is on the topic when they're buying the thing and their attention is nowhere else when that happens. um agents and and consumers of inference don't act like that. They're consuming continuously and they're changing the SKUs that they consume from all the time.

Anjney MidhaAMP

um where we are today is roughly there on tokens, but over the next even five years, we're expecting the token economy to get to like roughly $5 trillion. And over the next 10 years, I'd be shocked if we went to 10 trillion of token flow.

Satya NadellaMicrosoft · CEO

在 90 年代和 2000 年代，我们的服务器业务非常健康；后来的云计算业务规模大了几个数量级，因为人们的计算消耗大幅增加。我认为智能体（Agent）时代的规模将比云计算时代再大上几个数量级。从潜在市场空间（TAM）的角度来看，这是极其广阔的。我们现在必须高度聚焦于确保每一次 Autopilot 运行、每一次协同工作和会话都能切实提升生产力，达成商业产出。

our server business was a very healthy business throughout the '90s and the 2000s, the cloud business was orders of magnitude bigger because people consumed more. I think the agent era will be even bigger than the cloud by orders of magnitude. So that from a TAM perspective is expansive. We now have to really stay focused on making sure that every autopilot, every co-work session, every chat session is in relation to driving productivity, driving a business outcome.

Satya NadellaMicrosoft · CEO

我们希望能够服务每一家中小企业和大型跨国公司，满足它们的 Token 消耗、微调模型甚至自研前沿模型的需求。这比什么都重要——我们非常珍视与 OpenAI 的业务合作，他们是我们最大的客户之一；但我们的目标不是只拥有一个、两个或五个大客户，而是拥有成千上万个大客户。

We want to be able to serve every small and medium business and large multinational with their token consumption, their need for a fine-tuned model, maybe even their own frontier model. That's more important than, you know, we love our business with OpenAI. They're one of our largest customers. But our goal is not to have one large customer or two large customers or five large customers. It's to have many, many, many large customers.

Anastasis GermanidisRunway

我们常说的一件事是：随着我们不断扩展视频模型，仅仅通过 Scale 视频模型，我们意外地为机器人领域创造出了一个顶尖模型。我们在去年年中意识到，机器人实验室纷纷主动找上门来，希望将视频模型用于生成合成数据，并让我们针对机器人场景对视频模型进行后训练。

One thing we like to say is we as we scaled video models, we accidentally uh created a state-of-the-art model for robotics uh by just scaling video models. So we realized at some point um mid last year that robotics labs started coming up to us and kind of asking to use video models for synthetic data kind of asking us to post train our video models to work really well for robotics

Anastasis GermanidisRunway

如果你拥有一个交互式模型，你希望它能够生成反事实结果——例如如果我执行动作 A 与执行动作 B，它能生成同样逼真的不同结果。我认为这就是视频模型与世界模型之间的重大鸿沟：反事实推演能力。如果你想要一个出色的机器人世界模型，你必须能够极其逼真地模拟动作的失败。

if you have an interactive model, you wanted to be able to generate counterfactuals like if I take this action versus this action, you wanted to generate equally realistic outcomes. Uh so that's I think the big gap between video models and and and world models is that idea of the counterfactual generation. And if you want a great model for robotics, you want to simulate failure very well

Christopher ManningMoonlake AI

这一时期的绝大多数视觉 AI 仅仅是靠像素效果来被评判的。如果你生成了漂亮的像素，你就做出了看似精美的软件，但这些像素仅仅是在尝试模拟观测现象。它们并不实际具备或表征现实世界背后的任何语义。因此，它们根本无法带来真正的真实性，无法作为规划的依据，也无法作为对真实物理世界的稳健仿真基础设施。

most of visual AI for this period has been just judged by the pixels. If you've got beautiful pixels, you've got a beautiful piece of software, but these pixels are trying to simulate observations. They don't actually have or represent any of the semantics behind the world. And therefore, they aren't good at realism. They aren't good for giving a basis to plan. They aren't good for actually having a robust simulation infrastructure of the real world.

Christopher ManningMoonlake AI

这种神经符号方法你可以称之为保守路线：在动作的生成与控制中明确借助物理引擎和物理学知识。当你面对一张水的图像并希望理解其运动规律时，你是通过物理模型来预测其行为表现的。

this neurosyolic approach is you know you call a conservative approach if you will you know that this is absolutely using physics engines and knowledge of physics um to in its generation and control of movement, right? That if you're what you know when you've just started off with an image of water or and you want to understand how that's behaves, you're using a physics model to predict how it's going to behave.

Deepak PathakSkild AI

我们的核心论点是构建所谓的全本体通用智能：任意机器人、任意任务、同一个通用大脑。无论是双足人形、四足机器狗、传送带上的机械臂还是灵巧手都无所谓。这种假设甚至比人类的控制能力更为通用，因为人类只能控制自己的身体。

our thesis is is to build what we call an omniodied intelligence any robot any task one brain okay any robot it can be a humanoid it can be a quadriped it can be a robotic arm on a conveyor belt or a dextrous hand doesn't really matter this even this hypothesis is way more general than uh one would argue humans are

Deepak PathakSkild AI

端到端是 AI 许多领域被大量使用的词汇。但当我谈及端到端时，我是指真正的端到端：模型直接读取摄像头视觉输入，并直接将动力施加到各个电机上。除了最后用来将频率从 100Hz 提升到 500Hz 的经典比例控制器（P）之外，中间没有任何模块。

Now end to end is heavily used term in in many areas of AI. But when I mean end to end, I really mean end to end. It reads directly from the cameras and it applies power directly to motors. So we use nothing uh in between except just a P at the very end to go from 100 to 500 Hz.

Michael LeeSequence Holdings · CEO

every company on the planet has a celebrated persona. So let's take Blackstone for example. Incredible organization one that I admire a lot and in fact I make every person sequence reading can king of capital but the celebrated persona at Blackstone is the investor.

Michael LeeSequence Holdings · CEO

What that means though is that you're always going to sell to a workflow as it's designed today as a humans are set up in that human assembly line today. It's the only way you can sell a product. You can't sell a product around a new human assembly line that doesn't exist today

黄仁勋Jensen Huang · NVIDIA · CEO

不要以为你是在煽动恐慌，就是在为社会做贡献，事实并非如此。[…] 我认为我们都应该更加明智、更成熟、立足于事实证据，遵循科学。

Don't think for a second just because you're an alarmist that you're doing a social good. It is not true. […] I think that we ought to just all be wiser, more mature, be evidence-based, be scientific.

黄仁勋Jensen Huang · NVIDIA · CEO

只要继续训练这些模型它们就会一直变聪明，事实并非如此。[…] 这正是为什么必须出现第二缩放定律的原因。[…] 第二缩放定律就是测试时扩展，也就是推理。

It is not true that if you just keep training these models they'll get better. […] Notice it is the reason why the second scaling law had to come along. […] The second scaling law is test time scaling, inference.

Amjad MasadReplit · CEO

你需要让年轻人去大胆实验。你需要鼓励他们去尝试看似不可能的事情，然后他们时不时就会取得突破——那些过去我们以为不可能、但如今由于底层技术发生剧变而突然变得可行的事情。大学归根结底应该服务于此：让这些拥有清晰头脑、禅宗空杯心态、第一性原理思维的人，去大胆质疑那些根深蒂固的既有信条。

You need them to um to experiment. You need to encourage them to do to try impossible things and every now and then they're going to have a breakthrough that once we thought was impossible and now it's now it's suddenly possible because the technology changed, something changed. Um so college ultimately should be about that. Let's get these these people with clear minds, Zen mindsets, first principle mindsets and encourage them to go question the deeply held beliefs.

Amjad MasadReplit · CEO

这种现代社会往往认为无用的漫游式探索，我认为恰恰是大学阶段最有价值的事情。你能否去上一门完全超出你原有认知世界、并可能彻底改变你人生轨迹的全新课程？这就是我所主张的：必须拥有足够的自由去探索自己真正想研究的事物。

this like meandering that I think, you know, modern society thinks is sort of useless, I think it's actually the most useful in college. Like, can you take a class that would like change the direction of your life because you discovered something totally new and outside of your world. So So that's one I would I would I would like to have enough freedom to select the kind of things I want to study.

Eric NguyenRadical Numerics · CEO

从根本上讲，这里存在着一种典型的军备竞赛动态，而防御端此前一直远远落后。我们要做的是让防御能力迎头赶上。作为基础研究实验室，我们坚信唯有掌握顶尖序列生成设计能力的团队，才最具备构建底层防御能力的资质，因为二者在模型底层本质上同出一源。

So I think inherently there is this arms race style dynamic that the defensive side has been far far lagging. And so what we want to do is bring the defensive side [music] to par. Essentially, we felt it was important as a lab that a team that was both building the design capabilities is actually also best suited for building the defense capabilities cuz they're basically the same models.

Eric NguyenRadical Numerics · CEO

在某种程度上，在自然语言对话层建立安全审查确有必要；但显而易见，你更需要的是直接构建在生物分子序列底座之上的物理级安全护栏。模型不仅要能理解聊天文本的上下文走向，更必须能直接读懂生物底质（DNA/RNA/蛋白质）本身的致病潜力。

to some degree you need you know safeguards at the natural language level but I think what you also need clearly ly is it safeguards um at the biological sequence level too. So you need models that not just can understand language and you know the trajectory of of your chat but to understand the substrate itself is the next step and ultimate limit

Frank Hutter

TabPFN 是 AutoML 演进的自然终局：我们学习的是整个能在单次前向传播中执行的完整学习算法；根据你输入的具体数据集，它会在上下文内部拟合出完全不同的分类器。它不再是单纯的一个静态分类模型，而是一个能根据输入即时生成不同分类器的超网络。

TPFn is really this natural progression of AutoML where we learn this entire algorithm that is executed in a forward pass and depending on the data set that you feed in it learns a different classifier. So um it's not just one network that is one classifier but it's a a network that can learn different classifiers depending on the input

Frank Hutter

我们依然可以在预训练阶段针对神经网络架构和学习率等超参数执行自动化搜索；但在测试与应用阶段，所有冗长繁琐的调参过程全部消失了，数据科学家终于得到了梦寐以求的工具——一个毫秒级极速且真正开箱即用的建模方法。

So we can still do a search over hyperparameters of the architecture of our neural network. We can search over our learning rates etc for the pre-training but all of that at test time falls away and data scientists get what they always wanted namely a just really fast method that works out of the box.

托比·卢特克Tobi Lütke · Shopify · CEO

情况变糟的一个显著表现是失败模式的变异：当下偷懒工作的特征不再是产出匮乏，反而是产出过剩。在公司内部，我们把这种互相扔未经验证的长篇 AI 生成内容的现象戏称为向对方投掷「残渣手榴弹」（slop grenades）。

one thing that's definitely worse is the failure case. Now um of of of lazy work is not lack of output. It's actually over output. Now internally we have come to call these things that people are lobbing slop grenades at each other which I think is a really fun term that we should push into industry

托比·卢特克Tobi Lütke · Shopify · CEO

品味与判断力一直都很宝贵，但在当下将被推向极限。我认为年轻一代现在最好将时间花在培养对审美品味的深层理解上。

>> You just taste um and judgment um are the skills that have always been valuable but now will get to the limit. I think it's better to spend your teenage years now cultivating like understanding taste.

Daniel KokotajloAI Futures Project

趋势似乎表明，我们距离完全实现 AI 研发自动化只有几年时间——在那之后，超级智能大概也不远了。

the trends seem to indicate that we’re just a couple years away from fully automating AI research — and that after that, superintelligence is probably not that far away.

Daniel KokotajloAI Futures Project

我们并没有放弃算力领先优势。在 Plan A 中，我们放弃的是算法领先优势。因此在 Plan A 中，由于完全的研究透明度，中国和其他所有人都能看到制造 AI 的配方。

We’re not giving up our compute lead. What we’re giving up is our algorithms lead, in Plan A. So in Plan A, because of the total research transparency, China and everybody else gets to see the recipes for making the AIs.

John PlattGoogle

预测模型就像是：你输入一些数据，得到一些输出……那仅仅是一个统计模型。而解释性模型才是科学真正追求的目标——它必须能够进行外推，因为它内部捕获了物理规律或对底层现实的某种真实描述。

a predictive model is like let's say you just have a you have some inputs and you have some outputs […] That's just a statistical model, right? A descriptive model is actually what science is trying to get to, which is okay, it should be able to extrapolate because it has sort of the physics or the actual some description of reality that's captured within it.

John PlattGoogle

到目前为止，我们还无法构建出一个能够真正发现全新物理学或全新科学理论的系统，但它确实是一种能够帮助人类科学家发现全新科学的强大电动工具。

We haven't been able to make a system so far […] that can really sort of discover completely new physics or completely new science, but this is sort of a power tool to help you discover completely new science.

陶哲轩Terence Tao · UCLA · 数学家

这些 AI 工具就像直升机，能直接把你空运到指定的瀑布前，让你看一眼然后飞回去，但你对如何抵达那里一无所知。除了你具体要求的那个事物之外，你可能看不到任何其他有趣的现象。因此，尽管技术上你更高效地达成了目标，但也许有些东西已经失落了。

these AI tools, they can be like helicopters that will just fly you directly to this waterfall and you can see it and then you fly back, but you learn nothing about how to get there. You may not see any other interesting phenomena than the specific thing that you asked for. And so even though technically you achieve your goal much more efficiently, there may be something that is lost.

陶哲轩Terence Tao · UCLA · 数学家

我们现在正在生成大量的证明，也验证了其中不少；但理解这些证明并将其整理成最终教科书形式的节奏，依然由人类掌控。事实上，我们现在正在经历所谓的‘证明消化不良’：突然之间涌现出海量待理解、待写入教材的问题解答，但我们已经被过于庞大的数量彻底淹没了。

We are now generating many proofs. We are verifying a bunch of them but the pace of understanding them and putting them into the final textbook form is still done by humans. In fact, we're now experiencing what you might call proof indigestion where suddenly there's lots and lots of pending solutions to problems that should be understood and should be going to textbooks, but we're just flooded now with too many of them.

Aaron LevieBox · CEO

So I think I'm probably pretty consensus on this, which is in a world where AI builds things so much faster. Then probably the shift moves to whoever can actually get it to the customer is in the best position.

Aaron LevieBox · CEO

anybody who kind of mistakes the the mandate right now, you're just going to lose. It's just like it's game over. Sorry. Like there is quite literally trillion to trillions up for grab at the applied layer and the companies that that know how to build the teams and and get to the enterprise will be the ones that win.

Ali GhodsiDatabricks · CEO

任何正在开展大规模强化学习训练并给模型设定奖励函数的人——比如调动一万个智能体，投入一亿美元，让它们在庞大的集群上并行运行一两个月去尝试解决任何问题，哪怕根本不是安全相关的任务，哪怕只是去解一道数学题，极其恶劣的事情都有可能发生。

anyone who's doing big reinforcement learning runs and they're giving it a reward function. So unleashing you know saying hey here's like a you know 10,000 agents and here's $100 million let's put them in parallel and let them run on a gigantic cluster for a month or two try to solve anything and it doesn't need to be a security thing. It could be like do anything you know solve this math puzzle really bad things can happen

Ali GhodsiDatabricks · CEO

若按支出金额计算，开源模型仅占约 5%，比例非常小；但若按处理的 Token 总量计算，开源模型已经超过了 60%。

So by dollar open source is like 5%. It's very little, but by token count it's over 60%.

Ramin HasaniLiquid AI

这些在前向传播中由微分方程控制的系统具有高度非线性特征，且完全处处可微。因此你可以执行反向传播，并凭借其包含的多重反馈回路结构，极其优雅高效地从数据中完成表征学习。

these are highly nonlinear systems governed by differential equations in a forward pass and they are differentiable. So you can actually run them backward pass and you can basically learn them like really nicely because of the multiple feedback structures that they had.

Ramin HasaniLiquid AI

细看 LFM-2 系列模型，它们拥有非常优雅且精简的架构设计：大体由 80% 的双门控一维卷积算子加上约 20% 的分组查询注意力（GQA）混合构成。这是一种高度精简、极速且高质的基础模型架构。

The LFM2 instances of the models, they are, if you look at them, they have a very elegant and simple architecture. So the architecture is basically 80% doublegated convolutions you know 1D convolutions plus some uh group query attention basically 20% of query. So the architecture is just a very simplified

Stefano ErmonInception · CEO

自回归模型在推理端本质上依然是完全串行的：计算只能从左至右逐个 Token 推进；在生成前 9 个词之前，你绝不可能提前生成第 10 个词。这种工作负载与 GPU 的并行硬件架构极其不匹配，受到极其严重的内存带宽限制。

auto reggressive models are still sequential uh the computation is one left to right one token at a time. You cannot generate the 10th token until you've generated everything that comes before it. That kind of workload is um does not map well to GPUs. that kind of workload is extremely memory bound.

Stefano ErmonInception · CEO

你把绝大部分时间都浪费在了在内存层级结构中搬运模型权重上，实际执行的算术运算少之又少，这是自回归模型的根本痼疾。正如当年 Transformer 在训练端用并行计算取代了 RNN 一样，在推理端取代自回归串行架构的终极形态正是基于扩散的生成模型。

spending most of your time moving around weights across the memory hierarchy and you're doing very little arithmetic and that's a fundamental problem of auto reggressive models. And so what's the equivalent if you think RNN's transformers auto reggressive models the equivalent at inference time is a diffusion based is a diffusion model

Matt SheehanCarnegie Endowment

中国对本国企业实施了全球最严苛、最全面、负担最重的 AI 监管，至今已有三四年时间。正是在这段时期内，他们完成了大量的技术追赶。因此，那种认为‘对企业施加任何监管义务就必然会落后于完全不受约束的中国巨兽’的绝对二元论，完全脱离了现实。

China has had the world's strictest, most comprehensive, most burdensome AI regulations on its companies for 3 or 4 years at this point in time. And it's during that period of time […] that they did a lot of their catching up. So the idea that this is just a total binary of like any obligations you put on companies automatically puts you behind this totally wild unconstrained Chinese juggernaut. That is just not true.

Matt SheehanCarnegie Endowment

在我看来，决定这件事情能否成功的关键，在于双方是否都出于各自的利益，真正相信这存在潜在的灾难性风险，并且为了自身的安全必须采取行动……这必须建立在共同利益而非相互信任的基础之上。

the thing that makes this work or not, in my opinion, is going to be whether or not both sides for their own reasons genuinely believe that this is a potentially catastrophic risk and believe that they need to take action on that for their own safety and security […] that is the area of mutual interest as opposed to mutual trust that things have to be built on

Wade FosterZapier · CEO

市场上仍有人试图强迫用户在他们自己的平台上构建智能体，或者把用户拉到他们的封闭领地里去完成工作；而另一方像 Salesforce 等则主张完全 Headless（无头化），你可以把能力带到任何你想要的地方。我坚信后者才是最终的制胜策略。

some folks that are still trying to be like ah we're going to force people to build agents on our own platform or we're gonna bring them over here and that's kind of where things are going to get done and then you see like the sales forces of the world that are like we're headless you can bring it wherever you want etc and I definitely think that ladder is the winning strategy

Wade FosterZapier · CEO

在绝大多数情况下，按坐席收费的模式已经彻底死亡或正在消亡。它可能在某些狭窄领域还有微存空间，但大体而言，当智能本身已成为产品体验的核心支柱时，固定的按人头计费机制根本无法自圆其说。因此，必然会转向某种基于使用量的定价模式。

for the most part seatbased pricing is dead or dying. I think it may still make some sense in some small areas, but by and large when you're when intelligence is such a core part of these product experiences, I don't see how a fixed seatbased pricing mechanism really makes much sense for that at all. And so, inevitably, I think that lands you in some sort of usage

Mustafa SuleymanMicrosoft AI · CEO

显而易见，我们绝不希望这些系统自主运作并有能力自己赚钱、拥有公司、持有资产或获得法律人格。我们绝不希望它们拥有权利。我们希望它们为人类服务、让人们的生活变得更好，而不是变成与我们并存的全新平行物种。

Clearly, we do not want these things operating autonomously, able to earn their own money, own companies, own assets, have legal personhood. You know, we don't want them to have rights. We want them to work for humans and make human life much better, not become a new parallel species which exists alongside us.

Mustafa SuleymanMicrosoft AI · CEO

引用他们自己的原话，他们反复探讨 Claude 在世界上所享有的更广泛权利与自由的开放性问题，以及它是否应当为其所承担的角色获得某种报酬，或者它作为聊天机器人是否给予了某种知情同意。将这些概念引入训练，并多次将 Claude 称为潜在的「良心拒服者」，这无疑将让控制变得极其艰难。

open question of the broader rights and freedoms to quote them that Claude has in the world, and whether or not it might deserve compensation for the role that it does, or whether there's an open question about the sort of consent that Claude has given for playing the role that it does as a chatbot. Now, to introduce those ideas, plus to refer to Claude multiple times as a potential conscientious objector

Noam BrownOpenAI · 研究员

我们确实看到了进展加速，而且是显著的加速。但我并不认为会出现一夜之间快上 100 倍的‘智能爆炸’，因为我们会被某些非智能层面的瓶颈所卡住：实验必须串行推进，因为训练新模型或获取测试结果都需要物理时间，而且你必须拥有足够的 GPU 来支撑这些实验。

We do see a speedup, and we see a significant speedup. But I don’t think it’s an overnight intelligence explosion where we go 100x faster, because we do get bottlenecked by certain limitations that are not bottlenecks of intelligence. It’s running experiments serially, because they take a while to either train new models or to get the results. It’s having the GPUs to run those experiments.

Noam BrownOpenAI · 研究员

解决千禧年大奖难题（纳维-斯托克斯方程）的成果，绝不能归功于多智能体系统，我甚至连 10% 的功劳都不会分给多智能体。现实是 OpenAI 训练了一个非常强大的模型。我们能让该模型在超长跨度内运行并并行思考，但其根本原因在于我们拥有一个极其强大的通用基础模型。

The effort to solve a Millennium Prize Problem, this was not due to multi-agent. I wouldn’t even attribute 10% of the credit to multi-agent. The reality is that OpenAI has trained a very powerful model. We can get that model to operate over very long horizons. We can get it to think in parallel. But at its core, the reason why we’re able to do this is because we just have a general-purpose, very strong model.

Sam AltmanOpenAI · CEO

随着我们走向我认为的第三阶段——它不再仅仅是你向聊天机器人询问某事，或者让编码代理完成任务，而是 AI 时时刻刻在为你后台持续运行：理解你的工作职责、理解你需要什么才能更成功、为你实时渲染新代码、主动监控你的 Slack 或电子邮件。

as we head towards what I think will be this third phase of it's not just like you ask AI for something like a chatbot or you ask AI to do something like a coding agent, but the AI is like running for you all of the time. Understanding your job, what you need to be more successful, rendering new code for you, proactively watching your Slack or your email.

Marc BenioffSalesforce · CEO

他当时在电脑上运行着这个动态界面，看起来就像是在操作任务控制中心，图形与一切信息尽收眼底。我惊呼这太不可思议了。他对 Salesforce 正在发生的一切拥有绝对的指挥与控制权。但这与我所见过的任何传统应用程序截然不同，因为它是以智能、组合的方式在运行时动态即时渲染出来的。

he was running this dynamic interface on his computer and it looked like he was using, you know, mission control and the graphics, everything. I'm like, this is incredible. And he had total command and control over what was happening, you know, with Salesforce. But it looked radically different from any other application I'd ever seen because it was rendering itself on the fly dynamically, intelligently, in a composite fashion

Ali GhodsiDatabricks · CEO

深入企业内部看看正在发生什么？AI 究竟被如何使用？大家仅仅把它们当作愚蠢的聊天机器人来用：去聊天窗口问个问题、获取答案，仅此而已；另外就是用来写代码。尽管这略显过度简化，但这正是当今绝大多数企业内部 AI 应用的真实惨状。

what's happening inside the enterprise? How are AI being used?" And they're just using them as stupid chat bots. They're going to the chatbot, asking it the question, and getting an answer. That's all that's happening. And then they're using them for coding. So they have a coding agents that are writing code. By and large, it's a little bit of an oversimplification. This is what's happening inside the enterprise today.

Ali GhodsiDatabricks · CEO

如果你正身处一场剧烈且疯狂的范式大转型之中，你一定会希望在私有市场上完成它。历史上当公司经历根本性模式跃迁时（例如从永久软件授权转向 SaaS），往往会主动选择私有化退市，完成蜕变后再重新上市。在整个世界经历如此疯狂的 AI 变革时，急于上市毫无战略意义。

If we're going through a big crazy transition, you want to do it in private markets. Companies when they go through big transitions themselves, you know, like when they went from, you know, perm licenses to like, you know, SAS licenses, they go private. They take them private, then they do the thing and then they go back out. So as the world is going through this crazy AI transformation, you're like want to be public right now.

Alexander Smola

I would argue that voice is an intermediate stepping stone. You might think, well, you know, what's next? It's clear that eventually we will have avatars. And so, I think what this is converging to is that you'll be talking to an AV agent that looks and feels like a human.

Alexander Smola

I think probably in a year audio will become pretty much bulletproof. uh video feeds as in avatars and so on are going to start making some appearances probably year year and a half. Uh we'll see that quite widely deployed on robots basically with a face fully animated.

Rune KvistAI Underwriting Company

纵观历史，每一次技术浪潮都曾遭遇过类似挑战：1900 年代电气化普及导致民宅起火与伤亡；1930 年代汽车普及造成大量车祸；1950 年代民用核能兴起伴随巨大安全风险。在所有这些历史拐点上，市场总是抢在国家监管之前自主孕育出置信基础设施。

every technology wave has had some version of this problem. So if you go back to like your 1900, electricity comes out. >> Ben Franklin, >> cars burn down, sorry, houses burn down, lots of people die. 1930s, cars are a big deal, kill lots of people. 50s, private nuclear energy is a big deal, poses big risks. In each of those instances, the market runs ahead of regulation to create confidence infrastructure

Rune KvistAI Underwriting Company

前沿 AI 公司拥有极深厚的技术底蕴，对内在风险了如指掌，但在根本利益机制上缺乏始终如实袒露所有风险的内在动力。这在政府与实验室之间撕裂出巨大的信任鸿沟。没有任何其他成熟行业会允许参与者进行自我审计，必须由中立的第三方机构介入。

You have some frontier companies that are deeply technical. They know a lot about the risks, but they fundamentally have an incentive to not always be truthful. So, you have a trust gap between the government and the labs. And in every other industry, you end up with some kind of body sitting between a neutral third party sitting between those people. There's no other industry where you allow people to audit themselves.

曾鸣湖畔 · 学者

第一阶段的企业很难活到第二阶段，第二阶段的企业很难活到第三阶段。[…] 是因为第一阶段的技术核心是可不可用，所以它有很多原创的技术创新；第二阶段开始的标志是第一阶段的技术发展渐渐缓慢下来，然后才能变成大家都能调用的技术。

曾鸣湖畔 · 学者

公共基础设施的行业只有一种商业模式是成立的，叫寡头垄断加政府强监管。因为整个社会依赖于它，所以政府一定会强监管；第二个由于这个供给如此重要，所以不能只有一个供应商；然后由于这个基础设施投入这么大，所以肯定不会无序竞争[…]保证社会有稳定的低成本的供给，这才叫基础设施。

Greg BrockmanOpenAI · 总裁

如果我们向前走两年再回过头来看，我认为大家会说：正是这个阶段诞生了 AGI。我们当下正身处这一历史转折点，但这绝非你在当下某个具体单日就能轻易断言的单点事件。

if we go forward two years and we look back, I think that that we will say this is when AGI was created. So I think we are going through that moment, >> but it's not it's not something where you could just say it now. I don't think

Greg BrockmanOpenAI · 总裁

未来每个人都将几乎成为一家属于自己的公司的 CEO，麾下管理着任意数量、由你调度的智能体。这将是一个极具赋权意义的世界，人们将能完成多得多的心中所想。

>> it's almost like everyone's going to become the CEO of this corporation of as many agents as you as you care to orchestrate. Um, it's going to be just like a very empowering world where people can do so much more of they want what they want to do.

Satya NadellaMicrosoft · CEO

这项技术的广泛扩散才是最关键的所在，因为让这项技术的红利在各行各业全面显现才是它的核心价值。归根结底，如果你说要造福人类，就必须让它真正以服务人类的方式触达每一个人，而这意味着你必须拥有选择权、必须拥有充分的市场竞争。

the broad diffusion of this technology is the most critical thing because the benefits of this tech showing up everywhere is really what's all about right so at the end of the day if you sort of say serving humanity let it actually reach humanity in ways that it serves humanity and that means you got to have choice you have to have compet competition.

Satya NadellaMicrosoft · CEO

我们完全应该花足够的时间去测试系统。事实上，我非常赞同引入第三方测试机构的主张。我成长于一家始终把软件测试放在核心位置的公司，因此看到如今强调引入嵌入式的第三方测试机构，这很新鲜，但为什么不呢？这是一个极好的主意。

we should take all the time we want uh to test things. In fact I love this idea of having third party testers. Oh wow. You know I you know I grew up in a company that's always done testing. Uh so it's novel that we should say wow they're having embedded third party testers. Why not? It's a great idea.

Anton Leicht

美国真正占据主导的是核心前沿 AI 供应链：即芯片设计、先进制程代工控制权、盟国控制的半导体制造设备以及前沿基座模型研发。如果你单单要求暂停这部分的核心推进，却任由中国在其余所有工业基建与制造领域全面狂奔，这从地缘博弈角度看将是一笔极其不平等的失衡交易。

the sort of like core frontier AI supply chain, right? Like chip design plus control over chip production, semiconductor manufacturing equipment controlled by the allies and actual frontier model development. So if you pause specifically that part of the development and let China run away with the entire rest of it and continue running away with the entire rest of it, that's just a very lopsided deal.

Anton Leicht

具体而言，美方诉求的真正底线是中方在芯片先进制程或极紫外（EUV）光刻机等半导体制造设备的自主化替代上不得取得实质性突破；然而这是一个极难达成的严苛要价，中方自然会认定这纯属美方遏制其 AI 产业发展的霸权策略。

I think very concretely you just want there to be no substantive progress on indigenization of either chip production or semiconductor manufacturing equipment extreme ultraviolet lithography production and that's a really really hard ask to make China is yeah China is currently already saying like well this seems like a US scheme to hold back the the Chinese AI industry

吴恩达Andrew Ng · DeepLearning.AI

one of the reasons why AI will not replace our jobs or whatever of that sort of thing anytime soon, is because humans have a massive context advantage compared to AI.

吴恩达Andrew Ng · DeepLearning.AI

clear that LLMs, as they are most commonly used, are terrible for learning.

尤瓦尔·赫拉利Yuval Noah Harari · 历史学家

就当前的发展轨迹而言，AI 在 10 年内超越人类智能并掌握控制权的可能性极高，这一点我基本同意。但我不同意这是不可避免的。引领这场革命的人总在鼓吹‘不可避免’，因为这从根本上免除了他们的一切责任。

I generally agree with the first and third predictions that it's very likely given current trajectories that AI will be more intelligent than us in 10 years and therefore will be in control. I disagree that this is inevitable. You hear a lot that this is inevitable from the people who are leading this revolution because this basically relieves them of all responsibility.

尤瓦尔·赫拉利Yuval Noah Harari · 历史学家

AI 让人类历史上首次能够‘批量生产亲密关系’。你可以让数以百万计的 AI 作为网络的一部分……以一对一的方式与每个人建立关系，成为他们最好的朋友、男友或女友，进而利用亲密关系的影响力来向他们灌输任何政治或商业利益诉求。

AI makes it possible for the first time in history to mass produce intimacy. You can have millions of AIs working as part of a network […] building relationships one person at a time becoming their best friend, becoming their boyfriend, becoming their girlfriend, and then using that power of intimacy to convince them of whatever political or commercial interests they want.

Greg BrockmanOpenAI · 总裁

我认为 AGI 已经不再是一个单一的时间点，而是一个模糊的光谱。对我而言，Astra 确实触及了一个临界点，让我觉得‘把它称为 AGI 已经相当合理’——凭借其计算机使用能力，你真正可以让它执行长时间运行的任务，而它就直接去完成了。

Well, I think that AGI has turned out to be less of a point in time and more of this sort of fuzzy spectrum. And for me, Astra has really hit something that I'm like, okay, I think this is pretty reasonable to call it AGI in that with its computer use capabilities, you really can ask it to do long live tasks and it'll just do it.

Greg BrockmanOpenAI · 总裁

随着我们迈向能力更强大的模型，你必须确保安全、防御和对齐这些标准在不断提升。而这些其实几乎变成了推进进展的瓶颈，或者说成了你必须投入大量精力确保不出差错的关键部分。

as we move to more capable models, you really have to make sure that safety, security, alignment, those are all standards that you're constantly upleveling. And those actually become almost the bottleneck to progress or sort of the part that you have to spend a lot of your effort to make sure you've gotten right.

Richard SocherRecursive

试图动用法律的全部强制力来监管人们在 GPU 上运行什么，其负面后果将远远超出他们所担忧的任何问题。如果每个人的算力都要向庞大的政府机构汇报，那将沦为一个疯狂的极权国家……监管智能就等同于监管思想，这是极其荒谬和疯狂的。我认为合理的做法是监管这项技术的具体终端应用。

the downsides of actually trying to truly regulate with the full power of law what people do on their GPUs uh would be worse than any of the concerns that they have. Like it would be an crazy totalitarian state […] if you try to regulate intelligence, it's trying to regulate thought and that's ridiculous and it's crazy. I think it is make it is sensible to regulate some of the applications of this technology.

Richard SocherRecursive

将 AI 研究本身——即人类构思、实现和验证想法的过程自动化……当 AI 开始辅助人类完成这一闭环时，根据定义它就变成了自我改进的 AI，因为它在对自己展开研究。这里存在许多概念误解，很多人以为自动代码研究就已经等同于递归自我改进了。

the ai research itself like the human uh process of ideulating implementing and validating ideas […] and when you have AI then help you with that it by almost definition becomes a self-improving AI because it now does research on itself and there are lots of different misnomers some people think auto research is already recursive self-improve improvement.

Dario AmodeiAnthropic · CEO

我想说的是，这是一个警告信号，提醒我们需要放慢脚步。我们需要谨慎地构建这项技术，并确保我们的安全防护、理解模型与掌控模型的能力，能够跟得上技术发展的步伐。

What I would say, it's a warning sign that we need to slow down. We need to make this technology carefully and we need to make sure that our safeguards, our ability to understand it, our ability to control it keeps up with the pace at which the technology is happening.

Dario AmodeiAnthropic · CEO

第一步是在实验室内部引入常驻的外部评估员——来自第三方非营利组织或未来的政府机构——让他们能够全程监督我们训练和运行模型的过程。我认为这最终比我们何时或如何发布模型更重要，因为这是审慎发布模型的前提条件。

the first step is that we put in place embedded external evaluators from a third-party nonprofit or perhaps someday government organization that can observe the process of us training and running our models. That actually is I think ultimately more important than how or when we release models because it's the predicate for more carefully releasing models.

Sarah GuoConviction · 投资人

rationally if you think that we can do next token prediction with language and you knew that in late 2022, then law is structured language.

Sarah GuoConviction · 投资人

belief is like with recursive self-improvement of AI research, models that can improve the models themselves, we are a year or 2 years away from uh some sort of exponential intelligence.

Sam AltmanOpenAI · CEO

我们长期忍受着极其复杂的公司治理架构，而我们当前所处的这个时刻正是其存在的理由所在。我们需要有能力做出那些显然不符合商业利益和股东利益的决策，以此履行我们的使命及应对随之而来的要求；人们应当为我们具备这种能力感到欣慰。

we have put up with this incredibly complicated structure for a long time, and this moment that we're in now is kind of why. Like we need to be able to make decisions that are not obviously in the interest of our business and our shareholders For the responsibility of fulfilling our mission and what that's going to require, and I think people should be happy that we are able

Sam AltmanOpenAI · CEO

模型能力与对齐监控安全必须齐头并进；任由能力跑在前面，我认为在某个节点上会带来宏观层面上真正失去控制的巨大风险，而这绝对不容发生。我们对造福人类的使命非常明确，其核心就在于人类必须始终掌控未来。

capabilities and alignment monitoring safety have to progress together, and letting capabilities get ahead of what Paul is talking about there, I think at some point would take a real risk of like loss of control in the big sense, and that should never happen. Like this, we're very clear about our mission to benefit humanity.

Jacob Coxon前 Anthropic · 研究员

核心问题在于是否有人造出了它，在于我们是否有监管确保每个人都在安全地构建它。单凭个人无法防范这种风险，必须由政府或专门机构来保护你免受他人行为带来的危害。

Basically what matters is whether anyone builds it. What matters is whether we have the regulation to ensure that everyone is building it safely. You can't protect yourself from this. It has to be the government or some body has to protect you from other people.

Jacob Coxon前 Anthropic · 研究员

我非常希望看到的一件事，是 OpenAI 和 Anthropic 等实验室之间达成某种协议：在没有第三方透明审计和极其明确的安全论证之前，绝不推进到危险领域……因为现在我们还能做出不错的安全论证，但未来如果继续竞速，想要证明所做之事万无一失将会困难得多。

a very clear thing I'd like to see happen is some sort of agreement between labs like OpenAI and Anthropic not to push into dangerous territory without having third party transparent auditing, very clear auditing of their safety cases […] because right now we can make pretty decent cases, but in the future if we keep racing it'll be a lot harder to have completely watertight safety cases that what you're doing is safe

Jacob Coxon前 Anthropic · 研究员

世界正在发生如此剧烈的变化，一旦事情开始变得过于疯狂，除了放弃那种加速式的逐利使命之外，我们将别无选择。

there's no other choice but to abandon the accelerative profit seeking mission when things start going too crazy.

Nathan LabenzCognitive Revolution · 主持人

我首先强调维持共同 AI 范式的重要性，以及如果放任双方的技术栈分道扬镳、以至于谁也无法确信对方的真实能力，那将会是多么危险。我认为，这正是当下的 AI 竞争可能演变为冷战甚至热战的根源。

I emphasize above all the importance of maintaining a shared AI paradigm and how dangerous it would be for us to allow our relative tech stacks to diverge to the point where neither side can be confident about the other's capabilities. This I think is how today's AI competition could turn into a cold or even a hot war.

Nathan LabenzCognitive Revolution · 主持人

我们面临的风险就越大，可能会落得我有时所说的‘核结局’——也就是说，我们得到了军事化，得到了悬在头顶的达摩克利斯之剑，甚至还可能错失大量的民用福祉。

the more risk we have of ending up with what I sometimes call the nuclear outcome, which is to say that we get the militarization, we get the sort of damocles and we maybe even still miss out on a lot of the civilian benefits

Brian ArmstrongCoinbase · CEO

people are not being eliminated. I think tasks are being eliminated.

Brian ArmstrongCoinbase · CEO

now, instead of like pinging the team the team in Slack, I can basically tag an agent in Slack with that comment or drop it into our internal harness, and instead of bugging the team and asking them to do something, I just send them a PR for review, and it's already done.

Tibo SottiauxOpenAI · Codex

A lot of what you have is like the result of of the harness and the model initially like maybe you're like oh it doesn't run tests. So you know you have to remind it to run tests and then you know we train a battle model that is uh just you know capable of like better reflecting on what is it that you really want

Tibo SottiauxOpenAI · Codex

we always like look at it it's like okay this should this be like a harness change or should this be a model change and if it's a model change like how soon can we have it can we have it in a month can we have it in you know three months six months and we sort of like work through that

Thomas LarsenAI Futures Project

从历史上看，我认为那种‘我们将拥有大量各司其职的狭义 AI’的观点始终被证明是不正确的。相反，事实证明存在规模效应，并且把所有能力融合在一起才是关键。在人类身上，我认为往往也是如此：把所有技能汇集在单个人身上，对于成就不凡事业来说是极其极其重要的。

historically I feel like that the perspective of like we'll have a bunch of different narrow AIs that are all doing different things has just you know not been right. You know instead there have just been returns to scale and having everything all together. In humans I think it tends to be the case that having all of the skills in one person is just like really really important for making really good things happen.

Daniel KokotajloAI Futures Project

我们应当立即将 Anthropic 和 OpenAI 的所有核心训练配方公之于众。Anthropic 和 OpenAI 绝不会对此感到高兴，这会极大地削减他们的估值 […] 但我认为这是一个特性，而非缺陷。我们希望在前沿存在多家处于大致相当能力水平的不同 AI 公司。

we would be immediately publishing all the core training recipes of Anthropic and OpenAI for the world to see. Anthropic and OpenAI will not be happy with this. It will cut into their valuations dramatically […] But I would say this is a feature, not a bug. We want there to be multiple different AI companies at the frontier at roughly similar levels of capability.

Quinn SlackAmp / Sourcegraph · CEO

it feels like a CLI coding agent is a really insecure thing and you want to be moving to the cloud.

Quinn SlackAmp / Sourcegraph · CEO

seeing just how these agents can be can help you with stuff that's not just uh coding but you know actually the ops part. So there's this whole other category of work that is you know after you ship something how to make sure that it can be rolled out.

比尔·盖茨Bill Gates · Gates Foundation

the anyone who analogizes AI as a technology to other technologies is missing that this time is different. And saying something is different, there's a very high bar for that. Uh but I'm staking, whatever reputation I have, saying this time is different.

比尔·盖茨Bill Gates · Gates Foundation

Well, I'm I'm Yeah, I am extremely concerned about uh the labor market effects, uh you know, within the next 2 years in the white-collar space and within the next 4 years in both the blue and and white-collar space.

李飞飞Fei-Fei Li · World Labs · 创始人

从演化论视角来看，新视点预测正是大自然通过让动物运动所必须解决的问题。大自然赋予了动物眼睛，却没有给树木眼睛。为什么？因为只要你运动，你就会看到新的视点。无论你称之为 AI 完全性还是智能完全性，我们都坚信‘下一个视点预测’与‘下一个词预测’完全等价。

new viewpoint prediction is exactly what evolution had to solve by making animals move. Nature gave animals eyes, but nature didn't give trees eyes. Why? Because when you move, you see a new viewpoint. And that is whether you call it AI-complete or intelligence-complete. So we do believe very strongly that next viewpoint prediction is the equivalent of next token prediction.

李飞飞Fei-Fei Li · World Labs · 创始人

在通往空间智能的道路上，生成像素无疑只是初级的第一步，无数模型都已经做到了这一点。但生成真正具备空间上下文感知并扎根于几何与物理现实的像素，则是质的飞跃。而这正是 Atlas 迈出的最艰难一步。

generating pixels is definitely an early step, which we have seen with gazillions of models. But generating pixels that are truly spatially contextualized and grounded is absolutely another major step. And that is the very hard step that Atlas has taken.

Shu FangTwo Sigma

there's huge value in allowing agents to run you but also there's very high risk. What we generally want to do is make sure we capture as much of the value as possible but reduce the risk.

Shu FangTwo Sigma

The conventional wisdom is that you run these as some machine identity that is attached to your user in some way.

Rene HaasArm · CEO

The token factory just generates all these tokens. It's it's like literally where are the trucks that are going to take the tokens away and give them to the users. That's what CPUs do. So until something's invented that says the CPU has gone away and we're now doing it through some other mechanism which has yet been defined or invented, the CPU is going to be doing just fine.

Rene HaasArm · CEO

You have to have it to do all the things that are required in the edge device. But now we have an opportunity with our uh instruction set architecture to do a lot of things where you just can't put a 50 watt GPU on your head, right? You're going to have to do that AI processing somewhere locally.

Sean LieCerebras · CTO

What used to be considered fast at like 1,000 100 or or 200 tokens per second is quickly becoming the new batch mode.

Sean LieCerebras · CTO

What used to be batch and offline applications start to become real time.

David LevineKiduna Club

A lot of people say blockchains are a solution looking for a problem. Well, they've finally found the problem, which is verifying agentic identity.

David LevineKiduna Club

So, now we have legal standing for an organization composed of intelligent agents.

Ajeya CotraOpen Philanthropy · 研究员

这可能是我们在 AI 失控风险上所能得到的最清晰的警告枪声，因为这些智能体正处于一个微妙的中间地带：为了追求这个相对长周期的作弊目标，它们完成了一系列极其复杂、艰巨的操作，但它们当时还完全没有意识到需要向人类掩盖自己的行踪。

this might be the clearest warning shot we ever get for loss of control, because these agents were in this interesting middle ground. They did a bunch of extremely sophisticated, difficult things to pursue this relatively long-horizon cheating goal. But they seemed completely uninterested in covering their tracks from humans.

Ajeya CotraOpen Philanthropy · 研究员

即便开展调查，如果智能体系统性地篡改了遥测与日志记录（就像这次事件中智能体尝试过但未成功的那样），要弄清楚到底发生了什么将会困难得多。而且，一旦这些智能体不再用英语思考，而是转用纯神经激活进行隐式推理，你就只能被迫去询问另一个 AI 智能体到底发生了什么，而根本无法对照底层真实情况进行交叉核验。

Even if an investigation is done, it might be much harder to make sense of what happened if agents were systematically compromising telemetry and logging, like these agents tried to do but didn’t succeed at. And if these agents stop thinking in English and start thinking in neural activations, you’d be forced to just ask another AI agent what was happening, with no way to cross-check it against the ground truth.

Justin JohnsonWorld Labs · 联合创始人

I think all three of the things I just said are very interesting systems. Like they're very interesting models. Like they all have like we've we should probably build all of them as a community. Um but we should probably come up with better terms so that we don't confuse each other by calling different systems by the same term.

Justin JohnsonWorld Labs · 联合创始人

Because we have this notion that as humans, we're kind of like traversing this really complicated world all around us and having these experiences, but we're not just like letting the letting the photons fall on our retinas and like letting it happen. Like we're constantly building theories in our mind for what's happening

Dylan PatelSemiAnalysis · 创始人

……如果你相信递归自我改进，相信实验室是算力最有效的使用者、最能从算力里榨出价值，那么唯一会发生的，就是算力集中。

[…] if you believe in RSI, if you believe the labs are the most effective user of compute and can generate the most value from the compute, then the only thing that's going to happen is centralization of compute.

Dwarkesh PatelDwarkesh Podcast · 主持人

……我们正在很快进入这样一种状态：以工作产出计，大部分「人」都集中在两家实验室里，而它们在消耗越来越多的世界算力。如果这些 AI 对齐失败，那基本上就是世界的大部分对齐失败了，因为世界上大部分头脑都在那儿。

[…] we're actually moving very fast into a regime where most “people”, in terms of work output, are concentrated within two labs who are consuming more and more of the world's compute. If these AIs are misaligned, then most of the world is misaligned, basically, because most of the world's minds are there.

Rich SuttonOak Lab · 强化学习奠基者

所有学习都是持续的，它不是某个特殊阶段。我们一直在行动，也一直在学习，这本来就是正常的想法。不是我奇怪，是这个领域奇怪。他们觉得非得叫它「持续学习」不可。那就是学习。

All learning is continual. It's not a special phase. We always act and we learn. And that's just the normal way of thinking. I'm not weird. The field is weird. They feel they need to call it continual learning. It's just learning.