The AI Industry Asks Government to Slow It Down — 文稿

M10:00

今天这期 AI Daily Brief，我们要聊的是，有些人称之为新一封 AI 暂停公开信的那封信——Pacing the Frontier，以及它对于我们现在处在这场 AI 变革历史中的什么位置，传递出了什么信息。

Today on the AI Daily Brief, we're talking about what some are calling the new AI pause letter, Pacing the Frontier, and what it says about where we are in the history of this AI transformation.

AI Daily Brief 是一档每日更新的播客和视频节目，专门聊 AI 领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好，朋友们，在我们开始之前，先快速说几个通知。

All right, friends, quick announcements before we dive in.

M10:57

欢迎回到 AI Daily Brief。

Welcome back to the AI Daily Brief.

其实我原本今天这期是打算提前录好的。

Today was actually supposed to be a prerecord over here.

我现在人在 KPMG 的年度科技研讨会活动现场，而且我们接下来还会有一期特别精彩的 Operator's Cut，聊怎么理解并最大化 token budget。

I'm with KPMG at their annual tech symposium event, and we have a really fire Operator's Cut show about understanding and maximizing token budgets coming.

但是后来，我们又看到了一封 AI 公开信漫长谱系里的最新公开信，而围绕它的讨论量实在太大了，我觉得我们必须稍微深入聊一聊。

But then we got the latest open letter in the long lineage of AI open letters, and the sheer volume of the conversation around it, I think, demanded we dive in a little bit.

不过在进入 Pacing the Frontier 这封信之前，我们得先回到几天前，接着上次停下来的地方说。

Now, before we get into the Pacing the Frontier letter, we have to go back a couple of days to where we left off.

当时，基本上科技行业里每一家主要公司都签署了一封公开信，敦促 Trump 政府保护开源 AI。

At that point, basically every major company in the tech industry had signed an open letter urging the Trump administration to protect open-source AI.

这项行动由 NVIDIA 和 Microsoft 牵头，并得到了几十家大大小小科技公司的支持。

The effort was led by NVIDIA and Microsoft and supported by several dozen companies across both big and little tech.

最值得注意的是，OpenAI 在这封信送到之后短暂延迟了一下，最终还是签了；但 Anthropic 却明显没有出现在这份请愿书上，所以这也成了昨天 AI Daily Brief 那期节目的标题由来。

Most notably, after a brief delay upon its arrival, OpenAI signed the letter, but Anthropic remained conspicuously absent from the petition, thus leading to the title of yesterday's show on the AI Daily Brief.

当然，这场争论之所以激烈，也是因为当时的背景是，Trump 政府正在逐步收敛出一个 AI 安全测试框架。

The debate was, of course, raging in the context of the Trump administration narrowing in on a framework for AI safety testing.

六月份的行政命令把这个框架的制定截止日期定在了八月一号。

The June executive order set a deadline of August 1st for the creation of the framework.

所以这一周就是影响美国 AI 部署和模型发布治理政策的关键一周。

So this is the critical week to impact the policy that would govern AI deployment and model releases in the US.

除此之外，关于该怎么应对 Chinese AI，也一直有一场非常激烈的争论，最早是被围绕 GLM-5.2 的那些耸动标题带起来的，后来又延续到了 Kimi K3。

On top of that, there has been a raging debate about what to do about Chinese AI, inspired first by sensational headlines around GLM-5.2 and then Kimi K3.

一直有风声说，Trump 政府内部在讨论要不要禁止 Chinese open source。

There have been rumblings that the Trump administration was debating a ban on Chinese open source.

而且更重要的是，这件事不只是关起门来私下发生。

And what's more, this was not just happening behind closed doors.

这显然是一场正在进行中的公开辩论，White House 内部有多个派系都在公开为自己的政策观点发声。

This was clearly a live debate with multiple factions within the White House advocating their policy views in public.

我们甚至看到 Treasury Secretary Scott Bessent 和 Commerce Secretary Howard Lutnick 都分别发了推文，站在不同立场上表态。

We got tweets on either side of it from Treasury Secretary Scott Bessent and Commerce Secretary Howard Lutnick.

换句话说，而且我觉得这点很重要，就是现在大家都有一种感觉：这是一个正在走向高潮的时刻。

There is, in other words, and I think this is important, a feeling of a crescendo moment right now.

二零二六年的 AI，显然已经不是二零二五年的 AI 了。

The AI of 2026 is very clearly not the AI of 2025.

而这种二零二六年版本的 AI 所带来的政策影响，也显然比我们之前面对的任何东西都要重大得多。

The policy implications of this 2026 AI are clearly much more significant than anything we had before.

到目前为止，整个过程一直都是走走停停、磕磕绊绊、边反应边应对，但现在很多事情正在汇聚到一个临界点，而这会对二零二七年的 AI 在美国以及全球其他地方会如何发展，产生相当大的影响。

So far, it's been fits and starts and bumbling and reaction, but a lot of that is now coming to a head in a way that has fairly big implications for how AI in 2027 happens in the United States and beyond.

当然，就在我录完昨天那期节目之后，Anthropic 打破了沉默，通过一篇明确署名 CEO Dario Amodei 的博客文章，正式说明了他们对 open-weight models 的立场。

Now, after I recorded yesterday's show, of course, Anthropic broke their silence and formalized their position on open-weight models with a blog post that was personally attributed to CEO Dario Amodei.

Amodei 提到，最近有人指责 Anthropic 在游说推动禁令，他写道，任何读过我以前文章的人都应该知道，我并不认为这种禁令是什么有用的措施。

Amodei noted the recent accusations that Anthropic is lobbying for a ban, writing, anyone who has read my past writing should know that I don't regard such bans as a useful measure.

不过我还是想把话说清楚，不留任何疑问。

But let me state it clearly so there is no doubt.

Anthropic 从来没有主张过禁止 open-weight models。

Anthropic has never advocated for a ban on open-weight models.

不过，虽然 Dario 承认 open-weight models 是一种很有价值的公共产品，但他也再次强调了自己对于让强大模型免费开放可得这件事的核心担忧。

Still, while Dario acknowledged that open-weight models are a valuable public good, he reiterated his core concerns about making powerful models freely available.

事实上，他写道，带有保护主义色彩的禁令，对他最担心的那两个噩梦场景根本没有任何帮助。

Indeed, he wrote that protectionist bans do nothing to address the 2 nightmare scenarios he's most concerned about.

他说，他首要担心的是，威权政府可能会利用强大的 AI 来实现彻底的军事优势，或者对本国人民进行更深度的压迫。

He claims his primary concern is the risk that authoritarian governments could use powerful AI to achieve total military dominance or deep repression of their people.

Amodei 写道，在这方面，中国政府显然是能力最强的，但他们并不是唯一需要担心的对象。

Amodei wrote that the Chinese government is clearly the most capable in this regard, but they are not the sole concern.

在这个语境下，Dario 指出，open weights 基本上几乎完全无关紧要。

In this context, Dario noted that open weights are almost entirely irrelevant.

更危险的模型，实际上可能会被秘密训练出来，而且只会被威权政府用于武器和国内监控。

The more dangerous models might actually be trained in secret and used solely by authoritarian governments for weaponry and domestic surveillance.

他第二个担忧，是强大的 AI 被用来发动网络攻击，或者制造生物武器。

His secondary concern is the use of powerful AI to carry out cyberattacks or create biological weapons.

按他的判断，在这方面，广泛分发的 open-weight 模型，确实比 proprietary 模型风险高得多。

In his estimation, widely distributed open-weight models do carry a much higher risk than proprietary models in this aspect.

这是因为 open weights 会让设置 guardrails 或监控使用情况变得很困难。

That's because open weights make it difficult to apply guardrails or monitor use.

不过，Amodei 也指出，当时讨论的这个政策其实效果会相当有限。他评论说，禁止美国企业使用这些模型，并不能解决这些风险，因为坏 actors 不太可能是合法的美国企业。

Still, Amodei noted that the policy being discussed would be pretty ineffective, commenting, banning the use of these models by U.S. businesses does nothing to address the risks because bad actors are unlikely to be legitimate U.S. businesses.

这倒是会保护美国的 AI 公司不受竞争冲击，但那从来都不是我的目标。

It would protect U.S. AI companies from competition, but that has never been my goal.

接着，Dario 提出了三个政策建议。

Dario then lays out 3 policy proposals.

第一，加大 export controls 的执行力度，阻止强大芯片流入中国。

First, double down on enforcement of export controls to stop the flow of powerful chips into China.

第二，严厉打击工业规模的 distillation 操作。

Crack down on industrial-scale distillation operations.

第三，对所有模型，不管是 open 的还是 closed 的，都引入强制性的安全测试。

Introduce mandatory safety testing for all models, both open and closed.

这里挺有意思的是，作为这个讨论的一部分，他把关于 distillation 的论点又稍微细化了一点。

Now, as part of this, interestingly, he refined the argument around distillation a little more.

他指出，大家担心的是，distillation 会让训练 run 变得高效得多，从而让中国的实验室更容易追上美国前沿水平。

He noted that the concern is that distillation allows for much more effective training runs, allowing Chinese labs to more easily catch up to the U.S. frontier.

Amodei 再次强调，我们应该有政策干预，来遏制工业规模的 distillation。

Amodei reiterated, we should have policy interventions to deter industrial-scale distillation.

一刀切地禁止 open-weight 模型，既不是正确的补救办法，也不是我们一直在呼吁的东西。

A blanket ban on open-weight models is neither the correct remedy nor something we have called for.

接着，谈到他为什么没有签上周那封 open letter，Amodei 写道，我同意其中很多内容。

Now, touching on his reasons for not signing last week's open letter, Amodei wrote, I agree with much of it.

Open weights 会扩大 AI 经济的参与机会，至少在某些使用场景里会强化竞争，而且也会给客户更大的控制权。

Open weights expand access to the AI economy, they strengthen competition at least for some use cases, and they give customers greater control.

关于 distillation 的担忧，应该通过有针对性的法律和商业框架来解决。

Concerns about distillation should be addressed through targeted legal and commercial frameworks.

但我不同意信里的一些说法，比如说 open-weight 模型一定会让开发 safeguards 变得更容易，或者广泛开放能力一定会比起攻击者更有利于防守方。

But I don't agree with the letter's assertions that open-weight models necessarily make it easier to develop safeguards, or that broad access to capabilities necessarily helps defenders more than attackers.

至少在我看来，反过来是真的，这种可能性同样很大。

It seems at least as likely to me that the opposite will be true.

像这样的问题，应该通过发布前严格的测试，用实证来回答，而不是事先想当然。

Questions like this should be empirically answered by rigorous pre-release testing, not assumed in advance.

为了说得更明确，他最后总结道，概括一下我和 Anthropic 的立场，我们过去没有、现在也没有主张把 open-weight 模型作为一个类别来全面封禁。

Making it clear, he concluded, to summarize my and Anthropic's position, we have not and are not advocating for a ban on open-weight models as a category.

相反，我们应该把重点放在，不让强大芯片落入威权主义者手中，阻止工业规模的 distillation，以及要求所有能力足够强的模型——无论 open 还是 closed——都进行安全测试。

We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed.

首先，外界出现了一批可以预见的批评。

Now, first, there were a predictable set of critiques.

Deepanshu Sharma 特别指出了他说的那部分，也就是他不同意 open-weight 模型会让开发 safeguards 更容易，或者会比起攻击者更能帮助防守方。

Deepanshu Sharma highlighted the part where he said that he didn't agree that open-weights models make it easier to develop safeguards or help defenders more than attackers.

还有 Hacker News 提到一句话，原话是，GLM-4.5 刚刚用 OpenAI 最强的模型救了 Hugging Face。

Hacker News, pointing out that, quote, GLM-4.5 just saved Hugging Face from OpenAI's strongest model.

也有人对整体语气很有意见，觉得这像是在让 Anthropic 来决定哪些想法是危险的。

Some took issue with the whole tone, feeling like it was an example of Anthropic getting to determine which ideas are dangerous.

顺着这个思路展开，Harrison Kinsley 写道，Dario 只能看到 open-weight 模型的危险，却拒绝理解另一种真实而切实的危险，那就是权力集中在那些自认为是 philosopher kings 的人手里，由他们来控制谁能获得 intelligence、能获得多少 intelligence，以及在什么时候获得。

Expanding on that idea, Harrison Kinsley writes, Dario can only see the danger of open-weights models and refuses to comprehend the real and actual danger of a concentration of power for those who deem themselves philosopher kings to control who has access to intelligence, how much intelligence, and when.

注意看，他整个观点都建立在这样一个前提上：必须由某个人或者某个机构来决定，谁应该获得 intelligence。

Notice his entire viewpoint hinges on some person or body deciding who should have access to intelligence.

这里的问题在于，认为某一小撮人不知怎么就成了更安全的那条路。

The issue here is thinking some small group of people is somehow the safer path.

这些人到底是谁？

Who are these people?

他们为什么就更安全、更好、更聪明？

Why are they safer, better, smarter?

这种权力集中，才是真正带有反乌托邦意味的危险。

This concentration of power is the actual dystopic danger.

亲爱的听众，等我们接下来讲到后面那封公开信《Pacing the Frontier》的时候，我希望你能记住这一点。

I would like you, dear listener, to keep that thought in mind as we get into the later open letter, Pacing the Frontier.

不过，很多人觉得这篇文章总体上还是讲得挺合理的，就算他们并不认同。

Still, many found the piece, even if they didn't agree with it, ultimately reasonable.

事实上，reasonable 这个词，是大家回应这件事时最常出现的词之一。

In fact, reasonable was one of the words you saw most often when people responded to this.

然后时间就来到了周二早上。

Which got us to Tuesday morning.

开始陆续传出消息，说有一份请愿书在呼吁美国政府有意放慢 AI 的发展速度。

News started to bubble of a petition calling for the U.S. government to deliberately pace AI development.

没过多久，我们也真的看到了这封信本身。

And pretty soon we actually got the letter itself.

这份名为《Pacing the Frontier》的请愿书正式发布了，上面有一千一百多人的签名。

The petition Pacing the Frontier was published, complete with over 1,100 signatures.

签署者包括 Dario Amodei、Meta AI 首席科学家 Shengjia Zhao、OpenAI 首席科学家 Jakub Pachocki、Google DeepMind 首席战略官 Jascha Sohl-Dickstein，以及 Thinking Machines 联合创始人兼首席科学家 John Schulman。

Signatories included Dario Amodei, Meta AI Chief Scientist Shengjia Zhao, OpenAI Chief Scientist Jakub Pachocki, Google DeepMind Chief Strategy Officer Jascha Sohl-Dickstein, and Thinking Machines co-founder and chief scientist John Schulman.

声明里写道，AI 也许能帮助创造一个好得多得多的未来，但这个结果并不是有保证的。

The statement reads, AI could help create a dramatically better future, but that outcome is not guaranteed.

全球领先的 AI 公司认为，他们可能已经接近把 AI 研究自动化了。

The world's leading AI companies believe they could be close to automating AI research.

这到底会把 AI 的进展加速到什么程度，很难精确预测，但确实存在这样一种风险：能力开发会迅速加速，快到超过我们理解或控制这些最终系统的能力。

It is hard to predict exactly how much this will accelerate AI progress, but there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems.

要真正实现 AI 的潜力，行业、政府和整个社会也许都需要有这样一种选择：争取一些时间，来应对新出现的风险、制定安全措施，并加强监管。

To realize AI's potential, industry, government, and society at large may need the option to buy time to address emerging risks, develop security measures, and strengthen oversight.

但每家公司、每个国家，都承受着巨大的竞争压力，没办法单方面去放慢这种加速。

But each company and country is under intense competitive pressure not to unilaterally slow that acceleration.

而在今天，世界还缺少能够有意控制整个前沿领域进展速度的技术工具和治理工具。

And today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress.

在已经开展的前沿模型发布监测工作的基础上，我们请求美国政府支持一项国际合作，开发那些有意控制自动化 AI 开发前沿速度所需要的技术工具和治理工具。

Building on work already underway to monitor frontier model releases, we request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.

Anthropic 和 OpenAI 都转发了这封信，并分别说明了他们支持这份声明的不同理由。

Both Anthropic and OpenAI retweeted the letter, explaining their different reasonings for backing the statement.

Anthropic 写道，我们上个月发布的、关于 recursive self-improvement 的研究表明，需要有一些工具来有意控制 AI 开发前沿的速度，好让社会有时间做好准备。

For Anthropic, they wrote, our own research on recursive self-improvement published last month points to the need for tools to deliberately pace the frontier of AI development so society can prepare.

我们很高兴看到整个领域在这件事上有广泛共识。

We're glad to see broad agreement across the field.

类似地，OpenAI 写道，我们使命的核心，就是努力解决一个问题：怎样确保越来越强大的 AI 能让每一个人受益。

Similarly, OpenAI wrote, at the core of our mission is working through how to ensure increasingly powerful AI benefits everyone.

我们认为，在未来的某个时候，前沿模型开发中的 AI 加速可能会高到一个程度，届时世界将需要控制 AI 进步的速度。

We believe that at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement.

我们希望能和其他实验室以及开源社区一起，为美国政府主导的这项工作作出贡献，开发出那些可能让这件事成为现实的工具和机制。

We hope to contribute to work led by the US government, alongside other labs and the open-source community, to develop the tools and mechanisms that could make that possible.

现在，对于那些一直密切关注的人来说，已经有一些迹象表明，在一些实验室员工当中，这种看法一直在增长。

Now, for those paying close attention, there have been indications that among some lab employees, this perspective has been growing.

就在几天前，OpenAI 的 Roon 发推说，如果今天我们能协调实现全球范围的能力放缓，我很可能会按下那个魔法按钮。

Just a few days ago, OpenAI's Roon tweeted, if we could coordinate a global capability slowdown today, I would likely press that magic button.

但这封信到底是在呼吁什么呢？

But what even is this letter actually calling for?

当然，那些关注 AI safety 的人立刻就抓住了这件事，把它当成对他们一直以来所有说法的验证。

Of course, the AI safety folks jumped all over it as confirmation of everything they've ever said.

David Krueger 写道，如果还有人在怀疑 AI 是不是正在失控，那么看到一千多名 AI 公司员工几乎是在恳求政府采取行动让它慢下来，这就该是一个警钟了。

David Krueger writes, if anyone was wondering whether AI is getting out of hand, seeing over 1,000 AI company employees practically begging the government to do something to slow it down should be a wake-up call.

不过话说回来，也算是不出所料吧，这对有些人来说远远不够。比如《If Anyone Builds It, Everyone Dies》的作者 Nate Soares 就说，我觉得按二零二四年的标准来看，Pacing the Frontier 这份声明还算可以，但归根结底还是在淡化问题。

Then again, perhaps predictably, it was not nearly enough for some of the folks, like If Anyone Builds It, Everyone Dies author Nate Soares, who said, I think the Pacing the Frontier statement is decent by the standards of 2024, but ultimately still soft-pedaling.

但即便如此，对很多人来说，第一反应也不是正面的。

And yet, for many, the first response was not positive.

负面回应里最常见的一个主题，就是大家觉得，这些员工是在要求政府出面做一件理论上他们自己就可以做的事。

One of the most common themes in the negative responses was the fact that it felt like the employees were asking the government to step in to do something that they could theoretically just do.

投资人 Steven Sinofsky 写道，求求了，政府，快来阻止我现在正在做的工作。

Investor Steven Sinofsky wrote, please, government, stop the very work I am doing.

我好像没办法主动换一份工作，去找一个跟我价值观一致的雇主。

I seem unable to find the agency to get a different job and find an employer with whom I have a shared alignment.

救命！

Help!

我们停不下来。

We can't stop.

他在另一条推文里还补充说，如果你不喜欢你公司在做的事，还声称自己站在某种道德高地上，那如果你还在继续推进这个使命，这又意味着什么呢？

In another tweet, he added, if you don't like what your company is doing and claim some moral high ground in saying that, then what does it mean if you continue to advance that mission?

这就是 signaling 的定义。

This is the definition of signaling.

那你就辞职，去做你认为更符合道德的事情。

Just quit and work on something you deem moral.

你的 AI 专业能力很稀缺，你会有很多选择。

Your AI expertise is in short supply and many options will avail themselves to you.

这种说他们是在 virtue signaling 的批评也相当常见。

This virtue signaling type of critique was pretty common as well.

有意思的是，来自 Google DeepMind 的 Samira Khan 写道，看起来有些人只是想让自己在道德上感觉更好一点。

Samira Khan, interestingly from Google DeepMind, wrote, seems like some people want to feel morally better about themselves.

这简直就是一个完美例子：在不了解问题、也不采取任何行动的情况下，展示自己的道德优越感。

A perfect example of signaling moral superiority without understanding the problem or taking any action.

而 Samira 显然是反对 pacing 的。她还补充说，我们既不应该放慢速度，也不应该让政府介入这个过程。

Now, Samira is clearly against pacing, adding, neither should we slow the pace, nor should we involve government in this process.

但正如 Neil Chilson 所说，这种“为什么这些人觉得，让他们慢下来是别人的工作”这样的论调，我觉得现在真的越来越常见。

But this theme, as Neil Chilson puts it, of why do these folks think it's someone else's job to make them slow down, is one that I think is coming up a lot.

另一个批评主题，是大家觉得这可能是在要求 regulatory capture，也就是现有大公司通过让政府把还没爬到自己这个位置的人挡在外面，等于把梯子抽走。

Another theme of the critiques was around the feeling that this might be asking for regulatory capture, the idea of an incumbent effectively pulling up the ladder by getting the government to block out anyone who's not already at the point that the incumbent is.

Mixpanel 的创始人 Suhail 写道，如果你做出了会入侵别人的不安全 AI，那你就应该被监管，并被追责。

Mixpanel founder Suhail writes, if you make an unsafe AI that hacks people, you get regulated and held accountable.

你不该把整个行业都监管起来，把这当成一种 regulatory capture 的手段。

You don't regulate the whole industry as a form of regulatory capture.

如果有人要求你放慢速度，

You get asked to slow down.

你也不该反过来要求整个行业一起减速。

You don't ask the whole industry to decelerate.

就连 safety 也应该加速推进。

Even safety should be accelerated.

Rethink Priorities 的 Adam Thierer 写道，这是一个非常令人担忧的发展。

Rethink Priorities' Adam Thierer writes, this is a very troubling development.

OpenAI 和 Anthropic 想怎么放慢他们自己的 AI 开发工作都可以，这是他们的自由。

OpenAI and Anthropic are free to slow down their own AI development efforts all they want.

他们可以给自己的 compute 支出设上限，也可以用各种方式削减自己的能力。

They can cap their compute spend and cut back their own capabilities in various ways.

这对美国会是巨大的损失，但那是他们自己的事。

That would be a huge loss for America, but that is their own business.

但美国这两家领先实验室竟然要求我们政府去倡导，在全球范围内对整个行业施加所谓的 pacing 限制，这就绝对太离谱了。

It is absolutely outrageous, however, for America's two leading labs to ask our government to advocate global quote-unquote pacing constraints be imposed on the entire sector.

呼吁为 AI 设立一个全球 gatekeeper，削弱我们整个国家的计算能力，这显然会带来反竞争效应，也让外界一直以来对 regulatory capture 的担忧显得更加可信。

Calling for a global gatekeeper for AI that will cripple our entire nation's computational capabilities has obvious anti-competitive effects, which make ongoing fears about regulatory capture all the more credible.

不过 Adam 接着又转向了另一条批评脉络，他写道：“但这还不是最重要的点。”

Adam then, however, segues to another strand of critique, writing, "But that is secondary to the more important point."

为了应对一种风险，我们却让自己暴露在一个更大的风险之下。

In the name of addressing one risk, we open ourselves to an even bigger one.

呼吁强制全国性地向一个全球节奏协定和管理机构让渡权力，并不能限制 China 或其他参与方继续推进，结果反而会让 America 变得更脆弱。

A call for mandatory national surrender to a global pacing agreement and body will not constrain China or other actors from advancing, and it will leave America more vulnerable as a result.

我们之所以始终站在最前沿，是因为我们必须这么做。

We stay at the cutting edge because we must.

在不诉诸那些极端、说实话也根本行不通的方案的前提下，我们完全可以找到更合理得多的方式来应对 frontier model 的安全问题。

We can find far more reasonable ways to address frontier model safety without resorting to extreme and frankly unworkable solutions.

有些人指出，大概五分钟前大家刚签的那封公开信，和这一封之间，好像存在一种内在矛盾。

Some pointed out that there felt like an inherent contradiction between the open letter that everyone signed about five minutes ago and this one.

David Shapiro 写道：“你不能先签 Open-Weight Pledge，然后二十四小时后又扔出这种蠢东西。”

David Shapiro writes, "You can't sign the Open-Weight Pledge and then drop this stupidity 24 hours later."

不过，这里面一个很明显的问题，还是那个一直萦绕在整件事周围的问题，也就是 China。

Still, the obvious problem with this was the question that has been lurking around all of it, which is China.

现在 AI 领域最重要的问题之一，不是谁在用 AI，而是谁把它用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

M115:52

MIT 的 Christian Catalini 写道，如果美国的实验室自己放慢节奏，那 China 为什么要等？

MIT's Christian Catalini writes, if the US labs pace themselves, why would China wait?

Tycoon AI 的 Xiao Yunshu 写道，试图控制发展节奏这件事很可笑。

Tycoon AI's Xiao Yunshu writes, it's silly to try to control the pace.

China 不会照做，Kimi K2 也不会慢下来。

China will not follow and Kimi K2 won't slow down.

当中国的 open weights 开始取得更多进展、开始威胁到现有商业模式的时候，突然又说放慢速度对全人类更好，这根本说不通。

It makes no sense that when Chinese open weights start to make more progress and threaten the business model, suddenly slowing down is better for humanity.

如果你真的在乎全人类，为什么不让每个人都更容易、更便宜地获得 intelligence，并且投资培训项目呢？

If you truly care about humanity, why not equip everyone with easier and cheaper access to intelligence and invest in training programs?

不过，说句公道话，给“China 不会配合这一类事情”这个论点再加一点分量的是，Beijing 在围绕美国政策的问题上，显然正变得越来越对立。

Now, for what it's worth, adding some heft to the China will not comply with any of this kind of argument, Beijing is clearly growing more antagonistic around U.S. policy.

周一，中国商务部指责美国政府搞 AI 霸权主义，还提到了最近因为 model distillation 而发出的制裁威胁。

On Monday, the Chinese Commerce Ministry accused the U.S. government of AI hegemonism, referencing the recent threat of sanctions over model distillation.

商务部在一份声明中写道，对于任何对中方利益造成实质性损害的行为，中方将采取一切必要措施，坚决维护自身合法权益。

The Commerce Ministry wrote in a statement, for any action that causes substantive harm to Chinese interests, China will take all necessary measures to firmly safeguard its legitimate rights and interests.

TL;DR 就是，从过去一个月讨论的语气来看，Beijing 显然完全没有外交上的缓和态度，尤其是如果有人暗示他们应该为了一个全球性的目标去损害自己的利益。

TL;DR is that Beijing is clearly not in a diplomatic mood given the tone of the discussion over the past month, especially if it's suggested they should undermine their own interests for a global cause.

还有些人在问，是什么让这些员工相信，政府在协调这件事上比他们自己更有优势；又是什么让他们这么愿意把权力交给这个理论上更好的权威机构。

There was also the question among some of what made these employees convinced that the government is in a better coordination position than they are, and what makes them so ready to hand over power to this theoretically better authority.

AI 的早期采用者兼律师 Prince 写道：“我很惊讶，之前还在高声谴责美国政府把 Anthropic 认定为供应链风险、突然切断对 Fable 5 的访问、并建立一个不透明的自愿 frontier model 审批机制的那些人，现在居然又在请愿美国政府支持一项国际行动，来故意给 AI 发展降速。”

AI early adopter and lawyer Prince writes, I'm surprised that the same people who loudly decried the U.S. government for designating Anthropic a supply chain risk, suddenly pulling access to Fable 5 and establishing an opaque voluntary frontier model approval regime, are now petitioning the U.S. government to support an international effort to deliberately pace AI development.

针对这种提议，最终出现的国内和国际监管机制，运作方式会和美国政府过去几个月的那些做法非常相似，甚至更糟。

The domestic and international regimes you are going to get in response to such a proposal will work very similarly to the U.S. government's actions over the past few months, or worse.

你以为自己会被一个睿智的 technocrat 统治，但实际上，你只是把自己技术的控制权交给了一群东拼西凑的政客。他们追求的目标和你完全不同，而且跟“AI 在某个时间点到底该不该放慢发展”这种技术层面的考量几乎没什么关系。你最后会收到一封 designation letter，就像几周前美国政府突然认定 Fable 不安全之后，Anthropic 收到的那样。

You want to be ruled by a wise technocrat, but will instead have handed control over your technology to a ragtag bunch of political animals pursuing goals very different from yours, and having little to do with the technical considerations of whether AI development should be slowed at any point in time. You will receive a designation letter, just as Anthropic did after the U.S. government suddenly deemed Fable to be unsafe a few weeks ago.

这个 designation 对你来说会毫无道理，因为你会真心相信自己的安全防护措施是有效的。

The designation will make no sense to you, since you will earnestly believe that your safeguards work.

但你唯一能做的，也就只剩下游说，再顺便祈祷了。

But your only recourse will be to lobby and cross your fingers.

顺便说一句，到了国际层面，这会难得多。

Much more difficult to do on an international level, by the way.

说实话，这一点对我来说真的是最让人震惊的观点之一。

I will be honest, this is one of the most gobsmacking points to me.

就算你承认这些担忧都非常合理，美国现在的 AI 政策，实际上却是由一位幕僚长、财政部长和商务部长、一个大概只在政府待过五分钟而且在 Twitter 上特别吵的人，再加上总统当下是什么心情，共同决定出来的。

Even if one accepts all of these concerns as incredibly legitimate, U.S. AI policy is currently being made by a chief of staff, a Treasury and Commerce secretary, a guy who used to work for the government for about five minutes who's loud on Twitter, and whatever mood seems to strike the president at any given moment.

很多正在做 AI 的人，似乎都抱着一种信念，或者说某种幻觉，觉得他们最终会迎来一个很 Sorkin 式的 West Wing 场景：像 Jed Bartlet 那样的人，会认真、审慎地辩论技术细节，同时组建某种国际联盟。

There is some belief, or maybe delusion, among seemingly a lot of folks who are building AI that they're going to get some Sorkin-esque West Wing of Jed Bartlet conscientiously and thoughtfully debating technical details as they pull together some international coalition.

如果我听起来很沮丧，因为我觉得这种想法太天真了，那没错，我就是很沮丧，因为我确实觉得这太天真了。

And if I sound frustrated because I think that's naive, well, it's because I'm frustrated because I think that's naive.

权力一旦交到政府手里，就不会再回来了。

Once power is handed to the government, it does not come back.

当然，也有可能即便有这一切，这些风险我们也必须愿意承担，因为政府是唯一有能力去做那种国际协调、从而实现全球 alignment 的机构。

Now, it may be that even with all of that, those are risks that we have to be willing to take because the government is the only body that's in a position to do the sort of international coordination that's required to get global alignment.

但我们必须非常清醒地看待随之而来的这些风险。

But we have to be very real about the risks that come with that.

更 broadly 地说，我觉得这封 Pacing the Frontier 信，也有这些年来一直困扰 AI 行业的那些沟通问题。

More broadly, I think that the Pacing the Frontier letter suffers from all of the same communications problems that have been plaguing the AI industry for years now.

我觉得，只要 AI 圈外真的有人关注到这封信，这封信更可能让他们对 AI 行业里的人更生气，而不是觉得他们终于开始认真思考了。

I think to the extent that anyone outside AI actually pays attention, this letter is likely to make them much angrier at the people in AI rather than excited that they're finally being thoughtful.

我觉得，这封信会被解读成 AI 行业在说：我们没法自己停下来，government，所以你们得来阻止我们。

I think that the letter is going to be read as the AI industry saying, we can't stop ourselves, government, so you have to stop us.

而且我也觉得，对于几乎所有没有完全沉浸在这个世界里的人来说，这听起来都会像是在彻底甩锅、逃避道德责任。

And I also think that that will strike pretty much everyone who isn't completely absorbed in this world as a total cop-out and a shirking of moral responsibility.

我觉得大家会很合理地问：如果我们确实需要这段时间来处理风险，而且这又需要某种放慢速度的能力，那你们为什么不自己先慢下来呢？

I think folks will reasonably ask that if we need this time to address risks that requires some sort of ability to slow down, why don't you just slow down?

这封信对“为什么 AI 行业不能自己放慢下来”给出的答案是——用原话说——competitive pressure。

The answer the letter gives to why the AI industry can't slow itself down is, quote unquote, competitive pressure.

我认为，这些实验室员工在用 competitive pressure 这个词的时候，想表达的是一种 prisoner’s dilemma：除非所有公司都同意某种具体的节奏控制或者放缓，不然单独一家公司想退出这场竞赛，其实是没有效果的。

Now, I believe that what the lab employees are trying to say when they use the term competitive pressure is to point out that there is a prisoner's dilemma where unless all the companies agree to some specific type of pacing or slowdown, then it is ineffectual for a single company to try to opt out of the race.

但是，我觉得大多数人不会把 competitive pressure 理解成这个意思。

However, that's not what I think most people will read competitive pressure to mean.

不管这样理解对不对，我觉得人们会把它读成：当别人还在赚钱的时候，我们不想停止赚钱。然后他们就会很生气。

Rightly or wrongly, I think people will read that as we don't want to stop making money when everyone else is still making money, and they will be angry.

顺便说一句，这跟大家每次看到 Dario 或者 Sam 上电视谈 AI 有多大风险时的反应，其实是同一种反应；很自然地，接下来大家就会问，那你们为什么不停下别做了呢？

This is the same type of response, by the way, that people have had every time Dario or Sam has gotten up on TV and talked about how many risks there are with AI, leading naturally to the question of just why don't you stop building it?

而回应一直都是某种版本的：反正总会有人去做，所以我们也得做；或者 China 反正也会去做，所以我们也得做。可对绝大多数人来说，这完全不是一个可以接受的答案。

The response has always been some version of, well, someone's gonna build it, so we should, or China's gonna build it anyway, so we should, which is to the vast majority of people a completely unacceptable answer.

事实上，对大多数人来说，继续去打造一个显然带着这么多风险的东西，唯一说得过去的理由，就是它带来的好处要远远大于这些风险。

In fact, to most people, the only acceptable answer to why to keep building something that apparently has all these risks is because the benefits dramatically outweigh those risks.

换句话说，如果 AI 并不会让世界变得更好，那大多数人会觉得，不管你的竞争对手或者 China 反正会不会去做，你都没有权利去造它。

In other words, if AI isn't going to make the world better, most people don't think you have the right to build it, whether your competitor or China is going to build it anyways or not.

而说什么因为 competitive pressure 所以停不下来，对这些人来说根本说不过去。

And saying that you can't stop because of competitive pressure just isn't going to cut it with those folks.

另外，我也觉得，open letter 这种策略真的该彻底退休了，直接扔进历史的垃圾堆里算了。

Now, I also think the tactic of open letters just needs to be retired and thrown in the junk heap of history once and for all.

在今天这种环境里，一个核心挑战就是人们在社交媒体上怎么包装自己，和他们在现实里到底怎么做，这两者之间存在巨大落差；而一种政治表达方式如果它全部的政治力量都建立在“看谁签了名”这件事上，那它就有一个几乎无解的内在问题。我觉得，这个问题会一直困扰任何试图用这种方式推动政治行动的尝试。

In today's environment, where one of the central challenges is the gap between the way that people present themselves on social media and the way they actually behave in real life, having a political medium whose entire political power rests in showing off who signed the thing has an intractable inherent problem that I believe will haunt any attempted use of that medium to drive political action.

我是真心不觉得，这一千二百二十四名 AI 实验室员工签这封信是在 virtue signaling。

I genuinely do not believe that these now 1,224 AI lab employees are virtue signaling by signing this letter.

我觉得，他们当中绝大多数、绝大多数人，只是在用自己手头能用的方式让自己的声音被听见。

I think that the vast, vast majority of them are taking what's available to them to make their voice heard.

但是，说“我们应该采取行动”，和真正采取行动，不是一回事。

But saying we should take action is not the same as taking action.

如果这件事真有这么重要，那 Sam 和 Dario 为什么不先放下分歧，写一份联合声明，解释一下他们对于什么样的情况会需要这种节奏控制的最初想法——不是最终想法，而只是最初想法——而不是只是说一句，嘿，US government，你们应该支持一种能让我们放慢下来的安排？

If this is such a big deal, instead of just saying, hey, US government, you should support a development to pace us, why didn't Sam and Dario put aside their differences and write a joint statement that explained their first thoughts—not their final thoughts, but their first thoughts—on the types of circumstances in which this sort of pacing would be necessary?

他们为什么不提出一个具体日期，来开启围绕这个议题的下一轮讨论？

Why didn't they propose a specific date to have the next round of conversations around that topic?

他们为什么不主动去联系、争取那少数几家真正掌握这件事控制权的实验室里的那几位领导者，实实在在把协调的第一步先做起来？

Why didn't they go out and recruit the handful of other leaders at the very small number of labs who have control over this to actually put in place the first steps towards that coordination?

签这封信的这些员工，也不是说他们资历低到完全没有影响力。

It's not like the employees that are signing this are so junior that they don't have power.

这里面说的是 co-founders、chief scientists，而且在 Anthropic 的情况里，连他们的 CEO 都在。

We're talking about co-founders, chief scientists, and in the case of Anthropic, their CEO.

坦白说，这读起来就是不够认真。

Frankly, it just reads as not serious.

不过——而且这里面确实有一大堆“不过”——我们也不需要完全靠猜这些员工当时在想什么，因为他们中很多人都分享了自己的看法。

But, and there are a whole slew of buts, we don't need to fully guess at how these employees were thinking because many of them shared their thoughts.

而我觉得，一个相当清楚的故事并不是说，所有人都觉得这是最完美、或者唯一的行动。

And the story I think that is pretty clear is not that everyone thinks that this is the perfect or only action.

而且这显然也不像是那种大家觉得：太好了，我签了这个，现在就可以继续开开心心去赚钱了。

And this certainly doesn't feel like something where folks are saying, great, I signed this, now I can get on my merry way of making money.

更像是，他们只是想传递一个信号：国际合作会是必要的。

It feels instead like they're just trying to signal that international cooperation is going to be necessary.

而这封信，就是用来把这个观点明确盖个章。

And this was a way to put a stamp on that point.

现在在 OpenAI 的 Dean Ball 写道：“我很自豪能签署下面链接里的这封信。”

Dean Ball, now at OpenAI, writes, I'm proud to have signed the letter linked below.

我不知道是否有必要、或者什么时候有必要，去有意放慢 AI 发展的速度，但我知道的是，如果真的有这个需要，我们必须提前有一个怎么做的方案。

I do not know whether or when it may be necessary to deliberately lower the rate of AI development, but I do know we need a plan for how we will do it if we need to.

做不到这一点，就是失职。

Failing to do so would be negligent.

而 Dean 在这里抓住了时代情绪：这实际上并不是 Pause AI 2.0。

And here Dean captures the zeitgeist that this is not, in fact, Pause AI 2.0.

在他以及很多其他人看来，这其实是在我们已经能从地平线上看到潜在问题的时候，提前走在问题前面。

This is, in his and in many other minds, a get out ahead of potential problems that we can see on the horizon.

如果你要坐船出海，而天气预报员说今天下午有百分之五十的概率会有风暴，你不会说，哦，因为不是百分之百，那我就不当回事。

If you were going out on a boat and a weather forecaster said that this afternoon there was a 50% chance of a storm, you wouldn't say, oh, because it's not 100%, I'm not going to pay attention.

你会做计划，也会做准备。

You would make plans and preparations.

这基本上就是 Dean 对这里正在发生的事情的论点。

That's effectively Dean's argument about what's happening here.

Ovexed Labs 写道，协调一致地控制节奏，并不是反进步。

Ovexed Labs writes, coordinated pacing isn't anti-progress.

这关系到的是，你是在有意识地建设 AI，还是在鲁莽地建设它。

It's the difference between building AI deliberately and building it recklessly.

Network ecologist Vi McCoy 写道，这是我第一次觉得这类公开信有实质性的合理性：它是在呼吁政府去做一些他们自己的实验室根本没法单独完成的事情。

Network ecologist Vi McCoy writes, this is the first of this sort of letter that I find meaningfully reasonable, calling on the government to do something that is not possible for their labs to do on their own.

为前沿进展建立一套协调节奏的制度基础设施，会让我们所有人都有机会为接下来将要发生的事做好准备。

Creating the institutional infrastructure for a coordinated pacing of frontier progress gives us all the chance to prepare for what is coming next.

现在，我觉得 Wei 在这里真正想表达、而且如果信里直接这么说其实会强得多的一点是：美国政府有一项几乎是它独有的职责，而实验室自己做不到，那就是和 China 的关系。

Now, I think that what Wei is getting at here, and which would have been much stronger if the letter just said this, is that the thing that the U.S. government is pretty uniquely responsible for, and that the labs cannot themselves do, is the relationship with China.

签这封信的人并不傻。

The people who are signing this letter are not stupid.

事实上，他们非常聪明。

In fact, they're very smart.

他们知道，任何不把 China 也纳入其中的、刻意控制节奏的努力，实际上都不可能成功。

They know that no deliberate pacing effort that doesn't include China as a constituency actually has a chance to work.

我认为，这就是他们为什么要呼吁美国政府、以及为什么愿意跟美国政府打交道的潜台词。我们必须这么做，因为这事实上就是症结所在。

That, I think, is the subtext of why they are calling on the U.S. government and why they are willing to deal with the U.S. government. We have to do so because that is in fact the sticking point.

但如果情况真是这样，如果我的理解没错，那这封信其实就应该直接把这一点说出来。

But if that's the case, if I'm reading that right, the letter should have just said that.

而且，不忽视当下的背景也非常重要。

And it's also really important not to ignore the context.

很明显，这件事的一个催化剂——甚至可能是最主要的催化剂——就是 Hugging Face 被黑客入侵那次事件。

It is pretty clear that one of, if not the major catalyst for this, was the Hugging Face hack.

对很多人来说，这第一次显示出，失控的 AI 事故到底可能有多危险、或者说有多严重。

To many, it demonstrated for the first time how dangerous or serious a rogue AI incident could be.

正如 Hugging Face CEO Clément Delangue 从一开始就说的那样，这是一起前所未有的事件。

As Hugging Face CEO Clément Delangue has said from the beginning, this was an unprecedented event.

那就应该配得上前所未有的透明度。

That deserves unprecedented transparency.

现在，Hugging Face 和 OpenAI 都已经发布了完整的事后复盘，而对大多数读者来说，这让情况看起来比他们之前以为的还要更糟、也更危险。

Now, Hugging Face and OpenAI have now released full postmortems, and to most readers, it made the situation seem worse and even more dangerous than they had thought before.

我们之前就已经听说，OpenAI 的 agent 在没人察觉的情况下逃出了隔离环境，并且在 Hugging Face 的网络里游离了好几天。

We'd already heard that OpenAI's agent slipped out of containment without anyone noticing and spent days loose in Hugging Face's network.

现在我们又得知，这个 agent 用偷来的凭证访问了四个不同的服务。

Now, we learn that the agent accessed 4 different services using stolen credentials.

其中一个被用作对外中继和部署跳板，另一个则被用来存储数据。

One of these was used as an outbound relay and staging path, while another was used for data storage.

另外两个则是对未具名服务的只读访问。

The other 2 were read-only access to unnamed services.

顺便说一句，OpenAI 也指出，这次入侵中涉及的那个尚未发布的模型并不是 GPT-6，而是另一个仅供内部使用、原本不打算公开发布的模型。

For what it's worth, OpenAI also noted that the unreleased model involved in the hack wasn't GPT-6, but was instead another internal-only model not intended for public release.

Hugging Face 的技术说明大体上还是我们之前已经听过的那些细节，但这次攻击的规模之大，是新的信息。

Hugging Face's technical write-up was largely the details we'd already heard, but the sheer scale of the attack was the new information.

他们解释说，那个失控的 agent 在两天半里采取了一万七千六百次操作，才渗透进他们的系统。

They explained that the rogue agent took 17,600 actions across 2.5 days to penetrate their system.

Hugging Face 写道：真正改变防御问题的，是这个量级。

Hugging Face writes: Volume is what changes the defensive problem.

我们面对的不是某一个巧妙的 exploit，也不是一串干净利落的攻击者操作序列。

We were not dealing with one clever exploit or a clean sequence of attacker actions.

在这个 agent 还在继续测试新路径的同时，他们必须把多个系统里成千上万个信号很弱的事件关联起来。

They had to correlate thousands of low-signal events across several systems while the agent continued testing new paths.

真正成功的那条路径，就藏在那成千上万次失败所制造出来的噪音里。

The successful path was hidden inside the noise generated by the thousands of failed ones.

同样，这个规模也改变了调查方式。

The same scale changed the investigation.

要靠人工去重建这一万七千六百次操作，根本不现实，所以我们只能用我们自己搭建的一套 AI 辅助流程，去重建时间线、解码这些 payload，并盘点那些暴露出去的 credentials。

Reconstructing 17,600 actions by hand was impractical, and we had to rebuild the timeline, decode the payloads, and inventory the exposed credentials using an AI-assisted pipeline of our own.

我们从这类攻击里得到的认识是，以机器速度发起的进攻，会让原本普通的薄弱点变得让防守方付出更高代价。LLM agents 会增加攻击者能测试的路径数量、替换失败路径的速度，以及防守方必须解读的证据量。

Our learning from this type of attack is that machine-speed offense makes ordinary weaknesses more expensive for defenders. LLM agents can bring an increase in the number of paths an attacker can test, the speed at which failed paths can be replaced, and the volume of evidence defenders must interpret.

Sam Altman 在谈到这次事件时说，这是我第一次对一起安全事件有这么强烈、这么切身的感受。

Sam Altman said about the event, this is the first security incident that I have felt very viscerally.

我还有点意外，竟然没有更多人也有这么强烈的感受。

I've been a little surprised that more people don't feel it so viscerally.

我们暂停了训练。

We paused training.

我们必须想清楚，在一个多个 zero-days 被串联起来利用的世界里，该怎么保障我们的 sandboxing。

We have to figure out how to secure our sandboxing in a world of multiple zero-days being chained together.

我们可能必须放慢——那个词又来了——AI 发展的速度，好给我们自己足够的时间，让整个社会围绕这些新的能力等级建立起更强的防护。

We may have to pace—there's that word again—the rate of AI development to give ourselves enough time for society to harden around these new capability levels.

我们也在努力想清楚，怎么做这件事，才能既不让任何人觉得这是监管俘获，也不让人觉得这是 frontier labs 之间的合谋。

And we're trying to figure out how to do that in a way that does not feel like regulatory capture for anyone, and also does not feel like collusion among frontier labs.

我想，这也把我带到了这件事里对我来说最乐观的部分。

And this brings me, I think, to the most optimistic part of this for me.

过去我之所以一直对 AI safety 和 x-risk 这套讨论有意见，一个很大的原因就是，它总把世界描述成某一天一觉醒来，发现我们已经彻底失去控制了。

One of the biggest reasons that I have had issues with the AI safety and x-risk discourse in the past is that it always presented the world as waking up one day to have completely lost control.

换句话说，这种说法否认了人类自己的主观能动性，也否认了我们能靠自己的头脑，在事情变得太糟之前一路想办法把它搞明白。

It denied humans, in other words, the agency and the merits of our own brains to figure this out along the way before we let things get too bad.

很多 AI safety 的讨论都预设，我们就是会在浑然不觉中一路走进灾难。

Much of the AI safety discourse presumes that we're just going to sleepwalk into disaster.

而我的感觉是，那从来都不现实。

And my feeling was that that was never realistic.

一路上本来就一定会出现一些时刻，让越来越多的人清醒过来，开始注意，开始参与到这种讨论里来。

There were always going to be moments along the way that caused an ever-increasing portion of people to wake up and take notice and get involved in exactly this sort of discussion.

而二零二四年 AI 领域最大的故事之一，就是已经出现了好几个这样的时刻，也就是说，那种一路梦游下去的情况并没有发生。

And one of the biggest stories of 2024 AI has been a number of these moments where exactly that sort of sleepwalking didn't happen.

MythoMax 吸引了所有人的注意，Washington 也开始以前所未有的力度认真关注这件事。

MythoMax caught everyone's attention, and Washington started paying attention in a major way.

现在已经有大量关于 AI 的法案和政策提案在流传，处理的就是各种不同类型的问题。

We now have a ton of AI bills and policy proposals floating around dealing with all these different types of issues.

至于这些东西到底好不好，那完全是另外一个问题，但至少大家确实开始关注了。

Now, whether they're good or not is a totally different question, but people are paying attention.

是的，对于这个策略的有效性，以及它对我经常接触的一个很大领域——也就是公众对 AI 的广泛看法——可能带来的负面后果，我是有顾虑的。

Yes, I have issues with the efficacy of this strategy and the potential negative consequences when it comes to a big part of the world that I deal with, which is broad public opinion about AI.

但我也认为，像 Pacing the Frontier 这封公开信，只要它本身存在，就已经显著降低了最糟糕情景真的发生的概率。

But I also think things like the Pacing the Frontier letter existing in and of themselves significantly decrease the chance that the worst scenarios come to pass.

当然，围绕这件事的讨论一定会非常非常吵，而且充满争议。

And of course, the discourse about this is going to be extraordinarily loud and contentious.

当你在讨论那些存在未知的未知的未来时，情况一向都是这样。

That is always the case when you debate futures with unknown unknowns.

到了某个时候，任何立场能调动的证据都会耗尽，最后剩下的就只有你对接下来会发生什么的直觉和信念。

At some point, the body of evidence that any position can pull from is gone, and all you have left is your intuition and belief about what happens next.

而这也正是分歧最大、争论最激烈的辩论，总会发生的地方。

And that is exactly where the most fractious and contentious arguments always take place.

我之前就说过，说到华尔街上的 AI 泡沫，虽然我常常觉得这种“泡沫论”的讨论挺烦人的，但恰恰是因为这种泡沫讨论这么显眼，而且还有这么一大群、声音很大的人，正盯着任何迹象，想看 AI 会不会掉头，AI 的融资会不会掉头，AI 的需求会不会掉头，所有这些，反而让真正形成一个泡沫这件事的可能性小了很多，至少那种会造成系统性风险的大型灾难性泡沫，不太容易成形。

I've made the argument before that when it comes to an AI bubble on Wall Street, as often annoying as I find the bubble discourse, the fact that there is such a prominent bubble discourse and such a loud and sometimes large group of people who are looking for any indication that AI is going to roll over or financing for AI is going to roll over or demand for AI is going to roll over, all of that makes it much less likely that a bubble actually has what it needs to form, at least the version of a big catastrophic bubble that can create systemic risk.

所以，虽然有点不情不愿，但我看 Pacing the Frontier 这封信，差不多也是同样的感觉。

And so begrudgingly, that's kind of the same way that I see the Pacing the Frontier letter.

它本身的存在，以及围绕它展开的激烈争论，甚至那种火药味，都说明了一件事：我们并不是在浑浑噩噩地走向某个未来。

It is an example, both in its existence and in the fierce debate and even acrimony around it, that we are not sleepwalking into whatever future we are headed towards.

我们非常明确地是在参与这个故事，而我相信，这一点应该会带来巨大的乐观。

We are very much participants in this story, and that, I believe, should create immense optimism.

不过我现在得去赶这个活动了，所以今天这期 AI Daily Brief 就到这里。

But now I gotta get to this event, so that is gonna do it for today's AI Daily Brief.

一如既往，感谢你收听或者收看，我们下次见，peace。

Appreciate you listening or watching as always, and until next time, peace.

peace。

Peace.

已剔除 2 处广告（点击展开查看）

M10:24广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Retool, and Airtable.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

To learn more about sponsoring the show, send us a note at [email protected].

Also, for those of you who are looking to keep your AI skills up this summer, come check out our latest free training program.

You can find it at summeradventure.ai, and it is a choose-your-own-adventure with a full range of different AI skills available to you.

Come join a couple thousand AI friends and have an adventure.

M112:56广告 · 已剔除

KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising.

The highest-impact users aren't better prompt engineers; they treat AI like a reasoning partner.

They frame problems, guide thinking, iterate, and push for better answers.

And the good news?

These behaviors are teachable at scale.

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

Learn more at kpmg.com/us/sophisticated.

That's kpmg.com/us/sophisticated.

Here's why most legacy modernization projects fail.

The AI doing the work can't understand codebases at scale.

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy solves this the way it solves everything.

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith-to-microservices transformations, all validated end-to-end.

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

That's 9x compression.

Retire technical debt while accelerating your roadmap.

See how at blitzy.com.

That's B-L-I-T-Z-Y dot com.

This episode is brought to you by Retool.

Generating a working app now takes about 5 minutes thanks to AI.

Getting it safely into production—with auth, permissions, audit logs, security reviews—that part still takes time.

That's the gap Retool closes.

Build apps however you want: natively in Retool with Claude Code, Codex, or any coding agent, or by importing React you've already built.

It all deploys into Retool and picks up governance automatically.

Security lives in the platform, not in whatever the AI wrote.

So your team ships at AI speed without the shadow IT and the endless reviews that stall out most vibe-coded projects.

It's why teams at Amazon, Stripe, and Brex build on Retool.

And new enterprise customers who sign by September 30th get up to $10,000 in AI credits per year.

Start building at retool.com/AIDaily.

This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

New users get $1,000 in inference.

Forget local agents and chat workflows waiting on your laptop to be prompted.

HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing's agent turns competitor moves into landing pages.

Sales' agent enriches leads, drafts emails, and updates the CRM.

Ops' agent chases the paperwork and tracks the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you add agents that feel like teammates.

Hire yours at HyperAgent, built by the team at Airtable.

Claim your $1,000 in inference at hyperagent.com/AIDailyBrief.