IBM的AI过山车，德米斯呼吁设立AI监管机构，纽约暂停AI数据中心 | Diet TBPN

IBM简直崩了。

IBM is absolutely nuking.

股价跌了百分之二十五。

The stock is down 25%.

完了。

Boom.

IBM。

IBM.

这走势真是够疯狂的。

Well, that is a crazy chart.

这是什么情况？

What is that?

这是过去一周的图表吗？

The that's the one week chart?

要是看五年图，情况就好多了，自从ChatGPT推出后，股价其实涨了不少。

It looks better on the five year because the stock is actually way up in the AI era since the launch of ChatGPT.

IBM 这段时间表现得非常好。

IBM has done really, really well.

自从 ChatGPT 推出、AI 时代开启以来，IBM 的股价几乎翻了一倍。

The stock has basically doubled since the introduction of ChatGPT during the AI era.

你打算当赢家，还是当输家？

Are you going to be a winner or a loser?

你打算被碾压、被甩开，类似这样的情况吗？

Are you going to get steamrolled, slopped, something like that?

但直到今天之前，它一直表现不错，直到公司重新调整了关于其服务器业务的叙事。

But it's been doing well up until today when the company reset the narrative around their server business specifically.

用布拉德·格斯特纳和加文·贝克的说法，IBM 在 AI 产业链里没占到优势，主要是因为目前的 AI 投资都流向了 GPU、内存、网络、超大规模云计算和前沿模型推理。

So the high level reason that IBM is not well positioned in the token path to use the Brad Gerstner and Gavin Baker parlance is that AI spending is currently flowing into GPUs, memory, networking, hyperscale cloud computing, and frontier model inference.

IBM 在这些领域并不是主要赢家，所以咱们简单回顾一下 IBM，这是一家名字很响亮、业务挺有意思的企业，全称是国际商业机器公司。

IBM is not a major winner in those So just to refresh on IBM because interesting business with a great name, International Business Machines.

他们最早生产的是打孔卡片系统。

The first business machines they made were punch card systems.

他们生产过钟表。

They made clocks.

这就像在经营一家公司一样。

Like, it's like you're running a business.

你需要一台机器。

You need a machine.

需要一块精准的钟表。

Need a great clock.

你得有个钟表。

You're gonna need a clock.

不是的。

No.

你做的是特别难做的生意。

You have really worst hard business.

不是随便什么钟表都行。

Not just any clock.

就像一个特别准的钟表。

Like a clock that works really well.

没错。

That's right.

专业级的钟表。

Professional clock.

钟表Pro Max。

Clock Pro Max.

你正在

You on

点上。

time.

正是如此。

Exactly.

钟表Pro Max。

Clock Pro Max.

滑块。

Slider.

它最轻、最薄，外观也最漂亮

It's the lightest, thinnest, best looking

速度最快。

fastest.

你想要一台速度快的钟。

You want a fast clock.

但那些制表机，本质上是各种机械处理信息的方式。

But tabulating machines, basically a bunch of different ways to process information mechanically.

而这个根本性的洞察，其实很简单。

And that foundational insight, you know, was pretty simple.

企业会一直愿意花钱来自动化记录保存。

It was businesses will continually pay forever to automate record keeping.

从宏观上看，这种模式已经持续了很长时间，而且还在不断

And at a high level, that's sort of been working forever, and they're continuously

你有没有想过，如果有人把时钟当成SaaS产品来卖，把时间变成一种服务？

know if they ever tried to sell a clock as a SaaS product, time as a service?

如果

If

你要是使劲儿盯着红帽的Kubernetes看，它其实就是在协调分布式系统之间的时间同步。

you really, really squint Red Hat Kubernetes, it's keeping time between distributed systems.

也许这里面真有门道。

Maybe there's something there.

但当你在多台服务器上运行数据库时，时间同步确实是个关键环节。

But when you're running a database across a bunch of different servers, there's some timekeeping aspect that's important.

不过，我觉得他们从来没这么干过。

But, no, I don't think they ever did.

人们现在熟悉的IBM，也就是大型机业务，是1964年才起步的。

The IBM that people know, the mainframe business, that started in 1964.

System/360，那是一整套兼容的设备，这点挺有意思的。

System three sixty, it was a compatible family of devices, which is interesting.

人们以为主计算机只有一种，但实际上它是一整套不同的系统，你可以逐个升级，而不用重新设计整个工作流程。

It's not just one people think one mainframe, but it was actually a whole bunch of different systems that you can upgrade piecemeal without redesigning the entire workflow.

如果需要更多存储空间，就升级存储部分。

So you need a little bit more storage, you upgrade that.

如果需要更多计算能力，就升级计算部分。

You need a little more compute, you upgrade that.

这使得IBM成为企业计算领域的主导供应商，覆盖了银行、保险公司、航空公司、制造商和政府机构。

And this turned IBM into the dominant supplier of corporate computing banks, insurers, airlines, manufacturers, governments.

这些机构都把IBM当作它们硬件和软件的核心系统。

They all used IBM as the central system for their hardware and software.

这就是主计算机的时代。

This was the mainframe era.

IBM之所以在主计算机领域占据主导地位，关键在于它专注于高可靠性、长期客户关系，以及高昂的切换成本。

And the whole reason that IBM in particular became dominant in mainframes was they focused on high reliability, long customer relationships, expensive switching costs.

一旦你进入了IBM的生态系统，想要退出就非常困难，因为软件和硬件是深度绑定的。

It's very difficult once you're in the IBM ecosystem to weed your way out proprietary software tied to the hardware.

有些软件只能在IBM的硬件上运行，根本没法迁移到别的平台。

Certain software would only run on IBM hardware, so you couldn't replatform.

你得把整个系统推倒重来，这在六七十年代对大型银行或航空公司来说简直难如登天。

You had to rip everything out, which is very difficult for a large bank or a large airline in the sixties and seventies.

这生意做得真不错。

Good for business.

他们还附带了庞大的技术支持和咨询服务合同，覆盖所有软硬件，某种程度上就像福特派现场工程师的前身，你稍微想象一下就明白了。

And they also had huge huge support and consulting contracts associated with all the software and the hardware that they were delivering, sort of a precursor to the Ford deployed engineer if you squint a little bit.

但个人电脑时代的到来才是真正的转折点。

But the PC era was the real turning point.

IBM个人电脑是在1981年推出的。

So the IBM PC launched in 1981.

这彻底认可了个人计算市场，也让英特尔和微软这两家公司抓住了接下来这场计算浪潮的巨大机遇。

This legitimized the personal computing market and set up two new companies, Intel and Microsoft, to capture immense value during the next computing boom.

IBM个人电脑运行的是Windows系统，用的是英特尔的芯片。

So the IBM PC ran Windows and used an Intel chip.

那时候，IBM的年收入高达三百亿美元，英特尔还不到十亿，微软的销售额只有一千七百万。

At the time, IBM was doing 30,000,000,000 in revenue, Intel was doing less than 1,000,000,000, and Microsoft was only doing 17,000,000 in sales.

我记得微软当时大概就一百二十号人，结果突然之间，这两家公司变得巨大无比，大了十倍都不止。

And so I think Microsoft had like a 120 employees, and all of a sudden, two companies became ultimately way, way, way bigger, like 10 times as big.

后来市场开始分化，有人开始提议把IBM拆分成几家独立公司。

So the market eventually fractured and proposals to break IBM into separate companies started to pop up.

市场之所以分化，是因为有了英特尔的芯片和Windows系统之后，你可以在不同芯片上运行Windows，也可以用别的芯片搭配别的操作系统。这时候，其他PC厂商纷纷涌入，苹果更是打出‘挑战权威’的口号，直接跟IBM对着干。

The market fractured because once you had, you know, an Intel chipset and and Windows operating system, you could run Windows on a different chipset and you could you have a different chipset with different operating system and the value capture piece, there were just other PC manufacturers that came in and then obviously Apple with their, you know, anti IBM, like, challenge the man campaign.

市场不断分裂，整个八九十年代，不断有人提出要把IBM拆成几个独立部门。

So the market was fracturing and there was a bunch of proposals during the eighties and nineties to break up the company into separate units.

1993年上任的CEO路·格斯特纳否决了这个想法，他说：‘我们并不一定要自己制造每一种技术。’

Lou Gerstner, who became CEO in 1993, rejected that idea and he said, quote, we do not necessarily need to manufacture every piece of technology.

我们的任务是让所有这些技术协同工作。

We need to be the company that makes all of it work together.

所以我们必须整合在一起。

So, we have to work together.

我们要做的是整合者，系统集成商。

We're going to be the integrator, the systems integrator.

他的战略最终带来了三样东西：IBM全球服务、大型外包合同，以及庞大的咨询团队。

His strategy ultimately produced three things: IBM Global Services, large outsourcing contracts and a vast consulting organization.

今天我们所了解的IBM，很大一部分就来自这些。

And that's a lot of what we know about IBM today.

不过，服务业务也有局限：利润率较低、人员规模大、内生增长慢、价格竞争激烈等等。

So services businesses do have limitations though: lower margins, higher headcount, slower organic growth, price competition, etcetera.

2019年，IBM以340亿美元收购了Red Hat，随后在2021年剥离了传统的基础设施托管外包业务。

In 2019, IBM acquired Red Hat for $34,000,000,000 and spun off its traditional managed infrastructure outsourcing business in 2021.

所以今天，你可以把IBM看作三个核心业务板块。

So today, you can think of IBM as three key businesses.

他们有软件业务，占公司整体收入的44%。

They have software, which is 44% of the business.

这个板块的毛利率高达80%。

That's at 80% gross margins.

这生意不错。

Great business.

他们31%的业务来自咨询。

31% of their business is consulting.

不过毛利率还不到30%。

That's under 30% gross margins, though.

剩下的23%是基础设施业务，毛利率接近60%。

And then 23% of the business is infrastructure, which is just shy of 60% gross margin.

过去三年，股价表现很好，扣除股息前涨了77%，收购红帽的效益也开始显现。

And so for the last three years, the stock's been doing really well, up 77% before dividends, and the Red Hat acquisition started paying off.

z17大型机的周期出人意料地稳健，但问题是，他们刚指出客户正减少对大型机的投入，转而把资本开支转向实体AI基础设施建设。

And the z 17 mainframe cycle was surprisingly solid, but the problem is that they just called out a shift away from mainframe spending with customers shifting capital spending towards the physical AI build out.

AI和相关硬件的需求很旺盛，但IBM正在失去客户技术预算中的份额。

Demand for AI and associated hardware is strong, but IBM is losing share of their customers' technology budget.

IBM在AI时代依然拥有一个强大的资产——红帽OpenShift，这是他们的企业级Kubernetes平台，用于跨多台计算机协调工作负载。

IBM still does have a strong asset for the AI era, Red Hat OpenShift, which is their enterprise Kubernetes platform for orchestrating workloads across multiple computers.

但如今有太多公司都在全栈提供AI能力，导致IBM今天遭遇了自115年历史上最大的股价跌幅。

But there are so many other companies offering AI capabilities up and down the stack that they're getting a little hammered today with the the biggest share drop in its one hundred and fifteen year history.

IBM这天确实不顺，但整个故事依然很有意思。

Rough day for IBM, but an interesting story nonetheless.

DeepMind的首席执行官德米斯·哈萨比斯呼吁成立一个由美国主导的机构，来测试前沿AI模型。

Demis Hays from DeepMind, the DeepMind chief, he has called for a US led body to test frontier AI models.

他说，社会正处在宝贵的时间窗口里，必须为技术以历史罕见速度前进做好准备。

He says society has a precious window to prepare for technology advancing at historic speed.

他是一位诺贝尔奖得主。

He's a Nobel laureate.

《金融时报》报道了这件事，他还发了一篇帖子在X上，点进去就能看到核心观点。

The Financial Times has the story, and there's an article that he posted on X that will sort of click through and give you the takeaways.

根据《金融时报》报道，谷歌DeepMind首席执行官德米斯·哈萨比斯呼吁设立一个美国标准机构，用于测试新型前沿AI模型是否构成国家安全威胁，并强调国际监管机构必须立即采取行动，应对技术飞速发展带来的风险。

From the Financial Times, Google DeepMind chief executive Demis Hassabis has called for the creation of a US standards body to test new frontier class AI models for national security threats, arguing that urgent action from international regulators is needed to address the risks posed by rapidly advancing technology.

我有点意外，他怎么以前从来没提过这个建议？

I'm surprised has he never proposed this before?

这感觉像是早就被提过无数次了，可能我只是记混了那些AI 2027、AI 2040的人，还有OpenAI的白皮书，以及Anthropic说过的话。

This feels like something that has been proposed many, many times, but maybe I'm just misremembering the AI twenty twenty seven people and the AI 2,040 people and the OpenAI white paper and what Anthropics said.

好像我们以前就听过类似的说法，比如需要设立一个监管机构——早在两三年前，All In播客就提过要给AI搞个‘药监局’。

Like, it feels like we've seen this before, like we need to have a regulatory body of some sort, all the way going back to when the All In podcast was talking about an FDA for AI back, like, two or three years ago.

但现在这事终于从DeepMind的高管嘴里说出来了，分量就不一样了。

But it's now here and it's coming from a DeepMind executive, which hits a little harder.

哈萨比斯——这位领导谷歌AI业务的诺贝尔奖得主——发出警告，恰逢白宫上个月突然禁止向Anthropic最先进的模型出口，同时各界也再次发出警告，称AI可能对全球经济和金融体系造成冲击。

The warning from Hays, Nobel laureate who leads Google's AI efforts, follows the White House's abrupt export ban on anthropic's most advanced models last month, alongside a fresh wave of warnings about the potential for AI to disrupt the global economy and financial system.

我们之前简单聊过《经济学人》。

We talked a little bit about The Economist.

他们提出的意见要温和得多，我觉得，因为他们根本没打算对AI的开发做任何实质性的调整。

They got to see got together with a much more moderate proposal, I think, because it wasn't actually it wasn't actually calling for any sort of change to the development of AI whatsoever

也没对部署方式做任何改变。

or the rollout.

只是说，未来十年AI可能会变得更好。

Just saying it it could get AI could get better in the next ten years

是的。

Yes.

这其实是一种相当

Which is a very sort of

可能会变得好很多。

Could get a lot better.

他们就是这么说的。

That's what they said.

他们可不是只说‘好很多’。

They didn't just say a lot better.

说的是‘好很多’。

Said a lot better.

但五月。

But May.

但《经济学人》真正的建议是，经济学家和政府官员得认真考虑应对措施。

But the actual pitch from The Economist was we need to have economists and government officials think about responses.

如果人工智能导致大量岗位消失，会带来什么影响？社会又会如何反应？我们需要提前制定政策，做好准备——无论是再培训、经济刺激，还是就业计划，各种应对措施都得提前考虑。

If there is job displacement from AI, what is the impact, what will the reaction be, to sort of like prep the legislation so you can be more ready when things start to happen, whether that's retraining or stimulus or jobs programs or all sorts of different things.

这次德米斯提出的方案，是谷歌迄今为止最详尽的AI监管提议，表明它正与Anthropic和OpenAI激烈争夺AI领域的领导地位。

So, this intervention from Demis is the most detailed proposal yet for AI regulation from Google, which is vying for AI leadership with Anthropic and OpenAI.

我们已经看到前沿模型对网络安全带来的挑战，随着技术持续进步，核风险、生物风险等更严重的威胁也可能很快浮现。

We've already seen the challenges frontier models pose for cybersecurity and other threats, including nuclear and bio risks, may soon emerge as capabilities continue to advance.

AI领域当前的飞速发展，要求我们建立一种全新的测试机制，用来评估前沿模型的能力——这种机制必须是动态的、灵活的，而且足够严谨。

The rapid progress we're seeing in AI requires a new approach to testing frontier AI model capabilities that is dynamic, adaptable, and rigorous.

美国在经济和技术上的优势，使其具备率先建立这样一套框架的有利条件。

The US is well positioned, given its economic and technical standing, to take the first step in developing such a framework.

我最大的疑问是，找上那些顶尖实验室，跟他们说‘你们得走这个流程’，听起来似乎挺简单的。

My big question is, it seems pretty easy to go to the leading labs and say, hey.

但我们美国有没有一套成熟的机制，来审查中国那边随便开源扔过来的代码？

We you have to go through this process.

我们有没有一套靠谱的体系，能对中方开源发布的代码进行有效评估？

But do we have a good framework in The United States for reviewing code that China just sort of throws over here open sourced?

因为你看，像KimE-K2和GLM这些模型，如果真要像FDA那样拖上六个月，更别说一年，甚至像药物研发那样拖上五到十年，开源模型早就远远领先了。

Because, I mean, as we've seen with the Kim e k two and GLM, like, if you tie someone up in an FDA like review for even six months, let alone a year, let alone what the FDA timelines are for drug development, five years, ten years sometimes, you are gonna have open source models that are way, way more advanced.

这就是它的名字。

Here's the name

新前沿模型。

of it.

新前沿模型。

New frontier model.

是的。

Yeah.

它现在正在接受为期六个月的审查。

It's going through a a six month review.

对。

Yeah.

假设审查时间其实只有一个月，但之后还需要

Let's say, maybe it's really like a one month review and then there needs to be

用AI的话，可能两分钟就能审完。

With AI, it could be a two minute review.

对。

Yeah.

我希望情况会是这样。

Who I would hope.

但假设要花三个月，嗯。

But let's say it's like a a three month Yeah.

三个月或者六个月的延迟。

A three month delay or six month delay.

等它最终发布的时候，嗯，

Then when it eventually does get released Mhmm.

它很可能在更短的时间内就被简化了，

It it gets distilled likely within even less time than that,

比那个时间还要短。

than it's

对公众开放。

publicly available.

是的。

Yeah.

我希望看到更多这样的信件和提案，确实一直在出现。

I would like to see so we keep seeing these, like, letters and proposals Yep.

而且它们总是附带紧急行动的请求。

And they always come, one, with a request for urgent action.

对。

Mhmm.

但很少会提供具体的、近期可能发生的场景。

But they rarely come with super concrete scenarios, like near term scenarios.

我希望看到的是，六个月后会发生什么。

I want, here's what's gonna happen in six months.

十二个月后又会怎样。

Here's what's gonna happen in twelve months.

或者干脆设个触发条件。

Or even just a a like a trigger.

如果有人提出，只要失业率超过10%，就给每个人发1000美元的救济金，而且要根据收入情况筛选，只发给中低收入群体，这想法挺有意思的。

Like, it would be interesting if somebody said, if the unemployment rate goes above 10%, I would recommend a stimulus check of $1,000 be sent to everyone and means tested so it only goes to the middle class and lower class.

这种做法其实很合理。

Like, that is a very reasonable thing.

新冠疫情那会儿，不就是这么干的吗？

That's basically what happened during COVID.

对吧？

Right?

当时失业率飙到15%，结果立马就有支票寄到家里了。

Like, the unemployment rate went to 15%, and then boom, there were checks in the mail.

这种明确的方案就很好，只要发生某种情况，就自动启动相应措施。

And that's a very concrete proposal that you could say, if this happens, then this happens.

是的。

Yeah.

我希望能有像德米斯这样的人。

I want I want someone like Demis.

对。

Yeah.

主要是关于少犯错的世界，还有2027年和2040年的AI前景。

Basically, the the the world of less wrong and AI 2027 and 2040.

没错。

Yep.

他们愿意把各种可能性清清楚楚地列出来。

They're willing to lay out super Sure.

当然。

Sure.

当然。

Sure.

各种可能的情景。

Scenarios.

而且他们有时候听起来确实挺自信的。

And they they can, at times, come across as Yeah.

特别像科幻片。

Very sci fi.

是的。

Yeah.

对。

Yeah.

但到目前为止，这些设想里总还是有那么一点现实依据。

But there's always, at least so far, some element of reality in them.

我明白你的意思。

I hear what you're saying.

所以我想要的是这样一个人，

So I I I want somebody who's,

就像那样。

like Yeah.

通常更偏向于一种比较温和的态度。

Generally more, like, kind of moderate Mhmm.

进来直接说，这里有几种可能的情况，这是我个人的看法。

To come in and just say, like, here's here's a few potential scenarios and this is what I think.

因为我不觉得德米斯能替政府建议该怎么做。

Because I don't believe it's, you know, Demis could suggest what he thinks that the government should do.

对。

Mhmm.

在这种情况下，美国政府。

The US government, in this case Mhmm.

他是在鼓励，当然可以。

He's, you know, encouraging, like Sure.

美国的监督机构。

US watchdog.

是的。

Yeah.

但他人在伦敦。

He's in London, though.

所以我觉得，关键还是得靠我们的立法者和政府，去思考这些不同的情境下，我们该怎么应对。

So I think it's gonna be on our lawmakers and our government to understand in these different scenarios, at least start thinking through in these different scenarios, how would we approach them.

对。

Mhmm.

对。

Mhmm.

是的。

Yeah.

我一向对时间表和预测感到头疼，因为这些太容易被挑刺，也实在太难了。

I just I always have a problem with the timelines and predictions because those can be get so nitpicked and they're so hard.

我更关心的是，别太纠结于

I'd be more interested in less of, like

不对。

No.

但你不觉得这样会更有帮助吗，如果

But don't you think that'd be helpful if if

我觉得这没什么帮助。

I don't think it's helpful.

不对。

No.

不对。

No.

不对。

No.

我确实不觉得有帮助。

I actually don't.

我觉得更有效的方式是，如果失业率升到百分之十，就成立一个新机构，专门雇人做点事。

I think it's much more helpful to say, if the unemployment rate goes to 10%, create a new government body that hires people to do something.

比如说，成立下一个运输安全管理局。

Like, create the next TSA.

我只是觉得可以发类似的

I just think I just think Send out similar

支票或者降低利率。

checks or lower interest rates.

对吧？

Right?

你告诉华盛顿，AI模型在破解计算机系统方面非常在行

You tell Washington DC, AI models are very good at hacking computer systems

没错。

Mhmm.

而且它们在这方面还会越来越厉害。

And they're gonna get better at hacking computer systems.

但这些能力其实没什么用武之地，因为破解计算机系统本来就是违法的。

There's not really much for them to do with that because hacking computer systems are already it's already illegal.

是的。

Yeah.

所以解决办法是企业得加强自身的网络安全，确保使用最前沿的模型。

And the solution there is for companies to beef up their own cybersecurity, make sure they're using the most advanced models.

对。

Yeah.

如果你能具体设想一些场景，比如货运行业的时间表，以及可能面临的岗位流失，或者其他类似领域，我觉得这能帮助华盛顿的立法者真正开始思考。

And so if you play out more concrete scenarios where, like, here's a timeline for the trucking industry and potential job displacement within trucking or any of these other categories, I just think it allows people in Washington, actual lawmakers, to start thinking about

我觉得这种想法是错的。

I just think that's wrong.

他们对这类预测总是错的。

Like, they're always wrong about those predictions.

是的。

Yeah.

但更好的做法是直接说：如果货运行业真的出现大规模失业，那我具体建议怎么做。

But it's so much better to just say, look, if the trucking industry goes through mass job displacement, then here's what I actually propose.

这就是解决方案。

Here's the solution.

而不是光说可能有问题，我觉得接下来会出问题。

As opposed to just saying, like, there might be a problem, and I think that there's a problem coming down the pipe.

我不确定。

I don't know.

你到底在主张什么？难道只是说天要塌下来吗？

Like, it's like, what are you actually advocating for other than just being like, the sky might fall.

你给我一个数字，然后让我自己去琢磨。

I have a p doom of this number, and it's your job to go figure it out.

你这么聪明。

It's like, you're smart.

你有什么建议？

What do you suggest?

全民基本收入？

UBI?

提高税收？

Higher taxes?

我觉得预测也不总是错的。

I just don't think predictions are always wrong.

过去十年里，有很多例子证明人们的预测完全准确。

There's been so much so many examples over the last decade where people have gotten predictions, like, dead on.

是的。

Yeah.

对。

Yeah.

没错。

Yeah.

就是要有形势判断力。

Situational awareness.

泰勒，你有考虑过这个问题吗？

Tyler, do think about this?

是的。

Yeah.

我的意思是，我大概属于支持有效监管的那一派。

I mean, I I'm probably in the camp of, like, productive regulation.

通常来说，如果监管不到位，往往会带来一些负面后果。

Like, usually, it, like, has some bad consequences if it doesn't really work out.

不过，我想补充一下，他描述的其实就是 Casey，也就是商务部下属的 AI 中心、CNN 和创新办公室。

But, also, I I was just gonna say, like, what he's describing is basically just Casey, the Center for AI, CNN, and Innovation, which is under the commerce department.

没错。

Yep.

而且现在 Casey 的运作方式基本上是自愿参与，但新版本似乎应该加强一些。

And it's, like, a slightly beefed up version because right now, Casey is, like, very much you opt in.

对。

Yeah.

但他还不如直接说，应该明确加强 Casey 的职能。

But it seems like he should just have said like, you should specifically beef up Casey Yeah.

推出这些政策，具体落实几项措施。

Add add these policies, do these like specific things.

对。

Yeah.

我觉得这样会更让人接受，也更实际一些，比如我们到底该怎么处理这封信？

And I think that would be much more palatable or or well received or like something like like what do we actually do with this letter?

这事儿有点模糊，说实话，我不太清楚我们到底该拿它怎么办。

It's kind of very, you know, like, I'm not sure what we actually do with this.

是的。

Yeah.

我觉得，要是回到网络安全这个问题上，如果是国家层面的事，比如国安局在处理这类问题，那就增加他们的预算，也许提高税收，或者发更多国债来拨款。

It feels like if I mean, to go back to cybersecurity, it's like if it's a national issue, like, the NSA works on this stuff, increase their budget, maybe, raise taxes to increase their budget, issue more debt to to increase their budget.

如果这能靠私营市场解决，那就去支持CrowdStrike，或者成立一家新公司来帮忙做网络安全。

If it can be solved by the private market, it's like go support CrowdStrike or start a new company that can help with cybersecurity.

我不确定。

I don't know.

从德米斯写的内容里提炼出来的实际建议，大概就是这些方向。

The actual concrete recommendation boiled down from what Demis wrote is something along these lines.

成立一个美国前沿人工智能标准机构，类似Casey，但规模和权限要更大一些。

Create a US frontier AI standards body that's like Casey, but probably more beefed up.

他还主张，这个机构应由联邦政府监管，但资金由人工智能公司提供。

He's also advocating that it's overseen federally but funded by AI companies.

制定并定期更新评估标准，以确定哪些模型和实验室属于前沿级别。

Define and regularly update benchmarks to determine which models and labs qualify as frontier.

目前还没有这样的机制。

So, that's something that doesn't yet.

要求前沿实验室在模型发布前至少三十天提交模型接受测试。

Require Frontier Labs to submit models for testing up to thirty days before release.

这个做法挺好的，因为这样就能把那些只是在Netflix上做推荐系统的公司排除在外——虽然技术上用了AI，但根本算不上前沿模型。

That's sort of nice because that would allow someone who's just going and building a recommender system on Netflix that's actually it is using AI technically, but it's not a frontier model because it doesn't qualify for that.

他们完全可以继续发布最新的推荐算法，完全不受影响。

So, they can just go and ship the latest recommendation algorithm on Netflix.

没什么大不了的。

No big deal.

对涉及网络、生物、核能、欺骗、自主性及安全防护能力的模型进行测试，需要严格的网络安全审查、模型档案、水印技术以及深入的安全研究。

Test models for cyber, biological, nuclear, deception, autonomy, and guardrail passing capabilities require strong cybersecurity personal vetting, model cards, watermarking, and substantial safety research.

由国家实验室、联邦机构和独立第三方审计机构来执行评估。

Use national labs, federal agencies, and independent third party auditors to conduct evaluations.

开发独立且保密的测试机制，防止实验室专门针对基准进行训练；要求实验室在发布后修复已发现的重大漏洞；所有在美国部署的前沿模型，包括外国和开源模型，都必须遵守规定，但小型模型可豁免。

Develop independent, confidential tests so labs cannot train specifically against the benchmarks, require labs to fix serious vulnerabilities discovered after release, apply the rules to all frontier models deployed in The United States, including foreign and open source models while exempting smaller models.

明白了。

Okay.

所以他想把这套规则也用在外国的开源模型上。

So he wants to apply it to foreign open source models.

这确实挺棘手的，不过也许可以通过类似DMCA的警告，让Hugging Face和GitHub下架相关内容，避免它们在互联网上广泛传播。

That feels very tricky, but I guess you can get, like, sort of DMCA notices to Hugging Face and GitHub so that it doesn't proliferate across the web.

但确实如此。

But Yeah.

我的意思是，真正下载和使用这些模型的人群，可能集中在某些特定的地方。

I mean, there's probably some power lot to, like, where people are actually downloading and Yeah.

运行模型推理。

Inferencing the model.

如果你去找所有那些云服务商和开源社区，跟他们说：这个模型其实有问题。

And I guess if you go to all the neo clouds and all the open source folks and you're like, okay, this model is actually a bad model.

你得坚持用 GLM 5.2。

You gotta go you gotta stick to GLM 5.2.

对。

Yeah.

监管算力要容易得多。

It's it's much easier to regulate the compute.

这对开源爱好者来说肯定会引起很大争议。

This is gonna be very controversial to the open source fans.

没错。

Yeah.

这有点像乔治·海斯的情况。

This is kind of like the George Hays

噩梦级别的表现。

nightmare tape.

攻击每一台GPU。

Attacking every single GPU.

当然。

For sure.

当然。

For sure.

如果测试发现存在足够严重的风险，就协调前沿实验室放缓进度，并将美国的框架扩展为国际共享的前沿AI标准体系。

Coordinate a slowdown among Frontier Labs if testing reveals sufficiently serious risks and turn The US framework into an international system of shared frontier AI standards.

嗯，我喜欢这个总体方向。

Well, I like the general direction.

我喜欢他更广泛地分享自己观点的做法。

I like the idea that he's just sharing his viewpoint more broadly.

我觉得这些都很好。

I think all of that is good.

我不确定这里有多少实质内容能说清楚，具体该怎么落实，实际会对行业产生什么影响。

I'm not sure that there's enough to dig into here exactly how this would like, where the rubber meets the road, how this would be implemented, or what effect this would actually have on the industry.

这可能会对开源非常有利，因为它能拖慢那些封闭源码的前沿实验室。

Like, this could be really good for open source because it could just slow down the frontier closed source labs.

但如果流程变得太繁琐，开源项目可能没有足够的预算去应对监管审批，那对开源反而可能是灾难。

It could also be really bad for open source if it's much more cumbersome because an open source project might not have a regulatory budget to actually massage a model through the approval process.

这就像为什么小型生物技术公司总是在推出新药前被大药企收购。

Like, there's a reason why small biotech companies get acquired by big pharma before they launch their drugs.

因为大药企在华盛顿有办公室，能亲自带议员走完整个流程。

It's because the big pharma companies have offices in Washington DC and can walk the legislators through the whole process.

总的来说，我理解那些说监管总是有利于巨头的观点，毕竟历史上一直都是这样。

So in general, I'm I'm sympathetic to the view where people say, oh, regulation benefits the biggest companies in the world, because historically, that's how it's played out.

也许这次会不一样。

Maybe that's different this time.

谁知道呢？

Who knows?

这可能会拖慢前沿进展。

It could just slow down the frontier.

但他确实是在一家顶尖实验室工作。

But, you know, he does work for a a leading lab.

所以

So

纽约那边现在怎么样？

What's going on in New York?

我会

I will

告诉你纽约的情况。

tell you what's going on in New York.

今天，纽约州长凯西·霍赫尔签署了一项行政命令，暂停该州新建人工智能数据中心一年。

Today, New York governor Kathy Hoechel signed an executive order placing a one year pause on new AI data centers in the state.

这是针对所有人的，除了

This is the for everyone except

美国老百姓。

Americans.

那些……不过，大概吧。

People that are But Probably.

这项命令规定，在纽约制定监管框架并开展环境影响评估期间，暂停新建数据中心。评估内容包括数据中心对能源需求、用水量、水质、空气质量以及电网的影响。

The order establishes a moratorium while New York develops a regulatory framework and conducts environmental impact assessments examining data centers, energy demand, water use, water quality, air quality, and effects on the electric grid.

你可能会觉得，这些方面本来就有不少监管了。比如空气质量，不管你是开一家新的烧烤店，还是建一座燃煤电厂，按理说都有一条总规矩，就是不能污染大气。

You would think there's a decent amount of oversight around those things generally already, like air quality, like, whether you start a new barbecue restaurant or a coal plant, you would imagine that there's just a general rule about not polluting the atmosphere.

这本来默认是适用于所有数据中心的，但看起来这里有个特殊情况，所以他们正在专门处理这个问题。

That would apply to data centers by default, but it seems like there's a little bit of a special case here, and so they're working on this in particular.

这一举动立刻遭到了科技行业的批评，他们认为限制数据中心建设会让当地社区失去工作机会，并削弱美国在全球人工智能竞赛中的地位。

The move immediately drew criticism from the tech industry, which argues that restricting data center construction will cost local communities jobs and weaken America's position in the global AI race.

今年早些时候，缅因州也考虑过类似的暂停令，但民主党州长珍妮特·米尔斯在担心这会阻碍一个大型数据中心项目后，否决了该提案。该项目计划建在一个因当地造纸厂关闭而仍在努力恢复的城镇。

Earlier this year, Maine considered a similar moratorium, but Democratic governor Janet Mills vetoed the proposal after concerns it would block a major data center planned for a town still struggling after the closure of a local paper mill.

霍楚尔的共和党挑战者布鲁斯·布莱克曼也反对这项暂停令，他认为是否批准那些能带来巨大经济效益的项目，应该由地方政府决定，而不是州政府。

Hochul's Republican challenger, Bruce Blakeman, also opposes the moratorium, arguing that local governments, not the state, should decide whether to approve projects that promise significant economic benefits.

如果这项命令最终生效，纽约将成为全美首个对大规模AI数据中心实施全面暂停的州。

If it stands, the order would make New York the first state to impose a broad moratorium on large scale AI data centers.

据我所知，纽约州最近并没有出现太多数据中心的建设热潮。

There hasn't been that much of a data center boom in New York State that I'm aware of.

挺有意思的是，他们会怎么界定‘AI数据中心’这个概念。

It'll be interesting to see how they how they define AI data center.

是按耗电量来定，还是看用了什么型号的显卡，或者别的什么标准。

Will they do it on energy or what type of GPUs you're racking or what's going on there.

但更重要的是

But more to

深入探讨一下。

dig in.

肯·格里芬上了高盛的播客《交易所》，嗯。

Ken Griffin was on Goldman Sachs' podcast, the exchange Mhmm.

就是那个《交易所》播客，这周大家都在传，虽然我记得那期节目是上个月录的。

Exchanges podcast, and it was circulating this week even though it was, I think, recorded last month.

他当时提到，对，

And he was talking about how yeah.

在他看来，如果数据中心真的建起来，那这规定就是个错误。

In his view, what what an error this would be to the data centers are gonna get built.

而且如果这些数据中心不在美国建，那就意味着数千亿美元的收入会流向其他国家。

And if they're not built here, that means hundreds of billions of dollars of revenue basically flowing flowing through other countries.

对吧？

Right?

而且不是。

And No.

其他国家。

Other countries.

我是说，可能其他州会先跟进。

I mean, it'd probably go to other states first.

对吧？

Right?

他当时在说

Well, he was talking about

哦，如果这事发生

Oh, if it goes

在全国范围。

national.

你知道，要是纽约先行动，其他很多州也会跟进。

You know, if if New York does it, there's gonna be a lot of other states.

是的。

Yeah.

这种情况下，网络上都说中国赢了。

The the meme is like China wins in this scenario.

美国参议员约翰

US senator John

我们和核能合作。

We Coogan with nuclear.