Trump Asks OpenAI to Stagger Release of New Model, Google Pressures Publishers on AI Licensing

欢迎各位收看《信息》频道的TITV节目。

Welcome everyone to The Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是6月26日，星期五。

It is Friday, June 26.

本期节目，我们将深入解读《信息》独家报道：政府出于安全考虑，要求OpenAI分阶段发布其新模型。

Today on the show, we are breaking down the information's exclusive reporting that the government has asked OpenAI to stagger the release of its new model over security concerns.

我们邀请到了两位揭露这一消息的记者做客节目。

We'll be joined by two of the reporters who broke that story.

我们还会带来独家报道，揭秘谷歌在与新闻出版商的授权谈判中采取的强硬立场。

We also have some exclusive reporting about how Google is playing hardball in negotiations with news publishers over licensing deals.

接下来，我们将迎来艾莉·K。

We've then got Allie K.

米勒加入我们的节目。

Miller coming on the show.

她是一位非常受欢迎的内容创作者，也是人工智能的重度使用者。

She is a very popular creator and AI power user.

她还为财富500强企业提供建议，帮助它们思考如何布局人工智能。

She also advises Fortune 500 companies on how they should be thinking about adopting AI.

此外，《信息报》的劳拉·布拉顿最新披露，一些Anthropic和OpenAI的客户正在悄悄被多收费。

And finally, The Information's Laura Bratton has new reporting on how some Anthropic and OpenAI customers are quietly getting overcharged.

今天这期节目内容非常精彩，我们马上开始。

It's going to be a great show, so let's get right on into it.

《信息报》昨天独家报道，政府已要求OpenAI分阶段发布其新模型GPT-5.6。

The Information reported exclusively yesterday that the government has requested that OpenAI stagger the release of its new model, GPT-5.6.

我们昨天采访了《AI议程》的作者斯蒂芬妮·帕拉佐洛，她拿到了那份披露细节的备忘录。

We spoke yesterday with Stephanie Palazzolo, author of AI Agenda, about the memo that she got ahold of revealing those details.

但为了更全面地探讨这件事的影响，我想请出我们的科技与政治记者莱奥·施瓦茨，以及联合执行主编阿米尔·埃弗拉蒂，他也参与了这篇报道的采写。

But for a broader discussion on the implications of all this, I want to bring on Leo Schwartz, our tech and politics reporter, and Amir Efrati, co executive editor, who also contributed reporting to the story.

欢迎两位。

Welcome to you both.

很高兴你能来。

It's great to have you here.

Leo，我先从你开始聊。

So, Leo, I want to start with you here.

咱们先回顾一下。

I mean, let's recap.

Stephanie拿到了萨姆·阿尔特曼发给OpenAI的那份备忘录。

Stephanie gets ahold of this memo that Sam Altman put out to OpenAI.

5.6版本即将发布。

5.6 is coming.

得把发布节奏错开。

We have to stagger it.

这是政府的要求。

That's what the government has asked.

政府还提到，他们希望在分批发布时对客户进行审核，这个细节挺有意思。

The government has also said that they would like to approve customers in the staggered release, which is an interesting detail.

你看到这份备忘录的时候感到意外吗？

Were you surprised by this memo that we got ahold of?

自从白宫对Anthropic及其强大的新模型Fable和私有版Mythos实施出口管制以来，我们已经等了好几个星期，就等着另一只靴子落地。

We've been waiting for this other shoe to drop for weeks ever since the White House imposed export controls on Anthropic and its powerful new model Fable and the private version Mythos.

大家一直都在猜，这些管制会不会也适用于其他实验室。

There was always this looming question of whether this was going to apply to other labs.

迟早会有公司，比如OpenAI或谷歌，推出这么强大的模型，这没什么好意外的。

It was only a matter of time before OpenAI or Google or another company created a model that was this powerful.

我觉得政府采取类似的处理方式，其实并不让人意外。

And I don't think it's that much of a surprise that the government would treat it in a similar way.

这一切其实都回到一个更大的问题：公司想发布新模型，到底该遵循什么样的许可机制或流程？

And this all goes back to this bigger question of what the licensing regime or the process looks like for companies who want to release new models.

我们现在发现，这些规则其实是边走边定的。

As we're finding out, it's sort of being decided as we go along.

你能说说这份备忘录出台之前，到底发生了什么吗？

Walk us through what happened in the run up to this memo.

这件事是突然冒出来的，还是之前一个月里已经有讨论了？

Did it come out of nowhere, or were there discussions happening in the month prior?

正如我们报道的，OpenAI 已经和政府合作，提前展示了他们的新模型，持续了大约一个月。

As we reported, OpenAI has been working with the government to preview its new model for about a month.

这甚至发生在白宫出台出口管制决定之前。

This is even before the export control decision by the White House.

萨姆·阿尔特曼一直向高层官员展示这个模型。

Sam Altman had been showing the model to top officials.

但正如我们报道的，特朗普政府的各个部门和高级官员基本表示，如果他们真要发布，必须采取分阶段的方式——先只给核心合作伙伴使用，之后才能向公众开放。

But as we reported, agencies and different top officials in the Trump White House basically said that if they want to release it at all, they would have to do it in this very staggered way where it would only go to top partners before it could do a wider public release.

明白了。

Okay.

阿米尔，我们把视角拉远一点。

Amir, zoom out for us a little bit.

你那边

You're out

晴空万里，蓝天一片。

there in the sunny skies and the blue skies.

你这是从三万英尺的高空往下看。

You're looking at this from 30,000 feet above.

你对这件事有什么反应？

What was your reaction to this?

这个发展有多离谱？

How how crazy was this development?

那些多年来一直认真从事人工智能研究、尤其关注安全问题的人，大概会觉得自己的担忧终于得到了验证。

Well, the people who've been working in AI for, the better part of a decade and taking it seriously from a safety perspective probably feel a lot of validation for some of their concerns.

因为现在，不少政府机构，特别是国家安全领域的部门，都认同他们的看法——这些技术可能成为武器，不能随意公开发布。

Because now, a good portion of government agencies, especially in the national security realm, agree with them that these are, you know, potential weapons or that these models, are just something that you can't release willy nilly.

如果你回溯到OpenAI刚成立的时候，那时候他们讨论的正是这些事。

And so if you go back to, you know, when OpenAI started, these were the discussions they were having at the time.

当时大家就在讨论，政府什么时候会把我们收归国有，那会是什么样子？

There were discussions around, you know, when will the government nationalize us, you know, and how was that gonna look like?

因此，随着这些能力不断提升，相应的担忧也在加剧。

So, as a result of of these fears, there was a as a result of these capabilities getting getting better and better.

当然，我们离所谓的奇点还远得很。

Now, obviously, we're we're nowhere near the singularity, if you will.

但很明显，尤其是国家安全机构，早就开始有效利用这些模型了。

But, clearly, the national security establishment in particular, has been using these models to some effect for for some time.

正如利奥之前提到的，OpenAI、Anthropic 和政府机构之间长期保持着密切的沟通与协作。

As as, Leo stated before, there's been a tremendous amount of discussion and coordination between OpenAI, Anthropic, and government agencies for a long time.

这根本不是什么新话题。

This is not a a new theme.

甚至在 OpenAI 开发他们的推理模型——当时代号叫 'Strawberry'——的时候，那还是 2024 年左右，这些讨论就已经在进行了。

Even when, OpenAI was developing, an you know, their reasoning model, which was code named Strawberry at the time, this was back in, like, 2024 even, these discussions were happening.

所以我觉得，这其实是一个缓慢积累的过程。

So I I think it's been a a slow moving thing.

也许很多人没意识到事情会发展到这一步，但如今事实摆在眼前，想回头已经很难了。

Maybe a lot of people didn't realize that this was gonna come to a head, but but here we are, and it's it's hard to see how this goes back.

现在，尤其是国家安全机构，会持续担心这些技术不能随意公开，因为一旦泄露，会给大量企业和其他国家带来严重后果。

There's now gonna be a persistent, concern by the national security establishment, in particular, that these things can't just be released out there because there will be a significant consequence to a lot of companies and a lot of other governments, if that happens.

所以，我们现在才处于这样的局面。

So that's that's why we are where we are.

明白了。

Okay.

这就是国家安全机构对此的看法。

So that's the national security community's perspective on this.

那OpenAI现在的立场是什么？

What about OpenAI's perspective today?

他们对这一新动向有何反应？

How are they reacting to this new development?

是的。

Yeah.

说到底，没人真正对此感到高兴。

I mean, nobody's happy about this inherently.

对吧？

Right?

作为一家公司，或者任何想公开发布技术、让所有人都能使用的机构，不可能对这种情况感到高兴。

You don't you know, as a as a company or as an entity that's doing its thing and trying to release technology publicly so that everyone can use it, they're they're not gonna be happy about this.

这其实就是现实，你得想清楚了。

It's just sort of a a reality, and and you have to decide, okay.

我真要为了这些事跟自己的联邦政府对着干吗？

Am I gonna go to war with my own federal government over these things?

答案大概率是否定的。

The answer is probably not.

所以你只能想办法跟他们合作。

And so you just have to have to find a way to work with them.

我觉得很多人其实挺意外的，没想到事情会这么快就发生。利奥之前就说过，这将是下一个要发生的事件，但还是有不少人没想到，居然会这么混乱、缺乏协调。

I think there has been some surprise, around just how quickly I mean, Leo was talking about, you know, this is the next shoe to drop, But it still comes as a surprise to to some folks just the level of, I suppose, disorganization, coordination.

现在有这么多不同部门在介入，比如我们之前提到的商务部，它正试图扮演一个统筹协调的角色。

There are all these different agencies where, you know, we reported about the commerce department in particular that's trying to be a kind of central coordinating factor here.

但说到底，这些公司——谷歌、Anthropic、OpenAI，还有其他几家——都得面对联邦政府内部各种不同的力量。

But in, you know, in essence, they're just all these different, forces within the federal government that all these companies are having to deal with, you know, Google, Anthropic, OpenAI, and others.

这真是个特别混乱的时期。

It's a very, very confusing time.

我觉得Leo应该能讲清楚刚签下的那个行政令，说不定能给这团乱麻理出点头绪来。

I think Leo can probably break down this executive order that was just signed that is, like, maybe gonna start to create some order some order here in the chaos.

昨天那则新闻引发的反应还挺有意思的。

And, you know, the reaction to the story yesterday was pretty fascinating.

人们对政府如此深度介入技术发布时机这件事，看法两极分化得厉害。

People have very, very strong views about what it means for the government to be this involved in deciding when technology is gonna be released.

有人直接说：天啊，我们这不跟中国一样了吗？

Some people are like, ugh, we're China now.

也有人觉得，这对开源社区和初创公司简直是灾难——这一点Leo其实有更多深入的报道可以分享。

Or other people are like, this is terrible for the open source community or for the startup community, which I think Leo actually has a lot more reporting that he can he can share about that.

所以Leo，你来聊聊吧，Amir刚才提到的那个行政令。

So Leo Leo, chime chime in here then because, I mean, Amir was talking about this executive order.

上次请您来聊这个行政令的时候，我们特别提到‘自愿’这个词非常关键。

When we had you on last talking about the executive order, I mean, we we talked about how voluntary was a word that was very important in that order.

阿米尔觉得这或许能带来一些清晰度，但您之前给我的印象是，它根本没增加多少明确性。

Amir is suggesting maybe that could provide clarity, but the last view I got from you is that it wasn't adding much clarity at all.

是的。

Yeah.

我觉得咱们得退一步，看看特朗普政府过去是怎么说这件事的——他们一开始就说，要让产业界先行。

I think it's important to take a step back here and just frame how the Trump administration had talked about this in the past, which is when it came in, it said, We want to have industry first.

他们不想对产业施加任何会扼杀创新的监管。

We don't want to be imposing any regulation on industry that's going to stop quote unquote innovation.

就连这个行政令出台时，主要是回应Anthropic的强模型Mythos，特朗普政府依然坚持说，我们做的任何事都是自愿的。

Even when this executive order was being decided, which was basically reflex to Anthropic's powerful model mythos, the Trump administration kept saying, anything we do will be voluntary.

行政令里甚至有明确的措辞。

There is even language in the executive order.

里面专门有一条写明：任何内容都不能被解释为强制性要求，也不能构成许可制度。

There was a whole clause that said, nothing in here can be construed to be mandatory or to be a licensing regime.

我觉得你看到的AI公司和整个AI圈的焦虑，其实恰恰相反——按理说，这个行政命令里提出的框架还要一个月左右才正式实施。

I think a lot of the headache that you're seeing or consternation you're seeing with AI companies and people in the AI community is it seems like the opposite is happening here, which is we still have about another month until this framework that was laid out in the executive order is actually implemented.

预计八月左右才开始。

It's supposed to start around August.

但在这期间，政府到底该怎么处理这个问题，根本没个明确的说法。

And in the meantime, what's happening is there's no clear framework for how the government's dealing with this.

结果呢，政府反而对Anthropic实施了出口管制。

And instead, the government is imposing export controls on Anthropic.

好几个部门都在各自决定OpenAI该怎么发布它的模型。

It's having a number of agencies basically decide how OpenAI should be releasing its model.

现在的情况特别不确定。

There's a lot of uncertainty.

就像我们这篇文章里提到的，几周前，也就是6月9号，政府开过一次行业反馈会，就在白宫对Anthropic实施出口管制前几天，当时还在收集业界对这个框架的建议，但到底怎么落地，现在还完全没谱。

And as we reported in this article, there was one feedback session a couple weeks ago on on June 9, which was a few days before the White House imposed its export control on Anthropic, where it was starting to get feedback from industry about what this framework should look like, but it's very much uncertain how that's actually gonna be implemented.

而在这期间，我们只能面对一种新的常态：一个没有规则的、事实上的许可制度。

And in the meantime, we just have this this new normal, which is essentially a de facto licensing regime with no rules.

所以我昨天跟斯蒂芬妮聊的时候提到，这种情况跟加密货币时代政府那种‘先执法、后立规’的做法有点像。

So what I put forth to Stephanie yesterday when we had her on is this felt a little bit similar to the regulation by enforcement mechanism of legislation, I guess, or action from the government that we saw during the crypto era.

利奥，我想问问你，你长期报道过加密货币，你觉得这个类比对吗？

And, Leo, I want to ask you if you agree with that because you covered crypto for a long time.

在加密货币领域，当时充满了监管不确定性。

In the crypto world, there was a lot of regulatory uncertainty.

根本就没有明确的规则出台。

There were no rules that were actually made.

SEC 总是等公司做了之后才跳出来，说‘我们对这个不太满意’。

The SEC kind of came out after the companies tried stuff and said, Well, we're not so happy about that.

但其实从来就没定过什么正式规则。

But there were really no rules that were set.

现在的情况可能也差不多：政府压根没立规矩，Anthropic 一推出模型，他们就突然出台出口管制，什么规则都没定，可大家都吓得不敢动了。

It feels like here maybe it could be similar in that the government never set any rules, Anthropic comes out with a model, and then they put these expert controls out now, and nothing has really been set, but everyone's scared to do anything.

说实话，我这话可能得罪人，但我感觉确实有点像。

I mean, I might make some enemies here, but I would say it might.

情况可能更糟，因为我亲身经历过拜登政府时期加密货币领域的拉锯战，那时候国会根本没出台任何立法。

It might even be worse because having covered the trenches of the crypto wars with the Biden administration, what was happening then was there was no legislation coming out of Congress.

金融监管机构，比如证交会和商品期货交易委员会，也没制定任何明确的规则。

There was no real rule making coming out of the financial regulatory agencies, the SEC and the CFTC.

相反，他们直接对加密行业提起诉讼。

But instead, they were doing lawsuits against the crypto industry.

至少在诉讼里，你会看到一份六十页的诉状，清楚说明监管机构的立场。

With a lawsuit, at least you have a 60 page complaint that lays out the agency's argument.

而这次，至少在对Anthropic的处理上，白宫已经实施了出口管制，算是某种形式的执法了。

In this case, we're getting some degree of enforcement, at least against Anthropic, where the White House imposed these export controls.

但这次没有六十页的诉状来解释政府这么做的理由，只有一封三页的信，来自商务部长霍华德·卢特尼克，里面提出了一套全新的法律论点，却完全没说清楚这办法怎么操作，或者为什么非得这么做。

But rather than having a 60 page complaint where the government lays out its argument for why it's doing it, it was a three page letter that came from Howard Lutnick, the commerce secretary, with a novel legal theory and no real explanation as to how it would work or why it was doing it.

我觉得，等一个月后这套框架正式出台，或许能给像OpenAI或Anthropic这样的公司明确出一条路，知道该怎么发布新模型。

So I think there's some hope that when this framework does come into place in about a month, it will create rules of the road for how a company like OpenAI or Anthropic is supposed to release new models.

但在这之前，我们实际上处在一种模糊的状态，像OpenAI这样的公司只能猜着政府和各个不同机构到底想要什么，然后硬着头皮去配合。

But in the meantime, we basically have this shadowy system where a company like OpenAI has to comply with what the government and all of these different competing agencies want.

如果他们不遵守，表面上看，后果就是会被施加出口限制，公司里所有外籍员工都将无法使用这个模型。

And if they don't, ostensibly, what would happen is they would have exporting controls imposed on them where no foreign employees at their company would even be able to use the model anymore.

阿卡什，我们确实得承认，迟早会有一场诉讼，现在可能还不是时候，也不一定就针对这件事，但终究会有一个节点，政府想放缓模型发布，而这些公司又急着推出产品，双方必然会发生冲突。

Akash, we we should we should definitely say that inevitably, there's going to be some type of litigation, not right now, not necessarily about this, but, you know, inevitably, there will be some moment where something is gonna you know, something's gonna have to give between the government's desire to slow down, a release or a set of releases and these companies, you know, desire to, to actually get, you know, get their material out.

就像我们之前说的，现在中国的一些开源模型已经在市场上出现了。

And you have, like we talked about before, I mean, you have, like, you know, Chinese models, open source models that are getting out there.

如果美国本土的专有模型被任何形式地拖慢了，我担心会越来越多的人转向这些中国模型。

If, you know, if the proprietary models built here in The US are being slowed down in any way, I think there's gonna be concern that, more people are gonna shift to these Chinese models.

那到时候你怎么办？

And then what do you do about that?

这些可是开源模型。

These are open source models.

它们的权重参数本来就是公开的。

These open weights exist.

你该怎么去监管它们？

How do you regulate that?

这听起来更像一场噩梦，Leo 之前也报道过有关这些会议的情况。

That seems like an even bigger nightmare that, you know, Leo's done some reporting on meetings around that as well.

所以我觉得，冲突终究是不可避免的。

So I I think there there is gonna be inevitably some amount of conflict.

但归根结底，这些公司早就预料到会有这么一天。

But, but at the end of the day, like, these companies have been thinking about this moment for a long time.

他们发布过一页又一页的报告，详细讨论这些模型的安全风险，尤其是网络安全方面。

They release pages upon pages upon pages about the safety risks of these models, including specifically around cybersecurity.

所以这些都不是什么新问题。

So none of this is new.

只是可能这个共和党政府里，很多人现在才第一次认真考虑这个问题。

It just, you know, a lot of people in this Republican administration maybe are thinking about it for the first time.

就像 Leo 前面说的，他们上台时还说要放手不管。

And and like Leo was saying before, they came in saying, like, hands off.

但实际上，他们现在逐渐意识到，这些模型如果推到极致，其实非常强大。

But, actually, they are being informed that actually these models are quite effective or can be quite effective if you play this out to the logical extreme.

所以我们得赶紧处理这个问题。

So we have to deal with this sooner than later.

是的。

Yeah.

我的意思是，我昨天一直在想，到底谁手里有筹码，AI公司到底有多大谈判能力。

I mean, you know, I I sort of I was trying to sort of think in my mind yesterday about what bargaining power who has the leverage here and the bargaining power that the AI companies have.

阿米尔，也许你正是想说这一点：虽然没有政府批准，他们什么都做不了，但你也知道，硅谷一向是硅谷的样子。

And, Amir, maybe this is the point that you're getting at, but it feels like although they can't do anything without the government's green light, mean, we've seen Silicon Valley wants to Silicon Valley, right?

他们对政府显然不是毫无影响力。

And it's not like they don't have any influence, I guess, over government.

所以我觉得，从长远来看，可能会出现一种情况，大家会明确传达一个信息：我们必须继续竞争。

So I mean, I feel like in my mind, and I'm asking Amir, I feel like in the long run, maybe there's a scenario here where people do make the message clear that we have to continue to compete.

你怎么看？

What do you think?

对。

Yeah.

我是说，为了让你清楚，OpenAI 和 Anthropic 虽然是两家不同的公司，但在安全风险的看法上却非常相似。

I mean, they are though just so you're very clear, like, OpenAI and Anthropic, they're not the same company, but they they do think very similarly about safety risks.

所以，两家公司在发布最新模型时，都决定在一段时间内主动削弱模型的网络安全能力。

And so both of them have moved forward with their latest model with an idea to blunt the cybersecurity capabilities of the model when it's released to the public, at least for some period of time.

这完全是他们自愿的做法，我觉得他们希望借此安抚一些政府机构，比如国家安全局，让它们明白谁可以、谁不可以接触这类网络安全能力。

So this is a voluntary thing that they are wanting to do, which I think they they would think would mollify the, you know, various government agencies, namely the NSA, who can and who cannot have access to these types of cybersecurity, you know, capabilities.

但归根结底，关于这条线该划在哪里、规则该是什么，冲突是不可避免的。

But, yeah, at the end of the day, like, there there is going to be an inevitable conflict about where this line is gonna be and what the rules are.

我觉得他们会继续沟通，不断调整行政令的具体条款。

And I think they're gonna continue to talk, they're gonna continue to try to refine, you know, the the executive order, the terms of it.

之后，还会有其他一大批公司陆续加入。

And then there's gonna be a whole host of other companies that that come later.

迟早会有人对这些规定提出挑战，最终很可能闹到最高法院。

Inevitably, somebody's gonna challenge, some of this, and it will go up to the Supreme Court one way or the other.

核心问题始终围绕着安全——尤其是国家安全：这么做是不是在无谓地冒险？

And it's always gonna be around, security, national security, you know, is this taking a risk unnecessarily or not?

很好。

Great.

谢谢Leo和Amir今天来做客。

Well, Leo and Amir, I wanna thank you for coming on.

这位是来自《The Information》新闻部的Leo Schwartz和Amir Efrati。

That is Leo Schwartz and Amir Efrati from our newsroom here at The Information.

谷歌在与新闻机构谈判AI授权协议时态度强硬。

Google is playing tough with news publishers as it continues to negotiate AI licensing deals.

我的同事Ann Gehan负责报道数字媒体与电子商务的交叉领域，她打听到这些谈判的独家细节。

My colleague Ann Gehan, who covers the intersection of digital media and e commerce, caught wind of exclusive details of those negotiations.

我请她来跟大家分享一下她了解到的情况。

I want to bring her on to share with us what she learned.

欢迎再次回到节目。

Welcome back to the show.

很高兴你能来。

It's great to have you here.

你好，阿卡什。

Hi, Akash.

你了解到谷歌向新闻出版商推出的这个新项目有哪些内容？

What did you learn about this new program that Google is pitching to news publishers?

过去几个月，谷歌一直在向出版商推广一项新的AI功能试点，这些功能旨在融入谷歌所有产品中，以展示新闻网站的内容。

Google has been pitching publishers over the past several months on a new pilot of AI features that are kind of designed to be woven into all of Google's products and how they feature content from news sites.

谷歌最近向出版商重点推介的三大功能包括：出现在谷歌新闻和Gemini回复中的AI生成文章摘要，以及一种AI生成的音频简报，它会汇总多篇新闻文章，相当于每日回顾，帮助用户发现他们感兴趣或常读的媒体发布的新内容。

So the three kind of key features that Google has been pitching publishers on recently are, these AI powered article summaries that appear in Google News and also in Gemini responses, as well as an AI generated audio briefing that summarizes various news articles as kind of like a daily recap or something that can surface new articles for people you know, that they from either publications that they like or topics that they're interested in.

这仅仅是谷歌为新闻出版商推出的最新计划。

And this is just the latest program that Google has had for news publishers.

多年来，谷歌尝试过各种不同的项目、合作模式和试点方案，比如‘谷歌新闻精选’，过去几年已在多个国家，包括美国，陆续推出。

They have tried all kinds of variations of different programs and relationships and pilots over the years, including something called Google News Showcase, which it's been rolling out in several different countries, including The US over the past few years.

这可以说是他们在疫情期间推出的一项新闻计划。

This was kind of their COVID era, you know, news program.

‘新闻精选’项目为出版商提供每年一笔固定费用，作为交换，出版商允许谷歌使用其内容，将内容整合到谷歌的各种产品中，比如安卓手机或谷歌应用等。

And so Showcase gives publishers basically a flat fee each year in exchange for the publishers sharing their content with Google so that Google can kind of weave that into their various products whether it's Android phones or the Google app, things like that.

过去几个月，谷歌告诉一些出版商，最终会逐步停用Showcase。

And so over the past couple of months, has told some publishers that eventually it's going to phase out Showcase.

所以它一直在极力推动尽可能多的出版商加入这个新的AI试点项目。

So it's really been pushing as many publishers as it possibly can to sign on to this new AI pilot.

但到目前为止，主要都是一些大媒体，比如《华盛顿邮报》、《卫报》和《金融时报》。

But so far, it's mostly been big names like The Washington Post, The Guardian, and The Financial Times.

一些规模较小或中等的出版商，可能对和谷歌达成这样的协议还有一些顾虑。

And some of the smaller or kind of more mid sized publishers, you know, maybe have some reservations about entering into a deal like this with Google.

那再帮我回忆一下，Showcase到底是什么？

And so so remind me, so what what was Showcase?

就是那个要被停用的项目，具体是怎么回事？

What was the thing that was getting that's getting phased out?

Showcase就是谷歌每年给出版商一笔固定费用，作为交换，获得在谷歌所有产品中使用他们内容的权限。

So Showcase, Google pays the publishers a flat annual fee in exchange for being able to use their content kind of across all of Google's different products.

明白了。

Got it.

所以现在他们想弱化Showcase，转而重点推这个新的AI概要语音简报功能。说实话，作为记者，我本不该这么说，但我真觉得这功能挺不错的。

So trying to deemphasize that and emphasize this new AI overview audio briefing feature, which I mean, look, maybe I shouldn't say this as a journalist, but I think it's a pretty good feature.

我平时经常用Google新闻，尤其是那个新闻标签页，特别希望有个准确、有出处、基于真实新闻报道且经过核实的摘要。

I mean, I use Google News all the time, especially the Google News tab, and I would love a summary that is correct and cited based on actual news reports that I know is verified.

新闻出版商们对这个怎么看？

How do the news publishers feel about this?

对出版商来说，这事儿有利有弊。

It's kind of a double edged sword for the publishers.

对很多出版商来说，这不过是谷歌又推出的新花样罢了。

I mean, for a lot of them, this is kind of just the latest thing that Google is trying.

过去，很多出版商和谷歌的关系一直挺紧张的，毕竟他们太依赖谷歌带来的流量了。

You know, there's always been a pretty contentious relationship between a lot of publishers and Google just because in the past they've been so dependent on Google as a major source of traffic.

现在大家越来越多地用Gemini、ChatGPT、Claude这类AI工具，不再像以前那样在谷歌搜索结果里点链接了。

And as people are using more AI tools like Gemini, like ChatGPT, like Claude, they're not searching on Google and clicking on links that come up in Google search results as much.

所以过去几个月，很多出版商的流量直接断崖式下跌。

So a lot of publishers have seen their traffic basically just fall off a cliff over the past couple of months.

很多出版方并不觉得，那些在谷歌应用或谷歌新闻里自动生成的摘要能有效把读者引回他们的网站。

And a lot of publishers don't necessarily feel that AI generated summaries of their content that live on Google's app or in Google News and don't bring people to their sites.

对他们来说，这根本算不上一个有效的解决方案。

A lot of them don't really feel like that's an effective solution.

明白了。

Okay.

这里还有一个问题，这个协议并不仅仅是靠这种方式，让谷歌在谈判中占据优势。

And there's another catch here, right, with this deal, which is that's not the only way that Google is trying to take the upper hand in these negotiations.

是的。

Yeah.

根据这个试点项目的条款，谷歌向出版方索要的是非常宽泛的内容使用授权，基本上想怎么用就怎么用。

So under the terms of this pilot, Google is asking for pretty broad permissions from the publishers to use their content, you know, kind of however it sees fit.

因此，一些出版方对此心存疑虑，因为这些协议在法律上确实赋予了谷歌使用其内容训练AI模型的权利。

And so some publishers have had reservations because these agreements do technically give Google the right to use their content for training AI models.

所以很多有顾虑的出版方认为，谷歌提出的这种一次性固定付费模式，根本配不上他们所要求的广泛授权。

And so a lot of publishers feel that, you know, who have reservations feel that just this kind of flat fee lump sum structure isn't necessarily appropriate compensation for the broad permissions that Google is asking for.

你知道，这种一次性付清的模式，以前其实并不太受欢迎。

You know, the flat fee structure I think was something that wasn't always super popular.

还有一些新闻或科技公司，比如苹果新闻，它们是根据使用量或者带来的流量来付费的。

There are other news or kind of tech companies that have news initiatives like Apple News that pay based on usage or the traffic that they drive to the articles.

所以我觉得，这种一次性付费模式加上这么宽泛的授权，让很多出版方都犹豫了，不确定该不该签这个协议。

So I think the flat fee structure paired with these broad permissions has really given a lot of publishers pause about whether or not this is something that they want to sign on to.

他们确实犹豫了，但我好奇你觉得最后会怎样。

It's given them pause, but I wonder what you think ultimately happens here.

因为如果他们发现，虽然没从谷歌新闻获得多少流量，但很多人确实在用这些平台，你觉得他们最终还有别的选择吗？

Because if they're seeing traffic if they're not seeing the traffic benefit of being included in Google News, for example, and they are seeing that a lot of people are using these platforms, do you think they have a choice ultimately?

是的。

Yeah.

这迫使很多出版方重新思考谷歌在他们业务模式中的位置，以及他们到底该在哪里触达读者。

I mean, it's causing a lot of publishers to really reconsider how Google fits into their business model and where they can reach their audience.

对于一些规模很小的本地出版方来说，这笔钱真的很重要。

I mean, for some really small local publishers, they really depend on this funding.

这笔资金支付了员工的工资，让他们能够继续运营下去。

The funding pays salaries for their employees, it allows them to continue to exist.

而且他们的内容往往很有价值，因为非常独特、聚焦，只有这家出版商才提供这些信息。

But also, their content many times is pretty valuable because it's very unique, it's very focused and that publisher is the only place that that information exists.

他们该不该接受一次性付费，让谷歌随意使用他们的内容？

Should they agree to a flat fee and let Google use their content however they want?

我觉得现在很多人都在认真思考这个问题。

I think a lot of them are really thinking very hard about that right now.

你也会看到一些其他出版商并没有完全拒绝谷歌，而是开始把谷歌从他们的商业模式中剔除。

And I mean, you see other publishers that have kind of started to not turn their back on Google, but really kind of take Google out of their business model.

比如，更加专注于订阅模式。

Double double down on on subscriptions for example.

没错。

Exactly.

是的。

Yeah.

通过订阅、电子报等方式，直接接触受众，而不是依赖来自谷歌的流量和广告收入。

Subscriptions, newsletters, anything they can do to reach their audience more directly versus relying on traffic from Google and and ad revenue.

你已经看到不少出版商，尤其是那些拥有较大受众群体的，正在朝这个方向转变。

You've seen a lot of publishers, especially know, ones that do have a bigger audience, you know, move more in that direction.

所以我认为，谷歌依然是房间里最大的巨头，很多出版商都感到有不小的压力，不得不与他们合作。

So I think Google is still certainly the giant in the room and I think a lot of publishers feel a lot of pressure to work with them.

但确实也有一些出版商，正抓住这个机会，重新思考他们与谷歌的整体关系。

But there are definitely some publishers out there who are using this moment as a way to rethink their relationship with Google more generally.

很好。

Great.

非常感谢你今天来做客。

Well, I want to thank you for coming on.

这位是安·吉汉，我们《信息》杂志的电子商务记者。

That is Ann Gehan, our e commerce reporter here at The Information.

我们的下一位嘉宾是一位由人工智能驱动的用户。

Our next guest is an AI powered user.

艾莉·K.

Allie K.

米勒曾与众多财富500强企业探讨如何有效运用人工智能。

Miller has spoken to a wide array of Fortune 500 companies about how to use AI effectively.

她还是这一领域的热门内容创作者，在各平台拥有两百万粉丝。

She's also a very popular content creator on the topic with 2,000,000 followers across her channels.

她还通过自己的公司Open Machine投资人工智能企业，并提供战略支持。

She also invests in AI companies and helps with strategy with her company Open Machine.

我想请艾莉来深入聊聊当前人工智能领域的一些热点话题。

I want to bring on Allie to unpack some of the hot button topics in AI right now.

艾莉，欢迎来到节目。

Allie, welcome to the show.

很高兴你能来。

It's great to have you here.

是的，谢谢邀请我。

Yeah, thanks for having me.

我特别期待。

I'm excited.

好的，这周我看了你的一些内容，发现你说你有一支由34个AI代理组成的团队。

Okay, so I was watching some of your stuff this week, and I saw that you said that you have an AI agent workforce of 34 different agents.

你能给我讲讲你是怎么组建这支团队的吗？它到底帮你做些什么？

Can you walk me through how you built this workforce, what it even does for you?

这到底是什么东西？

What is it?

我整个团队都是在Claude Code里用自然语言搭建的，就是直接说：‘给我造一个能做X、Y、Z的代理。’

So I built the whole thing inside of Claude Code, all with natural language and just saying, Make me an agent that does X, Y, Z.

我花了大概几个月不断调整，才慢慢摸清了这个结构。

And I basically iterated on this for a couple months to figure out what this structure is.

我直接跟西蒙沟通，他是我的AI首席助理。

I speak directly to Simon, my AI chief of staff.

西蒙下面有六个直接下属，名字都取自《老友记》里的角色。

Simon has six direct reports all named after the Friends characters.

它们都有子代理。

They all have sub agents.

在任何时刻，它们都可能临时生成数百个代理来完成任务。

And at any given point, they might also spin up hundreds of temporary agents to be able to get something done.

更重要的是，西蒙还有一个小助手叫托比，负责管理记忆，像项目经理一样整体协调，提升效率。

Importantly, Simon also has a little assistant, Toby, who manages memory, who is looking across kind of as a project manager to make it more efficient.

这其实是个巨大的机会。

And actually, think that's a massive opportunity.

大家都对多代理和AI团队充满热情。

Everyone's very excited about multi agents and AI workforce.

但我们往往忽略了它枯燥、不光彩的那一面——怎么在整个公司里保持上下文连贯？

We're kinda missing the boring, ugly side of it, which is how do you maintain context across an entire company?

怎么确保这个AI团队不会跑偏？

How do you make sure that this AI workforce doesn't even drift?

所以我现在反而喜欢处理这些琐碎的事，比如保持系统更新。

And so I actually like handling the boring stuff now, which is keeping it updated.

我们团队最近在做的一个非常有趣的用例是，团队里的每个人都有一个专属的Slack频道，专门用来和他们的AI团队协作。比如，我可以直接进入苏西的频道，让她帮忙处理一件事。

And one very exciting use case that we've been working on as a team is that every single person on our team has one Slack channel dedicated to them and their AI workforce, which means I can jump in to Susie's channel, for instance, and ask Susie to do something.

我其实是在和她的AI首席助理对话。

And I'm talking to her AI chief of staff.

所以我们现在说的苏西，是真人还是AI助手？

So we've already Susie the is Susie a real person or Susie an agent?

是的。

Yeah.

在这个场景里，苏西是个真实的人。

Susie in this in this case would be a human.

明白了。

So Okay.

好的。

Okay.

苏西是真人。

Susie is real person.

明白了。

Okay.

对。

Yeah.

我觉得，本质上，一家AI优先的公司，应该在人员、流程和产品上全面贯彻AI优先。

I think, like, basically, an AI first org should be AI first across people, process, product.

每个人都会忽略流程这回事。

Every single person forgets about process.

你怎么在团队之间、人与人之间保持协同？

How do you just maintain things between teams, between people?

如果我让人发封邮件，结果等了七个小时才收到回复，这合理吗？

If I ask someone to send an email or whether they sent an email, why am I waiting seven hours to get that response?

所以对于低风险的事，我们就直接丢进那个频道，对方的AI团队会自动回复我们。

And so for low risk things, we're able to throw it in that channel and that person's AI workforce is going to respond back to us.

也就是说，对方的AI首席助理——可能她自己也有个西蒙——会先评估、分类这个请求，判断风险，再给我们反馈。

So that person's AI chief of staff, which again, maybe she has her own Simon, is qualifying and classifying that ask, figuring out the risk and coming back to us.

实际上，我对话的不只是一个人。

Essentially, I'm not just talking to one human.

我是在和一个人加他的AI团队交流。

I am talking to one human plus their AI workforce.

每个一线员工都得到了倍增，我们每个人也都有一套接口，可以和他们的团队沟通。

So every single IC is multiplied and we each have an interface, mode to be able to speak with their workforce as well.

好的。

Okay.

所以是34个不同的智能体。

So 34 different agents.

你到底有多少个？

You've what got is it?

罗斯、瑞秋、乔伊、钱德勒？

Ross, Rachel, Joey, Chandler?

约翰。

John.

是的

Yep.

是的

Yep.

是的

Yep.

那是PB。

That's PB.

我是谁？

Who am I?

莫妮卡。

Monica.

不好意思。

Sorry.

好的。

Okay.

那他们到底帮你做些什么呢？

So, like, what do they actually do for you?

你到底交给它哪些任务？

What what what are the tasks that you give it?

是帮你报税之类的吗？

Is it filing your taxes or what?

我的会计师特别棒，所以我从来不用操心这些事。

So I my accountant is amazing, so I never do that.

不过目前来说，我可能会调整一下任务的分工方式。

But for now, at least, the the way that I break them down might actually change.

我们目前设计多智能体系统时，大多是把人类的工作方式直接套用到AI代理身上。

So I think by default, a lot of the way that we work with multi agent systems is taking what we know about humans and just applying it into the AI agent space.

所以它看起来就像一个公司组织。

That's why it looks like an org.

所以才会觉得是Monica在负责运营。

That's why it looks like Monica runs operations.

瑞秋负责客户事务。

Rachel runs client work.

罗斯负责教育板块。

Ross runs education.

我觉得有意思的是，要找出一些我们以前从未设过的全新角色。

What I think is interesting is figuring out net new roles that we didn't have before.

比如西蒙手下的托比，这种角色以前根本不存在。

So Simon having that Toby, that's a role that we wouldn't have had.

就像有个AI监督员在旁盯着，确保一切顺利，这样效率会更高。

This like AI watchdog just looking over things going, that can be more efficient.

这样效率会更高。

That can be more efficient.

菲比是我后来加进来的，一开始我没她，因为我觉得自己用AI代理太保守了。

And Phoebe, I actually started without Phoebe and added her because I was telling myself that I'm not being creative enough.

我觉得自己在运用AI代理这件事上，还不够有远见。

I'm not being ambitious enough with how I'm using AI agents.

我加了Phoebe。

I added in a Phoebe.

Phoebe的全部工作就是当那个爱做梦的怪人，盯着输出结果看，

Phoebe's entire job is just to be the weirdo that dreams, the weirdo that looks at output from

所以现在他们更像是顾问，而不是实际动手的人，

all So it's kind like sort of acting more so as advisers than than as doers right now, sort of

不对。

No.

他们在完成具体任务。

They're doing task completion.

比如客户提案这类属于客户业务的工作，Rachel和她的子代理会先起草，再跟之前的做对比。

So if there's, for example, a client proposal that would fall under the client work, Rachel and her sub agents would draft that first thing, would move over it compared to previous ones.

你是得专门提醒他们去做，还是他们看到邮件就自动处理了？

And do you have to prompt them to do it or do they just see the email come through and just do it?

每半小时，我的AI团队就会循环执行一个提示，内容就一句话：做点聪明的事。

Every thirty minutes, I have my AI workforce on a loop with the prompt that literally just says do smart things.

我醒来了。

I have woken up.

我的日程里出现了些我根本没记得安排的事项。

I have had things show up in my calendar that I didn't even realize that I'd put in.

西蒙希望我多做些物理治疗。

Simon wanted me to do more physical therapy.

所以他每天给我安排了两个十分钟的时段。

And so he put in two ten minute blocks every single day.

我现在能分辨出这些安排是他做的，还是人类助理做的。

I now have him, cite whether the meeting was from him or, or from a human agent, human assistant.

他们都在主动完成任务。

And, and so they're, they're completing tasks proactively.

我觉得这是从OpenCLaw时代以来最大的转变之一——我们正进入一个更主动、更异步的世界，我不再总是得提醒别人‘请做这个’。

I think that's one of the bigger shifts that has come from the OpenCLaw era, which is that we're moving into a world that is much more proactive, much more asynchronous, and I am really not always going to be the one that's saying, Hey, please do this.

你每个月运行这34个代理要花多少钱？

How much does it cost for you to run these 34 agents per month?

我是在订阅范围内使用的，还没用到每月200美元的限额。

I'm doing it within my subscription, so I have not yet hit my $200 a month limit.

如果真用超了，我还有另外200美元的额度。

If I did, I also have the 200

你是指Claude里的额度吗？

Within dollars Claude?

在Claude里，我会主动设置一些任务，通过Codex来执行。

Within Claude, there'll be some proactive proactive things from me that I put into Codex.

我同时在用Claude Code和Codex，两个账户各有200美元额度，目前都没用完，还不需要动用API。

So I'm using both Claude Code and Codex, and I have $200 accounts in both, and I have not hit that max to force myself to have to use, you know, APIs.

我们有些产品是靠API驱动的，但所有个人使用场景——比如效率提升或自我成长——都控制在订阅额度内。

We do have some products that is API driven, but anything that is personal, you know, productivity or personal growth, all of that is within subscription amounts.

那你是怎么决定让这些代理访问哪些内容的？我猜这34个代理能访问你的邮件、文件和电脑吧？

And then how do you think about like what you give I mean, I assume these 34 agents have access to your email, your files, your computer.

你是不是把线上所有足迹都开放给他们了？

Do you give it access to literally everything in your online footprint?

还是有些东西你不给？

Or are there things that you don't?

是的。

Yeah.

我的风险级别会比一般企业高一些。

So I'm going to be higher risk than the average enterprise.

我先说说我的做法，再跟你聊聊财富500强公司是怎么做的。

So I'll give you my answer, but then I'll also kind of share what Fortune 500 companies are doing.

在我的工作空间里，不同的助手有不同的访问权限。

So in my workspace, different agents have different access points.

比如罗斯，他可以访问我整个教育文件夹。

So Ross, for example, can access my entire education folder.

但乔伊可能没必要有这个权限，因为他负责产品。

But maybe Joey doesn't have to do that because Joey's running product.

乔伊可以部署代码，其他助手则不行。

Joey can deploy code, the other ones cannot.

所以它们之间的规则和权限确实不一样。

So there's definitely different rules and restrictions between them.

这跟我们对待人类的方式完全一样。

Again, exactly how we think about humans.

我们有身份和访问管理机制。

We have IAM.

我们可以设定，某位高管只能访问这500项内容，而其他内容则绝对不能碰。

We have the ability to say this executive can access these 500 things that in turn should not be able to.

所以你的代理也该有层级分明的访问权限设置。

And so your agent should have stratified, sets of access, different permissions.

这肯定是其中一层。

That's definitely one layer.

第二层是逐步建立信任。

And then the second is progressive trust overall.

我根本没有给这些代理任何向外部世界发布内容的权限。

So I did not give any of these agents the ability to to put anything out into the external world.