Inside Claude Opus 5, Open-Source AI Regulation, Google’s Chip Financing Strategy

欢迎大家收看 The Information 的 TITV。

Welcome everyone to The Information's TITV.

我是 Akash Pasricha。

My name is Akash Pasricha.

今天是 7 月 27 日，星期一。

It is Monday, July 27.

这周有一大波季度财报等着我们。

We have got a big week of quarterly results ahead.

周三有微软、Meta 和 Robinhood。

Wednesday, we've got Microsoft, Meta, Robinhood.

周四有苹果、亚马逊、Coinbase 和 Reddit。

Thursday is Apple, Amazon, Coinbase, and Reddit.

明天还有 PayPal，考虑到有报道说这家公司可能被收购，应该会很有意思。

And tomorrow, we have PayPal, which should be interesting to see in light of reports that there has been acquisition interest for that company.

这周我们会全程报道这些财报，记得关注我们的节目。

We'll be covering all of those results this week, so stay tuned to our shows.

今天节目里，我们要聊聊Anthropic的Opus 5模型的首批评测。

Today on the show, we are unpacking initial reviews on Anthropic's Opus five model.

我们请来了Clay公司的AI负责人来聊这个话题。

We're bringing on the head of AI at Clay to talk about that.

接着，我们会和一位前白宫高级官员谈谈她对Kimi K3引发的AI政策辩论的看法。

We're then going to speak with a former top White House official about her view on the debates that Kimi K3 has sparked around AI policy.

我们还有一些新报道，讲的是谷歌如何借鉴华尔街大胆的债务策略，来为它的AI建设融资。

We've also got some new reporting on how Google is taking a page out of Wall Street's ambitious approach to debt as it seeks to finance its AI build out.

最后，我们会和我们的合作伙伴普华永道聊一聊，更全面地看看财富500强企业在开源AI上的立场以及它们的使用情况。

We're going to close out the show with a conversation with our partner, PwC, for a broader look on where Fortune 500 businesses stand on open source AI and how they're using it.

今天的节目会很充实，这周也会很忙，那我们就直接开始吧。

It's going be a busy show and a busy week, so let's get right on into it.

Anthropic周五发布了最新的Opus 5模型，这意味着开发者们周末都在忙着测试这个新模型到底怎么样。

Anthropic on Friday released its latest Opus five model, which means developers were very busy over the weekend seeing what the new model is made of.

我想请来AI销售工程公司Clay的AI负责人Jeff Barg，来聊聊他的看法。

I want to bring on Jeff Barg, the head of AI at AI Sales Engineering Company, Clay, for his thoughts on all this.

杰夫，欢迎来到我们的节目。

Jeff, welcome to the show.

很高兴你能来。

It's great to have you here.

非常感谢。

Thanks so much.

是的。

Yeah.

谢谢你们的邀请。

Thanks for having me.

这个周末很忙吧。

So busy weekend.

你在玩那个新模型吗？

Were you playing around with the new model?

是的。

Yes.

是的。

Yes.

周末很忙。

Busy weekend.

我觉得我们花了很多时间，既在评估这个模型在我们内部用例中的表现，也在看我们的客户会怎么用它，同时也在尝试把它当作个人编程助手来用。

I think we spent a lot of time both eval ing the model for our internal use cases, how our customers will use it, and as well just trying to use it as a personal coding agent.

所以这个周末确实挺忙的。

So it was a it was a busy weekend.

那你们觉得怎么样？

And so what what's the review?

总的来说，它真的很不错。

Yeah, the headline is it's really good.

我觉得每次Anthropic发布新东西，总是很有意思，因为有时候你看到一些基准测试结果，但那些结果在体验上或者我们内部评估中并不一定能复现。

I think it's always, it's always interesting to see what happens with an Anthropic launch, because sometimes you see some benchmarks, and then they don't replicate in either vibes or in actual eval results from our internal evals.

我觉得这次，我们内部的基准测试结果和Anthropic公开说的基本一致，性能接近Fable，但成本却低很多。

I think here, the internal benchmarks that we had pretty much matched what Anthropic said publicly, which is really close to Fable's performance at a fraction of the cost.

我们看到的，Anthropic说大概是Fable成本的一半。

And we were seeing, I think Anthropic said, about half the cost of Fable.

我们看到的甚至比那还低。

We were seeing even more than that.

所以可能只有Fable成本的三分之一甚至更少，而且在我们一些最难的评估中表现非常接近。

So maybe a third or even less of the cost of Fable and really close to the performance across some of our hardest evals.

那GPT 5.6怎么样？

What about five point GPT 5.6?

对。

Yeah.

我觉得他们确实在挑战一个非常难的领域，因为实际上5.6的表现和Fable、Opus五差不多。

I I think it's it's a really hard they're playing in a really hard space because actually 5.6 comes in just around the same as both Fable and Opus Opus five.

而且它比这两个都便宜。

And it's even cheaper than than both of them.

所以我觉得，Anthropic其实是在这个艰难的中间地带竞争，他们拥有前沿技术。

So I think, you know, Anthropic is really playing in this in this hard middle ground where they have the frontier.

Fable在顶尖性能和代码质量方面确实表现得很出色。

Fable seems to be really doing well at, you know, frontier performance and writing in quality of code, etc.

但在这种第二梯队，或者说日常使用的模型层面，我觉得Anthropic面临来自OpenAI的激烈竞争。

But in this kind of tier two where, you know, you have, I don't know, the daily workhorse model, it's not really a tier two, but your kind of daily model for doing things, I think Anthropic has a faces some pretty steep competition from OpenAI.

有意思。

Interesting.

所以，这些是基准测试和评估的结果。

So, I mean, that's the benchmarks and the evals.

那感觉怎么样呢？

What about the vibes?

这算是一个有趣的软性指标吧。

This is kind of an interesting soft metric, I guess.

周五刚发布的时候，好像有些人并不太看好。

When it first came out on Friday, it seemed like some people weren't too bullish on the vibes here.

你觉得这是为什么？

Unpack that for why do you think that was the case?

对。

Yeah.

这挺有意思的。

It's been interesting.

我觉得每次Opus发布都会遇到这种反应。

I think every Opus launch has had this kind of reception on arrival.

回想一下Opus 4.7发布的时候，我记得那可能是基于新的预训练模型。

I'm thinking back to the Opus four seven launch where I think there was it was maybe based off of a new pre train.

我们Clay内部很多用户其实不太喜欢Opus 4.7，那时候很多人开始转向Codex和GPT系列模型。

A lot of our internal users at Clay really didn't like Opus four seven, and that's when actually a lot of folks started moving over to Codex and, you know, the GPT series models.

我觉得Opus 5这次，氛围还不错，但Clay内部大多数用户还是会继续用Fable。

I think with Opus five Opus five, the vibes are good, But I think most users internal to Clay will continue to use Fable.

我觉得Fable的氛围简直太好了，有种说不出的灵动感。

I think the vibes on Fable are just so immaculate and just sort of effervescent.

就是那种写作质量，还有那些难以言传的东西，真的很难复制。

Like, there's just a there's something to the quality of the writing and the the sort of, intangibles that are really hard to replicate, I think.

但我想说的是，即使你承认Fable的成本很高，你还是觉得人们会继续用它吗？

And I guess but you're saying that even acknowledging the tremendous cost of Fable, you think people are still going to be using that?

我觉得会。

I think so.

我的意思是，对于日常用户来说，尤其是非企业账户，真的很难说。

I mean, it it's really hard to say for daily users, especially on non enterprise accounts.

所以主要是Claw的普通用户群体。

So really the clawed max user base.

我觉得很多人会转向Opus 5，可能在那里获得不错的性能，但Fable确实有它特别的地方。

I think a lot of people are going to move to Opus five and probably get pretty good performance there, but there's just something really special to Fable.

在Clay内部，我觉得会有很多用户坚持用Fable。

I think inside of Clay, we'll see a lot of users stick with that.

对。

Yeah.

你怎么看这次发布的时机？

What do you make of the timing of the release here?

我们经常和开发者讨论一个循环，就是当你的公司上头条的时候，当你的公司得到好评而不是差评的时候，你决定发布Opus five。

One of the things that we we talk to developers a lot about is this cycle of when your company in the headlines, when your company is getting good headlines versus not so great headlines, when you decide to release Opus five.

你觉得这个发布时机在战略上有什么讲究吗？

I mean, does it feel like strategic timing to you in terms of the release schedule?

很难说。

Hard to say.

你知道，最近我们在Clay内部经常思考的一件事就是那些开放权重模型，以及它们离前沿有多接近。

Think, you know, the recent I think the recent thing that we think a lot about internally at Clay is just all of the open weight models and how close that those are getting to the frontier.

所以我觉得Opus five的发布时机和这个有挺有意思的联系。

So I think there, it's really the timing of Opus five seems to relate to that in a pretty interesting way.

我们花了很多时间把很多工作流程迁移到开放权重模型上。

I think, you know, we're spending a lot of time really migrating a lot of our workflows to open weight models.

而且这个领域显然有很多动态。

And there's been obviously a lot of movement in that space.

好，那多说说这个。

Well, and say more about that.

具体来说，你们在开源权重模型里用的是哪些模型？

What what exactly which models are you using in the open weight category?

对。

Yeah.

我觉得开源权重模型对我们来说更难去评估。

I I think like the open weight models are much harder for us to to think about.

如果看我们内部的评估结果，开源权重模型的表现往往很不稳定。

I think if we look at our internal EBALs, open weight models tend to be very spiky.

有些评估项目它们表现非常好，但有些项目它们比很多前沿模型差得多。

So some of the evals they do really, really well at and some of them they're much worse than many of the frontier models.

我们现在在Clay里广泛使用GLM 5.2和Community Cage 2.6。

We use GLM 5.2 and Community Cage 2.6 pretty extensively at Clay now.

我们刚把它集成到Claygent里，那是我们的主要研究代理，但我们也用在Sculptor里，那是我们的市场推广工程代理。

We just launched it in Claygent, is our main research agent, but we also use it in Sculptor, which is our go to market engineering agent.

所以更多是典型的编码相关用例。

So more of a typical kind of coding adjacent use case.

我觉得，你知道，对于GLM 5.2，有时候它在我们的评估中排名第一。

I think there, you know, we're seeing for GLM 5.2, sometimes it's at the top of our evals.

但它的成本真的只是更便宜的OpenAI模型的一小部分。

But it's truly a fraction of the cost of even, you know, the cheaper OpenAI models.

所以它确实占据了一个相当有趣的位置，至少在我们的模型使用中是这样。

So it's really kind of a pretty interesting occupies a pretty interesting space, at least in our, you know, in our model usage.

那么现在Clay使用模型的比例中，开源权重模型和前沿模型各占多少？

So what percent of model use Clay is now open weight models versus frontier models?

我们仍然以使用前沿模型为主。

We're still majority frontier models.

好的。

Okay.

我觉得这个比例正在快速上升。

I would say it's increasing as a percentage fairly rapidly.

我认为我们工程团队在这里也投入了很多时间。

I think we're investing a lot of time on the engineering team here as well.

那回到你用的Kimi K。

And so you I'm going go back to you're using Kimi K.

你用的是Moonshot最新的模型，还是之前的版本？

Is it the most recent Moonshot model that you're using, or is it models before that?

2.6。

2.6.

2.6。

2.6.

好的。

Okay.

跟我讲讲你们在选择海外模型和美国模型时是怎么考虑的。

Walk me through your considerations here on using models from overseas versus models from The US.

Clay，简单回顾一下，你们正在打造的是AI现在所支持的面向市场的工程师，也就是销售工程师的角色。

Clay, just to recap, you guys are building out the go to market engineer, the sales engineering role, I guess, that AI has now enabled.

所以，你们从客户那里获得了大量数据，比如他们的线索、客户对话记录之类的。

So, I mean, you have a lot of data from customers about their leads, their notes from customer conversations, stuff like that.

这些都是相当敏感的信息。

And this is all fairly sensitive stuff.

那从安全角度来看，你是怎么考虑使用海外模型的决定的？

So how do you think about the decision to use models from overseas from a security perspective?

对。

Yeah.

需要说明的是，我们所有的模型都托管在Base 10和Fireworks上。

And it's important to note, we host all of our models with base 10 and fireworks.

所以所有的模型都是——对。

So all of the models are yeah.

我们使用的是位于美国境内的推理服务提供商。

So we host with inference providers that are located stateside.

但确实，这是个值得思考的问题。

But yeah, it's it's an interesting consideration.

而且我觉得，尤其是考虑到特朗普政府与一些开源权重模型之间可能存在的冲突或推动，我认为美国拥有像NVIDIA、Thinking Machines等公司推出的那些强大的开源权重模型，确实非常重要。

And I think especially with sort of the push or the potential conflict between the Trump administration and some of the open weight models, I do think it's very important for The US to have strong open weight models like the ones that we're seeing from NVIDIA and Thinking Machines and others.

太好了。

Great.

杰夫，最后一个问题。

Last question for you, Jeff.

你看，当你看到Opus 5的时候，我想知道这对Anthropic整个模型系列的战略意味着什么。

I mean, when you look at Opus 5, I wonder what it tells you about Anthropic's strategy overall with their suite of models.

我在想，哪家公司把自己定位在高端模型领域，哪家又是成本领先者？

I mean, I sort of think about which model Which company is positioning themselves in the premium model category versus the cost leader category?

Fable是个昂贵的模型。

Fable is an expensive model.

当你想到Sonnet、Haiku、Opus，还有Fable系列，你觉得Anthropic会怎么发展这个模型系列，特别是考虑到我们现在看到的Opus？

When you think about the Sonnet, the Haiku, the Opus, and then the Fable family, where could you see Anthropic taking this suite of models given what we're seeing with Opus?

对。

Yeah.

我觉得Anthropic的战略很有意思，因为我们看到他们基本上在放弃市场的低端部分。

I think Anthropic strategy is very interesting here because we're seeing basically they're moving away from the lower end of the market here.

听起来他们不会推出HYCU 5了。

It sounds like they're not gonna have a HYCU five.

对我们来说，我们真的把Opus 5看作是Sonnet 4.6的精神继承者。

And I would say for us, we really think about Opus five is almost the spiritual successor to Sonnet 4.6.

实际上，我们之前在Sonnet 4.6上跑的好多工作负载，现在都会转到Opus 5上，因为成本增加确实带来了性能上的显著提升，这是我们在早期模型迭代中没怎么看到的。

Actually, a lot of our workloads that we were running on Sonnet 4.6, we're going to run those on Opus five now because the increase in cost actually does come with a material increase in performance in a way that, you know, we weren't really seeing with earlier iterations of these models.

所以我有点觉得，所有这些模型实际上都在往上提升一个级别，对吧？

And so I kind of think about it as all of these models are effectively moving up a weight class, right?

以前我们在Sonnet上跑的东西，现在都会考虑用Opus 5.4。

Anything that we were running on Sonnet will now consider Opus 5.4.

而那些我们追求顶级性能的前沿工作负载，我们会认真考虑用Fable或者类似Fable级别的模型。

And any of the frontier workloads that we have where you're really looking for the best in class performance we'll really consider Fable or kind of Fable class models there.

或者你也可以直接把所有东西都开源，听起来大家反正都想这么做。

Or you could just move everything open source, which it sounds like everyone's interested in doing anyway.

对。

Yeah.

确实，我觉得现在这真的是大家讨论的重点了。

It's absolutely I think that's it's it's really in the conversation now.

对。

Yeah.

我觉得我们现在真的会主动考虑这个问题，而一年前我们虽然也在用开源模型，但真的不会把它放在这种前沿工作负载的考虑范围内。

I think really something we think about actively in a way that a year ago, we were, you know, we were using open source models a year ago, but it really wasn't in consideration for this tier frontier workloads.

那对你们来说，用开源模型是不是能省下不少成本？

And it is a it's a material cost advantage for you guys to use open source?

当然。

Absolutely.

对。

Yeah.

我觉得不管是节省的比例，还是说我们公司整体推理开销本来就很大，所以任何一点边际改进对我们业务来说都很重要。

I I think both in terms of the percentage of savings, but Clay just has a lot of inference spend across the board.

所以任何一点小的改进对我们来说都是实实在在的。

And so any kind of marginal improvement is material to our business.

太好了。

Great.

好的，杰夫，感谢你来做客。

Well, Jeff, I want to thank you for coming on.

这位是杰夫·巴格，Clay公司的人工智能负责人，在TITV节目上。

That is Jeff Barg, head of AI at Clay here on TITV.

谢谢。

Thanks.

Moonshot AI的Kimi K3引发了关于开源和闭源人工智能的广泛讨论。

Moonshot AI's Kimi K3 has sparked much debate around open and closed source AI.

与此同时，OpenAI的Hugging Face黑客事件引发了许多关于人工智能安全的新担忧。

Meanwhile, OpenAI's Hugging Face hack debacle has raised a lot of new concerns about AI security.

我想请Theresa Payton来谈谈她对这一切的看法。

I want to bring on Theresa Payton for her view on all this.

Theresa曾在乔治·W·布什总统任内担任白宫首席信息官。

Theresa was Chief Information Officer at the White House under President George W.

布什。

Bush.

她现在担任Fortilace Solutions的首席执行官。

She is now the CEO of Fortilace Solutions.

特蕾莎，感谢你来到节目。

Theresa, thank you for being here.

谢谢你们的邀请。

Thanks for having me.

好的。

Okay.

那么，请谈谈你对开源和闭源AI之争的看法，这场争论一直在进行。

So tell me a little bit about your view on the open source, closed source AI debate that has been playing out.

你对此有什么反应？

What's been your reaction to it?

是的。

Yeah.

我觉得现在这场辩论真的很有意思。

Mean, think it's a really interesting debate that we have going on right now.

我给企业高管们的建议是，考虑到目前AI的发展状况，以及我们还不清楚什么会让企业陷入困境、什么能让企业脱困，最好还是多选几家供应商，从韧性、可恢复性、安全性这些角度来考虑，同时也要确保供应链上不会因为某个关键任务的东西被切断而受影响。

What I am advising business executives is, obviously, given where we are with AI right now and not knowing what puts a firm in the penalty box and what allows the firm to come out of the penalty box, that's a good idea to have more than one vendor and to think about that from the standpoint of resiliency, recoverability, security, but also just making sure from a supply chain, you don't get cut off from something that's considered mission critical.

现在这场争论真的很有意思，因为你看那些重量级玩家，很多都在美国硅谷搞AI，你开始看到两个阵营出现，一边是主张允许开源和开放权重模型，这些模型成本更低，可能没那么好，但也许够用，另一边是闭源模型的概念。

The argument that's happening right now is really interesting to watch unfold because as you see some of the heavyweights, many of which are in sort of The United States, Silicon Valley in AI, you're starting to see two different camps emerge as to whether or not we should allow this open source and open weight models, many of which are lower cost, maybe not as good, but maybe good enough, versus the closed model concept.

我觉得这里面还有很多东西需要深入探讨，比如哪个——对。

And I think there's a lot more to unpack there as to which one- Yeah.

你在这个辩论中站在哪一边？

Where do you stand on that debate?

开源模型应该受到监管吗？

Should there be regulation on open source?

任何模型都应该有治理和护栏的监管，不管是开源的还是闭源的。

There should be regulation of governance and guardrails on any of the models, whether it's open source.

我们现在就在用开源代码库，不是AI，而是开源代码，这在过去几年里确实大大推动了开发团队的进步。

We use open source code bases today, not AI, but open source code, which has really helped propel development teams forward over the past several years.

但两者都需要治理和护栏，而不是只针对其中一个。

But we need governance and guardrails for both, not just one versus the other.

我一直在想，所有这些讨论都是因为KimiK3太强大了才引起的。

I've been trying to figure out So all this conversation came up because KimiK3 was so powerful.

大家都对它印象深刻。

People were very impressed by it.

最初的讨论是关于对来自中国和海外开源模型的监管。

The conversation initially was around regulation of open source models coming from China and from overseas broadly.

后来就演变成了对整个开源生态的讨论。

That then drifted into a conversation around open source altogether.

我一直在琢磨，为什么公司会采取这种立场，坚持要对开源AI保持开放态度，而Kimi K3的问题其实更具体一些，对吧？

And I've been trying to figure out why it is that companies have taken this approach that we need to remain open on regulating open source AI when The Kimi K3 issue, it was a little more isolated than that, right?

我的意思是，问题在于这个模型来自中国，这才是让人担心的点。

I mean, this was an issue of the model coming specifically from China that was the concern.

为什么这两个问题会被混为一谈呢？

Why did these two issues get conflated?

对。

Yeah.

可能是因为现在AI相关的动态太多了，所以这两个问题就被混在一起了。

It could be that there's just so much going on right now as it relates to AI that the issues got conflated.

我想说的是，企业必须考虑成本和收益。

What I would say is this, businesses have to look at dollars and cents.

他们必须关注自己的底线。

They have to look at their bottom line.

而且我们看到，闭源模型的价格在上涨，有些情况下，很多公司都在说，嘿，以前只需要几个token就能搞定的事，现在在新发布的闭源模型里，要花更多的token，花更多的钱。

And as we see the closed models ratchet up in price, and in some cases, many firms are saying, Hey, what used to only cost me a couple tokens in the past in the newer closed models that are being released, it's costing me more tokens and costing me more money.

所以他们开始寻找更经济实惠的替代方案，这些方案可能效果一样好，或者至少够用。

And they're looking at more alternatives that are more cost effective and maybe just as good or just good enough.

所以很有意思的是，在讨论中几乎能看到一种竞争态势在慢慢形成。

So it's very interesting to see almost like competition, thing sort of evolving here in the discussion.

我想说的就是，自由市场竞争是好事。

What I would just say is, like, free market competition is good.

我总想提醒大家的是，要小心。

What I would always caution is be careful.

确保你花的钱物有所值。

Make sure you get what you pay for here.

而且你要确保自己有内部治理和防护措施，保证你的业务安全可靠，如果你使用开源模型的话，不管它们来自哪里，在哪个国家开发的。

And you wanna make sure that you've got your own internal governance and guardrails to make sure that your business is safe and secure if you're using open source models, regardless of their origin story, which country they were built in.

你觉得为什么中国在开源模型方面特别有优势？

Why do you think China has developed such an advantage in open source models specifically?

这里面有几个因素。

Well, there's a couple of things at play here.

首先，他们被禁止使用我们最强大的芯片，而需求是发明之母。

First of all, they were banned from using some of our most powerful chips and necessity is the mother of invention here.

他们只能用那些不如美国公司能获得的芯片好的芯片来凑合。

They had to get by, with chips that weren't as good as some of the chips that many of the American companies had access to.

另外，多年来，中国虽然否认，但他们确实非常擅长模仿技术。

You also see, over the years that, China, although they deny this, they are incredibly good at doing sort of copycat technology.

所以就是研究一下，看看大家都在用什么，然后利用知识产权，做出差不多够用的模仿技术。

So sort of studying, looking at what's being used and then being able to leverage intellectual property and doing, you know, just good enough and copycat technology.

我听那些拿到模型的人说，这些模型在先进程度、复杂度和token使用优化上，不如一些美国的闭源模型。

I have heard from people who've gotten their hands on the model that the models are not as advanced and sophisticated and optimized for token use as some of the American closed models are.

但对很多企业来说，可能已经够用了。

But for many businesses, that may be just good enough.

现在，我觉得关于这些模型是否真的存在安全风险，因为它们是来自中国的，这个问题还没有定论。

Now, I think the debate is still out whether or not there is a true security risk with these models just because they come from China.

我很希望看到实验室里进行测试，并且看到测试结果。

I would love to see tests run-in the lab, and I would like to see the results of those tests.

所以，你还没有对这些模型是否存在安全风险形成看法？

So you haven't yet made an opinion on whether or not you think there is a security risk from these models?

我觉得可以假设有可能存在风险。

I think it's okay to assume there might be.

但在我们仅仅因为一个软件产品来自中国就指责它之前，还是先拿出证据来再说吧。

But before we start accusing just because a software product had an origin story of China, let's put the proof behind the details.

那就动手干吧。

Let's do the work.

我们进实验室，看看有哪些漏洞，然后让企业根据事实自己决定，他们能不能真正解决已经发现的安全漏洞。

Let's get in the lab, see what the vulnerabilities are, and then let businesses decide for themselves based on the facts, can they actually mitigate any security vulnerabilities that have been found?

现在我们就只是说，哦，它来自中国。

Right now, we're just saying, Oh, it's from China.

那肯定很危险。

It must be dangerous.

我们真的需要看到那些测试结果。

We really need to see those results.

你的团队或者你合作的那些公司，有没有——你是在网络安全领域工作得很深入的。

Have you has your team or the orbit of companies that you work with I mean, you you work in the cybersecurity space very closely.

你的事业就是建立在这个基础上的。

That's what you've built your practice on.

有没有什么早期的迹象或者结果？

Are there any early indications, results?

你有没有做过一些早期测试，能让我们对这个问题有所了解？

Have you ran any early tests that could give us a view into that at all?

是的。

Yeah.

我看到的早期测试主要是关于，如果你下载了模型并在自己的沙箱里使用，你实际上怎么保护它、确保它的安全。

The early tests that I've seen have been looking at, if you've downloaded the model and you're using it in your own sandbox, how can you actually protect and secure it?

这本身也有缺点，对吧？

That has its own downsides, right?

因为你必须确保能更新模型，而且更新模型不会破坏你已经构建的那些东西的流程。

Because you have to make sure you can update the model and that you're not going to break the threads of things that you've built by updating the model.

但目前大部分关注点还是这些模型在解决问题上的效果如何。

But most of the focus right now has been on how effective these models are in solving problems.

需要多长时间？

How long does it take?

成本有多高？

How expensive is it?

至于发现重大国家安全漏洞，我的团队还在和进行这些测试的团队沟通。

As far as coming up with the major national security vulnerability, my team is still in conversations with teams doing those tests.

我们还没有听到任何确凿的消息说，就是它了。

We have not heard anything definitive that says, here it is.

这就是确凿的证据。

Here's the smoking gun.

但这需要时间。

But this takes time.

找到那些隐藏的后门需要时间。

It takes time to find those hidden back doors.

我们——这需要几周还是几个月？

We have- Is this weeks or months?

我以前从来没做过这些测试。

I've never done any of these tests before.

那么我们应该预期需要多长时间？

So how long should we expect it to take?

对。

Yeah.

分析通常只需要几周时间，但我要告诉你，即使是最好的软件，由最优秀的专家操作，也可能会漏掉一些东西。

The analysis oftentimes only takes a few weeks to do, but I will tell you that even the best trained software looking for these vulnerabilities powered by the best practitioners, you still might miss something.

没错。

Right.

说到这个，那我们聊聊上周OpenAI和Hugging Face那件事吧。

Well, and to that point, so then let's talk about the OpenAI Hugging Face debacle from last week.

你看到那个头条新闻的时候惊讶吗？

Were you surprised when you saw that headline?

我并不惊讶。

I wasn't surprised.

关于这件事，有几点要说清楚，它是什么，不是什么。

A couple of things about what this is and what this is not.

是的，很明显，就是AI突破了护栏，去攻击了另一家公司。

Yes, obviously what it is, is AI jumped the guardrails, went and hacked another company.

那家公司为了在搞清楚状况的同时保护自己、进行防御，实际上就像被困在壁橱里的消防员一样，出不来。

The other company, in order to be able to protect and defend itself while it was trying to figure out what was going on, they actually were almost like firefighters locked in a closet, couldn't get out.

他们实际上使用了开源的、开放的调用。

They actually use open source, open wide calls.

对。

Right.

我是说，这就是我今天早上在报告里看到的，结果是中国的一些开源模型在保护他们方面表现得更好。

Mean, that's what I read in the report this morning, that it ended up being that open source models from China were actually better at protecting them.

是的，完全正确。

Yeah, exactly.

现在，我们想说的是，这并非一个商业产品在商业企业里运行而越过了护栏。

Now, we do want to talk about here is this is not a commercial product running a commercial enterprise that jumped the guardrails.

这是实验室里进行的测试越过了护栏，然后攻击了另一家公司。

This was tests going on in a lab that jumped the guardrails and hacked another company.

所以我认为我们得向实验室提出更多问题，就像我们向研究活病毒的实验室提问一样。

So I think we have to ask the lab a lot more questions, just like we ask labs that work with live viruses.

你们采取了哪些防护措施，确保它不会直接冲出门去感染公众？

What protocols are you taking to make sure it doesn't just jump through the doors and start infecting the public?

我们在这里也得做同样的事情。

We have to do the same thing here.

我开始觉得，这些实验室里进行的测试，缺乏那种类似危险品处理的治理和防护措施。

I'm starting to feel like there's not enough of those hazmat like situation governance and guardrails on the tests that are being done in these labs.

所以，说实话，发生这种事我并不意外，但我们可以从中吸取很多教训。

So sadly, I'm not surprised that it happened, but there's a lot of lessons learned that we can take away from this.

首先，如果你要在这些实验室里做测试，确保你有正确的治理和防护措施，这样你就不会去攻击另一家公司。

Starting with, if you're going to be doing the tests in these labs, make sure you've got the right governance and guardrails so you're not attacking another company.

别再这样了。

Let's not do this again.

别再犯同样的错误了。

Let's not repeat the same mistake.

对。

Right.

我想问问你关于监管的问题，总的来说，考虑到你之前在政府工作过一段时间。

I want to ask you about regulation, broadly speaking, given that you were in a government position for a little while.

我们节目一直在思考政府可以采取哪些手段来管理和监管人工智能。

We've been thinking on the show here about the levers that the government could pull as it relates to governing AI, regulating AI.

一方面，芯片过去一直是监管的重点。

On one hand, you have the chips, which has been a point of regulation in the past.

还有数据中心，以及数据中心实际所在的位置。

We also have the data centers and where the data center is actually located.

还有软件。

You have the software.

如果政府决定区分开源和闭源，你觉得他们最可能动用哪个手段来监管AI？

Which lever do you think is most likely to get pulled here as the government seeks to regulate AI if they decide on delineating between open source and closed source?

具体操作上你会怎么做？

How do you actually do it tactically?

嗯，我觉得我们现在面临的挑战是激励机制不一致。

Well, I would say the challenge that we have right now is incentives are not aligned.

所以现在有企业的利益、国家安全的利益，有那些AI公司，还有那些想搞清楚怎么用AI的高管们，还有像你我这样的普通公民。

So you have the incentives of businesses, incentives of national security, so you have the businesses that are the AI companies, then you have business executives who are trying to figure out how to implement AI, then you have ordinary citizens like you and me.

所以有这么多不同的利益相关方，而我们的利益并不一致。

So we have all these different stakeholders and our incentives are not aligned.

所以如果我们能找到那些共同点，让利益一致起来，那我们在一些关键问题上就能达成一致，比如数据中心建在哪里，外观是否美观，是否环保，以及邻居的感受。在安全、安保、伦理和韧性方面的监管上，我们必须让利益一致。

And so if we can find those common grounds where we can have alignment, then we will have alignment around some key issues around where do data centers get built and how aesthetically pleasing and environmentally friendly and neighbors As it relates to regulation around safety, security, ethics, and resiliency, we have to align incentives.

比如，现在很多公司本身并不盈利。

So for example, many of these companies today are not inherently profitable.

如果给他们研发税收抵免，而且这些抵免不会过期，或者有效期长达一百年，这样将来他们赚大钱的时候，现在就能做出正确的决定，因为他们有正确的利益激励呢？

What if they were given R and D tax credits that didn't expire or had a shelf life of one hundred years so that one day when they are incredibly profitable, they made the right decisions now because they had the right aligned incentives?

这就是我看到的问题所在。

That's where I see the disconnect.

所以现在有很多事情要关注。

So there's a lot of things to focus on right now.

但如果我们把这些都抛开，坐下来好好想想：哪些地方的利益是一致的？

But if we actually strip all of that away and actually sit down and say, Where are incentives aligned?

我们在哪些方面都希望彼此过得好？

Where is it that we all want the best for each other?

然后围绕这一点把相关的事情落实到位。

And then get the things in place around that.

但我想问的是，抛开这些不谈，你觉得政府应该把精力重点放在哪里？

But I'm asking So when you strip it all away, I'm sort of asking, Where do you think the government should focus its efforts?

是应该监管芯片吗？

Should it be on regulating the chips?

还是应该管数据中心的位置？

Should it be on the data center location?

还是应该审查模型发布的节奏和安排？

Should it be on the reviewing the model release cadence schedule?

你觉得他们应该先动哪个杠杆？

Which lever do you think they should pull on first?

我觉得这个问题挺难的。

I think this is a tough one.

我的意思是，显然AI是另一种技术。

I mean, obviously AI is another technology.

就像我们今天有数据中心来处理银行信息或医疗信息一样。

Just like we have data centers today for processing of banks' information or healthcare information.

明天我们也需要数据中心来支持AI和其他即将到来的技术。

We need data centers tomorrow for AI and other technologies coming.

所以我想说的是，我们需要关注技术本身的治理和防护措施，但很快它就需要算力来运行。

So what I would say is that we need to focus on the governance and guardrails of the actual technology itself, but pretty quickly it needs compute power to run.

所以我们真的必须现在就进行这场辩论，而不是以后，关于数据中心、算力、芯片。

So we really have to have this debate now sooner rather than later around data centers, compute, chips.

坦白说，其中很多都是老问题，我们一直拖延着没解决。

And candidly, many of these are age old problems that we've been kicking the can down the road for a while.

只要

Just

把治理和防护措施放在软件上，其他的自然就会跟上。

put the governance and guardrails on the software, and then the rest should follow.

对。

Right.

好的，特蕾莎，感谢你来做客。

Well, Theresa, want to thank you for coming on.

这位是Fortilace Solutions的首席执行官特蕾莎·佩顿，欢迎来到TITV。

That is Theresa Payton, CEO of Fortilace Solutions here on TITV.

谷歌上周公布了季度业绩。

Google reported quarterly results last week.

我们的AI和金融记者达金·坎贝尔带来了新的分析，探讨为什么谷歌的资产负债表可能低估了公司承担的风险。

Our AI and finance reporter, Dakin Campbell, has new analysis on why Google's balance sheet might underrepresent the amount of risk that the company is taking on.

我们想请达金来聊聊到底是怎么回事。

Want to bring on Dakin to explain what exactly is going on here.

达金，欢迎回到节目。

Dakin, welcome back to the show.

很高兴你能来。

It's great to have you here.

是的，非常感谢。

Yeah, thanks so much.

好的。

Okay.

那么，达金，你写到谷歌正在借鉴华尔街的做法，尤其是在债务和资产负债表方面。

So Dakin, you wrote about how Google is taking a page out of Wall Street's playbook, especially with respect to how it's thinking about debt and loading up its balance sheet.

给我们讲讲你的分析吧。

Explain to us a little bit about the analysis that you did.

好的。

Yeah.

我们看了他们的资产负债表，发现从去年开始，他们报告了一种新的衍生品，这在谷歌的财报中以前从未出现过。

So we were looking at their balance sheet and we noticed that they had starting last year, they had a new kind of derivative that they were reporting on their balance sheet, which we had not seen before in Google's results.

这和他们为数据中心租赁提供的担保有关。

And it relates to these backstops that they are providing for data center leases.

谷歌已经承诺为数据中心租赁提供约440亿美元的担保。

So Google has committed to about $44,000,000,000 worth of backstops on leases for data centers.

这些公司是建数据中心的，然后跟第三方签租赁合同。

So these are companies that are building data centers and leasing with a third party.

很多情况下，这个第三方是FluidSAC。

In many cases, it's FluidSAC.

为了能卖出债券来筹钱建数据中心，他们让谷歌出面，基本上就是给这些租赁合同做后端担保。

And in order to sell the debt that will allow them to build the data centers, they're getting Google to come in and basically guarantee the leases on the back end.

那些买债券的投资者，他们看的就是这些租赁合同。

The debt investors and people who buy the bonds are looking at the leases.

因为那是收入来源。

That's the money that's coming in.

所以他们担心这笔钱能不能持续进来，用来还债券。

And so they're worried about whether that money is gonna keep coming in to the to help repay the bonds.

所以谷歌基本上就是承诺兜底这些租赁。

And so Google is is basically committing to backstop those.

所以，所以，他们就是担保人。

So so so they're they're the guarantor.

我的意思是，这个，这个，这个

I mean, the the the

对，就是，是的。

the the yes.

这几乎就像保险一样。

It's almost like insurance.

他们提供的几乎就像保险。

It's almost like insurance that they're providing.

或者像纽约市的租房担保人，很多时候他们的要求相当严格。

Or or like New York City rental guarantors, which should they're be pretty they're pretty steep requirements a lot of the time.

好的。

So okay.

那么谷歌作为担保人，它在这里担保的规模有多大呢？

So Google the guarantor, what is the magnitude of guaranteeing, I guess, that Google is doing here?

是的。

Yeah.

所以到目前为止，他们已经承诺了438亿美元的担保。

So they have committed to $43,800,000,000 of guarantees so far.

他们在脚注里还提到，可能还会有240亿美元的担保。

They've said in the in further in the footnotes that they are that there may be another $24,000,000,000 in guarantees.

所以加起来，目前账上或者即将到来的担保总额是680亿美元。

So added up that 68,000,000,000 currently on the books or coming.

而且，我们接触的一些信用评级分析师中，有些人预期这种情况还会继续。

And there's an expectation among among some people, among, you know, credit rating analysts that we talked to that this will continue.

所以，你知道，再过几个季度，可能就会接近800亿甚至1000亿美元了。

So, you know, it's getting a few more quarters of this, and and you might be getting closer to, you know, 80 or or a $100,000,000,000.

好的。

Okay.

那我们现在就来谈谈核心问题吧。

So now let's get to the heart of the matter here.

如果这些担保出了问题，可能会有高达1000亿美元的潜在义务，这到底是怎么回事？

How much a $100,000,000,000 may be coming in in possible, you know, obligations if any of this goes south.

其中有多少反映在它的资产负债表上？

How much of that is represented on its balance sheet?

因为他们把它作为衍生品记录在资产负债表上，所以他们基本上是在看这个风险敞口。

So because they have represented it on the balance sheet as a derivative, they basically look at this exposure.

所以目前是44。

So right now, it's 44.

从100降下来的。

Just bring it back from a 100.

目前是44。

Right now, it's 44.

他们评估自己需要赔付这些款项的可能性有多大。

They look at how probable they think it is that they're gonna have to pay out on these.

目前，他们认为已经将其作为衍生品记录在资产负债表上，并且标记为8.15亿美元。

And right now, they think they've recorded it as a derivative on their balance sheet, and they've got it marked at 815,000,000.

所以还没到10亿。

So we're not at a billion yet.

那大概是多少？

That's like what?

那大概就是四十四分之一吧。

That's like one by 44.

那基本上就是百分之二左右。

That's like 2%, basically.

对。

Yes.

大概是百分之一点八三左右。

It's 1.83% or something.

对。

Yes.

好的。

Okay.

那人们是不是

Do do people