本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
欢迎各位收看Information的TI TV。
Welcome everyone to the Information's TI TV.
我叫阿卡什·马斯里查。
My name is Akash Masricha.
今天是2月24日，星期二。
It is Tuesday, February 24.
今天首先带来一些重要的芯片新闻。
First up today, some big chip news.
Meta与AMD达成了一项六吉瓦的计算协议。
Meta has struck a six gigawatt compute deal with AMD.
我们将邀请分析师为您详细解读。
We'll break it down with an analyst.
此外，Information还发布了深度报道，探讨软件公司如何反驳SaaSpocalypse这一说法。
Plus, the information published in-depth reporting around how software companies are pushing back against the SaaSpocalypse narrative.
我们将与我们的企业软件记者对话，聊聊他的报道以及他所观察到的情况。
We'll talk with our enterprise software reporter about his story and what he is seeing.
接下来，我们将与Notion的AI负责人对话，了解他们今天推出的全新智能代理，以及在AI时代对按使用量计费的思考。
We'll then speak with Notion's AI lead about new agents it is launching today and how it is thinking about usage based pricing in the era of AI.
我们还独家报道了Anthropic公司内部的50个研究项目，以及该公司对AI代理安全的专注。
We also have exclusive reporting on 50 of Anthropic's internal research projects and the company's focus on AI agent security.
最后，我们将与布拉德利·塔斯克对话，深入探讨硅谷当前一些重大的监管议题。
We will wrap with a conversation with Bradley Tusk and get into some of the big regulatory stories of the moment in Silicon Valley.
这将是一场精彩的节目，让我们马上开始。
It's going be a fun show, so let's get right on into it.
Meta与AMD达成了一项重要的芯片交易。
Meta has struck a major chips deal with AMD.
Meta同意向AMD采购价值6吉瓦的AI芯片算力，用于数据中心运营，作为交换，Meta将获得AMD最多10%的股份。
Meta has agreed to buy six gigawatts worth of compute of AI chips from AMD to power data centers, and in exchange, Meta will get as much as 10% of AMD stock.
这对AMD而言意义重大，因为其在AI芯片市场上的表现仍落后于英伟达，而这一消息也紧随上周Meta与英伟达宣布建立长期战略合作之后。
This is a big deal for AMD, which is trailing NVIDIA in the AI chip market, and it follows last week's announcement that Meta and NVIDIA are entering into a long term strategic partnership.
现在加入我们、对这笔交易做出分析的是创意策略公司的高级分析师奥斯汀·莱恩斯。
Joining me now to react to the deal is Austin Lyons, a senior analyst at Creative Strategies.
奥斯汀，欢迎来到节目。
Austin, welcome to the show.
很高兴你今天能来。
It's great to have you here.
嘿，卡什。
Hey, Kash.
谢谢您今天早上邀请我。
Thanks for having me this morning.
那你对这笔交易怎么看？
So what did you make of the deal?
我对这笔交易怎么看？
What did I make of the deal?
首先，英伟达的财报明天发布，所以我原本以为今天会比较平静，但我学到了一课：今天其实有很多新闻，因为我觉得大家都想在英伟达财报发布前把消息放出来。
Well, first of all, NVIDIA's, earnings are tomorrow, and so I thought maybe today would be quiet, but I think I've learned a lesson, which is there's lots of news today because I think everyone wanted to get their news out ahead of NVIDIA's earnings.
一大早我们就得知了AMD和Meta的消息，我认为这很好。
Bright and early, we found out about AMD and Meta, and I think good.
这对双方都有好处。
It's good for both of them.
Meta最近，实际上已经很久了，一直在说他们将不断增加资本支出，因为他们需要越来越多的计算能力来推动其依赖人工智能的广告业务。
Meta's been talking recently, and honestly, a long time, that they are gonna spend more and more CapEx because they need more and more compute to drive their advertising based business, which uses lots of AI.
我们知道这涉及NVIDIA。
We know that that involves NVIDIA.
我们知道这还涉及他们的MTIA定制芯片。
We know it involves their MTIA custom silicon.
但当然，他们长期以来一直是AMD的合作伙伴。
But, of course, they've long been partners with AMD.
这次的重大公告是关于Meta和AMD共同设计芯片和系统，真正地协调双方的利益，使这种合作关系能够延续多代，我认为这对双方都是双赢的。
And this was a big announcement about, really, Meta and AMD co designing chips to and systems together, and and sort of, like, aligning their incentives such that this will be a long term partnership for many generations, and I think it's win win for both sides.
现在，我想请你帮我们把这件事放在Meta签署的所有交易的大背景下来看。
Now, I want you to help us put this into context of all of the deals that Meta is signing.
我们看到了NVIDIA的交易。
So we saw the Nvidia deal.
我们还报道过Meta可能与谷歌达成协议，采购他们的TPU。
We also did some reporting on Meta's potential deal with Google to sign a contract to get their TPUs.
我的意思是，现在企业分散投资、对冲风险是否已成常态？还是Meta的做法有什么独特之处？
I mean, is this sort of par for the course now in terms of companies spreading their bets and hedging their bets, or is there anything unique about what Meta is doing?
是的，这里有很多事情在发生。
Yeah, there's a lot going on here.
归根结底，我们知道Meta为了提升业务，对算力有着巨大的需求。
So at the end of the day, we know that there's a ton of demand for compute for Meta to improve their business.
他们需要各种类型的AI算力。
They need all sorts of AI compute.
从供应端来看，台积电生产了所有这些算力芯片。
On the other end of the spectrum, from a supply perspective, TSMC makes all that compute.
因此，在这个供需两端之间，人们正在深入思考如何尽可能多地获取芯片资源。
And so there is a lot of thinking going on in between that spectrum of how do we get access to as many chips as possible.
许多公司都与台积电签订了晶圆协议，而Meta正与所有这些公司合作，只为争取获得尽可能多的算力。
There's lots of companies that have agreements with TSMC for wafers, and Meta is working with all of them to really just get us access to as much compute as possible.
当然，这里存在一些权衡。
Now, there are trade offs, of course.
你需要编写能在所有这些不同系统上运行的软件，但显然，Meta 对这种从不同公司获取多种计算资源组合的策略感到安心，并正在设法让他们的工作负载在这些系统间兼容。
You have to write software that works on all these different systems, but clearly Meta is comfortable with this idea of having a portfolio of different compute offerings from different companies, and they're figuring out how to make their workloads work across them.
你们有没有做过任何分析，或者想过六千兆瓦的计算成本会是多少？
Have you done any analysis or have you thought at all about what the cost of six gigawatts of compute would be?
我还在想，Meta 通过这种安排，获得了 AMD 股份的一定比例。
And the way I'm sort of thinking about this as well is, okay, Meta is getting a certain portion of AMD shares in exchange for this arrangement.
这能在多大程度上抵消成本呢？
Does that kind of offset the cost at all?
他们是否得到了某种回报？
Is it getting something in return?
你怎么看待这个问题？
How do you think about that?
是的。
Yeah.
是的。
Yeah.
所以AMD表示，每吉瓦的计算能力能带来数百亿美元的收入。
So AMD said this is tens of billions of dollars of revenue per gigawatt.
如果你仔细计算一下，并且把那些认股权证看作是一种折扣，你可以说，好吧。
If you crunch the numbers and if you're just looking at if you think about these warrants ultimately as sort of a discount, you could say, okay.
那么Meta因此获得了公司多少股份？
Well, how much of the company is Meta getting in return?
这部分价值多少？
How much is that worth?
相对于芯片的原价，这个折扣是多少？
And what is sort of the discount off the price of the chips?
今天上午的电话会议上，丽莎对此提出了反驳，说：
On the call this morning, Lisa pushed back on that thinking, say saying, hey.
这种想法太狭隘了。
That's too narrow minded.
尽可能从Meta今天提取更多收入。
Extract as much revenue from Meta today as possible.
目标是真正与Meta共同设计和开发，让AMD以Meta作为首个重要客户取得成功。
The goal is to really codesign and develop and for AMD to really win with Meta as a first big best customer.
他们也在与OpenAI这样做，AMD将从长远中受益，并在未来与其他客户合作时捕捉这部分价值。
They're also doing this with OpenAI, and that AMD will benefit from that in the long run and sort of capture that value later with all the other customers.
所以，想象一下，AMD所有潜在客户都在看到：哇，OpenAI正在大量部署AMD。
So think about it as all of AMD's potential customers looking and saying, wow, OpenAI is deploying a lot of AMD.
Meta也在大量部署AMD。
Meta's deploying a lot of AMD.
AMD通过与这些重要客户合作开发下一代产品，学到了大量知识。
AMD's learning a ton by working with these big best customers in developing future generations.
我感觉相当有信心。
I feel pretty confident.
也许我应该把AMD视为一个潜在的投资标的。
Maybe I should look at AMD as a potential offering.
所以，拥有这些巨型交易和巨型客户，几乎就像是让AMD能够真正地——我们在这个节目中已经讨论过这一点，当时是在NVIDIA的背景下谈的，但很难预见到在如此大规模的部署中会出现的所有挑战。
So it's almost like having these giant deals and having these giant customers will allow AMD to really and we've talked on this show about how we talked about it in the context of NVIDIA, but it's hard to anticipate all the challenges that will come up when you have these large scale implementations.
而能合作的顶级客户数量其实非常有限。
And there's only a couple of the biggest clients to go around.
所以我听到你的意思是，AMD与Meta达成了这笔交易，得以大规模部署他们的芯片，并从中学习，这实际上能帮助他们在未来开发出更先进的芯片。
And so what I hear you saying is AMD gets this deal with Meta, they get to deploy their chips at scale, they get to learn from that, that could actually help them develop further chips down the line.
是的，正是如此。
Yes, exactly.
关键是我们能否争取到几位这样的客户，在最大规模的生产环境中部署我们的芯片，并从中快速学习。
It's, can we get a few of those customers, deploy our chips in production at the biggest scale possible, and learn as rapidly from that.
你说得完全正确。
So you're exactly right.
他们本质上是获得了大规模运行和快速学习的许可，并以此为基础不断成长。
They're basically getting permission to run at scale and to learn quickly and to grow from there.
那么，奥斯汀，这是否还为芯片领域的初创公司，甚至更小的芯片公司留下了空间？
Now, Austin, does this leave any room for startups in the chip sector or not even startups, but smaller chip companies?
我的意思是，我们看到超大规模云服务商和像Meta这样的公司与英伟达、AMD之间的大单。
I mean, we see the big deals between the hyperscalers and the metas of the world, and then the Nvidias, the AMD's.
我的意思是，某种程度上，这真的变成了大者愈大的局面，对吧？
I mean, this is really turning into just the big getting bigger in some ways, right?
确实如此，情况比较复杂，但初创公司仍然有空间。
It is, and it's nuanced, but there's definitely still room for startups.
如今这种情况的背景是，AMD表示他们将为Meta的工作负载专门定制一款GPU，这实际上与其它初创公司试图参与的领域是一致的，比如最近Cerebras和Grok就频频上新闻。
So what the framing for what's happening today is AMD said we are going to make a custom variant of our GPU specifically tuned to Meta's workloads, And that's actually along the spectrum of where other startups are trying to play ball, whether, you know, Cerebras, Grok have been in the news recently.
但就在今天，我们看到Maddox融资了，SambaNova也融资了，最近Etched也完成了融资。
But then even today, we saw Maddox raised money, SambaNova raised money, recently Etched has raised money.
每个人都从不同角度定位自己，声称自己在某些特定工作负载、某些KPI或指标上表现最佳，比如超低延迟、超高吞吐量或超低成本。
And everyone is counter positioning and saying, hey, we are best for certain workloads, for certain sort of KPIs or metrics that you're tracking, whether that's ultra low latency or ultra high throughput or ultra low cost.
AMD历来专注于更灵活的GPU，你得到的就是你所看到的。
AMD has historically been focused on GPUs that are more flexible, and you just basically get what you get.
它有多少HBM、多少计算能力，就这么定了。
It's got this much HBM, this much compute, that's what you get.
AMD正在做的，是说我们可以根据你的工作负载来调整这些系统。
What AMD is doing is saying, Hey, we can actually tune these systems to your workload.
他们有一个名为芯片粒的概念，就像乐高积木一样。
So they have this idea of chiplets, which are like Lego blocks.
也许Meta的特定工作负载需要稍微少一点计算能力或更多内存。
Maybe Meta's particular workload needs a little bit less compute or a little bit more memory.
因此，他们正逐渐从仅仅提供现成的SKU，比如MI450X，转向根据你的工作负载进行一些定制调整。
And so they are starting to move down that spectrum away from just you get the particular SKU off the shelf, MI450X, to now like, oh, we're gonna tune that a little bit for your workloads.
这在一定程度上挤压了那些已经宣称要为你的特定指标打造高度定制化芯片的初创公司。
That does encroach a little bit on startups who are already saying, well, we'll build custom chips that are very hyper tuned to your specific metrics that you care about.
因此，这确实是一个大公司开始侵入这一领域的明显信号。
And so it definitely is the big sign of kind of encroaching on that.
但我认为，我们看到生成式AI所呈现的是，从实时语音到视频生成，再到前沿模型，各种各样的工作负载。
But I think what we're seeing with generative AI is there are all sorts of workloads from real time voice to video generation to frontier models.
因此，芯片对不同工作负载的需求将继续存在。
And so there'll continue to be different workload demands from chips.
我认为，初创公司仍然有机会进入市场，专注于某个特定的细分领域。
And I think there will still be opportunities for startups to come in and really hyper focus on a specific niche, if you will.
我们看到了OpenAI的交易。
So we saw the OpenAI deal.
我们看到了Meta的交易。
We saw the we saw the Meta deal.
你认为接下来谁会与AMD签下这样的大型交易？
Who do you think could be next to sign one of these giant deals for AMD?
你在关注谁？
Who are you watching?
这一切都关乎AI实验室大规模运行推理任务。
So this is all about AI labs running inference at scale.
Meta，我们看到了Meta的交易。
Meta, we saw Meta.
我们看到了OpenAI的交易。
We saw OpenAI.
这就引出了一个问题：Anthropic会是下一个吗？
It begs the question, could Anthropic be next?
显然，Anthropic正在与亚马逊就Trainium紧密合作，但每个人都需要更多算力。
Obviously, Anthropic's working closely with Amazon on the Trainium, but everyone wants more compute.
因此，我会关注的是，我们是否会听到Anthropic与AMD达成类似的消息。
So that's what I'd be watching is, will we hear something similar from Anthropic with AMD?
AMD确实表示，还会有更多战略合作即将公布，这就是我关注的重点。
AMD did say there's more strategic partnerships coming, so that's what I'd be listening for.
很好。
Great.
好了，Austin，感谢你前来做客。
Well, Austin, I want to thank you for coming on.
这位是Creative Strategies的高级分析师Austin Lyons，正在TI TV为您带来报道。
That is Austin Lyons, a senior analyst at Creative Strategies here on TI TV.
SaaSpocalypse仍在继续，本周多家大型企业软件公司的股价持续下跌。
The SaaSpocalypse continues as shares of several big enterprise software companies continue to fall this week.
《今日信息》刊登了一篇专题文章，探讨了这些大型公司如何应对，并反驳人工智能对它们业务构成威胁的说法。
The Information Today published a feature story looking at the ways in which several of these big companies are responding and how they are countering the narrative that AI is a risk to their businesses.
我想邀请我们的企业软件记者凯文·麦克拉林，帮助我们理清这一切。
I want to bring on Kevin McLaughlin, our enterprise software reporter, to help us make sense of all of it.
凯文，欢迎再次做客节目。
Kevin, welcome back to the show.
很高兴你能来。
It's great to have you here.
谢谢，阿卡什。
Thanks, Akash.
你今天写了这篇报道，开头选了HubSpot，这家公司在你看来是众多正经历这场震荡的SaaS公司之一。
So you wrote this story today, and you started with HubSpot, which was a company that you kind of lasered in as one of the many SaaS companies that are going through this reckoning right now.
你为什么选择从HubSpot开始？这家公司目前发生了什么？
Why did you start with HubSpot and what's happening at that company?
是的。
Yeah.
HubSpot是一家相比Salesforce等公司获得关注较少的企业，尽管它在客户管理软件领域与Salesforce竞争。
So HubSpot is a company that doesn't get a ton of attention compared to others like Salesforce, even though it competes with Salesforce and customer management software.
基本上，几周前他们在财报电话会议上，首席执行官亚米尼·兰根被问到：当HubSpot客户希望将他们的HubSpot客户数据与第三方供应商的代理结合使用时，你们会怎么做？
Basically, what happened is on their earnings call a couple weeks ago, the CEO, Yamini Rangan, was asked, what are you going to do in situations where HubSpot customers want to use their HubSpot customer data in conjunction with agents from third party vendors?
她的回应非常有趣，因为这可能是我迄今为止见过的最直白的表态——他们打算追踪这些行为、将其变现，并且不会成为允许他人随意从HubSpot提取数据的免费通道。
And her response was very interesting because they are probably the most unvarnished comments that I've seen so far, basically saying that we're gonna track it, we're gonna monetize it, and we're not gonna be a free pipeline for people to take data out of HubSpot.
因此，当你审视HubSpot所处的处境，以及几乎所有其他SaaS公司的现状时，这一点非常能说明问题。
And so I think it's very telling when you look at the situation that HubSpot is in and also basically every other SaaS company.
我认为，它们正竭尽全力向股东证明，它们不会坐视不管、任由这种情况发生。
They're looking to do anything they can, I think, to show shareholders that they're not gonna sit back and just let this happen?
它们将采取措施，对过去未曾变现的资源进行商业化。
They're going to take steps to monetize things that perhaps they haven't monetized before.
所以，为了总结一下，听众朋友们，我们此前已广泛报道过企业数据战争，以及这些企业软件公司掌握着所有客户数据的事实，我们也曾探讨过这些代理带来的风险——因为现在它们可以获取这些数据并用于自己的产品。
So just to recap, listeners, I mean, we've reported extensively information, you've reported extensively on the corporate data wars and the idea that these enterprise software companies are sitting on data that all their customers have, and we've written about the risks that these agents pose, given that now they can sort of take the data and make use of it for their own products.
而HubSpot现在说的是：你们要是想这么做，那就尽管去做吧。
This is HubSpot saying, You can do it if you want.
我们不会筑起高墙，但会为此收费。
We're not going to put up a wall, but we're going to charge you for it.
我们并不确切知道，因为HubSpot没有就此事做出回应，也没有说明他们具体将如何操作。
Well, we don't exactly know because HubSpot didn't engage on the story and they wouldn't say anything about specifically how they're going to do this.
因此，关于这一点有无数个问题。
So there are a million questions around this.
目前情况非常模糊。
It's very murky at this point.
但某种程度上，很明显，过去那些免费获取数据、客户数据的途径——HubSpot以在这方面非常慷慨而闻名——将发生变化。
But at some level, it's clear that, yes, things the the access to data, customer data that used to be free, HubSpot is well known for being very generous in that regard.
这种状况将会改变，但我们不知道会如何改变。
That's going to change, and we don't know how it's going to change.
但毫无疑问，对于希望访问HubSpot客户数据的第三方来说，情况将大不相同。
But certainly things are going to look different for third parties that want to access HubSpot customer data.
投资者对管理层这种明确表态有何反应？
How did investors react to type that type of clarity from management?
有趣的是，当CEO在回答问题时，HubSpot的股价在盘后交易中开始上涨。
The interesting thing was as, the CEO was speaking, as she was answering the question, HubSpot shares started to tick up in after hours trading.
我们现在不知道这是否正是原因，但值得注意的是，也许投资者希望看到她对此事采取强硬立场，做一些事情来表明HubSpot不会坐以待毙，任由SaaS行业的危机席卷而来。
Now we don't know if that's specifically why, but it was kind of interesting to see, perhaps investors wanted to see her, you know, take a fighting stance on this situation and and do something that, again, shows them that HubSpot's not going to just lay down and let the SaaS apocalypse wash over it.
这就是投资者的反应。
Now, that's how investors were reacting.
你有和HubSpot的客户交谈过吗？他们对这一切有什么感受？
Did you speak with any customers of HubSpot and how they might feel about all this?
我们联系了他们的一位合作伙伴，我们在报道中引用了他的话，他基本上表示，这种情况不会顺利，因为通常来说，涉及客户数据的保护主义举措都不会得到好评。
Well, we spoke with one of their partners who we quoted in the story, and he was basically saying that this isn't going to go well because generally when it comes to customer data, protectionist types of moves don't get a good response.
我们去年曾报道过，Salesforce的Slack部门更改了API条款，禁止第三方大规模导出和存储Slack中的数据。
We wrote about last year when Salesforce's Slack unit changed their API terms to prevent third parties from mass exporting and storing data from Slack.
是的。
Yeah.
目前没有迹象表明这损害了Slack的业务，但公众的反应确实不太好。
There's no indication that it hurts Slack's business in any way, but certainly the the public reaction wasn't great.
因此，当我们将这一立场扩展到整个企业软件行业时，HubSpot 已经采取了这样的态度。
So HubSpot has taken this stance on it as we broaden this out to the rest of the enterprise software sector.
它是唯一一家发表如此大胆言论的公司吗？
Is it the only company that has made comments that were as bold as this?
还是说，我知道现在企业软件高管们频繁上播客，尤其是在 SaaS 末日之际。
Or are there I know CEOs go on podcasts galore these days, enterprise software executives, especially amidst the SaaS pocalypse.
每个人都急于传达自己的观点。
Everyone is eager to get their message out.
我的意思是，他们传达的信息是否类似？他们整体的反应如何？
I mean, are they delivering similar messages, or how are they all reacting?
据我所知，还没有人说过这么直接的话。
No one has said anything this direct, to my knowledge anyway.
也许 somewhere 有某期播客我没看到或没听过，但我认为其他人可能会效仿。
Maybe there's a podcast somewhere that I haven't seen or heard, but I do think that it might be something that others follow.
我认为，所有所谓的‘记录系统’供应商——也就是企业存储客户信息、人力资源信息或财务数据的数据库——可能是最早对智能代理和第三方访问做出反应的一批公司。
I think all of the so called system of record vendors, which are basically databases where companies store customer information or human resources information or financial data, these are probably the the first ones to react to agents and third party access and things like that.
因此，看到其他类似公司的首席执行官说，‘嘿，我们也这么干’，我一点也不惊讶。
So I wouldn't be surprised at all to see, CEOs of other similar companies saying, you know what?
我们会这么做，因为他们可能有他们的理由。
We're gonna do this too because they might have an argument.
有几个理由。
There's a couple of arguments.
一个是安全性。
One is security.
你不能随意允许第三方访问你的数据。
You can't just allow third parties to access your data willy nilly.
这是一个长期存在的论点。
That's a perennial argument.
他们可能会在竞争中利用这个论点。
It might be convenient for them to use that argument in a competitive way.
但与此同时，这些供应商自己掌握着客户数据，保护数据是他们的责任。
But at the same time, the vendors themselves, they are in charge of the customer data, and it is their responsibility to protect the data.
所以你可以说，如果出了问题，责任在他们。
So you could make the case that if something goes wrong, it's on them.
当大家都在努力弄清关于智能代理的安全问题时，也许他们最好的做法是暂时暂停一下。
And while everyone's trying to figure out the security around agents, maybe it is in their best interest to just hit pause for a little while on this.
很好。
Great.
好了，凯文，感谢你前来做客。
Well, Kevin, I want to thank you for coming on.
这位是凯文·麦克劳林，我们《The Information》的企业软件记者。
That is Kevin McLaughlin, our enterprise software reporter, here at The Information.
Notion 正在迅速扩展其人工智能战略，今天正式推出自定义智能代理，让用户能够在平台上自动化工作，并从邮件、日历应用以及其他程序如 Figma 和 Slack 中提取数据。
Notion is quickly expanding its AI strategy, and today is launching custom agents to let people automate work within its platform and also pull data from mail and calendar apps and other programs like Figma and Slack as well.
他们还推出了新的定价策略，与客户实际使用智能代理的量挂钩。
They're also launching a new pricing strategy to go along with these agents that depend on how much customers actually use them.
我们邀请 Notion 的人工智能负责人莎拉·萨克斯，来为我们详细解读这些新工具。
Want to bring on Sarah Sachs, AI lead at Notion, to help us break down the new tools.
萨拉，欢迎来到节目。
Sarah, welcome to the show.
很高兴你来参加。
It's great to have you here.
早上好。
Good morning.
谢谢你们邀请我。
Thank you for having me.
告诉我们Notion今天发布了什么。
Tell us what Notion is launching today.
Notion一直是一个AI驱动的工作空间，我们将文档和项目整合到一个协作环境中，但我们也越来越成为一个人与AI代理可以协作的记录系统。
Notion's always been that AI powered workspace where we bring docs, projects into a collaborative space, but we're also really increasingly becoming that system of record where agents and humans can collaborate.
今天，我们推出了自定义代理，我认为这是实现企业级高效协作的重要一步，你可以将最重复的流程转化为受管控的AI同事。
Today we're launching custom agents, which I think are probably the largest step in the direction of doing that well at enterprise scale, where you can turn your best repeatable workflows into governed AI teammates.
它们可以监控触发条件，你可以设置它们按特定周期运行，并从你工作场所中所有连接的系统中提取数据，这样即使你离线，工作也能持续进行。
They can monitor for triggers, you can set them up on a certain cadence, and they can pull from all of their different connected aspects of your workplace, so work keeps moving even when you're offline.
我们今年春天在Notion制造大会上进行了预览，之后有许多Alpha用户取得了显著成果，经过大量调试后，我们非常兴奋地今天正式发布。
We previewed it at Make with Notion in the spring, and we've had a lot of Alpha users with a lot of success, and we're really excited after all that tinkering to have it GA today.
给我举个例子吧。
So give me an example here.
自定义代理，能举个它能完成的典型任务吗？
Custom agent, what is a sample task that this custom agent could do?
用户该如何构建这样一个代理？
How does one go about building it?
是的。
Yeah.
我们设计的自定义代理特别适合三类主要工作场景。
So we really built our custom agents to work particularly well for three main categories of work.
第一类是问答。
The first is Q and A.
比如Ramp公司就使用自定义代理实现了产品问答功能。
So Ramp, for instance, has a product Q and A feature of a custom agent.
因此，公司内部任何员工都可以向这个产品问答系统提问。
So internally, any employee can ask that product Oracle.
仅仅几周时间，它就已经回答了大约4000个问题，对吧？
Only in a couple of weeks, it's already answered about 4,000 questions, right?
如果你想想，每个人回答一个问题平均需要三十分钟，去处理、理解并回复，那么Ramp仅在一周内就节省了大约2000个人工小时。
So if you think about thirty minutes a question of someone being able to route that, understand, answer that, that's about 2,000 human hours just in one week that Ramp has saved.
我们第二个例子，作为工程经理我特别喜欢的是路由和触发机制。
The second example that we have, which I love as an engineering manager, is routing and triggers.
尤其是在工程或产品团队，或者任何公司里，有人可能会在Slack、邮件或其他工作渠道中提出问题。
Particularly in engineering product teams, or really any company, someone might ask a question in Slack, over email, wherever you're doing your work.
但工程经理或产品经理往往需要把问题转发到正确的Slack频道，去听、去跟进成千上万个Slack频道，才能真正搞清楚情况。
But an engineering manager or product manager, lots of people have to forward it to the right Slack channel, listen to it, follow 10,000 Slack channels to actually understand.
如今，通过自定义代理，你的代理可以被配置为监听这些Slack频道，理解问题内容，将其路由到正确的频道，创建任务，可能还回复相关的信息，并在Slack频道关闭时自动关闭任务。
Today with custom agents, your custom agent can be configured to actually listen into those Slack channels, understand what the question is, route it to the right Slack channel, create a task, maybe reply with the relevant information, close the task when the Slack channel is closed.
这种集成程度是Notion以前从未有过的吗？
Is this a level of integration that Notion has not had before?
我的意思是，它能够从 Slack 获取数据，并在 Slack 中创建任务、实现这些重复性操作，以前能做到吗？
I mean, idea that it can take data from Slack and then also create tasks in Slack, create some of these repeatable motions, was that possible before?
任务是在 Notion 中创建的。
So the tasks are created in Notion.
Notion 仍然是企业工作所依赖的核心记录系统。
Notion remains that system of record of where you're doing the enterprise work.
我们一直都能搜索 Slack 中的内容。
We've always been able to search over Slack.
我们的连接搜索和问答功能是 2023 年和 2024 年最重要的发布之一，这正是我们的切入点。
Our connected search and that Q and A functionality was one of our largest launches in 2024, 2023 that really was our entry point.
这一点独特之处在于，我们正在脱离聊天模式，对吧？
What's unique about this is the idea that we're moving away from the chat, right?
我们不再局限于在 Notion 内部的聊天中进行交互，而是让它成为一个可以在任何地方被触发的同事。
We're moving away from it being in a chat inside Notion, but really having it act as a coworker that can be triggered everywhere.
而许多开发工作都集中在企业级规模下将各种系统整合在一起的基础设施上。
And that's where a lot of this development has been the infrastructure of plugging things together at enterprise scale.
所以它是可管理的。
So it's governable.
它具有版本控制。
It has version control.
它具有审计日志。
It has audit logs.
在存在多个连接的情况下，如何尊重权限是很多工作所关注的，所有这些权限是如何被尊重的？
A lot of the work around respecting permissions if you have multiple connections, how are all of those permissions respected?
在过去十年中，Notion 在这方面积累了丰富的经验，这使我们在代理式工作流领域占据了有利位置。
That's the enterprise work that Notion has built an expertise on over the past decade that's really made us well positioned for agentic workflow.
所以我们
So we
我们刚刚和我们的企业软件记者讨论了企业数据之争，我们谈到了一些系统记录公司明确或隐含地表示：嘿，你可以访问我们的数据，但我们会向你收费。
just had a conversation with our enterprise software reporter about the corporate data wars, and we were talking about the tolls that some of these system of record companies have said either explicitly or in some ways implicitly, the idea that, Hey, you can access our data, but we're going to charge you for it.
这正是某些企业软件公司维持其业务的方式。
And that's sort of the way some of these enterprise software companies are going to be able to sustain their businesses.
但在这所有一切中，客户究竟想要什么，这是一个很大的问题，因为最终，许多客户都认为这些数据是属于他们的。
And there's a big question around what customers will want in all this, because ultimately a lot of these customers do believe it's their data.
所以，莎拉，我的问题是：在Notion与Slack以及其他所有这些程序和公司进行的所有集成中，是否遇到过任何公司设置壁垒，或者合作障碍日益增加的情况？
But my question for you, Sarah, is in all of these integrations that you have at Notion, with Slack, with all of these other programs and companies, has there been any challenges there with companies putting up walls at all, or maybe any increasing hurdles between cooperation at all?
我认为，在智能代理工作时代，每家公司都在摸索自己的定价模式。
I think every company is figuring out their pricing model in the age of agentic work.
Notion非常幸运，因为我们的许多核心企业用户，其大部分工作都是在Notion内部完成的。
Notion is really lucky because a lot of our core enterprise users, a majority of their work is happening inside Notion.
因此，我们拥有这种原生的第一方优势。
And so we have that native first party advantage.
我们已经看到，Notion并不一定试图取代你其他工作场所，比如Slack。
And we already see that Notion is not necessarily trying to replace some of your other areas where you work, for instance, Slack.
然而，我们是代理们进行协作的系统性记录中心。
However, we are that system of record where the agents are collaborating.
而这一点，是许多其他信息存储平台所不具备的。
And that isn't something that a lot of these information stores are holding.
它们可能是为个人工作提供记录的系统，但还没有人成功推出过适用于代理协作的企业级记录系统。
They might be a system of record for the individuals working, but no one has really been able to launch an enterprise scale system of record for agent collaboration.
一个很好的例子是离线模式。
A great example is offline mode.
Notion 开发了离线模式。
So Notion built offline mode.
我们在今年夏天推出了它。
We launched it this summer.
离线模式的妙处在于，当两个人在飞机上工作并上传文档时，你实际上能够解决这些冲突。
And the beauty of offline mode is if two people are working on an airplane and you upload a document, you're able to actually resolve those conflicts.
那
What
你
do
当两个代理同时进行并行工作时，你该怎么办？
you do when two agents are trying to do parallel work?
你如何解决这种冲突？
How do you resolve that conflict?
你如何查看和检查谁做了什么？
How do you look and inspect who did what?
这些代理不一定要基于Notion构建。
And those don't have to be Notion powered agents.
如果Slack创建了一个代理，或者有人在使用第一方前沿模型自己的代理，它们仍然需要一个地方来与企业工作协作。
If Slack created an agent, if someone's using a first party frontier model's own agent, they all still need a place to collaborate with their enterprise work.
这正是我们瞄准的市场，而今天这个市场还不存在。
And that's the market we're going after that doesn't really exist today.
我们看到了采用率的急剧增长，因为企业最终都渴望推出AI产品，但无法安全地大规模实现。
And we've seen just a hockey stick of adoption because ultimately enterprises are thirsty to launch AI products, but they aren't able to do it safely at scale.
无论你的信息存储在哪里，这都是一个不存在的市场。
That's a market that doesn't exist regardless of where your information lives.
这是一个记录代理协作情况的记录系统。
It's a system of record for where agents are collaborating.
即使在今天，Notion 上大约
Even today on Notion, about
现在，我真正想探讨的是你们正在推进的按使用量计费模式。
Now, 10 what What I actually wanted to get into was the usage based pricing that you guys are diving into here.
这是 Notion 首次采用按使用量计费吗？
Is this the first time that Notion has done usage based pricing?
是的。
Yeah.
目前，只有自定义代理会采用按使用量计费。
So only the custom agents today will have usage based pricing.
我们为商业版和企业版提供了基于席位的定价模式，该模式将保持不变。
We have a seat based model for business and enterprise plans, which will remain.
但随着代理完成更多工作，其价值真正与使用量成正比，而非席位数量。
But as agents do more work, the value really scales with usage, not with seats.
因此，我们非常希望确保我们的定价模型与实际价值相匹配。
And so we really want to make sure that our pricing model is correlated with the value.
所以我们希望根据成果来定价。
So we want to price on outcomes.
这是我们首次成功实现这一点。
And this is the first time we've been able to successfully do it.
话虽如此，这对我们的用户来说是一个新的调整。
That being said, it's the new adjustment for our users.
用户今天就可以使用自定义代理，并查看他们的用量统计，但我们实际上要到五月才开始对自定义代理的使用收费，因为我们正在与市场上的用户共同学习，了解他们希望拥有多少可见性和控制权——毕竟，我们不希望像你在推特上看到的那些代理用量失控的极端案例发生。
Users will have access to custom agents today and they'll be able to see their metered usage, but we're not actually charging on that custom agent usage till May because we're learning too with our users in the market on what amount of inspectability and control they want, because we don't want, as you see over Twitter, of these crazy examples of runaway agent usage.
那正是我想问你的，你们怎么知道代理在后台会工作多少呢？
What we Well, that's what I was going to ask you about is, how do you know how much the agent is going to be working in the background?
你们输入了支付信息。
You put your payment information in.
说‘我买这么多积分’是一回事，但最终，如果我的积分被快速消耗，我怎么知道原因呢？
It's one thing to say, Hey, I'm buying this many credits, but at the end of the day too, if my credits are getting drained, then how do I know that
有资金吗？
there's capital?
这正是Notion如此占据优势的原因，因为我们理解企业管理员希望做什么，以及那些只是希望在Mac Mini上使用时不耗尽额度、信用卡不被拒的普通用户之间的冲突。
So that's exactly why Notion is so well positioned because we understand these conflicts of what an enterprise administrator wants to do versus someone on their Mac Mini just hoping that they don't run out, that their credit card doesn't get declined.
对吧？
Right?
因此，第一点，我们多年来一直在为客户提供核心代理功能，并试图理解这会对我们的定价、利润率和收入产生怎样的影响。
And so number one, we've been building core agent for our customers and trying to understand what that pricing implication is and how it hits our margins and revenues for years.
所以我们对用户希望的可观察性水平——无论是模型选择、触发频率等——已经有了直观的把握。
So we kind of have an intuition on what level of inspectability, whether it be model selection, trigger rate, etcetera, that our users want.
我们拥有第一手的同理心，对吧？
We have first party empathy, right?
所以你们会给他们一个大致的预估，比如这个……你们是利用手头的数据来提供预估的，对吧？
So you give them kind of like an estimate saying this So is what this you use the data that you have on hand to give them an estimate.
而这里的理念是，因为你们拥有这些数据，所以你们的预估会比其他公司更准确。
Then the idea here is that because you have that data, the estimate will be more accurate than maybe other companies.
对。
Right.
然后我们还围绕着提醒你来构建产品，比如：‘你已经触发了这个功能超过500次。’
And then we also build product around telling you like, Hey, you've triggered this over 500 times.
你有没有考虑过使用一个更便宜的模型？
Do you want to think about maybe a cheaper model?
你有没有考虑过调整这个触发条件？
Do you want to think about changing what that trigger is?
最终，我们的目标是确保为客户提供真正的价值。
Ultimately, we're here to make sure that we're delivering value to our customers.
按使用量计费的模式设计得非常积极，因为无论怎样，只要客户获得了价值，我们的利润率就有保障。
The way usage based pricing is set up is it's a really positive incentive because no matter what, we're making solid margins on our customers having value.
我们并不是想靠用户为没用的功能付费来赚钱。
We're not trying to make money on people paying for something that they don't use.
因此，这是一种绝佳的价值对齐，我认为大多数公司都在努力寻找这种平衡。
And so it's a great value alignment that I think most companies are trying to figure out.
我们正在与客户一起不断迭代。
And we're iterating with our customers together.
我们之所以把截止日期定在五月，还有另一个原因，那就是我们希望真正为客户提供价值，确保他们只为实际使用的功能和取得的成果付费。
That's another reason we have till May, is that we want to do right by our customers and make sure that they're paying for the use that they're getting and paying for outcomes.
当自定义代理充当第二位同事时，我们的互动将不再局限于聊天界面。
Custom agents, when they act as that second coworker, we're moving out of the chat.
即使你离线时，它也会继续运行。
It's moving even when you're offline.
用户需要理解他们能从这些功能中获得什么价值，以及愿意为此支付多少费用。
Then users need to understand what value they describe to that and what they'd be willing to pay.
好了，莎拉，非常感谢你来参加。
Well, Sarah, I want to thank you for coming on.
这位是莎拉·萨克森，Notion的AI负责人，正在TI TV上做客。
That is Sarah Saxon, AI Lead at Notion, here on TI TV.
玛丽。
Mary.
Anthropic公司正将其部分研究重点放在网络安全上，一份最新独家报告披露了该公司正在开展的50项研究项目，旨在研究失控的智能代理。
Anthropic is focusing some of its research efforts on cybersecurity, and a new exclusive report from the information revealed 50 of the company's research projects that try to study rogue agents.
我想邀请报道这则故事的记者罗基特·德鲁，帮我们深入解析他的报道。
I want to bring on Rocket Drew, the reporter behind that story, help us break down his reporting.
罗基特，欢迎再次来到节目。
Rocket, welcome back to the show.
很高兴你能来。
It's great to have you here.
嗨，阿卡什。
Hi, Akash.
谢谢。
Thanks.
能来这里真的很棒。
It's great to be here.
能跟我讲讲你发现的Anthropic正在进行的这50个研究项目吗？
Tell me about these 50 research projects that you found out about are happening at Anthropic.
是的，当然可以。
Yeah, absolutely.
Anthropic 已经运行这个奖学金项目将近一年了。
So Anthropic has been running fellowship program for going on a year at this point.
通过这个项目，他们邀请年轻的研究人员与公司内更资深的研究人员合作。
And through this program, they bring on younger researchers to collaborate with their more senior researchers at the company.
这些研究人员通常是大学生或博士生，对与 Anthropic 合作感兴趣。
These researchers tend to be say college age or PhD students who are interested in collaborating at Anthropic.
通过该项目，Anthropic 的导师——即资深研究人员——会提出一些研究课题，供这些年轻人参与。
And through the program, Anthropic's mentors, their senior researchers, propose research projects that these people can work on.
他们提出了 49 个课题。
So they proposed 49.
其中大概只有一半最终被推进了，因为这取决于当前参与项目的这批人具体兴趣所在。
Probably of those, about only half ended up getting pursued because of the interests of the people in the particular cohort that they have running right now.
但也许最好的方式是向你介绍这些课题所归属的六大类别，你觉得如何？
But maybe the best way to give you a window into these projects is to tell you about sort of the six categories that they fall into, if that sounds good.
当然
Of
首先是安全，就像你提到的。
the So the first is security, like you mentioned.
确保在网页上工作的代理不会暴露于黑客攻击或被诱导泄露用户个人信息的提示。
Ensuring that, say, agents that are working on the web don't get exposed to a hack or a prompt that convinces them to expose a user's personal information.
第二个类别是AI控制，这是一类研究，旨在解决当我们不确定AI模型的目标是否与我们一致时，如何从AI模型中获得有用的工作成果。
A second category is AI control, which is a category of research that refers to figuring out how we can get useful work out of an AI model when we're not sure that the AI model has the same goals as us.
例如，你可能希望使用一个较弱但更可信的模型来监督其尝试采取的行动，并标记任何可疑的行为。
So for example, you might want to use a weaker but more trusted model to supervise the actions that it tries to take and flag anything that looks suspicious.
还有一类项目称为可扩展监督，这是一个研究议程，同样侧重于使用较弱的AI模型来监督和训练更强的AI模型。
There's another bucket of projects called scalable oversight, which is a research agenda that focuses similarly on using weaker AI models to supervise and train stronger AI models.
还有一个类别叫做模型内部机制，用更专业的术语来说，就是机制可解释性。
There's a category called Model Internals, which is in the more jargon we would call that mechanistic interpretability.
这是Anthropic研究的核心内容。
It's a real mainstay of anthropics research.
它涉及理解模型内部的运行机制。
It involves understanding what's going on under the hood.
在模型内部，是什么导致了我们所看到的输出。
Inside the models, what's responsible for the outputs that we see.
因为这些模型就像一个黑箱，我们通常这么描述它。
Because these models sort of a black box, is how we would talk about it.
最后两个类别是：一个是模型生物体，这是Anthropic的术语，指我们使用现有模型进行类似在小鼠身上做科学实验的操作，希望结果能推广到未来的人类模型。
And then the last two are: one is model organisms, which is Anthropic's term for when we use existing models as sort of, you know, doing a science experiment on a mouse and hoping that the results generalize to a human someday.
当我们检查当今的模型，寻找未来更强大模型可能出现的风险和情景迹象时，就是这种情况。
It's when we inspect modern day models for signs of the kind of risks and scenarios that could arise in future, more powerful models.
因此，Anthropic的许多项目都专注于这一点。
So a number of anthropics projects are focused on that.
最后一个类别是专注于理解中国模型、评估它们、弄清它们的能力，同时提升Anthropic自身托管和运行这些模型的能力。
And then the last category is projects that focus on understanding Chinese models, evaluating them, figuring out what their capabilities are, also improving Anthropic's ability to host and run those models itself.
这就是分类情况。
That's the breakdown
好了。
there.
这非常有趣，因为它让我们得以一窥Anthropic如何优先安排其研究工作，特别是与其AI的风险与安全功能相关的内容。
So this is kind of fascinating because this is a window into basically how Anthropic is prioritizing its research efforts, specifically to do with sort of the risk and safety functions of its AI.
我想问问你关于你写过的那些失控代理，因为我觉得你和我之前讨论过OpenClaw，以及让代理自由运行、可能赋予它访问你整个电脑乃至所有账户的权限，并说‘随你怎么做’所带来的安全风险。
I want to ask you about these rogue agents that you wrote about, because I think you and I have talked about OpenClaw and some of the security risks that come with letting an agent loose and potentially giving it access to not just your whole computer, but your accounts too, and saying, Do what you want.
你的报道很有意思，因为它似乎暗示Anthropic正在正视这种风险，并试图找出如何防范它。
Your reporting was interesting because it actually seemed to suggest that Anthropic is leaning into that risk and trying to figure out how to protect against it.
他们是怎么做到的呢？
How is it doing that?
是的，你说得对。
Yeah, that's exactly right.
我认为这确实是他们非常关注的话题。
I think it's a topic that's certainly on their minds.
我不知道你有没有看到昨天的事，Meta的研究员Summer Yu说，她的OpenClaw代理竟然开始删除她所有的邮件，尽管她并不希望如此。
I don't know if you saw yesterday, Summer Yu, who's a researcher at Meta, said that her OpenClaw agent went and started deleting all of her emails even though she didn't want it to.
这是一个典型的例子，说明代理明显在做与她意愿相悖的事，但这次甚至都不是被黑客入侵造成的。
And that was an example where the agent was clearly doing something that was not aligned with her desires, but it wasn't even hacked in that case.
只是它没有正确地遵循指令。
It was just that it wasn't following instructions properly.
所以这里有很多需要担心的情况。
So there's a lot of scenarios to be worried about here.
有些情况是代理误解了它应该做什么，无意中犯了错误。
Some of them are just the agent misunderstanding what it's supposed to do and sort of making mistakes by accident.
有些情况则是它遵循了指令，但这些指令本身是恶意的。
Some of them involve it following instructions, but those instructions being malicious.
有人试图让他们的代理去做它不该做的事，比如发动网络攻击。
Someone tries to get their agent to do something it's not supposed to, say, to conduct a cyber attack.
还有一种情况涉及黑客攻击，比如有人在网上诱导你，试图窃取你的信息或加密货币
And then there's this sort of other genre that involves hacking, like someone being prompt out on the web to try to steal your information, steal your crypto
钱包。
wallets.
那么，Anthropic 是如何防范或试图防范这些问题的呢？
So how is Anthropic protecting against that or trying to protect against it then?
是的
Yeah.
我认为有多种应对方式。
I think there's a range of approaches.
其中一些涉及研究当前模型的能力。
Some of it involves studying what models are capable of right now.
例如，我们觉得Claude可能擅长网络攻击，但还没有人真正努力让Claude参加网络安全挑战赛。
For example, we have a sense that Claude could be good at cyber attacks, but no one has tried that hard to get Claude to compete in cybersecurity challenges.
人们已经做过一些尝试，但它还有很长的路要走。
People have tinkered with it, but it has a ways to go.
因此，研究的一个方向就是弄清楚Claude到底能做什么？
So one thrust of the research is just figuring out what is Claude even capable of?
这些模型到底能做什么？
What are these models capable of?
它们可能有多危险？
How dangerous could they be?
展开剩余字幕（还有 266 条）
然后还要创建基准或标准化测试，以更严谨的方式衡量这些能力，从而能够察觉到这些能力的提升。
And then also creating benchmarks or standardized tests to measure that in a more rigorous way so we can notice when there's an improvement in those capabilities.
当然，人们也对降低这些风险感兴趣，并探索如何训练模型表现得更好。
And then of course there's also interest in mitigating those risks and figuring out how we can train the models to behave better.
例如，Anthropic的研究人员为他们的研究员提出的一个项目是：当野外出现某种网络安全事件或黑客攻击时，我们能否将其捕获并自动复现，从而训练Claude在未来不再重蹈覆辙？
So for example, one of the projects that anthropics researchers proposed for their fellows is when a sort of cybersecurity incident or one of these hacking incidents appears in the wild, can we take that and can we automatically reproduce it in a way that we could train Claude to not fall for the same trap again in the future?
目前这个过程还相当手动。
Right now that's a pretty manual process.
如果野外发现此类事件，比如某个智能体在与银行网站交互时被攻破，一名Anthropic员工必须坐下来重建该银行网站并复现同样的攻击，确保其有效。
If one of these things gets spotted in the wild, say, an agent gets hacked when it's interacting with a banking website, an Anthropic employee has to sit down and recreate the banking website with the same attack and ensure that it works.
这有点繁琐。
It's a little tedious.
这个项目提出：如果我们能自动完成这一切，会怎么样？
The project proposes, Well, what if we could just do that automatically?
你之前和Anthropic的一位安全负责人进行过对话。
Now, you had a conversation with one of Anthropic's safety leaders.
你去参加那次面试时带着哪些问题？最终从那次对话中学到了什么？
What were the questions that you went into that interview with, and what ultimately did you learn from that conversation?
是的，我一直在想，这些研究员所从事的研究有多重要？
Yeah, I think one question on the back of my mind is how significant is the research that these fellows are pursuing?
我的意思是，我们谈论的是一些经验不多的研究人员。
Mean, we're talking about researchers that sometimes don't have a lot of experience.
他们可能只工作四到六个月。
They're maybe only working for four to six months.
Anthropic 很可能为他们挑选了一些在短时间内能完成的项目，也许他们对模型的访问权限有限，计算资源也不多。
Anthropic is probably picking projects for them that they can accomplish in that short of a time, maybe without a lot of access to their models or without a ton of compute.
所以我在想，这个项目产出的研究成果到底能有多重要？
So I was wondering, you know, how significant can the research really be that's come out of the program?
我了解到，在过去几个月里，他们一个核心安全团队发布的研究成果中，研究员们贡献了超过一半。
And I was informed that of the research that one of their key safety teams has put out over the past several months, the fellows account for more than half of that output.
因此，我惊讶地发现，这个实习项目正成为 Anthropic 工作的重要组成部分，并极大提升了他们在安全与防护研究方面的成果。
So I was surprised that the fellowship program is becoming a significant portion of Anthropics' work and a big uplift to this kind of safety and security research that they're working on.
这个项目已经产生了一些备受瞩目的研究成果。
And there have been a couple pretty high profile examples of research coming out of the program.
既然我们已经了解了导师们为下一届研究员提出的计划，我认为我们有一些值得关注的事项。
Now that we have a sense of what the mentors are proposing for the next batch of fellows, I think we have some things to keep an eye out for.
很好。
Great.
好了，Rocket，感谢你来做客。
Well, Rocket, I want to thank you for coming on.
这位是Rocket True，我们《The Information》的AI与机器人记者。
That is Rocket True, our AI and robotics reporter here at The Information.
谢谢，Akash。
Thanks, Akash.
政府监管已成为全球许多科技公司增长策略中的核心挑战。
Government regulations have become a central challenge to many tech companies' growth playbooks around the world.
十年前，参与这一讨论的主要是Uber和Airbnb这样的初创公司。
A decade ago, the biggest companies involved in that conversation were startups like Uber and Airbnb.
这些公司仍然积极参与这一议题，但自那时以来，这一领域已经大大扩展，涵盖了加密货币、人工智能、预测市场、数据中心等更多内容。
Those companies remain actively involved in that story, but the arena has now very much expanded since then to include crypto, AI, prediction markets, data centers, and more.
我想邀请一位硅谷最知名的初创企业监管事务顾问。
I want to bring on one of Silicon Valley's best known advisors on startup regulatory affairs.
布拉德利·塔斯克是塔斯克风险投资公司的创始人兼首席执行官。
Bradley Tusk is the founder and CEO of Tusk Ventures.
布拉德利，欢迎来到TI TV。
Bradley, welcome to TI TV.
很高兴你能来。
It's great to have you here.
嘿，谢谢邀请我。
Hey, thanks for having me.
你一周前在《纽约每日新闻》上发表了一篇关于数据中心的文章，你的观点让我觉得很有意思，因为我觉得我们节目上请过的许多风险投资人都非常支持建设人工智能及其所需的所有基础设施。
So you wrote a piece about a week ago in the New York Daily News about data centers, and Your view was kind of interesting to me because, look, I think a lot of venture capitalists that we had on the show, they're very much for building out AI and everything that it will take to build the capacity.
但你对这个问题的看法有些微妙。
You had bit of a nuanced view on this.
你当时想传达的信息是什么？
What was your message that you were trying to send?
是的。
Yeah.
当我们考虑数据数据中心，或者更广泛的科技领域时，我们不能只考虑对我们自己最有利的东西。
That when we think about whether it's data centers or really tech in general, we can't just think about what's best for us.
如果我们开发的产品完全不受监管，不涉及政府或政治，那当然可以放手去做。
If we are developing a product that has no regulatory oversight at all, doesn't involve government or politics, sure, knock yourself out.
但如果你所做的任何事情本来就受到政府监管，那你为什么不考虑一下这个问题的政治影响呢？
But if you are doing anything that is regulated by government anyway, for you to not think about what are the politics of this issue?
这会对消费者产生什么影响？
How will this impact consumers?
这会对选民产生什么影响？
How will it impact voters?
我该如何确保我想要的东西既能被允许，又不会被禁止？
How do I ensure that what I want will be both allowed and not banned?
你必须把这些因素考虑进去。
You have to take these things into account.
当超大规模云服务商决定投入数万亿美元建设数据中心时，他们似乎完全没有意识到，这将给使用各种电网的消费者带来大幅上涨的电费。
And somehow when the hyperscalers decided to spend trillions of dollars building data centers, the notion that they're going to impose massively higher electricity costs on consumers who were, you know, in whatever grid they're using didn't seem to occur to them.
当人们认为‘我们在拯救世界，我们在开发新技术，所有人都会喜欢我们，所以我们不必考虑自己工作的负面外部性’时，这种想法极其天真且适得其反。
And I think it is incredibly naive and counterproductive when you think, Oh, we're saving the world, we're developing new technology, everyone's just going to love us, and we don't have to sort of think about the negative externalities of our work.
而事实正是如此。
And that's what's been the case.
但与此同时，你看到在所有设有数据中心的地方，消费者都面临着更高的电费账单。
But at the same time, you're seeing consumers facing much higher electricity bills anywhere where there are data centers.
无论是蓝州还是红州，各州都在提出立法，表示如果新建数据中心会大幅增加他人的成本，就不会批准其许可和用地规划。
And states proposing legislation, both blue and red states, saying, We're not going to allow the permitting and zoning of new data centers if they're going to impose much higher costs on everyone else.
因此，对我来说，如果我们希望数据中心得以建设，希望人工智能得以发展，就必须解决这个问题。
And so to me, if we want data centers to happen, if we want AI to happen, we have to address this issue.
而这正是那篇文章的核心观点。
And that was the point of the piece.
那么我们该如何解决这个问题呢？
So how do we address the issue then?
我认为，没有哪个政客会为了让萨姆·阿尔特曼成为万亿富翁而牺牲自己的政治生涯。
I think that, you know, no politician is going to sacrifice their career so that Sam Altman can become a trillionaire.
没人关心萨姆·阿尔特曼、黄仁勋，或者这些公司里的任何人。
Nobody cares about Sam Altman or Jensen Huang or anyone at any of these companies.
他们只关心能否连任。
They care about getting reelected.
所以，如果你的数据中心会导致在你下次选举中真正投票的选民电费大幅上涨，那它就不会被允许。
And so if your data center is going to impose meaningfully higher electricity costs on the people who actually vote in your next election, then that's not going to be allowed.
我们需要想出一些办法，让芯片更高效、数据中心耗能更少、用水更少，这样它们就能在我们需要的任何地方存在、获批和建设，同时又不至于成为政治斗争的靶子——让两党都因为政治利益而反对它。
And we need to come up with ideas that allow for far more efficient chips and far less energy usage and far less water usage for data centers so that they can exist and be permitted and be built everywhere that we need them, but at the same time not become this political football that everybody on both sides sort of decides to oppose simply because it's not good for them politically.
说实话，自从我2011年开始为优步工作以来，这种情况一直存在，这往往是硅谷和科技界的一个盲点：完全忽视政府、政治和监管。
And so look, mean, this has been the case since I started working with Uber back in 2011, which is one of the real blind spots often that you see in Silicon Valley and you see in tech is a just ignorance of government and politics and regulation.
就像错误的工程设计会毁掉一家公司，错误的融资策略会毁掉一家公司，错误的市场策略会毁掉一家公司一样，如果你是一家处于高度监管行业的初创公司，却没能准备好应对迎面而来的政治问题，那你很可能无法生存下去。
And just like the wrong engineering can kill your company or the wrong fundraising strategy can kill your company or the wrong, go to market strategy can kill your company, If you are a startup in a highly regulated industry and you are not prepared to be able to deal with the political issues, you know, coming at you, there's a good chance you're not going to make it.
我认为，特别是在人工智能领域，人们根本没怎么认真思考过这个问题。
And I think that especially in the AI sector, there has just been not really any thought put into it.
你在数据中心上能看到这一点，比如现在已经有二十多个州准备禁止心理健康聊天机器人。
You see this with data centers, you see it, for example, there are two dozen states now that are poised to ban mental health chatbots.
这个行业的表现糟糕透顶，既没有好好解释他们做什么，也没有说明为什么对社会有价值，更没有说明如何在受监管的方式下保护儿童并解决立法者的核心关切。
And that industry has done a pitifully poor job in explaining what they do, why it can be valuable to society, and how they can do it in a regulated way that protects kids and takes care of the underlying concerns that legislators have.
但你必须认真思考这些问题。
But you have to think about this stuff.
所以你是一位风险投资人，而人工智能是你作为投资者不得不参与的领域。
So you are a venture capitalist and AI is sort of this unavoidable area that you have to partake in as an investor.
基于你对数据中心建设方式的这些观点，那你具体在投资什么呢？
What are you investing in then, given that you have these views about how the data center build outs should happen?
是的。
Yep.
更节能的计算方式，以及能够为数据中心提供现场电力、无需接入电网的能源设施。
Alternate forms of compute that are far more energy efficient and energy facilities that can provide on-site power to data centers that don't require being attached to the grid.
因此，这可能涉及那些采用不同设计的内存模块的公司，能够显著提升能源效率。
So that could be companies that have memory cylinders that are built differently, that can really increase energy efficiency.
还可能存在完全不同的芯片类型，它们更加节能。
There can be different types of chips entirely that are more energy efficient.
我们是一家名为生物计算公司的企业，利用大鼠大脑的神经元来驱动人工智能。
We're one company called, biological computing company that uses neurons from rat brains, to power AI.
还有一些公司正在部署微电网核能、氢燃料电池或可安装在本地的涡轮机。
There are companies that do microgrid nuclear or, hydrogen fuel cells or turbines that can be put on-site.
因此，我认为在降低人工智能数据中心能源需求方面，存在大量投资机会。
So I think that there's a lot of investment opportunities specifically in reducing the energy needs of data centers for AI.
这些对我来说都是不错的投资机会。
And those are good opportunities to me.
我想了解，在即将到来的中期选举前，你认为科技领域将面临哪些关键问题。
I want to understand what you think will be the defining issues in technology heading into the midterm elections.
你提到了数据中心，这将是一个重大议题。
You talked about data centers, that will be a big one.
我们还有预测市场，这目前也是另一个热点问题。
We also have prediction markets, which is sort of this other hot button issue right now.
我正在思考政客们在中期选举周期中可能采取的立场。
I'm trying to think about the stances that politicians might take heading into the midterm cycle.
你觉得这两个是主要问题吗？
Are those the two big issues you think of?
你认为什么才会真正定义这次选举？
What do you think Let would really define this
让我稍微重新表述一下你的问题，可以吗？在我看来，中期选举有一个关键的流行词。
me kind of redefine your question a little bit, if that's okay, which is to me, there's one key buzzword for the midterms.
请记住，当我们谈到中期选举时，我们想到的是国会，即众议院的全体议员和参议院的三分之一成员。
And keep in mind, the midterms, when we think about it, we think about Congress, which is every member of the House and a third of the Senate.
但今年还有多达36位州长面临选举或连任，以及大多数州议会。
But there's up to 36 governors that are for election or reelection this year, as well as, you know, most state legislatures.
因此，这实际上是全面的。
So it's really across the board.
可负担性是2026年的核心议题。
Affordability is the issue of 2026.
我们在纽约市已经看到了预兆，班达米意识到这一点，并将这一议题转化为胜选关键，成功当选下一任纽约市长。
We got a preview of that in New York City with Bandhami realizing that and writing that issue to to victory to become the next mayor of New York City.
这正是未来关注的焦点。
That's what it's going to be about.
例如，数据中心问题之所以重要，是因为如果它推高了普通选民的电费，这就是一个可负担性问题。
So for example, the data center issue matters because if it does increase electricity prices for your average voter, that's an affordability question.
因此，从宏观角度看，如果你是一家科技公司，所做的事被认定为增加了普通民众的开支，你就需要意识到这一点，准备好进行辩护，并思考如何调整策略，为政客提供他们需要的解决方案，以免被彻底打压。
And so from a big picture standpoint, if you are a tech company that is doing something that is deemed to be increasing costs for average people, you need to be aware of it and ready to be able to defend it and think about ways that you can change and pivot to give politicians the one that they need so they don't put you out of business entirely.
或者，相反地，你很可能是一家能够提升可负担性的初创公司。
Or on the flip side, you very well might be a startup that can increase affordability.
在这种情况下，你就有绝佳的机会宣称：‘我们正是解决方案的一部分。’
And in which case you've got a real opportunity to try to say, Hey, we're part of the solution here.
比如，我投资了一家名为OwnWell的公司，它帮助人们处理房产税申诉，我们正在多个州推动立法，让民众更容易接触到这一系统。
So like, for example, I'm an investor in a company called OwnWell that helps people with property tax appeals, and we're working on legislation in a bunch of different states so that people kind of access to that system.
你知道，这在政治上是个赢家，因为虽然地方政府当然不希望失去财产税收入，但人们确实希望减少在财产税上的支出。
You know, that is a political winner in the sense that, while obviously local governments don't want to lose property tax revenue, people do want to spend less money on their property taxes.
而在一个 affordability 是核心问题的年份里，这显然是一个可以取得成功的关键领域。
And in a year where affordability is the question, that's really an area for success.
所以，比如预测市场，我们在这方面投入了很多，我想我对这个话题有很多想法可以深入探讨，但我认为今年它不会成为选举议题，因为它对可负担性并没有实质影响。
And so, so prediction markets, for example, which we work on a lot, and, I think I have a lot of thoughts on that we can get into, but I don't think it's going to be an electoral issue this year, in that it doesn't really impact affordability one way or the other.
你对预测市场有什么看法？
What are your thoughts on prediction markets?
因为这在某种程度上是典型的联邦与州之间的权力博弈。
Because this is sort of a classic federal versus state battle in some ways.
但我知道，这么说可能过于简化了。
But For that's, I know, probably oversimplifying it.
我觉得其实并没有简化。
I don't think it is actually.
那你就详细说说其中的细微差别吧。
Well, unpack the nuances here.
你认为这件事会走向哪里？
Where do you think this go?
听我说，人们都在说，你之前也做过这个话题的播客，这件事最终可能会上诉到最高法院，并在那里做出裁决。
Look, people are saying, and you've done a podcast on this, it could end up in the Supreme Court, and there will be decisions there.
但我的意思是，有没有什么迹象能告诉我们这件事会往哪个方向发展？
But, I mean, are there signals that tell us which way this is going?
是的。
Yeah.
我的意思是，这是个很好的问题。
I mean, it's a great question.
为了给观众们提供一些背景信息，有一些游戏公司是受各州监管的。
So just to give the viewers a little bit of a level set here, there are gaming companies that are regulated by state.
也就是说，通常是像FanDuel和DraftKings这样的公司从事体育博彩，它们在各个州拥有经营许可证。
So that's typically the FanDuels, the DraftKings of the world that do sports betting, and they have licenses in individual states to conduct that.
然后它们要接受每个州的监管和税收体系。
And then they're subject to the regulatory and taxation system of each individual state.
然后还有预测市场公司，它们被视为商品，由商品期货交易委员会监管。
Then there are prediction market companies, that are seen as commodities and they're regulated by the CFTC.
所以这是联邦层面的监管。
So that's at the federal level.
因此它们有着不同的监管和税收结构。
So they've got a different regulatory and taxation structure.
比如Kelsey、Polymarket等公司就属于这一类。
So that's companies like Kelsey and Polymarket and others.
而各州认为，预测市场实际上应该属于它们的管辖范围，因为最高法院在2017年推翻了一项名为PASPA的国会法案。
And states are saying, you know, that prediction markets really should be under their jurisdiction because, the Supreme Court in 2000, I think it was 'seventeen, overturned a congressional law known as PASPA.
PASPA的内容是禁止各州允许体育博彩。
And what PASPA was, was a ban on states allowing sports betting.
最高法院表示，这本质上是各州的事务，各州有权决定如何处理。
And the Supreme Court said, no, this is really a state issue, and their right to decide what to do.
这不应该是国会的决定。
It's not up to Congress.
这打开了整个行业，导致体育博彩遍布美国。
And that opened up the entire industry and led to sports betting all over The US.
各州认为，根据PASPA的裁决，他们显然也应对预测市场拥有管辖权。
States are saying under the PASPA ruling, they should clearly have jurisdiction over prediction markets as well.
商品期货交易委员会（CFTC）和联邦政府表示：不，这些是商品。
The CFTC, the federal government is saying, no, these are commodities.
这些是历史上一直由CFTC监管的合约。
These are contracts that have been historically and traditionally regulated by the CFTC.
这属于我们的管辖范围。
It is our jurisdiction.
因此，你现在看到的是各州与联邦政府之间涉及不同利益方的法律诉讼。
So what you're seeing now is litigation between, states and the federal government with different types of parties involved.
有些裁决支持CFTC的监管权。
And some of the rulings have been on the side of CFTC regulation.
有些则支持州级监管。
Some have been on the side of state regulation.
通常，当全国范围内出现分歧时，最高法院就会介入这一议题。
Usually when there are divided opinions across the country, that's when the Supreme Court takes up an issue.
很有趣的是，看看他们将如何推进，因为一方面，这个法院总体上曾作出过PASPA裁决，而且它是一个更为保守的法院，通常更倾向于支持州权。
It'll be really interesting to see how they, proceed here because on one hand, this is the court that by and large did PASPA, and is a court that is more conservative, which tends to be more favored towards states' rights.
另一方面，联邦层面已经存在一个清晰且有效的监管框架和体系。
On the other hand, there is a clear regulatory framework and system that is working at the federal level.
因此，实际上可能根本没有理由去改变现状。
So there really may not be any reason to change that.
所以，你知道，我觉得这有点像五五开的情况，意味着目前，大家只是继续维持现状，因为我们谁也无法预知最高法院何时会介入，或者他们会如何裁决。
And so, you know, I think it's kind of one of those fiftyfifty things, which means that for the moment, you know, everyone is just proceeding with the status quo because none of us have the ability to know when the Supreme Court's gonna take up an issue or how they're going to decide it.
但我要这么说。
But I will say this.
如果法院裁定支持商品期货交易委员会，我认为这至少有50%的可能性，那么从事体育博彩的公司将会转变为，我认为，预测市场公司，因为它们会说，我们宁愿只有一个监管机构，以及一个更优惠的税收体系，而不是面对42个不同的州或其他任何情况。
If the court rules in favor of the CFTC, which I think there's at least a 50% chance of that, then the companies that are doing sports betting are going to become, I would argue, prediction market companies instead, because they're going to say we would far rather have one regulator and a far more preferential tax system than, you know, 42 different states or whatever it might be.
这将给各州政府造成巨大的收入缺口。
That's going to create a giant hole in revenue for state governments.
他们必须填补这个缺口。
They're going to have to fill that hole.
所以，如果你展望未来，iGaming（即在线赌场游戏、扑克、二十一点、骰子等）我认为将是下一波浪潮。
And so I think that if you're looking towards the future, iGaming, which is just online casino gaming, poker, blackjack craps, whatever you want, I think will be the next wave.
因此，如果你在寻找可能提供有趣iGaming形式的初创公司——这些形式能吸引客户和消费者，那可能就是下一个机会所在。
So if you're looking at who are the startups that might be able to provide interesting forms of iGaming that, you know, customers and consumers might like, that's where kind of the next opportunity might sit.
这里面内容很多。
So there's a lot there.
而且，我想问一下你的看法：这部分指的是体育博彩业务，对吧？
And, you know, I think one question I did want to get your take here on is, this is the sports betting part of these operations, right?
当我们谈到预测市场时，我的意思是，你现在可以对任何事情下注。
When we think prediction markets, I mean, now you could bet on anything.
我一直以来都试图将体育博彩和预测市场业务区分开来，因为两者的受欢迎程度不同，监管方式也不同。
Could be betting on what you're going to say here on But show I have always tried to sort of segment the sports betting prediction market business from the rest of it, and where the popularity is, and also what is actually being regulated.
你在这里谈到的，只是体育博彩这一部分，对吧？
What you're talking about here, that's only the sports betting side of the Right?
我的意思是，嗯，
I mean, Well,
不一定。
not necessarily.
我的意思是，法院可能会认为，所有这些在他们看来都属于博彩形式，应当由各州根据第十修正案进行监管，而事实上，一些州已经出台了立法，禁止预测市场中非体育博彩的部分。
I mean, the the the court could say that all of this in their view is a form of gaming that is you know, should be should be controlled by the states that within the tenth amendment fits into state jurisdiction, then states, you could you've already seen legislation in different states to ban the non sports betting parts of prediction markets.
但目前这些法案毫无意义，因为它们本来就没有管辖权。
Right now, those bills don't mean anything because they don't have jurisdiction over it anyway.
但若最高法院将整个市场交还各州，确实有可能允许某些形式的预测市场，而禁止其他形式。
But it is certainly possible that if the Supreme Court sends the entire market back to the states, that it might allow certain forms of proxy markets and not others.
不过话说回来，我曾担任伊利诺伊州副州长四年，负责管理州预算。
But with that said, you know, I was the deputy governor of Illinois for four years, and I ran the state's budget.
你总是需要更多收入。
You always need more revenue.
在我看来，如果存在增加收入的方式，你大概不会说：‘我不想要。’
And it seems to me that if there is a way to generate more revenue, you're probably not going to say, I don't want it.
他们可能会规定某些特定类型的合约不被允许。
There might be very specific types of contracts that they say are not allowed.
但我认为，这些网站上提供的各种活动，无论由谁监管，都可能继续存在。
But my guess is that the variety of activities offered on the sites, will likely continue regardless of who regulates it.
那么，你认为谁应该监管这个领域呢？
So what's your view on who should regulate this?
你认为应该朝哪个方向发展？
Which direction do you think
我觉得目前的运作方式是有效的。
it You know, I think it's working.
我理解各州的立场。
So I understand the state's argument.
作为投资者和FanDuel的成员，当《体育博彩法案》被推翻时，我们从中获益良多。
Certainly as an investor and FanDuel, we benefited tremendously when passport was overturned.
但与此同时，我认为商品期货交易委员会在这方面做得不错。
But at the same time, I think the CFTC is doing a good job with it.
我认为消费者对它有真实的需求。
I think that there's real consumer demand for it.
我认为像CalShear这样的公司在运营其业务方面做得非常好。
I think companies like CalShear are doing a really good job running their businesses.
所以，从我个人的角度来看，我对目前的运作方式非常认可。
And so, you know, from my personal perspective, I am very much okay with the way it works right now.
但我同时也认为，如果最高法院确认了这一点，那么目前受各州监管的公司将会改变其行为，这既会给各州带来预算压力，也会为新型博彩和新兴初创企业创造机会。
But I also think that if that is affirmed by the Supreme Court, you're going to see shifts in behavior by the companies that are currently regulated by the states, and that will both create budget holds for the states and then opportunities for new types of gaming and new types of startups.
让我们来梳理一下当前所有热点监管问题。你曾深度参与过优步。
Let just run through the gamut of hot button regulatory issues right now So very you were very involved with Uber.
现在我们正经历自动驾驶的转型。
Now we have this autonomous vehicle shift happening.
目前你是否看到自动驾驶领域存在任何由监管造成的障碍，可能对这些自动驾驶公司构成影响？
Is there any roadblock right now that you see in terms of autonomous vehicles that the regulations could pose for the CO2 companies?
是的。
Yeah.
首先，这发生在联邦层面。
So one would be at the federal level.
联邦政府在监管自动驾驶汽车方面表现得极其糟糕，也就是说，他们根本什么都没做。
The federal government has done an abysmal job at regulating autonomous vehicles, meaning they've just done nothing.
因此，2015年，一项两党一致通过的法案在众议院能源小组委员会全票通过，为自动驾驶汽车建立了监管框架。
So in 2015, there was bipartisan legislation that passed the House Energy Subcommittee unanimously that created a regulatory framework for autonomous cars.
但自那以后，这项法案就再也没有进展。
It has never moved since.
为什么？
Why?
因为卡车司机工会害怕自动驾驶卡车，尽管事实上这毫无道理，因为目前司机严重短缺，我们迫切需要技术来填补这一缺口。
Because the Teamsters are scared of autonomous trucking, even though in reality it doesn't make sense because there's a massive shortage of truckers and you desperately need technology to fill the gaps.
但无论如何，特朗普自视为卡车司机工会的人，于是让他的交通部全面叫停了一切。
But nonetheless, Trump saw himself as a Teamsters guy, so he had his DOT just halt everything.
然后拜登也自视为卡车司机工会的人，于是他的交通部也这么做了。
Then Biden saw himself as a Teamsters guy, so he had his DOT do it.
现在特朗普又回来了。
And now we have Trump again.
因此，尽管各州内部对自动驾驶汽车和卡车已有监管框架，但我们仍然缺乏跨州的监管框架。
And so, while there are regulatory frameworks for autonomous cars and trucks intrastate, we still don't have an interstate regulatory framework.
所以，你知道，这是一个问题，因为人们需要能够跨越州界。
So, you know, that's a problem because people need to be able to cross state lines.
所以这是第一个问题。
And so that's number one.
第二个问题则与自动驾驶出租车有关，无论是Waymo、特斯拉、Uber，还是其他公司，出租车行业和出租车司机的政治因素都会成为阻碍。
Number two would be specific to, autonomous taxis, whether it's Waymo or Tesla or Uber drawing something out or whoever else, where, the politics of taxis and taxi drivers will get in the way.
例如，在纽约市，我们的市长佐拉姆·蒙达米并不希望Waymo进入，因为他自视为出租车司机的代言人，认为自动驾驶出租车会剥夺出租车司机的工作机会。
So for example, here in New York City, Zoram Mondami, who's our mayor, does not want Waymo because he sees himself as an advocate for the taxi drivers and sees that autonomous, taxis will remove potential jobs for taxi drivers.
从宏观角度来看，他可能没错，但问题在于，你该如何应对？
You know, he's probably not wrong about that in the big picture, but the question is how do you handle it?
答案是，你无法把释放出来的精灵再塞回瓶子里。
Which is you can't put the genie back in the bottle.
一旦出现了消费者需求更旺盛的更好方式，这件事就一定会发生。
Once there is a better way to do something that there's consumer demand for, it's going to happen.
依我看来，这里有两种可能的应对方式。
And so there's one of two approaches in the way I see it.
一种是蒙达尼可以选择无视这一点，尽全力去压制它。
One is, Mondani can choose to ignore that and just do his best to stifle it.
但它终究还是会发生的。
It will happen anyway.
而当它发生时，出租车司机们将遭受毁灭性打击。
And then when it does, the taxi drivers will be devastated.
当我在美国各地推动优步合法化的所有竞选活动中，出租车行业采取的就是这种策略，他们让政客们彻底禁止优步。
When I ran all the campaigns to legalize Uber around The US, that was the approach that the taxi industry took and got their politicians to take, which was just ban Uber completely.
而当我们赢得每一个市场的这场斗争后，出租车行业也因此遭受了严重打击。
And then when we won all of those fights in every market, taxi was really devastated as a result.
或者，你可以更明智地应对这个问题：在无人驾驶出租车的世界里，我们该如何思考出租车司机的未来？
Or you could approach this far more intelligently and say, how do we think about the future of taxi drivers in a world of driverless taxis?
而且可以设立自动驾驶出租车的牌照，并与此挂钩收取费用。
And there can be things like licenses for autonomous taxis and revenue attached to that.
这笔收入流可以用于设立出租车司机基金。
And that revenue stream could pay into a fund for taxi drivers.
因此，有更明智的方式来处理这个问题。
So there are intelligent ways to do it.
我希望像蒙达尼这样的进步派市长们会选择以这种方式参与进来。
I hope that people like Mondami and other ultra progressive mayors choose to engage that way.
但根据他们的言论、政治立场以及其他方面来看，你可能会看到他们只是说：科技不好，自动驾驶出租车不好，试图阻止这一切。
But, you know, from terms of their rhetoric and their politics and everything else, you could see them just saying, you know, technology bad, autonomous taxis bad, and just trying to stop it.
对。
Right.
在你走之前，我最后问你一个问题。
Let me ask you one last question before you go.
当你为一家公司提供建议，或者当你投资一家公司并担任监管顾问时——我们可以就此聊上很久——但布拉德利·塔斯克的策略是什么？
So when you are advising a company, or when you do invest in a company and you sort of take on this role as a regulatory advisor, and we could talk for a lot longer about this, but what is the Bradley Tusk playbook?
你向他们提出的战略是什么？
What what is the strategy that you put in front of them?
它包括哪些内容？
What does it include?
他们需要做些什么？
What do they need to do?
你如何建议初创公司应对监管？
Do you how do you advise, the startups to work around regulations?
是游说吗？
Is it lobbying?
那到底是什么？
What is it?
是的。
Yeah.
这是个很好的问题。
It's a great question.
因此，我们这里所有工作的核心理念非常根本：我们认为每一项政策结果都是政治投入的产物。
So there's a really kind of a fundamental thesis that drives everything that we do here, which is we believe that every policy output is the result of a political input.
每位政客做出的每一个决定，都仅仅基于赢得下一次选举，别无他求。
Every politician makes every decision solely based on winning the next election and nothing else.
因此，在任何特定情况下，你都必须弄清楚如何向相关当选官员或监管机构展示：如果你希望他们做的事，能提升他们连任的机会；或者，如果他们不按你的意愿行事，将会损害他们连任的可能性。
And as a result, you have to figure out in any given situation how to show the elected officials or regulators you need that doing what you want will further their chances of reelection, or if they don't do what you want, it's going to hurt their chances of reelection.
这可能意味着游说。
That might mean lobbying.
也可能意味着赢得媒体关注、付费媒体或社交媒体。
It might mean earned media or paid media or social media.
可能意味着一场大规模的草根运动。
It might mean a big grassroots campaign.
也可能意味着对对方进行负面攻击。
It might be going negative on the person.
有时，这可能意味着寻求许可。
Sometimes it might mean asking for permission.
有时候这意味着请求原谅。
Sometimes it might mean begging for forgiveness.
具体情况完全取决于公司、现有法律法规、管辖区域、你所颠覆的对象及其相对政治权力，以及其他诸多变量。
The context depends entirely on the company, the existing laws on the books, the jurisdiction, who you're disrupting, their relative political power, and a whole bunch of other variables.
我确实写了一本关于这一切的书。
And I did write a book about all.
这本书叫《修补者》，如果有人感兴趣的话。
It's called The Fixer if anyone's interested.
但它不仅在不同行业之间截然不同，在不同初创企业之间和不同司法管辖区之间也大相径庭。
But it totally differs not only from industry to industry, but from startup to startup and jurisdiction to jurisdiction.
我们可能对同一家投资组合公司，在一个州采用一种策略，而在另一个州采用完全不同的策略，仅仅因为在那里你需要这样做才能获胜。
We could have the same portfolio company and pursue one strategy in one state and a totally different strategy in another, simply because that's what you need to win there.
但从根本上说，任何政客都不关心你，也不关心你的初创公司。
But fundamentally, no politician cares about you, the startup.
他们不在乎你是否成为亿万富翁。
They don't care if you become a billionaire.
他们不在乎你是否成功退出。
They don't care if you have an exit.
他们不在乎你的关键绩效指标。
They don't care about your KPIs.
他们只关心自己的政治前途。
All they care about is their own political future.
我不知道你是否还记得亚马逊当初想把第二总部设在纽约市的时候。
I don't know if you remember when Amazon wanted to put their second headquarters in New York City.
那是2019年的事了。
Think was back in 2019.
而且，你知道的，成百上千的
And, you know, hundreds of
我住在多伦多。
cities I was living in Toronto.
多伦多，我们当时
Toronto, we were
对。
Right.
你们在侧街项目上也搞砸了。
You guys screwed it up too with with Sidewalk.
好吧。
Okay.
你们也搞砸了，但纽约市也搞砸了，亚马逊也是，因为有几百个城市在竞争这个项目。
You guys also messed it But but New York City messed it up because as did Amazon because, you know, a couple 100 cities competed for this.
纽约莫名其妙地稍微赢了一点。
New York kind of won a little bit out of nowhere.
我们真没指望能赢。
Don't think we were really expecting to win.
大家都非常兴奋。
And everybody's very excited.
然后奥卡西奥-科尔特斯出来反对，因为她的政治立场是不喜欢资本主义、工作和科技这类东西。
Then AOC comes out against it because her politics are she doesn't like capitalism, jobs, technology, things like that.
好吧。
Fine.
她就是这样的。
That's who she is.
所以，那个皇后区街区的州参议员名叫迈克·简里罗斯。
So the state senator for that neighborhood in Queens is a guy named Mike Janeiros.
他就是一个普通人。
And he's just a guy.
他并不优秀。
He's not good.
他也不邪恶。
He's not evil.
他不是保守派。
He's not conservative.
他也不是自由派。
He's not liberal.
他只想保住自己的职位。
He just wants to stay in office.
就是这样。
That's it.
他对自己说：‘天哪，现在AOC反对这个，如果我继续推动并让它实现，会发生什么？’
And he says to himself, Oh no, now that AOC is opposing this, what happens if I continue to push it forward and make it happen?
他根据历史投票数据知道，下一次初选的投票率很低，而且由于选区划分不公，只有初选才重要，投票率可能只有9%。
And he knew, based on historical turnout data, that turnout in his next primary, and because of gerrymandering, only the primaries matter, were probably about 9%.
这些人是该选区最左翼的选民。
Those are the most left wing voters in the district.
如果有人从左翼挑战他，尤其是得到AOC的支持，他们完全可以利用这个问题将他拉下马。
And if someone ran against him from the left, especially with AOC's backing, they could use this issue to take him out.
他面临一个选择。
And he had a choice.
四万个来自纽约人的新工作岗位，高薪且有福利的工作，或者仅仅保住他自己的职位，他选择了自己。
40,000 new jobs from New Yorkers, good paying jobs with benefits, or one job, his own, and he picked himself.
问题不在于迈克·热内里斯是个例外。
And the problem isn't that Mike Generis is the exception.
问题在于迈克·热内里斯才是常态。
The problem is that Mike Generis is the norm.
这就是政客们的做法。
That's what politicians do.
所以，除非你能向世界上那些迈克·热内里斯们表明，嘿，如果我们想要的你做了，不仅有助于你连任，如果你不做，就会损害你连任的机会。
And so unless you can show the Mike Generis of the world that, Hey, doing what we want will not only will help you get reelected, reelected, or if you don't do what we want, it will hurt your chance to get reelected.
否则他们根本不在乎，也不会按你的意愿行事。
Then they don't care, and they're not going to do what you want.
所以，关键总是要理解他们需要什么才能连任，以及你的主张如何被包装，才能让他们相信帮助你就是帮助他们自己。
So it's always about understanding what do they need for reelection and how does your thing get framed in a way that convinces them that helping you helps them.
很好。
Great.
好了，布拉德利，感谢你前来。
Well, Bradley, I want thank you for coming on.
这就是Tusk Ventures的创始人兼首席执行官布拉德利·塔斯克，正在TI TV上。
That is Bradley Tusk, founder and CEO of Tusk Ventures here on TI TV.
是的。
Yes.
今天的节目就到这里。
That does it for today's show.
提醒一下，我们每周一至周五上午10点（太平洋时间）/下午1点（东部时间）在此直播。
A reminder, we are on this stream Monday through Friday at 10AM Pacific, 1PM Eastern.
感谢大家收看。
I want to thank you all for tuning in.
我们非常感谢您的观看。
We really do appreciate your viewership.
请务必在YouTube上订阅我们的频道，并在X、Instagram、TikTok上关注我们，同时在您收听播客的平台关注我们。
Make sure you subscribe to the information on YouTube and follow us on X, Instagram, TikTok, and check us out wherever you get your podcasts.
我已经迫不及待想看明天的节目了。
I'm already excited for our next show tomorrow.
祝你周二剩下的时间愉快。
Have a great rest of your Tuesday.
暂时再见。
Bye bye for now.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。