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。
这是我们见过最具活力的市场之一，尤其是在软件采购方面，每个月都能看到老牌企业被新进者取代。
This is one of the most dynamic markets we've seen, particularly for buying software, where month over month, you will see large incumbents be replaced by the newcomers.
Anthropic最近就做到了这一点，取代了OpenAI，现在根据Ramp的数据，它已成为企业最常使用的模型。
Anthropic just did that with OpenAI, now the most popular model used by businesses according to Ramp data.
Cursor也对GitHub Copilot做了同样的事。
Cursor did that with GitHub Copilot.
我的主要看法是，所谓‘SaaS末日’的论调来得太早了，而且通常并不基于真实的商业行为。
My main take is that SaaSpocalypse as a pronouncement has come way too soon and is typically not informed by actual business behavior.
你的意思是，‘SaaS末日’在数据里根本找不到？
You're saying that SaaSpocalypse is not in the data?
从数据上看，‘SaaS末日’的两个关键主张都没有得到实际企业支出的支持。
I'd say quantitatively neither aspect of SaaSpocalypse is supported by actual business spend.
我说的两个关键主张，一个是它会彻底改变我们的采购方式。
Neither aspect as in it's going to change the way we buy it.
但目前来看，它并没有在任何实质层面改变采购方式，也没能淘汰那些经常被提及的公司。
Currently, it has not in any meaningful way nor has it killed off at least the companies that are are frequently mentioned.
本集最初播出于MTS栏目《货币事务》。
This episode originally aired on the MTS segment Monetary Matters.
过去两年，软件领域的主流说法是，人工智能将彻底颠覆SaaS的按席位定价模式，并让少数前沿模型提供商垄断一切。
For the last two years, the dominant story in software has been that AI will wipe out SaaSpocalypse collapse seat based pricing, and centralize everything around a handful of frontier model providers.
但当你查看实际的企业支出数据时，这个说法就很难站得住脚了。
But when you look at actual business spending data, that story becomes much harder to defend.
许多增长最快的AI公司并不是模型实验室本身，而是围绕这些模型构建的基础设施、工作流和应用层。
Many of the fastest growing AI companies are not model labs themselves, but the infrastructure, workflow, and application layers forming around them.
与此同时，企业越来越多地使用多种模型，变得更加注重成本，也在以各种方式尝试AI，而这些做法并不完全符合当前关于自动化和替代人力的主流叙事。
At the same time, businesses are increasingly using multiple models, becoming more cost conscious, and experimenting with AI in ways that don't neatly fit the prevailing narrative around automation and labor replacement.
杰克·法利和马克斯·维特与兰普公司的首席经济学家阿拉·哈拉齐安进行了对话。
Jack Farley and Max Wiethe speak with Ara Kharazian, lead economist at Ramp.
我们请来了兰普经济团队的首席经济学家。
We have the lead economist from Ramp Economics.
阿拉·哈拉齐安，非常感谢您今天做客我们的节目。
Ara Kharazian, thank you so much for joining us today.
阿拉
Ara
谢谢你们。
Thank you, guys.
你写过好几篇文章，都是围绕这个话题的。
Written you've written a number of pieces looking at this very topic.
人们的支出模式有变化吗？
Are people changing their spending patterns?
你们在Ramp那边观察到什么情况？
What are you guys seeing at Ramp?
是的。
Yeah.
我觉得这项工作特别重要，因为我们身处科技行业，人人都喜欢对市场走向做出宏大预测。
I mean, I think that's what's so important about this kind of work is that, you know, we live in this world, particularly in tech specifically, where everyone wants to make these big pronouncements about where the market's gonna go.
每个人都像专家一样。
Everyone seems like a pundit.
没人手头有真实的数据来了解市场到底发生了什么。
No one's really armed with any data to inform them about what's actually happening in the market.
所以我这份在Ramp的工作很特别，我们能接触到五万家企业的支出数据，年支出总额高达一千亿美元。
So I have this unique job at Ramp where, you know, we see the spend data from 50,000 businesses, a $100,000,000,000 of annual spend.
于是我们决定深入研究市场真实的情况。
And so we did set out to research what is actually happening in the market.
新的AI公司正在推出各种产品，跟许多知名的SaaS公司竞争。
New AI companies are coming out with competitor products to a lot of prominent SaaS firms.
我们有没有看到这些传统SaaS公司的使用率在下降？
Are we seeing any declines in the in adoption for those traditional SaaS companies?
人们购买SaaS产品的方式有没有发生变化？
Are we seeing any changes in how people are buying SaaS?
正因如此，我把所谓的SaaSpocalypse分成两类。
And that's really where I divide SaaSpocalypse into two different categories.
对吧？
Right?
有一种说法是，
There's one that's, hey.
人们是不是正在从传统SaaS转向那些AI公司推出的竞品？
Are people shifting away from traditional SaaS over to over to, you know, competitors provided by the model companies.
第二个是，人们购买软件的方式有没有变化？
And then number two is the way people are buying software changing.
会不会出现一种新模式，比如你买的是智能代理功能，或者买的是积分，而不是付SaaS平台的订阅费或席位费？
Are people shifting to a new model where you buy, like, agentic, you know, capabilities or you buy tokens instead of paying a a SaaS platform fee or seats?
但目前来看，这两种趋势都没真正发生，至少还没形成任何实质影响。
And right now, we see that neither of those trends are happening, at least not happening in any meaningful way.
按席位收费的合同仍然占到六到七成的支出，固定平台费用约占三成；即便是那些推出了积分定价模式的公司，比如HubSpot和Adobe，这类模式的使用率也只占总支出的不到百分之零点五。
Seat based contracts are still about 65, 75% spend, flat platform fees about 30%, even amongst the companies that have offered some kind of token based pricing for their SaaS tools, we're only seeing uptake about half a percent of spend on those platforms.
像HubSpot和Adobe这样的公司。
That's companies like HubSpot and Adobe.
它们都为自己的SaaS产品提供了一些基于积分的选项。
Both of them offer some kinds of token based offering for their SaaS offerings.
如果你想讨论人们是否在离开 SaaS 提供商，可以看看 Figma，毕竟他们推出了 Cloud Design。
And then you wanna talk about whether or not people are shifting away from SaaS providers, you can look at Figma where, you know, Cloud Design came out.
大家都只在谈论 Figma 强劲的业绩和财报。
Everyone was just talking about Figma's strong performance and earnings.
但我们一个月前在数据里就已经报告过，Figma 是我们平台上增长最快的供应商之一，企业还在持续购买它的服务。
But we'd reported in our dataset already a month ago that Figma was one of the fastest growing vendors on our platform, that businesses were continuing to buy from it.
所以我觉得这说明，很多这类说法其实只是人们对产品未来发展的推测，但并没有真正反映现实。
And so I think it just goes to show you that a lot of these pronouncements, they're coming from people's projections about how how products may develop, but they're
这些判断往往没有基于企业实际支出的变化。
not often informed by the actual changes happening with respect to business spend.
所以听起来，实际发生的变化比人们说的要少得多。
So it sounds like less has changed than people are talking about.
你觉得真正的变化发生在哪儿？
Where would you say the real change is happening?
哪些方面正在发生细微的改变？
What is changing on the margin?
我确实觉得，越来越多的软件供应商开始推出某种基于代币的产品。
I think it's definitely true that an increasing number of vendor software companies are starting to offer some kinds of token based products.
比如我刚提到的Adobe，它很清楚，自己的产品正越来越多地被使用AI功能的用户所采用，而这些功能的边际成本是存在的。
Adobe, which I just mentioned, is one of them, which knows that its product is increasingly going to be used by people who using the AI capabilities that has a marginal cost associated with it.
未来或许会出现AI代理的使用场景，那时候你就没法像以前那样按座位收费了。
Maybe there will be some down the line usage by an AI agent that you can't charge a a seat in the traditional way.
所以越来越多公司开始涉足这一领域，但就目前企业采购SaaS产品的实际支出来看，代币模式还只占不到百分之零点五。
So more companies are starting to offer that, but whether or not that's actually playing out today as far as how businesses are buying SaaSpoc, right now it's still half a percent of actual spend.
如果这种模式真要大规模普及，理论上，企业推出它一定是希望借此抢占市场份额或提升内部使用率，对吧？
So for that to really take off, I mean, in theory, you would only offer that as a company if you believe it's gonna allow you to take market share or increase usage, right, among like, within an org.
那是不是得有公司主动去抢别人的饭碗，用代币模式逼着整个市场接受它？
So is somebody gonna have to go after everybody else's lunch with a token based model to to force that upon the rest of the market?
到底什么样的情况，才会让所有人都彻底转向这种模式？
Like, what scenario would people fully switch?
我们会看到它真正兴起吗
Would we see that take
结束吗？
over?
嗯，也不是。
Well, no.
你说到点上了，这一切最终都会是这个领域里那些深思熟虑、资金雄厚的公司相互竞争的结果。
You got a really good point too that all of this is going to be the result of a series of competitive responses by all of the very thoughtful, well capitalized firms in this space.
你看Anthropic和Figma，现在两家都推出了设计类产品。
I mean, you look at Anthropic versus Figma, both of them now offering a design product.
Anthropic本身并没有什么特别之处，让它比别人更懂设计师的需求，或者更能做出设计师真正需要的软件。
There's nothing inherent about Anthropic that makes it more capable of reaching designers where they're at or producing software that designers need.
它只是拥有那个模型。
You know, it has the model.
它有快速迭代产品的能力，能持续优化。
It has the velocity product velocity to continuously make improvements.
但Figma已经是广受欢迎的产品，而且也能用上Anthropic开发和销售的同样模型。
But Figma is a very popular product, which has access to the same models that Anthropic produces and sells.
所以我一直没太明白，那些说‘SaaS末日论’的人，他们为什么觉得会是这样。
And so I I haven't really understood from SaaSpocalypse people what the intuition is to suggest that, hey.
说我们会看到这么多SaaS公司倒闭。
We're gonna see all these SaaS companies die off.
不过我确实觉得会有竞争。
Now I do think that there's gonna be competition.
对吧？
Right?
Anthropic 确实是有实力的新玩家，能进入不少市场，但这跟说大家会立刻转向这些新公司，完全是两码事。
Like, Anthropic is a worthy new entrant into a lot of different markets, but that's a very different claim than saying that, hey.
毕竟企业的采购习惯往往比较固定，不太容易轻易改变。
We're immediately gonna start to see businesses shift over to to these to these new players, and especially because business spend can be pretty sticky.
不过反过来说，这也是我们见过最活跃的市场之一，特别是在软件采购方面，每个月都能看到老牌巨头被新面孔取代。
Now the counterpoint to that, though, is that this is one of the most dynamic markets we've seen, particularly for buying software where month over month, you will see large incumbents be replaced by the newcomers.
Anthropic 最近就做到了这一点，根据 Ramp 的数据，它现在已经取代 OpenAI，成为企业界使用最广泛的模型。
Anthropic just did that with OpenAI, now the most popular model used by businesses according to Ramp data.
Cursor 也是这么取代了 GitHub Copilot。
Cursor did that with GitHub Copilot.
所以我觉得这事儿确实值得留意。
So I think it's, you know, something worth keeping track of.
但我主要的观点是，所谓‘SaaS末日’这个说法来得太早了，而且通常根本没考虑实际的商业行为。
But my main take is that SaaSpocalypse as a pronouncement has come way too soon and is typically not informed by actual business behavior.
你是说‘SaaS末日’在数据里根本看不出来？
You're saying that SaaSpocalypse is not in the data?
如果我们能看一下这个图表的话
That that that if we could share this this chart
那我可以明确地说，从实际的企业支出来看，‘SaaS末日’的两个说法都站不住脚。
that I'd say quantitatively neither aspect of SaaSpocalypse is supported by actual business spend.
两个说法指的是：一是会彻底改变我们采购软件的方式，二是会淘汰掉那些常被提及的公司。
Neither aspect as in it's going to change the way we buy it.
目前来看，它根本没有在实质上改变采购方式，也还没让那些常被提到的公司倒闭。
Currently, it has not in any meaningful way, and nor has it killed off at least the companies that are frequently mentioned.
再看看Perplexity这家公司。
And even look at Perplexity.
人们总说Perplexity在纯AI模型模式里风险很高，比Figma还危险，但事实上，Perplexity在Ramp平台上的增长速度却是最快的供应商之一。
People talk about Perplexity being an at risk company more in the pure play model space than a than a Figma, and yet Perplexity is also one of the fastest growing vendors on Ramp.
它的用户增长特别快，主要是因为推出了那些大模型公司还没跟上的产品。
Seeing a lot of uptake specifically because it's offering products that the model companies have not competed on yet.
你刚才还提到，这些软件公司基于令牌的支出在增加。
So so you talked about, as well, like, the increase in token based spend with some of these software companies.
这种增长到底是在给现有的产品体系增添AI功能，还是说这些功能其实流向了所谓的竞争对手，而不是直接取代了按座位收费的模式？
To what extent is that adding AI capabilities to the existing sort of stack of products and and many some of that even being, like, flowing through to these supposed competitors versus it replacing a seat based a a seat based consumption.
对吧？
Right?
本质上是给他们的产品线加了AI功能，用户可能按令牌付费，但背后还是保留着某种按座位收费的结构。
It it's it's adding AI capabilities to their suite of products and that you might pay token based, but you still have some sort of, like, seat based thing on top
上面还叠加着座位收费的部分。
of it.
是的。
Yeah.
确实是叠加的，但规模还是太小了。
It is additive, but it is still so small.
再说一遍，我们说的是那些第一次认真尝试的公司，它们在平台上的实际支出中，AI相关花费还不到百分之一。
Again, we're talking about at the companies that are doing it for the first time, but doing it in an earnest way, less than a percent of the actual spend that's happening on their platform.
所以很难判断这是否真的为他们的利润带来了额外增长。
So it's hard to measure whether or not it is truly something additive to their profits.
Anthropic超越OpenAI，现在大家应该都对这个故事挺熟悉的。
Anthropic surpassing OpenAI, I think everybody is is kind of familiar with with that story at this point.
那其他一些模型呢？
What about some of the other models?
你在模型支出方面还看到哪些趋势，是大多数人还没怎么注意到的？
What are the trends that you're seeing in terms of spend on models that maybe people aren't nearly as aware of?
是的。
Yeah.
其实我在关于Anthropic的那篇文章里，语气上还留了点余地，对它未来的前景没敢太肯定。
Well, I in my post about Anthropic, I actually was a little bit hedging, if you would say it, about Anthropic's prospects going forward.
我们从去年开始推出一个叫Ramp AI指数的指标，用来追踪Anthropic和OpenAI在企业中的采用率。
I think when we put out because we started putting out this metric last year called the Ramp AI Index, which is tracking the sort of business adoption rates for Anthropic versus OpenAI.
很长一段时间里，OpenAI一直领先。
And for a long time, OpenAI was the leader.
最近Anthropic才刚超过它，拿下了第一名。
Anthropic just took the first place spot.
但在发布这些研究的过程中，我发现很多人会误解我的意思，以为我在说‘完了，胜负已定’。
And what I found over the course of putting out this kind of research is that people will often interpret my findings as me saying, oh, it's over.
谁第一、谁第二，他们就当成了最终结局。
OpenAI is the leader or Anthropic's the leader.
但这不是我的本意，我也根本不觉得市场会这样发展。
That's typically not first, this is not my intention, but it's also not how I see this market playing out.
现在我们平台上越来越多的公司，都在同时使用多个模型，以各种方式部署到员工工作中，不管是通过订阅服务，还是集成到他们自己的AI原生产品里。
An increasing share of firms on our platform are using more than one model in in some deployed way across workers, whether that's a subscription based deployment or if it's something that they're building into their AI native product.
早期采用这些技术的公司最有可能同时使用多个模型，并随着时间推移不断增加其他AI供应商。
Firms that were early adopters are the most likely to use multiple models and add more soft add more AI vendors over time.
所以，如果你想从早期采用者身上看出整个市场未来的发展趋势——这在我们的数据中一直是AI采用的典型模式——那就可以合理推测，普通企业最终也会接入两到三个模型。
So if you wanna consider the early adopters of sign of where the rest of the markets are gonna go, and that has tended to be how AI adoption has developed in our dataset, then you can assume that the average business is also most likely going to have a couple onboarded models.
同时，也有迹象表明，企业在模型使用上的支出正变得越来越谨慎。
And then with that, there are signs that companies are becoming increasingly cost conscious as far as how much they're spending on their models.
对于那些按使用量付费的普通企业来说，过去一年里，API调用和高用量AI代理的token成本整整涨了13倍。
For the typical business that spends on tokens, so APIs and then also, like, the very high usage agent to coding, co token costs for that business have increased 13 x just over the last year.
不过对这些高投入企业而言，这部分支出占总运营成本（不含工资）的比例大概还是只有2%左右。
It's still about for those high intensity spenders, maybe about 2% of business spend, excluding payroll.
也就是说，AI支出在整个企业开销里占的份额其实很小。
So it's a very small share of actual spend that's happening.
但如果你把这13倍的增长趋势继续外推下去，就会看到一个完全不可持续的支出路径。
But you project out that 13 x, and you get to an extremely unsustainable spend path.
对吧？
Right?
大多数企业做不到，也不该这么做。
Most businesses can't do that and probably shouldn't.
正是这些公司正在把一部分AI支出转向像OpenRouter这样的平台。
And it's those firms specifically that are increasingly shifting some of their AI spend over to, you know, platforms like OpenRouter.
任何能让你在多个模型间选择、最好还能利用价格更便宜的开源模型的做法，都会有所帮助。
So anything that would allow you to select between multiple models and ideally take advantage of free and open source models that are priced a little bit cheaper when they're deployed.
即便如此，这仍然只占支出的很小一部分。
It's even that is still a very small share of spend.
在我们的平台上，有大约3%的AI支出直接通过OpenRouter，而不是Anthropic，而且这个比例正在上升，尤其集中在那些最希望使用OpenRouter这类平台的公司——也就是AI支出最高的那些企业。
They've got, like, 3% of spend on our platform goes directly through an OpenRouter versus an Anthropic, at least the AI spend on our platform, but increasing and particularly concentrated amongst those firms that are most incentivized to use platforms like OpenRouter, the ones that spend the most on AI.
他们在那儿找到了最佳性价比，我认为这些做法迟早会向下渗透，被更主流的企业采纳。
So they're finding the best value there, and I think it's likely that those practices will start to move down the chain and be adopted by the more mainstream firm too.
开始选择OpenAI和Anthropic之外的模型，比如用OpenRouter的模型或Anthropic的模型，但根据任务需求，自动切换到更便宜的选项。
Start to opt for models outside of OpenAI OpenAI and Anthropic or, you know, use OpenEye's models and Anthropix models, but use something that routes them over to the cheaper options when that makes most sense for the task.
这些就是我看得到的、对模型公司的主要挑战。
So those are the main headwinds I see on the model companies.
成本和竞争，尤其是来自那些更便宜的模型，还有可用的开源版本。
Cost and competition, particularly from arguably themselves in terms of the cheaper models that are available and then also the open source versions available.
这听起来像是对AI支出的一个相当强烈的看空观点。
I mean, that sounds like a pretty big bear case for just for spend on AI.
你觉得人们现在多大程度上意识到，随着我们越来越擅长创建和训练自己的智能体，很多任务其实根本不需要顶尖模型？
To what extent do you think people are figuring out that, oh, these tasks, now that we're getting better at creating and training our own agents, that they just don't need the frontier models?
我们做的工作中，有多少其实根本用不上顶尖模型？
Like, how much of our work that we do doesn't require a frontier model?
到目前为止，大家总是习惯直接堆算力来解决问题。
And up to this point, it's always just been like, let's just throw compute at the problem and that will solve it.
但实际上，更关键的是要多花点心思，想想流程该怎么优化。
And that it's actually just gonna be like thinking a little harder about what the process should be.
是的。
Yeah.
并不是什么时候都需要顶尖模型。
You don't need the frontier model all the time.
我的意思是，很多时候，你想要的效果反而更差，因为又贵又慢。
I mean, oftentimes, it's worse for what you want because it's expensive and slow.
你懂的。
You know?
我平时工作中经常不自觉地把任务直接丢给最贵的那个模型，尽管它更慢，其实我更希望在不需要的时候用Haiku或者Sonnet就行。
I find myself in my day to day oftentimes, you know, automatically routing things to the most expensive model even though it's slower, and I would actually prefer it just use Haiku or Sonnet when it's necessary.
但你不能指望普通员工自己做这种判断。
But you can expect the typical worker to make that decision.
我的工作就是关注AI领域的动态，所以我对新出的模型都了如指掌。
I mean, I'm fairly my job is to be informed about AI trends, so I know all the new models that are coming out.
我也还算懂怎么用最佳实践，但这背后需要花大量时间跟踪市场变化。
And I'm relatively good at about applying best practices, but I don't think that that but that comes with a lot of tracking of the market.
我不觉得普通上班族，或者一般打工人，应该被要求去做这些事，他们也没必要这么做。
I don't think that you can expect most American office workers or workers in general to to do that, and they shouldn't have to.
但目前市场还没跟上这种需求。
And the market hasn't caught up to that use case yet.
OpenAI 和 Anthropic 完全没有动力推出自动路由产品来帮你降低 AI 开支，因为他们的收入来自 token。
OpenAI and Anthropic have no incentive to offer an auto routing product that allows you to lower your AI spend because they make money on tokens.
顺便说一句，他们靠 token 赚的钱比其他 AI 前沿公司还多。
And by the way, they make more money on tokens than other AI frontier labs and firms.
谷歌倒是靠很多其他业务赚钱。
Like, Google makes money from a lot of stuff.
他们并不指望你非得在 token 上花大钱。
They don't need you to spend money on tokens explicitly.
Anthropic 和 OpenAI，差不多 80% 的收入都来自 token。
Anthropic and OpenAI, like, 80% of their business revenue is token based.
所以他们的收益直接跟使用量挂钩，自然希望你用得越多越好。
So it's directly tied to usage and incentivizing you to do that.
我们有没有
Have we
看到过
seen a
我们有没有见过这样的软件插件，能自动判断你的任务该用哪个模型？
have we seen, like, a a software plug in that that basically determines what sort of model your task should go to?
Cursor 就是这么做的。
I mean, Cursor does that.
Cursor 会自动处理，我觉得模型公司迟早得应对这种竞争压力，尤其是 Anthropic。
Cursor will automatically I mean, I think the model companies will have to respond to that kind of competitive pressure increasingly because especially Anthropic.
对吧？
Right?
现在越来越多的人已经碰到算力瓶颈了。
People are increasingly hitting these compute limits.
确实如此。
Mhmm.
所以他们最好早点做出调整，这样更多人才能更高效地使用这些服务。
And so it behooves them to make that change first so that at least more people can use it most effectively.
需求肯定是有的，但我认为第一步要看有没有其他公司率先推出这类功能，从而形成竞争压力。
They certainly have the demand for it, but I think it's gonna come from whether from it's gonna the first step is gonna be whether or not other firms that are in this space start offering it first and end up being that competitive pressure.
顺便说一下，这就是像 Cursor 这样的产品之所以有优势的原因——它能从开发者体验的角度去竞争，而 OpenAI 和 Anthropic 在没有 Cursor 的情况下，并没有动力去做这件事。
And by the way, that's the bull case for something like Cursor, which can compete on the developer experience in that way, in a way that OpenAI and Anthropic in the absence of Cursor are not incentivized to do.
它们可以把模型做得更便宜、性能更强，或者把路由功能做得更好，直接整合到产品体验里。
They can offer a cheaper, more performant model or better routing as part of the product experience.
嘿。
Hey.
好的。
Alright.
你这张图里显示得很清楚，现在 Anthropic 领先于 OpenAI，然后是 Google，再是 x AI。
So in this great chart you've got, we've got Anthropic now leading OpenAI, then Google, then x AI.
最底下那条蓝线，差得太多了，几乎都看不出来。
Way, way, way at the bottom is the blue line, people can barely see it,
那是 DeepSeek。
is DeepSeek.
这是怎么回事啊，Ara？
What happened, Ara?
我记得五个月前，大家还说有80%的风投支持的科技初创公司都在用DeepSeek，因为它的成本实在太低了。
I thought, you know, five months ago we were hearing about how 80% of venture capital backed tech startups were actually using DeepSeek because DeepSeek was so cheap.
如果真是这样，那为什么我在图表上几乎都看不到这条蓝线呢？
If that's the case, how come I can barely see this blue line on my chart?
这到底是怎么回事？
What's going on?
我不认为有80%的风投支持公司真的在用DeepSeek。
I don't think 80% of venture backed companies are using DeepSeek.
我们确实用过。
We did.
你看。
Look.
DeepSeek刚出来的时候，大概一年前，甚至更久一点，我们确实看到它的使用量有个小高峰，但从来没超过平台上任何一家公司的1%。
When when DeepSeek first came out to the scene, now about a year ago, maybe a little bit over, we did see a spike in adoption for DeepSeek, but it never hit even 1% of firms on the platform.
我们现在还把它保留在这个追踪列表里，因为觉得对大家有点参考价值，不过说实话，现在也有人觉得该删了。
Now we keep it in this tracker because I think it's helpful context for people, though there was a case to say we should remove it at this point.
但DeepSeek并不是唯一一个前沿模型，也不是非前沿模型。
But DeepSeek is not the only model frontier not the not frontier.
DeepSeek并不是市场上唯一可用的开源或低成本模型。
DeepSeek's not the only open source or cheaper model available out there.
所以在Ramp AI指数的未来版本中，我们会扩展追踪范围，去监测市场上其他一些低成本模型的使用情况。
So in future versions of Ramp AI Index, we're gonna expand, though, to track adoption of a bunch of other cheaper models that are available on the market.
即使模型是开源的，你在云上部署时可能仍然要付钱。
And even if it's open source, you know, you might still be paying for it in some way in the cloud implementation of it.
这正是你从类似开放路由这样的服务中可能得到的情况。
That's something exactly what you might get from, like, an open router, for example.
我的下一个问题是，DeepSeek这个案例其实挺特别的，因为它确实可能很便宜。
My next question I don't so DeepSeek DeepSeek is a tricky case, though, because it might be cheap.
但即便如此，也没人愿意为了安全原因使用DeepSeek。
But even if you are no one wants to use DeepSeek for security purposes.
哪怕你本地部署，大家对它的印象还是有顾虑，尤其是当你在开发面向企业或消费者的产品时，现在已经有其他更稳妥的选择了。
Even if you're locally hosting it, there's just this, like, perception around it, particularly if you're building a product that appeals to businesses or consumers, and there are other options available at this point.
所以我真不觉得 DeepSeek 能很快被广泛采用。
So I don't I don't really see DeepSeek gaining adoption very quickly ever.
而且。
And
是的。
yeah.
谢谢。
Thank you.
那谷歌的 Gemini 模型呢？
And and what about Google's Gemini model?
我是说，这条橙色线在图表上低得有点明显。
I mean, the yellow the the orange line, quite low on this chart.
你觉得今天的新闻会改变这个情况吗？
Is today's news gonna change that, you think?
谷歌其实被低估了，我承认。我们这项研究的一个关键点是，我们只追踪付费使用的情况。
Google's underrated, I'll admit, because let's I'll point out one of the features of our research here is that we only track paid adoption.
所以有很多公司都在用谷歌，但他们是通过谷歌工作区来使用的，里面免费集成了Gemini。
So there are a lot of firms that are using Google, but they're using it through Google Workspace, which integrates Gemini for free.
我们就是其中之一。
That's us.
你可以
You can
我们也是这么做的。
That's what we do.
对。
Yep.
画得完全正确。
Draw exactly.
所以你可以划定一条线，看看什么样的情况才算作AI的采用，以及这种采用需要达到什么程度才算有意义。
So you can draw the line about what counts as AI adoption and how significant that adoption needs to be.
也许你需要一定规模的付费采用，才能真正看到一家公司在生产力上的提升。
Maybe you need some amount of paid adoption to really see productivity gains across a firm.
目前谷歌提供的服务，我觉得大多数经济学研究者都怀疑，单靠聊天订阅能否带来我们期望的经济生产力提升。
Google's offering right now I think most researchers in econ are skeptical that chat subscriptions on their own will be driving the kinds of productivity gains we wanna see in the economy.
我们可能需要的不只是谷歌 Workspace 那种使用方式，大多数人也都这么认为，得有更全面的方案。
We might so we'll probably need something more comprehensive than just the kind of usage you get out of Google Workspace, at least most people do.
但谷歌确实被低估了。
But but Google's definitely underrated.
它确实有分布上的优势，因为所有使用谷歌 Workspace 的企业都接触得到它。
It certainly has a distributional advantage in that it's at all businesses that use Google Workspace.
在 Ramp AI 指数里，你知道的，入选公开市场的那一天可是大事。
So on on the Ramp AI Index, you know, in public markets, index inclusion day is a big deal.
大家都会猜测哪家公司会被纳入标普 500，那会带来巨大影响。
People speculate what's the company that's gonna be added to the S and P 500, it's gonna create all this.
到底是什么才能让一个模型真正脱颖而出？
What what is gonna make a model?
显而易见，Anthropic、OpenAI 和其他公司之间差距巨大。
Obviously, there's a huge gap between Anthropic, OpenAI, and everybody else.
那谁在入围名单里，有望被加入这个AI指数？你们具体看哪些标准来判断，嘿，这家公司值得上榜？
Who who's on the shortlist to be added to the Ramp AI Index, and what are the sort of criteria that you're looking at to say, hey.
就是哪些公司配得上跟那些行业巨头们一起列在榜单上？
This is a company that that deserves to be, you know, on the list with the big boys?
嗯，我是说，从衡量经济生产力的角度来看，大家真正关心的其实不是看哪家公司排第一、第二还是第三。
Well, I mean, I'll I'll I'll suggest actually that from a economic productivity measurement standpoint, the main interest is not to see, hey.
而是看哪家模型公司领先，哪家排第二，哪家排第三。
Which model company is ahead versus number two versus number three?
采用趋势会向几家大型玩家集中。
Adoption is going to trend toward multiple large players.
我们研究的下一阶段是不仅要衡量企业的采用情况，还要衡量采用的强度，并尝试给出一个定义，比如，嘿。
The next stage for our research is to measure not just business adoption, but the intensity of that adoption and try to come up with some definition of, hey.
一家企业如何能特别有效地采用AI，以及实现这一目标的路径是什么？
What does it look like to be a firm that adopts AI particularly well, and what is the path to that?
这个问题并不容易回答，因为大家的本能反应是，好吧，让我们看看那些在AI上投入最多的企业，但投入并不等同于生产力的提升。
And that's not an easy question to answer, you know, because the instinct is, well, let's look at the firms that spend the most on AI, but spend does not translate to productivity gains.
如果一年前问大家，他们会说，AI的成本迟早会降到零，根本没必要去衡量支出。
If you ask people a year ago, they would say, well, the price of AI is gonna go to zero, so there's no point in measuring spend.
不过，现在有不少很有趣的研究在探讨企业能获得哪些生产力提升，以及我们该用哪些具体指标来衡量这些成果。
So there's a lot of really cool research happening, though, about kind of productivity gains that a firm might get and trying to come up with the outcomes that we would measure.
人员数量当然是一个重要指标。
Headcount, of course, is a big one.
AI会对工作岗位产生什么影响？
How is it gonna affect jobs?
哪些岗位会增加？
Which jobs go up?
哪些岗位会减少？
Which jobs pull back?
特别是在软件工程领域，已经有一些很有趣的研究在分析代码提交的数量，还有代码质量的评估方法。
In engineering, specifically, software engineering, you can there's some really interesting research about quantity of PRs, some quality measurement of PRs too.
我觉得，这正是我们研究接下来要深入的方向。
So I think that's really the next stage of our research.
在这项研究的早期阶段，你观察到什么了？
What have you seen in this early stages of this research?
很多人猜测，那些CEO把裁员归咎于AI的举措，其实只是普通的裁员而已。
There's been speculation a lot of these layoffs that CEOs are attributing to AI adoption are really just regular old layoffs.
之前招人太多了，现在把裁员说成是AI造成的，听起来是个方便的借口，比如裁掉10%的员工。
There was overhiring, and it's a convenient way to say, this is why we're laying off, you know, 10% of the workforce.
你真看到那些声称因为采用AI而提升效率的公司，生产力有实际增长了吗？
Are you really seeing that productivity come through from these firms that are saying it's because of AI adoption?
现在还太早了。
Well, it's really early.
我可以说，目前还没有任何论文能有效回答这个问题。
I'll say that no paper has answered it effectively.
因为你真正需要的，是一个能清楚展示的数据集。
Because what you really want is you want a dataset that shows, hey.
这些公司采用了AI之后，他们的员工数量是怎么变化的。
These firms adopted AI, and then here's how their head counts change.
我可以分享一下我早期的直觉，目前有两件事在同时发生。
I can tell you my ins my my early sense of it, two things are going on.
第一，收入增长和员工人数增长之间的关联正在减弱，尤其是在软件公司，增加员工并不明显带来收入增长。
One, there has been a decoupling from of of revenue growth to head from headcount growth, particularly at software companies where there's not an obvious there's a less less obvious connection between growing your headcount and growing your revenue.
这是其一。
That's one.
但另一方面，那些积极采用人工智能的公司通常增长很快，未来还有很多发展空间。
The second thing, though, that counteracts it is that firms that adopt AI are typically quite fast growing, and they have a lot of opportunity ahead of them.
这部分是因为它们采用了人工智能。
And that is in part because they're adopting AI.
也有一部分是因为这些公司本身就有某些优势。
It's in part because of features of that firm already.
但我的感觉是，那些特别擅长应用人工智能的公司，其实也给员工留下了大量工作要做。
But my instinct is that the firms that are adopting AI particularly well probably have a lot of work for people to do as well.
但这并不意味着裁员这个现象会从世界上彻底消失。
Doesn't mean that we're ever gonna get rid of layoffs completely in the world.
你懂的。
You know?
在某种情况下，总归能找出理由来。
There's always some reasoning for that in some world, in some way.
是的。
Oh, yeah.
但我可不是那种对人工智能持悲观态度的人。
But I'm not I'm not one of those, like, doomers about AI.
我很想看到一些研究，解释为什么那些公司总说AI会摧毁所有工作——我实在想不通这对他们有什么好处。
I would love to see some research about why the model companies keep saying that it's going to destroy all jobs because I don't see why that is helpful to them.
顺便说一句，大多数经济学家并不这么看。
By the way, that's not even the position of most economists.
所以，我更希望看到有人去研究这个问题。
So that's that's where I would like to see research efforts.
那是TAM发的帖子。
Well, it's TAM posting.
我是说，劳动力的总潜在市场有多大？
I mean, what's the TAM of labor?
对吧？
Right?
他们实际上就是在这么做。
Like, that's what they're doing.
他们就是在
They're they're
大家都劝他们别这么干，可结果听起来反而怪吓人的。
they're basically Everyone tells them to stop doing that, and then it sounds weird and scary.
我觉得这正是大家普遍的看法。
I feel like that's what everyone's saying.
但你才是
But but you're the
99.99%的人都觉得这很奇怪、很可怕，但要是那0.01%的人恰恰是出钱和做决策的呢？
99.99% of people think it's weird and scary, but what if the, you know, the 0.01% of people are the people who are funding it and making these decisions?
也许他们才是目标用户。
Maybe they're the they're the audience.
对。
Yeah.
对。
Yeah.
我们根本不是目标用户。
We're we're not the audience.
那些大学毕业生并不是目标用户。
The the college grads aren't the audience.
真正的目标用户是副总裁、银行和掌握资本的人。
It's it's VPs and banks and the controllers of capital.
你们是想说科技行业可能存在某种群体思维吗？
Are you are you guys suggesting that there might be some element of groupthink in the tech sector?
也不一定。
Not necessarily.
我不确定。
I don't no.
不是群体思维。
Not groupthink.
这更像是
It's more
因为我同意。
Because I agree.
不是。
No.
我们是内容创作者。
Mean, we're content creators.
对吧？
Right?
有时候我们会请来一个人，他们聊的话题特别体制化。
Like, sometimes we you'll do you'll have somebody on, and they're talking about something, and, you know, it's just very institutional.
对吧？
Right?
就是说，它其实是针对一小部分人的。
Like, it's for a very small group of people.
可能全世界只有大概一万人真正需要看这类内容，但它却跑到了很多散户面前，然后就是这些散户在评论它。
There's maybe 10,000 people in the world that that piece of content is actually for, but yet it finds its way in front of a bunch of retail investors, and they're the ones who are commenting on it.
我觉得这在很多方面就是这种情况的翻版。
Think that this is in many ways the version of that.
这些人并不是在谈论如何用AI触达大众并为他们做科普，他们其实是在为两万亿甚至十万亿美元的估值找理由。
These people are not They're not talking about AI to reach the masses and explain to them, like they're trying to justify a $2,000,000,000,000 valuation, a $10,000,000,000,000 valuation.
这才是他们真正的目标。
That's what they're shooting towards.
所以，如果你想为这种高估值辩护，那好吧，整个劳动力市场确实是一个规模庞大的潜在市场。
And so if you want if you wanna justify that, well, you know, the entire labor market is a pretty big total addressable market.
没错。
Yep.
不对。
No.
我同意。
Agree.
我们看看他们能不能成功。
We'll see if it works out for them.
是的。
Yeah.
我的意思是，这确实是个问题，而且你已经开始看到一些监管和自我监管的迹象了。
I mean, that's that's a problem, and you are starting to see regulation, self regulation.
今天我看到有消息说，AI内容要加水印，标明是人工智能生成的。
I saw there was something today about AI watermarking, basically, that that content has been created by AI.
所以，现在确实出现了一些自我监管的苗头。
So there is some degree of, like, self regulation happening here.
对。
Yep.
不过我想问问，在Ramp你们还关注哪些其他趋势？特别是在SaaSpocalypse这个背景下，还有哪些软件相关的趋势值得大家留意？
But I guess as far as, like, other trends that you are interested in tracking at Ramp, what what are the other, like, software related trends that that you think people should be paying attention to as it relates to the SaaSpocalypse?
现在很多供应商的增长，并不是来自那些头部公司本身，真正亮眼的增长往往来自别处。
Well, a lot of the growth that's happening amongst vendors is not and a lot of the really compelling growth is not necessarily with the model companies themselves.
所以我劝那些特别担心SaaSpocalypse的人，别太紧张。
And so that's where I caution people who who are extremely concerned about a SaaSpocalypse.
市场肯定会有变化，但SaaS生态里其实有一些被低估的领域，正因为AI而快速增长。
Not to say there won't be transformations in the market, but to say that there are underrated parts of the SaaS market that are growing in large part because of AI.
比如AEO，也就是答案引擎优化，大概就是这个概念。
AEO as a category, so answer engine optimization, I guess.
你知道SEO是让你在谷歌搜索结果里露脸，对吧？
It's you know, SEO is how you show up in Google results.
AEO则是企业用来追踪自己的内容在AI模型中表现如何、有没有被推荐的工具。
AEO is the software that firms use to track their performance in AI models and whether or not they're being recommended.
增长空间巨大。
Huge growth.
这个领域以前根本不存在。
I mean, that sector did not exist.
开发这类软件的公司，和以前卖SEO软件的不是同一拨人。
The companies that are making that software are not the same companies that were selling the SEO software.
SEO公司确实在入场，但AEO领域增长最快的那些公司都是新面孔。
I mean, SEO ones are getting into it, but but the very fast growing ones in AEO are new.
ProFound就是一家新公司，增长速度非常快。
ProFound is a new one, growing extremely quickly.
这并不是模型公司直接提供的产品，而且说实话，模型公司根本没法做这个——Anthropic可以为自家模型做，但不可能有效覆盖所有不同的模型。
So, again, not a product offered directly by the model companies, and, arguably, the model companies could never offer that product because Anthropic can offer it for Anthropic, but they wouldn't be able to offer it across all of the different models in any effective way.
这本质上应该由外部的其他厂商来完成。
I mean, it arguably something that should happen externally by some other vendor.
所以这个领域正在快速增长，但很多人根本注意不到，也不会讨论它，因为它不是两大巨头之间的竞争，但在我的看法里，这恰恰能反驳所谓的SaaSpocalypse。
And so there's a lot of growth in this sector that that people would just ignore or or it doesn't make the discourse because it's not about the jockeying between, you know, two large players, but it certainly counters SaaSpocalypse in my mind.
不过有意思的是，我们之前提到的另一个例子，是跟踪你的任务该交给哪个模型。
Well, but it is interesting that, like, the the other example we talked about was keeping track of what model your tasks should go to.
我们看到增长最明显的领域，其实就是搞清楚AI到底在干嘛。
Like the biggest areas of growth that we all see are sort of just making sense of AI.
对吧？
Right?
弄明白AI生态里到底发生了什么。
Making sense of what's happening in the AI ecosystem.
你看SaaS的增长，有没有哪个领域不是跟AI沾边的？
When you look at the growth in SaaS, is there anywhere where it's not tied to AI that you're seeing growth where there isn't some connection?
这得看你怎么定义什么是AI了。
Well, it depends where you draw the line on what is AI.
对吧？
Right?
比如，用AI功能来升级老软件的产品，我们当然可以说它是AI驱动的，但这跟SaaSpocalypse的说法完全是两码事。
Like, software that uses AI features to improve an existing legacy software product, sure, we can call that AI native, AI driven, but it that's a different narrative from SaaSpocalypse.
对吧？
Right?
比如Adio，这是家总部在伦敦、增长特别快的CRM公司。
Like, Adio is this extremely fast growing London based CRM.
它的市场份额还只有个位数，但增长速度非常快。
It's like a couple single digits of market share, but it's growing really quickly.
Salesforce占了市场80%的份额。
Salesforce is 80% of the market.
你懂的。
You know?
所以它短期内不可能被轻易取代。
So it's not something that's gonna be unseated very quickly.
但Adio确实是原生AI的替代方案。
But Adio is, like, the AI native alternative.
我觉得，不能因为这些公司用了AI功能，就简单地认为它们的增长全靠大模型公司，这种看法是错的。
And, you know, I think it would be a mistake to discount the growth of these kinds of firms just because they have AI features and and and say that, oh, it's it's entirely because of it's entirely tied to the performance of the model companies.
这些产品非常独特，大模型公司目前还没法真正跟它们竞争，或者还没来得及竞争。
These are very different unique products the model companies can't really compete on or haven't yet.
所以你看，AI会渗透到很多正在发生的变化中，但它也会在这些模型公司主营业务之外，开辟出新的市场。
And so, you know, AI is gonna touch a lot of these developments that is that are happening, but it's also gonna grow a market outside of the explicit model company's business segment.
明白了。
Okay.
我很好奇Anthropic最近公布的那些集成功能。
I I I wonder the integrations that Anthropic recently announced.
嗯，对。
Okay.
你可以通过Anthropic访问Morningstar的数据。
You you can access Morningstar via Anthropic.
你也能通过Morningstar调用Anthropic的服务。
You can access and and you can access Anthropic via Morningstar.
保险行业用Verisk，还有伦敦证券交易所、Refinitiv，情况也差不多。
Same with Verisk for an insurance sector or, you know, the the the London Stock Exchange, Refinitiv.
这对软件公司来说是好事还是坏事呢？
Are is that good for the software companies or bad?
展开剩余字幕（还有 184 条）
为什么？
Why?
因为你可以
Well, you can
现在还能通过云连接器和OpenAI获取Ramp的数据。
also access Ramp data now through the Cloud Connector and through OpenAI.
哦，是这样。
Oh, yeah.
说明白点，这些数据和我们网站上发布的是一样的。
To be clear, that is it's the same data that we publish on our website.
数据是经过聚合和匿名处理的，但你可以把它导入进来，自己做分析，或者获得针对你企业的建议，比如：
It's aggregated and anonymized, but it allows you to pull in that data and build your own analysis or get a recommendation for your own business about, hey.
我该买什么？
What should I buy?
和我类似的公司都在买什么？
What are other firms like me buying?
差不多就是那样。
Something like that.
这对像Refinitiv这样的公司有帮助，还是对卖数据的公司更有用？
Is it good for a business like Refinitiv, or is it good for the data selling businesses?
我不清楚。
I don't know.
我的意思是，我觉得他们其实不得不这么做。
I mean, I I think they kind of have to do it.
对吧？
Right?
从我的角度看，Ramp并不卖数据，所有数据都是免费开放的。
I will say from my perspective, I don't we don't Ramp doesn't sell data, so it's all available for free.
所以我们希望更多人来使用，这本来就是我们的业务。
So we want people to use it, and we're in that business.
如果因为Ramp提供的免费数据，加上Anthropic提供的便捷访问方式，导致很多数据聚合商失去市场份额，我会很高兴。
I would be happy if a lot of data aggregators lost market share because of the both the free data that Ramp provides and then the ease of access that Anthropic provides to do data to produce data.
我并不是要贬低所有的数据聚合商，但很多这类业务其实都藏在幕后。
No shade necessarily to all data aggregators, but I think a lot of that market is, unfortunately behind closed doors.
它们的方法论往往不够透明。
The methodologies are not always clear.
总的来说，我觉得信息越开放，企业和消费者受益越多，决策质量也会更高。
And in general, I think information is better when more firms and consumers have access to it, but decision making tends to be better.
那这对卖数据的公司来说是好事吗？
And so is it good for the businesses that sell data?
可能不是，但我觉得对整个市场来说是更有利的。
Probably not, but it it is better for the market overall, I think.
那接下来，亚伦，软件行业的那些巨头呢？
And then, Aaron, what about the large the giants of the software business?
你提到很多软件公司增长非常快，表现也很不错。
You you mentioned many software companies that are growing quite quickly, that are doing quite well.
你说它们增长迅速，让我觉得它们应该是私营企业，可能有风投支持，我之前都没听说过，大概率是私企。
Because you said fast growing, that makes me think that they're private, perhaps venture capital backed, probably, I hadn't heard of them, they probably are private.
那像ServiceNow、微软这样的行业巨头呢？还有那些由私募股权支持的公司，它们阶段更靠后，增长没那么快，但现金流很充沛。
But then what about the large behemoths, the ServiceNows of the world, the Microsofts of the world, and then also there's the private equity backed thing where it's they're even later stage, they're not growing that quickly, but they're producing a lot of cash flow.
你之前提到过那些新兴的软件公司，但那些老牌的软件公司又怎么样呢？
You know, you you've talked about this, you know, growing budding world of, like, kinda new software companies, but what about the more legacy software?
嗯，说到底还是
Well, again, it's
得具体情况具体分析。
gonna be case by case.
很多老牌公司不仅会受到人工智能的冲击，还可能因为市场对它们软件的需求发生变化——要么消失，要么转型，要么暴增，都有可能。
And many of them will be disrupted not only by AI, but simply because of demand for their software, which may dry up or change or skyrocket or whatever.
对吧？
Right?
接下来几年最值得观察的，就是这个趋势。
So, I mean, that that's what's gonna be interesting about the next couple years.
人们很可能会下意识地认为，嘿。
There's gonna be an instinct to say that, hey.
这家公司是因为AI才失败的，但其实也可能是它对客户需求反应太慢，才导致倒闭的。
This company failed because of AI, whereas, you know, it very well may have just failed because it was not responsive to the needs of its customers.
也许是因为它没用AI，但另一方面，也可能是那些创新速度不够快的公司，本来就逃不过自然淘汰的命运。
Maybe that was because it didn't apply AI, but, also, maybe that was just because of the natural churn that ends up happening amongst firms that aren't aren't innovating fast enough.
人们总喜欢对整个软件市场下定论，但AI确实是个颠覆者，而且每个公司都知道，自己必须以某种方式做出回应。
So I think there's a tendency to make proclamations about the entire software market, but AI will certainly be a disruptor, but it's also one that every firm knows it has to respond to in some way.
所以在这方面，大家掌握的信息其实都是一样的。
So we all have access to the same information there.
颠覆可以有各种各样的形式。
So disruption can take a lot of different forms.
比如一家公司超越了它那些老牌竞争对手。
It can take the form of a company overtaking one of their legacy competitors.
也可能表现为一种防御性的并购。
It can also take the form of sort of, like, defensive M and A.
这通常取决于那些老牌企业有多快意识到增长速度的变化。
It tends to depend on, like, how quickly those legacy players are recognizing the speed of growth.
我肯定有不少人来找你们，想拿这些数据评估自己职位的风险有多高。
I'm sure you have people coming to you guys trying to get this data to understand how in danger their positions are.
那些老牌企业，在你们和他们讨论数据的时候，有多担心这些新兴公司？你们觉得未来会看到更多防御性并购，还是老牌企业会被彻底甩在后面？
Are these legacy players, like in the conversations you're having about your data, like, how concerned are they about these upstarts, and and do you think we're gonna see more defensive M and A or sort of legacy players getting left in the dust?
嗯，我倒是觉得。
Well, I look.
我觉得部分原因是，很多老牌企业并不真觉得AI会有那么大的颠覆性。
I think part of it is that a lot of legacy players aren't really sold on AI being that transformative.
也许他们觉得AI终将改变一切，但并不认为自己必须立刻做出强烈回应。
Or maybe they believe it'll be transformed, but they don't necessarily think it's something they need to be extremely responsive to.
其实有个很好的例子，就在一个多月前登过《华尔街日报》。
There's a really good example, actually, because it was in the Wall Street Journal, like, a month or two ago.
德勤赞助了《华尔街日报》上一个叫‘首席财务官期刊’的专栏。
Deloitte sponsors, like, a a note in The Wall Street Journal.
那个专栏叫‘CFO Journal’。
It's called, like, CFO Journal.
所以我看了，因为这上面通常写得都挺不错的。
And so I read it because it's normally just, like, pretty good write ups.
这不是广告。
Like, it's not it's not an ad.
只是由德勤赞助的，讲讲CFO们都在想些什么。
It's just, like, sponsored by Deloitte about what CFOs are thinking about.
标题大概是说，四大会计师事务所都在怎么搞AI。
And so the headline was something about how, like, this is how the big four all the big four accounting firms are doing AI.
好的。
Okay.
德勤赞助这个栏目挺有意思的。
That's an interesting thing for Deloitte to sponsor.
文章里提到，毕马威正在为这项任务部署智能体，普华永道也在做类似的事情，还开展了一系列实验，让员工接触各种工具。
And it goes talks about how KPMG is implementing agents for this task, and p y p PwC is doing this thing, and is doing all these experiments where they're giving their employees access to this and that thing.
而德勤则采取了更为谨慎的策略。
And then it said, Deloitte is actually taking a much more hedged approach.
我们认为，这会提升和增强员工的能力，但不会取代他们。
We think it'll be something that improves our workers and augments them, but not replaces them.
这挺有意思的，因为我本来一直不明白，直到看到德勤的部分，才搞清楚他们为什么在为其他会计公司做广告。
And it was fascinating because I actually wasn't sure until I got to the Deloitte part why they were doing this advertisement for for the other accounting firms.
我这才明白，这是德勤在刻意表明自己并不反对人工智能，只是在应用上更谨慎一些。
And then I realized, oh, this is an explicit effort by Deloitte to position itself as not anti AI, but certainly a little bit more conservative about its implementation of AI.
这跟其他会计公司的立场正好相反。
And it's directly counter how the other accounting firms are positioning themselves.
顺便说一句，很多公司都在宣传自己积极拥抱这项技术，想向客户传递一个信息：
And by the way, how a lot of companies are positioning themselves as firms that are actively adopting it and trying to sell to their clients and let firms their customers know that, hey.
我们已经跟上了这项新技术。
That we're we're sort of on top of this new technology.
我们会提供相应的解决方案，让我们的会计服务变得更高效，诸如此类的说法。
We're gonna offer you something that accounts for it, and, you know, it's gonna make us better accountants at the end of the day, something like that.
而我个人倾向于认为，还是应该谨慎一点。
Now I'm on the side of I generally airs to the side of, hey.
技术采纳是好事，尤其是人工智能方面的。
Technological adoption is good, more from AI.
当然，得聪明地用。
Obviously, do it intelligently.
别随便推出一个没道理的产品，更别让客户的审计出问题，特别是在金融和会计这种领域。
Don't, you know, unleash a product that doesn't make sense or that, you know, messes up your customers' audits, especially in something like finance and accounting.
我对劳埃德的立场感到挺意外的。
But I was surprised about the Lloyd's positioning.
这说明了，不管是不是老牌公司，对这些变革的反应都不一样。
And so I think it goes to show you that not every firm, legacy or not, is responding to these transformations in the same way.
很多公司其实对采用人工智能相当谨慎，尤其是不敢太快推进。
And many firms are actually quite hesitant to to implement AI, at least implement AI at least implementing AI particularly quickly.
如果你一年前问我，你觉得德勤在采用人工智能上会更谨慎还是更激进？
I think if you would have asked me, like, a year ago, do you think Deloitte will be more cautious or or less cautious about implementing AI?
我那时候大概会猜他们会更激进。
I think I probably would have gone toward the less cautious side.
哪些公司 historically 和德勤属于一类，却在AI应用上走在了最前面？
Who are the companies that you would maybe historically group with Deloitte that are actually at the forefront of adoption?
你觉得有哪些公司在AI adoption上出人意料地领先了？
Like, who are the surprise leaders in the adoption race that you're seeing?
这是个好问题。
Oh, it's a good question.
说实话，你提到的这点很关键，很多传统媒体公司其实特别值得关注，它们很可能在这些变革中脱颖而出。
I mean, so what comes to top of mind this is exactly what you're getting at, but a lot of the legacy media companies that are interesting because they're really gonna come out on top of some of the transformations.
你看《纽约时报》这类报纸，它们直接跟AI模型公司谈合作，授权使用内容。
I mean, you look at, like, the New York Times, a lot of newspapers which are doing deals directly with the model companies to sort of license their work.
这是否意味着它们在AI应用上会更占优势，能更有效地把AI融入团队？
Are they does that mean they're gonna be well positioned to be ahead on AI and to adopt them effectively as a workforce?
已经有了一些迹象。
There's some signs.
我跟不少记者聊过，他们随口提到，虽然AI在新闻编辑部里还是挺有争议的，但越来越多的人开始讨论，比如说——
I mean, I talked to a bunch of reporters who who have sort of offhand said that despite AI is quite controversial still in the newsroom, but there's also an increasing amount of talk about, like, hey.
有些写作其实更适合交给AI来做，这样记者就能专心做采访和报道这类工作。
Some writing is actually more appropriately done with an AI model, and reporters can focus on reporting, something like that.
各个行业现在都在做尝试，而且很多地方根本不是你想象中会率先行动的那些地方。
So there's experimentation happening across sectors, and it's not exactly in the places where you'd expect it.
说到这个，确实挺有意思，毕竟新闻界现在到处都是内容被盗用的情况，总有人问：‘免费版在哪？’
Well, on that front, I mean, that's very interesting, because obviously there's tons of, like, IP theft out there in the journalism world of of people like, oh, where's the free version?
发我个免费链接。
Send me the free link.
为什么这个要付费？
Why is this paywalled?
现在有些大公司开始付费获取这些内容，虽然跟所有盗用者都转成全额订阅还差得远，但至少总算能回收一部分了。
And if some of that is now being captured by these large companies that are paying a fee, it's not the same as if all of those people who were stealing the content are paying, you know, a full subscription price, but at least they're capturing some of that.
另外，AI不像人，它不会受限于阅读时间，但人类的注意力却越来越短。
And then also there's the aspect of, like, the AI has not, like, unlimited time to read, but as, like, human beings, like, our attention spans are going the opposite direction.
我觉得AI在这方面并不存在这个问题。
I don't think AI necessarily has that problem.
它能很快读完一篇长文章。
It can consume a large article very quickly.
你觉得这会不会真正提升报道的质量？因为现在不用再跟人抢注意力了，人类是通过某种已经提炼过长文内容的工具来消费信息的。
Do you think it has the potential to actually improve the quality of reporting because it's not being you're not competing in the attention economy for human eyeballs anymore because the humans are consuming it through something that kind of like already parses the long form Yeah.
毕竟人类的注意力持续时间有限。
Thing for human attention spans.
我这么说吧。
I'll say this way.
你看。
Look.
我觉得AI在大多数写作上其实很糟糕。
I think AI is pretty bad at most writing.
写博客或者有观点的文章，它特别差。
I think it's really bad at writing blogs or anything opinionated.
它生成的东西往往……唉，这话题大家已经讨论得够多了。
I think it comes out with with results that are I mean, everyone's people have talked about this extensively.
我一读就知道是AI写的。
Like, I know when I'm reading something that's written by an AI.
不只是这样
It's not just
是这一点。
this.
还有那一点。
It's that.
是哪一点？
It's what?
不是因为那个。
It's not because it's that.
对。
Yeah.
没错。
Yeah.
把心里话直接说出来了。
The quiet part out loud.
是的。
Yeah.
但我得说，它在某些类型的写作上确实不错。
But I will say it's pretty good at certain kinds of writing.
只要是结构特别清晰的，你只要跟它说，嘿。
Anything, like, highly structured if you just tell it like, hey.
你就直接给我列个要点，把发生的事总结一下——其实很多健康零食和笔记应用就是这么干的，它在这类事情上挺在行的。
Just, like, give me straight up bullet points, summarize this thing that happened, and that's essentially what granola and note taking apps do, it's pretty good at that.
它不会掺杂那种奇怪的、说话语气上的波动。
It doesn't add the weird, you know, vicissitudes of how it talks.
它在全面性方面做得相当不错。
It it does a pretty good job being comprehensive.
是的。
Yeah.
是的
Yeah.
用于实际应用时。
For an implementation.
对
Yeah.
还有大量报道工作，比如记者做采访，会接触很多消息源，亲自跑现场。
And a lot of reporting you know, let's say a reporter, like, does their job and they talk to a bunch of sources and they, you know, they go on the ground.
他们真的会面对面跟人交谈。
They literally talk to people in person.
AI机器人永远做不到这一点。
The AI robots will never do that.
它们不会去参加活动，亲自去采访人。
They're not gonna go to an event and interview people.
它们不具备这种能力。
They don't have that ability.
但作为记者，如果你能先做调查、搞清楚来龙去脉，再让AI把你的报道整理成一个简洁的摘要，这倒没什么问题。
But, you know, if as a reporter, you can sort of ask you can do your sort of research process, the actual reporting will work, and then have an AI write the straight summary of your reporting.
这完全说得通。
That's not unreasonable.
我的意思是，它能写一篇人物专访吗？
I mean, can it write a profile of someone?
可能不行。
Maybe not.
它能亲自跑一趟，去跟县档案员打交道，拿到那些机密文件吗？
Can it do the investigative journalism aspect of having to go in person and talk to the county clerk and get some documents that are, you know, that are sealed?
不能。
No.
它做不到这一点。
It can't do that.
但它可以把你的发现整理成文字。
But it can write out what you found.
所以我觉得，有些写作部分由AI来完成是有可能的。
So I think there is a world in which some parts of writing are done by AI models.
我谈的其实不是写作本身，而是人们怎么读它。
I was less talking about the writing itself and more the reading of it.
对吧？
Right?
以前没有AI的时候，真正读完十页深度报道的人本来就不多，然后这些内容通过社交媒体层层传递，等传到大多数人那儿，很多背景已经丢了，事实也悄悄变了。但如果AI来读这些内容，把关键信息统一传达给每个人，那反而会让人懒得深入调查、认真写作，因为到头来——
If you were to say, like, before AI, there's only so many people who were actually reading the 10 page deep dive article, and then it gets, like, telephoned out through social media, and by the time it reaches the majority of people, like, a lot of the context has been lost, the facts have slightly changed, if instead of the writing being consumed by, like, a small number of people and then miscommunicated out, like, it's that the AI is reading and communicating the key details uniformly to everyone, and thus, like, it incentivizes like, you're kind of disincentivized to go deep and do good writing because by the time
哦，有意思。
Oh, interesting.
它直接传到了每个人那里。
It's reaching everybody.
所以我的意思是，AI负责阅读这件事，说不定反而能推动新闻报道质量的提升。
So I'm saying, like, is it it's the fact that AI is doing the reading is maybe going to incentivize, like, higher quality journalism.
这个嘛，我不确定。
I mean, I don't know.
我不确定AI是否真的降低了《纽约时报》这类文章的阅读量，这倒是个值得看看的现象。
I I I I don't know if AI is actually going to it'd be interesting to see if AI has actually lowered the sort of the viewership on, like, a New York Times typical article, for example.
对吧？
Right?
我觉得订阅报纸的人大概还是会继续读报纸，跟以前一样。
Like, I think that the people who subscribe to newspapers are probably gonna keep reading newspapers as they do.
也许会有小幅下降，但我也不确定。
Maybe on the margin, there will be some reduction, but I don't know.
我说的是其他所有人。
I'm I'm talking about the everyone else.
如果那些平时不付费的人现在至少能通过大模型公司支付的费用接触到内容，而不是从推特上三手四手的信息或者假新闻机器人那儿获取新闻，那他们可能会直接去问Claude或者Gemini：这个新闻是真的吗？
Like, if if there's everyone else and they're not being monetized, and now they're at least being monetized via the the big model companies paying something to get that, and instead of, like, like your third degree connection on Twitter being where you get your news or like a fake news bot, like I'm just gonna go to Claude or I'm gonna go to Gemini and say like, this story true?
我的意思是，这已经成了常态。
I mean, you know, at rock.
对吧？
Right?
是的
Yeah.
这是
Is this
真的吗？
true?
但是
But
我其实想问问你关于xAI的事。
I do wanna ask you actually about x AI.
最后一个问题。
Sort of a closing question here.
你看看你那个指数图表，xAI的数值低得有点危险。
You know, going back to your chart of the index, like x AI dangerously low on that graph.
是的
Yeah.
你觉得这个情况会改变吗？
Do you think that is gonna change?
为什么XAI这么低呢？
Why why is XAI so low?
而且很明显，它在SpaceX上市中占了很大一部分。
And obviously, it's such a huge part of the SpaceX IPO.
是的。
Yeah.
如果他不觉得重要，根本不会把这两家公司合并在一起。
Wouldn't have combined the companies if he didn't think it was important.
那到底是什么原因呢？
Like, what's going on there?
XAI的业务增长还没起来，首先，它进入这个市场的时间比较晚。
XAI has not seen the rise in biz I mean, first of XAI was a relatively late entrant into this market.
说实话，它在上线几个月内就达到了2%到3%的使用率，这已经是很了不起的成就了。
I mean, the fact that it even grew to two, I think, 3% adoption within a couple months of launch is a pretty significant feat.
但目前为止，它还没能像Anthropic在过去六到十二个月那样实现垂直增长。
However, it has not been able to translate that into sort of vertical growth that Anthropic saw over the last six, twelve months.
但我觉得这些数据并不是没被报告或者被低估了。
But I don't think that these are unreported or underreported, certainly.
我觉得如果他们收购Cursor的交易最终成行，那我们就得把Cursor的市场份额算进XAI里了。
I think if their acquisition of Cursor goes through, well, actually, then we're gonna have to start integrating Cursor into XAI's market share.
不过我觉得他们当初收购这个公司肯定有充分的理由。
But I think there was a good reason for that acquisition as well.
你看，XAI真正拥有的是超强的算力，这也促成了他们和Anthropic的新合作。
I mean, when you look at what XAI really had available to it, which is this incredible CompuPower, also informed its its their sort of a new agreement with Anthropic.
通过收购Cursor来推动模型的实际应用，对XAI来说会是个非常明智的决定。
Acquiring something like Cursor to actually drive some kind of model adoption, I think will be a really good move for XAI.
明白了。
Alright.
Ara，非常感谢你参与这次对话。
Well, Ara, thank you so much for joining us.
大家可以在Substack上找到我的内容，网址是econlab.substack.com。
Everybody can find your stuff at it looks like it's on Substack, econlab.substack.com.
Substack。
Substack.
我在Twitter上的名字是Ara Kharazian。
It's my name on Twitter, Ara Kharazian.
现在我还在LinkedIn和Instagram上活跃，可以在所有平台找到我。
It's on nowadays, I'm on LinkedIn and Instagram too, so find me on all platforms.
全平台都上线了。
Out now on all plats.
Ara Kharazian。
Ara Kharazian.
非常感谢你参与我们的对话。
Thank you so much for joining us.
感谢收听本集的a16z播客。
Thanks for listening to this episode of the a 16 z podcast.
如果你喜欢这期节目，别忘了点赞、评论、订阅，给我们打分或写评价，也分享给你的朋友和家人。
If you like this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.
想听更多节目，可以去YouTube、Apple Podcast和Spotify查找。
For more episodes, go to YouTube, Apple Podcast, and Spotify.
在X上关注我们，账号是a sixteen z；也可以在a16z.substack.com订阅我们的Substack。
Follow us on x at a sixteen z, and subscribe to our Substack at a16z.substack.com.
再次感谢收听，我们下期节目再见。
Thanks again for listening, and I'll see you in the next episode.
本内容仅供教育参考，不构成任何投资或金融产品的购买、持有或出售建议。
This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product.
本播客由第三方制作，可能包含付费推广广告、其他公司提及以及与a sixteen z无关的个人内容。
This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with a sixteen z.
这些广告、公司和个人均未经Ah Capital Management LLC、a16z或其任何关联方认可。
Such advertisements, companies, and individuals are not endorsed by Ah Capital Management LLC, A16z, or any of its affiliates.
所载信息基于发布时可信赖的来源，但a16z不保证其绝对准确。
Information is from sources deemed reliable on the date of publication, but A16z does not guarantee its accuracy.
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。