Anthropic’s $900B Valuation Beats OpenAI, Claude 4.8 Drops, Former Shopify CTO on AI Risk

欢迎各位收看《信息》旗下的TITV。

Welcome everyone to The Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是5月29日，星期五。

It is Friday, May 29.

在开始今天的节目之前，我要提一下《信息》亚洲团队今天早上发布的独家报道：字节跳动正在研发一种AI芯片，风格可能跟Groq芯片类似，而Nvidia目前正在授权这种技术。

Before we get to today's show, I want to flag some exclusive reporting The Information's Asia Bureau published this morning about ByteDance working on AI chips that could take on a similar flavor to the Groq chips, which Nvidia is licensing.

这篇报道现在已经在我们的网站上发布了。

That is up on our website right now.

建议大家去读一读。

I encourage you to check it out.

今天节目中，我们来深入聊聊Anthropic最新推出的Opus 4.8模型，以及开发者们对它的看法。

Today on the show, we are unpacking Anthropic's latest Opus 4.8 model and what developers are thinking of it.

我们还会分析Anthropic最近这笔巨额融资，它让这家公司的规模已经超过了OpenAI。

We'll also break down Anthropic's big funding round that puts it ahead of OpenAI in terms of size.

这周的《编辑精选》就讲这个话题。

That is the subject of this week's Editor's Cut.

我们还请来了前Shopify首席技术官，聊聊他新加入一家AI法律与合同公司的经历。

And we're also bringing on the former CTO of Shopify to talk about his new role at an AI legal and contracting company.

我们也会聊聊整个电子商务行业的情况。

We'll also talk about the e commerce sector as a whole.

节目内容很精彩，咱们马上开始。

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

Anthropic发布了最新模型Claude Opus 4.8。

Anthropic released its newest model, Claude Opus 4.8.

每次有新模型发布，都有很多值得分析的地方，所以我请来了AI医保经纪平台Chapter的联合创始人兼CEO科比·布鲁门菲尔德·甘茨，听听他团队对这个模型的初步反馈。

Anytime there's a new model released, there is a lot to break down, and so I want to bring on Cobi Blumenfeld Gantz, the co founder and CEO of AI powered Medicare Broker, Chapter, to share with us his team's earliest thoughts on the model.

科比，欢迎来到节目。

Cobi, welcome to the show.

很高兴你能来。

It's great to have you here.

谢谢邀请我来。

Thanks for having me.

你们好像已经用了几天Opus 4.8了，是吧？

So you guys have been playing with Opus 4.8 for a couple days now, sounds like.

初步感觉怎么样？

What's the initial reaction?

还不错。

It's yeah.

用起来挺有意思的。

It's it's been fun to play with.

从用户角度和我们能看到的效果来说，跟4.7差别不大。

It's not that different a model in terms of the user experience and what we can see than 4.7.

我们注意到的一点变化是，它出错的次数减少了。

What we have seen a little bit of a difference on is the decrease in mistakes that it's making.

我们做的工作涉及帮助老年人处理医疗和医保，这种错误减少真的很重要。

In a high consequence environment that we operate in, as helping seniors with their healthcare, with their health insurance, that does actually matter.

但目前来看，这些改进似乎并不太明显。

But the improvements do seem rather marginal at this point.

你们是在技术团队里用这个模型吗？

And are you guys using it where in your techs are?

你们会用它来写代码吗？

Are you using it for coding?

你们会用它帮用户自己找保险计划吗？

Are you using it to help people find plans themselves?

能跟我们说说具体是怎么用的吗？

Just tell us a little bit about the use case.

我们其实会用很多不同的模型来应对各种场景，但针对4.7x和4.8x，我们确实用它来写代码。

We use a bunch of different models for a lot of different use cases, but specifically for 4.7x and 4.8x, we do use it for coding.

实际上，我们团队最近有一部分人转用Kerser的Composer 2.5了，因为成本低很多，而且它是基于开源模型的。

Actually, some of our team recently switched to Kerser's Composer 2.5 because it's just much more cost efficient, which is actually built on an open source model.

总的来说，我们用大模型和Opus来自动化各种流程，处理大量后台工作，这样我们的顾问和老年用户就不用填那么多表单，也能更顺畅地完成端到端的服务流程。

But broadly, we use LLMs and a lot of Opus for automating workflows, for doing a lot of background work so that our advisors and the end user, the senior, don't have to fill out as much work, and really stringing together end to end workflows.

那你们在使用这些模型的时候，会评估它们的效果吗？

Now, do you guys assess the effectiveness of these models when you're using it?

当然有各种基准测试，但更重要的是实际应用中的表现。

There's obviously the benchmarks, but then there's the effectiveness on the ground.

你们公司有没有一套量化的指标来衡量这些模型的成功与否，还是全靠主观感受？

Do you have a quantitative way of measuring the success of these models at your company, or is it just anecdotal?

算是介于两者之间吧。

It's somewhere in between.

我觉得大多数公司其实没必要搞复杂的基准测试，因为现实情况很简单：模型能解决问题就行，不能就淘汰。

I think most companies don't actually need to build robust benchmarking, because the reality is either the model solves the problem or it doesn't.

只要你们有靠谱的检查机制和质量验证，这才是最关键的。

As long as you have good checks and good quality verification, that's actually the most important thing.

模型是否符合某个抽象的基准，对大多数实际应用场景来说，反而是次要的。

Whether it's living up to some arbitrary or abstract benchmark, think is secondary in most operational real use cases.

但我们也发现，随着新模型不断推出，真正关键的其实是上下文。

But we are seeing that as the models come out, what actually matters the most is the context.

关键在于输入的数据。

It's the data that goes into it.

如今大多数模型的推理能力其实已经很不错了，而且还会继续提升。

Most of the models are actually quite good at reasoning today, and there will continue to be improvements.

但真正重要的是，我们已经有了大脑，接下来该怎么给它配眼睛、腿、鼻子、耳朵和听觉，把完整的上下文传递给模型，让它能给出高质量的回应。

But it's really, we have the brain, how do we build the eyes, how do we give it the legs, the nose, the hearing, the ears, so that the full context can be provided to the model so it can actually give good output.

如今，决定模型表现好坏的，更多是这些外部条件，而不是模型本身。

And that's much more determinant of quality of model than the model itself today.

我之前还真没听过把模型比作有腿有耳有眼的说法，不过我很喜欢这个比喻，也完全理解你的意思。

I've not heard yet the analogy of giving the model legs, ears, and eyes, but I like it, and I see your point with it.

我想问问，从你的角度看，它们和OpenAI的模型相比怎么样？

Tell me, how does it compare to OpenAI's models from what you can see?

其实差别没那么大。

Again, it's not so different.

我们经常根据具体场景的效果，来回切换使用不同的模型。

We constantly switch back and forth, frankly, based on what's most effective for a given use case.

我觉得现在的现实是，大多数模型已经达到了一个阶段，未来还会继续改进，但进步都是渐进的，至少我们目前看到的是这样。

Think the reality today is most models have just gotten to a point, and they will continue to improve again, but it's rather incremental, at least that's what we're seeing.

4.8 比最新的 OpenAI 模型稍微好一点，但优势并不明显，如果几个月后 OpenAI 推出一个比 4.8 还强的新模型，我一点都不意外。

So, think 4.8 is a little bit better than the latest OpenAI model, but it's not materially better, and I wouldn't be surprised if in a few months there's a new OpenAI model that is a little bit better than 4.8.

对。

Right.

不过我想再回头聊聊你之前说的关于编程模型的事。

Now, I want to go back to what you were saying earlier though about coding models.

你说你们在用开源模型做编程，而我理解的开源模型通常不是最强的，但成本更低。

You said that you were working with open source models for coding, and my understanding of open source models is they're often not the best models, but they can be more cost effective.

听起来你们确实采取了这种策略，目前在编程方面并不一定要用最顶尖的模型。

So, it sounds like you guys have taken that approach, and you're actually not needing to use the best in class right now for coding.

我们公司并不固执于某一种选择。

We're not dogmatic about it as a company.

我们的想法是，给团队在不同任务上配备最合适工具，因为各个模型在不同场景下的表现各有优劣。

Our view is we want to give our team the best tools for a given purpose, and the different models do have different performance areas in different contexts.

我觉得这取决于你做的代码类型。

So, I think it depends on the type of coding you're doing.

如果你在开发一些相对简单的项目，不需要太多上下文，那一些开源模型反而会非常高效。

If you're building things that are relatively simple, that don't require nearly as much context, then some of the open source models can be very efficient and effective.

因为它们不会把简单的事情搞得太复杂，有时候甚至比最新的模型表现得更好，那些模型反而容易把简单任务过度复杂化。

Because they aren't trying to overcomplicate things, sometimes they're actually even better than some of the latest models, which can overcomplicate some very simple tasks.

但如果你要构建更复杂或更困难的系统，我们通常还是会用最新的模型。

But yes, if you're trying to build a more difficult or more complex system, we generally then do rely on the latest models.

说到底，就是用对的工具做对的事。

So, it's using the right tool for the right purpose.

你觉得自己团队用开源模型的比例，未来会增加吗？

Do you see your use of open source as a proportion of your model usage, do you see that going up in the future?

这不好说。

It's hard to say.

我猜会吧。

My guess is yes.

目前我们用的开源模型占比还是很小。

It's still open source is still a very small percent of what we use.

我们绝大多数还是在用 OpenAI、Anthropic，偶尔用 Gemini，但主要是前两者。

We still are vast majority on OpenAI, Anthropic, and sometimes Gemini, but more and more OpenAI and Anthropic.

我们刚开始用一些越来越好的开源模型，不过如果它们以后变得更普遍，我也不觉得奇怪。

We are just starting to use some of the open source models as they get better and better, but I wouldn't be surprised if that becomes a more common trend.

我问这个是因为前几天我跟一位高管喝咖啡，聊到企业可能会更倾向于用旧模型，毕竟大家用得更熟了。

The reason I'm asking because I was having coffee with an executive the other day, we were talking about the extent to which enterprises may in fact see more use out of using the older models because of the fact that there's just more practice with them.

你更清楚怎么把它们整合进系统。

You have a better sense for how to integrate them.

有些情况下，它们也更便宜。

In some cases, they're cheaper too.

这让我想起芯片那回事，对吧？

Kind of reminds me of the chip discussion, right?

我的意思是，你未必非得用最顶尖的那款。

I mean, you might not need the Vera Rubin.

可能一块Blackwell芯片就能满足你的所有需求了。

You might be getting everything you need out of a Blackwell.

我是说，你觉得这会是未来的趋势吗？

I mean, do you see that being the future?

随着这些前沿模型不断推出，我们真的会看到企业更多地采用一年前或一年半前发布的模型吗？毕竟它们已经足够好用了。

While all these frontier models are coming out, are we actually going to see enterprises more adopt models that came out maybe a year ago, eighteen months ago, because they're good enough?

我觉得这种可能性很高。特别是因为Anthropic似乎借鉴了苹果的套路，在新模型发布前，故意降低旧模型的质量，这样新模型的效果就显得更出色了。

I think there's a high likelihood of that, in particular, because it seems like Anthropic has taken a page out of the Apple playbook, where they have Well, they're still delivered or not, I don't know, but they have decreased the quality of the prior models right before they launch the new model, so that the new model feels better.

你这话是

What do

什么意思？

mean by that?

能再详细说说吗？

Explain that a little bit more.

从三月到四月，Anthropic 后来承认，他们的 Claude 4.7 模型当时遇到了算力问题。

From March to April, Anthropic said that they were they later claimed that they were acknowledged that they were having a compute issue on 4.7.

所以他们试图让人觉得4.7表现得非常好。

And so they had tried to make it seem like 4.7 was doing very well.

但实际上他们遇到了计算资源的问题。

They were actually having compute issues.

他们把更多计算资源转向了4.8。

They've moved more compute to 4.8.

这让4.7显得慢了很多，体验也差了很多。

It makes 4.7 feel a lot slower and feel a lot worse.

事实上，我们团队在三月和四月就感受到了。

In fact, our team felt it in March and April.

这也是我们开始关注开源模型的原因，因为Anthropic的云模型表现不如以前了。

That's part of why we started looking at open source models, because the cloud models, the Anthropic models, were were not performing as well.

苹果以通过新软件让旧硬件显得过时而闻名，以此促使用户购买新设备。

And so they are you know, Apple is famous for making its older hardware seem much older with new software upgrades so that they incentivize you to buy the new hardware.

Anthropic似乎也在用类似的方式推广新模型。

It seems like Anthropic is doing something similar with the new models.

所以，当他们试图把用户推向最新款模型时——这些模型可能会越来越贵——我不觉得奇怪，有些公司会转向开源模型，或者回到那些没被人为压低性能的旧模型。

And so, as they try to push users to the newest models, which are probably going to be more and more expensive, I wouldn't be surprised if companies both go to more open source models and go to older models that aren't being arbitrarily depressed.

我以前没听过这种说法，居然觉得苹果的这套做法在这里也适用。

I have not heard that before, and this idea that the Apple playbook could be relevant in models here.

那你觉得这会对最终的赢家产生什么影响呢？

I wonder what implication you think that has then on who ultimately wins.

我问这个问题，是因为我们刚看到新闻，说Anthropic目前的估值已经超过了OpenAI。

And I ask this question because we saw the news that Anthropic is now valued higher than OpenAI for the moment at least.

我们还不知道这些事最后会怎么发展。

We don't know how these things are going to shake out.

但话说回来，在这个公司必须不断自我创新、同时又面临激烈竞争的世界里，你觉得这会对最终的赢家产生什么影响？

But I mean, in this world where companies need to keep innovating themselves, and also they're facing so much competition, wonder, I mean, how do you think it affects who ultimately wins?

我觉得这就像看着四季更替、花朵绽放一样。

I think it's kind of like watching four seasons and flowers bloom.

现在我们正处于Anthropic的季节，它的花朵正在盛开。

We're certainly in the Anthropic season, and the Anthropic flower is blooming.

九个月前，一切都围着OpenAI转，那时候是它的时代。

Rewind nine months ago, everything was OpenAI, and that was sort of season.

如果三个月后我们进入谷歌时代或者XAI时代，我一点都不会意外，等它们有了突破。

I wouldn't be surprised if in three months we're in the Google season or the XAI season when those companies have an advancement.

但说到底，它们都有大把资金。

But really, they all have so much money.

也都拥有顶尖的人才。

They all have strong talent.

没错。

Sure.

在专注方向和执行细节上，确实有些微小差别。

There are there are differences in maybe focus and execution on the margin.

但现阶段，我觉得未来两年内，四大巨头之间不会出现实质性的差距。

But at this point, I really don't think there's going to be a material difference between the four largest players, at least over the next two years, call it.

这反而造就了一个特别有意思的竞争格局，对大语言模型的使用者来说绝对是好事，因为产品会越来越强，价格也越来越低。

And I think that that just creates a really interesting competitive dynamic, which is probably very good for consumers of LLMs because they'll get better and better and cheaper and cheaper.

我不太清楚这对那四家公司意味着什么。

I don't know what that means for those four companies.

我觉得市场需求很大，所以它们都会过得不错。

Think there's a lot of demand out there, so they'll all be fine.

这个市场足够大，能容得下所有人，但我认为不会只有一个赢家。

It's a plenty big market for all of them, but I don't think there will be just one winner.

没错。

Right.

我最后问你一个问题。

Let me ask you one last question.

你的业务是帮人们找到最适合他们的医疗保险计划，对吧？

So, business, I mean, you are in the business of helping people find the best health insurance plan that they can.

你在工作中肯定经常跟这些保险公司打交道。

You must talk to a lot of these health insurers frequently throughout the course of your business.

这些保险公司愿意用最新最先进的模型吗？

Are health insurance companies, are they willing to use the latest and greatest models?

他们更倾向于用那些虽然老一点但更安全的模型吗？

Do they prefer the older models that sometimes can be safer?

我们知道，这些模型可能会有被黑客攻击的风险。

We know that there could be risks that the models could be susceptible to hacks.

他们对此怎么看？

What's their view on it?

大多数健康保险公司跟其他大型风险规避型机构差不多，不管是政府还是金融机构，都一样。

Most health insurance companies are pretty similar to most large risk averse enterprises, whether it's governments or financial institutions, you name it.

它们的运作方式都挺相似的，现在还处在很初级的阶段，正在摸索怎么用大语言模型。

They all operate fairly similarly, in that they're still very much in the moving phase, early stages of how do we move to even using LLMs.

对他们来说，一部分原因是害怕风险，不想用最新模型，但更大的原因是，他们还在努力推动组织适应生成式AI的使用。

For them, certainly a part of it is risk aversion, that they don't want to use the latest model, but a large part of it is they're still trying to manage that change to get their organizations to be comfortable using generative AI at all.

因为他们本身转型速度比较慢，跟科技公司和初创企业比起来差得远，所以自然会用更老的模型，一方面是出于安全考虑，另一方面就是他们就是转得慢。

Because they move relatively slowly, especially relative to the way tech companies and startups move, they are naturally going to be on much older models, partly for security reasons and partly just because they don't move that fast.

对，你说得一点都没错。

So, yes, I think what you're saying is spot on.

太好了。

Great.

科比，非常感谢你今天来做客。

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

这位是科比·布鲁门费尔德-甘茨，Chapter公司的联合创始人兼CEO，出现在TITV上。

That is Cobi Blumenfeld-Gantz, the co founder and CEO of Chapter here on TITV.

正如我们之前提到的，Anthropic在最近一轮融资中的估值已经超过了OpenAI。

Anthropic, like we said, has marched ahead of OpenAI in terms of valuation with its latest funding round.

Anthropic在融资前的估值达到了9000亿美元，本轮融资筹集了650亿美元。

Anthropic raised $65,000,000,000 at a $900,000,000,000 valuation before the investment.

关于这场竞争的更多详情，我们接下来邀请劳拉·曼达罗和杰森·迪恩，带来本期《编辑精选》的访谈。

For more on that rivalry, I want to bring on Laura Mandaro and Jason Dean for this week's edition of The Editor's Cut.

劳拉、杰森，欢迎来到节目。

Laura and Jason, welcome to the show.

很高兴你们能来。

It's great to have you here.

很高兴

Glad to

能来

be Okay.

到这里

Here.

我们刚才聊了 Opus 4.8，也就是最新的模型。

So we just talked about Opus 4.8, the latest model.

Laura，我想请你帮我们分析一下Anthropic这次估值的情况。

Laura, I want to come to you to help us unpack this valuation from Anthropic.

我的意思是，我们上一位嘉宾提到过，这跟当前的市场周期有关。

I mean, our last guest was saying, I mean, this is just the season we're in.

这确实要看你融资的具体时机，但到底是怎么回事呢？

It kind of depends when you fundraise, but what?

所以Anthropic现在是领先的一方，对吧？

Anthropic is winning now, I guess?

不过话说回来，OpenAI 在融资方面一直很强劲，他们的估值也不低。

Well, I mean, OpenAI still has been a powerhouse in fundraising, and their valuation isn't that shabby either.

我觉得今年发生的事是，Claude Code 和那个虚拟助手真的火了，投资者和潜在投资者都看在眼里。

I think what happened this year is that Claude Code and Cowork, the virtual assistant, really took off and investors or potential investors saw that.

我们和其他人之前就写过，他们的估值其实非常低，尤其是跟 OpenAI 比起来。

And we and others were writing that their valuation looked very, very cheap, particularly compared to OpenAI.

投资者争相入场，这其实一点都不让人意外。

And it's kind of was probably the least surprising thing that happened that investors try to invest.

在这一轮融资启动之前，距离他们上一轮融资才几个月，当时大家吵得厉害，都在讨论投资者能不能买员工手里的二手股份。

I mean, I think before this funding round kicked off, and it was only a couple months after their last fundraise, there was a lot of hullabaloo about whether investors could invest in secondary shares or buy shares from employees.

那会儿市场需求特别大。

There was a lot of demand there.

你遇到的情况是，Anthropic 本身好像没打算现在融资，但投资人的兴趣却异常高涨。

And you had one of these situations where it didn't seem like Anthropic was planning a fundraise now, but there was intense investor interest.

与此同时，公司高管自己都说，收入增长来得让他们措手不及，说实话，有点出乎意料。

And at the same time, this revenue growth, which the company's executives had said took them by surprise and seemed a little bit disruptive, to be honest.

你知道吗，阿达里奥在台上说，这种指数级增长固然好，但要是能有个更可预测的节奏，可能更好一些。

Know, Adario was saying on stage that, you know, this exponential growth is great, but maybe, you know, it'd be good to have something that they could, you know, predict a little bit more readily.

这意味着他们需要更多算力，不然系统会变慢，或者不得不更严格地限制用户使用，用户体验就会变差。

That meant that they needed more compute capacity because without it, things are slow or they have to rate limit customers more, and it tends to be a bad customer experience.

而OpenAI正乐意接手这部分需求。

And OpenAI is ready to pick up that slack very happily.

没错。

Right.

杰森，我想听听你的看法。

So Jason, I want to get your thoughts here.

另外，我想问你一个问题，当我跟别人聊起这个话题时——你看，大家都知道Anthropic，对吧？

And also, want to ask you this question in the context of when I talk to people I mean, look, we know Anthropic, okay?

商界对Anthropic当然不陌生。

The business community knows Anthropic.

但你要是问普通人，他们用的是Claude还是ChatGPT，我觉得大多数人还是会说ChatGPT。

But still, ask people if they're using Claude or ChatGPT, I think most people will tell you ChatGPT.

所以这里有个挺有意思的现象，普通人用的和公司的估值之间存在脱节。

And so there is sort of like an interesting disconnect here between what the average person uses and then the valuation of the company.

对，确实如此。

Yeah, I think that's true.

说实话，OpenAI现在面临的难题是，他们还没想清楚怎么把品牌知名度转化成收入，而Groq和Anthropic已经摸清了如何在职场场景、尤其是编程领域变现，OpenAI现在正在追赶。

I mean, really the challenge for OpenAI on that front is that they are still kind of figuring out how to monetize that brand recognition, whereas Groq, Anthropic has figured out how to monetize its prowess in workplace applications, particularly coding, And that's where OpenAI is playing catch up.

但我完全同意Laura的看法，这是一场长跑，目前OpenAI确实落在后面了。

But I totally agree with Laura that, you know, this this is a long game and and OpenAI is, the laggard at the moment.

不过我不觉得这种落后会一直持续下去。

But, I I I don't think that that is that predicts that it will be forever.

而且，Dario他们和Anthropic没料到增长会这么快，这带来了不少后果。

And, you know, the fact that that Dario and company and Anthropic were caught off guard by the the pace of their growth has consequences.

我们不知道他们签的这些算力合同具体价格，但可以合理推测，这些合同肯定不便宜，毕竟他们当时是出于急迫才签的，就是为了追上增长的步伐。

We don't know the pricing of these compute deals that they've been signing, but it stands to reason that they're pretty expensive given, you know, that they were done out of desperation to to try to catch up with their growth.

而OpenAI不仅融了更多钱，还更高效地花出去了——他们早早锁定了更多算力，那时候他们还没那么着急，反而更有主动权。

Whereas OpenAI has been not only raising more money, but spending it, you know, more quickly in a good way that, you know, they've secured compute more compute earlier in a moment when they weren't necessarily as desperate to do so.

还有，阿卡什，有一点要注意，现在大家的关注点和噪音都集中在商业应用上，尤其是我们跟大型机构投资者交流时，他们特别看好Anthropic，觉得这家公司目标明确，一直稳稳地兑现承诺。

And, Akash, I mean, one thing to remember, there's a lot of noise and focus on the business applications, you know, and and that's certainly when we talk to big institutional investors, they really like Anthropic because it feels like they know what they're getting, a company that has been focused and is consistently delivered on that.

但每个商业周期都会经历一个轮换，有时偏爱专注的垂直企业，有时又更看重业务多元化的公司，对吧？

But there is, in all business cycles, you swing from valuing the very focused enterprise to valuing the company that has a lot of different business lines for diversification, right?

所以OpenAI一直走的是多元化路线，到现在还是这样。

And so OpenAI had been following the more diversified model and still is.

我们的同事安刚写了一篇挺详细的报道，讲的是广告业务。

Our colleague Ann just wrote a pretty detailed story about the ad business.

看来他们确实在持续推动这项业务，就像十年前Facebook那些大媒体公司做的那样。

And so it does seem like they're continuing to just execute on that and try to develop it in just the same way we've seen saw the big media companies a decade ago like Facebook do.

所以我觉得，单纯认为OpenAI的未来就是变成另一个Anthropic，这种想法太片面了。

So I wouldn't I mean, I think it would be unwise to just think, okay, the whole OpenAI's big future is sort of turning into an Anthropic type company.

我觉得他们走的是另一条路。

I think they laid out a different path.

不过情况有点模糊，因为他们频繁调整组织架构，不同业务负责人对这块的表述也不太一致，说法总在变。

It is a little bit muddy because they seem to you know, they reorg a lot and they have different business leaders leading this function and they talk about it in a sort of a changing way.

但他们本质上是一家同时拥有消费业务和企业业务的公司。

But they have basically a company that has a, you know, a consumer business and an enterprise business.

在某些时候，投资者可能会更看重这种多元化模式，而不是纯粹专注企业市场的公司。

And at a certain point, investors may value that even more highly than a very focused enterprise business.

没错。

Right.

我只是想说

Was just

我同意劳拉的看法。

going to say I agree with Laura on that.

我认为这对OpenAI来说既是优势也是挑战，因为这种模式带来的执行难度，在某些方面甚至比Anthropic还要大。

I think that is a blessing and a curse for OpenAI because the it also creates an execution challenge that is, in some ways much greater than Anthropix.

很少有大型科技公司能同时成功打造消费和企业两大业务。

There are very few companies, big tech companies that have successfully built both enterprise and consumer businesses.

而且，我也不确定，那些做到的人是不是真的同时起步的。

And, I'm not sure that any of those that have done so have done built them simultaneously.

你明白吧？

You know?

你看亚马逊，先做起了消费业务，后来才推出AWS进军企业市场；谷歌也是走的同样的路。

If you look at Amazon, they built consumer business first, and then the enterprise business with AWS, Google, followed the same pattern.

所以要是想同时兼顾两者，还把重心从原本的消费端转向企业端，再可能回头拉回消费端，这简直是在考验他们管理团队的极限。

So to try to do both at the same time and sort of shift the focus from, consumer historically now to enterprise perhaps back to consumer, you know, is is is going to to test their strength as a as a management organization.

我还想补充一点，消费端的竞争远不止两家在打。

And I would just also add that that the additional challenge on the consumer side, this is not a two horse race.

谷歌在消费端AI领域实力很强。

Google is very much a player in consumer AI.

Meta也不容小觑。

Meta is strong.

我们始终都应该在努力。

We should always be is trying.

没错。

Yeah.

至少他们在努力，想在企业市场打响名号，不过企业市场是另一回事，关键是他们现在想让消费者为消费级AI买单。

They're they're trying their best at least to to to make a name in I mean, enterprise is a whole separate question, but I mean, they're trying to get people to pay now for the consumer AI.

没错。

That's right.

杰森，你怎么看这个？

What did you think of that, Jason?

意思是，他们正在测试消费级AI产品。

Mean, saw the so they're testing the consumer AI products.

这算个试点项目。

It's a pilot program.

说实话，我看到他们发布了Spark模型，大家的反应是：好吧，至少你们给了点东西。

My reaction to it honestly was that I know they released the Spark models and people were kind of they kind of said, Okay, well, you gave us something.

总算有点进展了。

Thank God.

但他们现在要用户付费，我真不确定大家是否觉得他们的AI真到了值得掏钱的前沿水平。

But they're asking people to pay, and I'm not really sure that there's a consensus that their AI is really the cutting edge thing that people want to pay for.

你觉得呢？

What do you think?

我同意。

I I agree.

我觉得这是个大问题。

I think that's a huge question.

Meta花了很长时间培养用户，让他们习惯免费使用他们的各种应用。

People are very Meta has trained people over a very long period of time now, to use their their various apps for free.

要想让用户改变这个习惯，你必须让订阅服务提供真正出色的新功能，因为这里的定价并不便宜。

And, to break people out of that, you're either gonna have to provide an additional functionality with these subscription services that is really, really great because the cheap the price points here are not cheap.

或者，你得逐步淘汰现有的服务，把功能迁移到订阅模式里，但这可能会惹恼你的用户群，而广告商正是冲着这些用户来的。

Or you're gonna have to deprecate somehow the existing services and sort of shift functionality into these subscription models, which, you know, risks alienating your user base that the advertisers are all targeting.

过去有些公司就曾在这上面栽过跟头。

And companies have struggled with that sometimes in the past.

所以我觉得，他们花这么长时间才走到今天这一步，开始推订阅服务，也不是没道理的。

So I don't think that's, you know, they've got it's taken them a long time to get to the point where they're beginning to launch, subscription.

要让这个变成他们的主要收入来源，可能还得花上很长一段时间。

It could well take a long time before that turns into, you know, a major a major revenue source for them.

但这表明他们既看到了机会，也感受到必须为巨额计算投入找到合理回报的压力。

But it's indicative of the fact that they feel both the opportunity, but also the need to justify the enormous amount of money that they are spending on on their own compute.

所以劳拉，一年后，你觉得Meta在云计算上的机会会怎样？

So Laura, I mean, a year from now, do we think I mean, the cloud compute opportunity for Meta a year from now, what do you think?

一年后，这会是他们的重点吗？

Is it a focus for them a year from now?

我本来想先说点别的，等会儿再回答你这个问题。

Well, I was going respond to something else, and then I'll respond to that.

关于Meta的消费者订阅服务，有一点是，OpenAI也在做类似尝试。

I think one thing on the Meta consumer subscriptions is that OpenAI is trying this too.

Meta肯定注意到了，对吧？

I'm sure Meta is seeing this, Right?

广告支持的低价订阅，八美元、五美元这种。

Ad supported low price subscriptions, dollars 8, $5.

既然OpenAI都要试，那为什么消费公司不试试呢？

So, I mean, why wouldn't a consumer company try that if OpenAI is gonna try that?

Meta偶尔也会推消费订阅服务，而且他们用户基数这么大，哪怕只有很小一部分人买单，也能带来可观的收入。

And you know, Meta does try consumer subscriptions from time to time, and they're so big that they can generate meaningful revenue, even if a small percentage of their customers do that.

所以这事儿可能没看起来那么不可能。接下来六个月挺值得盯着看的——要是现在值20美元的聊天机器人订阅，跟奈飞差不多价位，结果变成大众普遍接受的支出，尤其是那种广告支持的低价版，那会怎么样。

So it may not be as unlikely as it would seem, but I think there'll be an interesting thing to watch in the next six months if, you know, what we've thought of as sort of a $20 consumer chatbot subscription, which is, you know, on par right now with Netflix, if that becomes a widespread consumer, you know, cost, but but mainly on the cheaper version, right, like the ad supported version.

我不确定Meta能不能在企业市场成功站稳脚跟。

Don't know if I can predict if Meta's gonna be successful becoming more of an enterprise player.

他们以前就试过类似的事。

I mean, that is something that they've tried before.

他们就是这种风格，你知道的。

They just seem to, you know, this is their culture.

他们总在不停地尝试各种新东西。

They try a lot of things.

这种黑客思维是扎克伯格一直推崇的，你能在他们一个又一个业务周期里看到，也正因如此，跟踪他们才特别有意思。

This is the hacker mentality that Zuckerberg has always espoused, and you just see it in business cycle after business cycle, which makes covering it fun.

我的意思是，要是哪家公司啥都不试，那才真没劲呢。

I mean, that's you know, you'd you'd hate to cover a company that didn't try anything.

没错。

Right.

杰森，你最后说一句吧。

Jason, last thought from you.

关于他们可能转型成新一代云服务这个想法，我觉得大家一听马克这么讲就挺兴奋的。

On this idea of them sort of turning into a neo cloud possibly, I mean, I think people get excited when Mark talks about it.

如果你仔细分析他说话的方式，会发现这已经是他第二次被问到这个问题了。

If you look if you break down the way he's talking about it, he's gotten asked this twice.

他 basically 回应说，这事还在考虑中——不这么说才奇怪呢。

And he basically said is like, it's on the table, which it would kind of be stupid to say otherwise.

但他也说了，目前我们没这么做，因为我们觉得现有的算力已经够用了。

But he also says we are not doing that now because we think we have enough use for our compute.

实际上，他们显然觉得自己还需要更多算力。

In fact, they think they need a lot more, obviously.

如果我们发现算力过剩、用不完，那出租就成为一个选择。

And if we get to the point where we think we've overbuilt and we can't use it, then that would be an option.

从某种角度说，这其实是个看跌信号。

So in some ways, it would be a bearish sign.

如果他们能把多余算力租出去，确实能带来新收入，但我觉得，这反而说明他们在AI未来的布局上处境不妙——毕竟他们砸了这么多钱，原本是想用这些算力支撑自己的产品，不管是消费级还是企业级的。

They would get new revenue if they could rent out their their compute, but it would be a pretty bearish sign for their position, I think, in the future of AI if they have spent all this money in the hopes that they can they can leverage this massive compute for their own products, whether they be consumer or some some additional enterprise offerings.

结果却发现根本用不上，只能把资源租给别人。

And then they find out that they can't, they have to rent the space out.

我不确定，到时候投资者会不会真觉得这是个好事。

I I I'm not sure investors will actually see that as positively if and when that time comes as as something now.

但这也能带来更多的收入。

But could mean more money or revenue.

说到底，这终究是他们可能走的一条路。

I mean, you know, at the end of the day, that's that's where they could go with it.

总之，大家说得都很到位。

Anyway, great points all around.

非常感谢你们两位今天来做客。

I want to thank you both for coming on.

这位是来自我们在旧金山分局的劳拉·曼达罗和杰森·迪恩，欢迎来到TITV。

That is Laura Mandaro and Jason Dean from our San Francisco bureau here on TITV.

Shopify的前首席技术官让·米歇尔·勒穆瓦克斯，现在新任AI合同与法律公司Spellbook的顾问。

Shopify's former CTO, Jean Michel Lemieux, has a new position advising AI contracts and legal company Spellbook.

我想请他来聊聊他的新角色——作为高级独立贡献者，以及他对当前AI技术一些问题的看法。

I want to bring him on to talk about his new role as executive individual contributor and how he's thinking about some of the technical topics in AI today.

让·米歇尔，欢迎来到节目。

Jean Michel, welcome to the show.

很高兴你能来。

It's great to have you here.

阿卡什，很高兴能参与这次对话。

Akash, great to be here.

能说说Spellbook是做什么的吗？还有你作为高级独立贡献者，具体会承担哪些工作？

So tell me about what Spellbook does, and what exactly is your role as executive individual contributor going to be?

Spellbook 本质上是在运行全球经济。

So Spellbook basically, in some ways, runs the economy.

世界上我们想完成的每一件事，都得通过合同来推进。

Everything we want to get done around the world has to run through contracts.

可以说，整个世界的运转效率，瓶颈就在合同上。

In some way, can say the throughput of the planet is bottlenecked by contracts.

作为工程师，你会意识到，必须有一层 API，而合同正是驱动这层 API 的核心。

And as an engineer, you realize that there's an API layer that has to be in place and contracts power that.

我早就以顾问和投资人的身份支持 Spellbook 了。

So I've been an advisor and investor in Spellbook for a while.

Spellbook 一直走在前面。

And Spellbook had a lead.

他们从 2022 年就开始做这件事了。

They've been doing this since 2022.

全球有四千五百多家公司，分布在八十多个国家，都在用它。

4,500 companies around the world, in 80 countries I think use it.

所以我看到了这里的潜力，我觉得这个世界需要我们。

So I see the potential, I think the world needs us.

有趣的是，作为顾问和投资人，我一直参与其中，帮着CEO斯科特做事。

Now what's interesting is as an advisor and an investor, I've been getting involved and helping Scott, who's the CEO in the company.

后来斯科特问我：你愿不愿意来直接帮我们？

And at some point Scott said, Do you want to come help us?

当时公司发展得不错。

Things are going well.

他说：我希望你能担任CTO。

And he's like, I'd love for you to have the role of CTO.

我回绝了。

And I said, No.

我说：我觉得你真正需要的不是这个角色。

I said, I don't think that's the right role you need.

原因不是我想摆架子，我是真心觉得，现在我看到的全球情况，还有我离开Shopify之后亲身经历的——我从零开始创办了一家公司。

And the reason why, I wasn't just trying to be an ass, I was really just trying to say, Listen, I think what I'm seeing around the world right now and what I experience So post Shopify, actually, I started a startup from scratch.

说实话，我想重新开始锻炼，就像你说的那样。

Know, I wanted to build, I wanted to go back to the gym, as you'd say.

我跟Scott说，我觉得我们未来取胜的关键，在于公司怎么构建。

I told Scott, I said, I think the way we're going to win now in the future is how the company is going to be shaped.

你懂的。

You know?

全世界都在发生这样的变化。

And you see it around the world.

别人可以抄你的代码，也能模仿出类似的东西。

People can copy your code and they can vibe code something.

我觉得，我们现在真正的核心资产，是公司未来的组织方式。

I'm like, I think our IP now is going to be how we're going to shape this company for the future.

你所说的‘构建’具体是指什么？怎么个不同法？

And what do you mean by I mean, shape so how is that going be different?

我不能把所有机密都告诉你，但我觉得，现在组织的运转效率，和产品本身一样重要，而这正是我想专注的方向。

Well, I can't give you all our secrets, but I think the idea now is the throughput of the organization is as important as the product, and that throughput is what I wanted to work on.

我给你举个小例子，好吧？

And I'll give you a small example, right?

如果你回顾一下历史上技术变革的时刻，比如电力的出现。

So if you look at history of when we've had technology shifts, electricity was introduced.

一百二十年前电力刚被引入，那时候你有个工厂。

So electricity is introduced one hundred and twenty years ago, and you've got a factory.

你第一件事就是说，咱们先把灯打开，让电来照明吧。

So the first thing you do is you say, Let's turn on the lights and let's have some electricity in the lights.

大家会觉得，这还挺酷的。

They're like, That's kind of cool.

挺有意思的。

That's interesting.

如果你去问工厂里的每个员工，你们的生活有什么不同吗？

Now, if you went to every employee in that factory and you asked them, Is your life different?

他们会说，嗯，至少现在亮堂多了。

They'd go, Well, it's a bit brighter.

挺有意思的，但他们不会说生活提升了五倍。

It's kind of cool, but they wouldn't say I've got a 5X improvement.

接下来工厂里发生的事是，嘿，我们可以让机器运转起来了。

Then the second thing that happened in the factories is that, Man, we can turn our machines.

那时候，每个工厂通常只有一台机器。

And at the time, every factory only had one machine.

咱们把电用在这台机器上吧。

Let's put electricity into that machine.

这样一来，产量提高了，成本也降下来了。

And they got some throughput, some cost savings.

真正带来改变的是第三阶段。

And the third phase is where things actually changed.

人们开始说，把电用到工厂里的每一台机器上。

People said, Let's put electricity in every machine in the factory.

正是这时候，工厂重新布局，效率提升了十倍。

And that's where the factory rearranged and I get a 10x improvement.

我觉得，我跟斯科特还有那本魔法书打的赌，现在正是企业们即将要做的事。

And I think the bet that I was making with Scott and the spell book is I think that's what companies are about to do right now.

我们还没到第三阶段。

We're not in phase three.

我觉得执行官的角色，说实话，我是想进来帮一把。

And I think the exec IC role was, you know, I think I want to come in here and help.

作为CTO，我当然能帮上工程方面的事，但我更想参与的是……

And as CTO, I could help with engineering, but I wanted to help

不过你在这个角色里，可能更多时间要花在管理团队、指导别人上，就像你说的，现在你得抽空去健身房了。

with I mean, you might be more busy managing people and coaching people in that role too, which, this you day and age, like you said, you got to get back to the gym.

我想听听你对这周几个热门话题的看法，特别是在你技术岗位上的观察。

I want to get your take on a couple of the trending topics this week, certainly in your technical role.

4.8是最新发布的模型。

4.8, it's the latest model out.

你用过吗？

Have you used it at all?

那你觉得，人们到底是需要这些前沿模型，还是用得上开源模型或者一年前发布的模型就够了？

And what's your view on whether or not people are going to actually need these frontier models versus actually making use of open source models and models that came out a year ago?

我觉得对公司来说，第一要紧的是得做好随时切换的准备。

Well, I think the first thing for companies is you have to be ready to move.

也就是说，一出新版本，我们就能立刻用上。

I mean, we can run on it the minute it comes out.

当然，我们得先做些测试。

Obviously, we have to do some testing.

得确保别让性能下滑。

We want to make sure we don't degrade.

新模型里面复杂的地方太多了。

There's a lot of complexity in a new model.

所以作为公司，你得准备好去尝试这些东西。

So I think as a company, you have to be ready to be able to try these things.

最好你自己有评估标准，能自己去测试验证。

And hopefully you have your own benchmarks and ability to try it.

我觉得 Spellbook 的经验表明，这些模型确实非常强大且实用，但我们之前也处理过上千万份合同。

I think what we've seen with Spellbook is the models are obviously really powerful and useful, but we also have 10 plus million contracts that we've seen before.

所以，我们拥有的数据对提升产品品质帮助很大。

So the data that we have, I think, is really helpful in terms of improving the product quality.

对我们来说，竞争优势一方面是能迅速跟进新模型，另一方面也要清楚自己能提供什么。

So for us, I think in some ways, competitive advantage is be ready to jump on the models, know what you're bringing to the table as well.

我觉得这两者结合得非常好。

And I think that combination is great.

换句话说，当像 4.8 这样的新模型发布时，我理解你的意思是，你并不会立刻去折腾它。

So in other words, you see a new model come out like 4.8, what I'm hearing from you is you're not jumping to try to tinker with it right away.

更重要的是，如何确保我们的数据能持续有序地组织起来，真正为 4.8、4.9 乃至下一个版本提供支持。

It's more about, Hey, how do we make sure that our data is sustainably organized in a way that can actually feed into 4.8-4.9- whatever the next thing is.

其实也不是这么说。

I mean, not really.

我希望你们能尽快跟进，而我们确实已经这么做了。

I hope you're jumping on it, and we definitely jumped on it.

就在开这个会之前，我还在我们公司的Slack频道里，其实我们已经抢先用了。

I was literally on our Slack channel before this meeting, and we had already jumped on it.

但关键是你得明白它到底在做什么。

But the idea is like, do you understand what it's doing?

我们和律师团队合作的流程非常复杂。

And we have very complex workflows that we run with our lawyers.

结果必须准确无误。

It has to be accurate.

我们还会跟踪提交后对方的接受率。

We track acceptance rates when we give subjections.

所以，我们不仅要给客户一个模型，还得把整个配套体系都搭好。

So we've got to make sure that it's not just a model that we bring to our customers, but it's the rest.

我希望公司里每个人都能迅速跟进，但更重要的是，得清楚这会对客户产生什么影响。

I hope if you're in the company, you just have to jump on it, but you have to know how it's going to impact your customers.

接下来我想问的是，随着模型越来越强，你认为模型本身的提升最终会推动企业级AI的普及吗？

And so the next question I have here is, as the models get better, do you think better models is ultimately what will accelerate adoption of AI in the enterprise?

我问这个问题，是因为你接触过不少不同类型的企业和客户。

And I ask this, you've worked with a number of different types of companies, customers, businesses.

你在加入Shopify之前，是在Atlassian工作的。

You were at Atlassian before you were at Shopify.

真的会是更好的模型在这里起决定性作用吗？

Is it really the better model that's going to make a difference here?

不如我们先回到最开始的问题。

Well, it's almost let's go back to the factory.

是电力更快了，还是电压更高了？

Was it faster electricity or higher voltage?

真正关键的是那些更先进的设备，以及会用它们的人。

It was literally better machines who know how to use it.

所以现在当然更好的模型会有帮助，但我们也都在搭建基于它的应用层。

So I think right now, obviously, better models are going to be good, but we're all also building the application on top of that.

我觉得真正重要的，是这两者的结合。

And I think that's where it's that combination that's really important.