探索谷歌人工智能的潜在层面

谷歌在Cloud Next大会一天之内就布下三步妙棋。

Google just ran a three layer strategy move in one day at their Cloud Next conference.

他们发布了新的TPU芯片，Chrome正变成你的AI同事，还和OpenAI的联合创始人米里亚姆·拉迪及其公司Thinking Machine Labs签了数十亿美元的算力大单。

They announced new TPU chips, Chrome is turning into an AI coworker, and a multibillion dollar compute deal with Miriam Radi, the former, cofounder of OpenAI, her company Thinking Machine Labs.

我觉得这可能是我到目前为止最明确的一个信号：谷歌在AI技术栈上，可能已经从结构上领先于OpenAI和亚马逊了。

I think this is one of the clearest signals from me so far that Google is structurally perhaps ahead of OpenAI and Amazon in the AI stack.

不过在那之前，我想先说说OpenAI和印孚瑟斯的合作，要把ChatGPT推入六十多个国家的企业市场。

Before that, though, I wanna talk about the fact that OpenAI is teaming up with Infosys to get ChatGPT into 60 plus countries of enterprise deals.

彭博社报道，一个未经授权的团队攻破了Anthropic新推出的网络安全工具Mythos。

Bloomberg reports an unauthorized group breached Anthropix new cyber tool, Mythos.

还有一个新成立的研究实验室叫NeoCognition，刚拿到四千万美元，专门打造像人类一样有专长的AI代理。

And there is a new research lab called NeoCognition that has landed $40,000,000 to build agents that actually specialize like humans.

今天这期节目，咱们就来聊聊这些事。

So we're gonna get into all of that on the show today.

咱们第一个故事是10x Science，这是诺贝尔奖得主卡罗琳·贝尔图齐实验室孵化出来的斯坦福衍生公司。

Our first story is that 10x Science, this is a Stanford spinout of a Nobel laureate Carolyn Bertuzzi's lab.

他们刚完成了480万美元的种子轮融资，领投的是Initialized Capital。

They just closed a $4,800,000 seed, which was led by Initialized Capital.

最让我觉得有意思的是，现在像DeepMind的蛋白质预测模型，能一次性生成成千上万种药物候选分子。

What they're doing that's so fascinating to me is that, basically, there's this problem where models like DeepMind's Protein Predictor, they're spitting out thousands of drug candidates.

这些候选分子多得数都数不清。

So there's there's just thousands and thousands of these drug candidates.

制药行业现在最大的瓶颈不是造出这么多分子，而是怎么从中筛选出真正值得继续推进的。

And there's a huge bottleneck in pharma where it's not just about, you know, getting all of these candidates, but it's actually triaging them.

关键是得挑出哪些最有希望做成药、做成疗法。

It's actually figuring out, like, which of all the candidates is worth pursuing to try to make medicine or therapeutics.

而目前主流的筛选工具，就是质谱仪。

And, basically, the standard triage tool is just mass spectrometry.

这玩意儿慢得要命。

So it's very slow.

而且结果特别难解读。

It's very hard to interpret.

而且这个方法只有专业领域的专家才能用。

It's, you know, domain experts only.

10x Science这家公司就是在这个基础上搭建了一套SaaS软件系统，他们把确定性化学原理和AI智能体结合起来，让整个分析过程可追溯、可解释。这一点至关重要，因为监管机构不会接受对分子作用的‘黑箱式’结论。

10x science is basically just building a SaaS layer on top of that, and they have deterministic chemistry plus AI agents to try and make the analysis traceable and explainable, which basically matters because regulators don't accept a black box answer on what the molecule does.

对吧？

Right?

哪怕AI模型一个劲地喊‘这太棒了’也没用。

So just because an AI model is like, oh my gosh.

我们发现了一个特别有意思的情况

We discovered this super cool thing.

这套方案根本行不通

Like, this isn't going to fly.

监管部门完全不会接受这种做法

Regulators don't want that.

我们绝对不能给这些分子弄一个无法解释的黑箱出来

There's no we can't have the black box on the molecules.

所以我们必须完全理解它、彻底测试它，而要决定该测试哪些化学物质、哪些药物，本身就是一个大难题。

So we have to be able to fully understand it, fully test it, and knowing which of these, which of these chemicals, which of these drugs to test is a big problem.

现在AI生物技术圈子里的人，都在聊生成式模型，但几乎没人去搞底下那些真正的‘铁锹和锄头’层。

So everyone in the AI biotech conversation is basically talking about the generative side, and almost nobody is building the kind of picks and shovels layer that's underneath of it.

所以我觉得10x Science这家公司特别有意思。

And so this is why I think 10x Science is an interesting company.

接下来我想聊聊NeoCognition。

The next thing I wanna talk about is NeoCognition.

这是一个刚从隐身状态出来的AI研究实验室，种子轮就融了四千万美元。

So this is an AI research lab that came out of stealth with about $40,000,000 in their seed round.

创始人是俄亥俄州立大学的教授余苏，他在那儿领导一个AI代理实验室，这一轮投资由Cambium Capital和Walden Catalyst Ventures领投。

Yu Su, an Ohio State professor who runs an AI agent lab there, is the founder, And the round was led by Cambium Capital and Walden Catalyst Ventures.

英特尔CEO唐立本，还有Databricks的联合创始人Ion Stoka，都个人投了天使轮。

Intel CEO, Lip Bu Tang, and also Databricks cofounder, Ion Stoka, both wrote angel checks into this.

我觉得这信号特别强，毕竟英特尔的CEO，还有Databricks这种级别的联合创始人亲自下注。

So I think that's really strong signal, right, when you when you have the CEO of Intel and when you have kind of these high profile cofounders of something like Databricks.

这太惊人了。

This is phenomenal.

但我觉得这家公司最关键的一点是，现在的AI代理成功率大概只有50%，因为它们本质上都是不靠谱的通才。

But essentially, the thesis on this company that I think is important is that the current AI agents succeed maybe 50% of the time because they're basically unreliable generalists.

他们认为，人类之所以能做好事情，不是因为我们啥都懂。

And so what they're arguing is that humans aren't great at doing tasks just because we know everything.

而是因为我们一进入新领域，就能迅速专精起来。

We're great because we specialize fast when we're dropped into a new domain.

所以NeoCognition想打造的是能自己快速专精的代理，而不是像现在这样，每个垂直领域都得手工定制一个代理。

So neocognition is trying to build agents that self specialize the same way instead of kind of the current model where you hand, you know, craft a custom agent for every vertical.

我做过不少定制代理的流程，太清楚这种‘一个领域一个代理’的做法根本没法规模化。

I've built enough custom agent workflows to know that kind of this per vertical approach doesn't really scale.

工程师还没用完，应用场景就已经堆满了。

You run out of engineers before you run out of use cases.

如果NeoCognition真能做出一个能自己学会新环境规则的代理——而且是完全自主学习，我觉得这才是真正的突破口，那绝对是巨大的突破。

So if NeoCognition can actually ship an agent that learns the rules of a new environment, and if it's doing this on its own, which I think is definitely gonna be the key, I think that is a massive win.

现在他们团队才15个人。

Right now, have only 15 people on their team.

大部分都是博士。

It's mostly PhDs.

虽然还很早期，但我觉得这家公司特别值得关注。

Definitely an early company, but I think this is one that's worth watching.

明白了。

Okay.

今天得跟大家说个坏消息，关于Anthropic的。

So we have some bad news for Anthropic today.

昨天Bloomberg爆料，有一群未经授权的人黑进了Mythos——那是Anthropic新推出的顶级企业级网络安全AI工具，据说能横扫整个安全圈。

There is a report from Bloomberg yesterday that an unauthorized group got access to Mythos, which is Anthropic's new exclusive enterprise cybersecurity AI tool that's going to take over the world.

对吧？

Right?

他们只给了少数几家企业用，毕竟这东西太危险，也太厉害了，能轻松挖出各种安全漏洞。

They gave it to a handful of enterprises because it's so dangerous and good at getting, you know, finding security vulnerabilities.

我觉得这肯定不是什么好消息，毕竟之前他们一直吹嘘这玩意儿有多危险，只给少数特定客户用。

I think this is probably not good news because they, you know, they've been a lot of hype about how dangerous it was and they only gave it to special people.

要是真被泄露了，对他们来说可就麻烦了。

And if it really is getting leaked, that's not great for them.

如果你仔细看彭博社的报道，其实是一群人躲在私人Discord频道里，专门盯未发布的AI模型，他们从Anthropic的一个第三方承包商那儿搞到了登录凭证。

I think what basically happened, if you if you read the reporting on Bloomberg, there was a group that operates out of a private Discord channel focused on unreleased AI models, and they got credentials from a third party contractor working with Anthropic.

然后他们根据Anthropic其他模型的命名规律，猜出了这个模型的访问地址，就这么简单。

They then guess the model's endpoint URL based on the pattern that Anthropic uses for other models, and that's it.

之后他们直接把截图和实时演示发给了彭博社，证明自己确实能访问。

They then gave Bloomberg screenshots and a live demo to prove access.

Anthropic的回应是正在调查，目前说没发现任何证据表明泄露影响了他们自己的系统，只波及了供应商的环境。

Anthropic's response to all of this is that they're investigating, and so far they say they found no evidence that it touched Anthropic's own systems, just the vendor environment.

我觉得这其实挺有代表性的，现在好多安全事件都是这么发生的。

I think this is interesting because this is kinda where a lot of these security incidents are happening right now.

根本不是模型本身被攻破了。

It's not really a a model exploit.

其实就是承包商的账号，加上一个很容易猜到的URL格式。

It's just a contractor credential plus kind of a predictable URL pattern.

我的建议是，如果你公司正在用AI工具，一定要好好审查一下供应商那边谁有权限、谁在用这些工具。

And my advice that I would give to people is that if you are a company that is running AI tools, make sure that you are auditing who is on your vendor side and who has access to all those tools right now.

更重要的是，这种事在别的地方也在发生。

Kind of the the bigger thing is that this is also happening in other places.

Anthropic目前正在和高盛、摩根大通、摩根士丹利积极洽谈，计划十月上市。

Anthropic is actively in early talks with Goldman Sachs, JPMorgan, Morgan Stanley about an October IPO.

所以现在任何对Anthropic不利的负面新闻，都只会让情况更糟。

So any sort of bad PR story for Anthropic right now is not gonna be great.

他们绝对得赶紧把这事压下去。

They they definitely want to get on top of this.

接下来，我有条关于OpenAI的消息。

Next up, I have some OpenAI news.

OpenAI和印孚瑟斯刚宣布了一项合作，要大力推广ChatGPT，尤其是他们的编程工具Codex。

OpenAI and Infosys just announced a deal to push ChatGPT and specifically Codex, their coding tool.

这相当于在Infosys的企业客户群里跟Cloud Code竞争。

It's kind of a competitor to Cloud Code into Infosys' enterprise client base.

覆盖了六十多个国家。

This is across more than 60 countries.

他们没怎么透露具体条款或交易金额，不过今天早些时候TechCrunch已经报道过这事了。

They don't they didn't really disclose the terms or how much, money is changing hands in this, but TechCrunch did report on this earlier today.

简单说下Infosys的情况，上个季度他们的AI服务收入达到了2.67亿美元，占总收入的55%左右，主要做的是那种超大规模的企业转型项目，微软和埃森哲通常就占着系统集成商的位置。

And basically for context on Infosys, they did about $267,000,000 in AI service revenue last quarter, about 55% of their total, and they play in kind of the massive enterprise transformation deals where Microsoft and Accenture often take the system integrator slot.

现在，Infosys把OpenAI的模型整合进了他们的Topaz AI平台，而OpenAI也借此搭上了 Fortune 500 级客户的渠道，这些客户他们自己直接触达不了。

So right now, Infosys is getting OpenAI models baked into their Topaz AI platform, and OpenAI gets a channel into Fortune tier accounts that they don't reach directly.

我觉得OpenAI现在正全力冲刺企业市场。

I think OpenAI is trying to make a big push for enterprise right now.

说实话，这目标他们早就有了，但眼下这一波猛攻是因为感觉Anthropic在企业市场跑得太快了。

I mean, obviously, this has been a a goal for a long time, but this big push is coming right now as it feels like Anthropic is really running away with the the market in enterprise.

我接触过不少企业，买了ChatGPT企业版授权，结果搁置了一年都没用起来，压根没跟实际业务打通。

I've worked with a lot of enterprises that bought the ChatGPT enterprise licenses, and they kind of sat on them for, like, a year because they didn't wire them into anything real.

这正是我觉得 Infosys 想填补的空白。

This is basically the gap that I think Infosys is trying to fill.

从战略上讲，这跟微软把 Azure OpenAI 和 Copilot 捆绑销售的做法很像，只不过 Infosys 能接触到很多微软拿不下的单子。

And strategically, I think this is kinda mirroring what Microsoft already does by bundling Azure OpenAI with Copilot, except Infosys plays in a lot of deals where Microsoft doesn't always win.

微软在这些单子里根本插不上手。

It's not it's not in the deals.

它没那个渠道。

It doesn't have the distribution.

所以这算是一条新路子。

So it's kind of a new channel.

我觉得这种合作能帮 OpenAI 把企业收入推上去。

I would expect to see a deal like this to help push OpenAI's revenue higher.

OpenAI 当然想追上 Anthropic，人家现在年收入都超过三百亿美元了。

OpenAI obviously wants, you know, enterprise revenue to catch up to Anthropic, which is now over $30,000,000,000 in annualized revenue.

要追上 Anthropic，他们非得有这种渠道不可。

They need this type of distribution in order to try to catch up to Anthropic.

明白了。

Okay.

如果你还在为ChatGPT、Claude、Gemini和Grok付费，顺便说一句，还有11 Labs的音频工具和一堆独立的图像生成器，那你真得听听我介绍的AI Box。

If you're still paying for ChatGPT, CLAUD, Gemini, and Grok, by the way, 11 labs for audio and maybe like separate image generators, I have to tell you about AI Box.

这是我亲自搞出来的玩意儿。

This is what I personally have built.

我朋友老问我怎么用AI又不花大钱买一堆订阅，我就推荐他们这个。

It's what I recommend to my friends who keep asking me how to actually use AI without going you know, without spending a ton of money on subscriptions.

AI Box能让你在一个地方访问80多个AI模型，全是顶尖的，你想用哪个就用哪个，或者直接问问题，系统自动帮你挑最合适的。

Basically, AI Box gives you access to over 80 AI models in one place, all of the top models, so you can pick whatever one is best for the tack task that you're working on, or just ask a question and we'll pick the best model for it.

我觉得最实用的是我们的自动化构建器。

The part that I think is super useful is our automation builder.

你只要用大白话描述你想要什么工具，它就能给你搭好工作流。

You could just describe a tool that you want in plain English, and it will build a workflow for you.

根本不用写代码。

So you don't need to code.

你根本不用费劲去适应新平台。

You don't need to wrestle with new a new platform.

你只要说清楚想要什么，它就能帮你搞定。

You just describe it, and it will build it.

每个月只要8.99美元。

It's $8.99 a month.

便宜得离谱。

Very cheap.

我觉得这能让你省下一大笔钱，不用再付一堆订阅费，还能用上所有值得用的AI模型。

I think this saves you a ton of money on just all of the different subscription platforms, and you get your hands on every single AI model that is worth using.

我们这边有不少谷歌的最新消息。

We have a ton of news from Google.

他们今天的Google Cloud Next大会开幕了，谷歌这次可是干货满满。

Their Google Cloud Next conference kicked off today, and Google did not come, empty handed.

他们公布了三个重磅消息：新一代TPU、Chrome当AI同事，还有和Thinking Machine Labs的算力合作。

They had three major announcements, new TPUs, Chrome as an AI coworker, and a Thinking Machine Labs compute deal.

我觉得把这些放在一起，谷歌现在在AI整个技术栈里的布局可以说是得天独厚。

And I think taken together, basically, they're gonna make Google the most structurally positioned player in the AI stack right now.

我来一一说说这些合作，不过说实话，我对他们的发布真的很佩服。

I'm gonna walk through all of these different deals, but, yeah, I've been pretty impressed by their announcements.

当然，其他公司也可能崛起，但我绝对不认为谷歌会在这行掉队。

Of course, we could see other players, but I would definitely say never count Google out in this industry.

先说芯片方面的合作。

So first, the chips deal.

谷歌发布了两款新的TPU。

Google announced two new TPUs.

一个是用于训练的TPU v8，另一个是用于推理的TPU i8。

They have the TPU, eight t for training and the TPU I eight I for inference.

我觉得这种训练和推理的分开设计，特别有意思。

The split alone is, I think, what's really interesting to me.

现在跑AI在生产环境里，推理的成本才是大头，专门针对推理做优化的芯片，经济上太关键了。

Inference is the dominant cost of running AI and production right now, and having these kind of dedicated inference silicon matters economically.

这对利润影响可不小。

It makes a big impact to the bottom line.

谷歌声称，他们的训练速度是英伟达的三倍，每美元的性能提升高达80%，而且单个集群能扩展到一百多万个TPU。

So Google's claims is that it is three times faster at training and 80 better at performance per dollar against the NVIDIA alternatives and the ability to scale more than a million TPUs in a single cluster.

他们这次真是方方面面都兼顾了，不过传统上来说——我也说不准这算不算英伟达的锅，虽然肯定有人不这么想。

So they're they're hitting a bunch of different things here, but NVIDIA has traditionally I don't know this is necessarily even NVIDIA's fault, although I'm sure some people will disagree with me here.

但英伟达过去在集群规模上一直有点短板。

But NVIDIA traditionally has had a a problem with the cluster sizing.

我记得以前不是OpenAI，是……

And I remember when OpenAI's or not OpenAI.

是xAI搞他们的第一个‘GigaCluster’那会儿，具体名字我记不清了。

When x AI did their first, like, giga terminal I don't know.

算了，不提了。

Whatever.

他们给那玩意儿起了个机器人名字，我现在真想不起来了。

They're they have a robot name for it, so now I I can't remember.

但他们还是搞了个十万台NVIDIA G200 GPU的集群。

But they put together a 100 g p NVIDIA g 200,000 NVIDIA GPUs.

这事儿当时可轰动了。

It was like this big deal.

他们算是摸出了门道，之后其他人也开始搞越来越大的集群。

They kinda figured this out, other people started building these bigger and bigger clusters.

说白了，他们得从头搭建基础设施，方方面面都得自己琢磨。

Well and it was kinda like they had to figure out a lot of the infrastructure and build it and think about it.

看来谷歌是直接冲着这个目标去了，单个集群能撑上百万个TPU，这简直太离谱了， scalability 简直炸裂。

Well, it looks like Google is building straight for that, and the ability to have more than a million TPUs in a single cluster is insane and, I mean, insanely scalable.

我觉得这些数字是谷歌自己说的。

So I think those are Google's own numbers.

目前还没人做过独立的性能测试。

We don't have any independent benchmarks yet.

不过谷歌也没打算跟NVIDIA断了合作。

Google is also not killing the NVIDIA partnership.

他们今年晚些时候会在谷歌云上推出维拉·鲁宾。

They're reselling Vera Rubin in Google Cloud later this year.

我觉得谷歌其实两头都要，想让客户自己选，这跟AWS现在孤注一掷搞全栈训练的做法正好相反。

So I think, basically, Google wants both chips and it wants to let customers choose, which is kind of the opposite of what AWS's all in one training bet is right now.

我觉得谷歌在这上面最后会赢。

I think Google will actually come out ahead on that.

还有个挺有意思的事，谷歌宣布要把Chrome变成你的AI同事。

The other thing that I think is interesting that Google announced is that they're turning Chrome into an AI coworkers.

他们新推了个功能，叫自动浏览。

They have a new feature which is called auto browse.

这功能由Gemini驱动，直接内嵌在你的Chrome工作版里。

Gemini powers it, and it's basically running inside of your Chrome for workspace.

它能跨你所有打开的标签页理解上下文，自动帮你完成日常工作。

And it reads context across your open tabs and automates actual workplace tasks.

比如帮你录入CRM系统里的数据。

So it can enter your CRM data.

它能比较不同供应商的报价。

It can compare vendor quotes.

它能总结候选人的履历。

It can summarize candidate portfolios.

它能整理竞争对手的调研报告。

It can write up competitor research.

说实话，这工具听起来挺有意思的。

If I'm being honest, it sounds like an interesting tool.

确实挺实用的。

It sounds useful.

我其实试过一些类似的东西。

And I've actually tried some similar things.

以前微软在Edge浏览器里也搞过类似的功能，能读取你所有打开的标签页内容，还能跟它们互动，特别方便。

Microsoft had something like this in Microsoft Edge a while back where, like, all of your tabs that are open, it can just get data and read and interact with them, which is super useful.

不过有时候你需要的信息不在标签页里，我觉得这时候Claude的协作者还是更厉害，毕竟它是个独立应用。

Sometimes though, you need things that are not, you know, in your tabs, and I think this is where Claude's co work is still gonna keep crushing because they're an app.

你给它访问你桌面的权限。

You give it access to your desktop.

它能处理你的文件。

It can files.

它还能在你的本地设备上下载东西。

It can download things on your local device.

所以我知道，像Open Google这样的工具，压根不是想直接跟这个竞争。

So I know that, like, open Google isn't trying to directly compete with that.

他们只是强调人要全程参与，不会让智能体在没得到你批准的情况下擅自行动。

They're all they're kind of saying, you know, they're just like this human in the loop, and they're not kinda letting the agents execute without your approval.

所以呢，他们其实是在把这个当成一个优势来宣传。

And so, anyways, I think, like, they're kind of pitching it as a as a plus.

依我看，他们这玩意儿已经落伍了。

In my personal opinion, it's behind the ball.

跟Claude Co-Work比起来，它差远了——有时候我真不想每做一件事都弹出一百个弹窗要我点确认。

It's it's kind of not as good as something like Cloud Cowork, which actually sometimes I don't wanna have to get a 100 pop ups asking for approval.

我就希望它能把事儿办了，尤其是我都看见它成功搞过三四回了。

I just want it to get the job done, especially when I've seen it do it, like, three or four times successfully.

我才不想一个个去点同意呢。

I don't wanna approve it.

我就想让它自动搞定。

I want it to automate it.

也许他们的目标用户群不太一样，但我还是更愿意用 Cloud Cowork。

And so maybe they're playing it kind of a different target audience there, but I still would probably use Cloud Cowork over that.

另外还有一项挺实用的功能，跟 Cloud Cowork 类似，你可以把常用的工作流程保存成技能，用斜杠就能直接调用。

You can also something that is very useful though and kind of like Cloud Cowork, you can save frequent workflows as skills and you can fire them with a forward slash.

这项功能现在正在美国上线。

This is rolling out in The US.

首先是 Workspace 用户能用上。

Workspace users are getting this first.

而且他们还承诺，企业级的提示词不会用来训练谷歌的模型，我觉得这决定挺对的。

And there's also a commitment that enterprise prompts are not gonna be used to train Google's models, which I think is the right call.

很多企业都会很担心这一点，尤其是如果他们想把机密数据放进去的话。

A lot of enterprises would be very concerned about that, especially if they wanted to put confidential data in there.

不过呢，我可不是要一味挑这事儿的毛病。

But again, and not to like, I don't know, say all of the downsides to this.

但说实话，你得手动去触发一个技能，而像Cloud Cowork这种，直接设好就能自动跑。

Again though, you have to go and fire off a skill manually, whereas something like Cloud Cowork, you can just set these things up to automatically run.

明白了。

Okay.

接下来我觉得挺有意思的是，谷歌云刚和米里亚姆·穆拉迪的Thinking Machine Labs签了个几十亿美元的大单。

The next thing that I thought was interesting is that Google Cloud has just signed a multibillion dollar deal with Miriam Muradi's Thinking Machine Labs.

是十亿美元级别的交易。

This is a single digit billions.

这笔合作让Thinking Machine Labs能用上谷歌云上的NVIDIA GB300系统，还有他们首款产品Tinker的训练和部署服务——Tinker是个用来打造定制化前沿模型的工具。

It gives Thinking Machine Labs access to NVIDIA's GB 300 system on Google Cloud, plus training and deployment services for their first product, is called Tinker, which is a tool for building custom frontier models.

Thinking Machines的估值高达120亿美元，他们跑的都是重强化学习任务，计算量大到离谱。

Thinking Machines is raising at a $12,000,000,000 valuation, and they run a heavy reinforcement learning workload, which is basically very insanely compute intensive.

米勒·奥特，这位联合创始人之一，多次提到谷歌让他们上手得特别快。

Mile Ott, one of the founding researchers, has basically talked a ton about how much Google got them up and running really quickly.

我觉得这对他们公司来说会很有帮助。

So I think this is something that's gonna be helpful for their company.

我觉得把这些事串起来的关键是，谷歌在玩一套三层策略。

The thing that I think ties all of these together is that Google is running a three layer strategy.

底层是芯片，也就是TPU加上他们转售的NVIDIA显卡，其实他们并不在意到底哪个芯片胜出。

The bottom layer is silicone, TPU plus these resold Nvidias, And then they don't really care which chip wins.

我的意思是，他们当然希望自家的芯片赢，但不管你买的是不是NVIDIA，只要从他们那儿买，他们都能赚钱。

I mean, obviously, they kinda want their own to win, but they're gonna make money either way if you're even if you buy Nvidia from them.

中间层是为前沿实验室提供算力托管服务。

The middle layer is being the compute host for Frontier Labs.

Anthropic早就用TPU了，现在Thinking Machine Labs也要在谷歌云上跑起来了。

Anthropic already runs on TPUs, and now Thinking Machine Labs is gonna be running on Google Cloud.

最上层呢，是浏览器和工作空间里的智能代理层。

The top layer is the agent layer in the browser and the workspace.

目前来说，只有谷歌真正把这三层都做扎实了。

Google is, you know, the only player that's hitting all three credibly right now.

Anthropic 在终端用户和顶层应用上确实做得特别好，但他们自己不造芯片。

Anthropic is obviously crushing it with end users and and, you know, kind of in that top layer, but they're not making their own chips.

对吧？

Right?

所以谷歌要是能把整个栈都掌握在手里，优势就太大了。

And so Google has a big advantage when they own the entire stack.

微软在第二层上全力押注 OpenAI，但在芯片层上却很弱。

Microsoft is really heavy on OpenAI at layer two, but really weak on silicone.

亚马逊把所有赌注都压在了 Anthropic 身上。

Amazon is betting the farm on Anthropic.

英伟达只做芯片层，根本不碰应用层。

NVIDIA is the chip layer and doesn't touch the application layer.

说到底，真正把全栈都做全的，就只有谷歌。

Google is basically the only company really doing the full stack.

所以我觉得最有力的反驳是，Gemini 在消费者真正关心的那些评测指标上，还没能追上 GPT-5.4 或 Cloud Opus 4.7。

So I think the strongest argument against this is that Gemini still hasn't closed the gap with GPT 5.4 or Cloud Opus 4.7 on the benchmarks that consumers actually care about.

我觉得吧，那些 TPU 的数据可能是谷歌自己搞的，而且说实话，除非有独立实验室在 8T 和 8I 上跑出真实的训练结果，否则他们说的 3X 优势，基本就是营销话术。

And so I think, you know, maybe those TPU numbers are Google's own, and you could also say, you know, until, like, an independent lab publishes real world training runs on the eight t and on the eight I, the three x claim that they've given is marketing.

我不是要故意质疑谷歌，但咱也得讲点公道。

I mean, I don't wanna doubt Google too much, but you have to be fair.

现在 AI 公司太多了，谷歌以前就爱这么干——吹得天花乱坠，结果现实里根本撑不住。

There is a lot of these AI companies, and Google's been known to do this in the past, to make big claims on marketing, and doesn't always stack up in the real world.

但即便如此，我依然觉得，谷歌现在的位置简直太强了。

But here's why I still think even with all of that, Google is in an incredible position right now.

客户根本不在乎 Gemini 能不能冲上 Evo 排行榜第一。

The customer really doesn't care whether Gemini tops the Evo leaderboards.

只要他们在谷歌的平台上跑推理更便宜，又能通过 Chrome 直接推给员工，还无缝对接 Workspace 的数据，不用再找一堆系统集成商，这种整合能力，未来十二个月比单纯跑分领先重要多了。

If they can run inference cheaper on Google's stack, and if they can serve it to employees through Chrome and wire it through their workspace data without, you know, having this big huge system integrator, structural position, I think, matters way more than just pure benchmark lead in the next twelve months.

而且谷歌现在早就嵌进了每个人都在用的每一个工具里，所以我相信，谷歌肯定能大获成功。

And Google is integrated with every tool that people are using right now, And so I think Google's gonna do quite well.

另一点我觉得挺有意思的是，谷歌现在正面临司法部的反垄断诉讼，这事儿还悬着呢。

The other thing that I think is interesting is that Google is doing this with a DOJ antitrust kind of suit overhanging them right now.

他们一直被盯着的搜索业务整改案还没完，现在把Chrome变成一个能调用Workspace数据的智能层，这招简直就是专门吸引监管注意的典型操作。

The search remedies case that they've been in for a while now is still live, and turning Chrome into an agent layer that pulls from workspace data across your browser is basically the exact kind of move that makes regulators pay attention.

我觉得你最好留心一下，接下来几周司法部或者欧盟会不会有什么表态。

I think you would, you know, be wise to watch for some comments coming from the DOJ or the EU in the next few weeks.

这事儿肯定会引来大量关注，就像微软和他们跟OpenAI的合作过去一年一直被盯着一样。

This is gonna draw scrutiny that Microsoft and kind of their OpenAI deal has already been dealing with for a year.

我个人最关注的是：独立的TPU 8T和8I基准测试结果、自动浏览功能会不会从试点真正推广到企业内部，还有没有更多前沿实验室会因为Thinking Machine上线、而Anthropic已经入驻，转投谷歌云。

What I'm watching personally is the independent TPU 8T and 8I benchmarks, whether auto browse graduates from pilots to real rollouts inside of enterprises, and whether more Frontier Labs migrate to Google Cloud now that Thinking Machine is in and Anthropic is already there.

所以，只要再有一个实验室跟进来，那就是一个巨大的结构性转变。

So I think if even one more lab follows, that is a huge structural shift.

好了。

Alright.

今天就聊到这儿。

That's it for the show today.

如果你觉得这期节目有收获，麻烦在苹果播客留个评论，或者在 Spotify 给个五星好评，我真的会特别感激。

If you've got something out of this, it would mean the world to me if you could drop a comment on Apple Podcasts or if you could hit the stars over on Spotify.

在 Spotify 是点‘关于’页面就能写评价、打星星，在苹果播客也一样，留个评论就行。

It's the about tab on Spotify where you can leave a leave a review, leave some stars, drop a comment on Apple.

这真的对节目帮助特别大。

It helps the show a ton.

对吧？

Right?

我可不是白让你帮忙的。

I don't ask for it for nothing.

真的特别有帮助。

It really helps.

更多人就能发现咱们的节目了。

More people find it.

对算法也有好处。

It helps in the algorithm.

另外，如果你想知道怎么用每月8.99美元就能解锁80多个AI模型，还能用纯英文的自动化工具，就去试试aibox.ai吧。

Also, if you wanna get access to over 80 models with an automation builder that runs in plain English for $8.99 a month, go check out aibox.ai.

简介里有链接。

There's a link in the description.

好了，咱们下期再见。

In any case, I'll catch you guys all in the next episode.