Claude赋能移动协作

Anthropic 终于要把 Claude Co-Work 推到你的手机上了，就在 Claude 应用里。

Anthropic is finally letting Claude Co-Work go to your mobile device, your phone on the Claude app.

你可能已经注意到，那边能用 Claude 写代码，但 Claude Co-Work 一直没上线。

You've probably noticed you can get Claude code over there, but Claude Co-Work was missing.

所以他们很快就会推出这个功能。

So they're gonna be releasing that soon.

DeepSeek 正在研发自己的推理芯片，打算彻底绕开 NVIDIA 和华为。

DeepSeek is planning their own inference chips to basically cut NVIDIA and Huawei out of the deal for them.

他们之前严重依赖这两家，现在想摆脱这种依赖。

They're trying to get around them, and they rely on them heavily.

SK 海力士计划进行高达 280 亿美元的 IPO，因为内存需求暴涨，股价已经翻了两倍多，这家公司正在疯狂扩张。

SK Hynix is planning a $28,000,000,000 US IPO or listing because the demand for their memory has made a 260% stock jump for them, so this company is growing like crazy.

Solos 要推出 AirGo 6 了。

Solos is launching AirGo a six.

这是款智能眼镜。

These are smart glasses.

它们只有19克重，本质上是带语音AI的智能眼镜，但完全没有摄像头。

They're only 19 grams, and essentially, they're smart glasses with voice AI, but there is no camera in them.

Vercel的Guerreroche主张把AI模型和智能代理分开，他认为这样能更高效、更划算，把模型和代理拆成两个独立的部分。

Vercel's Guerreroche is betting on splitting AI models from agents, so he thinks that you're gonna get more bang for your buck and more efficiency if you actually take the AI model away from the agent and have these as two separate things.

说点无关的，今天我收到北卡罗来纳家乡发来的短信，说我们那儿正在干旱，只有门牌号是偶数的住户，才能隔天浇一次草坪。

In completely unrelated news, I got a text today from my city in North Carolina that said that we are in a drought, and I'm only allowed to water my lawn if my house number is even every other day of the week.

结果呢，一整天都在下雨。

And then it has proceeded to rain all day long.

我不确定是不是为了给AI数据中心省水才限制浇灌，不过现在雨下得这么猛，干旱估计算是结束了。

I'm not sure if we were supposed to be saving water so the AI data centers don't run out, but I think we have enough rain now, the drought might be officially over.

我太兴奋了，好久没这么激动过了。

Okay, I'm super excited, maybe the most excited I've been in a while.

Anthropic终于让Claude Co-Work上线到手机App了，就是你手机里的Claude应用。

Anthropic has just allowed Claude Co-Work to go on your mobile device on your app, on the Claude app on your phone.

以前只能在电脑上用，挺不方便的——你用Claude Code的时候，虽然能从手机远程操控，但功能有限得很。

Previously, it was only on your desktop computer, and it was actually pretty annoying because if you're doing something with Claude Code, you could remote control it from your phone, although it had some limitations.

其中一个问题是，我发现手机上的所有权限设置都是一样的，所以还得像以前那样反复批准各种操作，真的挺烦的。

One of them being I found I had like, all the permissions were set the same way from my phone, so I still had to keep approving things over and over again like the like the olden days, which was kinda brutal.

但我特别兴奋，因为现在你可以在电脑上启动任务，然后离开，接着在手机上继续和 Claude Co-Work 互动，而且它似乎还能继续控制你的电脑，帮你完成各种操作。

But I'm really excited because now you can fire something off on your computer, leave, and you can keep, you know, iterating with Claude Co-Work on your phone, and I think it's actually gonna be able to take control of your computer still and do stuff.

这实际上对很多事情都特别有用。

So this is actually insanely useful for so many things.

与此同时，Claude 也公布了一份人们使用 Claude Co-Work 的用途清单。

At the same time, Claude has also published a list of what people are using Claude CoWork for.

如果你需要点灵感，或者只是好奇，里面有一张图特别值得一看。

So in case you need any ideas or you're curious, there's one graph in particular.

他们说，33% 的人用它来处理业务流程和运营，16% 用于内容创作和文案撰写，而只有 8.7%。

They said that 33% of people are using Claude CoWork for business processes and operations, 16% are using it for content creation and copywriting, and only 8.7.

也就是说，大约 9% 的人用它来做软件开发。

So basically, like, 9% are using it for software development.

我本人经常用它来开发软件，不过我会在 Claude Co-Work 和 Claude Code 之间来回切换。

I actually use it for software development a lot, but I kinda toggle between Claude Co-Work and Claude Code.

这两种各有优缺点。

There's, like, pros and cons to both of them.

如果我要在浏览器里登录、点来点去，或者从电脑上拿文件，我就会用 Claude Co-Work。

I find if I'm doing something where I need to be, like, logging in on the browser to stuff and clicking around and grabbing files off my computer, I'm gonna use Claude Co-Work.

如果是发推送、搞 GitHub 这类事情，我更倾向用 Claude Code。

If I'm doing something that's more in like, you know, just making pushes to my GitHub and stuff, I'm gonna use Claude Code.

总之，两者各有长短，但我觉得 Claude Code 最大的优势就是用起来更经久耐用。

So anyways, there's pros and cons, and I will say the biggest pro with Claude Code is you just get way more mileage out of it.

我早就把 Claude Co-Work 的配额用光了。

I burned through all my Claude Co-Work credits.

我干脆又注册了一个 Claude Max 账号，就怕 Fable 的配额用完，顺便也谢谢 Claude。

I actually made a new Claude Max account just because I I didn't wanna run out of all my Fable credits, which, shout out also, Claude.

Claude 说，如果你有 Claude Max 订阅，Fable 五的功能在 Claude 应用里会延长到第十二天。

Claude said that they're extending Fable five, the use inside of the Claude app, if you have like a Claude Max subscription.

他们会延长到第十二天。

They're extending it to the twelfth.

本来应该今天七号就结束了，但他们延长到了十二号。

They're they were supposed to end today on the seventh, but they're extending it to the twelfth.

所以如果你用了 Claude，就能多用不少 Fable Five 的额度。

So, if you have Claude, you're gonna get way more Fable five.

但问题是，如果你跟我一样，可能早就把 Fable Five 的额度用光了。

Now, the problem is, if you're anything like me, you burnt all of your Fable five credits already.

我昨天刚开了个新账号，结果这一周的额度今天一上午就用完了。

I made a new account yesterday, and I have officially burnt through all of my weekly credits already.

这确实是个问题。

So, that's definitely a problem.

我正想办法解决，但得等到明天周三，额度才会重置。

I'm trying to work on that, and I have to wait till tomorrow, Wednesday, before it reset.

现在大家用 Claude 做得比较多的还有运维和基础设施，占了百分之七。

The next things that people are using Claude for is DevOps and infrastructure, 7%.

研究和情报分析，占了百分之六点四。

Research and intelligence, 6.4%.

数据分析和商业智能，占5.8%。

Data analysis and business intelligence, 5.8%.

文档处理与信息提取，占4%。

Document processing and extraction, 4%.

销售与营收运营，占4%。

Sales and revenue operations, 4%.

其余所有用途加起来大约占14%。

And then everything else is about 14%.

最主要的应用还是业务流程与运营、内容创作和文案撰写。

So the biggest thing is business processes and operations, content creation, and copywriting.

这三项加起来差不多占了大家使用量的一半。

That is basically 50% of what everybody is using it for.

我知道你们都能在手机上完成这些操作，这真是个巨大的突破。

I know you'll be able to do all of that from your phone, is a huge unlock.

而且它全天候在线，你可以在手机上丢个任务，然后去睡觉，早上起来就完成了，特别方便。

It's also on twenty four seven, so you can actually give it a task on your phone, and then go to bed, and it's done in the morning, which is really nice.

我以前用手机操作这类软件时遇到过问题，比如一锁屏，程序就停了，这真的很烦。

I think I've had issues in the past when working with this similar type of software from my phone where it, like, if you turn your phone screen off, then it stops running, which is bad.

不过没关系，这次它能一直后台运行，而且完全通过手机操作，简直太棒了。

So anyways, it's always on, it's from your phone, this will be fantastic.

今天正式上线了，目前还在公测阶段，只有月费100美元的Max套餐用户才能用。

It is launching today, it is in beta, and it is for max plan subscribers if you're paying a $100 a month.

之后也会向月费20美元的Pro用户开放。

It's also gonna roll out to the pro users at $20 a month later on.

Co-Work展示的演示中，它从邮件、Slack和多个会议记录里提取信息，自动生成文档、草拟邮件，全都是在手机上操作的，特别酷——不过实际上它是在你的电脑上完成这些任务的，这点很聪明。

Co-Work's demo that they showed with this basically showed it pulling data from emails, Slack, and a bunch of different meeting transcripts, and it was writing a document, drafting emails, and it was all doing that from your phone, which is really cool, but it was using your computer to, like your desktop to do all those things, which is pretty sweet.

中国流行的大型语言模型DeepSeek，正计划自主研发芯片来运行推理任务。

DeepSeek, the popular Chinese LLM, is going to be designing their own chips to run inference.

他们这么做，主要是为了摆脱对NVIDIA的依赖，同时也要跟国内的主要竞争对手华为竞争——目前华为在中国数据中心芯片市场占据了约一半的份额。

This is something that essentially they're doing to get away from relying on NVIDIA, and of course, their big competitor, Huawei, over in China, which currently dominates about half of China's data center chip market.

我想说的是，推理芯片比训练芯片更便宜、迭代更快，这对DeepSeek来说是个优势，每生成一次响应，他们的利润率都能提高。

What I will say is that inference chips are cheaper and faster to iterate on than training chips, so that's kind of DeepSeek's something that's gonna help their own profit margin on every response that their model's actually generating.

阿里巴巴和百度也在推动类似的定制芯片。

Alibaba and Baidu are both kind of pushing for similar custom chips.

OpenAI和博通已经推出了自己的推理芯片，名叫Jalapeno。

OpenAI Broadcom have released their own inference chip, which is called Jalapeno.

一般来说，定制加速器从设计到量产需要三到五年时间，而DeepSeek还没透露任何代工合作伙伴的信息。

As far as the timeline goes, custom accelerators typically take about three to five years from design to production shipment, and DeepSeek, they haven't said, you know, what who any of their, like, foundry partners.

他们也没公布任何时间安排。

They haven't said anything about their timeline.

他们对性能目标也一字未提。

They haven't said anything about their performance targets.

虽然他们已经开始研发，但我认为短期内还看不到成果。

So while they are starting to develop this, I don't think we're gonna see it anytime very soon.

SK海力士本周将在美国股市发售股票。

SK Hynix is gonna be selling shares on The US Stock Market this week.

他们计划融资约280亿美元，让美国普通投资者也能参与投资全球第二大内存芯片制造商。

They're raising about $28,000,000,000, and they're letting anyone in The United States get access to the world's second largest memory chipmaker.

这还是头一回发生，如果你还记得我上一期说的，三星、SK海力士，还有爱达荷州的一家公司，目前几乎是全球唯一能大规模生产AI训练所需定制内存芯片的三家厂商。

This is the first time this is happening, and if you remember my last episode, Samsung, SK Hynix, then one company out of Idaho are basically the only three people that can mass produce these custom memory chips for AI training at scale.

这次上市，说白了就是想抓住最近几个月股价暴涨260%的势头趁热打铁。

This listing in particular is going to I mean, they're really just trying to capitalize on the fact that in the last, couple months, their stock has surged 260%.

这一切都源于AI数据中心——它们迫切需要高速内存芯片来训练和运行AI模型。

All of this is because of AI data centers, which desperately need high speed memory chips to train and run AI models.

SK海力士这次要发售1780万股，很多人把这种情况称为‘内存末日’，因为他们在第一季度财报里显示，营收同比暴涨了200%，全是因为AI任务对内存的需求太猛，市场严重缺货。

SK Hynix is gonna sell 17,800,000 shares, And a lot of people are calling this entire situation Ramageddon because when they came out with their first quarter revenue, they're up about 200% year over year just because all of the AI workloads have this massive shortage of RAM.

如果再算上三星那边的动向，SK海力士和三星两家公司现在都在各自投入超过5500亿美元，新建内存工厂。

And if you would take into account Samsung and everything going on there, SK Hynix and Samsung are both spending over $550,000,000,000 to build new memory factories.

这可以说是半导体历史上规模最大的一次协同设备投资了，两家正在齐心协力，眼下正砸下巨资全力推进。

This is one of the largest coordinated equipment investments in semiconductor history, So they're working together, and there is a massive investment that is going in right now.

Solos公司要推出一款叫Ergo的六代智能眼镜。

Solos is launching Ergo a six smart glasses.

眼镜重量只有19克，他们干脆取消了摄像头，但内置了语音AI功能。

They're only 19 grams, and they've decided to cut out the camera entirely, but they do have voice AI built into them.

这其实是个挺有意思的布局，因为竞争对手有Meta。

These are this is a really interesting play because we have the competitors, are Meta.

苹果也在开发智能眼镜。

We have Apple working on smart glasses.

谷歌目前也有智能眼镜在推进。

We have Google that has smart glasses going on right now.

但Meta和谷歌的产品里都内置了摄像头。

But Meta and Google both have cameras inside of theirs.

我不太确定苹果那边到底在搞什么，毕竟他们至少还得等几年才可能发布，而且他们一向动作很慢。

I'm not exactly sure what's going on with Apple because, you know, they aren't gonna come out for a few years, and and we'll see that in a couple years because they're pretty slow with all of this.

不过在这段空档期，就在我们等着这些大厂推出智能眼镜的时候，反而有一些新公司冒出来，做出了挺酷的产品。

But it's kind of cool that in the meantime, while we're waiting for a lot of these other big players to come out with their smart glasses, we're getting some new firms in here making some cool stuff.

这款眼镜特别轻。

This is super light.

他们说是从隐私保护出发的设计，能解决智能眼镜目前最大的一个痛点。

They say that they're privacy first design, and it's gonna solve one of the biggest problems that smart glasses have.

我注意到很多人在领英和其他地方抱怨这些摄像头，你走在街上，看到别人戴着这种类似雷朋的智能眼镜，根本不知道自己是不是正被录着。

I think a lot of people I've seen complaining on LinkedIn and other places about the cameras, and if you're walking around and people have these Matter Ray Bans, you never know if you're being recorded.

跟你说话的人可能会因此感到不舒服。

It might make people uncomfortable if you're talking to them.

而且如果你整天戴着它们，还担心摄像头会被黑，那确实有不少隐患。

And so or if it's if you're wearing them and you're worried that the cameras could be hacked while you're wearing them all day, right, there's a whole bunch of issues.

所以去掉摄像头后，感觉好多了——再也不用整天顶着个大摄像头到处走，但依然能享受到AI的大部分好处，比如语音转文字、跟你对话，还能实时告诉你眼前看到的东西或者你在聊的话题。

So cutting out the cameras on these, it feels great to not just be walking around with a giant, camera on your head all day long, and you get a lot of the benefit of the AI, which is, you know, you can voice to text, talk to it, and get answers back on what you're seeing now, or what you're, you know, talking about.

不过，我觉得最大的缺点是，比如我正盯着汽车发动机，想找某个特定的盖子。

Now, I think, of course, the biggest drawback here is if I'm looking at my car engine and I'm like, where is the specific cap I'm looking for?

如果有这副眼镜，它就能直接告诉我。

If I have the glasses on, can actually let me know.

谷歌的版本可以高亮显示目标物体，特别直观，但这款做不到这一点。

In in Google's case, they're gonna be able to, highlight things so you can actually see which is really cool, but this won't have that capability.

这款眼镜内置了语音AI助手，功能很强大，比如实时翻译，简直太棒了。

These glasses in particular have voice AI, and their assistant can handle anything from real time translation, which is amazing.

我觉得这正是智能眼镜上AI最棒的应用之一。

I think this is one of the greatest use cases of AI in these glasses.

日程提醒、电话功能，还能通过耳后扬声器让你听到周围的声音，能帮你放大重点信息，特别实用。

Calendar reminders, calls, they also let you, hear the surroundings through speakers behind your ear, so they can amplify different things for you, which is really useful.

我可能真得给自己配一副了。

I just might have to get myself a pair.

Vercel 是一个平台，每天有数百万个AI代理在这里部署代码，说到这个，我个人就是其中之一。

Vercel, which is the platform where millions of AI agents deploy code daily, and when I say that, like personally, I'm one of those people.

基本上，我现在所有用到的软件和应用，都是靠 Vercel 来托管网站的。

Basically, I think all of my different softwares and apps are using Vercel to host all of the different sites that I have running right now.

之所以选它，是因为它真的表现得太出色了。

And the reason why is because it works incredibly well.

我觉得还有其他不错的平台，比如 Netlify 应该也做这个，亚马逊好像也有类似的服务。

And I think I think there's like a bunch of other good ones like Netlify probably does this, and I think Amazon has a version as well.

但 Vercel 的优势在于它和 Claude 等工具有特别棒的集成、API 和 MCP 连接。

But Vercel seems to be one that has some incredible integrations and APIs and MCPs which connect to your Claude.

所以你可以把 Claude、Codex 或你用的任何编码工具无缝集成进来，部署代码变得特别简单。

And so you can have Claude or Codex or whatever you're using to code integrated really well so it's easy to deploy code.

Vercel 认为，未来属于那些坚持让 AI 模型和代理保持独立的公司。

So, Vercel is betting that the future belongs to companies that are gonna keep AI models and agents separate.

他们不把这些东西绑在同一个实验室里，而目前几乎所有公司都是这么做的。

So, you're not bundling these together by one lab, which is basically what we've seen everyone doing so far.

他们的 CEO 吉列尔莫·罗施说，客户已经在把 OpenAI、Anthropic、Gemini 和 DeepSeek 这类开源模型，混搭进同一个基础设施，把模型当成可替换的零件。

Their CEO, Guillermo Rausch, says that customers are already mixing OpenAI, Anthropic Gemini, and Open Models like DeepSeek into infrastructure that treats models as a swappable part.

他还说，谁能帮客户实现这种灵活组合，谁就能赢得市场。

And he says that companies that are facilitating that are the ones that are gonna win.

目前 Vercel 每天处理的部署量大约是六百万次。

Vercel right now, they say that they process about 6,000,000 deployments per day.

其中将近一半是由编码代理触发的，每天通过他们 AI 网关的令牌量超过一万亿。

That's roughly half triggered by coding agents and over 1,000,000,000,000 tokens flowing through their AI gateway daily.

所以说，这绝对是行业里的巨头了。

So, I mean, this really is a giant in the industry.

他们推出了 Ebe，这是一个用自然语言定义智能体的框架，同时 Vercel Sandbox 能控制内部智能体能访问哪些数据，我觉得这正好解决了很多人担心的代码泄露和训练问题。

They said that they shipped Ebe, which is a framework for defining agents in plain language, and Vercel Sandbox lets you control what data internal agents can actually access, which I think is kind of the code leaking and training problem that a lot of people are worried about.

他们把这种全新模式定位为直接与 OpenAI 等 AI 实验室竞争，因为那些实验室都在把自己的托管和部署工具直接集成到产品里。

They have framed this whole new paradigm by saying that this is putting them in direct competition with AI labs like OpenAI, which are building their own hosting and deployment tools directly into their products.

说实话，我绝对不会只用 OpenAI、Claude 或 Anthropic 这种自家完全掌控的托管平台。

Now, I will say, I would you know, I I don't think I would ever use a hosting platform exclusively owned by OpenAI or, Claude, this or, you know, Anthropic.

就像我其实也不太喜欢 Lovable 这类深度整合的服务。

The same way, I actually don't love Lovable, or other people that, like, are pretty deeply integrated.

虽然 Lovable 用的是 Supabase 当数据库，但我还是受不了它这么深的绑定。

Even though Lovable uses Supabase as the database service, I still don't like how deeply integrated it was.

我当初想从 Lovable 换出来，简直痛苦到不行，所以我相信，像 Vercel 这样独立于 AI 实验室的公司，永远会有市场——因为如果你哪天想从 OpenAI 换到 Claude，或者从 Claude 换到 Codex，你肯定希望有这种灵活性，不管用哪个编码智能体都不受影响。

For me to get off of Lovable was a huge pain, so I think there's gonna be always a place for companies like Vercel that are independent and separate from the actual labs because if you ever want to switch from OpenAI to Claude or from Claude to Codex, you wanna have that kind of flexibility where it doesn't really matter what coding agent you're using.

你得避开这种供应商锁定。

You've gotta stay away from that vendor lock in.

说到这个，如果你想在一个账号里用上 80 多种不同的 AI 模型，我强烈推荐你试试 aibox.ai。

And speaking of which, if you wanna get access to over 80 different AI models all on the same account, I love for you to try out aibox.ai.

这是我的平台。

That's my own platform.

你可以通过它访问所有类型的工具，包括音频、图像、视频，甚至音乐生成模型。

I let you access everything from audio, image, video, even music generation models.

一共有80多个。

There's over 80 of them.

当然，也包含你熟悉的那些主流模型，比如Claude、OpenAI和Gemini。

I mean, there's all the basic ones that you know like Claude and OpenAI and Gemini.

还有很多你没听说过的开源模型，都可以试试。

Tons of open source ones you haven't heard of or tried before.

上面还有一些特别酷的视频和音频类AI工具。

Some really cool video and audio powered, AI tools on there.

如果你想知道这些全部功能，每月只要8.99美元，就去aibox.ai看看吧。

So if you wanna check out all of that for only $8.99 a month, go check out aibox.ai.

我在简介里放了链接。

I'll leave a link in the description.

否则，祝大家今天剩下的时间都过得愉快。

Otherwise, I hope you all have a fantastic rest of your day.