对Anthropic在机器人领域的抱负的探讨

三星正在与机器人公司Physical Intelligence洽谈收购事宜，这让OpenAI在机器人领域的布局变得更加复杂。

Samsung held acquisition talks with Physical Intelligence, the robotics company, and this is making the OpenAI robotics push a little bit more complicated.

Meta推出了一个叫Content Seal AI水印的技术。

Meta has launched something called Content Seal AI Watermarking.

这距离他们承诺给内容加上AI标识，已经过去三年了。

This is three years after they promised to put AI labels on content.

Synthesia推出了角色扮演训练功能。

Synthesia launched role play sessions.

他们的目标是让员工通过和AI对话，提升在多个方面的沟通能力，系统还会对表现进行评估。

Their goal is to let employees talk with AI to improve their, conversations in a bunch of different areas, they get judged on it.

三星正在洽谈以200亿美元的估值投资Mistral。

Samsung is in talks to back Mistral at a €20,000,000,000 valuation.

NextEra、杜克能源等200多家公用事业公司签署了特朗普的AI电力承诺。

NextEra, Duke, and 200 other utility companies signed Trump's AI Electricity Pledge.

上周，我第一次为我的初创公司aibox.ai投放了Facebook广告，之前它一直都是靠自然增长的。

Last week, I just started running my first ever Facebook ad campaign for my startup aibox.ai, which up until now has grown completely organically.

我们准备稍微投点钱做付费推广了。

We're ready to, invest a little bit into paid marketing.

我想跟你们说说效果最好的那个广告，我觉得挺搞笑的。

The top performing ad I wanted to tell you about because I thought it was pretty funny.

它看起来就像一条新闻片段。

It's basically it looks like an a news clip.

上面写着‘科技快讯’，画面里是我一脸震惊的大图，角落里还放大了摄像头镜头的细节。

It's like says tech alert, and it's me in the kinda big banner image looking shocked, there's like, in the top corner, there's, a zoom in on, like, an image and camera lens.

广告标题写着：现在Claude也能生成图片了。

And the, headline on it says, now Claude makes images.

不需要额外下载任何应用。

No extra apps needed.

广告文案写着：让Claude帮你生成图片。

And the caption on it says, get Claude to generate images.

所有顶尖的图像模型，全都在聊天窗口里。

Every top image model right inside the chat.

不需要新应用。

No new apps.

不需要API密钥。

No API keys.

总之，我觉得这挺搞笑的。

Anyways, I thought it was funny.

我试了77种不同的广告形式和风格，这个就是77个里表现最好的。

This is basically, I've tried 77 different ad types and styles and variations, and that is the winner of 77.

所以我想，既然要跟你们说AI Box里新推出的MCP功能，能直接在Claude里生成80种AI模型，不如直接分享这个最火的Facebook广告。

So I figured if I was gonna tell you guys about the new MCP on AI Box that lets you generate 80 different AI models inside of Claude, I was gonna just tell you guys the top performing Facebook ad.

而且，这真的特别酷。

And, yeah, it's super cool.

我这一周一直在用。

I've been using it all week.

只要你想在Claude里生成图片，直接说‘Claude，生成一张x y z的图片’，把你平时给ChatGPT或其他图像生成器的提示词丢进去就行，图片就会直接出现在Claude里。

Anytime that you wanna generate an image inside of Claude, you could just say Claude generate an image of x y z, and you just basically put in the prompt that you give to ChatGPT or any other image generator, and it will pull the image straight into Claude.

Claude 随后就能直接用它。

Claude can then go use it.

我主要用它来做 Facebook 广告，这周我已经做了 77 个了。

The thing that I mostly am using it for, number one, Facebook ads because I've been creating 77 of them this week.

另外，我还用它做了好几个网站，比如我生成一个网站，大家都知道 Claude 做的网站通常都挺普通的。

Also though, I've done it with a bunch of different, websites where I'm like, hey, I generate this, you know, website and we all know what a Claude website looks like, it's pretty generic.

所以我会说：给这个网站的每个板块都生成一些专属的图片。

And so I'll say, hey, generate custom graphics for each section of this website.

有时候我还会截图，把网站现在的样子截下来。

Sometimes I'll even say, hey, go generate, you know, take a I'll I'll take a screenshot of what the website looks like.

然后我说：重新设计一下这个网站，画一张更美观的版本，它就会用图像生成器补上定制的图形，让整体看起来更精致。

And I'm like, regenerate this website, an image of this website where the design looks a lot better, and it will use an image generator to kind of fill it out with custom graphics and make it look better.

接着我会说：现在把这张图里的每个小图形单独提取出来，为每一个都生成一张新图，然后嵌回网站里。

And then I'll say, hey, now go and take each of the little custom graphics in this image and generate an image for each of those and embed it into the website.

这让我做的所有网站和应用，从原本特别普通、千篇一律，变得特别有质感、有个性。

This has basically turned my websites and apps that I've done from looking super basic and generic to looking super able or custom.

对吧？

Right?

所以呢，这个用法真的很棒。

So anyways, very cool use case.

如果你想试试，可以去 aibox.ai/mcp。

If you wanna check it out, it's aibox.ai/mcp.

我会在描述里放个链接。

I'll leave a link in the description.

说到克劳德和Anthropic，今年春天Anthropic曾与Physical Intelligence洽谈收购事宜。

Speaking of Claude and Anthropic, Anthropic held acquisition talks with Physical Intelligence this spring.

这是一家估值一千一百亿美元的机器人初创公司，但交易最终没成。不过现在Anthropic和OpenAI都在拼命追赶，想打造机器人能力，因为他们都瞄准了下一个大型IPO。

It is an $11,000,000,000 robotics startup, and the deal didn't actually close, but both Anthropic and OpenAI are now really racing to try and build robotic capabilities because they're both going for the the new big, IPO.

两家公司今年晚些时候都要上市，而机器人领域现在特别火。

They're both gonna debut on the stock market later this year, and robotics is such a hot area.

挺有意思的是，SpaceX走的是太空数据中心这条路，萨姆·奥尔特曼说这是个骗局，而埃隆·马斯克却坚持说这一定会成真。

It's kind of interesting SpaceX, you know, with the space data center route, which Sam Altman says is a scam, Elon Musk says is gonna happen.

但不管怎么讲，SpaceX当时确实拥有一个非常新颖、前所未有的竞争优势。

But whatever way you wanna cut it, it was a very novel, very new, kind of competitive advantage that SpaceX had.

我认为OpenAI和Anthropic都希望涉足实体领域。

OpenAI and Anthropic, I think both wanna get into something physical.

OpenAI已经在物理数据中心方面做了不少布局。

OpenAI has done a lot with, you know, physical data centers.

我觉得这对他们来说会非常有利。

I think that's gonna play you know, bode really well for them.

但我认为，他们俩都想去搞机器人，这将是AI发展的新阶段——机器人脑子里嵌入了AI。

But I think both of them wanna get into robotics, is kind of the new phase of AI where we're gonna have, you know, robots that have AI embedded into their brains.

我想澄清一点，虽然Physical Intelligence号称是一家机器人公司，但他们其实并没有制造机器人。

One thing I wanted to clarify is that while Physical Intelligence is a quote unquote robotics company, they don't have robots.

他们融了一笔十亿美元的资金，开发了一个非常受欢迎的AI模型，被广泛用于控制各种机器人。

They've raised a billion dollars, and they make a very popular AI model that's used for controlling a lot of robots.

这个模型在学术研究中也被广泛应用。

It's used in academic research.

它叫 PIE 0.5，这个 PIE 就是那个符号。

It's called PIE 0.5, and that's, you know, the symbol PIE.

自2023年以来，OpenAI 已经收购了17家公司，最近还在旧金山正式成立了机器人实验室。

OpenAI acquired 17 companies since 2023, and they just launched an official robotics lab in San Francisco.

Anthropic 正通过他们的安全研究团队测试机器人控制技术。

Anthropic is testing robot control through their safety research group.

这两家公司都和 Physical Intelligence 有共同的投资方，比如 Kochla Ventures、Thrive Capital 和 Founders Fund。

Both of these companies share investors in physical intelligence, including Kochla Ventures, Thrive Capital, and Founders Fund.

所以他们在机器人策略上确实有不少重叠之处。

So there is definitely some overlap in kind of their their different robotic strategies.

我觉得，要想理解物理世界，这对 OpenAI 和 Anthropic 正在推出的下一代 AI 系统来说至关重要。

I think if you understand the physical world, that's gonna be super essential for the next generation of these AI systems that OpenAI and Anthropic are both coming out with.

像这样的模型对机器人领域会非常有价值，不仅因为模型本身，更因为它能装进机器人里，去收集真实的现实世界数据。

So a model like this that can help robotics would be incredibly valuable, not just for the model, but because it has the ability to go and collect real world data when it's put inside of a robot.

现在有一些公司花钱雇人，比如 Swift 可能就是其中之一，他们直接付钱让人穿上制服，戴个带摄像头的帽子，到别人家里去打扫卫生。

You know, we have these different companies that are paying, I think, Swift might be one of them, but it's literally just paying, people to accept a robe or a a person in their home to come and clean their home while wearing a hat with a camera on it.

这些摄像头录制的视频，本质上是用来训练机器人的。

And it's essentially the camera's recording that's gonna be used for training robots.

但如果你真有一个专为机器人设计的模型，就可以把它嵌入机器人里，派它们去执行任务，让机器人自己收集数据，反过来再训练其他机器人。

But if you actually have a model that's used for the robots, then you can embed that in the robots, then you can, send them out to do tasks, you can have the robots capturing data to help train other robots.

说实话，这种应用场景多得是。

I mean, there's a lot of different interesting use cases.

不过话说回来，要是机器人自己训练，搞不好会练出错的方向。

I mean, may maybe at that point if the robots are training, they'll they'll train in a bad way.

所以现在还是得靠人类来提供训练数据。

So maybe we still wanna get the humans for the training data right now.

但不管怎样，这家公司是两家都非常看重的。

But either way, this is a really critical company that both of them are looking at.

Meta 刚刚推出了 Content Seal。

Meta has just launched Content Seal.

他们就是这个月才上线的。

They launched it, just this month.

这是一个隐形水印，用来标识用他们新推出的Muse模型生成的图片。

It is an invisible watermark that identifies images made with their new Muse model.

三年前，他们曾承诺说，嘿，

So three years ago, they promised you know, they were like, hey.

看好了，

Look.

我们会在所有发布在Facebook、Instagram这些平台上的AI生成内容上加上标签。

We're gonna get these labels on all of the different AI generated content that's on, like, Facebook and Instagram and stuff.

但这个水印只对Muse生成的图片有效，所以显然，自2023年以来Meta生成的数百万张AI图片，都无法被这个工具识别出来。

And this watermark in particular only works on Muse images, so obviously there's millions of AI photos that Meta generated since 2023, and those are not detectable by the tool.

我感觉网上有不少文章，都在呼吁Meta解决这个问题。

I think there's a bunch of different, articles online where people are kind of begging on Meta for that.

说实话，归根结底，我觉得现在大多数AI图像生成模型虽然进步很大，但还没到完美程度，而且其实很容易被看出来。

I'm gonna be honest, at the end of the day, I think most image generation models with AI, like, they're getting really good, but they're not perfect, and they're pretty I'd say they're detectable now.

我们还没进入完全无法识别的阶段。

We're not, like, past the undetectable phase.

所以很多人一直在找Meta理论，说你们当初承诺过，可2023年以来的那些图片根本没法检测出来。

So as far as, like, a lot of people are kinda begging on Meta saying, look, they they promised this, and all of these images since 2023 are undetectable.

在我看来，Meta自家的模型其实挺差的，所以我对它们以前的图像生成模型真没多大担心。

Meta's models in particular, in my opinion, are pretty bad, and so I'm not super concerned about all of their old image models.

这个新的Muse模型确实不错，我很高兴他们终于把这个功能加进来了。

This new Muse image model is really good, and so I'm happy that they're adding it here.

不过检测功能只能通过一个有访问限制的网页工具使用，并没有集成到Meta AI里，用户在和聊天机器人互动时根本看不到检测结果，Meta也没说什么时候会把检测功能加进聊天机器人。

Detection only runs through a rate limited web tool though, so it's not inside of Meta AI where users are actually seeing all the AI generated images, and Meta hasn't said when or if they're gonna add detection to their chatbot.

内容标识基本上是在模仿谷歌的Synth ID，而OpenAI早就用了这套技术，但Meta偏偏自己搞了一套标准。

Content seal is basically copying what Google did with Synth ID, which OpenAI already uses, but Meta built their own standard.

他们本来其实也不想这样，说起来还挺讽刺的。

They didn't really want I think they were all even it's actually funny.

他们明明参与了C2PA这个行业标准的制定，结果转身就自己另起炉灶。

So they were actually on the same steering committee for c two PA, which is kind of an industry wide standard, but they just went and rolled their own thing.

说到底，科技公司不都这样嘛。

I mean, tech companies are gonna do this.

Synthesia是一家我用了好多年的初创公司，早在ChatGPT出现之前我就在用了。他们周三推出了一款新产品，叫‘角色扮演对话’，能让员工跟AI虚拟形象练习对话，然后系统会根据一套评分标准给表现打分。

Synthesia, a startup that I have been using for many, many years, way before ChatGPT, launched a new product called Role Play Sessions on Wednesday, and basically it lets employees practice conversations with AI avatars, and then it's going to grade them against a rubric.

这家公司总部在英国，他们的想法是，企业愿意为‘培训真正改变了员工行为’这种证据付费，而不是只买便宜的视频片段。

So this is a company based out of The UK, and they're basically betting that enterprises are gonna pay more for proof that training actually changes behavior, not just cheaper video clips.

目前第一批客户包括欧洲市值前三的企业、全球财富100强前五的公司，还有世界最大的招聘机构之一。

So a lot of the early customers for this include one of Europe's top three firms by market value, a top five Fortune 100 firm, and one of the world's largest recruitment companies.

有意思的是，从来没人愿意公开说：‘我们用AI给全员做培训’。

It's always funny to me that, like, no one wants to put their names on, hey, we're using AI to do everyone's training.

但很明显，世界上一些最大的公司已经在用了。

But obviously, some of the biggest companies in the world are using it.

这个产品的评分功能用的是OpenAI的推理引擎，其实就是接入了OpenAI的服务。

The product uses OpenAI's reasoning engine for the actual grading, so it's just plugged into, you know, OpenAI.

Synthesia对自己的虚拟形象和语音技术都保留了 proprietary 控制权。

Synthesia keeps its avatars and voice technology proprietary.

所以在这些方面，他们竞争得很激烈。

So, you know, they're very competitive there.

他们一直以来都是做这些AI生成的虚拟形象。

That's kind of what they've always done is had these these, you know, avatars, these AI generated avatars.

他们的语音技术确实做得不错。

Voice technology, they do a great job there.

但一旦有人和这些虚拟形象做完测试，系统就会把结果发给OpenAI来打分。

But then, you know, once someone does like a test with one of these avatars, it's gonna spit it over to OpenAI to to give them a grade.

Synthesia计划从只做企业市场扩展出去。

Synthesia plans to expand from enterprise only.

接下来几个月，他们打算瞄准中小企业和学校。

They're gonna try to target small businesses and schools in the next couple months.

运行这些AI对话的成本一直在下降，尤其是当开源模型或者其他技术足够成熟之后。

The cost to run a lot of these AI conversations keeps going down, especially when you start using like open source models or other other things once they get good enough.

对吧？

Right?

他们不需要包揽所有事，对某些任务来说，成本一降下来，中小企业和学校就变成了更合适的目标客户。

They don't have to do everything, for certain tasks, the cost goes down, and so then all of a sudden small businesses and schools are a better target audience for them.

现在看来，Synthesia 正在把自己定位成一个绩效管理平台，而不仅仅是个AI视频工具。

Right now, I think this is positioning Synthesia as kind of a performance management platform, so it's not just an AI video vendor.

我一直觉得挺有意思的是，这些AI公司总会慢慢深入某个特定领域。

It's always interesting to me when these AI companies start going more deeper into one particular niche or another.

这家公司我早就用过他们的AI虚拟人了，简直用了很多年。

This is a company that, like, I I've used their AI avatars for a million years.

几年前我还拿他们家的虚拟人做播客呢。

I used to use them years ago for generating podcasts.

但现在他们彻底转型了，越来越聚焦在技能培养、领导力发展这些可量化的提升领域，还能提供数据分析。

But right now, they're making this big shift and they're going way more deep on kind of like skills and leadership and some of these measurable improvement areas and where you can, you know, get analytics.

你可以让人对着虚拟人演练特定场景中的表现，系统会评估他的领导力，或者充当销售教练。

So you can have someone go and talk through, you know, with their avatar on, what they do in a certain situation, and it will evaluate them for leadership, or it'll, you know, be like a sales coach.

他们卖的不只是虚拟人，而是基于虚拟人的评估服务。

They're selling the avatar something, and it will evaluate them on that.

说实话，这个应用场景真的很有意思。

So, honestly, very interesting use case.

三星目前正在与法国人工智能初创公司Mistral洽谈投资，估值高达200亿美元。

Samsung is currently in talks to invest in the French AI startup Mistral at a €20,000,000,000 valuation.

这比2024年年中的估值高出三倍多。

That is more than triple its value in mid twenty twenty four.

这笔投资将使Mistral成为欧洲唯一一家前沿AI实验室，如今已能与OpenAI和Anthropic抗衡。

This would make Mistral Europe's only Frontier AI lab, which is now competing with OpenAI and Anthropic.

同时，三星也将直接持有这家公司的股份，而这家公司其实早已在为三星所有Galaxy手机提供技术支持。

It would also give Samsung a direct stake in this new company that actually powers all of their Galaxy phones already.

Mistral今年六月刚完成了58亿美元的融资，领投方是通用资本。

Mistral also just raised $5,800,000,000 in June, so they raised that from General Catalyst that was leading that.

而这次200亿美元的估值，意味着不到十八个月内估值飙升了3.4倍。

And now this $20,000,000,000 valuation, it would be a 3.4 x jump in under eighteen months.

简直不可思议。

So pretty incredible.

增长势头完全跟上了美国前沿AI实验室的步伐。

Definitely on track with all of the US Frontier Lab growth rates.

三星已经在部分特定市场，将Mistral的模型集成到Galaxy设备的功能中。

Samsung already integrates Mistral models into their Galaxy device features in some specific markets.

他们想减少对谷歌Gemini的依赖，提供地区专属的选择，这点挺有意思。

They're trying to reduce reliance on Google Gemini and offer region specific choices, which is interesting.

我觉得在欧洲很多地方，这已经成了默认的选择。

I think a lot of places in Europe, this is what they're they're kind of defaulting to.

Mistral是2023年才成立的。

Mistral was just founded in 2023.

创始人来自Meta和DeepMind，专注开放权重模型和API服务。

It was a Meta and DeepMind researcher, and it focuses on open weight models and API services.

他们主要瞄准的是欧洲市场。

Really, they're focusing on kind of the European market.

他们在应对欧盟的AI法案、数据主权这些复杂问题。

They're navigating EU AI Act and data sovereignty concerns and all that kind of stuff.

近200家公用事业和数据中心公司，包括NextEra Energy和Duke Energy，都签署了特朗普的承诺，要保护家庭免受AI带来的电费上涨影响。

Nearly 200 utility and data center companies, including NextEra Energy and Duke Energy, signed Trump's pledge to shield households from AI electricity costs.

现在这已经覆盖了美国80%的电力输送。

That is now covering 80% of US power delivery.

这原本只是个自愿承诺，目的是让AI公司而不是普通民众来承担新建电厂和电网升级的费用。

It was really just this kind of voluntary commitment, and the goal with it was to make AI providers, not regular people, pay for new power plants and grid upgrades.

我觉得对这件事可以从好几个角度来理解。

I think there's a bunch of different ways to, you know, think about this.

很多数据中心项目被叫停了，或者根本推进不下去，因为公众反对声很大——尤其是当地方政府明明在大力补贴、鼓励建设数据中心的时候。

A lot of, data centers are getting kind of shut down or the data center projects are getting shut down from proceeding because it's very unpopular, especially when local governments are offering big incentives for data centers to get built and especially when they're offering energy incentives.

他们补贴了能源成本，但补贴意味着当地居民或纳税人的电费会上涨。

So, know, they're subsidizing their energy because subsidized energy just means that the taxpayer or whoever's locally living there, their energy costs go up.

所以很多人对此很不满，这个承诺不只是想绕开这个问题，更是想避免让这些项目变得更受欢迎——我们当然希望国内技术进步，但没人愿意最后为这一切买单。

So obviously a lot of people are not happy with this, and I think this is trying to not just get around to that issue but make them, you know, not make these projects more popular because we want the advancements in our country, but obviously people don't wanna get stuck paying the bill, for all of this.

最初在三月签署这项承诺的是谷歌、Meta、微软、甲骨文、OpenAI和亚马逊，后来又加入了像Equinix和Digital Realty这样的大型公用事业和数据中心公司。

The original signers from March were Google, Meta, Microsoft, Oracle, OpenAI, and Amazon, and they've expanded that list to a bunch of major utilities and data center companies like Equinox and Digital Realty.

希望这样一来能起作用，让数据中心企业不再把成本转嫁给消费者，真正解决问题。

So hopefully this makes a difference and data center providers are not gonna be passing on the cost to consumers, and this actually works out.

我觉得这件事争议挺大的。

I think there's a lot controversy on this.

有人觉得能行，有人觉得不行，但希望这项立法能顺利通过，带来好结果。

Some people think it'll work, some people don't think it will, but let's hope that this is a a good, bit of legislation that goes through.

谢谢大家今天收听我们的播客。

Guys, thank you so much for tuning in to the podcast today.

如果你喜欢这期节目，不妨去看看 AI Box MCP，想把图像、音频和视频直接接入 Claude，就用它。

If you enjoyed the episode, make sure to go check out the AI Box MCP if you wanna get AI, image, audio, and video straight into Claude.

如果你用的是 ChatGPT，他们很快就要下架图像和视频模型了。

Or if you're using something like ChatGPT, they're about to pull the image or the video model sort off of it.

如果你想把谷歌的 V3，或者任何其他 AI 模型嵌入到你的 AI 工具里，我们有 80 多个模型，MCP 都支持。

If you wanna get, Google's v o three inside of that or any AI model inside of any AI model, we have 80 plus AI models, and our MCP works with all of them.

赶紧去看看吧。

So go check it out.

描述里有链接。

There's a link in the description.

我们下一期再见。

I'll catch you guys all in the next episode.