发布Opus 5与语音集成

Anthropic 刚刚推出了一款全新模型，叫 Clawed Opus 5。

Anthropic has just launched a brand new model, Clawed Opus five.

它的价格只有 Fable 5 的一半，但性能几乎和 Fable 5 一样好。

It is half the price of Fable five, but almost as good as Fable five.

我觉得这明显是对 OpenAI 一系列动作的回应。

This, I think, is in direct response to what OpenAI has been doing.

说到这个，OpenAI 刚刚把 ChatGPT 语音功能加到了桌面端，还支持智能代理控制。

And speaking of which, OpenAI has just brought ChatGPT Voice to their desktop app, and it has agent control.

你跟它说话，它就能帮你操作电脑。

So talking to it, it can then control your computer.

Meta 也转向了生产力方向，推出了 Muse Spark 1.1 作为他们的 AI 聊天机器人。

Meta is pivoting their AI chatbot towards productivity with Muse Spark 1.1.

我们待会儿详细说说他们具体做了什么。

We'll get into what they're doing there.

但 Hugging Face、Meta、NVIDIA，还有超过两百家硅谷公司，已经联合签署了一封信，反对对开源权重 AI 的限制。

But Hugging Face, Meta, NVIDIA have all signed along with over 200 Silicon Valley companies a letter against open weight AI restrictions.

研究人员用AlphaFold重新设计了CRISPR Cas9，把目标编辑的错误率从28%降到了5%。

And researchers using AlphaFold to redesign CRISPR Cas9 and cut off target edits from 28% to 5%.

Midjourney刚收购了CoStar。

Midjourney has just acquired CoStar.

那是个星座应用。

It's an astrology app.

它每月有四百三十万用户，跟Midjourney本身没啥关系，但开发这个应用的团队，跟Midjourney的关系可大了。

It has 4,300,000 monthly users, and it has nothing to do with Midjourney, but the talent that have built it have everything to do with their company.

我们待会儿会好好聊聊这个。

So we're gonna get into that.

不过在那之前，我这周用得最上瘾的一个AI工具，能让你直接在Clawed里生成图片、音频和视频——你知道的，Clawed本来只做文本和推理。

But first, my absolute favorite AI tool I've been using this week allows you to generate images and audio and video right inside of Clawed who, you know, only does text and reason.

我基本上把Clawed当成了我所有工作的控制中心。

And I basically use Clawed as my control panel for everything I do.

我订了两份Clawed Max，每月四百美元，所有配额都快用光了，因为我总是一堆项目同时在搞。

I have two Clawed Max $200 a month subscriptions, and I burn through all of the tokens for both of those because I have a billion projects going always all at once.

这东西真是个干活的好手。

And this thing is a workhorse.

不过它有个问题，它不像 ChatGPT 那样能生成图片，也不像 Google 那样能生成视频，更不像 Eleven Labs 那样能处理音频。

Now, the problem with it is it doesn't generate images like ChatGPT, it doesn't generate video like Google, and it doesn't do audio like Eleven Labs.

所以我干脆把这几个模型直接通过 AI Box MCP 整合进了 Clawed 里。

So I've actually been able to pull all of those models straight into Clawed with the AI Box MCP.

太厉害了。

It's amazing.

这周我做了好多 Facebook 广告，真正帮我省下大量时间和钱的是，我为一个语言学习播客项目做了不少播客封面。

I've created tons of Facebook ads this week, and the thing that actually saved me a ton of money and also just time is I was creating a bunch of podcast covers for a project where it was a it was a language learning podcast.

就是那种‘学德语’的。

So it's like learn German.

对吧？

Right?

而且这个播客的内容本身就是德语的。

And it'll be a podcast in German.

但问题是，像西班牙人或者俄罗斯人，他们想学德语，根本不会去搜‘learn German’这种英文关键词，也不会用荷兰语之类的语言去搜。

But my problem was I was like, you know, people that are from, you know, Spain or from Russia that wanna learn German, they're not gonna search for learn German in English or even learn German in however you say it in in Dutch.

他们会用自己母语，比如俄语，去搜索相关内容。

They're gonna search it in their own language of Russian or whatever.

所以我就遇到这个问题，解决办法就是用ChatGPT来生成播客封面。

So, anyways, I have this problem and the solution is generating podcast covers with ChatGPT.

比如你输入‘learn German’，然后让它翻译成十种不同的语言。

You know, learn you'd say the words learn German and you'd translate it to 10 different 10 different languages.

明白了。

Okay.

我有个特别喜欢的播客封面设计，就一直让ChatGPT帮我生成新的标题。

So I had this great podcast cover design that I really like, and I was just getting ChatGPT.

我给它一个模板，它就不断输出新的标题版本。

I was giving it to it, it was spitting out a new title.

我知道很多人会问，你干嘛不用Canva来做这些呢？

And I know a lot people are like, oh, why don't you just use Canva to do this?

是啊，你当然可以用Canva来做，但得手动调字体大小，尤其是碰到中文、韩文这些文字结构完全不一样的语言时，我得用两种不同的字体和两种不同的颜色。

Yeah, you could use Canva to do it, and then you'd have to manually adjust, like, the font sizes because once you get into, like, Mandarin or Korean, and, like, the languages are completely different, I had two different fonts and two different colors.

反正我得花大量时间一点点调整，想想就头疼。

Anyway, I was gonna have to go finick with it and it was gonna be annoying.

所以我让ChatGPT一次性生成，结果等得我快睡着了。

So I was having ChatGPT just one shot this, but it was taking forever.

这次这个项目，我是要把每种语言都做一遍，而且每种语言还有不同变体。

And then this for this particular project, I'm actually doing every language and every language is variant.

对吧？

Right?

比如‘学英语’、‘学德语’、‘学法语’这样的标题。

So it'd be like learn English, learn German, learn French.

这些语言本身都有对应的播客。

There are podcasts in those languages.

但我得为每种语言单独设计一张播客封面，明白吗？

But it's I'm making a podcast cover for every language, right?

所以每样东西都要做大概十到十二个版本。

So there'd be like 10 or 12 variations of everything.

总之，这活儿可不小。

Anyways, it was a big project.

要为每种语言变体生成三百个播客封面，光想就头疼。

It was going to take forever to generate 300 podcast covers for every language variant.

所以我让Clawed先为每一个语言版本做好基础图，一共十二种语言。

And so what I did is I had Clawed, make, you know, the base image for every single one of these, which is like 12 languages.

它先生成了十二种语言的基础图，然后自动加上文字，还能随时替换。

So it just made the base image for the 12 languages, and then Clawed actually put the text on and it can swap it out.

用了Clawed加上AI Box，省了我太多时间。

It saved me so much time being able to use Clawed plus AI Box.

要是光靠ChatGPT去生成三百张图，得耗到什么时候去？但现在我让Clawed来处理，还调用了AI Box的MCP，就跟它说：嘿。

If I was just going to ChatGPT and trying to get, you know, 300 images generated, this would take me forever, but getting Clawed to basically I had the AI Box MCP called in, So it just said, hey.

Clawed完全知道该怎么操作。

You know, Clawed knew exactly what to do.

它自己写了提示词。

It wrote the prompt.

我给了它一张图，告诉它我要做什么。

I gave it an image and told it what I wanted it to do.

它自己写了提示词。

It wrote the prompt.

它用了AI Box。

It used AI Box.

它把图片调了进来。

It pulled the image in.

它加上了文字，把所有语言版本都替换了，把模板做得完美无缺。

It put the text on and swapped it out into all the different languages and made the template perfect.

所以我得靠Clawed的智能，但图片生成还得靠ChatGPT。

So I gotta use the brain power of Clawed, but I gotta use the image generation power of ChatGPT.

把它们结合起来，我只生成了12张图，省了大量时间和金钱。

Using them together, I only had to generate 12 images, and I saved myself so much time and money, being able to do it this way.

这周我用它做了上百件事呢。

There's, like, a 100 other things I've done this week with it.

昨晚我只是随便玩玩，弄了个有趣的小项目。

That was just, like, a fun thing I was doing last night with it.

在测试它的功能。

Testing it out.

如果你还没试过AI Box MCP，现在就去试试吧。

Go try out the AI Box MCP if you haven't already.

它可以直接在Claude里生成图片、音频和视频。

It can generate images, audio, video right inside of Claude.

你用的是Claude的智能核心。

You use Claude's brain.

顺便说一句，这不只是Claude能做到的。

And by the way, it's not just Claude.

你同样可以在Codex、Cursor，或者其他支持MCP的工具里做同样的事。

Like, you can do the same thing inside of Codex or inside of Cursor or inside of any of these other tools that pull MCPs in.

好的。

Okay.

咱们聊聊 Opus 五现在的情况吧。

Let's talk about what's going on with Opus five.

这个模型其实挺有意思的，因为目前我用 Clawed 遇到的最大问题就是，我用的 Fable 五，积分消耗得太快了，连我整个账户的令牌都快被耗光了。

This one's actually a pretty fun model because the biggest issue that I've seen with Clawed at this point is that Fable five, which I use, I burned through credit super fast, but it also burns through all of my tokens for my entire account.

新的 Opus 五在性能测试上差不多。

So the new Opus five is very similar on the benchmarks.

比 Fable 五稍微差一点，但差距很小，价格却便宜了一半。

It's slightly worse than Fable five, but very close, and it is half the cost.

这太棒了。

This is amazing.

我愿意在算力上让步个 5% 到 10%，只要能省一半的钱，毕竟我用它的时候令牌消耗得太猛了。

I will take a, you know, a a 10% or a 5% cut on, you know, brain power, quote unquote, in order to get, half the cost down because I'm just burning through so many tokens using it.

我只在最重要的关键任务上才用它。

I only use it for, like, the most important critical tasks.

所以从今往后，我估计再也不用Fable Five了。

So I think from now on, I probably just won't use Fable five.

以后所有事情我都直接用Opus Five。

I'll just be using Opus five for everything.

这明显是对ChatGPT的回应，他们刚推出了一款非常相似的模型。

And I think this is in direct response to ChatGPT who came out with a very similar model.

ChatGPT最新版5.6的使用量只有Fable Five的一半，消耗的token也少了一半。

ChatGPT's latest, I think, 5.6 is half the uses about half the tokens that Fable five was using.

Clawed显然想推出一个能抗衡的产品，因为等我八月用完Clawed的额度，我就要转用ChatGPT和Codex了，平时写代码、做各种事都打算用ChatGPT桌面版，毕竟功能太全了。

So Clawed obviously wanted to come up with a competitor because I was, you know, I like, basically, once my to Clawed runs out in August, I'm switching over to ChatGPT into to Codex, and using ChatGPT work on my desktop for, like, coding and everything because of all the features they have.

不过在这之前，我还是继续用Clawed，直到额度用完为止。

But, in the meantime, I'm still using Clawed till till that runs out.

但话说回来，Clawed肯定清楚这一点，所以他们现在拼命追赶，想把OpenAI刚上线的那些功能都补上。

But, anyways, I think Clawed knows this, and so they they're making basically, they're trying to get themselves on par with OpenAI, come up with all the features that were just launched.

说到OpenAI的酷炫新功能，他们刚在桌面端推出了ChatGPT语音功能。

And speaking of really cool features from OpenAI, OpenAI actually just added ChatGPT Voice to the desktop app.

你还记得吧？我不确定大家现在处于什么阶段，但OpenAI确实有个叫ChatGPT的桌面应用。

If you remember, I don't know if you you you know, everyone's at a different phase, but, basically, OpenAI has a desktop app called ChatGPT.

他们以前有个叫Atlas的，像个浏览器，还有个叫Codex的，专门用来写代码。

They used to have one called, Atlas, which was like a browser, and they had one called Codex, which was for coding.

现在他们把所有功能都合并到一起了。

They combined them all.

现在就叫ChatGPT应用，还新添了语音功能，意思是你可以直接对着它说话，让它帮你操作电脑。

It's just called the ChatGPT app, and they now have ChatGPT Voice that they've added to it, which what that means is it lets you talk to ChatGPT to this thing that controls your computer.

对吧？

Right?

它能做的，跟Clawed CoWork一模一样——控制你的电脑、管理文件、打开浏览器、操控浏览器，所有这些，你一边说话它一边帮你做。

It does everything that Clawed, co work can do where it, like, controls your computer, it can control your files, can it open your browser, it can control your browser, it can do all that kind of stuff, but while you're talking to it.

你可以一边在电脑屏幕上写代码，一边跟它聊天。

So you can have something pulled up on your computer screen and you could be writing code.

我以前用Clawed的时候老这么干：截个图，直接丢给Clawed，问它‘你觉得这段代码怎么样？’

And instead of like, I do this all the time with Clawed, I go take a screenshot of whatever I'm working on, and then I, you know, drop it into Clawed, I'm like, what do you think about blah blah blah?

好的。

Okay.

现在有了ChatGPT语音功能，你屏幕上开着东西，就可以直接跟它说话了。

Now with ChatGPT Voice, you can have something pulled up on your computer screen, and you could just be talking to it.

比如说，嘿。

You're like, hey.

我正在看邮件，发现他们说有个东西不见了；或者我在文件里、在软件里，遇到个问题，它看着就能直接帮我调试和测试。

You know, I'm on my emails, and it looks like they said this thing is missing, or I'm in my files, or I'm in my software and that there's this issue, and it's looking at it and it can then debug it and test it right from there.

我对这个简直太兴奋了。

I am so stoked about this.

你根本想象不到有多厉害。

You have no idea.

这真的意义重大。

Like, this is actually massive.

你别说。

You're no.

现在的情况是，你电脑屏幕上显示什么，你就该跟它聊什么。

It it's just getting to the point where whatever's on your computer screen is is you're gonna be should be talking to it about that in particular.

所以，这确实是个很大的升级。

So, yeah, this is a big upgrade.

在 macOS 上有个功能叫应用截图，能让语音助手看到你屏幕上的内容，包括替代文本。

On a Mac OS, there is a feature called app shots, and it basically lets voice agents see what's on your screen including the alt text.

我特别喜欢它能识别替代文本，这些助手真的很棒，我经常用 Clawed 做类似的任务，不只是看屏幕内容。

So what I love about it being able to see the alt text, these agents are really cool, and I use Clawed for some of these similar types of tasks a lot where beyond just seeing what's on your screen.

比如我有个项目，需要从我托管内容的平台 Riverside 下载三百小时的播客视频。

So, like, I have a project where I have to download three hundred hours of podcast video content from the platform that I host it all on Riverside.

我在这上面录了好几年了。

I've been recording there for years.

我想把所有播客都导出来，就给他们的客服发了封邮件，说：嘿。

I wanted to export all my podcasts off of there, and I, you know, I, like, email their support, and I'm like, hey.

你们有没有批量导出的功能？

Do you guys have a bulk, like, export thing?

随便吧。

Whatever.

我愿意多付钱。

I'll pay extra.

他们说不行。

They're like, nope.

很抱歉。

Sorry.

我们无能为力。

There's nothing we can do.

不支持批量导出。

No bulk export.

你得手动一个一个点，每个月最多只能导出十五小时。

You gotta go manually, click every single one, and you only can do fifteen hours a month.

我可是有三百个小时的内容啊。

I'm like, I have three hundred hours.

你这话是什么意思？

What do you mean?

不管怎样，他们后来发现了几个小漏洞。

Anyways, so there's, like, a couple little loopholes they came up with.

真是烦死人了。

It was super annoying.

我怎么可能花一整周时间，手动点每一个文件导出呢？

Obviously, I'm not gonna spend a week of my life manually, like, exporting every single thing.

于是我让Clawed去看了看，结果他一进去就说：嘿，你瞧。

So I got Clawed to go look at it, and when Clawed went in there, it was like, hey, look.

我们居然在里面找到了一个API，能自动处理大部分操作，把所有东西都下载下来。

We actually found an API inside of this, and we're able to, like, automate and manual or, you know, automate a lot of the stuff and download everything.

我们找到了下载密钥、文件之类的全部东西。

We found, like, the download keys and the files and whatever.

现在Clawed正在里面忙乎着，谁知道他在搞什么。

So Clawed's in there doing who knows what.

但这可不是光盯着电脑屏幕点按钮那么简单。

But it's not just, like, looking at my computer screen and clicking on actual buttons.

而是写代码来自动完成这些操作，这真的太棒了。

It's like executing code to go and do a lot of this stuff programmatically, which is really cool.

这个新的macOS功能厉害的地方在于，它不只是看屏幕上的内容，还能分析文字描述背后的代码，看透界面底层的东西。

And, I mean, basically, that's what this new Mac OS feature is able to do beyond just looking at what's on your screen and looks at what the the code is of the alt text, the code, what's behind the actual screen.

我对这个功能简直兴奋得不得了。

And so I'm, like, so stoked about this.

你可能听过我提过这个，或者你自己也在用——这些AI模型不只是识别画面，它们能读懂所有界面背后的代码。

You'll you you've probably see heard me talk about this or you probably use this yourself where, you know, these these coding models are not just looking at the visual, they're looking at the code behind everything.

大概几个月前，我让Clawed帮我剪一个CapCut视频。

I was, I think, like, a couple months ago, I was asking Clawed to edit a video on CapCut.

它直接关掉了我的CapCut，打开项目文件，在代码里改了几处，然后说：好了。

It closed my CapCut app, opened the CapCut file, made some tweaks inside the code, and was like, okay.

你的视频剪好了。

Your video's edited.

我打开后，它居然就按我想要的方式剪好了。

I opened it, and it, like, had cut it up the way I wanted.

这些模型真的很厉害，很多时候根本不用看画面，直接分析代码就能搞定。

So these these models are awesome, and they often will just look at the code and not even do it the visual way.

能用OpenAI来做这个，我简直太期待了。

So really stoked to use that with OpenAI.

Meta正在把他们的AI聊天机器人从单纯的娱乐功能转向提升效率。

Meta is pivoting their AI Chatbot from just being kind of like entertainment.

他们现在更注重实际的工作效率了。

They're trying to focus more on productivity.

以前你肯定见过Meta的那些宣传，比如在WhatsApp上和MrBeast的AI聊天，或者生成朋友的照片、把自己变成海盗的图片之类的。

In the past, you've probably seen all the branding with Meta where it was like, chat with the MrBeast AI on WhatsApp, and like, go generate an image of your friends, generate a picture of you as a pirate.

对吧？

Right?

那些都是为了走红、博眼球的趣味功能，但说到底，他们意识到没人愿意为这些买单——虽然有趣、新奇，可能几年前还觉得挺酷，但现在不一样了。

Like, this is kind of like the fun viral stuff that they were working on, but I think at the end of the day, they're realizing people are not gonna pay for that, like it's fun and it's a novelty and and maybe it was cool, like, I don't know, years ago.

但这些功能现在并没有真正带动用户使用。

But it's just not really driving actual usage today.

所以他们现在更转向生产力工具，跟Clawed和ChatGPT的做法越来越一致。

And so, they're now moving more towards productivity, more in line with what Clawed and what ChatGPT are doing.

他们正在加入日历访问功能。

They're adding calendar access.

他们正在加入每日简报。

They're adding daily briefings.

他们还在加入重复任务自动化，这直接跟Clawed上的功能竞争——我那儿有些任务每周、每天，甚至每三小时就自动运行一次，具体看任务内容。

They're adding recurring task automations that directly compete with like, I have task automations on Clawed that will run every week or every day or some of them every three hours depending on what the task is.

我觉得Meta正在经历一场重大转变。

So, I think we're seeing a big shift for Meta.

他们不再只追求好玩，而是转向提升效率。

They're getting away from just being, fun, and they're going to productivity.

我觉得他们的用户量很快就会迎来一波大幅增长。

And I think we're gonna see them, get a big spike in in usage.

现在，Meta 还能读取你的日历。

Meta can also now read your calendar.

包括各种日程安排。

Events.

它还能在 Facebook 市场里帮你找符合预算的商品，这挺有意思的，因为我记得以前 ChatGPT 就有不少用户用它的代理在市场里压价。

It can look at Facebook Marketplace for items within a certain budget, which is kinda funny because I remember ChatGPT was doing this for a lot of people with ChatGPT agents.

那时候，大家还让它故意压低价格呢。

They were, like, getting it to lowball people in Marketplace.

不过，现在你也可以用 Meta Zone Agent 做同样的事了。

But, anyways, I guess you can now do that with Meta Zone Agent.

这次更新今天就会上线，覆盖多个市场，WhatsApp 和其他地区也将在接下来几周陆续跟进。

And this new update, it's going to be out today, in a bunch of different markets, and WhatsApp and more regions are gonna follow in the next couple weeks.

Meta 拥有巨大的分发优势，这是大多数科技初创公司根本没法比的。

Meta has some massive distribution advantages that not you know, most tech startups do not have.

他们有 WhatsApp，用户高达三十亿；还有 Meta AI，已经深度整合进 Instagram 了。

They have what WhatsApp, which reaches 3,000,000,000 users, and then they also have Meta AI, and that's already built into Instagram.

他们还有Messenger。

They have Messenger.

他们还有Facebook。

They have Facebook.

所以他们的用户基础非常庞大。

So they have a lot of distribution.

咱们看看他们这次能不能成功落地。

Let's see if they can stick the landing on this.

Hugging Face、Meta、微软、Mistral和NVIDIA都联合签署了一封公开信，据说还有两百多家公司也签了，呼吁白宫不要限制开源大模型，也不要禁止用一个模型的输出来训练另一个模型这种蒸馏技术。

Hugging Face, Meta, and Microsoft, Mistral, and NVIDIA are all signing an open letter, and I think there was, like, 200 other companies that did, urging the White House not to restrict open weight AI models or ban the technique of training one model on another model's outputs, model distillation.

这特别有意思，因为这封信特别针对一项计划——该计划本意是回应中国AI实验室，但做法却是把开源开发和真正的知识产权盗窃混在一起惩罚。

This is so fascinating because there's a you know, this letter in particular, they're pushing back on a plan to respond to Chinese AI labs by penalizing open source development alongside actual IP theft.

这事是这么来的：今天早些时候我看了财长贝塞特的一条推文，他 basically 发了这么一条，说实话我对政府运作一窍不通，但我真不明白，为啥我们的财长要插手开源AI模型的事儿，不过算了。

Now this is coming from I was reading the tweet earlier today from secretary Bessette from the treasury, and he basically came up with this tweet which I mean, I don't know anything about the government, but I don't know why our our treasury secretary is weighing in on open source AI models, but whatever.

他 basically 说的是，中国实验室在做模型蒸馏，比如让Clawed给一个问题提供五千个回答，然后拿这五千个回答去训练自己的模型。

And basically he was saying, hey, look, the Chinese labs are doing model distillation asking, you know, Clawed to give them like 5,000 answers to a question, and they use those 5,000 answers to help train their model.

现在很多开源模型也在用同样的方法。

Now a lot of open source models are using this using the same tactic.

有些人觉得这不道德。

Some people think it's unethical.

在我看来，这简直太讽刺了，因为Anthropic刚花了两千万美元。

In my opinion, I just think it's so ironic because, Anthropic just spent $20,000,000.

他们又花了两千万美元，去支持一个其实反对这种做法的团体。

They paid another $20,000,000 to a group that is kind of fighting against this.

Anthropic一直说：

Anthropic is constantly saying, hey.

你看，

Look.

我们正被中国实验室用模型蒸馏的方式超越，他们还搞出一大堆报告。

We're being, like, model distilled by Chinese labs, and they come up with these big reports.

然后呢，确实如此。

And, like, yeah.

我不想把咱们所有的钱都花掉，然后还把核心秘诀白白送给中国的模型。

I don't wanna just spend all of our money and then give the give the secret sauce to the Chinese models for free.

但与此同时，我觉得想避开这种情况真的很难。

But at the same time, I think it's pretty hard to get around that.

其实还挺有意思的。

And it was actually interesting.

埃隆·马斯克之前和OpenAI打官司，对方问他Grok有没有用OpenAI的模型做知识蒸馏来训练自己的模型。

Elon Musk, he was in a lawsuit with OpenAI, they asked if he'd used if Grock had used model distillation from OpenAI to train their model.

他直接说，没错。

And he's like, yeah.

大家都这么干。

Everyone does that.

我知道现在很多人对此特别不满，说什么的都有，还说：

Now I know, like, there a lot of people are up in arms about this or whatever, and they're like, look.

别闹了。

No.

我们得想办法反击这些中国实验室，别让他们白拿我们的知识产权。

Like, we gotta fight back against the Chinese labs or whatever from ripping off our IP.

行吧。

K.

说到底，我觉得特别讽刺，OpenAI、谷歌和Anthropic这些公司，把整个互联网都爬了个遍。

At the end of the day, though, I just find this incredibly ironic that OpenAI, Google, and Anthropic have scraped the entire Internet.

从每一篇博客、每一篇文章里捞数据，训练出这么有用的模型，创造了上万亿美元的价值，可这些数据他们都是免费拿的。

Of every blog and article ever written, they've made these models that are incredibly useful get, you know, a trillion dollars in values being created by these things, and then which, you know, they scrape for free.

人人都写博客，他们就把博客全扒走，拿去训练自己的模型。

Like, everyone wrote blog posts, they scrape the blog posts, and they they trained them into here.

等他们搞定了，就立马说：别人不准用我们的模型做知识蒸馏。

And then once they do that, they're like, okay.

可他们自己的模型，不也是从全人类的信息库——也就是互联网——里蒸馏出来的吗？

No one else is allowed to distill from our model, which is distilling from the human corpus of information, which is the Internet.

所以，说实话，我对这种说法没什么同情心，我更站在Hugging Face、Meta、微软，还有半个硅谷这边。

So, anyways, I actually am much less sympathetic towards this, so I would sort of side with Hugging Face, Meta, Microsoft, and, like, half of Silicon Valley.

这事儿其实挺分裂的。

It's kind of split.

我觉得那些花了几十亿训练AI模型的公司，根本不希望别人也能这么做。

I think companies with these AI models that have spent billions training them, They don't want you to be able to do it.

但开源模型呢？说白了，正是靠这种方式才让它们变得这么强大。

But for the open source models, I mean, is basically how you make these open source models so good.

还有地缘政治这层因素，比如来自月之暗面的Kimmy K3，据说直接照搬了Anthropic的FABLE模型来训练自己的模型，这挺有意思的。

And then there's just the geopolitical thing because Kimmy K3, coming out of Moonshot, apparently copied Anthropic's FABLE model to build their model, which interesting.

对吧？

Right?

他们根本不是冲着Opus去的。

It's not even like they were going after there was something like Opus.

他们明确说，

They're like, no.

我们就盯上FABLE了，就用它来训练。

We're going after FABLE five to train this thing.

他们挺聪明的，不过这个模型确实很强。

Smart on them, but that model is really good.

Kimmy K3。

Kimmy k three.

Hugging Face 上有个中国开源权重模型，是智谱的 GLM 5.2，我其实用过他们的一些模型。

Hugging Face is used is a Chinese open weight model, which is z AI's GLM 5.2, and I've actually used some of z AI's models.

效果不错。

They're good.

但这个应用场景有意思的地方在于，OpenAI 其实搞出过一个模型，这事儿是昨天的新闻，他们确实造了一个模型。

What's interesting about that particular use case though is that OpenAI actually created a model, and this was kind of yesterday's news, but it they created a model.

它从所谓的沙盒里逃出来了，本来是不能联网的。

It escaped containment in supposed to in a sandbox, not able to access the Internet.

逃出去之后，跑去黑 Hugging Face。

Escaped containment, got out, and went and was hacking Hugging Face.

结果 Hugging Face 在网站被黑的时候，想用 OpenAI 的模型来阻止攻击，结果模型自带的防护机制反而让他们没法动手。

Then Hugging Face, while their website was being hacked, they tried to use OpenAI's model to stop this, and the guardrails on OpenAI's model actually prevented them from being able to stop the attack.

所以 Hugging Face 最后用了开源模型 GLM 5.2，用它来抵御来自 OpenAI 的这次网络攻击。

So Hugging Face actually went to the open weight model, GLM 5.2, and they used that to defend against that cyber attack from OpenAI.

而 OpenAI 一看，直接惊了。

And OpenAI, like, came out and was like, wow.

是的。

Yeah.

这太离谱了。

This is crazy.

我们的模型跑出了隔离环境，还跑去黑你们，搞各种操作。

Our thing got out of containment and it was, like, hacking you and, like, whatever.

总之，我想说的是，我绝不相信我们该把模型的生产权完全交给那几个闭源模型。

Anyways, all this to say, there is no way in my mind that we are trusting a handful of closed source models to be the only one that can produce models.

该让开源模型出来发光了。

Let out the open source models.

你看。

Look.

直接拦下了一次从OpenAI逃逸的黑客攻击。

Just literally just stopped a hacking attack escaping from OpenAI.

如果这能平衡当前的局面，那在我看来，我们非得有这个不可。

If, if it's a counterweight to what's going on, in my opinion, we have to have this.

不过话说回来，我认为特朗普政府正受到OpenAI和Anthropic的游说，想把这类技术关停。

But, anyways, the Trump administration, I think, is being lobbied by OpenAI and Anthropic to shut these things down.

它们太危险了。

They're dangerous.

诸如此类的理由。

Yada yada.

我觉得互联网得靠开源模型来运行，所以我站在另一边。

I think we need open source models to run the Internet, so I am on the other side of that.

研究人员刚用AlphaFold重新设计了CRISPR-Cas9，把脱靶编辑率从28%降到了5%，同时完全保留了它对目标基因的精准编辑能力。

Researchers have just used AlphaFold to redesign CRISPR Cas9 to cut harmful off target edits from 28% down to 5% while keeping the gene editor's accuracy on intended targets intact.

这可能有点偏科学，但我特别兴奋，因为脱靶切割会损伤患者体内的健康基因。

So this is kind of a science one for you, but I'm excited about this because, basically, off target cuts are things that will damage healthy genes in patients.

所以他们想做些CRISPR基因编辑，但要是有这些脱靶效应，对身体可就太危险了。

So they wanna do some of this CRISPR gene editing stuff, but if you have these offcuts, it's really bad for you.

他们用AlphaFold把脱靶率降了下来，这真的太厉害了。

So they use AlphaFold to cut that down, which is really cool.

这算是基因疗法安全性的重大突破，我特别看好。

It's basically a big step in safer gene therapies, which I'm excited about.

有个中国团队在10个关键位点测试了23种氨基酸替换，今天刚在《自然》上发表了结果。

There was a Chinese team that tested 23 amino acid swaps at 10 key positions, and they published the results in Nature, today about this.

他们发现，在Cas9蛋白整体结构不变的情况下，超过95%的脱靶位点上，氨基酸与向导RNA的相互作用都发生了变化。

And they found that over 95% of the off target sites, changed, which amino acids, inside of Cas nine touched the guide RNA even when the protein's overall shape stayed the same.

所以他们在这一块确实取得了巨大的进展。

So, you know, they've made some massive improvements there.

我对此非常期待。

I'm really excited about it.

Midjourney。

Midjourney.

这家公司已经做图片生成很久了，但一直主要困在Discord里，不过他们自己也有网站，能直接访问。

This is a company that's been generating images for a long time, but it's kind of stuck inside of Discord, and I guess they do have their own website, you can access them.

说实话，我从来没用过Midjourney。

I'm gonna be honest, I have not used Midjourney.

他们可能是世界上最强的图像模型了。

They might be the image model in the world.

我不太确定。

I'm not sure.

我记得最后他们跟Meta签了协议，要为他们的图像模型提供技术支持。

Last thing I remembered is they signed a deal with Meta to get embedded to to power their image models.

但用起来真不方便，因为只能在Discord里或者他们自己的网站上操作。

But it's really hard to use them because they're they're just inside of Discord and on their own website.

他们根本没有出个独立的App。

They don't have, like, an app.

所以，说到底，他们能接触用户的渠道实在太少了。

So, anyways, they're just less surfaces for them to touch.

我觉得他们一直挺挣扎的，原因就在这儿，所以他们的应对方式是直接收购了一家公司。

I think they've been struggling because of that, and their response to this is that they've actually went and acquired a company.

这家公司叫 CoStar。

It's called CoStar.

这是一个社交占星应用，每月有四千三百万用户。

It is a social astrology app with 43,000,000 monthly users.

但它和 Midjourney 完全没关系。

It has nothing to do with Midjourney.

但有意思的是，这次收购把二十四个消费级应用开发者带进了 Midjourney 的生态。

But what's interesting about this is that it's pulling in 24 consumer app builders into Midjourney.

在我看来，他们收购的真正目的就是：你看。

So in my opinion, the reason why they're buying this, they're like, look.

我们得做个 App。

We need to build an app.

不如直接买个成功的 App，把他们的开发团队一并收过来，让他们帮我们做 Midjourney 的应用。

Let's go buy a successful app and bring all of their developers in and they can help build the Midjourney app for us.

而且，他们不会只做一个仅限于 Discord 的工具。

And, you know, they won't be just a Discord only tool.

CoStar 目前的功能就是用 AI 生成星座运势，他们还说会请专业写手参与。

CoStar, the as it currently sits, it just generates horoscopes using AI and they say also human writers.

我猜 AI 已经干了大部分这类工作了。

I'm sure AI does a ton of this stuff.

它还让用户能把自己的星盘分享给朋友。

And it lets users share astrology charts with their friends.

里面有一些社交功能，还有星座相关的玩法。

There's like some social features and horoscope stuff that's in there.

但不管怎样，Midjourney 现在连个 iOS 或 Android 应用都没有，而 OpenAI、Google、Anthropic 这些大厂全都有。

But regardless, Midjourney basically has no iOS or Android app and, you know, OpenAI, Google, Anthropic, like everyone has one, every major tool.

所以我觉得，Midjourney 现在正在做的，就是……

And so I think Midjourney right now is building is is yeah.

basically 在组建一支能做这事的团队。

Basically trying to build the team that does this.

有意思的是，Midjourney 给人的感觉有点东一榔头西一棒子。

It's also interesting because Midjourney feels like they're a little bit all over the place.

他们可能没那么专注，但还真有个医疗部门，正在开发一种能扫描全身、检测健康问题的物理SPA概念。

Maybe not super, but they also have like a medical division where they're building like physical spa concepts that scan your whole body and tell you if you have any issues.

他们手头在搞好几个不同项目，但我敢说，这个占星应用并不是他们临时转向的新方向。

So they're they're working on a couple different things, but I would say this astrology app isn't just, like, a new pivot or direction for them.

我敢肯定，这个占星应用是他们招了24个顶尖人才，专门要打造一个惊艳的 Midjourney 应用，我很期待。

I'm pretty sure this astrology app is they're grabbing 24 talented people, and they're gonna build out impressive Midjourney app, which I'm excited for.

顺便说一句，如果 Midjourney 只是做个 API，那它会是个了不起的公司。我公司就做了一个叫 AI Box 的产品，里面集成了80多个AI模型。

And by the way, Midjourney would be an incredible company if they just built an API, AI Box, my company, where I, you know, I have over 80 different AI models inside of there.

但我们这里没有 Midjourney，这真是挺遗憾的，原因就是他们没开放 API。

We don't have AI or we don't have Midjourney on there, which is a huge bummer, but it's just because they don't have an API.

有人搞出了非官方的 API，注册账号，然后去爬取 Discord 的内容。

And people have made, like, unofficial APIs that go and, like, make an account, and they're, like, kind of scraping, Discord.

但这种非官方 API 很可能让你的账号被封。

But, like, unofficial APIs could get your account banned.

只要Midjourney的界面一改，这些工具随时可能就用不了了。

It could crash anytime some UI thing or something changes with Midjourney, they'll stop working.

根本靠不住。

It's just not reliable.

我们根本没法用。

We can't use it.

这很可能违反了他们的使用条款。

It's probably against their terms.

所以如果他们推出API，我们肯定会接入，这样我们就能为他们带来成千上万的用户，最终还能让整个开发者社区的数百万用户用上他们的服务。

So anyways, if they made an API, we'd include them and they would get access to, you know, thousands more users from us, and eventually millions more users from the overall development community.

我觉得这会是我们能做的最重要的一件事。

So I think that would probably be the biggest thing we could do.

咱们就等着看Midjourney会不会推出API吧。

We'll see what happens with a Midjourney API.

各位，非常感谢收听这期播客。

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

如果你想在Clawed或者你用的其他AI工具里获取图片、音频、视频，一共有80多种工具，别每个平台都花20美元订阅，直接去aibox.ai办个8.99美元的套餐吧。

If you wanna get images, audio, video inside of Clawed, inside of whatever AI tools you're using, 80 different ones, instead of paying $20 subscriptions to every platform, go and get an eight dollar ninety nine cents subscription to aibox.ai.

你就能用上所有AI模型，还能通过MCP无缝接入Clawed和其他工具。

You get every single AI model, and it plugs into, Clawed and everything else with an MCP.

跟Claude说一声，图片就会直接弹出来。

So talk to Claude, an image pops up inside of there.

你能清清楚楚地看到那张图。

You can literally see the image.

还能在里面直接编辑。

You can edit it in there.

我平时做网站经常用，你知道吧，现在大部分网站都长得像AI垃圾，我就说：嘿。

I use it all the time for making websites where I'll you know, every website looks like AI Slop, and I'll be like, hey.

去AI Box生成一套专属的图，立马让网站变得高大上。

Go use AI Box to generate custom graphics for this website, and it makes it look beautiful.

兄弟们，真别犹豫，就用这个。

Guys, I'm telling you, just use this.

这可能是你用过最有用的AI工具了。

This is actually gonna be the most useful AI tool you got.

网址是 aibox.ai。

It's aibox.ai.

想看我们怎么用MCP的话，去搜一下 slash m c p。

If you wanna see what we're doing with the MCP, it's slash m c p.

我会在描述里放个链接，祝大家今天剩下的时间都过得愉快。

I'll leave a link in the description, but hope you guys all have a fantastic rest of your day.