蒂姆·库克之后的苹果，OpenAI 的新 Mojo，Meta 的内部追踪闹剧

你觉得呢？

What do you think?

行吧。

Okay.

你知道吗？

You know what?

我得先降温了。

I'm already having to back off.

我本来还挺期待萨姆这次发布的新面孔的，但话说回来，我们支持你，也希望你能成功。

I was getting excited about Sam's new face in this launch, but we love you and want you to win.

得了吧。

Come on.

我整个职业生涯，说白了就是围着创业公司的魔力转。

My whole career has largely been about the magic of startups.

我觉得该说的都说了。

Like, I think enough has come out.

你觉得大家会认为这是真心的吗？

Do you think people are gonna take this as sincere?

肯定不少人会信，但你觉得他真能一直保持这种‘我们爱你，希望你赢’的真诚态度吗？

I'm sure plenty of people will, but do you think, like, he's gonna be able to maintain this sincerity of we love you and want you to win?

我还注意到，记得他回复完一位Anthropic的工程师后，发了条‘哦，老古董’，然后转推说：‘今晚我喝了点酒’，还把‘今晚’拼错了。

I also noted, and I remember there's this one after replying to an anthropic engineer, okay boomer, he quote tweeted, tonight, I have had a couple of drinks, misspelling tonight.

我其实去查了下，萨姆·阿尔特曼平时喝酒吗？

I actually looked up, like, does Sam Altman drink?

结果发现他公开说过很多次，为了优化睡眠，他不喝酒。

And it's saying he's had a lot of public statements about for sleep optimization.

他基本不碰酒精，偶尔可能破例一次。

He does not consume alcohol and then very rarely might.

所以，他到底是真喝微醺了在那儿发推，还是说——我现在再看，这感觉有点假了？

So, again, is he actually just sitting there a little tipsy tweeting, or is this now now that I'm looking at it, it's feeling a little bit insincere.

这招也许能奏效，但也可能适得其反。

It might work, but it might backfire too.

行，我懂了。

Well, okay.

那我们来聊聊这个事儿，说到底核心就是Mythos对阵Spud，或者说Mythos对阵5.5版本的冲突。

So let's talk about this because ultimately what this comes down to is Mythos versus Spud or Mythos versus 5.5.

我挺好奇你觉得正确的处理方式是什么，我也一直在琢磨这个问题。

And and I I I think that I'm curious what you think the right approach is.

我完全能理解Anthropic的立场，他们的顾虑就是自家的这个模型确实能被用来发动网络攻击。

I totally hear Anthropic's perspective on this, which is like, we know that this thing can do cyber attacks.

我们会和一批信任的合作伙伴一起，非常谨慎地逐步推行这个计划。

We're gonna roll it out really slowly with a series of trusted partners.

而且我也了解到了OpenAI这边的立场——我之前和格雷格·布洛克曼聊过这件事，这期节目在咱们的播客频道里就能听。

And then I kinda hear OpenAI's perspective as well, which I spoke with Greg Brockman about that, and you can listen to that show on the feed.

那期节目是昨天更新的。

That was yesterday.

他当时的说法大概是，他们对自己的治理模式很有信心。

And what he said is basically, like, we're pretty confident in our governance.

我们设计的时候就坚决不让它允许网络攻击，所以你可能会遇到更多拒绝，因为我们在网络攻击这块设了高墙，但我们还是希望每个人都能用上。

We've built this in a way that is not going to be permissive towards cyber attacks and you can, you know, you might get more refusals because of it because of our, the walls around cyber attacks, but we want everybody to have it.

你觉得这儿该咋办才对？

What do you think is the right approach here?

我觉得得在‘我们下一个模型发布就会毁灭人类、搞垮世界经济’和‘萨姆跟所有人说他爱他们’之间找点平衡。

I think there's gotta be a middle ground between our next model release will destroy humanity and crush the world economy and Sam telling everyone that he loves them.

我的意思是，你看看微软这周的公告，或者Adobe的峰会，

There's gotta be I mean, like, take a Microsoft announcement or an Adobe summit this week.

都有软件发布，

There's some software releases.

都有升级更新，

There's some upgrades.

不是每件事都得惊天动地、颠覆世界啊。

Not everything has to be earth shattering and world defying.

这不就是模型的下一个迭代嘛。

Like, it's just the next iteration of the model.

要是这些公司不用为了上市而一直制造 hype，我们真有必要为每个新模型都搞这么大阵仗吗？

Like, if these companies weren't in the position of having to kind of keep this drumbeat of hype being pushed until they go public in the markets, do we need this much hype for every new model release?

这话题是挺能聊的，但我希望这些更新能像发布说明那样平平淡淡就好。

It makes for good conversation fodder, but my my wish is these would just be kind of like they would be in release notes maybe.

也许发个简短的新闻稿就够了，别整那么复杂。

Maybe there's, like, a press release, and that's it.

没必要非得搞得非黑即白、轰轰烈烈的。

It doesn't have to be this all or nothing type of way of communicating around it.

这说法有道理，但说到底，还得看这些技术到底有没有真正的网络安全能力，或者进攻性能力。

Is that truther on the well, I guess it sort of depends on whether you believe these things have real cybersecurity capabilities or cyber offensive capabilities as well.

如果它们真有攻击性能力，那可就不能随便说‘去吧，随便用’了。

If they have cyber offensive capabilities, then you can't just, you know, sort of say, alright.

得慎重对待。

Go for it.

你怎么看这个问题？

So what's your perspective on that?

你觉得这事儿

Do you think it's

是真的吗？

real?

明白了。

Okay.

我觉得这事儿是真的。

Because I think it's real.

我接触过不少人，也做过不少相关报道，至少我觉得这事有一定程度的真实性。

From the people that I've spoken with, I've done enough reporting on it that I believe it's real at least to some degree.

如果这事儿是真的，这周彭博社报道了，Mythos系统被入侵了，一个未经授权的团队用了好几种方法，成功访问了模型。

So so if it is real, this week, Bloomberg reported there was a breach for Mythos where an unauthorized group, they tried a number of different strategies and were able to gain access to the model.

他们是Anthropic的第三方承包商的员工。

They work for a third party contractor that works for Anthropic.

他们就是靠猜，瞎蒙那个访问模型的网址，结果还真让他们猜中了。

And the way they did it is they literally, like, made educated guesses about the mark the the target URL to access the model.

如果你的模型真像你过去两周说的那么危险，我是Anthropic的话，我早就急得团团转了。

If your model is truly as dangerous as you have made it out to be over the last two weeks, If I am Anthropic, this announcement, you should be hair on fire running around.

要是他们真靠猜URL、靠外包人员就黑进去了，天知道还有多少人也能做到。

Because if they did it by guessing a URL and having contractor access, god knows who else has it.

中国早就拿到了。

China already has it.

什么Jensen的事儿，现在都无所谓了。

The whole Jensen thing just becomes moot.

如果这模型真这么强大又危险，这事儿该是头号新闻才对，他们早该站出来道歉，说正在尽全力补救——可现在呢？什么动静都没有。

Like, if it is so powerful and dangerous, this should be the biggest story, and they should be telling the world how not only are they incredibly sorry about what has happened, they are doing everything in their power to actually fix this and nothing.

这摆明了就是不在乎。

This is just it's like, they don't care.

我真的搞不懂。

I don't know.

要是真被这么简单的方式攻破了，他们难道不该更紧张吗？

Do do like, if it's truly breached in such a pedestrian way, like, shouldn't they care more?

首先，我得说，就因为 Discord 里几个宅男拿到了 Mythos 的访问权限，不代表 Mythos 的网络攻击能力是假的。

Well, first of all, I'll say just because a couple of dorks in the Discord got access to Mythos doesn't mean that the cyber offensive capabilities of Mythos are a lie.

要是你把它随便放给所有人，结果啥事没有，那我肯定得说点什么。

If you would have given it to everybody and and had a nothing burger, then I would have said something.

但光凭这个，还不能直接否定它。

But it's not on its face disqualifying.

现在都已经是 Mythos 真实曝光的第三周了。

Then I'll say it's we're already now in week three of this Mythos Real.

我理解你对这事的怀疑，真的。

And I appreciate your skepticism around it.

我真的很理解。

I really do.

我觉得吧，咱们现在就是面对这个新产品、这个模型了。

I think that we just we've kind of here are this is our new product or the model.

我觉得你我在这事儿上卡住了，只能等下去，看结果了。

I think you and I are kind of at an impasse, and we're just gonna have to wait and see to

你得给我个答复。

get that answer.

但不，不是这样的。

But but no.

不是。

No.

但我想知道，如果只是几个普通人在Discord里就能接触到你们那个可能毁灭世界级别的模型，这难道不更该引起重视吗？

But I wanna ask, shouldn't that be more important if a couple of guys and a couple of folks in the Discord are able to access your, like, potentially world destructive model?

如果你真在Anthropic工作，还负责领导这家公司，要是这事儿是真的，难道不应该是你最害怕的事吗？

Shouldn't that be if you are working for Anthropic, if you're leading Anthropic, shouldn't that be the most terrifying thing imaginable if it is real?

这么说吧。

I mean, let me put it this way.

Anthropic以前不也遇到过不少类似的情况吗？

Hasn't Anthropic had a number of similar situations?

比如Claude Code的源代码都泄露过，诸如此类的事。

The source code for, like, Claude Code leaked and all this stuff.

他们简直太依赖Claude了，我觉得他们本该更早想清楚这一点。

And it's just like they're I think it's almost like they're leaving too much to Claude, and they probably should have thought this through.

既然他们在做这个版本，Claude怎么没提醒他们改一下命名规范呢？

Why didn't Claude tell them to change the naming convention if they were working on this release?

太厉害了。

Powerful.

不对。

No.

你说得对。

You're right.

是的。

Yeah.

我觉得问题在于，大家的思维跳得太快了——刚才还在说‘地球上最危险的东西’，转眼又说‘模型在公园里长出了生命’，还觉得它会活过来抓你、发邮件、发帖子；可转头又说‘哦，对了’。

I think it's just more if like, it the the kind of whiplash from most dangerous thing on earth sandwich in the park, the model is coming to life and you're anthropomorphizing it, and it's, like, gonna come out and take you and send emails and post without your like, going from that to, oh, yeah.

顺便说一句，只要猜对一个URL，随便哪个承包商都能访问到它。

By the way, anyone can access it by guessing a URL and some contractor access and whatever.

对我来说，这根本说不通。

To me, it just doesn't square.

你和我昨天在跟格雷格·布罗克曼聊的时候也提过这个，你和萨姆·阿尔特曼对这事的看法挺像的。

You and, I brought this up in the Greg Brockman conversation yesterday, but you and Sam Altman have a similar perspective on this.

这事儿是这样的，萨姆·阿尔特曼这周上了阿什利·万斯的播客，聊了聊‘神话’这个事。

This is a just a so Sam Altman was at on Ashley Vance's podcast this week, and he talked about Mythos.

他说，说‘我们造了个核弹’，这营销简直太绝了。

And he said, it's clearly incredible marketing to say we've built a bomb.

我们本来就要把这玩意儿砸在你头上。

We were a we were about to drop it on your head.

我们卖你个避难所，一亿美金，包你全家安全，但前提是得我们挑中你当客户。

We will sell you a bomb shelter for a $100,000,000 thrown across all your stuff, but only if we pick you as a customer.

这招真高明。

That's good.

确实高明。

That is good.

好吧。

Okay.

也许……也许萨姆真的在乎我们。

May may maybe maybe Sam does love us.

也许他想大规模制造奇迹，我觉得那条推文后面的意思就是这个。

Maybe he wants to make magic at scale, which is I think what the rest of that tweet said.

对。

Yep.

嗯，

Well

在超大规模下制造奇迹。

Magic at hyperscale.

奇迹。

Magic

在超大规模下。

at hyperscale.

那也只能算是个愿望吧。

Be a would be Yeah.

我的意思是，这太让人失望了。

I mean letdown.

谁会想要这个啊？

That's who who wants that?

得了吧。

Come on.

不过我确实把达里奥又弄回了白宫。

Well, I did get Dario back into the White House.

看起来他们又要和白宫达成协议，开始跟整个政府合作了。

Looks like they're gonna have a deal with the White House again to start working with the whole government.

所以至少，这件事是办成了。

So if anything, it did that.

超大规模的魔法？

Magic at hyperscale?

我现在就只想着这件事。

That's all I'm thinking about right now.

我人生的新目标就是实现超大规模的魔法。

That's my new purpose in life, to deliver magic at hyperscale.

我不知道具体怎么做，但我就想干这个。

I don't know how, but that's what that's what I wanna do.

你现在对AI有点怀疑，那我正好可以念一段我们关于布罗克曼视频的评论，听听你咋想。

You're now having this skepticism around AI, so this might be a good opportunity for me to read a comment that we got on the, on the Brockman video and and see what you think.

LM模型面向公众都快四年了，你们还老在说下一个模型会有多牛。

We're closing in on four years since LM models were broadly released to the public, and you guys are still talking about how the next model will be so amazing.

永远都是‘下一个’。

Always the next one.

meanwhile，甲骨文裁掉了一半员工。

Meanwhile, Oracle laid off half its company.

脸书现在也正在裁掉一半员工。

Facebook is in the current currently in the process of laying off half its company.

数据中心一个接一个地被取消了。

Data centers are being canceled one after another.

Chat Chippy Tee 现在也放广告了。

Chat Chippy Tee now has ads.

这段日子挺不错的。

It was a good run.

咱们换下一个话题吧。

Try time to pivot to the next topic.

咱们听众真给力，有不少火药味十足的粉丝。

That's a good our listener we got some spicy listeners.

他们可喜欢了。

They like it.

对了，

By the way

我喜欢

I like

就是这个。

it.

我很感谢大家的反馈，不管是读者、听众还是观众的，只要别在苹果播客或Spotify上给两星就行。

I appreciate our reader feedback, our listener feedback, our viewer feedback just as long as it's not two stars in the Apple Podcast or Spotify app.

我们收到这些评论了。

I Send us these comments.

支持我们。

Agree with us.

我喜欢。

I I like it.

这让人眼界大开。

It's it's mind expanding.

你知道吗，我们现在又处在这个世界里，就是那种：好吧。

You know, we are because we are again in this world where it's like, alright.

下一个模型肯定会超棒。

The next model is gonna be so good.

但我觉得它们确实进步了。

But I I would say they've improved.

说实话，很难否认它们已经进步了。

I would say it's really hard to argue that they haven't improved.

我得说，它们的进步简直翻天覆地。

I I will say they've dramatically improved.

再说，我这行里的人，我清楚。

Again, I work in the industry.

就是那个说，当初大家都觉得5.0或5.1是失败作的时候，推理和工具调用才是下一个风口的人。

I that was the one who said when everyone said five o or five one was a dud, that reasoning and tool calling were gonna be the next big thing.

你说得对。

You're right.

这一路确实有飞跃式的进步，但我也完全理解听众的感受——我真的不想老谈下一个模型。

There's been dramatic step changes along the way, but but I also fully empathize with the listener that, like, I hate to talk about the next model.

这跟我刚才说的完全一样，对。

And it's just what I was saying Yeah.

就在刚才。

A few moments ago.

我真希望下一个模型发布时能像Adobe本周在Summit上推出的，或者微软在别的活动上发布的Copilot for Work那样平平无奇，你知道的，就该是这种感觉，别整什么‘人类小心’的戏码。

I want the next model release to be as boring as whatever Adobe launched at Summit this week or Microsoft launched at whatever else, Copilot for a co work or whatever you know, like, that's what it should be, not, like, beware humanity.

我们的下一个模型要发布了。

Our next model is dropping.

你知道吗，我觉得这周被低估的是什么？

You wanna know what I think is getting underplayed this week?

你其实已经提到了，但我真觉得咱们在休息前得把这话挑明了说。

And you already mentioned it, but I think we should just say this before we go to break.

ChatGPT图像2.0简直离谱。

The ChatGPT images two point o is insane.

对，太离谱了。

Oh, it's insane.

去网上搜一搜。

Search the web.

它还能编辑图片。

It can edit images.

我是说，我早就把从DALL·E开始的所有图像生成器都试了个遍。

I mean, it is I've I've tested like crazy every one of these image generators from Dolly on out.

这玩意儿简直离谱。

This thing is insane.

太离谱了。

Insane.

不可能。

No.

真的假的。

No.

这就是为什么我觉得，这在图像生成上绝对是质的飞跃。

This is why this was genuinely the I would put this at step change on the image generation side.

至于推理能力，我倒没觉得多惊艳，毕竟最近的模型只要能联网搜索，应该都能做这种多步推理。

The thinking side wasn't as interesting to me because any recent model should be able to do this kind of multistep reasoning and if web search is part of it.

但你看看这些输出结果，我之前看过好多像 Nano Banana 或 Nano Banana Gemini Flash 那些东西，跟现在的 Chatty PT Image 2 一比，简直像卡通画一样。

But, like, just seeing the outputs, and I've seen a lot of kinda what got us all excited about Nano Banana two or Nano Banana Gemini Flash two, like, looks cartoonish already compared to what, like, Chatty PT Image two is doing.

所以我觉得这确实是个大突破。

So I think this was this was big.

这真的非常了不起。

This this was actually a very impressive thing.

我觉得吧，挺有意思的。

Do I think but it's interesting.

这是一次很不错的软件更新。

It's a software update that is good.

你觉得它能改变人类吗？

Do I think it's, like, gonna change humanity?

不会。

No.

但确实很惊艳。

But it was very impressive.

这是第一个让我真开始担心平面设计师未来的工具。

This was the first one that actually had me worrying about the future of graphic designers.

真的有这么厉害。

It's that good.

我可不是在开玩笑。

I'm not even kidding.

哦，其实我一直都担心做传统平面设计的设计师们前景不妙。

Oh, I guess I mean, this is I've worried about the future of graphic designers who do traditional graphic design for a long time.

而且说实话，Claude Design 能做到这些，我觉得早就有预兆了。

And, actually, I mean, Claude design and what it's able to do, I think this has been coming for for a while.

像图片修改、优化，甚至网站的用户界面布局，这类技能早就跟写邮件标题一样，慢慢被淘汰了。

And I think, like, just being able to do some kind of, like, image alteration or improvement or even, like, UX layout of a website, I think that kind of skill has been going the way of someone kind of, like, writing email subject lines for a long time.

但对我来说，这变化其实没那么大。

But I don't I I guess to me, that didn't change that much.

挺有意思的，尤其是视觉传达这块——我注意到这次它真正擅长的是把一个想法清晰地表达出来。

It is interesting, like, visual communication and because what I saw happening much better with this is actually kind of communicating an idea.

比如 Nano Banana 能做出挺像样的卡通流程图，但 GBT Image 2 真的把视觉概念表达得好多了。

Like, Nano Banana could do some kind of, like, good cartoonish flowcharts, but, like, GBT Image two actually was communicating visual concepts much, much better.

或者说是用视觉来传达东西，

Or, like, communicating visually,

根本不是什么视觉概念的问题。

not even visual concepts.

没错。

Agreed.

行吧。

Alright.

咱们歇会儿，聊聊 Meta 这波裁员的事儿，还有别的能扯的，趁还没得走赶紧说。

Let's take a break and talk about these cuts at Meta and whatever else we can fit in until we have to go.

广告后马上回来。

We'll be back right after this.

你们 IT 团队一半时间都耗在重复的工单上，公司一扩张，请求就堆得没完。

Your IT team waste half their day on repetitive tickets, and the more your business grows, the more requests pile up.

密码重置、权限申请、新人入职，全在拖累他们做真正重要的事。

Password resets, access requests, onboarding, all pulling them away from meaningful work.

用了 Servo，你能把 80% 的客服工单直接砍掉。

With Servo, you can cut 80% of your help desk tickets.

别的老玩家只是硬塞 AI 进去，Servo 从一开始就是为 AI 代理设计的。

While legacy players bolt on AI, Servo was built for AI agents from the ground up.

来看看这个转变有多彻底。

Here's the transformation.

以前经理给新人入职，一套流程下来得花好几个小时。

When a manager used to onboard a new hire, the old process would take hours.

得在 Slack 里喊一嗓子，再发邮件找 IT，还得等审批。

They'd ping Slack, email IT, wait on approvals.

结果新人干等着，一等就是好几天。

Meanwhile, new hires sit around for days.

现在用 Servo，经理在 Slack 里说一声，就能启动入职流程。

With Servo, a manager asked to onboard a new hire in Slack.

AI会自动在几秒钟内完成所有权限分配，并确保必要的审批都到位。

The AI provisions access to everything automatically in seconds with necessary approvals.

IT部门根本不用插手。

IT never touches it.

像Perplexity、Verkada和Mercur这样的公司，上线后马上就把一半以上的工单给自动化了。

Companies like Perplexity, Verkada, and Mercur automated over half their tickets immediately after setup.

如果我要用这类产品，我肯定就选这个。

If I needed this product, it's exactly what I would use.

Servo正在为Perplexity、Mercur、Verkada和Clay这些全球增长最快的企业提供动力。

Servo powers the fastest growing companies in the world like Perplexity, Mercur, Verkada, and Clay.

让你的团队摆脱客服工单，回归真正喜欢的工作。

Get your team out of the help desk and back to the work they enjoy.

马上去serval.com/bigtech预约免费试用吧。

Book your free pilot at serval.com/bigtech.

就是serval.com/bigtech。

That's serval.com/bigtech.

我在这档节目里采访过很多优秀的科技创始人，有一个出人意料的共同难题反复出现——找一个合适的域名。

I've interviewed a lot of great tech founders on this show, and one surprisingly universal challenge comes up again and again, finding the right domain name.

我当初做‘Big Technology’的时候，自己也碰过这事儿。

It's something I ran into myself when launching big technology.

你想要的名字八成早被人抢了，这时候很容易就想凑合着用一个算了。

The names you want are often taken, and it's tempting to just settle and move on.

但我最敬佩的那些创始人，从不在根本问题上将就，域名就是其中之一。

But the founders I respect most don't settle on fundamentals, and your name is one of them.

你的域名得让人一眼就明白你到底在做什么。

It should immediately signal what you actually build.

这正是我喜欢.dot tech这种域名的地方。

That's what I appreciate about dot tech domain names.

顺理成章，再自然不过。

It just makes sense.

它直接告诉世界、你的客户、投资人，还有任何搜你的人都在说：我们是搞科技的。

It tells the world, your customers, your investors, anyone googling you that you're building in technology.

简洁直接，不加修饰。

Clean, direct, no qualifiers.

我看到越来越多认真的初创公司开始用这个。

And I'm seeing more serious startups lean into it.

Nothing.tech、1x.tech、aurora.ces.tech、ultra.tech、alice.tech、neuron.tech、blaze.tech、pie.tech，还有好多好多。

Nothing.tech, 1x.tech, aurora.ces.tech, ultra.tech, alice.tech, neuron.tech, blaze.tech, pie.tech, and so many more.

如果你做的东西是技术驱动的，别将就。

If you're building something tech first, don't settle.

赶紧从你喜欢的注册商那儿抢个.tech域名，从第一天起就让别人一眼看懂你是干嘛的。

Secure your dot tech domain from any registrar of your choice and make your positioning obvious from day one.

我们又回到《大科技》播客的周五特辑了。

And we're back here on big technology podcast Friday edition.

今天照例请来了Margins的Ranjan Roy，我们得聊聊从加州门洛帕克传来的几个让人心里发沉的消息。

We're here with Ranjan Roy of Margins as we typically are on Fridays, and we have kind of a depressing story or set of stories out of Menlo Park, California.

这事儿来自彭博社。

This is from Bloomberg.

Meta通知员工，公司将裁员10%，并推动提效。

Meta tells staff it will cut 10% of jobs and push for efficiency.

Meta平台计划裁员约8000人，占总员工数的10%，目的是提升效率，缓解在人工智能上巨额投入带来的压力。

Meta platforms plans to cut 10% of workers or roughly 8,000 employees in an effort to boost efficiency and offset its heavy spending on artificial intelligence.

公司已公布这一决定，并在周四向员工发送了内部备忘录。

The company disclosed the move, and a memo sent to employees Thursday.

Meta还将冻结6000个原本计划招聘的岗位。

Meta also won't hire employ hire workers for 6,000 open roles that it had intended to fill.

就在CEO马克·扎克伯格大手笔投入人才和基础设施，全力开发包括大模型和聊天机器人在内的顶尖AI产品之际，公司却开始裁员。

The cut the job cuts come as chief executive officer Mark Zuckerberg is spending aggressively on the talent and infrastructure needed to develop state of the art artificial intelligence products, including LMs and chatbots.

这还没完，再来看看路透社的报道。

Well, as if that was not enough, this is from Reuters.

Meta将开始收集员工的鼠标移动和键盘输入，用作AI训练数据。

Meta to start capturing employee mouse movements, keystrokes for AI training data.

据路透社看到的一份内部备忘录显示，Meta正在美国员工的电脑上安装新的行为追踪软件，记录鼠标移动、点击和键盘输入，用来训练AI模型，这是公司打造能自主完成工作任务的AI助手的更大计划的一部分。

Meta is installing new traffic tracking software on US based employees' computers to capture mouse movements, clicks, and keystrokes for use in training its artificial intelligence models, part of a broad initiative to build AI agents that can perform work tasks autonomy, the company told staffers in an internal memo seen by Reuters.

一方面，大裁员。

One hand, laying people off.

另一方面，留下的员工还得配合，让自己的鼠标移动、键盘输入，还有所有电脑操作都变成训练AI的数据。

On the other hand, those that stay have the pleasure of their movements and keystrokes and whatever they're doing on their computers being used to train AI.

你咋看？

Your reaction.

明白了。

Okay.

咱们分开说。

Let's separate out the two.

先说第一点，说实话，这事儿我没觉得多震惊，但真没想到，这么多年来，他们在元宇宙和现实实验室上砸了那么多钱，居然从没想过要控制成本。

The first, it actually shocked me, or it didn't shock me, but it's just it's incredible to me that for years on the metaverse spending and Reality Labs, they never would tie any kind of cost cutting efforts.

那所谓的‘效率年’到底算啥？

The entire year of efficiency what was that?

是2024年吗？

2024?

2023年？

2023?

对，2023年。

'23.

2023年？

'23?

他从来没直接说，这是因为我们在元宇宙上花太多钱了。

Like, he never directly attributed it to because we are spending so much on the metaverse.

更让人觉得不可思议的是，现在为了弥补在人工智能上的巨额投入，居然裁了八千人。

And it's kind of, like, amazing that now we are cutting 8,000 people to offset our heavy spending on artificial intelligence.

市场嘛，我懂，他们爱听这个。

So the market, I get, loves hearing that.

投资者也爱听这个。

So investors love hearing that.

大家肯定会继续这么干，但我真说不准。

So people will continue to do it, but I don't know.

我觉得，这些公司明明知道这么做能短期拉升股价，就拿这个当沟通策略，这真的对吗？

I I feel do you think it's the right way to communicate for these companies just because they know it'll pop their stock into the short run?

我觉得这根本不是沟通的问题。

Well, this is one I don't think is about comms at all.

我觉得他们真的是在践行自己说过的话。

I think this is legitimately, like, they are doing what they said they're doing.

对吧？

Right?

他们确实是在……

They are

是公司之前招太多人、人浮于事，现在该瘦身了？我觉得很多大公司，尤其是过去五年，确实如此。还是说，他们真就是想省钱，好把钱投到人工智能上？

Is is it overhiring in a bloated company that just needs to actually trim itself, which is, I think, very true for many, many of these big companies, especially over the last five years, or is it really, like, we need to cut costs so we can invest more in artificial intelligence?

我觉得是后者。

I think it's the latter.

我的意思是，他们在AI上已经砸了太多钱了。

I mean, they've spent so much money on AI.

他们正在投钱建数据中心。

They're spending on the data centers.

如果你没利润，市场可不会买账。

The market is actually much less forgiving if you don't have margins.

对吧？

Right?

他们确实能在AI上看到回报，毕竟AI帮他们优化了创意工作流。

And they they don't have they mean they have some ROI on the AI because it's helping optimize their creative stack.

最近不是有新闻说，他们马上就要超过谷歌，成为最大的广告业务了嘛。

And I think there was a headline recently that they're gonna pass Google as largest advertising business.

他们能看到这些成果，但他们自己并不是那种靠AI算力需求暴涨受益的平台。

So they see that the results there, but they're not a platform that's sort of benefiting from this surge in demand for AI compute.

他们也没搞出什么超级人工智能。

They haven't built super AI super intelligence.

所以他们现在处境尴尬，没法再这么臃肿下去了，尤其是还没拿到最顶尖的模型。

So they are in a position where they cannot remain this bloated, especially as they don't have the leading model.

但话说回来，

Well but,

他们平时花的钱可不少。

normally, they've spent a lot of money.

光说‘我们要多花钱’，这可没法让人安心。

So just saying we will spend more money to me isn't a reassuring message.

作为 headline 听起来，好像挺好的。

Like, at the as a headline, it sounds like, okay.

这事儿没错。

This is good.

钱根本不是问题。

This is whatever it needs to be doing, but, like, money has not been the problem.

几个月前，他们出多少钱请人来？

What were they paying people couple of months ago to join?

我的意思是，说来搞笑，他们为了请亚历山大·王，真花了整整一百五十亿美元。

I mean, they basically I mean, this is facetious, but they effectively spent, 15,000,000,000 to hire Alexander Wang.

是的。

Yeah.

我的意思是，再多花钱根本解决不了问题。

I mean, all so so more money is not the answer.

所以，如果公司真的臃肿了，他想瘦身，搞点扎克伯格式的效率年那种操作，我懂。

So, like, again, I get if the company is just bloated, if he wants to kind of, like, start to make things leaner, go into Zuck beast mode, year of efficiency type stuff.

要说谁能比别人更擅长干这事儿，那肯定是扎克伯格。

I I mean, I if anyone is able to do that well and better than others, it's Zuckerberg.

但我也不确定。

But, like, I don't know.

光说能多投钱，我还是不太信这套。

Just because it means we'll be able to invest more, I still I don't quite buy that.

但这跟核心追踪是两码事。

But that's separate from the key tracking.

那

That

是的。

Yeah.

说说这个。

Talk about that.

这事儿吧，从技术角度讲，其实还挺有意思的，虽然挺吓人的。

I it's one of those that it starts to makes like, purely technically, it's almost kinda interesting.

就是说，大家是不是都在训练模型，重复做那些事儿——效率是高了，但也就那样。

Like, if it's terrifying, but it's interesting in terms of is everyone essentially training models to do the things they're doing repetitively, which is efficient, I guess.

那既然这样，你干嘛还待在那儿呢？

I guess, like, yeah, why do you work there then?

我觉得，如果目标是能让人有点儿干劲儿，那或许还行。

I think if the goal is maybe there could be an inspiring alright.

这是我自己的想法。

Here's my attempt here.

以后，你真正要做的工作，我觉得，就是把信息从系统A传到系统B，做个简单的汇报，再加点你自己的小分析，当个大流程里的一颗螺丝钉——这种活儿，根本算不上什么知识工作。

In the future, the type of work you will need to be doing is, and I believe this, like, moving a little bit of information from system a to system b, making a little presentation around it, doing a little kind of, like, adding your own tiny bit of insider analysis and being that cog in a larger process is not going to be a lot of knowledge work.

所以扎克伯格可能站出来说：我是在为你们所有人铺路，让你们成为所有科技公司员工里，最能适应未来需求的人。

So maybe Zuckerberg can stand up, and he can just be like, I am preparing you all for the future so you can be the best positioned out of any kind of tech company employee to kind of meet the needs of this future.

这就是我想传达的励志观点。

That's my inspiring message behind this.

我的意思是，我希望他真这么干的时候，是躲在办公室里，对着会议电话说的——要是他在Meta食堂当众这么说，肯定得被蔬菜砸死。

Mean, I hope he'd I hope he'd be doing it, like, you know, locked in his office on like a conference room phone because he would get vegetables thrown at him from the Meta cafeteria if he said that.

回顾历史上这种劳动形态的转型，以前有个叫泰勒制的东西，工厂里工人的一举一动都被测量，逼着他们做到最高效。

You know, in in the history of of labor and trends transitions of this nature, there was a practice back in the day called Taylorism, where they they measured the movements of people working in the factory, and they got them to move as efficiently as possible.

他们 literally 控制工人的每一个动作，像机器一样，不许有一丝浪费。

They literally controlled their movement to, be like a machine, so there wouldn't be any wasted movement.

后来，这些人很多就被机器取代了。

And then eventually, they replaced many of them with machines.

我就是觉得，这种故事最后从来不会对打工人有利。

I just don't see stories like this, ending, in the right way for the worker.

不过呢，这里有个挺有意思的小插曲——你还记得Scale AI吗？亚历山大·王？当然记得。

And, I will say there's one interesting wrinkle here, which is that, do you remember Scale AI, Alexander Wang's Of course.

公司。

Company.

他们最近告诉我，他们大部分训练用的是强化学习，就是搭建一些环境，哦。

They told me recently that most of the training that they're doing is reinforcement learning, where you build environments Oh.

让机器人进去自己摸索该怎么做。

For the bots, and they go and they try to figure out what to do.

那你要怎么才能打造出这么棒的强化学习环境呢？

And, well, if you're what do you need to do to build these great reinforcement learning environments?

你得给他们看各种表单、网页操作这些，然后让他们模仿出来。

You sort of need to show them forms and web behavior and stuff like that, and then you try to get them to model it.

要是亚历山大·王在Meta内部干的就是这事儿，我一点都不惊讶——换句话说，这根本不是什么彻底的AI自动化，而是在给机器人建健身房，好让它们有更多环境练强化学习。

And, with Alexander Wang within Meta, I wouldn't be stunned if that is what's happening is that, you know, maybe the the other way to read this is instead of, like, a a complete AI automation move, it is effectively building gyms for bots that need more environments to do reinforcement learning within.

这太有意思了。

That is fascinating.

如果亚历山大·王被收购后真能这么创造价值，那150亿美元花得还真值。

And and if Alexander Wang is adding value post acquisition in this way, maybe that 15,000,000,000 was was worth it.

我其实挺震惊的。

I I was kind of fascinated.

他们压根就没否认这事儿真的在发生。

Like, they also there was no denial at all of this happening.

报道里提到，CTO安德鲁·博兹在评论区直接说，你用公司配的电脑，就没得选。

Like, part of the reporting was that the CTO, Andrew Boz, like, he responded in the thread that there's no option out of this on your work provided laptop.

这条评论底下全是哭脸、震惊脸和愤怒的表情符号。

This comment received a mix of crying, shocked, and angry face emojis.

不过，Meta官方对《商业内幕》的回应是，他们有安全措施保护敏感内容，数据不会挪作他用。

But, also, the official response from Meta to Business Insider was, there are safeguards in place to protect sensitive content, and the data is not used for any other purpose.

真让我没想到，他们居然就这么直说了。

Like, they it is amazing to me that they just they said it.

一点都没想退缩，就照着干。

There's no backing off of what they're doing.

这事儿简直太离谱了。

So so this is this is kinda nuts.

这事儿真在发生。

It's happening.

太疯狂了。

Crazy.

没错。

Yeah.

我以前干过的几份工作，真不希望这软件装在那台电脑上——反正我也没正事干，整天就刷collegehumor.com。

There's definitely a few jobs that I've held in my life that I really would not want this software to be installed on, because I had nothing to do and spent a lot of time on my collegehumor.com.

你想想，要是所有人都在Twitter上摸鱼，不干活，那AI训练的数据不就全是这些玩意儿？

See, this is actually, imagine if everyone is just on Twitter, is not not doing work, and that's what all the training data is received.

然后这些AI助手一上线，就开始干活，结果呢？刷着刷着就跑去Instagram了，接着就……

And, like, the the the agents just, like, they start he starts to put him to work, and then they're just scrolling, and then they go to Instagram, and then they're like

你迟早会看到这些助手的。

You're gonna see these agents.

你说得对。

You're right.

它们会被安排帮你做演示文稿之类的活儿。

They're gonna be tasked with, like, you know, writing a deck for you.

它会直接接管你的电脑。

You're gonna have it take over your computer.

干到一半，它就开始刷YouTube，看狗玩滑板的视频。

Midway through, it's gonna be on YouTube watching dogs on skateboards.

你肯定得问：这到底在搞什么鬼？

And you're gonna be like, what's going on here?

它还觉得：哦，原来这样才算正经上班。

It's like, well, I learned that this is the right way to do work.

相信数据吧。

Trust the data.

相信数据吧。

Trust the data.

相信训练结果吧。

Trust the training.

我敢打赌，最后肯定会冒出个荒唐的研究，说那些会拖延的AI模型反而比死守任务的更有效。

I'm sure we'll end up seeing some ridiculous study about this, and it'll be like AI models that procrastinate are actually more effective than AI models that stick to task.

说起来，你有没有看过那个采访？我觉得她好像是Anthropic的伦理专家。

Actually, did you see I think she was like the Anthropic Ethicist.

网上现在到处都在传那个采访视频。

There's all these videos going around around, like, this interview.

大概意思就是，她负责理解Claude这个模型背后的情感机制，你知道的，就是那种……

Basically, the idea was like, she was she's like the one who's under supposed to understand, like, the emotional underpinnings of Claude and the model and, like

哦，对对对。

Oh, yeah.

刚才在聊Amanda的事。

Was talking about Amanda

是Askel。

Askel.

没错。

Yeah.

而且它还会焦虑。

And how it is anxious.

所以啊，不如让你的助手看看YouTube，顺便刷刷X。

And so maybe you should let your agent watch a little YouTube, surf a little x.

这样反而能更高效地把事儿办了。

Just it will get the job done in a more efficient way.

你逼得太紧，它反而干不好。

When you push it too hard, it's it's just not gonna do a good job.

给它点喘息的空间。

Give it a break.

我们不都得歇歇吗？

Don't don't we all don't we all need a break?

谁不需要暂时放下手头的事，去做点完全没意义的事呢？

Don't we all just need some time to divert from task and try something completely meaningless?

这才叫真正的人类生活啊。

That's the that's what being human is all about.

如果你是Meta的员工，它会通过追踪你的行为来不断学习。

And it will learn as it tracks your behavior if you're a Meta employee.

我都能想象到，那些Meta员工整天拿着手机刷视频，却让AI代劳干些蠢事。

I could just imagine all these Meta employees, like, built having AI token max doing some dumb shit while spending the whole day, like, watching videos on the phone.

新的

The new

这就是2026年工作的未来。

that is that is the future of work in 2026.

如果现在《硅谷》这部剧还在播，我猜肯定会有新版，毕竟素材多得是。

If there's a Silicon Valley if that show existed now, I wonder if there'll be a new version of that, but there's just so much material.

现在想讽刺都难了，因为这事儿已经离谱到极致了。

Kinda hard to parody at this point because it is so

是啊。

Yeah.

荒唐得不行。

Ridiculous as it is.

没错。

Right.

根本没法调侃了。

There's no parody.

行吧。

Alright.

听好了。

So look.

我们这几个月一直在准备材料，对比流媒体平台的价格，现在正好有个机会好好吐槽一下。

As as we come to a close, we've had this basically comparison of streaming prices, in our prep doc for weeks now, and we have an opening for a rant.

那你就来收个尾吧，拉詹，跟大伙儿说说这流媒体涨价到底是怎么回事，意味着什么？

So why don't you take us home, Ranjan, with a little exposition here on the increase in streaming prices and what it means?

劳工统计局的数据说，去年消费者价格指数上涨了2.6%。

So the Bureau of Labor Statistics tells us that the consumer price index rise in the last year was 2.6%.

可我呢，这些年订了一堆流媒体服务，根本不知道该砍哪个——儿子死活不让停Disney+，老婆更离谱，说Netflix、HBO、Hulu、Peacock这些一个都不能少。

Now as someone who subscribed to too many streaming services over the last years and does not know when to cut what and has a son who won't let them cut Disney plus and a wife who said the idea of Netflix or HBO ever leaving or even Hulu or Peacock, you know, these things.

我很好奇，因为前几天奈飞刚把个人套餐涨到15.99美元，但同时把能同时用的设备数量砍得特别死。

So I I was curious because Netflix the other day just raised prices to $15.99 for the individual plan, but they severely restrict you in number of devices.

我现在每个月花27美元，简直离谱。

So I think I'm paying $27 now, which is insane.

所以我回头去查了下。

So I went back and looked.

从2019年开始算，那时候还没疫情，我想看看这些服务到底涨了多少。

It since 2019 because I was kind of, like, pre pandemic, how much have prices gone up?

大家都知道，以前流媒体这生意根本赚不到钱，对其他公司来说就是个引流工具，只有奈飞还算有点盈利。

And we all knew there was this moment where the streaming business did not make sense, and this was a loss leader for all these companies other than Netflix.

迪士尼Plus刚上线的时候才6.99美元。

So Disney Plus came out at $6.99.

现在涨到了18.99美元。

It's currently $18.99.

Hulu从11.99涨到了18.99。

Hulu went from $11.99 to $18.99.

HBO Max 其实主打高端定位，现在也从14.99涨到18.99了。

HBO Max, actually, to their premium positioning, only $14.99 to $18.99 now.

Peacock 从5块直接涨到11块。

Peacock, basically $5 to $11.

Paramount+ 从5块涨到8块。

Paramount plus, 5 to $8.

Apple TV 呢，还是老样子，我真受不了他们这服务模式，以前才5块，现在直接涨到13块一个月。

Apple TV, again, going back to my hatred of their services business now, comes in at $5, bundles you in.

现在一个月要13块了。

Now it's $13 a month.

这些服务我都搞不懂了，真不知道该怎么选。

Like, all of these things, I think I don't know.

一算每月开销，就得琢磨着该砍掉哪个。

When you look at your monthly expenses, then deciding what you're gonna have to cut.

这种事儿到处都是。

And you see this across everything.

这其实也是优步那种心态。

It was the Uber mentality as well.

但我真说不好。

But I think I don't know.

我一直在想，是不是该有个针对普通科技用户、科技圈内人、那种收入不错、懂科技又离不开这些服务的人的通胀指标？

I I would I was thinking, like, there needs to be an inflation metric relative to the average, probably big technology listener, tech industry participant, just kind of like, I don't know, just, like, upwardly mobile tech savvy person that is just stuck subscribing to all this.

Spotify的价格也涨了吗？

Do you Spotify has been jacked up as well?

可就是没人抱怨。

And there's no backlash.

根本没人组织起来大规模取消这些订阅。

There's no big consumer, like, movement to actually, like, cancel these services.

你有取消哪个吗？

Are you cutting any of these?

我试过一阵子，只留一个最常使用的，但最后还是放弃了。

I I have tried for a while to, like, just be subscribed to the one that I use most often, but I've given up.

我真是烦透了，老是忙着取消Netflix，过阵子又得重新订回来，搞来搞去的。

I'm, like, quite fatigued at, like, trying to cancel Netflix and then reinstalling it and stuff like that.

所以现在我又把Netflix续上了。

So now I have Netflix.

Prime Video、HBO我都有，还有那个Paramount Plus，我好像花了三十块买了一年的会员。

I have Prime Video, HBO, and I think I think I subscribed to, like, Paramount plus for, like, $30 for the year.

我想看那个用AI做的《南方公园》。

I wanted to watch the South Parks with the AI.

那你还得有Peacock看英超，Paramount Plus还得看NFL。

So But then you have Peacock for Premier League, Paramount and NFL, Paramount plus for NFL.

我也不晓得。

I don't know.

我去爸妈家的时候，他们还在用有线电视，那会儿我居然有点想回去了。

I like, it is interesting to me that I was at my parents' place who still have cable, and it actually kind of made me miss cable.

我也不懂。

And I don't know.

我觉得他们花了一百八十块，但我现在花得比这还多。

I think they're spending, like, a $180, which I think I'm spending more now.

真是疯了。

But Crazy.

这一波AI热潮最后会不会让我们反而怀念起以前流媒体刚兴起时的日子，甚至想回过去有线电视的时代？

Is is all this AI hype gonna end up with us longing for the days of basically what streaming has done, make me wish and miss cable?

真会这样吗？

Is that what's gonna happen?

首先，情况只会变得更糟，有两个原因。

Well, first of all, it's just gonna get worse for two reasons.

一个是行业开始合并了。

One is we're starting to see consolidation in the space.

你看Netflix已经明显胜出了，接下来派拉蒙和华纳兄弟探索频道也要合并了。

Like, you have Netflix as this clear winner, and then, and then Paramount and Warner Brothers Discovery are gonna tie up.

对吧？

Right?

所以啊，你会发现，其实这两家互相制衡倒也不是坏事。

So, like, you're gonna see well, it is actually good that the two will balance each other out.

但那种每个流媒体平台都互相乱打价格战、谁也不服谁的日子，估计要没了。

But like the days of every streamer competing with every streamer kind of uneven footing in price matters.

这种局面好像正在消失。

It seems to be away.

真的快消失了。

Seems to be going away.

再加上AI靠分析人的观看习惯，现在还逼着员工每天上班至少看五小时奈飞，这简直就是给我们所有人挖的坑，迟早要被坑惨。

And then, of course, with AI trained on human screen behavior, they now are required to watch at least five hours of Netflix during the workday, and that is a demand signal that we're all gonna get screwed by.

然后苹果和Ternus说不定能搞定一切，把所有订阅服务都拆了，搞成一个真让人期待、不让人觉得被绑架的产品。

And then Apple and Ternus will somehow solve it all and unbundle all their subscription services and actually make it just a product I'm excited about and I don't feel trapped by.

苹果确实有这个能力这么做。

That is something that Apple could do.

所以也许他们真会干。

So maybe they will.

苹果，再说说那个iCloud照片的事儿吧，真是服了。

Apple Again, iCloud freaking photos.

我每个月付40刀，真不知道图啥。

I'm paying $40 a month.

我甚至都不明白这到底是怎么运作的。

I don't even know how.

它居然直接警告我，说我们家是家庭套餐，如果不续费，你这辈子就再也看不到自己的照片了。

It literally was telling me that, Like, it was telling me telling my wife we're on a family plan, like, you will lose access to your photos your entire life.

这意思就是，不交钱就跟你拼命，苹果现在就是这么搞服务的。

Like, if you do not pay us gun to head, that is the Apple services business model now.

他们确实有这套操作，我得承认。

They certainly have I will say this.

他们确实有这么些商业手段。

They certainly have some business practices in there.

Ternus当然也得赚钱，这事儿得闭环，但希望他能想想，这要是成了他的标签，那可太丢人了，别干这种事儿了，真恶心。

So, I mean, Ternus, of course, will have to make make money to bring a full circle, but, hopefully, he, like, looks at this and just realizes it's gonna be his legacy and decides not to do stuff like this because that sucks.

埃迪，你觉得这个桌面机器人会带来订阅收入吗？

Eddie, do you think there's subscription revenue tied to the tabletop robot?

但愿如此。

I hope so.

但愿如此。

I hope so.

这是按使用量计费，按令牌收费的。

It's consumption based, token based.

就这模式。

That's it.

它每动一下，就得扣一次费。

Every time it moves Oh

我的天啊。

my god.

它越复杂，你手机上收到的提醒就越多。

The more complex it does, you just get an iPhone notification.

又扣你钱了。

You've been charged another.

我的天。

Oh my god.

我都能想象出那个广告了。

I can just imagine the ad.

画面从黑屏缓缓亮起。

The scene fades in from black.

旁边摆着一张漂亮的玻璃桌，桌上放着一个桌面机器人，它手上还连着一块屏幕，站着的正是拉詹·罗伊。

Standing in next to a beautiful glass table with a tabletop robot and a screen attached to its hand is one Ranjan Roy.

广告就这么开头。

Here's how the ad begins.

嗨。

Hi.

我是拉詹·罗伊。

I'm Ranjan Roy.

希望你挺享受你那个桌面机器人，也希望你继续每月付10.99美元，就为了能把你家东西都存下来。

I hope you're enjoying your tabletop robot, and I hope you keep paying $10.99 a month for the for the pleasure of being able to preserve everything in your house.

等你一停订阅，我就让这些机器人把你家砸个稀巴烂。

Because once your subscription lasts, I will be using these robots to smash your shit up.

这就是Ternus的商业模式，股价肯定飞涨。

That's that is the Ternus business model, and the stock is gonna skyrocket.

就这。

That's it.

就这。

That's it.

现在咱们总算搞明白这桌面机器人是干嘛的了。

Now we figured it we at least figured it out figured out what the tabletop robot is for.

瞧见没。

There it is.

纯粹是吓唬人罢了。

Merely a threat.

这纯粹就是个威胁，逼你继续交服务费。

It's merely a threat to keep paying your services bill.

总得有点啥吧。

Gotta have something.

那玩意儿能把你直接干飞。

That thing will knock you the fuck out of you.

这才叫创新，咱们这期《科技大观》周五特辑，又得这么收尾了。

Now that is innovation, and that is how we will end another week here on big technology podcast, Friday edition.

拉詹，老样子，见到你真高兴。

Ranjan, great to see you as always.

再次感谢你

Thank you again for

来参加节目。

coming on.

下周见。

Next week.

没什么是安全的。

Nothing is safe.

好了，各位。

Alright, everybody.

谢谢收听，我们下期再见，欢迎回来收听《科技大观》。

Thanks for listening, and we'll see you next time on big technology podcast.