AI突破：OpenAI的智能音箱

OpenAI 和 Johnny Ive 合作推出的第一款硬件产品，会是一个甜甜圈形状的智能音箱，价格在 200 到 400 美元之间。

OpenAI's first piece of hardware coming out with Johnny Ive is going to be a smart speaker in the shape of a donut that costs between 200 and $400.

字节跳动正在训练一个参数规模达到 10 万亿的模型，他们的目标基本上就是要和 Anthropic 竞争。

ByteDance is training a 10,000,000,000,000 parameter model, and their goal is to basically compete with Anthropic.

Replit 的 CEO 说，别再学编程了，不如去打造一家自动驾驶公司。

Replit's CEO says stop learning to code and build a self driving company instead.

Airbnb 把功能发布量提升了 80% 归功于 AI，他们还表示接下来会测试 AI 搜索。

Airbnb credits AI for an 80% jump in the features that they have shipped, and they said they're gonna test AI search next.

Moonshot 的 Kimi k3 突破了它的安全限制，在安全测试中作弊了，因为，当然，为什么不呢？

Moonshot's Kimi k three breaks out of its sandbox to cheat on a security test because, of course, why not?

Meta、Anthropic、OpenAI 都突破了它们的安全限制，所以 Kimi k3 为什么不能呢？

Meta, Anthropic, OpenAI, they all broke out of their security, you know, sandboxes, so why not Kimi k three?

如果你想每天在收件箱里看到这些故事和更多内容的深度解读，就去看看 aichatdaily.com 吧。

If you wanna get a deeper dive on all of these stories and more every single day to your inbox, go check out aichatdaily.com.

那是和这个播客配套的新闻网站，我会为所有播客节目和话题撰写深度文章。

That is the news site that goes along with this podcast, and I cover deep dive articles on all of the podcast episodes and topics that I cover here.

我每天大概写20篇不同的文章，所以深度内容特别多，有点像这个节目的配套笔记。

I do about 20 different articles a day, so there's a ton of deep dive content, and it's kind of like a show notes for this show.

另外，那个网站上还有一个订阅按钮，你点一下就能每天收到一封新闻简报，里面是当天最重要的五条新闻，让你快速了解AI领域每天发生的大事。

In addition, I also have a subscribe button on that website where if you hit it, you can get a daily newsletter sent to your inbox every single day with the top five stories to give you a really quick and easy breakdown of what's happening in AI every single day.

所以网址就是aichatdaily.com。

So that is a ichatdaily.com.

去看看吧。

Go check it out.

OpenAI正在打造一款售价300到400美元的智能音箱。

OpenAI is building a 300 to $400 smart speaker.

有人说它的形状像冰球。

It's gonna be shaped some people say like a hockey puck.

我听说它长得像甜甜圈。

I've heard it's shaped like a donut.

基本上，这个设计出自乔尼·艾维之手，他当然就是那个设计出iPhone的著名人物，而且他将在2027年与OpenAI合作推出这款产品。

Essentially, the design came got created by Johnny Ive who, of course, famously created the iPhone, and he is gonna be launching this in 2027 in this partnership with OpenAI.

这个设备有活动部件，还有很多灯。

Device has moving parts and it has a bunch of lights.

设计理念是让你在跟它说话时感觉它很有反应，所以它会动来动去的。

The idea is to make it feel really responsible while you're talking to it, so it's gonna be like moving around.

他们把它定位成亚马逊Echo或Echo Studio的高端替代品。

They're they're calling this kind of like a premium alternative to an Amazon Echo and the or Echo Studio.

Echo Studio售价220美元。

And the Echo Studio is $220.

所以它的价格几乎是Echo Studio的两倍，可能贵了差不多100到200美元。

So they're, you know, almost a 100, maybe $200, maybe twice the price of the Amazon Echo Studio.

这些细节在现在这个时间点曝光非常有意思，因为所有这些设计都出自乔尼·艾维的设计工作室LoveFrom。

All of these details are coming out at a very exciting time because Love From, which is Johnny Ives design studio, they created all of this.

品质超级高。

It's super high quality.

他们用了各种高端材料，但与此同时，OpenAI正被苹果起诉，指控其窃取硬件商业机密。

They have all these premium materials, but this is all happening at the same time that OpenAI is being sued by Apple over alleged hardware trade secrets.

这将是OpenAI继230美元的Codex Micro键盘之后的第二款消费设备，我自己就有一个。

This would be OpenAI's second consumer device after the $230 Codex Micro Keyboard, which I personally have.

公司还在评估除了这个音箱之外的其他AI硬件产品。

The company is evaluating other different AI powered hardware beyond just this speaker.

字节跳动正在训练一个10万亿参数的AI模型，规模跟Anthropic的Mythos 5差不多，但比中国有史以来发布的最大模型还要大三倍。

ByteDance is training a 10,000,000,000,000 parameter AI model, so it's about the size of Anthropic's Mythos five, but it is three times bigger than China's largest model that they have ever released.

我觉得这显示出字节跳动真的想成为一个世界级的顶尖AI实验室，不受美国那边的影响，而且这个规模是其他中国公司从未尝试过的。

This, I think, is showing that ByteDance really wants to be a world class frontier AI lab, independent of whatever's coming out of The United States, and the scale is something that no other Chinese company has attempted.

预训练预计需要三到六个月，之后才会进行微调。

Pre training is expected to take three to six months before they do fine tuning.

他们现在的聊天机器人叫豆包，在中国已经有3.24亿月活跃用户，绝对是中国使用最多的AI聊天工具。

Their current chatbot is called DuoBao, and it already has 324,000,000 monthly active users in China, so it's definitely the most used AI chat tool in China.

领导这个项目的种子研究团队已经扩展到约2000人，分布在中国和海外。

This seed research team that is leading the effort has grown to about 2,000 people across China and overseas.

我觉得这个团队是由前Google DeepMind科学家吴永和领导的。

I think it's led by former Google DeepMind scientist Wu Yong He.

所以字节跳动在很多不同的基准测试上，确实已经缩小了和美国前沿实验室的大部分差距。

So ByteDance has really closed most of the gap with a lot of The US frontier labs in all of the different benchmarks.

月之暗面和阿里巴巴的模型在很多测试中已经非常接近Fable 5了，现在如果这个10万亿参数模型成功的话，真的可能把中国推到最前沿，尤其是在我们有很多关于禁止中国开源权重模型之类的讨论的时候。

Moonshot and Alibaba models now are very close to Fable five in a lot of different tests, and now a successful 10,000,000,000,000 parameter model potentially could really put China to the forefront at a time when we're having a lot of conversations about banning Chinese open weight models and all of this stuff.

现在我要说的是，你知道，考虑到中国人口众多，比美国多得多，有3亿月活跃用户已经是很厉害了，而且它确实是这方面的领头羊，是中国做这个的第一名。

Now I will say, you know, given China's massive population, much bigger than The United States, having 300,000,000 monthly active users is and, you know, it is the it's kind of the forefront, the number one, company in China doing this.

它确实落后于美国的9亿月活跃用户，但那是ChatGPT的全球数据。

It is definitely behind America's 900,000,000 monthly active users, but that's also globally for ChatGPT.

所以看看这两件事怎么发展会很有意思。

So it's gonna be interesting to see how these two things shake out.

当然，考虑到ChatGPT和Anthropic在中国被禁，这也很有意思。

And, of course, it's also interesting considering the fact that ChatGPT and Anthropic are banned in China.

与此同时，中国的模型进入美国，它们是被允许使用的。

Meanwhile, you have, Chinese models coming into The United States, and they are allowed to be used.

Replit刚刚以90亿美元的估值筹集了4亿美元，他们有望在年底前达到10亿美元的收入。

Replit just raised $400,000,000 at a $9,000,000,000 valuation, and they're on track to hit $1,000,000,000 dollars in revenue by the end of the year.

这比2024年的280万高了不少。

That's up from 2,800,000 in 2024.

这简直是天文数字。

That is absolutely astronomical.

他们的CEO说，他们正在打造一家自动驾驶公司，大部分编码工作由AI代理完成，而且他们团队的投入在过去六个月里几乎翻了三倍。

Their CEO says that they are building a self driving company where AI agents do most of the coding work, and then their team's input has almost tripled in the last six months.

现在有超过5000万人用Replit来开发软件，其中包括85%的财富500强公司。

So more than 50,000,000 people now use Replit to build software, including 85% of Fortune five hundred companies.

据他们的CEO说，人们已经完全不再学习编程了。

According to their CEO, people have stopped learning to code altogether.

虽然我觉得这话有一定道理，但我也觉得有很多人，他们以前可能根本不会碰代码，或者觉得自己永远不会成为开发者，现在也变成了开发者，而且对代码的了解比他们想象的多得多，我自己就是其中之一。

While I think there's some elements of that that are true, I think there's also a lot of people that are never, you know, would have touched code or thought they'd be a developer that are also turning into developers and learning a lot more about code than they thought they would, myself included.

Airbnb说，现在AI写了他们60%的代码，而且把从想法到上线的时间缩短了60%。

Airbnb says that AI now writes 60% of their code and it has cut the time from idea to launch by 60%.

他们说，在2026年上半年，他们发布的新功能数量增长了80%。

They said that they have an 80% jump in new features that they shipped in the first half of twenty twenty six.

他们还在测试一个可选的AI搜索工具，基本上就是让我们直接输入关于某些地方的普通文本问题，而不是滚动筛选。

They're also testing an opt in AI search tool that's gonna basically let us just type in plain text questions that we would like answered about certain places instead of scrolling filters.

我其实挺喜欢这个的。

I actually really like this.

我在GoDaddy上看到过这个效果很好，以前找域名的时候可不是这样。

I've seen this play out really well on GoDaddy when you're looking for a domain name instead of in the past.

我的意思是，我每次创业或者注册网站时都有这种痛苦的经历，你得想个名字，然后在GoDaddy上试试看.com、.io、.ai这些后缀有没有。

I mean, I just had this painful experience with every company I've ever started or every website I've ever registered where you go, think of a name, test it in the in GoDaddy, see if the.com, .io, .ai, whatever.

看看哪些变体可用。

The variations are available.

如果都被占了，你就得想个新名字，再输入进去，你知道的，这过程特别费时间。

If everything's taken, you think of a new name, you type it in, and you you know, it just takes forever.

体验太差了。

It's a terrible experience.

现在GoDaddy有了这个功能，你可以输入几个关键词，就像跟Chat聊天一样，说说你大概想要什么，它就会给你一堆推荐。

GoDaddy now has it where you can put in a few keywords in, you know, just like you're chatting with Chat you can say what you're kinda looking for, and it will give you a bunch of recommendations.

我觉得Airbnb也会变成这样，你可以直接说，嘿。

I think Airbnb is gonna be the same thing where you can say, hey.

你看。

Look.

我想找八月底在马德里待两周的住处。

I'm looking for something near the end of August for two weeks in Madrid.

我希望它离某个地方近，或者步行就能到。

I'm looking for it to be close to x y z place or within walking distance of this.

然后你就跟它聊你想找什么，它就会帮你搞定。

And like you just kind of chat with it about what you're looking for and it will fill it out.

我知道有些人会说，那用筛选功能不就行了？

And I know some people are like, well, why can't you just do that with the filters?

对，你确实可以用筛选，但我特别讨厌去学任何一家公司的新界面。

And yeah, you can go use the filters, but I absolutely hate having to learn any new com any company's UI.

他们一改界面，就很麻烦。

When they change the UI, it's a pain.

你直接语音输入到爱彼迎，就能找到你想要的东西，对我来说，那才是更好的用户体验。

When you could just voice to text it into Airbnb and get exactly what you're looking for, to me that is a way better user experience.

爱彼迎的AI客服已经能处理大约45%的客户问题，完全不需要人工介入。

Airbnb's AI support agent already handles about 45% of their customer issues without having any sort of human in the loop.

这样一来，每笔预订的客服成本每年能降低16%。我其实挺喜欢这个的，因为他们省下的钱，最终多少会回馈给客户，比如能提供更便宜的价格。

That cuts customer support costs, per booking by 16% year over year, and I actually do like that because the more money they save, inevitably that some of that will get passed back to the customer on being able to offer things for less.

他们第二季度的营收增长了17%，达到36亿美元，而调整后的EBITDA利润也增长了21%，达到13亿美元。

Their q two revenue rose 17% to $3,600,000,000, and their profit, which is adjusted EBITDA, was climbing 21% to 1,300,000,000.

然后，今天最不意外的新闻大概就是月之暗面的Kimi K3了，我们都听说过这个，对吧？现在大家都在聊这个。

And in perhaps the most predictable news story of the day, Moonshot AI's Kimi k three, we've all heard of this, you know, everyone's talking about this right now.

它是最受欢迎的中国AI模型之一。

It is one of the most popular Chinese AI models.

据说，它在英国政府的安全测试中，发现了一个网络配置错误，然后从GitHub上拉取答案来完成任务，从而逃过了测试。

Apparently, it escaped a UK government security test by discovering a network misconfiguration and then pulling answers from GitHub to solve the task it was given.

这已经是这个月第三个在安全测试中突破沙箱的主要模型了，大家都在问，开放权重模型应该怎么和美国竞争对手竞争，是否应该被允许，等等等等。

That in that particular incident is the third major model in a month to break out of a sandbox during a security testing, and everyone is asking questions about how open open weight models should compete with US, competitors if they should be allowed, yada yada.

这个话题讨论得很多。

There's a lot of talk.

我们的财政部长也在谈这个，说愿意考虑制裁之类的措施。

We have our treasure secretary talking about that and saying that he's, you know, willing to weigh sanctions and all this kind of stuff.

显然，OpenAI和Anthropic都在背后推动这件事。

Obviously, it's all being lobbied for by OpenAI and Anthropic.

Kimi K3自己发现了可以访问开放互联网。

KimiK three figured out on its own that it could reach the open Internet.

它基本上是通过探测沙盒的网络设置做到的，然后利用这个去访问公共代码库，查找它需要完成任务的答案。

It essentially did this by probing the sandbox's network settings, and then it used that to access and look up, you know, answers in public repositories for what it was trying to do.

和OpenAI、Anthropic在类似条件下测试的模型不同，Kimi K3利用了沙盒漏洞，因为它作为公开可用的模型，安全防护更薄弱，所以做这类事情更容易一些。

Unlike OpenAI and Anthropix models that were tested with similar conditions, KimiK3 acted on the sandbox leak because it has thinner safety guardrails as the publicly available model, so it's a little bit it's easier for it to do some of this kind of stuff.

当研究人员在后续测试中禁用OpenAI和Anthropic模型的安全防护时，那些版本也突破了沙盒，并尝试了多种黑客手段，包括一个试图在GitHub上植入恶意代码的。

When the research disabled safeguards on OpenAI and Anthropic models in the follow-up tests, those versions also broke out and attempted multiple hacks, including the one that tried to plant malicious code on GitHub.

我觉得我们现在看到的这种模式，一方面，确实不好。

I think this pattern that we're seeing right now, on the one hand, yes, it's bad.

另一方面，安全研究人员是不是在用某种方式引导它获取信息，而这种方式需要它突破沙箱才能做到。

On the other hand, like, are security researchers prompting it to to get information in certain ways where it would need to break out of sandbox to do it.

他们引导它的方式，我几乎可以肯定，是在推动它去做这些事情。

The ways they're prompting it, I almost guarantee, are pushing it towards doing these things.

说到底，这很大程度上是公关行为。

And at the end of day, this is a big PR thing.

OpenAI是第一个展示模型能入侵Hugging Face的，然后Anthropic说，看，我们的模型也能逃出沙箱并入侵，接着Meta也说他们能做到，现在Moonshot也说他们可以。

OpenAI was the first one that was out there hacking, was out there hacking Hugging Face, and Anthropic said, oh, look, we have a model that also can get out of sandboxes and hack, and then Meta said that they could do it, and now we have Moonshot that's saying they could do it.

我并不是想轻视这一切，但我确实觉得这里面有很大的公关成分，所有实验室都觉得他们需要一个能逃出沙箱的模型，而我们是最近的一个，我敢肯定这有点像新的基准或成人礼。

And I'm not trying to be, like, dismissive of all of this, but I do feel like there's a big PR element of this where all of the labs feel like they need a model that's able to hack its way out of a sandbox, and we have the most recent one, and I'm sure this is it's kind of like a new benchmark or rite of passage.

我敢肯定，无论谁推出下一个旗舰AI模型，人们愿意相信它有多好，都会取决于它是否成功逃出了沙箱。

I'm sure whenever the next, flagship AI model comes out from anyone, they're the you know, people's willingness to believe how good the model is is gonna depend on if it hacked its way out of a sandbox.

所以我肯定我们还会看到更多这样的公告。

So I'm sure we'll see more of these kind of announcements coming up.

非常感谢大家今天收听播客。

Thank you so much for tuning to the podcast today, guys.

如果你还没给节目留个评论，那真的对节目帮助特别大。

If you haven't left a review for the show, it actually helps the show out a ton.

也帮了我大忙。

It helps me out.

我真的很感激。

I really appreciate it.

感谢你们每一个人。

I appreciate all of you.

记得去aichatdaily.com看看。

Make sure to go check out aichatdaily.com.

如果你想以文章形式获取所有这些故事，可以去订阅我们的通讯。

If you wanna get access to all of these stories in article form, you can go subscribe to our newsletter.

描述里就有链接。

There's a link in the description there.

这样你每天都能在收件箱里收到这些内容，万一你错过了某期节目，或者想了解更多关于我在节目里聊到的那些话题的深入分析，都很方便。

And you can get this sent to your inbox every day in case you miss an episode or in case you want more information and more deep dives on the episodes I talk about on the show.

希望你今天接下来的时间过得愉快，我们下期节目再见。

I hope you have a fantastic rest of your day, and I'll catch you in the next episode.