探索 GPT 5.5 功能

OpenAI 刚刚发布了 GPT 5.5 的即时版本。

OpenAI has dropped GPT 5.5 instant.

他们还推出了自助广告管理平台，现在正在逐步上线。

They also have self serve ads manager that OpenAI has been rolling out.

格雷格·布罗克曼在法庭上承认，今年他们将在算力上投入五千亿美元。

Greg Brockman was in court admitting that they're gonna spend $50,000,000,000 on compute this year.

这一切都发生在短短六个小时之内。

This all happened in the span of about six hours.

在此之前，苹果终于决定开放 Siri，允许用户接入 Claude 和 Gemini，随便选自己喜欢的模型来运行。

Before that, we had Apple that is finally going to open Siri up to Claude and Gemini and let you basically pick any model you want to run inside of there.

还有其他一些关于苹果的有趣消息。

There's also some other interesting Apple news.

由于苹果智能的营销宣传，他们被罚了两亿五千万美元，涉及所有购买了 iPhone 15 Pro 或 iPhone 16 的用户。

They're being fined $250,000,000 for anyone that bought a iPhone 15 Pro or iPhone 16 because of their marketing campaign around Apple intelligence.

他们得给每位用户退还一百美元，因为宣传承诺没兑现，属于虚假广告。

They're gonna have to pay a $100 back to anyone for basically false advertising since they missed their deadline.

还有一位研究人员发现，Chrome 在你不知情的情况下悄悄下载了一个 4GB 的 Gemini 模型到你的电脑上，类似的情况还有很多。

There's also a researcher that caught Chrome silently downloading a four gigabyte Gemini model onto your computer without asking and so much more.

这些内容我们都会在播客里一一聊到。

We're gonna get into all of it on the podcast.

咱们先从苹果说起。

Kicking this off, let's talk about Apple.

他们终于要开放 Siri 了。

They are finally cracking Siri open.

马克·古尔曼最早在三月就报道过这个框架，但现在《9to5Mac》又出了新消息，确认这个功能真的会在 6 月 8 号的 WWDC 上上线。

Mark Gurman, he first was reporting on this whole framework back in March, but there is a new report that came from nine to five Mac that is confirming that this is actually going to ship at WWDC on June 8.

这会作为 iOS 27、iPadOS 27 和 macOS 27 的一部分推出。

This is gonna be part of iOS 27 and iPadOS 27 and MacOS 27.

所以所有最新的更新都会在六月发布，核心机制叫‘扩展功能’。

So all of the latest updates are gonna be coming out in June, and the mechanism is basically called extensions.

这意味着你可以在苹果智能和 Siri 的设置里，自己决定让哪个模型来处理哪些任务。

What this means is that you're gonna be able to go into settings in Apple Intelligence and Siri, they're and gonna pick what model handles what.

你可以接入 Gemini。

So you can plug Gemini in.

也可以接入 ChatGPT。

You can plug Chat GPT in.

还能接入 Claude。

You can plug Claude in.

有趣的是，你还可以根据响应的模型来设置不同的语音。

And what's interesting is you can actually set custom voices depending on which model is responding.

这种自定义功能应该会挺贴心的，或许也能稍微弥补一下大家的不满——毕竟拖了这么久才推出，偏偏这时候还赶上诉讼：苹果得给每个买了 iPhone 15 Pro 和 iPhone 16 的人退 100 美元，因为当初发布时宣传的 Apple Intelligence 根本没上线，营销全在暗示它会同步推出。

The customization is going to be nice and hopefully a good enough feature to apologize to everyone for why this has all taken so long, especially at the same moment that this lawsuit has come in where, basically, they're gonna have to pay a $100 back to everyone that bought an iPhone 15 Pro and an iPhone 16 because Apple Intelligence didn't ship when they those were launched, and the marketing all basically pointed to it shipping.

明白了。

Okay.

咱们聊聊 Chrome 现在出了啥事。

Let's talk about what's going on with Chrome.

有个叫亚历山大·哈姆的网络安全研究员。

There was a security researcher named Alexander Hamf.

他是一名居住在欧盟的计算机科学家兼律师，发现谷歌浏览器在未经用户同意的情况下，下载了一个四千兆字节的Gemini Nano模型到用户设备上。

He's a computer scientist and a lawyer based in The EU, and he caught Google Chrome downloading a four gigabyte Gemini Nano model onto the user device.

根本没有弹出任何授权提示。

There was no consent prompt.

它根本没问你能不能下载。

There was, like, it didn't ask you.

直接就把这个四千兆的模型悄悄装到了你的电脑里。

It just downloads this four gigabyte model onto your computer.

这个文件名叫weights.bin，存放在一个叫opt_guide_on_device_model的文件夹里。

The file is called weights dot b I n, and it lives in a folder called opt guide on device model.

据汉诺夫说，他是通过查看macOS内核的文件系统日志来确认这个安装行为的。

Apparently, according to Hanoff, he verified the install by reading macOS kernel file system logs.

这件事背后有好多让人担忧的问题。

There's a bunch of reasons why this is problematic.

这位名叫汉诺夫的研究人员指出。

According to this particular researcher, Hanoff.

他说这违反了《电子隐私指令》第5条第3款，该条款禁止在未经用户事先同意的情况下在设备上存储数据，同时也违反了GDPR第5条和第25条。

He says that this violates the e privacy directive article five three, which prohibits storing data on a user's equipment without prior consent, plus GDPR article five and twenty five.

所以，你想想，让十亿多Chrome用户在毫不知情的情况下，白白占用四个GB的存储空间，这实在不太妥当。

So, you know, this isn't really super great to give four gigabytes onto, you know, a billion plus Chrome users' computers, without any sort of, you know, telling them that they could do this.

而且，如果每个AI公司都效仿，随便往你电脑里下载模型，还说这只是软件升级的一部分，那这个先例可就太危险了。

And the precedent isn't great if, you know, every AI company could start downloading models onto your computer and say it's just, you know, part of their software upgrade.

这些模型，比如这个Nano版，虽然现在才四个GB，但以后只会越来越大。

I mean, these models, the nano one, even their four gigs, these things just get bigger and bigger.

所以，这个先例真的不靠谱。

So precedent, not good on this.

说到这么多AI模型，如果你已经为ChatGPT、Cloud Pro、Gemini、Grok，还有Eleven Labs这些音频和图像AI工具付费了，那我得跟你聊聊我开的公司，叫AIBox.ai。

And speaking of multiple different AI models, if you're already paying for ChatGPT and Cloud Pro and Gemini and Grok and maybe Eleven Labs and all of these different audio image AI tools out there, I'd love to tell you about my software company called AIBox dot AI.

我们平台整合了80多个主流AI模型，全都来自顶级实验室，每月只要8.99美元。

We have one platform with 80 different AI models on there, everything from the top AI labs, and it is $8.99 a month.

你只需一个订阅，就能一次性访问所有顶尖的图像、音频、视频和音乐生成模型。

So you get access to all of the top image, audio, video, music models in one place one subscription.

你不需要开一大堆账号，更重要的是，不用在二十个不同的服务上分别付二十美元的月费，花冤枉钱。

You don't have to have tons of different accounts, and most importantly, you don't have to have $20 subscriptions on 20 different accounts to cost you a ton of money.

希望这个能帮你省下不少钱。

So hope this saves you a lot of money.

网址是 aibox.ai。

It is aibox.ai.

这就是我推荐大家一站式使用所有AI模型的方式。

It's how I recommend people get access to all of the different AI models.

全都用一用，测一测，试一试。

Use them all, test them all, try them all.

我特别喜欢用Claude来写东西、处理各种任务，但它没法生成图片，我得切换到ChatGPT才行。

I love using Claude for, writing and tons of different tasks, but it doesn't have an image generator, I gotta go over to ChatGPT to do that.

而ChatGPT又没有生成MP3音频的功能，所以我又得去Eleven Labs那边处理。

And then that doesn't have, you know, an audio m p three file generator, so I go over to Eleven Labs for that.

总之，所有这些功能现在都能在一个平台里搞定。

Anyways, all of those in one platform.

你可以去试试看。

You can go check it out.

链接在简介里。

Link's in the description.

我们收到了不少来自OpenAI的最新更新。

We have a ton of updates from OpenAI specifically.

首先，他们推出了一款新模型，叫GPT-5.5 Instant。

First of all, they have a new model, GPT 5.5 instant.

这将成为ChatGPT的默认模型。

This is going to be the new default on CHAP GPT.

这个更新会覆盖所有免费用户、Pro用户和企业用户，直接取代之前的GPT-5.3 Instant。

This is rolling out to everyone on free plus pro enterprise, and it basically replaces GPT 5.3 instant.

Maxwell Zenith写了一篇详细分析，提到OpenAI一直强调的一个关键数据是——幻觉减少了52.5%。

Maxwell Zenith did a whole write up about this and basically said that the the headline number on this that OpenAI keeps telling everyone is that there is 52.5% fewer hallucinations.

这正是OpenAI对这款模型的宣传重点。

That's the claim that OpenAI is making on this.

而且，这些幻觉减少的情况主要出现在他们认为属于高风险的任务上。

And these like, the lower amount of few hallucinations are specifically on tasks that they are deeming as high stake prompts.

比如医疗、法律和金融领域。

So that's medicine, law, and finance.

他们还提到，在之前用户标记过的棘手对话中，不准确的表述减少了37%。

They also say that there is 37% fewer inaccurate claims on tough conversations that had, you know, previously been flagged by users.

他们仔细分析了那些被标记的内容。

So they were able to kind of look at what was flagged.

然后通过训练，把这个问题改进了。

They trained and made this better.

根据AIME数学测试，分数从65分提升到了81分。

According to the AIME math, it jumped from 65 to 81.

GPQA的分数从78分上升到了85分。

GPQA went from 78 to 85.

CharVIV的分数也从75分涨到了81分。

Char VIV went from 75 to 81.

所以，总的来说，我们的各项基准测试成绩都有了提升。

So, basically, a bunch of benchmarks, we we moved up.

显然，如果模型在这些测试上没进步，他们是不会发布新版本的。

We wouldn't obviously expect them to release a model if it didn't get better on the benchmarks.

我觉得真正重要的是，在法律、金融和医疗这些领域上取得实质进步。

The thing that I do think is is important is specifically getting better at law, finance, and medicine.

这些都是关键领域，一点都不能出错。

Those are some really critical areas that you can't have it, messing anything up on.

他们还提到新增了一个记忆来源面板。

They also said that there is a memory sources panel.

当ChatGPT根据你过去的聊天记录、保存的记忆、Gmail或其他个人数据来个性化回答时，你现在能清楚看到它引用了哪条记忆，还能删除或修正它。

So when ChatGPT personalizes an answer using any of your past chats or any of your kind of saved memory or your Gmail or any of that kind of data, you can now see which memory it pulled from, and you can also delete or correct it.

这一点我特别欣赏，因为我之前遇到过很多次这种情况——比如朋友或我妻子用ChatGPT帮他们做事。

Now this is actually something that I really appreciate because I've had many conversations where, you know, I'll have a friend that, like, used ChatGPT or my wife will use ChatGPT for something.

等我后来再聊起这些事时，它却突然提起那些对话的上下文，说什么‘你不是担心你的丰田卡罗拉吗’之类的。

And now when I'm talking about it, it's, like, grabbing the context of that conversation, and it's like, since you're, you know, I don't know, worried about this your Toyota Corolla, blah blah blah.

我就说，天哪。

And I'm like, oh, man.

我让朋友问ChatGPT关于他的车，结果它还以为那是我的车，一直记着。

Like, I let my friend ask a question to ChatGPT or ask a question to ChatGPT about his car and how it's know, thinks this is my car forever.

对吧？

Right?

这种情况其实挺多的。

So there's a whole bunch of cases where this can happen.

你可以删掉这些记录，我觉得这特别有用。

You can delete that, and I think that's super useful.

另外，彼得·格斯托夫在X上发帖说，他用GPT 5.5在Codex里跑了一个特别复杂的指令，结果一次就完美解决了。

In addition, Peter Ghostov was posting it over on X, and he said that he ran a really complex prompt in Codex with GPT 5.5, and he said that it nailed it in one shot.

所以说，现在各种吹捧都来了。

So, I mean, there's kind of all of the hype.

这事儿还打了广告。

It was that sponsored.

我不太清楚他和OpenAI之间到底是什么关系。

I don't really know what was the relationship between him and OpenAI.

我不确定，但确实有不少人说，Codex这个编程模型表现特别出色，最近进步非常明显。

I'm not sure, but I there's actually a bunch of people that have been saying that Codex, specifically the the coding model, is really good, and they're seeing some awesome, improvements with that.

所以很明显，OpenAI正瞄准编程领域，毕竟Claude之前在这一块几乎是垄断了。

And so, you know, it looks like OpenAI is going after the coding industry, obviously, since Claude has completely dominated that.

OpenAI的另一个大新闻，就是他们正在筹备IPO，华尔街和投资者都在密切关注。

The other big news that we have from OpenAI is obviously that they are targeting their IPO and everything Wall Street is going to be looking at and investors are gonna be looking at.

因此，伴随着这些动向，OpenAI刚刚为所有美国广告主开启了自助广告管理功能。

And so, in line with all of that, OpenAI just flipped the switch on self serve ads manager for all US advertisers.

现在这个功能已经正式上线了。

So, it is live right now.

你可以去 ads.openingeye.com 看看。

You can go to ads.openingeye.com.

我个人觉得，这可能是把产品推给以前那种小众受众的绝佳机会。

Personally, I think this is probably a incredible opportunity to get your products in front of people that previously, you know, I mean, is a very niche audience.

这是按点击付费的竞价模式。

It's cost per click bidding.

系统推荐的起始出价是每次点击3到5美元。

There's recommended starting bids of 3 to $5 a click.

可见OpenAI瞄准的主要是高端产品，毕竟一次点击要5美元呢。

So OpenAI is really targeting, you know, premium more premium products, $5 a click.

律师、牙医、医生这些人，负担得起这个价格。

I mean, lawyers, dentists, doctors, these people can afford it.

如果你卖的是按摩浴缸，那当然也付得起。

If you're selling a hot tub, yeah, you could afford it.

但很多低价商品，比如TMU的T恤，靠这5美元一次的点击根本赚不到钱。

But there's a lot of cheaper items that, you know, T shirts from TMU are probably not gonna be making any money on these $5 OpenAI clicks.

他们还推出了转化追踪API。

There is a conversions API.

有像素追踪，还有一整套为效果营销者设计的工具包。

There's a pixel measurement, and there's kind of a whole performance marketer toolkit.

他们之前设了五万美元的最低消费门槛，那会儿更像是在搞一个试点项目。

They dropped this, and they used to have a $50,000 minimum spend, so that was kind of for, like, a pilot that they were running.

现在连这个门槛也取消了。

They they don't have that either anymore.

根据Axios的Sarah Fisher透露，OpenAI内部的目标是今年靠这个自助广告项目赚25亿美元，到2030年达到1000亿美元。

And according to Sarah Fisher at Axios, OpenAI's internal goal is to get $2,500,000,000 in ad revenue this year and a $100,000,000,000 by 2030 from this self serve ad program.

如果拿这个来对比，2030年的目标差不多就是现在YouTube广告业务的规模了。

So, like, if we're looking at the context of this, the 2030 number is about the size of YouTube's ad business today.

很明显，他们在这件事上非常激进。

So, obviously, they're being very aggressive with this.

AI营销咨询公司TAU的负责人Robert Webster说过一句话，我觉得特别有意思。

Robert Webster, who runs the AI Marketing Consultancy, TAU, said something that I thought was really fascinating.

他说OpenAI现在的处境更接近Meta起步时的状态，而不是Google。

He said OpenAI is closer to Meta's starting position than Google's.

他的意思是，OpenAI目前还缺乏用户搜索意图这个基础。

His argument there is that they, you know, they don't have intent.

比如，用户会在谷歌的平台上搜索内容。

Like, people are doing searches on there.

他们会主动提出各种问题。

They're asking questions.

但其实我不太能理解他的这套说法，他说谷歌掌握用户的搜索意图。

But, you know, Google, they're saying actually, I'm I I don't really understand his argument that much because he's saying that Google has intent.

谷歌能通过用户的搜索记录清楚他们想找什么，但OpenAI的情况更像Facebook，只是坐拥一大堆用户而已。

They know what people are looking for based off of their searches, but OpenAI is more like, oh, we have a whole bunch of users like Facebook.

但大家和ChatGPT对话的时候，会像用谷歌搜索一样向它提问。

But when you're talking to Chad Chappier, you're asking it questions just like you would ask Google.

实际上现在很多人已经用ChatGPT和其他AI模型，来代替谷歌搜索了。

In fact, a lot of people have replaced Google searches with just talking to Chad Chappier and other AI models.

所以我有点不太理解他的这个说法。

So, I don't quite follow his argument there.

我觉得OpenAI的广告业务潜力非常大。

I think OpenAI has a massive potential for their ad business.

好的。

Okay.

还有另一件关于OpenAI的事。

Something else that has happened with OpenAI.

格雷格·布罗克曼刚刚在奥克兰的马斯克与阿尔特曼案中作完证，CNN有报道。

Greg Brockman just finished his testimony in the Musk Altman trial over in Oakland, over on CNBC.

他们当时就在讨论这件事。

They were talking about this.

他们就在现场。

They were in the room.

据格雷格·布罗克曼在宣誓下证实，他个人在OpenAI的股份现在价值约300亿美元。

Apparently, Greg Brockman confirmed under oath that personal stake in OpenAI is now worth about $30,000,000,000.

另外，我觉得有意思的是，他说OpenAI今年将在算力上投入大约500亿美元。

And, also, what I think is interesting, he said that OpenAI is gonna spend about $50,000,000,000 on compute this year.

这还只是算力方面的开支。

That is I mean, just compute.

这太惊人了。

That is wild.

2017年那会儿才三亿美元，现在到2026年就要飙到五千亿美元了。

The number was about 30,000,000 in 2017, and now we're going to 50,000,000,000 in 2026.

去年已经是几十亿了，我觉得今年就直接到五千亿。

It was tens of billions last year, and I think 50 this year.

这个增幅实在太大了。

That is a huge jump.

他们之前还公布过，到2030年之前要砸六万亿美元，现在看来，他们宣布的总支出已经突破一万亿了，只是我们一直不清楚今年具体会花多少。

I think they also previously disclosed a $600,000,000,000 spend pledge through 2030, and they said over a trillion and change in basically all of their spending has been announced, but we didn't know exactly what was happening this year.

这规模确实不小。

So that's that's pretty big.

顺便说一句，就在这个时候，三星今天也突破了一万亿美元市值，这主要得益于AI内存需求的爆发。

Also, as a just as a side note, this is happening at the exact same time that Samsung just crossed a trillion dollar market cap today, and this was all basically on the back of AI memory demand.

他们的利润翻了八倍。

They did an 8x profit.

他们的利润因为这个增长了八倍。

Their their profit went up 8x because of this.

现在苹果也在和美国的芯片制造商洽谈。

And now we have Apple that is in talks with US chip manufacturers.

我觉得OpenAI这笔开支的后半段，也会变成其他公司的收入。

So I think kind of the back end of OpenAI's bill is also gonna be someone else's revenue line.

对吧？

Right?

OpenAI打算花掉这五千亿美元。

Like, OpenAI is gonna spend this $50,000,000,000.

这笔钱会流向三星这样的公司，它们正拼命接单，为数据中心扩充内存。

It's going to companies like Samsung that are, you know, gobbling up that money to help them add memory into data centers.

所以每当你看到这种巨额支出的公告，也值得想想钱最终去了哪儿、谁受益了，以及整个行业、华尔街和市场会因此有什么反应。

So whenever you see these massive announcements of spending, it's also interesting to think about where that spending goes, who benefits, and what the industry and Wall Street and the market is gonna do because of that.

好了。

Alright.

今天的内容就到这里了。

That's it for the show today.

非常感谢你们收听。

Thank you so much for tuning in.

如果觉得这期播客对你有帮助，给个评分或评论会对节目有很大帮助。

If you got any value out of the podcast, it would help the show out an absolute metric ton.

无论你在哪个平台听播客，比如在Spotify上，或者在Apple的‘关于’页面，都欢迎留下评价。

If you could leave a rating or review wherever you get your podcasts on Spotify, it's the about tab on Apple.

写个评论，给几颗星都可以。

You can leave a review, drop some stars.

这真的能帮到节目。

It helps the show out.

再次感谢，祝大家剩下的时间都过得特别愉快。

Thank you so much, and I hope you all have an amazing rest of your day.

还没去过aibox.ai的话，记得去看看。

Make sure to go check out aibox.ai if you haven't already.

希望它能帮你省下一大笔钱，还能让你在一个地方访问80多个AI模型。

Hope it saves you a ton of money and gives you access to over 80 AI models in one place.

好的。

Alright.

下期节目见。

Catch you in the next episode.