OpenAI 推出 GPT-5.6，分三个层级，价格低于 Claude。

欢迎收听本节目。

Welcome to the podcast.

今天节目中提到，OpenAI 因为政府要求暂缓发布，决定推迟 GPT-5.6 的上线，以便进行安全测试。

Today on the show, OpenAI is limiting the GPT-five 0.6 rollout because the government has requested that they push it back while they test it out.

正因为如此，欧洲正在推动人工智能自主权，因为特朗普对 Fable 5 实施了出口管制，现在 OpenAI 的 GPT-5.6 也被要求必须经过测试才能向全球发布。

Because of that, Europe is pushing AI sovereignty because Trump is doing export controls on Fable 5and now we have OpenAI GPT-5.6 that they say they need a test before it can go out to the world.

另外，有一家叫 Patronus AI 的公司刚刚融资了五千万美元，专门用于在模拟环境中对 AI 代理进行压力测试。

And also, we have a company called Patronus AI that just raised $50,000,000 to stress test AI agents in simulated worlds.

所以你看，这些新推出的模型能力越来越强，政府也开始说：等等，让我们先看看。

So, again, we're having this thing where, you know, these these models are super powerful that are coming out, the government says, hey.

他们想亲自评估一下，同时也有新公司正在融资，专门做这些模型的压力测试。

Let us look over them, and we have these new companies that are raising money to stress test them.

OpenAI 将以三个不同层级推出 GPT-5.6，并且会在价格上低于 Claude，我认为这和他们的部分使命是相符的。

OpenAI is gonna be shipping GPT-5.6 in three different tiers, and they're gonna be undercutting Claude on price, which I think is important given parts of their mission.

还有一家新公司 Micron，因为 AI 内存需求激增，市值已经超越了 Meta 和特斯拉。

And Micron, a new company, is overtaking Meta and Tesla in market valuation because AI memory demand is just exploding.

我说的新公司，其实不是那种刚成立的初创企业，而是最近才进入大众视野的公司，因为它在供应链中的地位太关键了。

And when I say new company, this is a new company, but it's a new company for many people in the zeitgeist where we're starting to talk about it because of how important it is to the supply chain.

咱们来聊聊OpenAI最近在搞什么。

Let's get into what's going on with OpenAI.

说到OpenAI，如果你用过ChatGPT，但心里嘀咕：要是它能直接生成语音就好了，或者你心想：

And speaking of OpenAI, if you've ever wanted to use ChatGPT, but you're like, man, I wish it could do audio generation right inside of it, or you're like, shoot.

真遗憾ChatGPT要下架那个视频模型，你特别想在里头用Google的V3，或者你用Claude，纳闷它怎么就不支持视频、图片或音频功能？

I'm, you know, really bummed that ChatGPT is getting rid of sore other video model, and you wanna use Google v o three inside of it, or perhaps you use Claude, and you're like, man, why doesn't Claude do video image or audio?

我特别推荐你试试我自己的创业项目aibox.ai。

I would love for you to try out my own startup aibox.ai.

我们新推出了MCP连接功能，能直接接入你的Claude、ChatGPT或Gemini，让你在熟悉的那个工作平台里，一键调用80多个不同的AI模型。

We have a new MCP connection that you can connect into into your Claude, into your ChatGPT, or into your Gemini and give it access to over 80 different AI models right inside of whatever the workstation you use and whatever AI model you're the most familiar with.

我特别喜欢这个功能，因为每次用Claude，我都烦它不能生成图片，总得切换到ChatGPT去画图。

I love it because when I'm using Claude, it's super annoying that it doesn't generate images, and I frequently have to, you know, go over to ChatGPT and and generate the images over there.

所以我在AI Box上自己搭了几个图像生成工具，专门用来处理手头的几个项目。

So I've actually been able to create some tools on AI Box, some image generation boxes for different projects I have.

在Claude里，我只要让它调用AI Box，启动我创建的工具，就能直接生成图片，链接还会直接显示在Claude里。

And inside of Claude, I just tell it to go use AI Box and call the tool that I've created, and it can generate the images, and it gives me a link right inside of Claude.

如果你想让Claude也能生成图片、音频和视频，不妨试试AI Box的MCP功能。

So if you wanna give Claude access to generating images, audio, and video, go check out the AI Box MCP.

每月只要8.99美元，还不到十块，比ChatGPT或其他订阅服务都便宜。

It's $8.99 a month, so it's less than $10, less than ChatGPT or any of the other subscriptions.

而且说真的，你完全可以拿它来替代ChatGPT、Gemini或者其他平台。

And by the way, you can basically use it to replace ChatGPT or Gemini or anything.

它集成了八十多个不同的模型。

It has 80 plus different models.

里面还有Grok。

It has Grok.

也有Mistral。

It has Mistral.

还有很多优秀的开源模型，专门用于音频、图像和音乐生成。

It has tons of cool open source models for audio image and music.

去 aibox.ai/mcp 看看吧。

Go check it out aibox.ai/mcp.

好的。

Okay.

咱们聊聊 OpenAI 最近的情况吧。

Let's get into what's going on with OpenAI.

最大的新闻当然是，政府阻止他们发布最新模型 GPT-5.6。

The biggest news here, of course, everyone's talking about the fact that the government is, forcing them to not roll out the latest model, GPT-five point six.

他们认为这应该只是暂时的。

They're arguing that this should be temporary.

这是因为华盛顿现在正加强对各种 AI 模型的监管。

This is coming because Washington right now is, you know, really tightening their oversight on a lot of these different AI models.

我觉得 Anthropic 最早被叫停，是因为他们之前发布了 Fable 5，政府就出面了。

I think Anthropic had the biggest pullback because it released Fable five before, you know, and the government was like, hey.

你瞧。

Look.

绝对不能让任何有外国访问权限的人拿到这个。

You can't let anyone with foreign access get that.

Anthropic那边也说，喂。

And Anthropic was like, hey.

你看。

Look.

我们认为这完全是个误会。

We think this is a big misunderstanding.

你们不该来管我们。

You guys shouldn't be regulating us.

有意思的是，就在上周，或者前几天，Anthropic还向某个政府机构提交了一份报告，说：喂。

What's interesting is we also have Anthropic that just last week, put out a report or just a few days ago, put out a report to, some government agency where they're like, hey.

你看。

Look.

中国的阿里巴巴曾用大约两千万条消息对我们模型发动过模型蒸馏攻击。

Alibaba over in China did a model distillation attack on us with, like, 20,000,000 messages to our model.

他们注册了两万五千个虚假账号，发了两千万条消息，打算用这些数据低成本训练自己的模型，这太危险了，诸如此类。

They made 25,000 fake accounts, and they did 20,000,000 messages, and they're gonna use those messages to train their model for way cheaper, and this is dangerous, yada yada.

然后他们还说，搞不懂为什么政府不让我们上线Fable Five。

And then they're like they're like, well, I don't know why, you know, the government is trying to not let us roll out Fable five.

很明显，政府担心Fable Five一发布，阿里巴巴就会趁机下手——当然，也不只是阿里巴巴。

Well, obviously, the government's worried that when Fable five gets out, Alibaba's gonna go and, you know or not just Alibaba.

不过，他们可能还算个相对负责任的玩家。

They're probably a, I mean, semi good actor in this.

我不太确定。

I don't know.

但肯定还有更多恶意行为者，随时可能发起同样的模型蒸馏攻击。

But there's probably a lot of other bad actors that could do the same model distillation attack.

他们可以针对Fable Five，或者任何其他模型下手。

They could do Fable five or anything else.

罗伯特的立场其实有两面性。

So and to Robert has two sides of the messaging.

他们说，你瞧。

They're like, look.

我们的模型太厉害了。

Our models are so good.

这玩意儿迟早要毁了这个世界。

They're gonna destroy the world.

然后他们就说，哇哦。

And then they're like, woah.

那我们为啥还被监管？

Why are we being regulated?

现在呢，政府觉得自己能决定哪些模型好、哪些模型坏，也许能，也许不能。

Anyways, so now we have this environment where the government thinks it can decide what models are good and bad, and maybe they can, maybe they can't.

我不确定。

I don't know.

很多人对这事看法不一。

A lot of people have different opinions on that.

GPT-5.6 每百万输入令牌大约五美元，输出令牌则是三十美元。

GPT-five point six costs about $5 per million input tokens and $30 per million output tokens.

Terra 是他们下一个版本，抱歉，他们一共有三个。

Terra, which is their next one so they have three sorry.

他们分三个档次。

They have three tiers.

分别是 Sol、Terra 和 Luna。

They got Sol, Terra, and Luna.

Sol 是五美元那个。

Sol is the $5.

Terra 是两块五，输出十五美元；Luna 是一块和六美元。

Terra is $2.50, and it is then $15, and then Luna is $1 and $6.

Luna 是最便宜的。

So Luna is the cheapest.

Sol 是性能最好的。

Sol is the best.

OpenAI说，Sol在编程能力上基本和Anthropic的Claude 3 Mythos相当，而且他们提到，构建东西时只用了三分之一的输出token。

OpenAI says that Sol basically matches Anthropic Claude Mythos five on coding, and they say that they're using third a third of the tokens of the output tokens to build things.

他们有安全防护机制，而且强调这些防护是内置于核心模型行为里的，而不是额外加一个过滤层。

They have safety guardrails, and they say that's all inside of the core model behavior rather than adding this as a separate filter.

所以Anthropic的做法是，你看——

So with Anthropic, they're like, look.

我们有Mythos 5，

We have Mythos five.

它非常危险，但我们会在上面加一堆防护措施。

It's super dangerous, but we're gonna add all these guardrails on top of it.

这样就没问题了。

It'll be fine.

而OpenAI说，你看——

And OpenAI says, look.

我们早就把防护机制直接做进模型里了，根本不用额外处理。

We already built the guardrails inside the model so you don't need to do that.

我的意思是，这真的吗？

I mean, is it true?

那就得看时间怎么说了。

Well, time will tell.

对吧？

Right?

还是说有人能黑进去？

Or is anyone able to hack it?

我们迟早会知道的。

We'll find out.

他们说这模型和Mythos五代一样强，而且输出的token还少三分之一，如果真能降低成本、提升效率，那就太惊人了。

But they're saying that it's just as good as the Mythos five model, and they say that it is one third fewer output tokens, which would be incredible if it's actually cheaper and more efficient.

最近特朗普的一项行政命令要求所有家具公司必须在发布前三十天提交模型供政府审查。

A recent Trump executive order required all furniture companies to submit models for government review up to thirty days before they release.

但那里并没有明确的安全标准。

There is no defined safety standards specified there.

所以，不管怎样，这件事真是闹得沸沸扬扬。

So, anyways, again, just a lot of drama on that.

正因如此，欧洲现在开始大力推动人工智能自主化。

And because of all that drama, Europe is now pushing for AI sovereignty.

特朗普出台了出口管制，而法国刚刚拿到了一千亿欧元的人工智能基础设施承诺。

Trump put in these export controls, and France just secured a €100,000,000,000 in AI infrastructure pledges.

这笔投资主要由软银主导，他们正在欧洲建设一个七千五百亿美元的数据中心。

This is anchored by SoftBank's $75,000,000,000 data center bet that they're making right now inside of Europe.

欧洲这边，我觉得很多人把这事搞得特别夸张，说什么美国背叛了欧洲之类的。

Europe is, I think, getting honestly, people make this like a big, huge, dramatic thing where they're like, oh, you know, The United States has betrayed Europe or whatever.

美国该干嘛还干嘛，管不了那么多。

And The US is just gonna do what The US is gonna do.

咱们这儿乱七八糟的事一堆，但技术实力确实没得说。

I mean, we're we got all sorts of crazy stuff going on over here, and we also have all the technology.

归根结底，我觉得发展自主人工智能是件好事。

At the end of the day, I think that sovereign AI is a fantastic thing.

我觉得每个国家都该有自己的数据中心、自己的电力供应、自己的AI模型，不该依赖别人。

I think every country should focus on having their own data centers, their own power supply, their own AI models, and shouldn't be, you know, reliant on anyone else.

所以归根结底，这事儿其实没什么不好。

So I think at the end of the day, this isn't a bad thing.

Anthropic的Claude模型被下架，也不至于闹得这么大，搞得跟什么大事儿似的。

This isn't like some big drama needs to happen just because Anthropics Claude you know, Fable model got pulled.

很多人觉得，美国把Fable模型从欧洲用户那儿撤了，就是背叛了欧洲。

Because I think a lot of people are like, you know, The United States, like, betrayed Europe by pulling the Fable model from European users.

其实人家是全球一起撤的，我们这边现在正吵得厉害，主要是政策上的分歧。

I mean, they just pulled it from everyone in the whole world, and it seems like there's kind of a big policy debate that we're having over here.

但不管怎么说，自主的AI模型确实很棒。

But in any case, I think that sovereign AI models are fantastic.

我很高兴软银在支持这些项目。

I'm excited that SoftBank is backing a bunch of those programs.

马克龙搞了个叫‘选择法国’的计划。

Macron has this thing called Choose France.

他们推出了一项计划，打算建设一批自己的数据中心。

It's an initiative that they announced where they're be building out a bunch of their own data centers.

太棒了。

Fantastic.

Anthropics 单独一轮融资就筹集了 650 亿美元，超过了去年整个欧洲和英国所有 AI 初创公司融资总额的总和。

Anthropics' single $65,000,000,000 fundraise exceeded the entire sum, invested in all European and UK AI startups last year combined.

归根结底，虽然各国都在推自己的计划，但跟美国一家公司就能融到的钱相比，这些投入还是九牛一毛，更何况美国还有这么多其他公司。

So at the end of the day, like, there's all of these national programs that are going on, but there's still, like, a drop in the bucket compared to what just one company in The United States was able to raise, and then we have a lot of different companies.

所以我觉得，欧洲真正该专注的是打造自己的体系，这些国家其实有很多创新机会——因为资金终于会流向本地的 AI 公司，而不是一味地往 OpenAI、Anthropic 这样的巨头身上砸钱。

So I think at the end of the day, Europe should build out what focusing on, and I think that there's a lot of opportunity for innovation to come out of these countries because they'll you know, people will finally be funding national, like, AI, like, I guess smaller AI companies for geo geographic reasons rather than everyone just dumping more money into the mega caps, the OpenAI, and the Anthropic.

我很期待看到市场上出现更多竞争。

So I'm excited to see more competition in the marketplace.

就在大家热议这些 AI 模型的种种风波、安全性与能力的时候，一家叫 Petronas AI 的新公司刚完成了 5000 万美元的 B 轮融资。

One so with all of this drama around these AI models, you know, coming out and how dangerous are they and how good are they, there's a new company called Petronas AI, and they just raised $50,000,000 as their series b.

他们的做法是尝试模拟数字世界。

And the way the what they're doing is they're trying to simulate digital worlds.

他们基本上是在部署前对新的AI代理进行测试。

So they're basically testing the new AI agents before they get deployed.

你可以想象，政府那边说：等等。

You can imagine, right, the government was like, look.

你们得给我们三十天时间，别急着发布。

You gotta give us thirty days before you put it out.

我心里就想，你们这三十天能干啥？真能测出这些系统的问题？

And in my mind, I'm like, what are you guys doing in thirty days that you could possibly test these things?

看来Petronas AI可能是个解决方案，他们模拟真实世界，把AI模型放进去，故意让它们制造混乱，看看会出什么问题。

Apparently, Patronus AI could be an answer to that, where they're simulating the real world, and they're taking these AI models, and they're just telling them to wreak havoc and seeing what they could actually do.

也不是真让它们搞破坏，主要是做压力测试。

I mean, not really just wreak havoc, but they're trying to stress test it.

对吧？

Right?

他们建了个模拟环境，然后放进去让它自己运行。

So they're building a simulated world and letting it run there.

去年他们的收入增长了十五倍。

Their revenue grew 15 x in the last year.

现在有不少AI实验室都在做这类测试，我觉得这家公司在这方面做得特别出色。

There's a bunch of different AI labs that are trying to get these kind of, testings done, and I think this is, this is a great company that's working on it.

Petronas建立了网站和内部系统的数字副本，让AI代理可以在强化学习环境中训练。

Petronas has digital replicas of websites and internal systems where agents can train under reinforcement learning.

系统会奖励任务完成，也会惩罚错误操作。

It's gonna reward task completion and it's gonna penalize errors.

这挺有意思的，对吧？

So it's interesting, right?

他们说的数字世界，不是那种字面意义上的、像扎克伯格的元宇宙那样的东西。

They have like, you know, like when they say they've made a digital world, it's not like they've, like a literal, it's not like you know, Mark Zuckerberg's metaverse.

他们构建的是一个由计算机、系统和代码组成的世界，我猜他们很可能收集了一大堆开源项目，搭建了大量环境，让这些模型去互动、完成任务，大概就是做渗透测试这类工作。

They're making a world of computers and systems and code, and they've probably grabbed a whole bunch of open source projects, if I'm being honest, and spun up a lot of, you know, a lot of stuff where these models have to try to interact, get tasks done, to probably do penetration testing, that kind of stuff.

这一轮投资中，Notable Capital、Lightspeed、Datadog和三星都加入了Greenfield Partners，这家公司的总融资额大约七千万美元。

So Notable Capital, Lightspeed, Datadog, Samsung, all of them joined Greenfield Partners in this round, and the total funding for this company is about $70,000,000.

这家公司是2023年才成立的。

It was just founded in 2023.

我很期待看到他们接下来会走到哪一步。

I'm excited to see where they go with this.

由于AI内存需求激增，美光科技的市值刚刚超过了Meta和特斯拉，这可不是小事，实际上，这个问题关乎所有人。

A company called Micron has just overtaken Meta and Tesla in market value because of the crazy demand for, AI memory that's you know, I mean, this isn't just a this is a problem basically for everyone.

昨天苹果刚宣布，所有MacBook都要涨价——MacBook Air涨100美元，MacBook Pro涨300美元，说白了，现在到处都在涨价，全是因为AI导致的内存严重短缺。

We just had Apple yesterday say that they're raising the prices on all MacBooks by a $100 MacBook Airs and then $300 on the MacBook Pros, and, like, basically, everything's going up in price because there's this huge memory shortage everywhere which is caused by AI.

现在对高带宽内存——也就是HBM的需求简直爆炸式增长，而这正是AI训练集群的核心组件。

It's just, like, really explosive demand for high bandwidth memory or HBM, and this is, of course, used for AI training clusters.

很多人提到，这说明现在制约AI基础设施发展的，不只是GPU，内存同样成了关键瓶颈。

So something that a lot of people are bringing up is that this is basically showing that memory and not just GPUs alone are now kind of gating and or kind of like a constraint for the AI infrastructure build out.

过去大家总在担心：我们怎么才能搞到足够多的英伟达显卡？

I think in the past everyone was like, oh man, how are we ever gonna get enough GPUs from NVIDIA?

而英伟达对每家公司能卖多少显卡，本身也有严格限制。

And NVIDIA has like, you know, limits to how many they'll sell to a specific company.

过去几年，大家担心的其实就是这个问题。

And that was kind of, I guess, the last couple years what everyone's been worried about.

现在不再是GPU不够了，而是内存真的快用完了，这会严重拖慢AI基础设施的建设。

No longer is it GPUs, now we're just literally running out of memory and and that's gonna kind of hold back the AI infrastructure build out.

所以那些掌控内存基础设施瓶颈的供应商肯定会赚得盆满钵满，美光就是其中之一。

So I think these suppliers that are really kind of have a bottleneck on the memory infrastructure are gonna be making a ton of money, and Micron is one of them.

顺便说一下，HBM是专门为英伟达加速器定制的，都是通过多年供货协议来销售的。

And by the way, HBM is basically a custom something that is custom engineered for NVIDIA accelerators, and it's sold under multiyear supply agreements.

这种内存定价极高，传统DRAM根本没法比。

They sell this at a premium pricing, that basically traditional DRAM can never get.

目前全球能大规模生产HBM的公司只有三家。

So only three companies can actually manufacture HBM at scale.

分别是美光、SK海力士——我记得是韩国的——还有三星。

That's Micron, SK Hynix, which is, I believe, in Korea, and then also Samsung.

听说SK海力士最近打算在美国上市，三星也在考虑这件事。

And I believe SK Hynix is is planning an IPO in The United States soon or at least looking at this, and then Samsung.

所以，由于这两年需求激增，供应商现在拥有极大的定价权。

And so, basically, there is some enormous price power for suppliers because there's this kind of two year demand surge that's been coming.

能生产这种产品的供应商本来就没几个。

There's not a lot of suppliers that are able to make this.

像Meta和特斯拉这样的大型云服务商，现在估值反而低于它们所依赖的供应商美光。

Meta and Tesla both, which are huge hyperscalers, now rank below the component supplier Micron, which both of them are using.

对吧？

Right?

它们俩都离不开美光。

They're both dependent on Micron.

这恰恰说明，AI的未来增长多么依赖基础设施。

But I think that just goes to show how much projected AI growth is tied to infrastructure.

它们得从外面采购这些组件，所以美光股价这波涨势真是惊人。

You know, they actually have to go buy this externally, and so it's incredible to see the run up that Micron has been doing.

非常感谢大家今天收听这个播客。

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

如果你学到了新东西，或者喜欢这期节目，记得在苹果播客上给个评分和评论。

If you learned anything new, if you enjoyed the episode, make sure to leave a rating review wherever you get your podcast on Apple.

也欢迎在下面留言。

Leave a comment.

我每一条都会看。

I read them all.

如果有你感兴趣的话题或公司，或者想让我多聊的内容，别忘了点个赞。

Drop some stars if there's any topics or companies you want covered or or things you'd like me to talk about more.

别忘了去看看 aibox.ai/mcp。

And make sure to go check out aibox.ai/mcp.

简介里我会放个链接，帮你把图片、音频和视频功能接入 Claude 或其他模型。

I'll leave a link in the description if you wanna get images, audio, and video inside of Claude or any of your other models.

无论你用的是哪个 AI 工具或模型，都能通过它接入 80 多种不同的 AI 模型。

You can get access to over 80 different AI models inside whatever AI tool or ever whatever AI model you're using.

好了。

Alright.

非常感谢收听。

Thanks so much for tuning in.

我们下一期再见。

I'll catch you all in the next episode.