Anthropic 的 Mythos 回归，OpenAI 发布 GPT 5.6，苹果上调价格

Mythos回来了，OpenAI最强的模型GPT 5.6也发布了。

Mythos is back, and OpenAI's most powerful model, GPT 5.6, is out too.

但公众真能见到这两者吗？

But will the public ever get to see either of them?

苹果为了跟上内存价格上涨，大幅提高了价格。

And Apple hikes prices massively to keep pace with memory prices.

这是贪婪吗？

Is it greed?

这些内容将在本周末的《科技大观》节目中播出，马上继续。

That's coming up on a Big Technology podcast Friday edition right after this.

面对持续的变革与机遇，TMT领域的领导者需要拿出实实在在的成果，而不仅仅是想法。

In the face of ongoing disruption and opportunity, TMT leaders need to deliver tangible results, not just ideas.

当速度和表现至关重要时，普华永道将市场洞察、行业专长与人工智能、云计算和前沿技术结合，加速您的转型，从战略到执行，实现可衡量的投资回报。

When pace and performance matter most, PwC combines market insights and deep sector experience with AI, cloud, and emerging tech to accelerate your transformation and drive measurable ROI from strategy to execution.

普华永道能帮您预判下一步，超越变革，赢得竞争。

PwC can help you anticipate what's next, next, outpace disruption, and compete.

如需更多信息，请访问 pwc.com。

For more information, visit pwc.com.

欢迎收听《大科技》播客周五特辑，我们以一贯冷静、理性的风格为您深度解析最新动态。

Welcome to Big Technology podcast Friday edition where we break down the news in our traditional, cool headed, and nuanced format.

今天这期节目内容很精彩，虽然比平时晚了一些，但咱们还是准时上线了。

We have a great show for you today coming a bit later than usual, but we are here on the feed.

今天我们聊聊 Anthropics 的 Mythos 回归，以及 OpenAI 推出的 GPT 5.6。

We're gonna talk about the return of Anthropics Mythos, the debut of OpenAI's GPT 5.6.

顺便说一句，OpenAI 可能要把上市时间推迟到 2027 年，也可能根本不上市。

OpenAI, by the way, may be delaying its IPO to 2027 or maybe it won't.

这个我们待会儿细说。

We'll discuss that.

苹果把 MacBook 等设备的价格上调了百分之十几到百分之三十不等。

And Apple hikes the prices up on MacBooks and other devices by anywhere between the teens and 30%.

所以我们来聊聊这背后的原因。

So we'll talk about what's behind that.

很多消费电子产品涨价的背后，其实都是内存供应紧张的问题，这到底是不可避免的，还是像有些人说的那样，纯粹是贪婪作祟？

The memory crunch that's underlying a lot of the price increases you might see in consumer electronics, and whether it is unavoidable or as some might say, greed.

今天一如既往，我们请来了Margins的Ranjan Roy来和我们聊聊。

Joining us as always to do it is Ranjan Roy of Margins.

Ranjan，很高兴见到你。

Ranjan, great to see you.

我也很高兴见到你，Alex。

Great to see you, Alex.

你所在的机构在那一百家能用上Mythos的美国机构、大公司和政府单位里吗？

Is are you on the Mythos list out of the 100 US institutions, major companies, and government agencies that now have access to Mythos?

我不在。

I am not.

如果政府真在挑赢家和输家，那我毫无疑问是那个输家。

If the government is picking winners and losers, then I am without a doubt a loser.

那咱们先从这条 headline 开始聊起。

So let's go through the headline here.

这件事我们现在正在实时跟进。

It's something that we have breaking as we speak.

美国方面已通过Semaphore向部分美国企业发布Anthropic的先进模型Mythos。

The US releases powerful anthropic model Mythos to some US companies from Semaphore.

美国政府于周五解除对Anthropic旗下强大AI模型Claude Mythos五代的限制，允许该公司向超过一百家美国机构开放，包括大型企业和政府机构。

The US government Friday lifted its block on Anthropics' powerful Claude Mythos five AI model, allowing the company to release it to more than a 100 US institutions, including major companies and government agencies.

这封发给Anthropic的周五信函，标志着特朗普政府与全球最有价值的私营企业之一之间的紧张关系出现重大缓和。

The decision in a Friday letter sent to Anthropic is a major deescalation in the confrontation between the Trump administration and one of the world's most valuable private company.

这封信对Fable五代未置一词，那是Mythos的较弱版本，曾短暂成为消费者能接触到的最强AI模型。

The letter is silent on Fable five, a weaker version of Mythos that was briefly the most powerful AI model available to consumers.

知情人士表示，双方正朝着开放Fable五代的方向推进，但具体时间尚不明确。

People close to the talks said they are moving toward releasing Fable as well, though the timeline is unclear.

商务部长霍华德·卢特尼克已正式批准这一决定。

Howard Lutnick, the commerce secretary, has given his stamp of approval.

Ranjan，看来Fable的长期停用就要结束了，这场风波也该画上句号了。

Well, Ranjan, it looks like this long outage of fable is about to end and so ends the saga.

我不清楚有没有具体信息，说明美国政府要求的那些安全机制——比如防越狱的防护措施——是否真的已经落实了。

I don't know if we have any details about whether there have been actual structural fixes or whether the guardrails or the anti jailbreaking guardrails that the US government has asked for are there.

你对这则新闻怎么看？

What do you make of the news?

对我来说，最难理解的是整个监管格局：到底谁在什么时候掌握着什么权力。

It's for me, the most difficult part of trying to process all this is I mean, on the overall regulatory landscape of who has control of what and when.

我觉得我们之前讨论过，Mythos到底是营销噱头和制造恐慌，还是真的对全球秩序构成威胁？

I think, I mean, we've talked about is Mythos more marketing and fear mongering, or is it a true danger to kind of the global structural order?

我真的说不准。

And I don't I don't know.

你觉得政府现在对Anthropic到底有没有一个清晰的应对策略？

It's like, do you feel that the government is actually has any real plan around how they're approach dealing with anthropic now.

这看起来就像是当初‘Project Glasswing’的延续。

Like, it looks like this just feels like an extension of what originally Project Glasswing was.

当初是二三十家公司，现在一下子变成了一百多家。

Whatever that was, 20 to 30 companies, now we're at a 100.

但我真不太清楚这背后的方向是什么，到底有什么根本性的变化。

But I don't really understand where this is going or what's fundamentally changed.

没错。

Right.

说到底，结合这周OpenAI的新闻来看，政府实际上已经把自己变成了前沿模型的审批方。

Well, it does seem basically, you know, when you take it in combination with this week's OpenAI news that the government has installed itself as the approval of frontier models.

我觉得咱们得好好聊聊这背后的意义。

I think we should talk about the implications there.

咱们先说说OpenAI的新闻吧，我觉得它和Anthropic的新闻其实是紧密相关的。

Let's get into the OpenAI news because I feel like it really couples together with, the Anthropic news.

这是来自OpenAI一篇博客文章，预告了GPT-5.6 Soul这款新一代模型。

So this is from a blog post from OpenAI previewing GPT 5.6 Soul, a next generation model.

今天，我们正式启动GPT-5.6系列Soul——我们的旗舰模型的有限预览。

Today, we are beginning a limited preview of the GPT 5.6 series Soul, our flagship model.

Terra是一款均衡型模型，适合高效、高并发的任务；Luna则是快速又实惠的日常使用模型。

Terra, a balanced model for efficient, high volume work, and Luna, a fast affordable model for everyday work.

我先停一下。

I'm just gonna pause here for a moment.

他们把这两个或三个模型分别叫做Terra和Luna。

They're calling their two or their three models Terra and Luna.

这难道不是史上最糟糕的命名失误吗？

Isn't that the, like, the worst branding mistake ever made?

你。

You.

这第一个就让我觉得特别扎眼。

That that was the first thing that just jumped out dramatically to me.

对没记住的听众说一下，这两个名字以前是两种加密货币的名称，一个暴跌了99.999%，另一个Terra跌了98%，而且Terra还是个稳定币。

For for listeners who don't remember, those were the names of cryptocurrencies that crashed dramatically, one by 99.999%, the other by Terra by 98%, and Terra was a stable coin as well.

它们当年可是加密货币崩盘的典型代表。

These were kind of the poster children for the crypto fallout from a couple years ago.

所以，比起他们之前的代号‘Spud’，我宁可选这两个名字，但问题是，他们到底是怎么想出这两个名字的？

So, like, why I mean, I'll take that over Spud, I guess, their code name, but how do you think they came up with those names?

其实都是。

All of them, actually.

他们大概是想搞个类似的故事体系，像十四行诗、交响曲、寓言那种感觉。

Think they were trying to find some version of, like, the anthrop anthropics all are, like, a version of a story, like a sonnet and an opus and a fable.

所以就用了地球和太空相关的名字，比如Terra、Luna和Soul，想自己打造一套品牌叙事。

So they were trying to make their own with these earth and space celestial nicknames like Terra, Luna and Soul.

他们显然也知道Anthropic在品牌命名上已经赢了。

They're obviously like they have are wide open to the fact that Anthropic has beat them on branding.

但拜托了，千万别用Terra和Luna这俩名字。

But I would just say for goodness sakes, don't use Terra and Luna.

你们能不能看看形势？

Read the room, guys.

看看形势。

Read the room.

天哪。

My god.

求你了。

Please.

而且不只是他们搞砸了。

And and it's not only that they crashed it.

而且，如果我没理解错的话，

It's that and correct me if I'm wrong here.

这根本就是个骗局。

It was a scam.

对吧？

Right?

这根本是个骗局，骗所有人相信这两家公司很稳定，结果大家血本无归。

It was a scam that basically got everybody to invest in these two thinking that they were stable and people lost their shirts.

所以我才建议别用这种命名。

That's why I would stay away from that naming.

是的。

Yeah.

其实现在你这么一说，我倒没觉得那套‘地、风、火、太空’之类的说法有多突出。

Don't actually, now that you did explain it to me, it didn't jump out at me, the whole kind of like earth, wind, fire, space type stuff.

我可能只是在把加密货币和那些真正骗人、让很多人血本无归的骗局区分开来。

Maybe I am feeling it a little bit separating out the cryptocurrency association of actual full scale fraudulent scams that had a lot of people lose a lot of money.

也许吧。

Maybe

最好的电影。

best movie.

我倒是有点喜欢，说不上来为什么。

Am liking it a lit I don't know.

可能我就是有点喜欢。

Maybe I'm liking a little bit.

也许我们是想吸引非加密货币用户，找个更让人印象深刻的名称。

Maybe we're trying to appeal to a non cryptocurrency crowd and really try to get to something a little bit more memorable.

我不确定。

I don't know.

我对这些名字的态度，已经从明显负面转为犹豫不决了。

I've just gone to undecided from net negative on these names.

我本人还是反对，不过我觉得这总比被控制强。

Well, I'm still against, but I suppose it's better than harness.

明白了。

Okay.

咱们还是回到主线吧，OpenAI在博客里说，这个模型最初会限制发布。

So let's talk about let's get back to our main story here, which is that here's what OpenAI says in its blog post that the release of this model is going to be restricted at first.

我们支持广泛使用，但先会面向一小群经过审核的合作伙伴提供限量预览，这些合作伙伴之前已经向政府报备过，之后才会全面开放。

We believe in broad access, but we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government before releasing more broadly.

在预览阶段，我们会继续测试，并与合作伙伴紧密协作，为更大范围的发布做准备。

During this preview, we will continue testing and coordinating closely with the partners as we work toward broader availability.

我们不认为这种政府审查流程应该成为长期的常态。

We don't believe this kind of government access process should become the long term default.

显然，他们看到了Mythos和Fable那边发生的事。

So obviously, they saw what happened with Mythos and Fable.

他们不想再重演一次，所以干脆让政府来决定谁能合作、谁不能合作。

They don't want a repeat of that, so they're kind of letting the government handpick the customers that they're able to work with and those they can't.

我们接下来会聊聊这个模型的能力，确实挺厉害的，但我觉得眼下更关键的是它的发布方式。

We're going to get into the model's capabilities because it does seem pretty impressive, but I think the the way that it's being released is more important at the moment.

这背后显然有不少影响，对吧？

Obviously there are implications here, right?

如果企业或国家没法可靠地获得美国基础模型公司提供的前沿AI，他们就会赶紧转向开源方案。

If companies or countries cannot rely on their ability to access frontier AI from US foundational model companies, they will rush toward open source.

他们会寻找其他替代方案。

They will rush towards other solutions.

他们可能会推迟采用这些更昂贵的前沿模型，这样一来，投资回报和研发动力就会减弱。

They may be slower to go to these more expensive frontier models, which means the ROI and the incentive to build them will be less.

所以我们现在到了一个节点——我还不敢说整个商业模式都岌岌可危，但考虑到其中的利害关系和融到的资金，说不定真有这个风险。

And therefore we're now at a point where I won't say the entire business model is in question, but given the stakes and given the money that's being raised, maybe it is.

等等再说。

Well, hold on.

我们先来谈这个。

Let's take it.

我们把问题拆开来看。

Let's separate out.

我觉得这句话里其实包含了好几个不同的部分。

I think there's like a few different parts of that statement.

第一部分，我想说的是关于商业模式这个问题。

The first part is, I guess, like, on the or actually on the business model question.

别忘了，是Anthropic最先提出这些担忧，并主动邀请美国政府介入讨论的。

Let's remember that it was anthropic that raised the flag and the concerns around this and basically invited the US government to bring this up.

现在看起来，这似乎成了美国推出任何前沿模型的默认监管模式，但其实是这些实验室自己推动了这个叙事。

Now it feels like this is becoming the default American rollout, like, regulatory model for any kind of frontier model, but it was the lab itself that pushed the narrative.

OpenAI显然在跟进，但你不觉得不把前沿模型向全体公众开放，未必是件坏事吗？

OpenAI obviously appears that it's following it, but do do you think not releasing a frontier model to, like, a full scale audience is necessarily a bad thing?

你是不是觉得，应该先让最有能力、最负责任的个人和公司来率先测试，并探索它的边界？

Do you think, like, the most qualified responsible people and companies should be the ones that start testing it and pushing its limits first?

是的。

Yeah.

我们在这事儿上已经来回讨论过好多次了。

Well, we've gone back and forth in this.

我过去确实说过，如果你觉得这东西能有这么大的威力，那谨慎发布可能是对的策略。

I mean, I I have in the past said, okay, maybe this is the right strategy to if you think that this thing can be this powerful, then maybe you should be careful with your release.

所以我明白，我之前在这里也表达过这个观点。

So I understand I've said that here.

但另一方面，你也可能陷入一种局面：政府和企业直接决定了谁赢谁输。

But the other side of it is that like you could end up in a situation where the government and the companies really pick the winners and losers in this space.

从商业角度看，他们或许能定出更高的价格，对吧？

And from a business standpoint, maybe they're able to charge a lot more, right?

所以Mythos的初始预览版价格相当高。

So the initial preview of Mythos was quite expensive.

而GPT-5.6的预览版价格，其实只有Fable的一半左右。

And actually the preview of GPT-5.6 is about half the cost of Fable.

这挺有意思的。

So that's interesting.

不过从商业角度看，这里确实存在一些连锁反应。

But yeah, I think that from a business standpoint, there are second order effects here.

更广泛地说，我担心的是，如果我们真进入政府挑选赢家和输家的阶段，可能会走向一个更受管控的经济。

And I think more broadly, there's concern to me that if we're going to go into this moment where there is this government picking winners and losers situation that you could end up with a more controlled economy.

有人在推特上说，美国在人工智能上的管控比现在的中国还要自上而下。

There have been people on Twitter talking about how The US is much more top down on AI than even China is today.

我还看到一条挺有意思的推文。

Here's another tweet that I thought was interesting.

有人写道：‘我担心我们已经进入了一个AI模型开发与获取的黑暗时代。’另一个人转推并评论：‘底层永久阶层的人震惊了，原来上层永久阶层是只限邀请的。’

Someone wrote, I'm afraid we've entered a dark era in AI model development and access and somebody else quote tweeted them and wrote, breaking member of the permanent underclass left shocked after finding out the permanent upper class is invite only.

我觉得这个阶段终会过去，但如果持续下去，确实会留下一些令人担忧的先例。

So I imagine that this moment is going to pass, but certainly there's like some worrying precedence here if it continues.

你怎么看？

Your thoughts?

我觉得‘赢家和输家’这种说法，还有政府来挑选谁赢谁输，其实挺让人在意的，因为这正是本届政府，尤其是第二个任期以来一贯的做法。

Well, I I guess I found it a bit the whole phrase winners and losers and the government picking them because that is how the administration has essentially worked for the last I mean, for especially in this second term.

现在这种做法只是落在了一群长期高喊自由主义的人头上。

And it it's just coming down on, I think, a faction that has kind of, like, screamed for libertarianism for a long time.

但你看，无论是谁有资格收购TikTok，还是其他各种决策，比如Stargate怎么推出，这些事上，政府和行政部门一直都在主导。

But even though, like, when, you know, when it's who gets to buy TikTok or what any other decision or how does Stargate roll out or, you know, like, all of these things the government and the administration have been pushing.

所以在我看来，他们对前沿实验室也是用同样的套路。

So it's just kind of taking the same approach to frontier labs in my mind.

说实话，我也说不清。

And I don't know.

我对事情发展成这样一点都没感到意外。

I I guess I have not been surprised at how things have been playing out at all.

我不觉得这是好事。

I don't think it's a good thing.

但我确实认同，从全球AI格局来看，尤其是跟中国对比，整体趋势真的很不妙。

I I I do agree that overall, like, especially, like, vis a vis China and how things are playing out overall across the entire global AI landscape, I think it's a very bad thing.

但我其实没像很多风投圈的人那么惊讶。

But I guess I'm not as surprised as a lot of VC Twitter seems to be.

是的。

Yeah.

我倒不觉得用‘惊讶’来形容我的反应合适，但这件事确实很重要。

I wouldn't say surprised is the reaction I'm having, but you know it's important.

我觉得有必要讨论一下，将来可能引发的问题——比如，如果这就是新的规则，那当你没法用这些模型、而对手可以用的时候，你就处于明显劣势了。

I think it's important to talk about the the problems that could lead down the line both in terms of like, alright, if this is the new process, then certainly you're at a much greater disadvantage if you aren't able to use these models and your competitors are.

对吧？

Right?

那谁来决定这些呢？

So who gets to decide that?

对吧？

Right?

我觉得这不符合美国该有的自由市场原则。

I don't think that's the free market that The US needs.

另一方面，你是不是反而帮了竞争对手的忙？

Then the other side of it is do you end up playing into the hands of your competitors?

这是艾伦·列维的一个有趣观点，我们之前稍微提过，但值得再深入聊聊。

This is an interesting thought from Aaron Levy and we've kind of touched on this, but it's worth talking about.

如果美国始终站在技术前沿，同时对前沿智能的发布实施严格监管，我们就能在经济和地缘政治上占据优势，因为我们能控制谁有资格接触这些前沿技术。

If The US remains at the frontier at all times and has heavy regulation on the release of intelligence, then we end up with an economic and geopolitical edge because we can control who has access to frontier intelligence.

但如果我们推迟模型发布，而另一个玩家——特别是中国——并不放慢脚步，很快就能推出同样强大的模型，那我们的延迟反而会帮他们提升模型和整个技术栈的竞争力。

However, if we delay model releases and another player, specifically China, doesn't slow down and has equally strong models, not now but soon, then our delays end up advantaging their models and eventually their tech stack.

最近我们确实听到不少讨论，说人们正在从直接使用前沿模型，转向使用模型路由和更多开源模型。

And we've definitely seen conversations recently about how, you know, we're moving from people using a frontier model to using model routing, using more open source models.

我觉得这正好印证了你的观点。

And and I think this kind of goes to your point.

对吧？

Right?

如果你觉得现在真正的竞争优势是系统整合能力，而不是底层模型本身，那这种政策可能会把前沿实验室挤到边缘，让中国开源社区独占AI浪潮的大部分红利——谁知道呢？

Like, if if you think the the harness is the big competitive advantage now and not the underlying model, then you could end up with this type of policy sort of icing out the Frontier Labs and having a lot of the benefits of this AI moment go to open source in China, which who knows?

也就是说，如果开源AI资源很丰富，那在很多方面其实都挺有利的。

Mean, if there's a robust open source AI availability, then that could be really good in a lot of ways.

但这样的话，你在安全方面的管控能力也就跟着削弱了。

But you also then sort of lose your ability to control on the safety side of things.

所以这周才有人提议禁止开源模型，不过我觉得那样做也不太好。

And that's why you have people talking this week about banning open source models, which I don't think would be good either.

好，那我们来深入聊聊这一点。

Well, I think let's dig into that a little bit.

对我来说，这是两回事。尤其是现在听到的很多故事，或者大家在做的工作，甚至是我在一线客户中看到的，都在探索开源模型。

Like to me, they are two separate issues because especially a lot of the stories you're hearing about or work that's being done or even what I've been seeing kind of like with customers on the front lines around exploring open source models.

这些更多是关于成本，以及哪些模型适合哪些任务，而不是说大家真的会把很多现在路由到开源的工作交给像Mythos或Fable这样的模型。

A lot of that is more around cost and like what models are actually suited to what tasks versus the idea that anyone is gonna use Mythos or Fable for a lot of the work that's being routed now towards open source.

所以我认为，成本驱动因素（我们稍后会详细讨论）比监管方面的影响更大。

So I think, like, the cost driver, which we'll talk more about, is more of a factor rather than the regulatory side.

不过我也觉得，毕竟我们要记住，最顶尖的前沿模型，其价值主要体现在发现大量安全漏洞，或者推动科学创新方面。

But I do think, like, the because we we have to remember, like, the most state of the art cutting edge frontier model, that's really about whether it's, you know, finding tons of security loopholes or, like, pushing the boundaries of science, scientific innovation.

我觉得，那种工作和我们现在讨论的开源模型根本不是一回事。

Like, I don't know, like, that that that kind of work is still to me very different than what we're hearing about on the open source side.

没错。

Right.

那是因为开源模型在能力上还落后一些。

But I mean, that's because the open source models have trailed in capabilities.

但不管有没有监管，这种情况迟早都会发生，对吧？

But it could end up, I mean, this might have happened whether we see the regulation or not, right?

开源模型不仅会更便宜，还可能逐渐赶上主流水平。

That the open source models will not only be cheaper, but they can start to rival.

已经在发生了。

It's happening.

它们的能力正在逐步逼近。

They can start to rival in capabilities.

尤其是现在这些公司都准备上市了，你不得不想：如果你有一个闭源模型，性能相当于十五个博士，情商高，能帮你高效完成任务，再对比一个开源模型，那会怎么样？

That really, I mean especially now as these companies head towards IPO, it really makes you wonder and I did try to speak with Brockman about this a little bit when he was at the AI summit, but like makes you wonder if you have a model that's closed and is the equivalent of like 15 PhDs and has good EQ and can do stuff get stuff done for you and you have an open source model, right?

假设几年后，这种模型相当于十二个博士的水平，情商也不错，照样能帮你把事办了。

Let's say a couple years down the line that's like of 12 PhDs and has decent EQ and will still get things done from you.

那如果所有模型都变得这么聪明了，前沿模型相比开源模型还有什么竞争优势呢？

Then like what is the competitive advantage for the frontier models versus the open source if they all get that smart?

嗯，这么说吧。

Well, okay.

关于商业模式和IPO的问题，我们待会儿还会更深入聊OpenAI的财报和可能的上市，但我觉得这事儿影响太大了。

So on the business model and IPO question, and we're we're gonna talk more about OpenAI's, the reporting and on their potential IPO, but I think it's a massive deal.

这简直是个大到离谱的转折点。

I think it's like a absolutely ridiculously massive deal.

它彻底颠覆了整个叙事逻辑。

The fact that it completely destroys the entire story.

以前的说法一直是：前沿模型造价高昂。

The story has been, like, expensive frontier models.

从大约去年九月、十月到今年二月、三月，所有公布的收入增长，其实都来自人们不断使用最新、最贵的前沿模型。

All the ARR growth that was shown to the world from, call it, September ish, October ish to February, March was people just cranking on whatever the latest most expensive frontier model.

这一刻，虽然我们能一直争论这个悖论，但我坚信，模型和智能体AI的整体使用率将会呈指数级增长。

The moment, and again, we'll we can paradox it all day long and and I do firmly believe overall utilization of models and agentic AI is gonna be just, you know, like exponential.

但核心的IPO故事，我认为因此受到了巨大冲击。

But the core IPO story, I think, takes a huge hit from this.

所以，当人们意识到根本不需要法拉利的时候，哪怕只是短暂拥有更强大的技术，这种优势也会逐渐消失。

So I think, like, the the advantage of having something far superior for a short amount of time starts to go away anyways the moment people start realizing they don't need the Ferrari.

本田雅阁已经足够好了。

The Honda Accord is is great.

没错。

Right.

直到现在，情况还像以前买IBM那样，谁都不会因为选了它而被责怪，因为你根本没必要换别的。

And you know, up until this point, was like a nobody gets fired for buying IBM situation where like you wouldn't look elsewhere because you didn't need to.

毕竟，从这些封闭公司直接接入前沿模型，操作起来就是更省事。

Because it was just kind of easier to plug and play these frontier models from the closed companies.

但我说的是，现在前沿技术被限制了，反而让人更有动力去寻找其他替代方案。

But what I'm saying is because the frontier is now being restricted, it adds an incentive to look elsewhere.

当这种动力促使人们去寻找其他选择时，你可能会看到增长放缓。

And when it adds that incentive to look elsewhere, you might see a slowing.

不过，基础模型公司会说，他们的模型依然更优秀——确实如此，而且算力也更强——这一点也没错。

Now, the foundational model companies would argue that their models are still better, which they are, that they have more compute, which they do.

但问题在于，人们对他们的期望已经被抬得太高，投入的资金也太过庞大，以至于他们必须一击即中，甚至可能只需要两杆进洞，我也不确定。

But ultimately, the expectations have been built up so high and the money is so significant in that they basically need to shoot a hole in one here for everything to work or maybe a hole in two, I don't know.

他们必须近乎完美地执行，压力现在全在他们身上，这也就是为什么我觉得OpenAI会冒险表态：我们现在愿意配合政府的要求，但其实我们并不想这么做。

They have to basically be near flawless in their execution and the pressure is now on and that's why I think you saw OpenAI taking the risky stance of saying we're going to play with the way the government wants us to play right now, but we would really prefer not.

你说得对。

And you're right.

他们确实执行得近乎完美。

Like they have executed dearly flawlessly.

过去六个月到一年，几乎一直都是这样。

I mean, feels like for the last six to twelve months.

正因如此，大家的期望才被推到了这么高的地步。

I mean, that's why the expectations have been built so massively.

但说实话，我也说不准。

But, yeah, I don't know.

总的来说，我觉得这是个大问题。

Overall, I think this is a big issue.

不过我一直在想，现在大家总在说，一旦这些前沿模型真正大规模应用，模型蒸馏就会开始。

I guess one thing I have been wondering though is right now you keep hearing and reading about as these frontier models actually, know, like have widespread adoption, that's where the distilling starts to take place.

那时候竞争对手就能开始动手了——我听说有些团队搞什么‘蒸馏集群’， literally 发送数百万条提示，就为了搞清楚模型的全部行为，然后从开源角度试着把它重新做出来。

That's where competitors can start to I mean, like, you know, I've heard about these, like, distillation swarms where they're literally, like, asking you know, like, sending millions of prompts to try to, like, start understanding all of the model behavior and starting to work to actually recreate it from a open source standpoint.

有没有可能，监管政策能起到保护作用，只把模型提供给最赚钱的那一百家大型企业和政府客户，他们付的钱最多，同时也能降低被蒸馏的风险？

Is there some world where the regulation is good to kinda protect the frontier model and you only give it to your 100 most profitable large enterprise and government customers that pay you the most money and also minimize risk of distillation as well.

他们会不会反而觉得这样挺好？

Could that could they be almost a little happy that that's the case?

我的看法是，Anthropic 在 Fable 上加了这么多限制，与其说是担心会被滥用，不如说——

My hot take is that one of the things that we saw from Anthropic with all these safeguards that it put on Fable was not as much that they were worried that Fable would be misused, although I'm sure there was some worry there.

他们更担心的是竞争对手会去蒸馏这个模型，所以干脆不让他们在最有价值的地方得手。

It was more that like we know our competitors are going try to distill this model and we're just not going to let them do it in places that provide the most value.

事实上，我认为Anthropic是在明确表示：如果我们怀疑你用我们的技术来开发竞品模型，就会切断你的访问权限。

And in fact, I think Anthropic was effectively coming out and saying that if we think that you're using our technology to build a competing model, we're going to block your access to do that.

但他们后来稍微退了一步，因为大家说，Anthropic这样阻止别人用模型做AI研究，简直是反竞争、不妥当。

And they kind of backed off that a little bit because people were like Anthropic doing any preventing any AI research with its models is anti competitive and bad.

不过说实话，这确实是个真实的担忧，这种限量发布策略能在一定程度上阻止模型被蒸馏，但看起来，想完全保密这些模型实在太难了。

But yeah, think that was a real worry and this limited release strategy could prevent distillation to a degree although the models it seems like it's kind of hard to keep the models under wraps.

你觉得呢？

Don't you think?

我不确定。

I don't know.

这周我最喜欢的一条推文是比尔·盖利说的：如果你快实现AGI或ASI了，为什么你的模型还不够聪明，没法实时识别出窃取式蒸馏行为？

My favorite, Bill Gurley this week had tweeted, if you're on the verge of AGI or ASI, why isn't your model smart enough to recognize espionage distillation in real time?

你说你能在几年内治愈癌症，那察觉非法蒸馏难道不比治愈癌症简单多了吗？

You say you can cure cancer in a few years, isn't sniffing illicit distillation quite a bit easier than curing cancer?

我特别喜欢这个观点，因为它点出了一个矛盾：既然这种蒸馏威胁到他们的生死存亡，为什么他们还没投入足够资源去解决，反而只能靠开放更多访问来应对？

And I I I kind of loved it because it's like, why do you need why it is surprising to me that that is still such a threat when it is existential to their business that they haven't kind of invested invested or figured out a way to protect against this short of giving as many people access.

是的。

Yeah.

不是。

No.

这话说得很有道理。

It's a great point.

我的意思是，他们当初搞这些主动防护措施，大概就是想避免这种情况，但显然问题还是存在。

I mean, think that's what they were trying to do with these proactive safeguards, but obviously still a problem.

他们刚指责了阿里巴巴。

They just accused Alibaba.

Anthropic最近刚指控阿里巴巴在窃取他们的模型。

Anthropic just accused Alibaba of distilling their models recently.

明白了。

Okay.

再说一件关于GPT 5.6的事，之后我们再聊OpenAI的上市消息。

One thing on GPT 5.6, then we'll move on to the OpenAI IPO stories.

他们让 GPT 5.6 通过了一个叫 Exploit Bench 的测试，表现和 Mythos 基本持平。

So they ran GPT 5.6 through a test called Exploit Bench, and the capabilities were basically on par with Mythos.

这显然是个网络安全测试，但 Mythos 的令牌效率低得多。

This is obviously a cybersecurity test, but Mythos is much less token efficient.

Mythos 做这个测试用了大约三十万个令牌，而 GPT-5.6 只用了十万左右。

So Mythos used something like 300,000 tokens for the the test and GPT-5.6 used around 100,000.

所以我觉得它的效率至少是 Mythos Preview 的两倍，这挺有意思的，而且价格还不到它的一半。

So I would say it's at least two times more efficient than Mythos Preview, which is interesting and it's priced more cheaply than it, about half price.

是的。

Yeah.

你看，价格现在成了主要卖点，不再是顺带提一提的事了。

I mean, you see price constantly being now a highlight rather than kinda some afterthought.

过去大多数人根本说不清输入输出令牌具体花多少钱，现在人人都在关注这个了。

I don't think most people could have, you know, like, reasonably known what their input and output token costs were in the past, and now everyone is looking I mean, is looking at it.

等等，我还没说完。

Well, I've wait.

等一下。

Hold on.

现在完全没法公开使用。

It's not publicly available in any capacity now.

对吧？

Right?

5.6版也不行？

None of 5.6?

是的。

Yeah.

它确实没对外公开。

It it's not publicly available.

情况都一样。

It's the same thing.

只有少数几家公司能

There's a limited number of companies that that

有三个都能访问吗？

have access all three them?

三个都像

All three like

还有Terra吗？

and Terra?

三个都一样，不像Luna那样人人都能稍微用一下，现在还不行。

All three of So they're not doing like a Luna everybody gets a little bit of access to, Not yet.

得等着。

Gotta wait.

我的意思是，大家希望它能很快推出。

I mean, I think the hope is that it comes out fairly soon.

他们现在就是这么说的，但还得看实际情况。

That's just the way that they're talking, but we'll see.

我问这个是因为，在我看来，也许想法不对，我还是觉得模型越贵、越庞大，就越危险。

Well, I guess why why I'm asking that is in my mind, and maybe it's wrong, I still kind of associate the more expensive and heavy the model is, the more dangerous it is.

所以当Luna又快又便宜，适合日常使用，定位就这么定的时候，你可能会觉得，这应该很简单，直接给大家开放就行；而Soul才是旗舰模型，才是那个可能让我们全出问题的危险家伙。

And that's why when Luna is fast and affordable for everyday work and that's kind of the positioning, you would think, like, that should be easy and you just give it to everybody versus Soul is the flagship model and, that's the big dangerous one that can get us all into trouble.

但也许这个想法是错的。

But maybe that's the wrong assumption.

是的。

Yeah.

你得小心点。

You gotta be careful.

那些灵活的对手。

Those nimble fighters.

对吧？

Right?

你以为它们只是轻巧又随和，结果一不留神，就给你来一刀。

You think they're just lightweight and chill and next thing you know, cut you.

它们到处利用软件漏洞，正在拖垮全球金融系统。

They're they're they're exploiting software left and right, bringing down global financial infrastructure.

没错。

Exactly.

好的。

Okay.

说到全球金融基础设施，我们这边还有几家公司准备上市。

So speaking of global financial infrastructure, we do have IPOs on the way.

但问题是，到底什么时候呢？

This is from the but but the question is when?

这是《纽约时报》的报道。

This is from the New York Times.

OpenAI倾向于推迟到明年再进行IPO。

OpenAI leans toward waiting until next year for its IPO.

OpenAI正考虑将首次公开募股推迟到明年，这一转变凸显了快速崛起的人工智能巨头们前景的不确定性。

OpenAI is leaning towards holding off its initial public offering until next year, a turnabout that punctuates the uncertain future for fast rising artificial intelligence giants.

ChatGPT的开发公司已经聘请了投行和律师，原本计划在今年第三季度或第四季度就启动上市。

The maker of ChatGPT hired bankers and lawyers with an eye towards a public offering as soon as the third or fourth quarter of this year.

但因为这达不到萨姆·阿尔特曼想要的那1万亿美元估值，他们很可能要推迟上市计划。

But because that's not going to hit the $1,000,000,000,000 valuation that Sam Altman has been seeking, it looks like they are going to potentially postpone.

这是《时代》周刊的报道。

It's from the time story.

现在大家最关心的是，这个月埃隆·马斯克的SpaceX上市后发生了什么。

Top of mind is what happened to Elon Musk's SpaceX after its IPO this month.

它创下了史上最大规模的IPO，融资超过850亿美元，估值高达1.77万亿美元。

It had the largest ever largest IPO ever raising more than $85,000,000,000 at a $1,770,000,000,000 valuation.

但之后股价就大幅回落了。

However, it's fallen back to earth.

我先说说我的一个猜测，然后听听你怎么看这件事背后的影响。

I'll just share my conspiracy theory on this one and then turn it to you in terms of your views of the implications here.

我认为OpenAI很清楚，鉴于最近的迅猛增长，它对公开市场来说至关重要。

I think OpenAI knows it's important to the public markets given the growth that has shown of late.

它可能觉得现在是个更好的时机，毕竟之前因为政府的顾虑被拖住了。OpenAI或许只是放出风声说要等一等，但其实随时准备像轻量级拳手那样突然出击，抢在Anthropic前面拿下市场。

It probably sees a better opening now because has been in a way held back by the government because of the fable fumble and OpenAI might be just sending out signals that it's going to wait, but then it will pounce as a lightweight fighter does and cut anthropic and get there first.

这就是我的猜测。

That's my conspiracy theory.

把两家可能值万亿美元的公司比作轻量级拳手，不过……

Talking about two potentially trillion dollar companies as lightweight fighters here, but

得灵活才行。

Gotta be nimble.

确实得灵活。

Gotta be nimble.

明白了。

Okay.

我喜欢这个说法，因为居然能泄露出来，这让我挺意外的。

I I I do like that because it was surprising to me that that could even be leaked.

我觉得很多关于IPO的报道都是故意放出来的，毕竟以前我近距离看过太多这类事了，大家都会想尽办法阻止这种消息外泄，但这种情况我们见得还少吗？

And I feel a lot of this IPO reporting is very purposefully leaked because again, in the past, like, you mean having seen a lot of this very up close, like, everyone does whatever they can to not allow for any of this kind of reporting, and we've certainly seen plenty of it.

所以现在每一条消息出来，都感觉是经过精心设计的。

So it feels very like, anything that comes out is purposeful.

那问题来了，他们为什么要放出可能要推迟的信号呢？

So then, yeah, the question, why would they signal that they may be delaying?

我很喜欢你这个看法。

I do like your your take there.

我觉得这个说法挺有意思的。

I think that's a fun one.

但我觉得现在市场上真正有趣的是，两个月前，一切都还很有方向感。

But I also just think like, right now, what's to me what's interesting with the market is two months ago, everything felt like it had direction.

当时所有的营收数据都惊人得离谱，随便哪个数字都是如此。

Everything the ARR was just unreal like the any kind of number that was being reported out.

总的来说，大家对现状都挺乐观的。

Overall, you know, everyone's feeling good about things.

前沿实验室和最先进模型依然是主流叙事。

The frontier lab and state of the art models still are the kind of dominant narrative.

但与此同时，最大的情绪转变在于，当你看到比尔·格利或者Coinbase的CEO布莱恩·阿姆斯特朗，大家都在炫耀自己根本没用前沿模型。

Meanwhile, to meet the biggest vibe shift has been when you see Bill Gurley or the Coinbase CEO, all to Brian Armstrong, everyone is just bragging about not using the frontier models.

想想两个月前、三个月前，大家动不动就问，你用过Opus 48了吗？或者别的什么，那时候GPT-55已经全面超越48了。

Think about two months ago, three months ago, it was just, oh my god, have you used Opus four eight or whatever, like it was just GPT-five five is now ahead of four eight.

那时候人人都在聊这个。

That's all anyone talked about.

普通人压根就没听说过Quinn，更别说这周刚出的GLM 5.2了。

No one had ever even I mean most the average user had never even heard of Quinn or like now with GLM 5.2 this week.

对我来说，现在简直是个该按兵不动的时刻，因为市场从两三个月前就彻底变了，除了马斯克，没人及时抽身，而他确实做得特别好。

So to me, you it almost feels like it's this is a moment to wait because the entire market has shifted from two to three months ago and and no one got out other than Elon and he he did a great job at that.

所以我觉得，现在选择等等，反而是更明智的做法。

So but I I think waiting actually would be the smarter decision.

那你打算等多久？

But how long do you wanna wait?

要是等太久，你可能已经花了不少钱，还在等更多资金进来，可客户已经开始想办法提升效率了。

Because if you wait too much longer, you might be in the situation where you've spent a lot of money, you're waiting for more money, and your customers are starting to try to find ways to efficiency max.

但他们确实已经在这么做了。

But they are.

他们早就这么做了。

They already are.

他们早就这么做了。

They already are.

但你现在希望怎样？目前趋势是向上攀升，可你希望当客户开始寻找更高效使用你工具的方式时，这个趋势能趋于平缓吗？

But do you want that you know, right now the curve is looking up into the right, but do you want it to even out as these customers try to find more efficient ways to use your tools?

说实话，我对这方面其实也没有确切的信息。

Guess, I mean, and I have no direct knowledge either way on this.

对我来说，现在的趋势已经不像几个月前那样持续上扬了。

To me, the curve is no longer up into the right from where it was a few months ago.

所有方面都是如此。

Everything yeah.

我的意思是，当Coinbase的CEO在炫耀模型路由，说他们用了五个不同模型分别处理不同任务，人人都在讨论这件事的时候，整个增长曲线就已经变了。

I mean, just again, when the Coinbase CEO is bragging about model routing and here's five different models suited to different tasks and everyone is talking about that, it changes the entire the growth curve.

从九月到二月、三月、四月左右的整个故事，全都变了。

The the entire story from September to February, March, April ish, all of that has changed.

这事儿太明显了。

Like, it's just so clear.

所以你要出去宣传，就得拿出最新的数据。

So that the to go out, you're gonna have to show your latest results.

你得展示最近几个月的表现。

You're gonna have to show the last few months.

我觉得，既然你有这么多有潜力的业务线，不如先让它们再成熟一点，再讲一个更扎实的故事，而不是拿一个看起来夸张的ARR曲线出来。

So I do think versus you have a lot of promising different business lines, start allowing a little bit more maturity, and then go out with a stronger story rather than here's an ARR curve that looks insane.

没错。

Right.

这不光是Coinbase的CEO，也不光是社交媒体上有人在讨论怎么更高效地使用模型，或者在Anthropic和OpenAI身上找节省成本的方法。

And it's not just the Coinbase Coinbase CEO and it's not just, you know, Twitter voices who are talking about finding more efficient uses of their models and looking for savings in their engagements with Anthropic and OpenAI.

另外，还有报道说Anthropic和OpenAI可能在向客户多收费。

And by the way, there's also some reporting about Anthropic and OpenAI potentially overcharging customers.

所以这是一场更广泛的调整，不只是个别网红在喊，我们有报道支撑，等会儿就来聊这个。

So it is a broader pullback, not just a micro influencer pullback and we have the reporting and we're going to talk about it right after this.

我们继续回到《大科技》播客，今天邀请的是Margins的Ranjan Roy。

And we're back here on Big Technology Podcast with Ranjan Roy of Margins.

好的，Ranjan，刚才我们聊到，现在网上流行一种说法，说大家都在想办法减少在OpenAI和Anthropic上的开支，但实际报道显示，这不只是个网络段子，而是一个真实且可能比社交媒体上看到的还要普遍的现象。

Alright Ranjan, so before the break we talked a little bit about how there's this now it's it's become a meme that people are trying to pull back on their costs with OpenAI and Anthropic and finding ways to do it, but the reporting shows that it's more than a meme that it's a reality that seems like might be more widespread than you would even know by seeing the tweets.

以下是相关信息：AI客户如何降低在Anthropic和OpenAI上的支出。

This is from the information: How AI customers are lowering their Anthropic and OpenAI bills.

随着Anthropic和OpenAI核心产品的价格不断上涨，一些大客户开始转向这两家公司提供的更便宜的AI模型，甚至改用其他供应商的方案。

As prices for Anthropic and OpenAI's flagship products soar, some large customers are using cheaper AI models from those companies' menus as well as from other providers.

Ensemble Health Partners是一家为医院提供软件服务的公司，今年计划在AI上投入高达一亿美元，他们表示，改用OpenAI一个价格仅为高端模型二十分之一的模型后，效果非常好。

Ensemble Health Partners, a provider of software for hospitals that plans to spend up to a 100,000,000 on AI this year, said it's had success switching to an OpenAI model that's onetwentieth as expensive as the company's more advanced models.

这个AI工具用于撰写给保险公司提出的申诉信，因为这些保险公司拒绝为Ensemble的医院客户支付诊疗费用。

The AI powered tool for writing appeal letters to insurance firms that won't reimburse Ensemble's hospital clients for care they provide.

也就是说，AI驱动的这个工具专门处理这类申诉信。

So the AI powers a tool that does that.

Ensemble每个月要发出大约一万五千封这样的信，公司首席技术官表示，改用成本更低的模型后，每年能节省近70万美元。

Ensemble sends about 15,000 such letters every month, and the chief technology officer of the company said switching to a lower cost model will result in savings of nearly $700,000 per year.

很多使用Anthropic和OpenAI服务的客户虽然能接受价格上涨，但有些公司一超出预算就立刻停手了。

Many Anthropic and OpenAI customers are willing to stomach the rising costs of these two companies' products, but some are hitting the brakes after blowing through their budgets.

明白了。

Okay.

现在的情况是，那些轻量级模型已经足够好了，企业开始重新审视自己的开支，心想：等等，这完全说不通啊。

So basically what you're seeing now is it's almost as if these smaller flash models have gotten good enough that you can make these moves and companies are looking at what they're spending and saying, hey, wait a second, this makes zero sense.

我们干脆改用轻量模型，把任务分配出去，而把更强大的模型留作顶层决策，而不是直接干活。

We are actually just going to go to the smaller models and we're going to route and maybe use the more powerful models as a decision maker as a sort of top layer decision maker as opposed to a worker.

这确实挺有意思的，Ranjan，你觉得呢？

Pretty interesting Ranjan, wouldn't you say?

这其实跟我们上半年聊的是一回事——对这些公司来说，真正的经济压力正在形成，可不是什么好事。

I mean it does seem like this is going back to our discussion in the first half year, a real, not to use jargon, but a real economic headwind for these companies and not good.

如果你还指望它们在上市前展现出惊人的增长势头，那这可不是个好兆头。

Again, if you're rooting for them to show exceptional growth as they get to IPO.

你刚才说这是经济压力，但有意思的是，即使在这种情况下，他们还是在用OpenAI的模型，只是换成了价格只有原来的二十分之一的那个。

I mean, to me the most interesting part of what you just said is, you called it an economic headwind, but even within that reporting, they're still talking about using an OpenAI model, just one twentieth is expensive.

所以在我看来，这给整个叙事带来了更大的挑战：即使客户还在用 Anthropic 和 OpenAI 的产品，这对它们的上市故事和商业前景依然不利，因为它们的商业模式依赖于使用最昂贵的模型。

So to me that's an even bigger challenge overall to the narrative is is like you can be using Anthropic and OpenAI's own products and still that's not good for their overall, like, IPO narrative and the the business story because the it's dependent on using the most expensive models.

并不是说你们有一整套产品和模型，可以尽情使用，这对大家都有好处。

Not here's a whole suite of products and models and use all of them the best you can and that's good for all of us.

现在的情况更像是，这件事本身成了一个故事，而不是简单地在 OpenAI 生态里选用不同的工具。

It really feels like the fact that that becomes the a story in itself versus, oh, you're just using different parts of the OpenAI ecosystem.

我觉得这其实挺能说明问题的。

I think that is pretty telling thing.

这一切有没有可能涉及杰文斯悖论？比如，你们从只用我们最前沿的模型，转成用一些便宜点的模型，但反而会因此用得更多。

Is there a Jevons Paradox side to all this where it's like, yeah, like you switched from using only our frontier model for everything to some of our cheaper models, but you're going to do a lot more now.

于是，原本在 2023 年根本不会花一分钱的领域，现在却要投入大量资金。

And so there therefore, a category that you never would have even spent on in 2023, now you're going to spend a ton more.

我觉得现在这个时间点特别微妙。

I think like this is the crazy part of the timing of all this.

我坚信，再过几年，尤其是未来十年，整体的 token 消耗量会大幅增长，所有东西都会被智能化代理化。

To me, I firmly believe in a few years time and especially in like the next decade, there is going to be just massive increases in overall token consumption, agentification of everything.