本周AI五分钟：神界混乱特辑

今天在《AI每日简报》，为您带来给忙碌人士的本周AI要闻。

Today on The AI Daily Brief, This Week in AI for terrifyingly busy people.

《AI每日简报》是一档每天更新的播客和视频节目，聚焦AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

在进入本期内容之前，先说一句提醒。

A quick note before we get into this episode.

周五晚上，当大多数人正陪家人放松，或者观看世界杯美国队首场比赛时，AI领域的权力格局发生了根本性的变化。

On Friday evening, as many of us were settling in with our families or watching the first US game at the World Cup, the tides of power and AI shifted in a fairly fundamental way.

根据当时一份匿名消息源的报道——后来证实是亚马逊——美国政府要求Anthropic公司切断所有外国用户对Fable 5和Mythos 5的访问权限。

Based on a report from at the time an unnamed party, which later turned out to be Amazon, the US government directed Anthropic to cut off access to all foreign nationals for Fable 5 and Mythos 5.

为确保严格执行这一要求，Anthropic直接对所有人关闭了这两款模型。

In order to ensure compliance with the request, Anthropic took the models down entirely for everyone.

这件事影响深远，尤其是本集结尾推荐大家去试试Fable 5的建议，现在根本行不通。

There are many implications of this, not least of which is that the recommendations at the end of this five minute recap episode to go out and try Fable five don't exactly work at the moment.

我将在今天周六，6月13日，专门录制一期紧急特别节目，稍后发布。

Now I will be recording a full emergency episode to go out later today on Saturday, June 13.

不过，我还是想做这个五分钟的简报，因为我希望让这个节目对更多人来说更容易接触。

However, I did still want to have this five minute recap as I attempt to continue to make this show more accessible for an even wider audience.

所以，先享受这份对AI领域这关键一周的部分回顾吧，我稍后会推出完整报道，详解Anthropic对Fable 5的关停事件。

So enjoy this partial recap of what ended up being an extraordinarily consequential week in AI, and I'll be back soon with the full news about the anthropic Fable five shutdown.

好了，各位。

Alright, friends.

欢迎回到《AI每日简报》。

Welcome back to The AI Daily Brief.

今天，我们继续尝试这种超快节奏的每周五分钟回顾。

Today, we continue our experiment in these hyper fast five minute weekly recaps.

上周收到了很多积极反馈，如果你还没听过，这些节目的设计有两个目的。

I had a lot of positive response last week, and for those of you who missed that, the idea of these episodes is two part.

第一，很多听众的同事、朋友或家人没时间每周听一大堆二十五到三十分钟的长节目，但他们想大致了解AI领域最重要的动态，这就是你可以转发给他们的那一集。

First, a lot of you have colleagues, friends, family who don't have time to listen to a whole slew of twenty five or thirty minute episodes in a week, but who do want to know, broadly speaking, the most important stuff that happened in AI, this is the episode that you can send them.

第二，就连你们这些铁杆听众，有时候也会忙得不可开交，比如真正在上班，或者在地中海航行什么的。

Second, even you diehard listeners sometimes have weeks where you're just busy doing, I don't know, real work or sailing around the Mediterranean or something.

这些简报对你来说也同样适用。

Consider these episodes for you as well.

今天的第一条重磅新闻，同时也是本周最重大的发布——当然是Claude的Fable Five。

Now we kick off today with the biggest story of the week, which happens to also be the biggest release of the week, which is, of course, Claude's Fable five.

是的。

Yes.

我们终于迎来了神话级模型。

We finally have a mythos class model.

过去几个月，自从神话模型的消息传出并被Anthropic确认后，我们普通用户一直等着，看着各大机构通过Project Glasswing抢先体验，满心期待着有朝一日我们也能拥有神话级的能力。

For the last couple of months ever since the news of mythos broke and was then confirmed by anthropic, Us and the pleb class have just been waiting as big institutions get to check it out through Project Glasswing, eagerly anticipating the day when we too would have the powers of Mythos.

这周，我们终于等到了，而且这能力确实非同凡响。

Well, this week we got them, and powers they were indeed.

所有基准测试的结果都明显惊人，但它真的更好用吗？

All of the benchmarks were obviously massive, but was it actually better?

抛开争议不谈——我们稍后会聊这个——总的来说，我看到的那些深入体验过Fable的人，绝大多数都认为它极其强大，相比之前的产品是质的飞跃。

Well, aside from the controversy, which we will get to in just a minute, on average, I would say, the vast majority of people that I saw that really dug into Fable found it to be incredibly powerful and a distinct jump up from what was previously available.

事实上，网上流传着一种巧妙的说法，说AI已经先进到除非你拿它做特别复杂的事，否则根本感觉不出差别。

In fact, there was this clever little narrative running around that we'd reached the point where AI was so advanced that unless you were using it for really powerful things, you wouldn't necessarily know the difference.

Citrini Research发推说，我觉得普通人已经很难分辨新模型是不是比旧的更好了。

Citrini Research tweeted, I think we've reached the point where normal people can't really determine whether new models are better than previous ones.

对我来说，Fable也没觉得强多少，但我认识的那些智商超过150的人，个个都说：天啊，奇点来得比我想象的还早。

Like, Fable doesn't seem that much better to me, but every 150 IQ person I know is like, wow, the singularity came sooner than I thought.

我觉得这话有一定道理，因为Fable真正厉害的地方确实在于解决超级难的问题，但我也觉得这说法不太对——依我经验，Fable五在基础任务上也强出一大截。

Now I think this has a hint of truth to it in that where Fable uniquely shines is super hard things, but I also think that this is kind of wrong too because in my experience, Fable five is way, way better for the basic things as well.

在我看来，它是个更出色的策略思考者，也更擅长从第一性原理出发论证，不会轻易顺从或被提问者牵着鼻子走。

It is, in my opinion, a much better strategic thinker, and it is also a much better first principles arguer in that it doesn't kowtow and isn't as easily manipulable by the person prompting it.

这让它成了一个更有价值的思维伙伴，所以不管你是不是在写复杂的网页应用，只要你在做策略相关的事，都会从Fable五身上获益良多。

That makes it a much more valuable thought partner, so if you are doing anything with strategy, even if you are not coding big, unwieldy web apps, I think you will find a lot of benefit to Fable five.

简而言之，如果你还没试过，我强烈建议你尽快去体验一下Fable五。

TLDR, I definitely think that if you haven't yet, you should go try Fable five soon.

我说的‘尽快’，真的是马上——因为你只有很短的时间可以使用。

In fact, by soon, mean really soon because you only have a limited time to do so.

因为Fable Five不仅是本周最重磅的发布，也是争议最大的话题。

That's because Fable five was not only the biggest story and the biggest release of the week, but also the biggest controversy of the week.

我刚才提到的是，Fable Five目前只在普通套餐中提供，直到6月22日为止。

Now the part that I just alluded to is that Fable five is only available in normal plans until June 22.

之后你就得通过API按使用量付费了。

After that, you're gonna have to pay for it through the API on a usage basis.

这延续了我们一直看到的趋势——公司越来越倾向按使用量收费，而不是按用户数收费。

This certainly continues the trend that we've seen of companies moving to a usage based model instead of a seat based model.

不过，这种影响被削弱了，因为大家其实早有心理准备。

And so some of the impact of that was dampened by the fact that it wasn't all that unexpected.

但说到争议，那点小事根本算不上什么，真正让人震惊的是它设置了极其严格的限制。

And yet when it comes to the controversy, that was small potatoes compared to, first, the fact that there were incredibly stringent guardrails on the thing.

生物医学工程师达莉亚·阿努普·马兹说：‘我甚至没法跟Fable打招呼，除非开无痕模式，因为它知道我是生物医学研究者。’

Biomedical engineer Daria Anoop Maz wrote, I can't even say hello to Fable except in incognito mode because it knows I'm a biomedical researcher.

Fable Five上线时还附带了数据保留政策，说会把对话记录保存三十天，这直接让所有企业都坐不住了：‘这我们可接受不了。’

Fable five also launched with data retention policies that said that they could keep conversations for thirty days in ways that basically made every enterprise sit up and say, I don't think so.

早在那之前，微软就已经限制员工使用Fable Five，因为他们还在评估这些数据保留政策对自己意味着什么。

Already, we saw Microsoft limit employee use of Fable five as they figure out what the data retention policies mean for them.

但这场争议里最大的焦点，无疑是针对大模型研究者的暗中限制。

And yet the biggest deal in the controversy was for sure the secret sabotage for LLM researchers.

对于生物学和化学这类话题，系统会明确告诉用户，它已切换回之前的模型如Opus 4.8；但面对某些类型的LLM研究时，Anthropic却让Fable Five在不告知用户的情况下自行屏蔽回复。

While the approach to the other guardrails around things like biology and chemistry was to explicitly tell the user that it was shifting them back to a previous model like Opus 4.8, when it came to certain types of LLM research, Anthropic said that Fable five would nerve its responses without telling the user.

这引发了前所未有的强烈反弹，最终Anthropic在不到24小时内就撤回了这项政策。

This led to an uproar the likes of which I've never seen, ultimately leading Anthropic to walk back that policy in less than twenty four hours.

后来人们才明白，真正让人不满的不只是这条具体政策，而是意识到Anthropic，以及所有这些前沿实验室，究竟掌握着多大的权力——决定谁能接触到下一代经济的核心工具。

Now it turns out that what people were really turned off by wasn't just that specific policy, but the realization of just how much power Anthropic and, frankly, any of these frontier labs really has to control who gets access to the tools of the next stage of the economy.

事实上，这周我们看到AI实验室面临的三大挑战完全汇聚在一起了：权力——他们对经济命运的掌控力；政策——他们如何运用这种权力；还有公关——他们如何传达这些决定。老实说，Anthropic这周在这三方面全都搞砸了。

In fact, this week, I think we saw the confluence of the three great challenges for AI Labs, power, I e, the inherent power they have over our economic destiny, policy, in other words, the choices they make in terms of how they wield that power, and PR in terms of how they communicate that stuff, which, honestly, Anthropic got all three of them wrong this week.

而本周最后时刻悄悄冒出来的另一件大事，当然是SpaceX的IPO。

Now the one other biggest story sneaking in at the very end of the week was, of course, the SpaceX IPO.

如果你关注主流媒体，大家最热衷讨论的，就是埃隆·马斯克即将成为全球首位万亿富翁。

If you were watching mainstream media, the big thing that they wanted to talk about was Elon Musk becoming the world's first trillionaire.

但我们这些在人工智能行业里的人，更关心的是它的表现，以及这对未来IPO意味着什么。

But for us over in the AI industry, what we were interested in was the performance and what it might mean for future IPOs.

上市首日股价就飙升了19%，Allspring全球投资的罗伯特·格伦迪克说：这预示着市场前景良好，那些即将登场的大规模IPO也会有不错的表现。

Well, on the first day of trading, the stock popped 19%, leading Allspring Global Investments' Robert Grundyke to say: It bodes well for the market and for these other IPOs are coming that are going to be quite sizable as well.

至少在第一天，定价确实很到位。

They clearly priced it right at least on day one.

这应该让你对市场，尤其是成长型股票，更有信心。

It should make you optimistic for the markets, especially for growth stocks.

这周值得关注的另一个新话题，是代币恐慌。

In terms of narrative to watch this week, the big new one was token panic.

从狂热追捧代币到突然陷入恐慌，整个过程快得让人措手不及。

We've gone from token maxing to token panic in a whiplash inducingly short period of time.

所谓代币恐慌，具体指的是企业开始限制员工使用AI的额度，从而引发的负面市场反应——这种趋势未来会越来越常见。

The specific idea of token panic, and one that I think you're going to see coming up a lot more, is the negative market consequences of companies starting to put caps on the amount of AI that their employees can consume.

优步是第一个这么做的例子，但就在我说这话的时候，刚爆出Meta也要跟进实施同样的限制。

Uber was the first example of this, but just as I was recording this, it broke that Meta was gonna be putting those caps on as well.

嘉信证券发布了一份具体的研究报告，表达了这种观点，不过我在周五的节目里说过，他们实际说的和社交媒体上传得完全不一样。

This was given voice in a specific research note by Citadel Securities, although as I explained in my Friday episode, what they were saying was quite different than what social media said they were saying.

你从这场代币恐慌中不该得出的结论，是代币需求突然逆转了。

The thing you shouldn't take away from this token panic is some big idea of token demand rolling over.

真正值得关注的是，未来一段时间内，企业会大力推动代币使用的效率，这会直接影响到我们每个人。

The thing you should take away is that there's going to be a lot of push for token efficiency in the foreseeable future, with implications for all of us.

那下周我们该关注什么？

So what should we be watching for next week?

IPO上市首日的表现，其实根本说明不了什么。

One big thing is that the first day of trading doesn't really mean anything with an IPO.

至于这对其他AI公司IPO意味着什么，我得先看看SpaceX下周怎么行动，再下判断。

When it comes to the implications for the other AI IPOs, I want to see what SpaceX does next week before I really start to make any sort of assessments.

另一个值得关注的重点，是OpenAI会怎么应对Fable。

The other big thing to watch is how OpenAI responds to Fable.

我们会看到GPT 5.6吗？

Do we get GPT 5.6?

我们会看到《华尔街日报》报道的那些大幅降价真的出现吗？

Do we get what The Wall Street Journal reported might be coming in significant price cuts?

现在球在OpenAI这边，大家都在等他们下一步怎么走。

Basically, the ball is in OpenAI's court, and everyone's waiting to see what happens.

最后，这个周末你可以试试这五个针对Fable 5的测试。

Lastly, in terms of what you should try this weekend, here are five tests for Fable five.

如果你还没玩过Fable，我建议你试试这几点：第一，用它来处理策略问题，或者你最近一直在思考的重大问题；第二，来一场第一性原理的辩论，看看它能不能独立反驳你，而不是一味附和你的观点。

If I were you and I hadn't played with Fable yet, I would test it on, one, strategy, some sort of big question you're pondering, two, some sort of first principles debate where you can see how well it pushes back on you without being just subservient to your opinion.

第三和第四，试试让它帮你做研究和写作，就是你平时工作中会遇到的那些实际任务。

Three and four, I would test it on research and writing, some tasks that actually come from your normal work experience.

第五，没错，找一个老大的编程项目，看看它一次性能推进到什么程度。

And five, yes, I would try to find some big old coding project to really see how much farther it can go in a single shot.

Fable 5确实重新点燃了我对好几个项目的热情，因为它比之前任何模型都更接近直接完成任务。

Fable five has certainly reignited my excitement for a number of different projects because it gets much closer to completion than any of the previous models before.

好了，朋友们，这周就先聊到这里。

So friends, that's gonna do it for this week.

希望你这个初夏周末过得愉快，要么看世界杯，要么追《寡妇湾》，等着本周的 finale。

I hope you are heading into a beautiful early summer weekend, watching the World Cup or binging Widows Bay in anticipation of this week's final.

不过现在，今天的《AI每日简报》就先到这里。

But for now, that's gonna do it for the AI Daily Brief.

一如既往，感谢你的收听或观看，咱们下期再见，保重。

Appreciate you listening or watching as always, and until next time, peace.