AI刚刚改变的几种主要方式

今天在《AI每日简报》中，我们来聊聊为什么六月是近年来人工智能领域最重要的一个月。

Today on the AI Daily Brief, why June was the most significant month in AI in years.

《AI每日简报》是一档每天更新的播客和视频节目，聚焦人工智能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.

好了，各位。

Alright, friends.

在正式开始之前，先说几件事。

Quick announcements before we dive in.

首先，感谢今天节目的赞助商：毕马威、Robots and Pencils、Blitzy 和 Hyperagent。

First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Blitzy, and Hyperagent.

想收听无广告版本的节目？请访问 patreon.com/aideailybrief，或者在 Apple 播客上订阅。

To To get an ad free version of the show, go to patreon.com/aideallybrief or you can subscribe on Apple Podcasts.

如果想了解如何赞助本节目，请发送邮件到 sponsorsaidailybrief dot ai。

And to learn more about sponsoring the show, send us a note at sponsorsaidailybrief dot ai.

各位，现在是七月四日周末。

Well friends, it is July 4 weekend.

我大部分美国听众，至少眼下正沉浸在夏日的湖光山色、烟花盛典和浓浓的爱国氛围里，庆祝这个国家成立二百五十年。

Most of my American listeners, at least, are awash in summertime, lakes, fireworks, and patriotic feelings on the two hundred and fiftieth anniversary of this country.

可在这里，AI领域却正从一个月切换到另一个月。

And yet over here at AI, we are shifting from one month to another.

真有点诗意——就在月底最后一刻，Fable 5 及时回归，让我们能顺利在七月继续推进开发。

It is a little poetic that at the very, very end of the month, Fable five got back just in time for us to get back to building in July.

不过在开始之前，我觉得有必要花点时间回望一下上个月，我认为这可能是自ChatGPT以来AI史上最重要的月份之一。

And yet before we do so, it is worth spending a moment, I believe, looking back at the last month, which I would argue is one of the most significant in the post ChatGPT history of AI.

顺便说一句，如果你好奇这个网站配套体验是怎么做的——其实不是用Fable，而是用Codex和GPT 5.5做的。

By the way, for those of you wondering, this website companion experience was actually created not with Fable but with Codex and GPT 5.5.

在聊六月之前，咱们先回溯到五月。

Before we get into June, let's actually go back to May.

作为历史爱好者，我觉得这两个月像一对互补的拼图，讲的是同一个故事，只是角度不同。

The historian in me thinks that these two months kind of make a matched pair, telling the same story but from different angles.

五月的故事，核心是AI从补贴时代转向了令牌稀缺时代。

So the story of May was all about the shift from the AI subsidy era to the token scarcity era.

早在五月之前，我们就已经看到各家服务商逐渐放弃按座位收费的订阅模式，转而采用按使用量计费的方式。

Even before May, we had started to see providers shift away from their seat based subscription models and move towards more usage based models.

这当然是从非代理型工作负载转向代理型工作负载后不可避免的结果——后者消耗的智能资源远远超过了2024年和2025年我们常见的查询类型。

This was, of course, the inevitable consequence of shifting from pre agentic to agentic workloads, which consumed just an absolutely massive amount more of intelligence than the type of queries that we were running back in '24 and '25.

五月的时候，那些急于追求token最大化的企业终于开始尝到苦果。

May was also when we started to see the chickens coming home to roost when it came to enterprises that had run out to start token maximizing.

优步在过去几个月频频上新闻，因为他们在年初的前四个月就把AI预算花光了。

Uber had been in the news for a couple months as it burned through its AI budget in the first four months of the year.

我们不断听到更多公司关闭了他们的token使用排行榜。

We got more and more reports of companies turning off their token leaderboards.

总的来说，五月给人一种新范式即将开启的信号。

And all in all, May felt like the beginning of a shift to a new paradigm.

到了六月初，这种转变真正开始落地了。

Now, at the beginning of June, that started to become real.

月初刚一开始，沃尔玛就从内部工具的无限使用模式，改为了设定token预算。

At the very beginning of the month, we had Walmart moving from unlimited usage of their internal tools to token budgets.

优步宣布每月AI支出上限设为1500美元，这一消息登上了头条。

Uber made headlines when it set a $1,500 per month cap on AI spend.

这些新闻以及其他类似案例，都强化了一个共识：在企业领域，AI的使用效率和成本管控正成为越来越关键的新要素。

And these stories and the others like them reinforced the idea that token efficiency and token discipline were going to become important new aspects the AI landscape, particularly in the enterprise.

从那时起，整个月都在涌现新的方法、新的架构，效率成了重中之重。

And starting then and throughout the month, new approaches, new architectures, efficiency became the name of the game.

说之前每家公司都把最前沿的模型用在所有任务上，这确实有点简化了，但说实话，大多数公司的实际情况也差不多就是这样。

Now it's a little reductive to assume that before this, every company was just applying the most advanced frontier model to every workload, but that's honestly not that far off from what I think the average situation was with most companies.

坦白讲，对大多数公司来说，AI应用还没走到需要认真考虑效率的阶段，因为它们目前消耗的智能资源，只占未来总量的微不足道的一部分。

Frankly, at most companies, AI adoption hasn't proceeded to the point yet where they would really even need to be thinking about efficiency, because most companies are consuming a vanishingly small portion of the total intelligence that they will ultimately consume.

但对于那些走在前沿的公司来说，新效率、新模型架构、转向低成本模型的趋势已经非常明显，包括中国开源模型——这些模型在月底前后开始获得更多关注。

And yet for those companies on the vanguard, there was very clearly a new emphasis on new efficiencies, new model architectures, shifting to lower cost models, including Chinese open weight models, which would become a little more fraud as we would see later in the month.

我们还看到，AI相关的基础设施也在随之调整。

We also saw some indications of the infrastructure around AI adapting as well.

独立评测机构Artificial Analysis调整了其核心智能指数中的部分指标，以更准确地反映智能代理的使用情况。

Independent benchmarking company Artificial Analysis shifted around some of the metrics in its core intelligence index to better reflect agentic usage.

而且，有一件事悄无声息地发生了，我至今觉得它被严重忽视了：微软不仅推出了一套从零开始训练的专有模型，还推出了一款新产品，可以根据特定企业客户的需求和标准来定制训练模型。

And very quietly, in a story that I still think is wildly under discussed, is Microsoft pushing not only a new set of proprietary models that they had trained from the ground up, but a new product where they would post train models to the specific criteria and requirements of a particular enterprise customer.

我觉得这事被忽略了，一方面是因为它被淹没在微软成千上万的其他公告里，另一方面是因为它发生的时间，正好赶在大家开始把‘令牌效率’当成最新热点之前。

I think this missed notice: a) because it was surrounded by a million other Microsoft announcements, but b) it was just before we really started talking about token efficiency as the important idea du jour.

但这个月真正进入高潮，是在6月10日，Anthropic发布了Fable 5。

But the month really kicked into high gear when, on June 10, Anthropic released Fable five.

过去，实验室在跨入全新数字代际时常常让人失望，比如OpenAI从GPT-4过渡到首个GPT-5模型时就是这样，但Fable 5完全不是这种情况。

And while historically it has often been the case that labs have underwhelmed when they've shifted to entire new numerical categories, such as when OpenAI went from the GPT-four class to the first GPT-five model, Fable five was not that.

它一发布就明显强大得多，尤其在技术性和编程类任务上表现突出。

It was immediately and clearly much more powerful, particularly around technical and coding use cases.

不过，就像我之前说过好几次的，虽然最初几天的主流说法是，它对编程任务帮助最大，而在其他领域和Imbroglio Opus、GPT相比优势不明显，但我的实际体验并不支持这个观点。

But honestly, as I've said a couple of times, as much as the initial narrative in those first couple of days was that it made more difference for those coding tasks, and the Imbro Opus and GPT wouldn't be as apparent in other areas, I have found that not to be the case.

我认为，它在所有方面的提升都极其显著。

I think the improvement over the other models in every area is extraordinarily clear.

不过，在Fable 5发布后的头四十八小时左右，大家的焦点还是找那些最复杂、最具挑战性的问题，然后让Fable 5尽情发挥。

Still, for the first forty eight hours or so after Fable five was released, the name of the game was finding your most complex and challenging problems, and letting Fable five just absolutely rip on them.

我没法从技术角度比较4.8和Fable Five的代码有多优雅或高效，但要说它哪里不一样，我最好的描述方式是……

The best way that I can describe what was different about it given that I am nontechnical and can't compare the elegance or proficiency of the code of 4.8 to Fable five.

回顾2026年我用各种编码模型做过的所有项目，我经常做到八九成就停下了，再也推进不下去。

One of the ways that I can describe how it felt different was that if I look back over all the things that I've done throughout the course of 2026 with any of the various coding models, I very frequently get to 80 or 90% of a project and then just don't finish it.

有时候是因为结果没达到预期，或者我注意力转移了，但更多时候是因为，虽然这些模型让我容易起步，却没帮我把最后那点力气补上，没法真正收尾。

Now in some cases, that's because the initial test to results weren't exactly what I wanted, or just my priorities shifted elsewhere, but in a number of cases, it was because while the coding models had made the activation energy low enough to just get started, they hadn't obviated the completion energy to actually get the thing done.

Fable Five是第一个让我觉得，不仅启动大型编码项目变得轻而易举，连收尾都变得毫不费力的模型。

Fable five was the first model that made it feel fairly insignificant not only to start those big coding projects, but to just finish them as well.

你们如果用过新推出的AI Daily Brief网站，把每期内容拆成一个个能分享的小片段，那你们现在享受的便利，就来自这个项目。

And any of you who've enjoyed the new AI Daily Brief website that chunks every episode down into individual shareable components is living in the benefit of that.

这个项目我搁置了好几个星期，直到Fable Five出现，我才一鼓作气，一口气把它做完了。

This is a project that I had had kicking around for weeks at that point, and because Fable 5 came around, I just decided to get it done and done it got in one fell swoop.

真庆幸我当时这么做了，因为现在大家都清楚，Fable Five没存在多久。

And thank goodness because as we know now, Fable five wouldn't be around all that long.

在那最初的几天里，到处都能看到有人完成更复杂、更完整的项目，简直数不胜数。

And in those first couple days, there were basically infinite other versions of people really seeing much more complex and much more complete work getting done.

我们曾看到莱利·布朗用一招搞定了一款类似 Replet 的移动应用构建工具。

We had Riley Brown one shotting a replet mobile style app building app.

我们还看到创作者在测试三维世界功能。

We had creators testing three d World.

有一个客户通话的案例中，Fable 五在和客户对话的同时，直接生成了对方要求的功能。

And one customer call story had Fable five building a requested product feature while the conversation with the customer was still happening.

但这并不意味着 Fable 上线后一切都顺风顺水。

Which is not to say that everything was hunky dory when Fable came out.

几乎立刻就冒出了一些重大问题。

There were a bunch of big questions that surfaced almost immediately.

其中一些问题涉及人们对生物学等领域过于严格的使用限制的担忧。

Now, some of this was about what people thought were overaggressive guardrails around topics like biology.

但另一项限制政策是，Anthropic 实行了三十天的数据保留机制，声称包括 Fable 在内的 Mythos 系列模型的输入和输出都会被保留，用于安全审查。

But one of the other guardrail policies was that Anthropic was instituting a thirty day retention policy, where Anthropic said that prompts and outputs for Mythos class models, including Fable, would be retained for trust and safety review.

这一下子让许多企业直接说：绝对不行。

That immediately made many enterprises say, absolutely not.

如果你们要保留这类数据，我们根本没法用这个。

We can't use this if you're going to keep that sort of data.

这在很多方面预示了一个更深层的权力问题——不只是某一项政策，而是企业们开始意识到，未来商业世界最重要的资产之一，其访问权竟被一家或少数几家公司牢牢掌控。

It was a preview in many ways of a broader power issue where it wasn't just that one policy, but companies realizing how much their access to one of the most important assets in the business world going forward was mediated by a single or small handful of companies.

但回过头看，这些都显得太温和了，因为就在第一周的星期五，Fable 的故事彻底转变了，这个模型反而成了政府直接干预前沿AI访问权的先例。

And yet all of that seems quaint in retrospect because by Friday of that first week, the Fable story had transformed, and the model instead became a precedent for direct government intervention in frontier AI access.

美国政府通过出口管制指令，要求Anthropic暂停外国人士对Fable Five和Mythos Five的访问权限。

The US government used an export control directive to demand that Anthropic suspend Fable five and Mythos five access for foreign nationals.

Anthropic表示，他们唯一能遵守这项要求的方式，就是彻底关闭所有人的访问权限。

Anthropic said that the only way that they could actually comply with that was to shut down access to the model for everyone.

起初，大家以为这事来得特别突然，Anthropic根本没时间反应，但后来的报道显示，实际情况在接下来几天要复杂得多。

Now initially, the story was that this was all extremely abrupt, and that Anthropic had had almost no time to react, although reporting made it clear that it was a little bit more complex over the next few days.

我就不一一细说了，毕竟过去几周已经讲得够多了，但可以肯定的是，我们后来才知道，是亚马逊提交的一份有限的越狱报告，触发了美国政府的这一连串行动，但说到底，这更像是一个导火索，让美国政府各个部门突然意识到，这类模型的威力远超我们以往接触过的任何技术。

I'm not going to recount all of it given that it's been so much of the substance of the last few weeks, but suffice it to say, we would later learn that a narrow jailbreak report from Amazon triggered this flurry of activity in The US Government, but that in many ways it feels like it was a catalyst for various parts of the US government to wake up and realize that this class of models was significantly more powerful than what we had had access to before.

没错，那个具体的越狱漏洞在谈判过程中一直是个争议点，但与此同时，显然还有一场更广泛的追赶和调整正在发生。

While, yes, the specific jailbreak did remain a point of contention throughout the negotiations, there was clearly a broader catch up process happening at the same time.

接下来的几周，整个行业都在等待，看着政府和Anthropic进行谈判。

Now for the next couple of weeks, the industry waited while the government in Anthropic negotiated.

很快，禁令就超出了Fable的范围，GPT 5.6的发布也被推迟了。

Pretty quickly the ban extended beyond Fable as GPT 5.6 got delayed too.

OpenAI宣布，GPT 5.6实际上会包含三个不同的模型，但现阶段，美国政府将逐批审批哪些公司和人员可以访问这些模型。

OpenAI announced that GPT 5.6 would actually be a set of three different models, but that for the time being, the US government would be approving every wave of new companies and people to have access to the model.

对很多人来说，这感觉像是一个毫无章法、临时拼凑的AI授权体系，既没有明确的法律依据，也没有既定的先例，纯粹是凭感觉在拍脑袋决策。

To many, it felt like the beginning of a messy, ad hoc AI licensing regime not based in any sort of determination or legal precedent, but instead a licensing regime that was very much just shooting from the hip.

当然，行业并没有坐以待毙，眼睁睁看着这一切发生。

Now of course, the industry wasn't just sitting around as this all happened.

虽然有人认为Fable 5太强大了，干脆先休息两周再回来，反正它一上线就能解决GPT-5.0或Opus 4.8这段时间里所有没搞定的问题。

Although there were some folks who argued that Fable five was so much more powerful that it made more sense to just take a vacation for two weeks and then come back to it, since Fable five was going be fixing everything that GPT-five point or Opus four point eight had done in the meantime anyway.

对大多数其他人来说，这成了继成本因素之后，另一个促使企业和个人重新考虑绕开封闭式前沿模型的重要原因。

For most others, this became the second major reason after cost considerations for why individuals and companies needed to take another look at alternative approaches to just frontier closed source models.

换句话说，现在企业不仅要考虑成本，还得考虑技术主权，必须重新评估如何摆脱对OpenAI或Anthropic的单一依赖。

Basically, we now had both a cost and a sovereignty dimension for companies to think about diversifying their architecture away from just OpenAI or Anthropic.

现在最重要的AI问题，不是谁在用AI，而是谁用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

普华永道和德克萨斯大学奥斯汀分校刚分析了140万次真实职场中的AI互动，发现一个意外的结果：最有影响力的人并不是擅长写提示词的人，而是把AI当成思维伙伴来用。

KPMG and the University of Texas at Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会明确问题、引导思考、反复迭代，并不断追求更好的答案。

They frame problems, guide thinking, iterate, and push for better answers.

好消息是？

And the good news?

这些行为是可以大规模培养的。

These behaviors are teachable at scale.

如果你正想从单纯使用AI转向真正掌握AI能力，普华永道关于高级AI协作的研究值得你花时间看看。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

了解更多，请访问 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated.

网址是 kpmg.com/us/sophisticated。

That's kpmg.com/us/sophisticated.

我每天在这档节目里探讨AI潜力与实际应用之间的差距。

I cover the capability gap between AI potential and AI reality every day on this show.

大多数公司还在摸索怎么起步。

Most companies are still figuring out how to start.

而Robots and Pencils已经启动并快速扩大规模了。

Robots and Pencils is already launching and scaling.

Agendaq和生成式AI在大型企业中也已在几周内投入生产。

Agendaq and Generative AI in production at large enterprises in weeks.

AWS高级层级合作伙伴的数量在过去一年里翻了一倍多。

AWS advanced tier pattern partner more than doubled in a year.

他们现在正在招聘。

And they're hiring.

有五十个职位空缺。

50 open roles.

如果你意识到这个时代与众不同，想置身其中而不是旁观，那这值得你认真看看。

If you're someone who knows this moment is different, who wants to be inside it, not watching it, this is worth a look.

在 Robots and Pencils，最好的点子才能胜出，团队也始终坚持顶尖水准。

At Robots and Pencils, the best ideas win, and the team is purposefully kept super high quality.

这里是你将来回望时，会觉得这辈子最明智决定的地方。

This is the kind of place you look back on as the best decision you ever made.

去看看 robotsandpencils.com/careers 吧。

Take a look at robotsandpencils.com/careers.

Blitzy 正在帮助大型企业把工程效率提升五倍以上。

Blitzy is driving over 5x engineering velocity for large scale enterprises.

一家上市公司保险公司用 Blitzy 开发了一个定制支付系统，原本预计要十三个月，结果只用了六周就上线运行了。

A publicly traded insurance provider leveraged Blitzy to build a bespoke payments processing application, an estimated thirteen month project, and with Blitzy, the application was completed and live in production in six weeks.

另一家上市的垂直 SaaS 公司用 Blitzy 从一个五十万行代码的单体系统中拆分服务，速度比原先预估快了二十一倍，而且完全没影响线上业务。

A publicly traded vertical SaaS provider used Blitzy to extract services from a 500,000 line monolith without disrupting production 21 times faster than their pre Blitzy estimates.

这些不是实验，是真正在用的方案。

These aren't experiments.

这就是 2026 年全球最具创新力的企业交付软件的方式。

This is how the world's most innovative enterprises are shipping software in 2026.

你可以在《现代CTO与CIO精选播客》中直接听到其他财富500强首席技术官对Blitzy的评价。

You You can hear directly about Blitzy from other Fortune 500 CTOs on the Modern CTO or CIO classified podcasts.

想了解更多Blitzy如何提升你的软件开发生命周期？请前往blitzy.com预约一位AI解决方案顾问。

To learn more about how Blitzy can impact your SDLC, book a meeting with an AI solutions consultant at blitzy.com.

网址是blitzy.com。

That's blitzy.com.

本集《AI每日简报》由Hyperagent赞助播出，你可以在这里管理团队共同使用的智能代理群组。

This episode of the AI Daily Brief is brought to you by Hyperagent, where you run fleets of agents your team can manage together.

新用户可获得1000美元的推理算力额度。

New users get $1,000 in inference.

别再等着本地代理或聊天流程在你的电脑上被触发了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent将智能代理部署在云端，7x24小时运行，无缝对接你团队正在使用的各类工具。

Hyperagent deploys always on agents in the cloud, doing real work across the tools your team already uses.

市场团队的代理能自动把竞争对手的动态转化为落地页面。

Marketing's agent turns competitor moves into landing pages.

销售团队的代理负责完善潜在客户信息、撰写邮件，并更新客户关系管理系统。

Sales's agent enriches leads, drafts emails, and updates the CRM.

运营代理则跟进文书工作，追踪预算使用情况。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享的上下文信息，并严格遵循你设定的范围和审批规则。

Every agent has access to shared context and follows your rules about scope and approvals.

现在是时候引入那些像同事一样默契的智能代理了。

It's time you add agents that feel like teammates.

立即前往 Hyperagent 选购，这款产品由 Airtable 原班团队打造。

Hire yours at Hyperagent, built by the team at Airtable.

前往 hyperagent.com/aidailybrief 领取你的 1000 美元推理额度。

Claim your $1,000 in inference at hyperagent.com/aidailybrief.

整个月里，我们看到大量关于任务路由的探索，许多公司正在构建更复杂的 AI 架构，能更精准地把不同类型的任务分配给合适的模型层级。

Throughout the month, we saw a ton of experimentation with routing companies that would help create more complex AI architectures that were better adept at routing different types of tasks to the right level of model.

但与此同时，新模型也备受关注，其中最引人注目的无疑是 z.dot.ai 推出的 GLM 5.2。

But we also saw a lot of interest in new models, with no model capturing more attention than z dot ai's GLM 5.2.

自从2025年1月深序时刻到来，人们首次接触到DeepSeek R1，体验到了推理模型的强大，随之而来的是NVIDIA市值瞬间蒸发数千亿美元。

Ever since January 2025, when the deepseq moment happened where people discovered DeepSeek R1 and experienced, in many cases, reasoning models for the first time, leading to, among other things, hundreds of billions of dollars being ripped off NVIDIA's market cap.

之后每隔几个月，总有人宣称某个中国新模型也迎来了自己的深序时刻，但直到GLM 5.2出现，才真正配得上这个说法。

Every few months after that, someone was proclaiming that some new China model was having a deepseq moment, but it really wasn't until GLM 5.2 that I think you could legitimately apply that label.

GLM 5.2的性能未必能比得上Fable 5，甚至未必稳超OPUS 4.8或GPT 5.5。

Now, it wasn't that GLM 5.2 was as good as Fable five, or even necessarily OPUS 4.8 or GPT 5.5.

但它确实超越了OPUS 4.6和GPT 5.2的水平，正是这个突破，拉开了2025年底到2026年初智能体时代的序幕，也开启了我们至今仍在经历的这个阶段。

But what it was, was a model that exceeded the OPUS 4.6 GPT 5.2 sort of level, which initiated the agentic era at the end of twenty twenty five and the beginning of twenty twenty six that jumpstarted the period that we've been living through ever since.

对很多人来说，GLM 5.2是第一个让‘退而求其次’不再像妥协、而真能与前沿模型正面竞争的开源权重模型。

For many, GLM 5.2 was the first open weight model that made the fallback strategy feel less like compromise and more like genuine competition for the frontier.

更值得注意的不只是GLM 5.2这类原始模型本身，还有基于它们二次训练的定制模型，比如Cursor的Composer 2.5，它基于Kimi构建，以及那些整合了多模型系统的架构。

And what's more, it wasn't just models like GLM 5.2 and their raw state that were getting attention, but also custom post trained models that were built on top of those open weights things like Cursor's Composer 2.5, which was built off of Kimi as well as integrated architectures that basically had multi model systems built into them.

我们看到Harvey和Fireworks把开源的GLM模型和Opus顾问搭配使用，处理法律任务时效果比单用Opus更好，成本却低得多；Open Router的Fusion系统也出现了，它用一组模型、一个裁判和一个合成器来处理复杂任务，同样承诺以更低的成本达到顶尖水平。

We saw Harvey and Fireworks pair an open weight GLM worker with an Opus Advisor for legal tasks, seeing improved performance over Opus alone for a fraction of the cost, and we also got Open Router's Fusion which used a panel of model, a judge, and a synthesizer for hard tasks, once again promising state of the art level capacity at a lower cost.

当然，说在这段Fable暂停期间所有人都集体转向，那就夸张得离谱了。

Now, it would be wildly overstating it to say that everyone switched en masse during this period of forced pause from Fable.

但这是我做这个节目以来，第一次真正看到本地AI成为一大群此前从未关注过它的人认真考虑的问题。

But for honestly the first time since I've been doing this show, local AI became a serious question for a much wider set of actors than it ever had before.

全球各地的企业高管们都在认真讨论：我们对本地AI和开源模型的政策是否需要重新评估。

You had genuine enterprise boardroom conversations all around the world, asking what their policy relative to local AI and open weight models was, and whether that should be reevaluated.

在没有新模型可用的这段时间里，大家把更多注意力放在了围绕模型的工具链和生态系统上——这些同样关键，决定了AI如何真正融入实际工作流程。

Now the other thing that happened in the absence of Fable was that since we didn't have a new model to play with, there was a lot more emphasis on the harnesses and ecosystems that surrounded the models as an equally important part of how AI gets integrated into real world work systems.

显然，2026年的一大主题就是工具链工程的兴起，从OpenCLaw开始，一路延续至今。

Now, obviously, one of the big themes of 2026 has been harness engineering kicking off from OpenCLaw and running right on through.

从这个角度看，其实也没什么新鲜的。

So in some ways, that's nothing new.

但在Fable停摆的这段时期，各种新功能、新发布和新实验，都让这一切变得更加清晰和具体。

But there were a slew of new features and announcements and experiments that really put a fine point on all of this in that Fable Pause period.

Anthropic和OpenAI都推出了更专注的网页或网站生成工具，促使知识工作者重新思考他们日常使用的传统工作成果形式。

Both Anthropic and OpenAI pushed some version of a more dedicated HTML or website artifact builder, getting knowledge workers to think differently about the traditional artifacts that they'd use to do their work.

我专门做了一期节目，讲了各种类型的知识工作，为什么现在该考虑用网站替代过去常用的表格或幻灯片。

I did a whole episode about all the different types of knowledge work where you should think about building websites instead of the spreadsheets or slide decks that you used to use.

人们也越来越意识到，人工智能战略本质上应该是一种生态系统战略。

There was also a growing sense that AI strategy needed to be an ecosystem strategy.

Fable 5下线后不久，微软CEO萨提亚·纳德拉在X平台上发了一篇长文，指出每家公司都必须围绕AI使用建立一个持续学习的机制和系统。

In the immediate aftermath of Fable 5 being taken offline, Microsoft CEO Satya Nadella wrote a long post on X about how every company needed to build a learning loop and a learning system around its AI usage.

企业不能只挑对的模型，还得掌控模型使用过程中那些不断累积的上下文、评估和记忆体系。

Basically, firms didn't just need to choose the right model they needed to own the compounding context decisions, evaluations, memory that surrounded the usage of models.

还有一个值得特别提一下的新功能，叫Claude Tag。

One more feature that was announced that I do think is worth specific note was Claude Tag.

Claude Tag不只是通过Slack与Claude互动的另一种方式，而是让Slack里任何地方的人都能调用Claude Code的强大能力。

But Claude Tag wasn't just another way to interact with Claude via Slack it was instead a way for people in any part of Slack to call upon the power of Claude code.

这让更多人能接触到Claude Code的高级技术能力，也让它能获取更持久的上下文，开始把AI从个人工具转变为团队协作工具，效果相当显著。

This democratizes access to the advanced technical capabilities of Claude code, it gives Claude code access to more persistent context, and it started to shift AI from an individual experience to a group experience in ways that apparently have had fairly dramatic impacts.

人们之所以特别关注这一点，是因为Anthropic团队谈起它时那种近乎敬畏的态度。

One of the reasons that people took notice of this was the reverence almost with which Anthropic's team was talking about it.

其中最引人注目的说法之一是，Anthropic的产品团队现在有65%的代码，不是在Claude应用或终端里写的，而是通过Slack启动Claude Code生成的。

One of the biggest headline grabbing claims was Anthropic saying that 65% of its product team code was now being produced not in the Claude app or in the Claude code terminal experience, but by initiating Claude code from Slack.

回到五月，从代币补贴转向代币稀缺阶段，这一转变的直接原因不仅仅是代理式使用带来的算力需求激增。

Now going back to May, and the shift from the token subsidy to the token scarcity era, the proximate causes of that shift were not just the increase in workloads that came along with agentic usage.

更关键的是，这些短缺会随着我们触及现有计算基础设施及其扩展所需的物理设施的极限而进一步加剧。

It was also the fact that those shortages are going to be amplified as we run up against the limits not only of our existing compute infrastructure, but the surrounding physical infrastructure that's needed to expand that compute.

本月市场的一个重要趋势是内存厂商表现突出，因为内存短缺问题逐渐浮出水面。

One of the big themes in markets this month was the outperformance of memory companies as the memory shortage came into focus.

计算本身正在成为一个独立的市场。

Compute itself is becoming a market of its own.

这一趋势最初由SpaceX推动，他们扩大了与Anthropic的合作，随后谷歌和Reflection AI也陆续跟进。

This has certainly been led by SpaceX, who expanded their Anthropic deal, to other similar deals with Google and Reflection AI.

现在有报道称，Meta和扎克伯格正步SpaceX和埃隆的后尘，无意中进入了这片所谓的‘新云’领域。

And now it's being reported that Meta and Zuckerberg are following Elon and SpaceX into that sort of accidental Neo Cloud space.

随着每个月过去，人工智能——尤其是通过数据中心——在政治讨论中变得越来越敏感。

And with every month that goes on, AI, specifically via data centers, become more and more of a hot button when it comes to political discourse.

六月在某种程度上算是个低谷，但你能感觉到左右两派都在暗中酝酿着什么。

June in some ways was actually a fairly low ebb, but you can feel things brewing from the left and from the right.

一边是艾伦·布罗科维奇，另一边是前茶党保守派，他们都反对同一件事，这种事自然会成为政治讨论的焦点。

Anytime you've got Aaron Brockovich on the one hand and former Tea Party conservatives on the other mobilizing against the same thing, it's gonna be part of the political discourse.

对于企业来说，有些关于前沿技术的讨论，可能离他们的实际经验太远了，毕竟他们还处在AI采用的早期阶段。

Now for enterprises, some of this discourse around the frontier is going to seem so far outside of their lived experience just based on where they are in the adoption cycle.

虽然少数早期采用的企业正在应对令牌效率这类问题，但大多数处于平均水平的企业，正面临代理工作带来的新挑战，比如Glean报告中提到的‘机器人陪护’现象。

Indeed, while a tiny sliver of early adopter companies are dealing with things like token efficiency, companies that fall more in the average band when it comes to AI adoption are uncovering new challenges around agentic work, like this new phenomenon of bot sitting that was identified in a Glean report.

所谓‘机器人陪护’，就是为让AI真正能用起来所做的一系列工作。Glean的调查显示，员工平均每周要花6.4小时，给AI提供上下文、检查它的输出、重跑那些不理想的结果。

Bot sitting is basically all the work that goes around making agents work, and Glean found workers in their survey spending an average of six point four hours per week making AI usable through things like feeding it context, checking its outputs, and rerunning underwhelming results.

某种程度上，六月再次证明，能力过剩的问题不会靠新模型自己解决——事实上，新模型只会让问题更糟，真正解决它还得靠实实在在的管理变革。

In many ways, June reinforced that the capability overhang is not just going to be solved by new models in fact, new models are going to make it worsenseand that it's only going to be solved by real change management.

最近一期普华永道季度脉搏调查中，一个明显的变化是，越来越多的首席执行官开始亲自把AI当作战略重点来抓。

One big jump we saw in the most recent KPMG Quarterly Pulse survey was the growth of CEOs actively owning AI as a strategic priority.

他们还发现，当CEO对AI负责时，企业更有可能从AI应用中获得切实的业务价值——这种可能性是CEO不负责时的两倍以上。

They also uncovered some value around that, finding that organizations where CEOs were accountable for AI versus CEOs were not accountable for AI were more than twice as likely to report meaningful business value being gained from using AI.

那么，进入七月，我们最该关注哪些关键问题？

So as we head into July, what are the big questions?

尽管Fable已经回来了，但情况依然没有解决。

Despite the fact that we have Fable back, the situation remains very unresolved.

比如，我们现在还不清楚政府和Anthropic达成的协议会对GPT 5.6的发布产生什么影响。

We don't, for example, know right now how whatever agreement the government reached with Anthropic impacts the release of GPT 5.6.

而且既然有报告称，更先进的GPT和Anthropic模型已经在路上了，这种临时的、非正式的授权机制能否应对这些新模型，也还不清楚。

And given that we've got reports that there are now even more advanced GPT and Anthropic models waiting in the wings, it isn't clear how this ad hoc, informal licensing regime is going to deal with those new models either.

所以接下来的一个方向，必然是政策层面的问题会越来越多。

So one strand of what happens next is going to be inevitably more and more questions on the policy side.

与此同时，对企业来说，我认为这段时期会留下一个持久的影响——关于AI的讨论边界已经发生了显著变化，不再只是死守最先进的封闭模型。

Meanwhile, for companies, I do think that there will be a lasting legacy of this period of a pretty significant shift in Overton window when it comes to not just being locked into whatever the state of the art closed frontier model is.

但你也注意到，像政策、AI授权机制、企业重新设计令牌架构、考虑开源模型这类问题，都不是短期能解决的。

And yet you'll note that these types of questions policy and AI licensing regimes, companies redesigning their token architectures or thinking about customer open models these are not short term changes.

相反，它们正在为今年余下的时间乃至更长远的未来绘制路线图。

Instead, these are setting a map for the rest of the year and beyond.

不过眼下，Fable 5确实已经恢复了。

Now, in the immediate term, we do have Fable five back.

我认为七月和八月其实是个难得的机会，趁这个时候抢先一步会很有优势。

And I think there is actually a fairly unique opportunity in July and August for people to take advantage of that to race out ahead.

不管新技术多酷，企业界很多部门一到夏天这个时候就完全进入休眠状态。

No matter how cool the new technology is, there are big chunks of the corporate world that really turn off for this part of the summer.

尤其是如果你就在这个圈子里做事的话。

And honestly, especially if you are operating in that world.

如果你不休假，反而利用这段时间搞清楚这批新模型到底能做什么，你就有机会大幅提升自己对关键方的价值。

If you do not turn off, and instead use this time to really see what this new class of models can do, you have a chance to significantly increase your value to whoever you need to be valuable for.

和往常一样，我一直在想我们能怎么帮上忙，但在我看来，这才是整个夏天真正的机会。

As always, I am thinking about ways we might be able to help with that, but that I see as the real opportunity for the rest of the summer.

说到底，大家应该不会惊讶，我觉得六月在历史上会是极其重要的一笔。

Anyways, guys, it'll come as no surprise just how significant I think June will go down in history.

如果说2026年初是实时智能应用爆发的起点，那年中就是开始正视这种能力暴涨带来的后果和挑战。

Appropriately, if the beginning of twenty twenty six was all about the explosion of RealAgentic use cases, the middle of twenty six was about recognizing the consequences and challenge of that increase in capacity.

而2026年剩下的时间，就是要搞清楚接下来该怎么走。

And the rest of 2026 is gonna be all about figuring it out from here.

好了，各位，希望你们无论在哪里，都能迎来一个美好的周末。

Anyways, friends, I hope that wherever you are, you are heading into a wonderful weekend.

好好享受吧，和家人朋友一起。

Enjoy, friends and family.

美国独立日快乐。

Happy birthday, America.

一如既往，感谢你们的收听或观看。

Appreciate you listening or watching as always.

那我们下次再见，保重。

And until next time, peace.