如何应对下一波人工智能竞争浪潮

上周晚些时候，OpenAI 宣布将切断对 Cursor 中其模型的访问。

Late last week, OpenAI announced that they would be cutting off access to their models in Cursor.

对许多 Cursor 用户来说，这无疑是个巨大的打击。

Now for many Cursor users, this is a huge blow.

他们投入了大量资源在这个工具生态上，认为 OpenAI 此举有些小家子气，像是在搞竞争报复。

They invested in that harness ecosystem and view OpenAI's move as some version of competitive pettiness.

但也有观察者认为，从 SpaceX AI 决定收购 Cursor 那一刻起，这一切就注定会发生。

Other observers think that this was always inevitable as soon as SpaceX AI decided to buy Cursor.

他们举了其他例子，比如 Anthropic 在 2025 年年中切断对 Windsurf 的支持，来说明这不过是前沿实验室之间竞争的新常态。

They point to other examples like Anthropic cutting off Windsurf in mid twenty twenty five to say that this is just the way that Frontier Lab competition is going to work from here on out.

虽然这一切看起来可能只是实验室之间的心理战，但对于企业级 AI 用户来说，影响却非常深远。

And while this all might seem like just psychodrama competition between the labs, for enterprise AI users, it has very significant implications.

其实，企业界早就在努力理解和更好地利用开源权重模型，希望构建更复杂的模型架构，让用户能更精准地匹配任务和模型能力，从而降低成本。

Already, there was a push in enterprises to understand and to better be able to take advantage of open weights models, To build the capability to have more complex model architectures that allow their users to better match the task with the capability level in a way that is more cost effective.

这些举动清楚地表明，这不仅仅是成本效益的问题，更关乎控制权和抗风险能力。

What these moves make clear is that this is not just a cost efficiency conversation, but is more broadly about control and resilience.

这些最新的举动清楚地表明，企业不仅要在模型层面考虑这些问题，还要在工具链层面考虑。

And what these latest moves make clear is that enterprises have to think not only about these questions in terms of models, but in terms of harnesses as well.

《AI每日简报》是一档每日更新的播客和视频节目，内容涵盖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.

好了，朋友们。

Alright, friends.

在开始之前，先快速说几个通知。

Quick announcements before we dive in.

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

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

想要收听无广告版本，请访问 patreon.com/aidailybrief，或者也可以在Apple Podcasts上订阅。

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

如果想了解如何赞助本节目，请发送邮件至 sponsors@aidailybrief.ai。

And to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai.

最后，随着大家准备返校或重返工作岗位，如果你需要更多支持，最新一期超级智能高管代理培训项目将在劳动节后不久启动。

Lastly, as you guys get prepared for back to school, back to work, if you are looking for more support, the latest crop of super intelligent executive agent training programs is coming up just after Labor Day.

你可以从 idailybrief.ai 上找到那个信息，或者直接去 training.bsuper.ai。

You can get that information linked off of a idailybrief.ai, or you can go to training.bsuper.ai.

如果你想要提升自己，或者喜欢我和 New Far Gaspar 的嘉宾对话，我强烈推荐你去看看那些课程。

For anyone who's looking for a boost or who has just liked my guest episodes with New Far Gaspar, I highly recommend you check those programs out.

今天，我们先来聊聊数据中心争论的最新进展，蓝领工会正在组织起来支持数据中心的建设。

Today, we begin with a follow-up in the ongoing saga of the data center debate where blue collar labor unions are organizing to support data center construction.

随着数据中心成为中期选举的关键政治议题，工会正在成为反对声浪中的反制力量，他们威胁要撤回对反对数据中心的候选人的支持。

As data centers become a pivotal political issue for the midterms, labor unions are becoming the backlash to the backlash and are threatening to withhold support from candidates who oppose data centers.

《华尔街日报》看到了一份来自蒸汽管道工工会 UA Local 602 的备忘录，该工会的成员在弗吉尼亚州和马里兰州安装工业管道，覆盖了被称为“数据中心走廊”的地区。

The Wall Street Journal viewed a memo from steamfitter UA Local six zero two whose members install industrial piping across Virginia and Maryland, covering the region known as Data Center Alley.

工会表示，他们划出了一条“明确的界线”，不会支持任何持反数据中心立场的政客。

The union said that they were drawing a, quote, clear line and won't back any politicians running on an anti data center platform.

他们还补充说，这对 Local 602 来说是一个生死攸关的时刻。

They added, this is an existential moment for Local six zero two.

在某些情况下，工会甚至因为这个议题，打破了与当地民主党长期以来的联盟关系。

In some cases, unions are breaking their long standing alliance with local Democrats over the issue.

堪萨斯州的暖通空调和铁路工人工会几十年来首次因为这个问题支持共和党州长候选人。

The Kansas HVAC and Railroad Workers Union have supported the Republican gubernatorial candidate for the first time in decades over the issue.

悉尼·博尼拉，蒸汽管道工地方602分会的财务主管，强调工会的支持不仅仅是选票的问题。

Sydney Bonilla, the treasurer for Steamfitters Local six zero two, emphasized that union support isn't just about votes.

他说，他们的工会还参与上门拉票和筹款，以支持地方竞选活动。

He said his union also participates in door knocking and fundraising to support local campaigns.

博尼拉预计全国各地的工会都会效仿，拒绝支持反对数据中心的候选人，并评论说：“我们依赖这些工作。”

Bonilla expects unions across the country to follow suit in rejecting anti data center candidates commenting, We are dependent on these jobs.

其他工会领导人也在质问政客们站在哪一边。

Other union leaders are asking politicians which side they are on.

弗吉尼亚和马里兰电工工会地方26分会的政治协调员唐·斯拉门说：“如果你剥夺了绝佳的职业机会，你就不是朋友。”

Don Slamen, the political coordinator for Electrician Workers Local twenty six in Virginia and Maryland said, You're not a friend if you're taking away great career opportunities.

这是一个千载难逢的机会，真正进入中上层阶级。

This is a once in a generation opportunity to really get in the upper middle class.

如果你过去几个月一直关注我对这个问题的报道，你会看到这个问题开始冒头并逐渐显现。

Now if you have been paying attention to my coverage of this issue for the last several months, you will have seen this start to bubble and emerge.

其实，在我上一期节目里，我已经对这个话题做了个铺垫。我当时的一个观点就是，工会最适合在社区和建数据中心的科技公司之间充当中间人，因为工会既深深扎根于这些社区，又能从它们带来的变革中获得经济利益。

In fact, in my last episode that was a primer all about this issue, one of my arguments was that these labor unions were most uniquely suited to intercede in the middle between these communities and the tech companies that are building the data centers, given that the unions are both deeply rooted in those communities, but also stand to benefit economically from the transformation that they bring.

总的来说，我觉得这是一个非常积极的进展，能给目前缺乏理性的讨论带来很多共识。

Overall, I think it's an extremely positive development that could bring a lot of common rationality to the discussion that has been lacking thus far.

对某些人来说，这是一个扭转局面的时刻。

For some, this is a moment to turn the tide.

投资人加文·贝克写道：大约18个月前，人们对数据中心有一些合理的担忧，比如用水、税收、就业、电价、环境，以及对小镇的影响。

Investor Gavin Baker wrote: There were reasonable concerns about data centers eighteen ish months ago: water, taxes, jobs, electricity prices, the environment, and what they would do to small towns.

如今，规划良好的数据中心项目在很大程度上已经解决了这些问题，我们应该为此感到高兴。

Well structured data center projects have largely addressed these concerns today and we should be celebrating this.

总的来说，数据中心对美国来说，从各方面看都是非常有益的。

On balance, data centers are awesome for America in every way.

就连英伟达的黄仁勋也转发了加文的帖子，并说：说得太对了。

No less than NVIDIA's Jensen Huang reposted Gavin and said, Spot on.

人工智能正在把制造业带回美国，让这个国家在经历了数十年的外包之后重新工业化。

AI is bringing manufacturing back to America and re industrializing the nation after decades of offshoring.

我们有机会为全美各地的社区带来持久的利益，并帮助美国引领下一场工业革命。

We have the opportunity to create lasting benefits for communities across America and help America lead the next industrial revolution.

接下来，另一个频繁登上头条的话题。

Next up, another frequent topic in the headlines.

特朗普政府正在制定规则，以防止中国实验室远程访问AI芯片。

The Trump administration is developing rules to prevent Chinese labs from getting remote access to AI chips now.

本月早些时候，CNBC报道称，多家中国公司通过泰国、马来西亚和日本的数据中心枢纽，获得了英伟达尖端芯片的使用权。

Earlier this month, CNBC reported that multiple Chinese firms had access to cutting edge NVIDIA chips through data center hubs in Thailand, Malaysia, and Japan.

除了从第三方运营方租用计算能力外，报道还称，其中一些数据中心由阿里巴巴和字节跳动所有。

Alongside renting compute from third party operations, the reporting claimed that some of the data centers were owned by Alibaba and ByteDance.

重要的是，这些行为并不违法，因为出口管制只针对物理出口到中国的芯片。

Importantly, none of this was illegal as the export controls only govern physical exports into China.

但当然，如果政府的目标是阻止尖端芯片的获取，那么这种安排就破坏了整个政策。

But of course, if the administration's goal was to block access to cutting edge chips, then this sort of arrangement undermines the entire policy.

据The Information报道，商务部目前正在制定一项新规则，以堵住这一漏洞。

The information reports that the Commerce Department is currently working on a new rule aimed at closing the loophole.

消息人士称，这是AI扩散规则的简化版，该规则在拜登政府最后一周推出，特朗普上任后立即被废除。

Sources described it as a slimmed down version of the AI diffusion rule, which was introduced in the final week of the Biden administration and immediately scrapped once Trump took office.

该规则因巨大的执法和行政负担而受到严厉批评，但它本会让中国实验室难以在第三国建立数据中心枢纽。

The rule was heavily criticized for having a huge enforcement and administrative burden, but it would have made it difficult to establish third country data center hubs for the Chinese labs.

其中，扩散规则将非盟友国家的AI芯片进口上限设得非常低。

Among other things, the diffusion rule capped AI chip imports at a very low level for unaligned nations.

如果各国政府与美国合作，确保其数据中心不会为中国公司服务，这些上限可以提高。

Those caps could be raised if national governments worked with The US to ensure their data centers wouldn't service Chinese companies.

目前，报道中尚不清楚新规则的具体范围，也不清楚它是否在政府内部获得支持。

Now, it's unclear from the reporting exactly how far the new rule would go, and whether it even has support within the administration.

扩散规则的批评者认为，它会削弱美国在芯片行业的主导地位，促使世界大部分地区采用中国技术。

Critics of the diffusion rule argued that it would weaken America's dominance in the chip industry, pushing most of the world to adopt Chinese technology.

商务部高级官员对扩散规则公开批评，副部长杰弗里·凯斯勒在7月告诉国会：“我不想替换扩散规则，因为我认为它不值得替换。”

Senior officials at the Commerce Department have been vocally critical of the diffusion rule, with Under Secretary Jeffrey Kessler telling Congress in July, I don't want to replace the diffusion rule because I don't think the rule is worth replacing.

这是个糟糕的规则，我们很高兴它没有被执行。

It's a bad rule, and we're glad that it's not being enforced.

但如果你关注过这届白宫在AI政策上的任何动向，你就会知道，三个人里能冒出四种意见，所以我们只能等着看最终结果如何。

But if you have watched anything when it comes to AI policy out of this particular White House, you know that among three people there's going be four opinions, so we'll just have to wait and see where it lands.

说到这届政府，Anthropic赢得了对五角大楼的诉讼，联邦法官裁定政府没有理由将他们列为供应链风险。

Speaking of this administration, Anthropic has won their lawsuit against the Pentagon with a federal judge ruling that the government had no basis for declaring them a supply chain risk.

在她的裁决中，美国地区法官丽塔·林认为，政府没有提供证据表明Anthropic对国家安全构成真正的威胁。

In her order, US district judge Rita Lin found the government hadn't provided evidence that Anthropic represented a genuine threat to national security.

相反，林法官写道，被告当时的言行证实，受到质疑的行动是基于一种意图，即拿Anthropic当众做文章，因为它所谓的“傲慢”批评政府。

Instead, judge Lynn wrote, defendant's contemporaneous words and deeds confirm that the challenged actions were based on a desire to make a public example out of Anthropic for its, quote, unquote, arrogance in criticizing the government.

空洞地搬出国家安全，并不是一张可以随意惩罚和报复政府批评者的空白支票。

The empty invocation of national security is not a blank check to punish and retaliate against government critics.

特别值得注意的是，林法官指出，政府继续使用Anthropic的模型，这削弱了五角大楼的说法。

In particular, Judge Lynn noted that the government's continued use of Anthropic's models undermined the Pentagon's claims.

该裁决认定，政府在做出这一认定时违反了第一修正案和第五修正案。

The order found that the government had violated the first and fifth amendments in making the designation.

林法官写道，尽管战争部无疑可以自由选择其AI供应商，但证据表明，对Anthropic施加的广泛措施是非法的、毫无根据的。

Judge Lynn wrote, though the Department of War is undisputedly free to select the AI vendor of its choice, the evidence demonstrates that the broad measures imposed on Anthropic were illegal and baseless.

该命令要求政府撤销所有将Anthropic列入供应链风险黑名单的指导和指令。

The order directed the government to rescind all guidance and directives that blacklisted Anthropic as a supply chain risk.

因此，该命令要求政府撤销所有将Anthropic列入供应链风险黑名单的指导和指令。

Consequently, the order directed the government to rescind all guidance and directives that blacklisted Anthropic as a supply chain risk.

虽然这对Anthropic来说是个大胜利，但他们肯定还没完全脱离困境。

Now, while this is a big win for Anthropic, they are certainly not out of the woods just yet.

第二起诉讼在华盛顿特区上诉法院仍在等待裁决，而该案的法官到目前为止对五角大楼的论点稍微更接受一些。

A second lawsuit in the DC Appeals Court is still awaiting a ruling, and the judge in that case has been so far a bit more receptive to the Pentagon's arguments.

说到今年早些时候Mac Mini热潮的后续，有个有趣的小故事。

One funny little story in a follow-up to the Mac Mini craze of earlier this year.

在苹果最近一次的财报中，他们表示Mac的销量在过去一年增长了29%，这是公司所有产品线中增长最快的。

In their most recent earnings report, Apple said that Mac sales were up 29% over the past year, which was the fastest growth of any product line in the company.

当然，对于知情者来说，OpenClaw的爆发是其中很大一部分原因，估计带来了超过一亿美元的Mac Mini销售额。

And, of course, for those in the know, the OpenClaw boom was a big part of that, leading to an estimated $100,000,000 plus of Mac Mini sales.

大多数人以为这只是消费者趋势，爱好者和早期采用者抢购Mac Mini来运行他们新的智能体套件。

Now most assumed this was just a consumer trend, with hobbyists and early adopters snatching up Mac Minis to run their new suite of agents.

上周新款Mac Mini发布时，虽然配置提升了，但价格也明显上涨，可能让那些普通消费者有点望而却步了。

When the new Mac Mini line was unveiled last week, although the specs were up, the price was up meaningfully as well, making it potentially a little bit more out of reach for that generalist sort of audience.

不过，周末的时候，《The Information》发了一篇深度报道，说Mac Mini火爆的背后，其实很大一部分是企业需求。

Over the weekend, however, The Information published a deep dive on how, in fact, a big part of the Mac Mini explosion has been enterprise demand.

六月份，苹果举办了一场面向企业的硬件销售活动，重点推Mac Mini，把它定位成一种节省成本的方案，让简单的AI代理和模型可以在本地运行，而不是让云账单越堆越高。

In June, Apple held an enterprise focused hardware sales event with significant emphasis on the Mac Mini, which was pitched as a cost cutting measure, allowing simple agents and AI models to be run locally instead of contributing to rising cloud bills.

苹果前企业营销经理Todd Daily说，这跟苹果一贯的做法太不一样了。

Former Apple Enterprise Marketing Manager Todd Daily remarked on how out of character this was for Apple.

苹果没有专门面向企业客户的工程团队，甚至连开发者关系团队都没有。

Apple doesn't have a dedicated engineering team for business customers or even a developer relations team.

Daily评论说，要说苹果哪个团队真有什么拥抱企业AI的计划，那纯属开玩笑。

Daily commented, The idea that any team at Apple has an actual plan for embracing enterprise AI is a joke.

不过，在某些领域，企业需求确实非常真实。

Still, enterprise demand seems to be very real in some pockets.

有消息人士告诉《The Information》，OpenAI已经买了数万台Mac Mini和Mac Studio，用来做强化学习。

Sources told the information that OpenAI has purchased tens of thousands of Mac Minis and Mac Studios for reinforcement learning.

这些机器用来训练计算机使用代理，消息人士说OpenAI非常想买更多。

The machines are used to train computer use agents, and sources said OpenAI is desperate to buy more.

据知情人士透露，Anthropic似乎也在从AWS租用Mac Mini。

Anthropic is apparently also renting Mac Minis from AWS according to people familiar with the operations.

换句话说，即使价格上涨，这款不起眼的Mac Mini似乎仍将在下一波AI浪潮中扮演重要角色。

In other words, even with the price increases, it seems that the humble Mac Mini will continue to play a significant role in the next wave of AI.

说到OpenAI，今天节目的主题是关于OpenAI上周末做出的一个重大决定，以及这对企业和公司意味着什么，他们必须为新的竞争现实做好准备。

Now speaking of OpenAI, the subject of today's main episode is about a big decision that OpenAI made at the end of last week, and what it means for enterprises and companies who have to position themselves for a new competitive reality.

说到竞争现实，OpenAI最近推出了一些大幅降价措施。

Well, speaking of competitive realities, OpenAI recently introduced some significant price cuts.

他们通过API将GPT56 Luna的价格降低了80%，而更大的Terra版本降低了20%。

They cut prices on GPT56 Luna by 80%, and the larger Terra version by 20% through the API.

他们后来宣布SOL降价20%，不过这个价格只生效了几个星期。

They later announced a 20% price on SOL, although that pricing has only been in effect for a couple of weeks.

这一切的预期目标显然是推动OpenRouter等第三方平台上的使用量。

Now the assumed goal of all of this is to drive up usage on third party platforms like OpenRouter.

这很可能说明前沿实验室现在不仅要在原始能力上竞争，还要在效率的前沿上竞争。

This could be part and parcel of a recognition that the Frontier Labs now have to compete not just on the Frontier when it comes to raw capability, but also when it comes to the frontier of efficiency.

不过也有人猜测，因为媒体把OpenRouter这样的第三方平台使用量当作整体token消耗的代理指标，虽然这其实是对数据的严重误读，OpenAI的目的可能是想通过一个相对小的动作获得超大的公关效果。

However, some also speculated that because media uses third party platform usage like OpenRouter as a proxy for overall token consumption, even though that's a pretty massive misread of the data, that perhaps OpenAI's goal was to get an outsized PR effect from a relatively minor move.

随着Anthropic未来几个月IPO前竞争加剧，这可能是OpenAI制造一些令人担忧的新闻标题的方式。

With the battle heating up ahead of Anthropic's IPO in the coming months, this could be a way for OpenAI to generate some concerning headlines.

不管动机和最终目标是什么，OpenRouter报告称OpenAI模型的使用量大幅增加。

Whatever the motivation and the ultimate goal, OpenRouter is reporting a massive boost in usage for OpenAI's models.

他们报告说，折扣后Terra的日使用量增长了5.6倍，而Luna更是暴涨了13.8倍。

They report that daily usage of Terra is up 5.6 x after the discount, and Luna rose a massive 13.8 x.

在他们观察的时间窗口内，SOL还没有打折，使用量相对平稳，只增长了10%。

SOL was not yet discounted during the window they looked at and its usage was relatively flat with just a 10% gain.

而且，虽然Terra和Luna的折扣在8月14日到期，但OpenRouter报告说用户留了下来，近三分之一的人继续以全价使用OpenAI的模型。

What s more, while the discounts on Terra and Luna expired on August 14, Open Router reports that users stuck around, with nearly a third of them continuing with OpenAI s models at full price.

Box的Aaron Levy转发了这张图表，并说：有时候人们对J形曲线悖论在token消耗上的表现没有直观感受。

Box's Aaron Levy reposted the chart and said: Sometimes people don't have an intuitive sense of what J Bond's Paradox looks like for token consumption.

我们这些跟企业打交道的人，每天都能亲眼看到这一点。

Those of us working with enterprises get to see this firsthand every day.

基本上，企业有源源不断的任务，他们很想把这些任务自动化，但每个任务是否值得自动化，取决于自动化的成本，也就是投资回报率是否为正。

Basically, enterprises have an unending stream of tasks that they'd love to be able to bring automation to, but for each individual task, it's either ROI positive or not, based on the cost of bringing automation to it.

当令牌在某个能力水平上变得更便宜时，企业就能负担得起将更多令牌用于他们的工作。

As tokens get cheaper at a certain capability threshold, enterprises can afford bringing more of them to the work that they do.

这可能包括处理每一份合同、阅读每一条日志、监控新的数据流以获取洞察、让后台代理执行工作流等等。

This could be processing every contract, reading every log, watching new streams of data for insights, having background agents execute workflows, and so on.

每次我们降低令牌的成本，都会看到消费量不成比例地增长。

Any time we can lower the cost of tokens we will see a disproportionate increase in consumption.

即使令牌价格下降50%，对于这类工作负载，令牌使用量也可能增加5倍。

Even a 50% drop in token prices could result in a 5x increase in tokens for these kinds of workloads.

这就是为什么持续降低AI成本至关重要，也是为什么这对所有市场参与者都有利。

That's why it's critical to keep bringing down the cost of AI, and why that's good for all market participants.

Vercel的Brandon Gayling补充说：除了杰文斯悖论，更便宜的令牌意味着，考虑到组织的支出意愿和对实验的风险承受能力，整个用例的可行性从零变成了一。

Vercels Brandon Gayling added: Beyond just JVONS paradox, cheaper tokens means entire use cases go from zero to one in viability, given an organization's willingness to spend and their risk tolerance for experiments.

很多任务实际上需要一定数量的令牌才能完成。

Many tasks require some minimum quantity of tokens to actually do the job.

如果定价不允许达到这个最低可行性，那就不会去做。

If the pricing doesn t allow you to hit that minimum feasibility, it won be done.

当定价下降时，企业能够达到价值点，并批准之前不可行的用例。

When the pricing drops, enterprises are able to get to the point of value and greenlight use cases that were previously not viable before.

他们的总令牌支出增加了，但从中获得的价值和投资回报率也增加了。

Their total token spend increases, but so does the value and ROI they get from it.

现在，就像我说的，今天的主要节目都是关于一些新的竞争举措。

Now like I said, today's main episode is all about some new competitive moves.

我想以OpenAI竞争性定价实验的影响这个故事作为结尾，是因为它体现了我认为我们正在进入的实验性时刻。

And the reason I wanted to end on this story about the impact of OpenAI's competitive pricing experiments is that it's exemplary of the experimental moment that I think we're heading into.

但就这样，让我们结束头条新闻，进入主要节目。

But with that, let's close the headlines and move on over into that main episode.

毕马威和德克萨斯大学奥斯汀分校的一项新研究发现，当人们与AI合作时，相似的技能并不能保证相似的结果。

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

研究人员研究了500多名早期职业人士，发现表现最好的人通过引导、评估和优化AI的输出，持续放大了AI的价值。

Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

这些顶尖表现者，被称为AI放大器，他们的优势不在于单纯的知识，而在于他们与AI合作的方式。

These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

想了解更多关于AI放大器与其他人的区别，请访问kpmg.com/us/aiamplifiers。

Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers.

市面上每个AI编程工具做的第一件事都一样。

Every AI coding tool on the market does the same thing first.

那就是开始写代码。

It starts writing code.

而Blitzy恰恰相反。

Blitzy does the opposite.

在写一行代码之前，Blitzy会花几天时间逆向工程你的整个代码库。

Before writing a single line, Blitzy spends days reverse engineering your entire codebase.

数千个智能体处理数百万行代码，映射出每一个依赖关系、每一个未记录的约束，以及过去十年中做出的每一个架构决策。

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

结果就是一个动态的知识图谱，它能像一位在系统里干了三十年的首席工程师那样理解你的软件。

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after thirty years in the building.

其他工具靠grep搜索和markdown文件来猜测上下文。

Other tools guess at context with grep searches and markdown files.

Blitzy从不猜测。

Blitzy never guesses.

它先建立真正的理解，然后自主交付超过80%的完整软件史诗。

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

经过验证、端到端测试、生产级质量的拉取请求。

Validated, end to end tested, production grade pull requests.

这就是为什么财富500强的工程团队把最重要的代码库交给Blitzy。

That's why Fortune 500 engineering teams trust Blitzy with the code bases that matter most.

你自己去blitzy.com看看吧。

See for yourself at blitzy.com.

就是blitzy.com。

That's blitzy.com.

我经常看到一些公司在各种云、各种模型、各种框架之间摇摆不定，或者花钱请大型服务商做个没完没了的试点项目。

One thing I keep seeing in companies hedging across every cloud, every model, every framework or paying a GSI for a pilot that never ends.

真正在交付的团队，他们选定了方向，然后快速推进。

The team's actually shipping, they've picked a lane, and they move fast.

这也是我喜欢今天赞助商Robots and Pencils的原因之一。

That's one of the reasons I like today's sponsor Robots and Pencils.

他们全力投入了AWS。

They've gone all in on AWS.

他们是AWS的高级合作伙伴，能在四十五天内交付生产级的AI同事。

They're an advanced tier and AWS pattern partner and they ship production AI coworkers in forty five days.

这让他们在AI同事方面做出了一些我见过的最有意思的工作。

That's led to them doing some of the more interesting work I've seen on AI coworkers.

我说的不是聊天机器人，而是真正的智能体系统，它们嵌入企业架构中，做实际的工作。

And by that I'm not talking about chatbots, I'm talking about actual agentic systems that sit inside a business architecture and do real work.

如果你是企业领导者，想把真正的东西投入生产，或者你是AWS的销售代表，想推动客户从感兴趣到实际部署，这种专注很重要。

That kind of focus matters if you're an enterprise leader trying to get something real into production, or an AWS rep trying to move a customer from interested to deployed.

在robotsandpencils.com上申请AI简报。

Request an AI briefing at robotsandpencils.com.

和Robots and Pencils聊一次，你就明白了。

One conversation with Robots and Pencils and you'll know.

本期《AI每日简报》由Hyperagent赞助，你可以在这里运行你的团队能共同管理的智能体集群。

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

新用户可以获得一千美元的推理额度。

New users get a thousand dollars in inference.

别再依赖本地智能体和等待你笔记本电脑提示的聊天工作流了。

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

Hyperagent在云端部署始终在线的智能体，在你团队已经使用的工具中完成实际工作。

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

营销部门的智能体把竞争对手的动态变成落地页。

Marketing's agent turns competitor moves into landing pages.

销售部门的智能体丰富潜在客户信息、起草邮件并更新CRM。

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

运营代理负责处理文书工作和跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享上下文，并遵循你关于范围和审批的规定。

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

是时候让代理感觉像队友一样了。

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

在Hyperagent雇佣你的代理，它由Airtable团队打造。

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

在hyperagent.com/aideallybrief领取你的1000美元推理额度。

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

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

周五晚上，OpenAI宣布将终止与Cursor的合作关系。

On Friday night, OpenAI announced that they would be ending their relationship with Cursor.

这是一个影响深远的决定，虽然并不特别出人意料。

This is a highly consequential, if not necessarily particularly surprising decision.

一方面，今天我们要讨论这对前沿实验室之间AI竞争格局意味着什么，但也会深入我认为更重要的讨论，至少对我们所有人来说，从实际应用的角度看，那就是如何让我们自己和我们的公司为下一阶段竞争带来的必然趋势做好准备。

And on the one hand today, we are going to discuss what this means for the state of AI competition among the frontier labs, but we will also get into what is I think the more important discussion, at least for all of us in an applied sort of way, which is how to position ourselves and our companies for the inevitabilities that that next phase of competition bring.

虽然我们无法预知未来，但我认为有一些相当明显的模式，这些模式对我们如何思考企业AI战略有着相当重要的影响。

While we can't read the future, I think that there are some pretty clear patterns that have some fairly significant implications for how we think about enterprise AI strategy.

但首先，让我们谈谈OpenAI采取的这个行动。

But first, let's talk about the move that OpenAI made.

在周五晚些时候的公告中，OpenAI基本上是说，我们爱Cursor，但我们恨Elon。

In a late Friday announcement, OpenAI basically said we love Cursor, but we hate Elon.

所以，抱歉了，Cursor用户们。

So sorry, Cursor users.

你们不能再使用OpenAI的模型了。

You don't get to use OpenAI models anymore.

不过，他们在新闻稿中试图把这件事说得稍微好听一点。

Now they tried to frame it a little bit nicer in the press release.

他们写道，要与像SpaceX这样的大型合作伙伴合作，我们通常依赖定制合同来确保遵守我们的服务条款，并且这种集成提供了大规模的安全性。

They wrote, To work with a large partner like SpaceX we typically rely on custom contracts to ensure compliance with our terms of service and the integration provides for safety at scale.

马斯克收购推特后，现在推特属于SpaceX的一部分，公司违反了我们的合同条款，还有很多其他公司也一样。

After Musk acquired Twitter now part of SpaceX the company broke the terms of our contract alongside many others.

今年早些时候，马斯克在宣誓作证时承认，xAI（现在也是SpaceX的一部分）违反了OpenAI的服务条款。

Under oath earlier this year Musk admitted that xAI, now also part of SpaceX, had violated OpenAI's terms of service.

另一方面，关于Cursor，他们说：我们和Cursor合作了将近四年，非常尊重他们的团队、他们的产品，以及他们为开发者社区所打造的一切。

On the flip side when it comes to Cursor they said We've worked with Cursor for nearly four years and have enormous respect for their team, their product, and what they've built for the developer community.

我们知道，受这个决定影响最大的是那些在Cursor中依赖OpenAI模型的开发者。

We know that the people most affected by this decision are the developers who rely on OpenAI models in Cursor.

我们关心他们在这次过渡中的体验，并且我们准备竭尽全力支持他们。

We care about their experience in this transition and we're ready to go above and beyond to support them.

不过，这个截止日期实际上要到11月12日才生效，这么长的期限看起来是故意的，OpenAI说他们给出了合同规定的最长通知期。

Now the cutoff will not actually come until November 12, and that long deadline seems to be intentional, with OpenAI saying that they are giving the maximum notice provided by their contract.

现在，埃隆的回应可能不会让人意外。

Now Elon's response will likely come as no surprise.

他在X上回复了一个帖子，说“我根本不在乎”，并且用了他最喜欢的称呼，把山姆·奥特曼叫做“骗子奥特曼”。

He responded to a post on X saying I couldn't care less and using his favorite moniker of calling Sam Altman Scam Altman.

Cursor的首席执行官迈克尔·特鲁尔说，他们看到这封信感到很遗憾，正在看看能不能达成其他协议。

Cursor CEO Michael Trull said that they were sorry to see the note and were seeing if they couldn't come to some different agreement.

他还提到，OpenAI的模型大约占Cursor用户流量的5%。

Now he also included a note that OpenAI models serve about 5% of Cursor user traffic.

显然，他的意思是，即使他们无法达成协议，这也不会太糟糕。

Clearly with the implication that even if they weren't able to come to some agreement that this wouldn't be all that bad.

OpenAI的产品经理蒂博，不知为何觉得有必要澄清那5%的数据，他转发了迈克尔·特鲁尔的帖子并争辩说：代币数量并不能代表收入或创造的价值，而且OpenAI的模型在代币效率方面处于最前沿。

OpenAI product manager Thibault, for some reason felt the need to clarify on that 5%, reposting Michael Truell and arguing: Tokens are not a proxy for revenue nor value created and the OpenAI models are on the very frontier of token efficiency.

较小或较弱的模型需要更多的代币才能完成任务，因此会显著夸大流量份额。

Smaller or less strong models require many more tokens to achieve a task and therefore will inflate traffic share significantly.

换句话说，他的观点是，即使那5%的代币流量在事实上是准确的，它可能代表的是10%或15%或20%甚至更多的价值创造，因为人们使用OpenAI模型处理特定任务，而且这些模型完成任务的效率要高得多。

His point, in other words, is that even if that 5% of token traffic is factually accurate, it might represent more like 10 or 15 or 20 or even more percentage of the value created because of what particular tasks people are using OpenAI models for, and how much more efficiently they do them.

至于他为什么决定要让大家更清楚地意识到他们给Cursor用户带来的痛苦，这并不完全清楚，但Epic的创始人蒂姆·斯威尼很乐意插话说：也许不要坑那些使用Cursor的ChatGPT客户？

Now why he decided that he had to increase everyone's awareness of the pain that they were causing for Cursor users isn't exactly clear, but Epic founder Tim Sweeney was happy to jump in and say: maybe don't screw over chat gbt customers who use Cursor then?

地球上没有开发者希望你们在我们的电脑里打公司战争。

No developers on Earth want you guys waging corporate warfare inside our computers.

现在一个自然的问题是，Anthropic 会不会也跟着这么做？

Now a natural question might be, is Anthropic going to follow suit?

不过，联合创始人兼首席计算机官汤姆·布朗说绝对不会。

However, co founder and chief computer officer Tom Brown said absolutely not.

他在 X 上发帖说，自从 Sonnet 3.5 以来，Cursor 一直是 Anthropic 的信任伙伴。

He posted on X, Cursor has been a trusted partner of Anthropic since Sonnet 3.5.

我们会继续增加算力，支持 Cursor 中的云模型，并且很期待接下来与他们在 SpaceX 的合作。

We'll continue to increase compute to support cloud models in Cursor and are excited for what comes next with them at SpaceX.

谷歌的洛根·基尔帕特里克发了一条推文，说得含糊但也不算太含糊：如果你的对手正在犯错，别去打断他们。

Google's Logan Kilpatrick vague but not that vague tweeted, If your opponent is busy making a mistake, don't interrupt them.

当然，对社区里的很多人来说，这次犯错的确就是 OpenAI。

And certainly for many in the community, it IS OpenAI who's making a mistake here.

Dazeemon X 写道：最大的输家是 OpenAI。

Dazeemon X writes: Biggest loser here is OpenAI.

Cursor 已经有 Grok 4.7 在路上，还有 Composer 3，以及一堆开源模型。

Cursor already has Grok 4.7 on the way, Composer three, a bunch of open models.

OpenAI 刚才的做法会让所有合作伙伴开始考虑备选方案。

What OpenAI just did is going to make every partner start thinking about plan B.

Youssef Altuqi 写道：这种行为真的很差劲。

Youssef Altuqi writes: This behavior is genuinely so poor.

我大部分使用 OpenAI 模型都是在 Cursor 上。

Most of my usage of OpenAI models was on Cursor.

Cursor 不像 Codex，它有一个快速模式开关，对模型真的有效，能带来两倍的速度提升。

Cursor, unlike Codex, has a fast mode toggle for the models that actually works and causes a 2x speed increase.

Cursor IDE 的用户体验非常棒。

Cursor IDE has the most gorgeous UX.

OpenAI 却因为小心眼，把这种体验从付费用户那里剥夺了。

OpenAI is stripping this away from paying users due to nothing but pettiness.

没有任何正当理由阻止付费用户在他们选择的工具中使用他们选择的模型。

There is no valid reason to stop paying users from using the model of their choice in the harness of their choice.

更气人的是，他们还回复 Cursor 说：实际上，从我们的效率来看，我们贡献了超过 5%。

Then to add insult to injury, they reply to Cursor, Actually, when you look at our efficiency, we contributed more than 5%.

除了在客户伤口上撒盐，没别的意义。

No point other than rubbing salt in the wound of customers.

但对其他人来说，OpenAI 这么做显然是有一些原因的。

And yet for others, there are pretty clearly some reasons for OpenAI to make this move.

Benjamin Decracker 写道：SpaceXAI 收购 Cursor 的全部目的，就是收集人们用它写代码时的训练数据。

Benjamin Decracker writes: The entire reason SpaceXAI bought Cursor is to harvest training data from people using it for coding.

这很明显，而且收购宣布时，这个优势还被公开吹嘘成 xAI 的一大胜利。

This is very obvious, and this advantage was openly touted as a huge win for xAI when the acquisition was announced.

OpenAI 不让自己的高级模型被卷进那个训练数据收集器，不也挺合理的吗？

Isn't it somewhat reasonable for OpenAI to opt out their advanced models being included in that training data harvester?

Pragmatic Engineering 的 Googlia Rose 写道：前沿模型公司不会为其他前沿模型公司打造的工具提供自家模型的直接集成。

Pragmatic Engineering's Googlia Rose writes, Frontier model companies don't offer direct integration of their models for tools built by other frontier model companies.

比如，你不能直接在 Claude Code 里用 GPT-56，也不能在 Codex 里直接用 Opus 5。

EG, you can't use GPT-fifty six from Claude Code or Opus five from Codex out of the box.

Cursor 现在是 SpaceX 的了，所以 OpenAI 撤掉 GPT，也没啥好惊讶的。

Cursor is now SpaceX, so OpenAI pulling GPT, not all that surprising.

盖尔·维纳写道：OpenAI 应该怎么做？

Gail Wiener writes, What was OpenAI supposed to do?

这不只是竞争对手收购了Cursor，而是这个竞争对手曾把他们告上法庭，试图摧毁他们。

This isn't just a competitor bought Cursor, but it's a competitor who took them to court to try to destroy them.

更重要的是，很多人指出，这并非孤立事件，而是前沿实验室现在的常态。

And much more significantly of all, many pointed out that this is not some isolated incident, but this is just the norm now of how Frontier Labs behave.

一个早期的例子发生在2025年6月。

An early example of this came back in June 2025.

在彭博社报道OpenAI即将达成收购Windsurf的交易后，Anthropic切断了Windsurf对其模型的访问。

After Bloomberg reported that OpenAI was close to nearing a deal to acquire Windsurf, Anthropic cut off Windsurf's access to their models.

2025年6月3日，Windsurf的瓦伦·莫汉写道：在不到五天的通知期内，Anthropic决定切断我们几乎所有的第一方能力，无法使用Claude 3.x模型。

On 06/03/2025, Windsurf's Varen Mohan wrote, with less than five days of notice, Anthropic decided to cut off nearly all of our first party capacity to Claude three point x models.

当汤姆·布朗在评论中提到Cursor自Sonnet 3.5以来一直是Anthropic的信任伙伴时，很多人向他提出了这一点。

This was a point that many brought up with Tom Brown in the comments when he posted about Cursor being a trusted partner of Anthropic since Sonnet 3.5.

Replit CEO阿姆贾德·马萨德写道：也许你已经改变了做法，但我们都记得你对Windsurf做的事，那要恶劣得多。

Replit CEO, Amjad Masad, wrote, maybe you've changed your ways, but we all remember what you did to Windsurf, which was infinitely nastier.

阿姆贾德接着指出，Anthropic在这里的做法之所以不同，很可能是因为他们现在与SpaceX AI有计算合作关系，这对他们未来模型的训练至关重要。

Amjad continued pointing out that it is likely that part of Anthropic s difference of approach here has to do with the fact that they now have a compute relationship with SpaceX AI that is integral to the training of their future models.

AI研究员梅胡尔·莫汉写道：我不明白OpenAI因为封禁Cursor而遭到这么多批评，而Anthropic怎么反而成了好人？

AI researcher Mehul Mohan writes: I don get the hate OpenAI is getting for banning Cursor and how Anthropic is somehow trying to be the good guy here?

Anthropic一年前对Windsurf做了完全一样的事，而且如果Anthropic没有每月向SpaceX AI支付十亿美元的计算费用，他们百分之百也会对SpaceX AI做同样的事。

Anthropic literally did the same thing with Windsurf one year ago, and would 100% have done the same with SpaceX AI if Anthropic was not paying a billion dollars per month for compute to them.

这有什么好争议的？

Why is this controversial?

而且Windsurf那件事并不是孤立的。

And the Windsurf thing wasn't an isolated incident.

2025年8月，Anthropic封锁了OpenAI对API的访问。他们声称违反了服务条款，但大多数人认为这是因为OpenAI使用Anthropic的模型来构建竞争性的AI模型。

In August 2025, Anthropic blocked OpenAI's access to the API, They claimed a violation of terms of service which most believed was about using Anthropic's models to build a competing AI model.

OpenAI的说法是他们只是在测试模型，但事后看来，整个举动显然是出于对蒸馏的担忧。

OpenAI's story was that they were just benchmarking the models, but clearly in retrospect the whole move was about concerns about distillation.

然后在今年一月，Anthropic也封锁了xAI。

Then in January of this year, Anthropic also blocked xAI.

在一月份的Slack消息中，xAI的联合创始人Tony Wu说：大家好。

In a Slack message in January, xAI co founder Tony Wu said, Hi team.

我相信你们很多人已经发现，Anthropic的模型在Cursor上不响应了。

I believe many of you have already discovered that Anthropic models are not responding on Cursor.

根据Cursor的说法，这是Anthropic对所有主要竞争对手实施的新政策。

According to Cursor, this is a new policy Anthropic is enforcing for all its major competitors.

这既是坏消息也是好消息。

This is both bad and good news.

我们的生产力会受到影响，但这真的会推动我们开发自己的编码模型和产品。

We will get a hit on productivity, but it really pushes us to develop our own coding models and products.

我们正处于AI成为我们自身生产力关键技术的时代。

We're at a time in which AI is now a critical technology for our own productivity.

团队正在快速开发我们自己的模型和产品。

The team is rapidly developing our own models and product.

我们很快就会有东西和大家分享。

We will have something to share with everyone soon.

与此同时，你仍然可以在Grok Build里试试不同的模型。

In the meantime, you may still try different kinds of models in Grok Build.

然后，当然，在接下来的几个月里，Anthropic改变了他们的订阅方式，不再覆盖第三方工具，特别是像OpenClaw和Hermes这样的东西。

Then, of course, over the next couple of months, Anthropic changed the way their subscriptions work to not cover third party tools, specifically things like OpenClaw and Hermes.

Hermes的联合创始人Technium也很乐意把这个指给Tom Brown看。

Hermes co founder Technium was happy to point this one out to Tom Brown as well.

几个月后，Anthropic公开与Figma合作，然后挖走了一位高管，并推出了Claw Design来直接竞争，这导致了《华尔街日报》七月的一篇报道，担心Anthropic会利用用户使用他们模型时获得的专有信息，来构建与客户竞争的服务。

A couple months later, Anthropic partnered with Figma publicly and then poached an executive and launched Claw Design to compete directly, which led to a July story in The Wall Street Journal about a concern that Anthropic was going to use proprietary information that they got from people using their models to build competing services to their customers.

虽然这些例子都是关于Anthropic的，但很多人都在敲响警钟，说这就是下一阶段竞争的方式。

And while these examples have all been Anthropic, many have been beating the drum that this is just the way that it's going to work in the next phase of competition.

Palantir的Alex Karp一直在向任何愿意听的人和任何新闻媒体大声疾呼这件事。

Palantir's Alex Karp has been screeching about this to anyone and any news outlet that will listen.

而前沿实验室与其客户之间存在根本性错位的假设，现在看来至少是微软战略的核心。

And the presumption that there is a fundamental misalignment between the Frontier Labs and their customers now seems to be core to no less than Microsoft's strategy.

几个月前，微软CEO萨蒂亚·纳德拉发布了一篇名为《反向信息悖论》的博客文章。

A couple of months ago, Microsoft CEO Satya Nadella released a blog post called the Reverse Information Paradox.

在那篇文章里，他写道，你其实是为智能付了两次钱，一次是金钱，另一次是更宝贵的东西。

In it, he wrote, you essentially pay for intelligence twice, once with money, and again with something even more valuable.

那就是为了让智能发挥作用，你必须透露的专有知识。

The proprietary knowledge you must reveal to make that intelligence useful.

你越希望模型表现得好，就得喂给它越多的这类知识。

The better you want the model to perform, the more of that knowledge you have to feed it.

随着时间推移，信息不对称会变得越来越严重。

Over time, the information asymmetry becomes increasingly skewed.

卖家在你使用所购产品时，对你了解得越来越多，而你却几乎不知道卖家在反过来了解你什么。

The seller learns more and more about you as you use what you purchased, while you learn very little about what the seller is learning in return.

这就是我所说的“反向信息悖论”。

That is what I think of as the reverse information paradox.

这需要的不仅仅是数据保护。

This requires more than data protection.

模型从各种痕迹中学习，比如人们写的提示词、智能体使用的工具，尤其是当模型出错时人们所做的纠正。

Models learn from exhaust, the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong.

每一次修正都被提炼成机构的知识。

Every correction is distilled into institutional know how.

这种知识是竞争对手永远买不到的，而且几乎察觉不到地泄露出去。

It's the kind of knowledge a competitor could never buy, and the kind that leaks almost imperceptibly.

一点一滴，一次修正，一次评估。

Trace by trace, correction by correction, eval by eval.

我们必须把学习基础设施分发给每家公司，让他们能掌控自己的学习循环。

It's imperative that we distribute the learning infrastructure to every firm so that they can control their own learning loop.

如果你一直关注的话，这已经为微软后续的所有动作定下了基调。

Now if you've been paying attention, this has set the tone for every move that Microsoft has made subsequently.

他们最近发布的新模型，很大程度上就是为了定制化和后期训练而设计的，虽然它们肯定不是开放权重，但设计上更可定制，更归客户所有，而不是仅仅把信息发送到一个只有前沿实验室才能访问的黑箱里。

Their new models that they recently released are very much designed to be the base for customizations and post training, which while they are certainly not open weights, are designed to be more customizable and more owned by the customer as opposed to just sending off information into a black box that only the Frontier lab can access.

尽管OpenAI试图说这一举动不是针对所有用户，而是针对埃隆的具体问题，基于已展示的行为模式，但对很多人来说，含义很明确。

And as much as OpenAI tries to say that this is a move that is not general but is specific to issues with Elon based on a demonstrated pattern, the implication for many is clear.

正如金宇辰所说，当埃隆收购Cursor时，OpenAI就切断了Cursor。

As Yuchen Jin put it, When Elon acquires Cursor, OpenAI cuts off Cursor.

当OpenAI试图收购Windsurf时，Anthropic就切断了Windsurf。

When OpenAI tried to acquire Windsurf, Anthropic cuts off Windsurf.

不是你的权重，就不是你的产品。

Not your weights, not your product.

实际上对企业来说，这两件事有没有实质区别并不重要。

Now functionally for enterprises, it doesn't matter if there is substantive difference in these two things.

他们必须做出的反应是非常相似的。

The response that they're going to have to have is very similar.

AI内容创作者Theo说：如果你想避免被公司之间的争斗伤害，确保你拥有自己的工具，并且你与所依赖产品的关系是直接的，而不是通过其他层来中转。

As AI content creator Theo put it: If you want to avoid getting hurt by companies beefing with each other, make sure you own your tools, and your relations with the products you rely on are direct and not routed through other layers like this.

我觉得这不是一次性的事情。

I have a feeling this is not a one off thing.

如果要说的话，这可能是终结的开始。

If anything, it's the start of the end.

我们可能会看到Anthropic和OpenAI，甚至像Google和SpaceX这样的公司，采取越来越多这样的行动。

We're probably going to see more and more moves like this from Anthropic and OpenAI, and probably even companies like Google and SpaceX.

而且，当然，在这之前，我们已经看到一些公司开始更熟悉开放权重模型了。

And of course, already, even before this, we had seen companies starting to get more acquainted with open weight models.

早在八月份，《华尔街日报》报道了AT&T如何开始使用开放模型，而Business Insider也谈到了汤森路透如何基于阿里巴巴的千问模型做了类似的事情。

Earlier in August, The Wall Street Journal profiled how AT and T had started working with open models, and Business Insider also talked about how Thomson Reuters had done something similar building off of a base of Alibaba's Quen.

这些报道大多把这当作一种成本控制措施。

Now these stories mostly posited this as a cost control measure.

显然，过去几个月的一个主要话题就是公司如何构建更复杂的模型架构，以便更经济地匹配任务难度和模型能力。

Obviously, a big part of the narrative for the last several months has been how companies build more complex model architectures that allow them to match task difficulty with model capability in a way that is more cost effective.

不过，现在这些举措显然也涉及主权和控制的问题，这表明这些问题正从单纯的模型层扩展到所谓的“驾驭层”。

However, there is clearly now a sovereignty and control aspect of those moves, and what this shows is that those questions are moving from strictly the model layer to also the harness layer as well.

再次提到《华尔街日报》的CIO Journal，最近的一篇文章向更广泛的读者介绍了AI模型“驾驭”的概念。

Again, in The Wall Street Journal's CIO Journal, a recent post introduced the idea of an AI model harness to a wider audience.

在解释为什么企业需要模型驾驭时，穆迪的David Pan称这是一种让公司重新掌控自己AI的方式。

In explaining why businesses need a model harness, Moody's David Pan called it a way for companies to take back control of their AI.

该报写道：围绕AI模型开发自己的软件，Pan称之为“驾驭工程”，这能让企业将工作流程与模型本身解耦，从而减少对单一AI提供商的依赖。

The journal writes: Developing their own software around AI models, a practice Pan calls harness engineering, gives businesses a way to decouple their workflows from the models themselves, and that helps them become less reliant on a single AI provider.

如果你把那个控制框架带进公司内部并掌控它，你就在很大程度上增强了业务的韧性。

If you bring that harness in house and control it, you're baking in a lot more business resilience.

而且考虑到这篇文章是一周前发布的，潘建议企业自己构建控制框架，而不是依赖OpenAI和Anthropic等实验室提供的框架，这似乎特别有先见之明。

And given that this article came out a week ago, PAN's advice that companies build their own harnesses rather than rely on those offered by labs like OpenAI and Anthropic seems particularly prescient.

那么，如果你是企业AI的买家，你该怎么做呢？

So if you are an enterprise AI buyer, what is the move here?

首先，这再次提醒你，如果你还没有关于开放权重模型的政策，你需要去搞清楚。

First of all, this is another reminder that if you don't have a policy around open weights models yet, you need to go figure it out.

需要明确的是，我并没有看到企业完全放弃前沿的封闭模型。

Now to be clear, what I do not see is companies abandoning frontier closed models entirely.

坦白说，即使是经验丰富的用户，也并不是大多数用例都已经转向那些新模型。

Frankly, even among sophisticated users, it's not like some big majority of use cases have even moved over to those new models yet.

但经验丰富的企业AI用户明白，能够将某些类型的任务整合并路由到更轻量、更便宜、更可控的开放权重模型上，将是他们必须具备的关键能力，而且像任何能力一样，这需要时间，他们需要现在就开始。

But what the sophisticated enterprise AI users understand is that being able to integrate and route certain types of tasks to lighter, cheaper, more controllable, open weights type models is going to be a key capability that they need to have, and that like any capability, it's going to take time and they need to start now.

我认为这个Cursor的举动所强调的一点是，这不仅仅是关于模型的讨论，也是关于控制框架的讨论。

What I think that this cursor move puts a point on is that this is not just a model conversation, but also a harness conversation as well.

我非常肯定地预测，你会看到更多关于开放模型和开放框架的讨论。

My very strong prediction is that you're going to see a lot more discourse around not just open models but open harnesses.

一个早期的例子是，本月初DeepSeek发布了DeepSeek Harness。

An early example of this is that once again earlier this month, DeepSeek released DeepSeek Harness.

在他们的公告中写道：DeepSea Carnus是一个围绕一个核心思想构建的代理框架。

In their announcement post they wrote: DeepSea Carnus is an agent harness built around one core idea.

一切都是插件。

Everything is a plug in.

模型、工具、技能、会话、沙箱、文件系统、循环、编排和用户界面都作为插件实现，可以混合、匹配、替换和扩展。

Models, tools, skills, sessions, sandboxes, file systems, loops, orchestration, and UI are all implemented as plug ins and can be mixed, matched, replaced, and extended.

对DeepSea Carness的第一印象相当不错。

And first impressions of DeepSea Carness are pretty good.

但我认为DeepSea Carness本身并不重要，重要的是开放框架现在也成为企业可以使用的工具。

But I think that the DeepSea Carness itself matters less than the fact that open harnesses are now a tool that enterprises are going to have access to as well.

最终，我认为Cursor的举动并不特别令人惊讶，这当然不意味着Cursor用户不应该对OpenAI大喊大叫，看看他们能否改变立场。

Ultimately, I don't think anything about the Cursor move is particularly surprising, Which certainly doesn't mean that Cursor users shouldn't yell and scream at OpenAI and see if they can't change their position.

我认为，一个高度碎片化的模型和工具链世界，本质上对每个人来说都比我们现在正在进入的那个世界更糟糕，所以我完全支持人们利用市场压力来试图改变这一政策。

I think a highly balkanized world of models and harnesses is inherently worse for everyone than the one we seem to be going into, so I am fully in support of people using market pressure to try to change the policy.

但对企业来说，教训很明确：如果你不想受制于那些控制你模型和工具链的公司的反复无常，你基本上就不能依赖任何一家公司来控制你的模型或工具链。

But for enterprises, the lesson is clear: if you want to not be subject to the whims of the companies that control your models and control your harnesses, you basically can't be reliant on any one company to control your models or to control your harnesses.

所以，你知道，对企业来说，要让AI为你所用，你还需要掌握一整套新的技能。

So you know, for enterprises, just a whole additional set of things that you have to get good at to make AI work for you.

总之，有趣的时刻。

Anyways, interesting times.

这是我们将会密切关注的一个趋势。

This is a trend that we will watch closely.

今天就到这里，这是今天的AI每日简报。

For now, that is gonna do it for today's AI Daily Brief.

一如既往，感谢你的收听或观看。

Appreciate you listening or watching as always.

下次再见。

Until next time.

再见。

Peace.