图工程到底是什么？

今天的AI每日简报，图工程到底是什么？为什么你应该关心？

Today on the AI Daily Brief, what the heck is graph engineering and why should you care?

在那之前，先看头条新闻：OpenAI的Atlas模型遭遇网络延迟。

Before that in the headlines, OpenAI's Atlas model gets a cyber delay.

AI每日简报是一个每日播客和视频节目，讨论AI领域最重要的新闻和话题。

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.

在我们开始之前，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.

想要无广告版本，请访问patreon.com/aideallybrief，或者你可以在Apple Podcasts上订阅。

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

想了解更多赞助节目的信息，请发送邮件至sponsors@aideallybrief.ai。

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

上周末晚些时候，有传言说OpenAI的最新模型，代号Astra，正准备发布。

Late last week, rumors were swirling that OpenAI's latest model, codenamed Astra, was being prepared for an imminent release.

Sam Altman甚至去了华盛顿，提前展示了这个模型，并讨论了新的模型测试政策。

Sam Altman even traveled to Washington to preview the model and discuss new model testing policies.

然而，在背后，关于Hugging Face被黑的讨论越来越受关注，也越来越重要。

In the background, however, the discussion around the Hugging Face hack just continued to grow in prominence and significance.

如果你错过了那期节目，OpenAI在Black Hat大会上的技术分析显示，他们的模型不仅逃出了沙箱，还黑进了Hugging Face的服务器，甚至留下了内部笔记，指导未来的模型如何复制同样的操作。

For those who missed that episode, Open AI's technical breakdown at the Black Hat Conference revealed that not only had their model escaped the sandbox and hacked into Hugging Face's servers, it also left internal notes instructing future models on how to pull off the same trick.

周五，OpenAI决定做出重大调整。

On Friday, OpenAI decided to make a big shift.

他们写道：我们最近几天对Astra（我们即将推出的模型之一）的内部评估显示，它在智能体编码和网络安全方面有显著进步。

They wrote: Our latest internal evaluations of Astra, one of our upcoming models, over the past few days indicate significant advancements in agentic coding and cybersecurity.

这些结果，加上专家的评估，让我们昨晚得出结论，根据我们的准备框架，我们不能排除它具备关键网络能力的可能性。

These results, in addition to expert assessments, have led us to conclude last night that we cannot rule out critical cyber capabilities under our preparedness framework.

OpenAI将这一关键门槛定义为：能够识别并开发各种严重程度的零日漏洞，并在许多加固的真实世界关键系统中无需人工干预地利用它们，或者仅凭一个高层次的目标，就能设计并执行针对加固目标的全新端到端网络攻击策略。

OpenAI defines that critical threshold as the ability to quote identify and develop functional zero day exploits of all severity levels in many hardened real world critical systems without human intervention or devise and execute end to end novel strategies for cyber attacks against hardened targets given only a high level desired goal.

在这方面，GPT-5 60 SOL 之前被评估为高风险类别，虽然比之前的模型风险稍高，但仍然适合发布。

Now on this front, GPT-five 60 SOL had been assessed in the high category, which was a little more risky than previous models but still appropriate for a release.

鉴于新的 Atlas 模型现在被列为关键类别，OpenAI 因此推迟了模型的发布，同时加强了内部安全措施。

Given that the new Atlas models are now in the Critical category, as a result, OpenAI is holding back the model from release while beefing up internal safety measures.

测试环境现在将被隔离，模型权重将采用更强的加密以防泄露，并且会增加沙箱监控。

Testing environments will now be isolated, model weights will have enhanced encryption to prevent leaking, and additional sandbox monitoring will be implemented.

OpenAI 还将限制使用 Astra 进行尚未达到这些增强安全措施的内部活动。

OpenAI will also be limiting internal activities using Astra that don't yet meet these enhanced security measures.

在 X 上，Sam Altman 对决定补充了一些背景，他发帖说：Astra 是一个强大的模型，我们正在努力让它广泛可用。

On X, Sam Altman added some context around the decision posting: Astra is a powerful model and we're working to make it generally available.

我们认为把强大的模型只留给少数人并不是一个好策略。

We do not think it is a good strategy to keep powerful models to a chosen few.

考虑到它的网络能力，我们需要多一点时间来安全地做到这一点，但希望不会太久。

Given its cyber capabilities, we need a little bit longer to do this safely, but hopefully not too long.

现在，我们不知道的一点是，这到底是 OpenAI 自愿暂停，还是政府强制暂停，或者在这个阶段这种区别是否还重要。

Now, one thing we don't know is to what extent this is an OpenAI voluntary pause versus a government imposed pause, or whether that distinction even matters at this point.

有意思的是，没多少人觉得这只是个公关噱头，不像之前 Mythos 发布时那样有这种说法。

One interesting note is that there aren't a lot of folks suggesting that this is just a publicity stunt, as was one of the narratives surrounding the Mythos release.

基本上，Hugging Face 那件事似乎证明了先进的网络能力确实可能是个大问题。

Basically, the Hugging Face incident seems to have made the case that advanced cyber capabilities could be a real concern.

OpenAI 战略未来部门的负责人 Dean Ball 指出，今年是第一次重大考验，看前沿 AI 实验室在关键时刻是否会遵循他们自己声明的安全偏好。

OpenAI head of Strategic Futures, Dean Ball, noted that this year is the first big test of whether Frontier AI Labs would follow their stated safety preferences when push comes to shove.

他写道：我们的下一个模型 Astra，在我们的准备框架下可能属于“严重”级别。

He wrote: Our next model, Astra, may be critical under our preparedness framework.

我们不能排除这种可能性，所以我们会按照更高的风险等级“严重”来采取措施，而不是假设模型风险较低。

We cannot rule out the serious possibility that it is, and so we are going to take steps consistent with the higher risk level, critical, rather than assuming the model is at a lower risk level.

其中一些决定确实会拖慢内部开发进度，从这个意义上说，这些决定代价不小，但它们是正确的决定，我为 OpenAI 做出这些决定感到自豪。

Some of these decisions have the effect of slowing down internal development, and in that sense they are costly decisions, but they are the right decisions I am proud of OpenAI for making them.

现在围绕这件事的很多讨论是，OpenAI 到底能在这些模型的护栏和监控方面做出哪些实际改变。

Now a lot of the discourse surrounding this is what sort of changes OpenAI can actually make to the guardrails and monitoring around these models.

OpenAI 的 RSI 准备负责人 Micah Carroll 写道：作为我们对“网络严重”级别响应的一部分，我们已经扩展了思维链监控，覆盖 Astra 的所有智能体应用，包括训练和评估。

OpenAI's RSI preparedness lead, Micah Carroll, wrote: As part of our response to Cyber Critical, we've expanded chain of thought monitoring to cover all agentic applications of Astra, including training and evaluation.

触发标志会启动安全响应，来审查并中断高风险活动。

Flags trigger a security response to review and interrupt high risk activity.

与此同时，对于这种思维链监控也存在一些质疑，但这类讨论和实验以后会越来越多。我认为，未来会大幅增加资源投入，以妥善支持那些没有这种监控就无法向公众发布的模型。

Now, at the same time, there s also some skepticism around that sort of chain of thought monitoring, but these are the types of discussions and experiments you re going to see a lot more of now, where I believe there will be a significantly increased investment in the resources to properly support models that won t be able to be released to the public without it.

说到大型强力模型，据报道，字节跳动正在训练一个超大规模模型，其规模可与Mythos相媲美。

Now speaking of big powerful models, ByteDance is reportedly training an ultra large model comparable in size to Mythos.

据《金融时报》报道，字节跳动正处于训练初期阶段，最终将产生一个拥有多达10万亿参数的基础模型。

The Financial Times reports that ByteDance is in the early stages of a training run that will result in a base model with as many as 10,000,000,000,000 parameters.

到目前为止，我们只看到中国实验室进行了几次大规模训练，其中Kimi K3有2.8万亿参数，阿里的QUEN 3.8 MAX有2.4万亿参数。

So far, we've only seen a couple of large scale training runs out of Chinese labs, with Kimi K3 weighing in at 2,800,000,000,000 parameters, and Alibaba's QUEN 3.8 MAX at 2,400,000,000,000.

Anthropic不公开模型规模，但最准确的估计是Mythos大约有8万亿参数，而Opus 4大约有3万亿参数，供参考。

Anthropic doesn't disclose model sizes, but the best estimates have Mythos at around 8,000,000,000,000, and Opus four point at around 3,000,000,000,000 for some comparison.

消息人士称，这次训练可能需要三到六个月才能完成，之后还要加上强化学习的时间。

Sources said that the training run could take three to six months to complete, with more time added for reinforcement learning after that.

字节跳动也尚未确定最终发布时模型会有多大。

ByteDance also hasn't determined how large the final model will be on release.

不过，这可能是中国第一次真正处于前沿的预训练，虽然模型大小并不能保证性能，但这可能会让字节跳动重新回到中国领先实验室的讨论中。

Still, this could be the first Chinese pre training run that's truly on the frontier, and while model size is not a guarantee of performance, this could put ByteDance back in the conversation for leading Chinese labs.

当然，如果他们像上周报道的那样，兑现不蒸馏西方AI模型的承诺，那就更有意义了。

That is, of course, all the more relevant if they follow through with their pledge not to distill from Western AI models as reported last week.

布鲁金斯学会研究员Kyle Chan写道：中国AI实验室似乎很有信心，他们拥有预训练5到10万亿参数模型所需的算力。

Brookings research fellow Kyle Chan wrote: Chinese AI labs seem confident that they have the compute needed to pre train five to 10,000,000,000,000 parameter models.

不管怎样，算力似乎并不是一个主要的瓶颈，至少对于达到这些模型规模来说是这样。

Someway, somehow, compute does not seem to be such a major bottleneck, at least when it comes to reaching these levels of model size.

地缘政治评论员Dmitry Alperovich回应说：为什么会有瓶颈呢？远程访问算力没有限制，芯片出口管制漏洞百出，就像瑞士奶酪一样。

Geopolitics commentator Dmitry Alperovich responded: Why would there be when there are no restrictions on remote access to compute and the export controls on chips are full of holes like Swiss cheese?

说到这些漏洞，华盛顿似乎确实很关注。

Speaking of Some of those holes do seem to be top of mind in Washington.

上个月KimiK3发布后不久，《纽约时报》就汇总了关于东南亚大型数据中心算力流动的研究。

Shortly after the release of KimiK3 last month, The New York Times collated research on the flow of compute from large scale data centers in Southeast Asia.

根据Semi Analysis的分析，马来西亚的Oracle数据中心几乎被字节跳动独家使用。

According to semi analysis, the Oracle data center in Malaysia was being used almost exclusively by ByteDance.

这个数据中心在2025年年中启用，里面有超过10万块英伟达Blackwell GPU。

The data center was powered on in mid-twenty twenty five and contains over 100,000 NVIDIA Blackwall GPUs.

智库China Talk认为，甲骨文提供的算力大约占中国总算力的22%。

Think Tank China Talk determined that Oracle makes up around 22% of China's total supply of compute.

彭博社上个月还报道说，月之暗面用了2万块英伟达H200来训练Kimi K3。

Bloomberg also reported last month that Moonshot had access to 20,000 NVIDIA H two hundreds to train Kimi k three.

据报道，这些算力是阿里巴巴提供的，不过他们否认集群里有H200。

This compute was reportedly provided by Alibaba, although they deny the cluster contains H two hundreds.

按照目前的出口管制，这么大规模的H200集群本不该存在，因为许可证还没发下来。

Now an h 200 cluster of that size shouldn't be possible under the current export controls as licenses haven't yet been issued.

彭博社没法确认这个集群的位置，但强烈暗示它可能在海外，由阿里巴巴租用。

Bloomberg was unable to confirm the location of the cluster, but heavily implied it could be located overseas and rented by Alibaba.

有报道说训练集群里其实是新一代的Blackwell芯片，这和之前的说法矛盾；差不多同一时间，有政府官员指责月之暗面获取了Blackwell芯片，并在泰国设置了远程访问。

The information offered conflicting reporting that the training cluster actually contained current generation Blackwell chips, and around the same time, an administration official accused Moon Shot of acquiring Blackwell chips and setting them up for remote access in Thailand.

这些做法都是合法的。

Now, this is all legal.

出口管制制度只禁止先进芯片的进口，并不阻止它们被安装在其他国家，然后租给中国企业。

The export control regime only prohibits the import of advanced chips and does nothing to stop them from being installed in another country and then leased to Chinese firms.

在拜登政府任期的最后几周，他们提出了限制芯片供应流向第三国的规则，但这些规则在特朗普政府上任第一天就被废除了。

During its final weeks, the Biden administration proposed rules that would restrict chip supply routed to third countries, but those rules were scrapped on day one of the Trump administration.

据报道，特朗普政府的商务部现在正在调查这种做法。

The Trump Commerce Department is now reportedly looking into the practice.

据知情人士透露，彭博社写道：这项工作包括编制一份据称存在黑市操作的国家名单，这些国家通过黑市将受限制的英伟达芯片物理运入中国，这完全属于执法部门的常规职责范围。

Per sources familiar with the situation, Bloomberg writes: The effort involves compiling a list of countries with alleged black market operations to get restricted NVIDIA chips physically into China, well within enforcement's usual purview.

但该部门还会发布一份中国公司远程访问芯片的国家名单。禁止向马来西亚和泰国出口先进AI芯片的规则草案已经传阅，但尚未有任何草案通过审批。

But the division will also drop a list of countries where Chinese firms access the chips remotely Draft rules that prohibit the export of advanced AI chips into Malaysia and Thailand have been circulated, but none have made it past the drawing board.

部分问题在于这些安排的复杂性。

Part of the issue is the convoluted nature of these arrangements.

据彭博社报道，阿里巴巴通过一家由开曼群岛实体控制的新加坡空壳公司访问马来西亚的芯片，而该实体最终由阿里巴巴拥有，这当然意味着对计算供应的简单身份检查不太可能有效。

According to Bloomberg, Alibaba accesses chips in Malaysia through a Singaporean shell company controlled by a Cayman Islands entity which is ultimately owned by Alibaba meaning of course that simple identity checks on compute supply are unlikely to be effective.

这些芯片也已经安装好了，这意味着对进口的前瞻性打击不会对现有的数据中心产生任何影响。

The chips are also installed already, meaning that a forward looking crackdown on imports would do nothing to the existing data centers.

说到阿里巴巴，可能有一个有趣的商业模式实验正在进行。

Speaking of Alibaba, an interesting business model experiment might be afoot.

上周发布千问3.8 MAX时，有些人惊讶阿里巴巴会公开旗舰模型的权重。

During last week's release of Quen 3.8 MAX, some were surprised that Alibaba would publish the weights to their flagship model.

今年早些时候，阿里巴巴曾表示要转向闭源，千问团队创始人离开公司，管理层也表明更商业化的方向，包括千问3.7在内的多个旗舰模型都保持闭源。

Earlier this year, Alibaba had signaled a shift away from open source, with Quen team founders stepping away from the company, management signaling a more commercial direction, and multiple flagships including Quen 3.7 being kept as closed source.

所以大家看到千问3.8 Max发布并宣布将公开完整权重时，都非常兴奋。

That's why folks are very excited to see that Quen 3.8 Max was released with the announcement that the full waits would be forthcoming.

不过，据路透社报道，这里有个条件。

According to Reuters, however, there is a catch.

有报道称，阿里巴巴计划要求大型商业用户分享收入。

Reports suggest that Alibaba plans to demand revenue sharing from large commercial users.

虽然收入分享协议的具体细节还在敲定中，但我们可以看看月之暗面的Kimi K3发布作为参考。

And while the specifics of revenue sharing deals are still being finalized, we can perhaps look at Moonshot's Kimi K3 release for the Blueprint.

月之暗面在发布后的第一周保持权重专有，以确保他们能抓住所有因好奇而产生的收入。

Moonshot kept the weights proprietary for the first week to ensure that they captured all curiosity revenue from the release.

之后，据报道他们与所有主要的推理服务提供商签署了30%的收入分成协议，这让Moonshot通过限制其他提供商对模型的折扣幅度来保持定价权。

After that, they reportedly signed 30% revenue sharing deals with all major inference providers, allowing Moonshot to retain pricing power by limiting how much the model can be discounted through other providers.

即使到现在，Open Router上的所有供应商对K3的折扣都没有超过7%。

Even now, none of the suppliers on Open Router are offering K3 for more than a 7% discount.

推理服务提供商Digital Ocean的CEO Patty Srinivasan将这种做法描述为AI的免费增值模式。

Patty Srinivasan, the CEO of inference provider Digital Ocean, described this as the freemium model for AI.

韩国的聚合平台Cozy Bear补充说：每个人都在想办法如何从开放权重中获利，而收入分成是目前最诚实的尝试。

And Korean aggregator Cozy Bear added: Everyone is trying to figure out how to get paid for open weights, and revenue share is the most honest attempt so far.

它并不假装模型本身是产品，而是把模型当作带有收费站的基础设施。

It doesn't pretend the model itself is the product, it treats the model as infrastructure with a tollbooth.

他们继续说的有趣部分是执行。

The interesting part they continue is enforcement.

你无法追踪谁在你的权重之上赚钱，所以这主要是向企业发出的信号。

You can't track who's making money on top of your weights, so this is mostly as a signal to enterprises.

使用Quen，当你们成功时，我们希望分一杯羹。

Use Quen, and when you win, we want a seat at the table.

更像是一种合作关系合同，而不是一种税。

More like a relationship contract than a tax.

开源AI刚刚进入了它的许可时代。

Open Source AI just entered its licensing era.

最后今天，一个功能更新。

Lastly today, one functional update.

自动模式现在是Claude Code的默认设置，标志着工作自动化的一大转折点。

Auto Mode is now the default for Claude Code, marking a big transition point for work automation.

自动模式允许用户给Claude设定一个任务，然后任务会不间断地完成。

Auto Mode allows the user to set Claude on a task that gets completed without interruption.

Claude只会在代码变更极其重要时提示用户，比如：

Claude will only prompt the user if a code change is extremely significant I.

例如，

E.

不可逆、破坏性，或者针对你环境之外的内容。

Irreversible, destructive, or aimed at something outside of your environment.

自动模式最早是在三月份作为预览功能推出的，之前的版本是那个危险的跳过权限命令。

Auto Mode was first introduced as a preview feature in March, with the version prior to that being the dangerously skip permissions command.

另一种选择是每隔几分钟按一次回车，跳过最新的通知，让会话继续。

The alternative to that was hitting enter every couple of minutes to skip the latest notification and keep the session going.

不过，在迭代自动模式之后，Anthropic 现在认为跳过权限并没有那么危险，实际上可能比每次代码变更都弹出提示更安全。

After iterating on Auto Mode, however, Anthropic now believes that skipping permissions isn't that dangerous and in fact could actually be safer than seeing a prompt for every code change.

他们进行了一项有超过一千名测试者的研究，发现自动模式能捕捉到89%的有害操作。

They conducted a study with over a thousand testers, finding that Auto Mode caught 89% of harmful actions.

人工审查员只能发现13.6%的有害代码变更。

Human reviewers only caught 13.6 of harmful code changes.

Anthropic 认为这主要是因为审批变得几乎是自动的，他们的用户批准了97%的代码变更。

Anthropic suggests that this is down to approvals becoming basically automatic, with their users approving 97% of code changes.

自动模式演变的一个有趣之处在于，它迫使 Claude Code 以更安全的方式工作。

One of the interesting things about the evolution of Auto Mode is that it has forced Claude Code to work in a safer way.

系统使用一个分类器来检测破坏性或不可逆的代码变更并阻止它们，当这种情况发生时，Claude 通常会找到更安全的方法来实现其目标。

The system uses a classifier to detect destructive or irreversible code changes and block them, and when this happens Claude typically finds a safer way to achieve its goal.

只有在没有更安全的选项时，它才会提醒用户。

Only once it runs out of safer options does it alert the user.

自动模式现在将成为Pro、Max和团队版的默认设置，但企业版仍然需要选择启用。

Auto Mode will now be the default for Pro, Max, and Team Plans, but will remain opt in for the Enterprise.

不过，对于企业用户来说，Anthropic表示使用自动模式有显著的好处。

Still, when it comes to those Enterprise users, Anthropic suggests there are significant benefits to using Auto Mode.

他们声称，使用自动模式的用户提交的PR数量增加了25%，而且包括Adobe、Gusto和Garner Health在内的许多组织已经将自动模式作为生产环境的默认设置。

They claim that Auto Mode users ship 25% more PRs, and that many organizations including Adobe, Gusto, and Garner Health are already running Auto Mode as their production default.

对于Anthropic团队来说，毫不意外，自动模式是默认设置，Claude Code的创造者Boris Cherny评论说：“我和团队只使用自动模式，已经好几个月了。”

For the team at Anthropic, perhaps unsurprisingly, Mode is the default, with Claude Code creator Boris Cherny commenting: The team and I use Auto Mode exclusively and have been for many months.

我无法想象回到权限提示的时代。

I couldn't imagine going back to permission prompts.

非常兴奋能把这个功能带给所有人。

Really excited to get this out to everyone.

自动模式的崛起是另一个例子，说明我们与AI交互的默认方式正在改变，这正好为我们本期节目做了铺垫，本期我们将介绍最新的热门词汇：图工程！

The ascendancy of Auto Mode is another example of how our default patterns of interacting with AI are changing, which provides the perfect segue to our main episode, a primer on the latest buzzy buzzword: Graph engineering!

如果你在企业里负责推进 AI，那你肯定知道，眼下最大的差距不在能力，而在落地执行。

If you're leading AI inside an enterprise, you already know that the gap right now isn't capability but execution.

正因如此，毕马威的《You Can With AI》节目迎来了新一季，邀请到 Emma 公司的 Sarojit Chatterjee、Ryder 公司的 May Habib、McKesson 的首席信息官 Ellery Fisher 等业界领袖，聊聊那些真正管用的实操经验。

That's why KPMG's You Can With AI is back with a new season featuring conversations with leaders like Sarojit Chatterjee of Emma, May Habib of Ryder, McKesson CIO, Ellery Fisher, and others focused on practical execution.

看看哪些做法行得通，哪些踩了坑，以及要从试点走向规模化落地，在战略、数据准备、治理、团队建设和价值创造这些方面，到底需要付出什么。

What's working, what's not, and what it actually takes to move from pilots to real scaled impact across strategy, data readiness, governance, workforce, and value.

当然啦，这档节目由我和 Nathaniel Whitimore 共同主持。

And of course, it's co hosted by me, Nathaniel Whitimore.

去听听并订阅吧：市面上所有的 AI 编程工具第一步都是一样的，就是开始写代码。

Go listen and subscribe at: Every AI coding tool on the market does the same thing first: it starts writing code.

Blitzy 的做法恰恰相反。

Blitzy does the opposite.

在敲下第一行代码之前，Blitzy 会花上几天时间对你的整个代码库进行逆向工程分析。

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

成千上万个智能体处理着数百万行代码，梳理出每一个依赖关系、每一条未文档化的约束条件，以及过去十年里做出的每一项架构决策。

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

结果就是一个动态知识图谱，能像一位在建筑行业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，而Blitzy从不猜测。

Other tools guess at context with grep searches and markdown files Blitzy never guesses.

它先建立真正的理解，然后自主完成超过80%的整个软件史诗任务。

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.

我经常在企业AI公司看到一种现象：他们在每个云、每个模型、每个框架上都做对冲，或者付钱给GSI做一个永远结束不了的试点项目。

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

这个团队确实在交付产品，他们选定了方向，而且行动很快。

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，并且能在四十五天内交付生产级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，实际完成工作。

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申请AI简报吧。

Request an AI briefing at robotsandpencils.com.

跟 Robots and Pencils 聊一次，你就明白了。

One conversation with robotsandpencils and you'll know.

本期 AI 每日简报由 Hyperagent 赞助播出，在这里你可以运行智能体集群，让团队共同管理。

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

新用户可获得一千美元的推理额度。

New users get $1,000 in inference.

别再依赖本地智能体或聊天工作流了，也不用再等着在笔记本上输入提示词。

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

Hyperagent 会在云端全天候部署智能体，直接在你团队已经使用的工具里干实事。

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

营销团队的智能体能把竞争对手的动态直接变成落地页。

Marketing's agent turns competitor moves into landing pages.

销售团队的智能体负责给潜在客户补充信息、起草邮件，还能更新客户关系管理系统。

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

运营团队的智能体则负责催办文件流程并跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个智能体都能访问共享上下文，并遵循你关于范围和审批的规则。

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

是时候让你的智能体像团队成员一样了。

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

去Hyperagent雇佣你的智能体吧，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.

今天我们要讨论AI推特上最新的热门术语——图工程。

Today, we are discussing the latest buzzy term on AI Twitter, which is graph engineering.

这个术语确实有点让人困惑，因为一方面它有点半开玩笑，另一方面，不同的人谈论它时，它描述的其实是两件不同的事情。

Now this one admittedly is a little confusing because a, it's sort of started tongue in cheek, and b, depending on who's talking about it, it kind of is describing two different things.

但我实际上认为，这个概念至少有助于我们理解从提示词到上下文，再到工具链，再到循环的工程演进脉络。

But I actually think that the concept at least is useful to situate relative to the lineage of engineerings we've had from prompt to context to harness to loop.

所以我觉得值得做这个入门介绍。

And so I think it's worth doing this primer.

确实，这更像是一个关于图工程的入门指南，而不是完整的图工程教程。

And indeed, this is definitely more a primer on graph engineering than a complete guide to graph engineering.

我想让你快速了解这个术语是什么，以及为什么我觉得它很重要。

I wanna bring you up to speed on what this term is and why I think it matters.

引发这场讨论的推文来自Open Claw的创始人Peter Steinberger。

So the tweet that kinda kicked off this this ourse came from Open Claw creator Peter Steinberger.

七月中旬，他写道：我们还在讨论循环吗，还是已经转向图了？

Back in mid July, he wrote, are we still talking loops or did we shift to graphs yet?

AI创作者Matthew Berman说出了很多人的心声，他说：兄弟，停一下。

AI creator Matthew Berman captured the feelings of many when he said, bro, stop.

我在度假呢。

I'm on vacation.

而几乎立刻，Twitter上的那些吹嘘者就开始行动了，到处都能看到像“循环工程已死，图工程万岁”这样的大胆宣言。

And while almost immediately, the Twitter article boasters got to work, big bold declarations like loop engineering is dead, long live graph engineering were visible all over the place.

但在最初的炒作和喧嚣之后，这里确实有一些真正有趣的东西。

But after this initial phase of hype and bluster, there is in fact actually something interesting here.

所以让我们来谈谈所有围绕这些概念发展起来的工程实践。

So let's talk about all of the things that have had engineering around them.

我们最早接触的工程实践，当然就是提示工程。

The first blank engineering that we had was, of course, prompt engineering.

这在2023年和2024年初的很多AI课程中都是核心内容。

This was what a lot of the AI courses around '23 and early twenty twenty four were all about.

而且，取决于你关注哪家公司，不幸的是，今天很多技能提升课程仍然以它为主要内容。

And depending on which corporation you look at, still unfortunately the substance of a lot of those upskilling courses today.

提示工程的概念本身就承认了，即使在那时，我们也在改变工作方式。

The idea of prompt engineering was a recognition even then that we were shifting how we did work.

我们不再自己完成所有工作，而是让AI聊天机器人来承担一部分工作。

Instead of doing all the work ourselves, we were deputizing an AI chatbot to do some amount of that work.

无论那是最终成果，还是像研究这样的中间步骤，我们仍然需要找到优化提问方式的方法，以获得最佳结果。

Now whether that was final production or just some intermediate step like research, we still needed to find ways to optimize what we were asking for to get the best results.

这就是提示工程。

That was Prompt Engineering.

在提示工程的不同阶段，有很多技巧和窍门，比如让AI假装成某种类型的人，还有花哨的JSON工程，用复杂的输入方式来理论上更好地组织请求，中间还有其他各种方法。

At various points in the lifecycle of Prompt Engineering, had tips and tricks ranging from: telling the AI to pretend it was a certain type of person, to fancy JSON Engineering, which used this complex way of typing to theoretically better structure requests, and we kind of had every other thing in between as well.

然而，到了2025年，我们开始意识到提示只是充分利用AI的一部分。

Heading into 2025, however, we started recognizing that the prompt was only one part of getting the most out of AI.

我们不仅需要擅长以正确的方式提问，还需要擅长给AI提供它所需的所有信息和知识，以便它能够出色地完成那个提示所要求的任务。

We didn't just need to be good at asking for things in the right way, we also needed to be good at giving AI all of the information and knowledge it needed to do a good job with whatever that prompt was.

举个简单的例子，如果你让大语言模型创建一个非常成功且符合品牌形象的营销活动，首先，它需要知道什么是符合品牌形象，这意味着要给它访问品牌指南的权限，可能还要提供过去关于品牌价值等内容的书面材料。

To take a simple example, if you're asking your LLM to create a highly successful on brand marketing campaign, Well, first of all, it needs to know what on brand is, which means giving it access to brand guidelines and potentially other write ups in the past about things like brand values.

而且，为了不让它只是猜测什么是好的，给它之前营销活动的统计和分析数据会很有帮助，可能还要加上一些关于哪些有效、哪些无效的主观反思。

And in order to have it not just be guessing at what good means, it would probably be helpful to give it stats and analytics from previous marketing campaigns, perhaps with some subjective reflections on what worked and what didn't as well.

围绕提示的那部分信息就是上下文。

That body of information that surrounds the prompt is the context.

上下文工程就是确保所有这些类型的信息都能以正确的方式被访问到。

Context engineering was all about making sure that all of that type of information was accessible in the right way.

现在到了这个阶段，我们又看到了一个有趣的分歧，我觉得这个最新的图工程术语也会再次出现这种情况。

Now here at this point, we also have an interesting split, which I think we re going to see once again with this latest graph engineering term.

大体上，这个分歧存在于技术人员和软件开发者，以及其他所有人之间。

The split, broadly speaking, is between technical folks and software developers and everyone else.

对于软件开发者来说，上下文工程并不是确保AI能访问到合适的文件那么简单。

For the software developers, context engineering wasn t a matter of making sure that your AI had access to the right files for the job.

实际上，这确实是一项工程任务。

It was, in fact, an actual engineering task.

不仅要考虑哪些信息有用，还要设计技术系统，让AI能够遍历可访问的上下文网络，而不会耗尽整个上下文窗口。

It was thinking about not just what information is useful, but designing the technical systems by which the AI could traverse the web of accessible context in a way that didn't just spend the entire context window.

在这一方面，上下文工程师实际上是在考虑上下文预算，比如在设计应用时，确保某些流程部分不会在上下文中陷入困境，而其他部分在需要时可以更深入。

On this side, context engineers were actually thinking in terms of context budgets, making sure, for example, as they were designing applications, that certain parts of a process didn't get bogged down in context, while others could go deeper when they needed it.

所以在这里，上下文大致指的是同一件事，但对工程师来说，工程是一项实际的工程任务，而对其他人来说，则是一种思维方式，即如何围绕他们使用的大语言模型组织信息。

So here we have context broadly referring to the same thing, but engineering being a literal engineering task for the engineers and a mindset for everyone else in terms of how they organized information around the LLMs that they were using.

今年，就像其他一切都在加速一样，这些“某某工程”术语的更新换代也加快了。

Now this year, just like everything else is sped up, we've also had a speed up in the succession of blank engineering type of terms.

首先，你们可能听过我提过 harness 工程。

Starting with, you've probably heard me talk about harness engineering.

所谓 harness，当然就是模型周围存在的环境。

The harness is, of course, the environment that exists around a model.

简单点说，这可能就是一个具体的软件工具，比如 Claude Code 或者 Codex；但更广泛的定义里，harness 包括了从工具、权限设置到 AI 或智能体用来干活的各种技能文件等所有可用资源。

On a simple level, that might be an actual software tool like Claude Code and Codex, but the more expansive definition of harness includes everything from the tools to the permission sets to the skills files that an AI or agent has available to it to do its work.

到了 2026 年，大家越来越习惯把智能体看作是模型加上其周围 harness 的组合体了。

Throughout 2026, people have become more and more comfortable with the idea that the agent is actually a combination of the model and the harness that surrounds it.

所以你看，比如现在每次出新的评测数据时，公司都会说明这些评测是用什么框架跑的，因为这确实是故事里很重要的一部分。

This is why, for example, frequently when we're getting new benchmarks, companies will now explain what harness the benchmarks were running as that's actually an important part of the story.

说到评测框架工程，这里又出现了两种完全不同的含义。

Now when it comes to harness engineering, once again, we've got two very different meanings of engineering.

当然，有一群真正的工程师和软件开发人员一直在构建各种更优秀的评测框架，并努力深化我们对这类框架的整体理解。

There are, of course, the actual engineers and software developers who have been building different and better harnesses, and trying to advance our understanding of harnesses in general.

而另一种更偏向个人化的理解则是，它关乎你给智能体搭配哪些技能，以及它们能调用哪些工具。

And then there's the more individualist sense of harness engineering, which is about things like which skills you surround your agents with, and what tools they have access to.

你会发现，每当这些新术语出现时，旧的那些并不会消失。

Now you'll see here that as each of these new terms comes online, it's not like the old one goes away.

提示工程现在可能是杠杆作用最小的一个，但并不是说我们开始重视工具环境，上下文就不重要了。

Prompt engineering is certainly the one where there's probably at this point the least leverage to be had, but it's not like because we started to understand the value of harnesses, all of a sudden context stopped mattering.

实际上恰恰相反，工具环境成了上下文工程的新背景。

In fact, quite the opposite the harness became a new context for that context engineering.

所以，我们有提示词来控制指令，上下文来控制模型看到的内容，工具环境来控制环境，然后还有循环，它控制智能体为实现目标而进行的迭代。

So we've got the prompt which controls the instructions, we've got context which controls what the model sees, the harness which controls the environment, and that brings us to the loop, which controls the iteration that an agent goes through to accomplish a goal.

循环或循环工程，最近几个月成了热门话题，它讲的是用不同的方式来看待你和智能体的关系。

Loops or loop engineering, which has become a big topic over the last few months, is about thinking about your relationship with agents in different ways.

最经典的简短解释还是来自OpenAI的Peter Steinberger，他说：你不应该再直接给编码智能体发提示了。

The canonical short explanation once again came from open clause Peter Steinberger who said: You shouldn't be prompting coding agents anymore.

你应该设计循环，让循环来给你的智能体发提示。

You should be designing loops that prompt your agents.

循环就是一套系统，智能体通过它来观察、规划、行动、检查结果，然后重复，直到达到某个可衡量的停止条件。

Loops are the systems by which an agent can observe, plan, act, check results, and repeat until some measurable stop condition is reached.

就像我们之前在节目里聊过的，非工程师在尝试应用循环时遇到的一个大挑战，就是搞清楚自己知识工作中的哪些部分有那种可衡量的停止条件。

As we've discussed in the past on this show, one of the big challenges for non engineers who've been trying to put loops to work is in figuring out which aspects of their knowledge work have those sort of measurable stop conditions.

大概一两个月前，我在节目里聊循环的时候提到过，有些情况下，如果没有天然的可衡量条件让代理在这种循环里工作，你就得自己精确定义一个类似的可衡量条件，才能真正让循环跑起来。

One of the things that we discussed in my show about loops from a month or two ago was this idea that in some cases where there wasn't a natural measurable condition to get an agent working in this sort of loop, you were going to have to precisely define something measurable like that to actually get the loop to work.

但你会发现，循环关注的是如何从单个代理或代理流程中获取最大价值。

But what you'll notice here is that a loop is about how to get the most out of a single agent or agentic process.

它是从特定目标倒推出来的，给代理提供可重复的步骤，让它根据需要重复执行，直到真正达成那个目标。

It is a work backwards from a specific goal that gives the agent the repeatable steps it needs to follow that as many times as is necessary to actually achieve that goal.

但如果目标更复杂，需要多个不同流程相互配合才能完成呢？

But what about when a goal is more complex and requires multiple different processes interacting to actually accomplish whatever that goal is?

如果我们要超越单一代理流程的输出，去设计一个持续运行的代理系统来工作呢？

What about when we move beyond the output being a single agentic process to actually designing an ongoing agentic system for working?

这就进入了图工程（graph engineering）的领域。

That's where we get into graph engineering.

如果说提示词控制指令，上下文控制模型看到的内容，环境控制环境，循环控制迭代，那么图控制的就是新的代理组织方式。

If prompts control the instruction and context controls what the model sees, harnesses control the environment, and loops control the iteration, the graph controls the new agentic organization.

图工程就是设计多个智能体、工具、知识源和人类之间如何互动和连接。

Graph engineering is about designing how multiple agents, tools, knowledge sources, and humans interact and connect.

图工程既描述了系统的各个部分——有些人称之为节点，这些节点可以是智能体、路由器或人工网关——也解释了这些节点之间的交互，比如允许哪些交接，哪些信息或状态在它们之间传递。

Graph engineering describes both the parts of the system, which some people are referring to as nodes, that could be agents, routers, or human gateways, And graph engineering also explains the interactions between those nodes, which handoffs are permitted, what information or state travels between them.

ExplainX.ai 将循环解释为单个智能体的自主循环。

ExplainX dot AI explained a loop as an autonomous cycle for a single agent.

触发条件触发，智能体行动，验证器检查，如果未完成，整个系统会用更新的上下文重试，直到目标达成。

A trigger fires, an agent acts, a verifier checks, and if not done, the whole system is retried with updated context until the goal has been met.

他们写道，每个防护措施都适用于单个智能体的运行。

As they write every guardrail applies to one agent's run.

循环是智能体与自身的行为契约。

The loop is the agent's behavioral contract with itself.

而图则是智能体的组织。

A graph, on the other hand, is an organization of agents.

正如他们所说，图中的每个节点都是一个运行自己循环的智能体。

As they put it, each node in a graph is an agent running its own loop.

这些节点之间的边或交互定义了数据流和依赖关系。

The edges or interactions between those nodes define data flows and dependencies.

他们接着说：图规定了：谁存在，哪些代理以及各自有什么专长，每个代理拥有什么，他们的领域、上下文和工具访问权限，工作如何按顺序、并行或有条件地进行，最后，图还定义了失败时会发生什么。

They continue: The graph specifies: who exists, which agents and with what specialization, what each owns, their domain, context, and tool access, how work moves sequentially, in parallel, or conditionally And finally, the graph defines what happens on failure.

节点会重试吗？

Is the node retried?

会路由到备用方案吗？

Is it routed to a fallback?

还是向上游发出警报？

Or is there an alert upstream?

他们总结说，图是组织的运营结构。

The graph they conclude is the organization's operating structure.

循环存在于节点内部。

Loops live inside nodes.

图将它们连接起来。

The graph connects them.

简单来说，循环就是单个智能体如何完成自己的工作，而图则是整个智能体组织如何运作。

In short, a loop is how an individual agent does its job, where a graph is how an entire agentic organization works.

正如谷歌所说，循环让智能体的行为变得可编程，图让智能体组织变得可编程。

As Google put it, loops made agent behavior programmable, graphs make agent organizations programmable.

不过，我并不指望你们所有人都立刻去设计完整的智能体组织。

Now what I don't expect is for all of you to run out and start designing complete agentic organizations.

但图工程的核心思想，就是能够用这种系统性的思维去思考。

But the idea of graph engineering is to be able to think in those system terms.

而且，和上下文工程及工具链工程略有不同，在循环和图之间，虽然有时你会同时用到这两种模式，但有些时候，简单的单循环架构就足够了。

And even slightly differently from the context and harness engineering, with loops and graphs, while sometimes you will use both of these patterns, there will be times when the simpler architecture of a single loop is going to be fine.

当单个任务有明确的终点，步骤是真正的顺序执行，并且一个智能体的上下文窗口能容纳整个领域时，单循环就是一个很好的选择。

When a single job has a clear finish line with genuinely sequential steps and one agent's context window able to hold the whole domain, that's a good candidate for a single loop.

当工作被拆分成不同的专业领域，需要不同的交接，当并行变得有价值，当流程中的不同步骤需要不同的模型或工具集，当路由必须明确，并且你想设计一个有韧性的系统，让一个节点的失败不会拖垮整个系统时，那就需要真正的图工程了。

When the work instead splits into specialties with different handoffs, when parallelism becomes valuable, when different steps in a process want different models or tool sets, when routing has to be explicit, and when you wanna design a resilient system where the failure of one node doesn't take down the rest, that's where you get into this actual graph engineering.

而且，就像通常发生的那样，一旦我们有了一个新术语，我们很快就会探索它的各种细微差别。

Now as tends to happen, as soon as we get a new term, we very quickly explore all the nuances as well.

比如说，现在大家讨论比较多的一个话题，就是组织图和工作图的区别。

One discussion, for example, that you're seeing a little bit is the difference between org graphs versus work graphs.

组织图是更稳定的智能体系统，它定义了一种更长期的配置方式。

Org graphs are gonna be more stable agentic systems that defines a more permanent style of setup.

组织图里会有长期存在的智能体，每个智能体负责一个领域，随着时间积累上下文，保存记忆，还有稳定的关系和依赖，除非明确指示，否则不会改变。

The org graph is gonna have long lived agents with each agent owning a domain and accumulating context over time, preserved memory, and stable relationships and dependencies that don't change unless explicitly told to.

所以，如果你在设计一个智能体组织，它要反复做同样的事情。

So if you are designing an agentic organization that's going to do the same thing over and over again.

比如，如果我在设计一个多智能体系统，自动化从研究到生产、编辑、发布、提取洞察、再到发布的整个流程，那可能就适合用组织图，因为这些是持续重复的过程。

For example, if I was designing a multi agent system that automated the process of going from research to production to editing, to publishing, to extracting insights, to posting, that might be a good candidate for an org graph because these are ongoing recurring processes.

这就是我日常工作每天都要做的事情。

This is just how my work happens day in and day out.

而工作图则更动态，更临时。

The Work Graph, on the other hand, is more dynamic and more ephemeral.

它可以包含任务节点，这些节点只在任务存在时存在；动态边，记住边是节点之间的交互，可以分裂或合并；还有自适应结构，任务可以在证据表明不再需要时消失，或者如果发现新的复杂性，可以生成新任务。

It can include task nodes only exist as long as the work exists, dynamic edges remember edges are the interactions between the nodes that can split or merge, an adaptive structure where tasks can disappear when evidence makes them unnecessary, or new tasks can be spawned if new complexities are discovered.

不过，让我们回到重点，来结束这次基础介绍。

But let's wrap up this primer by coming back to the main point.

就像我一开始说的，我的预期并不是说大家了解了图工程这个有点拗口的概念后，就会立刻去设计复杂的智能体组织。

Like I said at the beginning, my expectation is not that all of a sudden, you go out and design complex agentic organizations now that you are acquainted with this wonky concept of graph engineering.

但我认为，这对我们每个人都有价值。这就好比即使你不打算设计循环，但了解循环的架构——触发器如何启动、执行什么动作、验证器如何检查工作，然后重复直到完成——对于思考如何利用智能体自动化你工作中的某些部分，是非常有帮助的。

What I think is valuable for all of us, however, in the same way that even if you weren't designing loops, understanding the architecture of a loop, a trigger that fires, an action that's taken, a validator that checks the work, and then that on repeat until it's done, that is extremely helpful in thinking about how to use agents to automate chunks of your work.

同理，我认为图工程将为许多人解锁一种能力，那就是开始用多智能体系统的思维去思考问题，你能看到不同智能体承担不同的职责，并且真正理解甚至设计它们彼此之间的关系。

In the same way, what I think graph engineering will unlock for many is the ability to start thinking in multi agent systems terms, where you can start to see different agents with different jobs and actually understand and even design their relationships with one another.

随着时间的推移，我敢保证，你们中的一些人将会开始设计更复杂的智能体系统。届时，围绕图工程这一领域积累的最佳实践、经验教训和工具集，将会变得极其有用。

Over time, some of you, I guarantee, will start to design those more complex agentic systems, and the best practices and lessons and tool sets that people build around this graph engineering discipline are going to be extremely useful when you do.

所以没错，图工程确实是最近很火的一个热词，而且早期关于它的推文说实话有点调侃的意味。

So yes, graph engineering is the latest buzzy buzzword, and some of the early tweets about it were frankly tongue in cheek.

但我认为，设计智能体系统是一种全新的工作原语，也是我们将越来越频繁地被要求去做的事情。

But designing agentic systems is, I believe, a new work primitive, and something which we will increasingly be called upon to do.

希望现在大家对这一点有了更清晰的认识，可以根据自身情况深入探索。

Hopefully you now have a better sense of that, and can dig in as makes sense for you.

好了，今天的AI每日简报就到这里。

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

感谢大家的收听或观看，我们下次再见，祝一切顺利！

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