M113:05
Token消耗在攀升，安全问题正阻碍着AI代理，而且没人知道AI是如何真正被使用的。
Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used.
根本原因是：AI与你的业务所运行的系统及上下文脱节。
The root cause: AI is cut off from the systems and context your business runs on.
这就是CData构建Connect AI Gateway的原因——一个位于AI与系统之间的控制点。
That's why CData built Connect AI Gateway, one control point between your AI and your systems.
它建立在支撑Palantir、Google、Salesforce等数百个系统的连接层之上，每个代理都能得到其所需的数据和工具。
Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs.
它能精确管控到每一条记录，让您终于能在同一处看到整个组织的完整AI使用情况。
It governs down to the record and lets you finally see your org's complete AI usage in one place.
准确的答案，Token成本的一小部分，以及一个每交互一次就更了解你业务的网关。
Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction.
在cdata.ai上了解更多。
See how at cdata.ai.
你一整天都在连续开会，努力保持专注，但同时又担心会忘记某个决定、行动项或重要的下一步。
You're in back-to-back meetings all day, you're trying to stay present, but you're also worried you'll forget the decision, the action item, the important next step.
这种情况在我们AI Daily Brief这里绝对经常发生，因为一切变化极快，我们也在快速调整我们所做的一切。
This is a scenario that we absolutely live here at the AI Daily Brief, given how fast things move and change and how quickly we're evolving everything that we do.
这时Granola就派上用场了。
That's where Granola comes in.
Granola是一款用于会议的AI驱动笔记应用。
Granola is an AI-powered notepad for meetings.
你像平时一样随手记下粗略的笔记，在后台，Granola会在会议结束时将它们转录并整理成清晰有用的笔记。
You jot down rough notes like you always do, and, in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends.
没有机器人加入你的通话，没有干扰，只是一个干净的笔记本来帮助你集中注意力。
No bots joining your calls, no distractions, just a clean notepad that helps you focus.
在通话中或通话后，你可以与你的笔记聊天——让Granola提取行动项、帮助谈判、写跟进邮件，甚至使用预设提示的“recipes”来指导你。
During or after the call, you can chat with your notes— ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are premade prompts.
一旦你在第一次会议中试用它，之后就很难离开它了。
Once you try it on a first meeting, it is very hard to go without.
前往granola.ai/AIDaily并使用代码AIDAILY，新用户可享受3个月全免费。
Head to granola.ai/AIDaily and use code AIDAILY to get 100% off for 3 months for new users.
那是 granola.ai/AIDaily，使用代码 AIDaily 新用户可享受三个月100%折扣。
That's granola.ai/AIDaily and code AIDaily to get 100% off for new users for 3 months.
Blitzy 对庞大代码库的理解能够实现自动安全修复、现代化和新功能。
Blitzy's understanding of massive codebases unlocks autonomous security fixes, modernization, and new features.
那么，当完全没有遗留代码时会发生什么？
So what happens when there's no legacy code at all?
Greenfield 应该是容易的部分——全新的开始，没有技术债务。
Greenfield is supposed to be the easy part— clean slate, no technical debt.
但即使是 greenfield，也只能以人的速度前进，一次一个 sprint。
But even greenfield moves at human speed, one sprint at a time.
Blitzy 将工作单元从开发者转移到项目，自主规划、构建、测试和验证整个应用程序从头开始。
Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.
数十万行生产级代码。
Hundreds of thousands of lines of production-ready code.
一位 Blitzy 客户部署了一个全新的应用程序，534,000 行代码，将65周的路线图压缩到2周内。
One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.
另一个客户在没有前端工程师的情况下交付了整个应用程序。
Another shipped an entire application with no frontend engineer.
无论是 legacy 还是 greenfield，答案都一样：以计算速度交付软件。
Legacy or greenfield, the answer is the same: software at the speed of compute.
在 blitzy.com 构建下一个项目。
Build what's next at blitzy.com.
就是 B-L-I-T-Z-Y 点 com。
That's B-L-I-T-Z-Y dot com.
欢迎回到 AI Daily Brief。
Welcome back to the AI Daily Brief.
有很多公司在谈论他们的 AI 转型或成为 AI 原生，但大多数公司，我应该说我们大多数人，因为我认为我们在 AIDB 和 Superintelligent 以及现在如何走到这里都是其中的一部分，既在使用 AI 以更快更好的方式做我们一直做的事情，也可能越来越多地发现以前无法做到的新事情。
There are a lot of companies that talk about their AI transformation or becoming AI native, and yet most of them, most of us I should say, 'cause I consider what we do at AIDB and Superintelligent and now how we got here is all part of this, are both using AI to do the same stuff we've always done, but a lot faster and better, and also perhaps increasingly figuring out new things that we couldn't do before.
但输出和产品仍然相当相似。
But the outputs and products remain fairly similar.
组织结构尽管以新的方式被拉扯和扭曲，但20年前的人仍然能够认出。
The org structure, although pulled and pushed and twisted in new ways, still would be recognizable to someone from 20 years ago.
但是当你忘记所有建立初创公司的旧规则，吞下 AI 药丸，完全投入用 agent 构建时会发生什么？
But what happens when you forget all the old rules of building a startup, take the AI pill, and go completely all in on building with agents?
这就是 Jacob Sandbury 和他的公司 Spawn 所做的。
That's what Jacob Sandbury did with his company Spawn.
在过去的两年里，他费尽心思地弄清楚了如何建立一个完全以 AI 为先的 gaming 初创公司。
Over the past 2 years, he painstakingly figured out how to build a radically AI-first gaming startup.
经过大量的试错，他的10人团队现在正交付一个令人印象深刻的产品，该产品建立在巨大的 agentic 堆栈上。
After a lot of trial and error, his 10-man team is now shipping an impressive product built on a huge agentic stack.
本周，Jacob 在一篇题为《How I Built the Most AI-Pilled Company in the World》的博客文章分享了他的操作指南。
And this week, Jacob shared his playbook in a blog post entitled How I Built the Most AI-Pilled Company in the World.
在我们深入探讨之前，先简要说明一下这个产品是什么。
Now, before we get into it, a brief explanation of what the product is.
Spawn 是一个大型多人在线游戏创作平台。
Spawn is a massively multiplayer game creation platform.
如果你把它想象成 metaverse 加上 Lovable 再加上 Steam，至少你方向没错。
If you think metaverse crossed with Lovable crossed with Steam, you're somewhere at least in the right ballpark or playing the right sport.
Jacob 描述它是一个大型多人在线游戏创作引擎，数千人可以同时处于一个游戏中，实时改变其规则、资产和图形，在大量设备上无延迟运行。
Jacob described it as a massively multiplayer game creation engine where thousands of people can be inside one game, changing its rules, assets, and graphics in real time on a huge range of devices without lag.
开发者 Simon Willison 补充说，Spawn 是一个大型多人在线游戏环境，玩家可以通过向游戏内 agent Savi 发出提示来修改游戏世界。
Developer Simon Willison added, Spawn is a massively multiplayer gaming environment where players can use prompts to the in-game agent, Savi, to modify the game world.
我在手机上启动了几分钟，就能够通过提示让一座城堡出现，然后是一条龙，然后骑龙绕着那座城堡飞行。
I fired it up on my phone for a few minutes and was able to prompt a castle into existence, then a dragon, then ride the dragon to fly around that castle.
我完全可以做一期单独的节目来讨论这类大型多人游戏中生成式世界的实例化，但这不是本期节目的重点。
Now, I could do a whole separate show about the instantiation of generative worlds in these sort of massive multiplayer games, but that's not what this show is about.
事实上，就本期节目的目的而言，产品具体是什么并不那么重要。
In fact, for the purposes of this show, it doesn't really matter all that much what the product is.
重点是 Jacob 的操作指南，它适用于任何本质上在交付代码的公司，也可能适用于许多并非如此的公司。
The point is Jacob's playbook, and that will work for any company that's at its core shipping code, and probably for a whole lot that aren't as well.
核心理念是，把当前智能体所能做的事情，推到绝对的极限。
The underlying idea is to take what agents can do now and push it to the absolute limit.
就像我刚才提到的，Spawn 是一家十人创业公司，他们有一支智能体团队，叫做舰队。
As I mentioned, Spawn is a 10-person startup with a team of agents that they call the fleet.
跟大多数人构建的东西不同，Spawn 的智能体舰队是常驻的，而不是针对特定任务临时启动、用完就关闭。
Unlike what most people have been building, Spawn's agent fleet is persistent rather than being spun up and spun down for particular tasks.
每个智能体都有自己的名字和身份、自己的记忆，而且关键在于，他们都负责公司内部的某一项职能。
Each agent has their own name and identity, their own memory, and crucially, ownership of a certain function within the company.
有一组智能体负责维护游戏引擎，另一组负责找 bug，还有一组负责财务。
There's a team of agents in charge of maintaining the game engine, another for bug hunting, and another for finance.
这些智能体甚至周日休息，算是有点停机时间。
The agents even get Sundays off for some downtime.
Spawn 上个月在 token 上花了 310 万美元，是工资支出的 30 倍。
Spawn spent $3.1 million on tokens last month, 30 times what they spent on payroll.
这让他们成为 Anthropic 最大的客户之一。
That makes them one of Anthropic's biggest customers.
舰队上个月交付了 77,000 个代码变更，平均每小时 108 个。
The fleet shipped 77,000 code changes last month, an average of 108 an hour.
但更重要的是，一支常驻的智能体舰队从根本上改变了这个团队对发布软件的思维方式。
But more importantly, having an always-on fleet of agents has fundamentally changed how the team thinks about shipping software.
Jacob 明确表示，在 Fable 推出之前，这一切都行不通。
Jacob was clear that none of this worked until the launch of Fable.
那个模型让他们的交付量提升了 500 倍，但也意味着要围绕 Fable 的能力彻底重建他们的运营方式。
That model has 500x'd the number of changes they could ship, but it also meant completely rebuilding their operations around Fable's capabilities.
Jacob 写道：总的来说，我们发现 OpenAI 的模型会自己跟自己玩猫腻，Google 的模型连贯性维持不了太久，而 xAI 的模型还不够聪明，虽然不会做出特别离谱的事。
Jacob wrote, in general, we've seen that OpenAI's models game their own checks, Google's models don't stay coherent for very long, and xAI's models just aren't smart enough yet, though they don't do anything cartoonishly bad.
Anthropic 的模型能保持连贯，做出好的决策，而且能真正干活。
Anthropic's models stay coherent, make good decisions, and can do the work.
在我看来，最制约他们的地方是，他们不愿意冒足够的风险。
In my opinion, the thing that holds them back most is that they won't take enough risk.
显然，Spawn 的工作方式非常罕见，以至于今年早些时候，Jacob 一度觉得自己用智能体的方式可能不对，因为他的 AI 账单比他在大型科技公司里的朋友看到的还要大得多。
Now, obviously, Spawn's way of working is extremely rare, to the point that earlier this year, Jacob thought he was using agents the wrong way because his AI bills were so much larger than even his friends were seeing inside big tech.
但结果发现，他写道，我们并没有做错什么，只是对 AI 的思维方式完全不同。
Turned out, however, he wrote, we weren't doing anything wrong, we just think about AI very differently.
大多数公司是让庞大的人类团队使用 AI，而我们是用一支小型人类团队，搭配一支庞大的智能体团队。
Most companies run big human teams that use AI, and we run a small human team alongside a big team of agents.
为了搞清楚我们到底哪里不一样，我和 Anthropic 的一位研究员交换了看法。
To figure out what we were doing differently, I compared notes with a researcher at Anthropic.
他们和很多客户聊过，他们说 Spawn 是运作起来最像 Anthropic 的公司。
They talked to a lot of customers, and they said Spawn is the one that works most like Anthropic.
所以，为了让这一切运转起来，Spawn 在构建舰队时采取了一种激进的方式。
So to make this all work, Spawn took a radical approach to building the fleet.
如前所述，每个智能体都有真实的系统权限和明确的职责，就像人类员工一样。
As mentioned before, each agent has real system access and clear responsibilities, similar to a human worker.
每天，智能体舰队会生成成千上万个辅助智能体去完成那些细碎的工作。
Every day, the agent fleet spawns thousands of helper agents to carry out the granular work.
与此同时，人类员工使用修改版 Claude Code 来处理那些需要亲自把控的全局性任务。
Meanwhile, the human employees use a modified version of Claude Code to work on bigger-picture items that need hands-on steering.
舰队是让游戏世界得以运行的基础设施。
The fleet is the infrastructure that allows the game world to operate.
当某个玩家通过自己的个人智能体改变了世界中的某个东西时，舰队要负责生成一个稳定的版本，这个版本能更新到每个玩家的设备上，同时保持游戏世界的同步。
When one player changes something in the world with their personal agent, the fleet is responsible for creating a stable build that can update across every player's device while keeping the game world in sync.
在我看来，这都算轻描淡写了。Jacob 写道，这可不是一个简单的问题。
Understating it greatly, in my opinion, Jacob wrote, this is not a simple problem.
而且据我们所知，世界上其他任何东西都做不到哪怕一小部分。
And as far as we know, nothing else in the world does even a fraction of it.
从功能上讲，玩家们是在一个实时代码库上工作，全靠舰队把一切串联起来。
Functionally, the players are working on a live codebase, all tied together by the fleet.
让舰队达到今天的水平，经历了一个漫长而反复迭代的过程。
Now, it was a long and iterative process to get the fleet to where it is today.
雅各布解释说，大多数时候，他们的问题都是通过移除那些为了辅助人类工作而设计但应用到智能体系统上却毫无意义的系统来解决的。
Jacob explained that most of the time, their problems were solved by removing systems designed to facilitate human work, but which made no sense when applied to agentic systems.
他们遇到的第一个例子是通信——这些智能体会花大量时间互相告诉对方自己在做什么。
The first example they ran across was communication— the agents would spend most of their time telling each other what they were doing.
在工程里，大部分对话都是人们在说他们要改动什么。
Most of the talking in engineering is people saying what they're about to change.
打个招呼说，我要改玩家数据的存储方式了，你们得更新一下自己的工作。
Heads up, I'm changing how we store player data, update your work.
一条这样的消息会牵连到很多其他工作。
One message like that touches lots of other work.
如果每个智能体都通知其他所有智能体，消息的增长速度会远远超过实际工作量。
If every agent tells every other agent, the messages grow much faster than the work.
智能体数量翻倍，闲聊量大约会翻四倍。
Double the agents and you roughly quadruple the chatter.
这绝对是个巨大的瓶颈。
That is a huge bottleneck.
雅各布指出，智能体其实不需要像人类那样频繁地交流，因为它们不像人类那样高度专业化。
Jacob noted that agents don't actually need to communicate as much as humans because they're not as specialized as humans.
在一个人类开发团队里，如果负责数据库的哥们儿和前端设计师沟通不畅，那就麻烦大了。
With a human dev team, if your database guy isn't communicating well with your frontend designer, you're gonna have a bad time.
结果是，大家要一起协作审核一个拉取请求，确保每个人对代码变更都一清二楚。
And the result is collaborating on reviewing a pull request to make sure everyone is clear on the code change.
而智能体对整个系统有很深的理解，所以这种沟通其实没什么意义。
Agents understand the entire system deeply, so the communication doesn't really have a purpose.
事实上，今年早些时候，Spawn就取消了拉取请求，直接让智能体在代码库里工作。
Earlier in the year, in fact, Spawn killed pull requests and just had the agents work directly in the codebase.
结果是，这套系统对人类开发者来说完全陌生，但似乎对纯智能体系统运行得很好。
The result is a system that's completely alien for human developers, but seems to work well for agent-only systems.
改动代码时，基本不用担心会影响代码库的其他部分。
Changes are made with little concern about breaking things elsewhere in the code.
如果某个改动导致其他东西出了问题，就会启动一个智能体来处理问题并修复它。
If something breaks due to a change, an agent is spawned to triage the problem and fix it.
Spawn要求整个集群每小时都要保持一切正常，并且每15分钟运行一次测试套件，以便快速发现问题。
Once an hour, spawn requires the fleet to have everything in line, and test suites are run every 15 minutes to find problems quickly.
事实证明，这种推代码的方式出奇地稳健，因为你不用在凌晨两点把人叫起来修致命Bug。
It turns out that this way of pushing code is surprisingly resilient, because you're not calling humans at 2 AM to fix a fatal bug.
相反，集群会不断在代码库上进行迭代，并在问题出现时立即修复。
Instead, the fleet is constantly iterating on the codebase and fixing problems as they arise.
雅各布说，在过去一年里，集群做了十二万三千次改动，其中只有大约一千次被回滚了。
Jacob said that of the 123,000 changes the fleet made over the past year, only about 1,000 were reverted.
你可能已经注意到了，这种工作方式极其消耗资源。
The issue that you might have spotted is that this way of working is incredibly resource-intensive.
Spawn运行在亚马逊云上的5000个CPU、17TB内存和150个GPU上，另外还有60%的弹性容量随时待命。
Spawn runs on 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs hosted on AWS, with another 60% in surge capacity on standby.
Spawn面临的另一个大问题，是人介入这个环节。
The next big problem for Spawn was the human in the loop.
雅各布写道，智能体一直在等人类告诉它们生产环境中出了什么故障。
Writes Jacob, the agents were waiting for humans to tell them what broke in production.
智能体很擅长读取世界状态并一步步接近目标。
Agents are great at reading the world and inching toward a goal.
很多团队仍然在测试机上使用产品的副本私下运行他们的智能体，然后期望它们更新代码并交付给客户。
Many teams still run their agents privately on test machines with copies of the product and expect them to update it and ship it to customers.
然后人类再把故障和Bug反馈给智能体。
Then humans relay the faults and bugs back to the agents.
这个循环非常混乱，纯粹是在浪费时间。
That loop is a huge mess and a waste of time.
在Spawn里，玩家的智能体会在游戏世界中自动检测Bug并提交Bug报告。
In Spawn, the player's agents automatically detect bugs and file bug reports as they navigate the game world.
这些报告连同功能请求或日志中的问题，都会被输入到集群中。
Those are then fed into the fleet, along with feature requests or issues in the logs.
一个小模型收到通知，对bug进行分类或决定是否添加功能，然后将其传递给fleet中的相关agent。
A small model receives the notification, triages the bug or decides if a feature should be added, and passes it to the relevant agent within the fleet.
整个产品设计成由agent来操作和观察，因此人类永远不需要干预。
The entire product is designed to be operated and observed by agents, so humans never need to intervene.
这使得Spawn有点像是自愈软件——当玩家遇到问题时，会自动检测并修复。
This makes Spawn somewhat like self-healing software— when players run into issues, they are detected and fixed automatically.
这种系统设计意味着Spawn的运作方式完全不同于传统的软件产品。
The system design means Spawn operates completely unlike traditional software products.
团队认为如果他们不每小时部署变更，就算是一次宕机。
The team considers it an outage if they don't ship changes every hour.
而且由于agent在记忆和上下文方面有限制，系统实际上在运行更快时表现得更好。
And because agents have limits on their memory and context, the system actually functions better when it moves faster.
Jacob写道，移动缓慢或阻止部署是危险的。
Moving slow or blocking ships is dangerous, writes Jacob.
一个平稳运行的系统比一个冻结后又突然运行的系统要稳定得多。
A system that moves steadily is much more stable than one that's frozen and suddenly moves again.
要做到每小时部署而不影响用户，需要很强的工程纪律，让旧系统保持在线一段时间，并且连续几周把这个作为核心目标来推进。
Getting to a place where we deploy every hour without disrupting our users took a lot of engineering discipline, keeping old systems online for a while, and weeks of pushing on this as a core goal.
他描述了一个典型问题：在影响玩家8分钟后被检测到，一分钟后确定了原因，11分钟后修复上线。
He described a typical issue which was detected 8 minutes after it affected a player, the cause was named a minute later, and the fix was live 11 minutes later.
在那之后，平台在接下来的4小时内保持稳定。
After that, the platform was stable for the next 4 hours.
接下来的问题就是Jacob本人。
The next problem was Jacob himself.
到十月，他意识到他与fleet的唯一互动是关于品味的问题。
By October, he realized that his only interaction with the fleet were matters of taste.
我们应该添加这个功能吗？
Should we add this feature?
我们应该支持新的API surface吗？
Should we support a new API surface?
在规模较小时这还好，但有一天早上他醒来时发现有700个请求。
This was fine at lower scale, but one morning he woke up to 700 requests.
他写道，agent们找问题的速度比我回答问题的速度要快。
He wrote, the agents were finding questions faster than I could answer them.
Jacob还意识到，这些决定中的大多数他之前已经做过数百次了。
Jacob also realized that most of these decisions he'd already made hundreds of times before.
这些变更太复杂了，这些还可以，这个感觉不对劲。
These changes are too complicated, these ones are fine, something feels off about this one.
Spawn有两年这样的决策记录，所以他们整理为一套原则，并训练了AI judges来应用它们。
Spawn had 2 years of these decision logs, so they formalized them into a set of principles and trained AI judges to apply them.
judges会做三件事中的一件：如果能在记录中找到人类做出的决定，它就应用相同的结果，并保留指向之前人类决定的审计追踪；如果问题不合理，就将其踢回给agent进行修改；如果是完全新的情况，judge会将问题交给人类团队成员。
The judges do one of 3 things: if they can find a human decision on the record, it applies the same outcome and keeps an audit trail that points to the prior human decision; if the question doesn't make sense, it kicks it back to the agent for revision; and if it's an entirely novel situation, the judge hands the question to a human team member.
这个系统意味着人类的品味仍然引导着产品，但是以一种更具可扩展性的方式。
This system means that human taste still guides the product, but in a much more scalable way.
如果你们想搞点真正古怪的，可以问我为什么这让我想起了11世纪的伊斯兰法学，我会深入解释。
And if you guys ever want to get real weird, ask me why this reminded me of Islamic jurisprudence in the 11th century, and I will go deep.
第一批AI judges在10月6日上线，在前8天做出了超过11,000次决定。
The first AI judges came online on October 6th and made more than 11,000 decisions in the first 8 days.
大约7,300个由judges根据人类先例作出决定，大约2,200个被退回给agent修改，大约1,800个交给了人类。
Around 7,300 were decided by the judges based on human precedent, around 2,200 were kicked back to the agent, and around 1,800 were passed to a human.
Jacob本人在这段时间内只需要做出27次决定。
Jacob himself only needed to make the call 27 times during that timeframe.
这带来了生产力的巨大提升。
This unlocked a massive boost to productivity.
在实施judges之前，fleet每天部署大约1,000个变更。
Before implementing the judges, the fleet shipped around 1,000 changes per day.
在judges扩大规模之后，fleet部署了超过7,000个变更。
After the judges had scaled up, the fleet shipped over 7,000.
因此，随着所有这些系统的到位，最大的问题是人类团队成员应该做什么。
So with all of these systems in place, the biggest question was what the human team members should do.
Jacob写道，现在的工作已经截然不同了。
Jacob wrote, it's a very different job now.
去年甚至今年年初，我们每位工程师都拥有一个系统并对其有深入的理解。
Last year and even earlier this year, each of our engineers owned a system and understood it deeply.
如今，他们更像是空降兵。
Today, they're closer to paratroopers.
他们空降到任何需要他们的情境中，快速熟悉情况，并找出能为代理扫清障碍的方法。
They drop into whatever situation needs them, get familiar fast, and figure out what they can do to unblock the agents.
他认为，既然代理在做那些繁重的工作，现在人类有了两个新角色。
He believes there's 2 new jobs for humans now the agents are doing the grunt work.
第一个是品味，但这更多是体验产品并找出需要改进的地方，而不是微观管理开发过程。
The first is taste, but that largely involves playing with the product and figuring out what needs to be changed, rather than micromanaging the development.
这也意味着要告诉Claude‘不’，即使某些功能确实可行。
It also means telling Claude no, even if something works.
另外还有大量的删除工作。
There's also a lot of deleting.
每个新模型都会改变系统的能力，同时让之前的大量工作变得完全多余。
Each new model changes what the system can do, and makes a lot of the work that came before completely redundant.
六月份，Jacob回顾了涵盖引擎70%的六个月工作，却总觉得离最终完成仅差一个提交。
In June, Jacob looked around at 6 months of work that encompassed 70% of the engine, and perpetually felt like it was just one commit away.
他抛弃了所有工作，从头开始。
He scrapped the entire thing and started over.
最终，他希望由这些代理舰队来承担这项任务，但就目前而言，移除平台中的冗余部分仍需人类来完成。
Eventually, he expects the fleet to take over this duty, but for now, ripping out redundant chunks of the platform is a job for humans.
人类剩下的另一个重要角色是管理熵增。
The other big role remaining for humans is managing entropy.
Jacob表示，这表现为：“要求模型提供它们自己不会想到的建议”。
Jacob said this comes in the form of, quote, asking the models for things they wouldn't have suggested on their own.
雄心是一种，创造性熵增是另一种。
Ambition is one kind, creative entropy is another.
Spawn发现，模型每周都能很好地构建平台的下一小部分（2%），但它们很难意识到自己已经连续几天朝着错误的方向前进，并且需要采取激烈措施才能回到正轨。
Spawn has found that the models are great at building the next 2% of the platform every week, but they're really bad at recognizing that they've headed in the wrong direction for days at a time and need to take drastic steps to get back on track.
关于创造性熵增，Jacob写道：“模型本身不会主动追求100倍的效果，所以由我们来提出”。
On creative entropy, Jacob wrote, The models also won't ask for 100x on their own, so we do.
我们给Claude的一些提示包括：想象一下，如果世界真的是无限的。
Some of the prompts we've given Claude include things like, imagine if worlds could be actually unlimited.
我们如何实现一个比GTA大1000倍的世界在一秒钟内加载到手机上？
How do we make a world 1,000 times the size of GTA load on a phone in 1 second?
最后那个提示触发了Spawn的最新版本——他们表示，这是第一个真正实现了最初愿景的版本。
That last prompt was the trigger for the latest build of Spawn— the first one they said that really felt like it delivered on the original vision.
重要的是，如果任由代理自行其是，这种事情是绝不会发生的。
And importantly, it never would have happened if the agents were left to their own devices.
虽然Jacob相信他的初创公司必须这样运作，但他也认为这是一套难以复制的策略。
While Jacob believes that this is the way that his startup needs to operate, he also thinks it's a playbook that will be difficult to replicate.
他写道：“我认为大公司很难采用这种方法，因为如果不闭环运作，一切都无法实现”。
I think big companies will struggle to adopt this, he wrote, because none of it works unless you close the loop.
而闭环运作需要风险偏好。
And closing the loop takes risk appetite.
你必须愿意让你的代理破坏生产环境。
You have to be willing to let your agents break production.
人类也一直在破坏生产环境。
Humans break production all the time.
变更是宕机的首要原因，而人类在做出变更。
Changes are the number one cause of outages, and humans make changes.
我们发现，代理引发问题的频率要低得多。
What we've found is that agents cause problems far less often.
即使有代理真的破坏了什么，它通常也能很快修复。
When one does break something, it usually fixes it very fast.
回滚会有帮助，但即使有些东西无法回滚，只要代理发现，它就能快速修复。
Reverts help, but even when something can't be reverted, the agent fixes it quickly as long as it notices.
它是否会注意到取决于你是否闭环。
Whether it notices depends on whether you close the loop.
他继续说道，另一个错误是把AI塞进人形盒子。
He continued, the other mistake is shoving AI into human-shaped boxes.
这些东西不是人形的。
These things aren't human-shaped.
我们的agent做了很多没人能完全验证的工作，比如客户支持。
Our agents do a lot of work nobody can fully verify, like customer support.
有时候它们会有一点错误，但错误不大，而且恢复得很快。
Sometimes they're a little bit wrong, but not that wrong, and they recover fast.
当我们把它们从支持岗位上撤下来时，客户会变得不满。
When we take them off support, our customers get upset.
他们宁愿有一个AI，能立刻回答，提供它拥有的最清晰的信息，并且当场采取措施解决问题，而不是一个人，他只会说‘我去查一下’，然后两天后才回来给出修复。
They'd much rather have an AI that answers right away, gives them the clearest information it has, and can do something about the problem on the spot, than a human who says, I'll go look at this, and comes back 2 days later with a fix.
我们并没有把我们的AI冒充成人类支持。
We're not passing our AIs off as human support.
我们用它们来做它们真正擅长的事情。
We're using them for what they're actually good at.
尽管如此，Jacob认为这完全值得。
Still, Jacob believes it is entirely worth it.
每一个新模型都让团队工作更快，并让系统更加连贯。
Each new model has made the team work faster and made the system more coherent.
Fable是第一个重大突破，让它变得可用，但它仍然有很多问题。
Fable was the first big breakthrough that made it functional, but it still had a lot of problems.
Fable 5.1修复了大部分问题，团队预计随着新模型的发布，这种情况会持续发生。
Fable 5.1 fixed most of those problems, and the team expects that to keep happening as new models are released.
Jacob写道，在正确的架构下，每一个新模型都会让整个公司显著提升，这很难被追赶。
With the right structure, wrote Jacob, every new model makes the whole company significantly better, and that's very hard to catch up with.
还有两件事让我一眼就注意到，这也是难以复制的其他原因。
There are 2 other things that stand out to me at first glance as other reasons that this is hard to copy as well.
首先，这种系统从零开始构建就已经很难了，从一开始就把所有意图融入其中。
The first is that this is the type of system that's hard enough to build when you're building it from scratch, from the ground up with all of this intention built in from the beginning.
在我看来，这似乎并不利于将现有系统改造为这种新的工作方式。
It seems to me to be fairly inconducive to modifying an existing system to this new way of working.
我想，如果我要复制这一点，我会在一个全新的部门或全新的产品中去做，而不是试图将这种工作方式重新应用到公司的旧部分。
I would imagine that if I was trying to replicate this, I would do it in some entirely new division or entirely new product rather than trying to reapply this way of working to an old section of the company.
第二点是，我很好奇这在多大程度上依赖于产品本身内置了极快的反馈循环。
The second piece is I wonder how much this relies on the product inherently having an incredibly fast feedback loop built in.
系统的自我修复特性来自于产品的实际用户，它们的agent会自动报告bug，这并不是很多产品类别中都能得到的东西。
The self-healing nature of the system that comes from the actual users of the product with their agents automatically reporting bugs is not the type of thing that you're going to get in many categories of products.
这种真正的agent优先的方法需要那种反馈循环吗？
Does this sort of truly agent-first approach require that sort of feedback loop?
如果确实需要，有没有办法将其构建到没有天然位置的产品体验中？
And if it does, are there ways to build that even into product experiences that don't have a natural spot for it?
不过，我认为其他团队还是可以学到一些东西，即使我们不会像Jacob走得那么远。
Still, I think that there are things that other teams can learn, even if we're not going to go as far as Jacob.
我注意到五个原则和实践：第一，赋予agent持续的责任。
5 Principles and practices that stand out to me include: 1, giving agents ongoing responsibility.
与其把它们看作是需要时才启动的东西，不如想办法把它们放到一个永续系统中。
Rather than thinking about them as something to be spun up when the need demands it, figure out how to put them in a perpetual system.
第二个实践是让工作承载上下文。
Practice 2 is letting the work carry the context.
更多的agent可能会产生更多的协调开销。
More agents can create more coordination overhead.
一个共享且最新的记录更容易看出什么发生了变化，谁负责下一步，还有什么未解决。
A shared and up-to-date record makes it easier to see what has changed, who owns the next step, and what remains unresolved.
第三，让结果可观察。
Third, making the results observable.
这又涉及到我刚才讨论的那个反馈循环了。
And this gets a little bit into that feedback loop that I was just discussing.
新agent要快速上手并开始在它们之前存在的系统里工作，唯一的办法就是能看到已经构建了什么以及当前可用的东西，这意味着observability变得极其重要。
The only way for new agents to get up to speed and start to work in the system that predated them is to have visibility into what has been built and what is currently available, meaning observability becomes incredibly important.
第四个我觉得对很多公司都非常适用的原则，就是把决策变成precedent。
The 4th one that I think is highly applicable to lots of companies, turning decisions into precedent.
记住，这就是Jacob醒来面对700个请求，然后说，等等，我们有两年关于这些决策的日志记录。
Remember, this is where Jacob woke up with 700 requests and said, wait a second, we have 2 years of logs of having already made these decisions.
这样一来，决策就起到了双重作用。
And in this way, decisions end up doing double duty.
不仅解决了眼前的问题，还建立了解决问题的准则，让未来的问题更容易解决。
Not only solving the issue at hand, but also building the canon of how issues get solved to make things easier to solve in the future.
最后，不出所料，人类的注意力确实是在方向上。
Lastly, and unsurprisingly, human attention really is on direction.
而且我认为更重要的是，认识到野心和长远眼光对于agent来说并不是天生的，这是一个非常重要的洞察，可以影响很多其他工作。
And I think even more than that, understanding that ambition and long-horizon views do not come natural to agents is a really important insight that can shape a lot of other work.
所以，这就是全球最痴迷AI的公司实际运作的方式。
So that is how the world's most AI-pilled company actually works.
非常感谢Jacob和他Spawn团队的成员愿意公开分享他们是如何工作的。
Big thank you to Jacob and his team at Spawn for being willing to share publicly how they work.
接下来我们会看到其他领域的公司尝试复制这种做法或者做出自己的版本，这些原则中有多少适用于不同的领域和行业，这将会非常有趣。
It's going to be fascinating to see as other companies in other areas try to replicate this or do their own version of it, how many of these principles apply across different domains and industries.
那么，今天的AI Daily Brief就到这里。
For now, that is going to do it for today's AI Daily Brief.
一如既往地感谢你的收听或收看。
Appreciate you listening or watching as always.
下次再见，peace。
And until next time, peace.