← 返回任务列表

How to Build an AI-Native Company Today

242 段 · 1 位说话人 · 原片 27:35
M1
M10:00

一年前,企业AI领域的情况和现在很不一样。

A year ago, it was a very different time in enterprise AI.

当时公司还在谈论他们有多少个AI用例之类的东西。

Companies were still talking about things like how many use cases they had for AI.

现在,一年过去了,我们不再谈论用例了。

Now, year on, we are no longer talking about use cases.

事实证明,一切都可以是AI的用例。

Everything, it turns out, is a use case for AI.

事实上,在2026年,那个被期待已久、讨论很多的向agentic AI的转变真的开始了。

And in fact, in 2026, the long-awaited, much-discussed transition to agentic AI actually began.

围绕这个转变,公司经历了一个重大的转型过程,基本上所有人都还在这个过程之中。

Surrounding that, companies have undergone a significant transformation process, one that pretty much everyone is still in the midst of.

然而,当公司努力变得更AI原生时,问题是,这到底意味着什么?

And yet, as companies try to become more AI native, the question is, what does that actually mean?

那些不只是把AI和agent生搬硬套到旧流程上,而是真正用新方式做事、从头重新设计的公司,有哪些标志和特征?

What are the hallmarks and characteristics of companies that are not just glomming AI and agents onto old processes, but are really doing things in new ways, redesigning from the ground up?

虽然这些都还在不断出现,但我觉得我们现在已经开始看到一系列定义AI原生公司的特征和特点了。

While it's all still emerging, I think we're at the point where we are starting to see a set of features and characteristics that define AI native companies.

The AI Daily Brief是一个每日播客和视频节目,内容是关于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,在我们正式开始之前,先快速说几件事。

All right, friends, quick announcements before we dive in.
M1
M11:36

欢迎回到AI Daily Brief。

Welcome back to the AI Daily Brief.

今天这期很有意思。

Today we have a fun one.

我觉得现在基本上你们所有人都处于两种情况之一:要么是在大公司里,试图适应这个AI赋能的世界,成为自己的下一个版本;要么就是被那些公司请来帮助他们完成这个转变的人。

I feel like at this point, pretty much all of you are either at big companies who are trying to adapt and become the next version of themselves in this AI-enabled world, or are the people who are being hired by those companies to help them with that adaptation.

不管你坐在桌子的哪一边,所有这些转型之下都有一个核心问题:我们到底要变成什么?

Whichever side of that table you find yourself on, a core question that exists underneath all of that transformation is what are we actually transforming into?

有一个经常被提到的词叫“AI原生”。

A term that gets thrown around a lot is AI native.

这个词吸引人的一个原因是,它把那些只是简单把AI和agent生搬硬套到旧流程上的公司,和那些真正从头重新思考、利用这个新时代的公司区分开了。

Part of the attraction of the term is that it separates companies that have simply glommed on AI and agents to old processes from those who have actually rethought from the ground up to take advantage of this new era.

但AI原生的实质是什么?

But what is the substance of AI nativeness?

它特别适合做播客是因为这里面内容很多,而且很多都值得讨论。

Part of what makes it great for a podcast is that there are a lot of things and many of them are debatable.

这就是Alex Lieberman登场的时候。

Enter Alex Lieberman.

Alex是10x Labs的创始人,这家公司帮助现有企业转型为AI原生公司。

Alex is the founder of 10x Labs, which is a company that helps transform existing companies into AI-native companies.

在那之前,他是Morning Brew的创始人。

And before that, he was a founder at Morning Brew.

你可能会想到,在Morning Brew的经历意味着他经常会产出很棒的内容。

As you might imagine, that experience at Morning Brew means that he is often a great source of content.

启发这一集的那篇帖子,其实直接来自他最近在X上的账号,叫《AI原生公司的30个特征》。

And the post that inspired this episode is actually a direct crib from his X account recently, 30 Features of an AI-Native Company.

我觉得好玩的做法是,先用Codex和GPT Image 2把它做成一个漂亮的1950年代复古未来主义风格的演示文稿,然后我们一条一条过这些特征。我会分享Alex帖子里关于AI原生的那些方面,再加上我自己的想法、补充和不同意见——虽然我觉得很多地方我也未必不同意——以及我在和企业合作中看到的一些观察。

What I thought would be fun after turning it into a beautiful 1950s retro futurist themed presentation with the help of Codex and GPT Image 2 is to go through these features one by one where I will share the aspect of AI nativeness that Alex posted about and then add any thoughts, qualifications, disagreements, although I don't think that I necessarily disagree in a lot of places, and other observations that I've seen in my work with enterprises as well.

我得说明一下,我觉得这些特征没有什么特定的顺序。

Now, I should note that I don't think that this is in any particular order.

实际上,你完全可以看出来这不是AI生成的,因为它读起来更像是意识流,坦白说,这比你现在平时看到的大多数内容都更意识流,让人感觉挺新鲜的。

In fact, you can very much tell that this is not AI generated, as it reads much more like a stream of consciousness than frankly most of the content that you're used to seeing these days, which is kind of a breath of fresh air.

AI原生公司的第一个特征是为每个流程制定蓝图,用Alex的话说,就是为整个业务创建一个逐功能的流程蓝图。

First feature of an AI-native company is to blueprint every process, to create, in Alex's words, a function-by-function process blueprint of the entire business.

我之前说过我不会太不同意,这里我也不是要反对。但有一个领域是,虽然我不反对这样做,我认为很多公司做这件事背后的理由,其实正让他们走上错误的路。

Now, I said I wasn't gonna disagree much, And I'm not exactly going to disagree here, but this is one area where, although I don't disagree with doing this, I think the reasoning behind it for a lot of companies is actually leading them down the wrong path.

为流程制定蓝图之所以有价值,换句话说,就是去画出工作实际上是怎么完成的,其中一个原因是,现在很多这样的信息都被锁在人们的脑子里。

Some of the reasons that it's valuable to blueprint processes, in other words, to map out how work actually gets done, is that a lot of that information right now lives locked inside people's heads.

有很多细微差别和边缘情况,人们一直以来都是自己处理,也许是通过走廊里的随机交谈或Slack上的私下聊天来口口相传,但这些从未真正进入操作手册。

There are a lot of nuances and edge cases that people have been handling on their own forever, And which are perhaps transmitted person to person through random spoken meetings in the hall or side chats on Slack that don't ever find their way into actual operating manuals.

这让你引进协助工作的AI和智能体处于劣势,因为它们要从头开始重新创造一切。

That puts the AI and agents that you're bringing in to assist with the work at a disadvantage because they're recreating things from the ground up.

因此,拥有更好的当前工作流程图,在你围绕AI和智能体重新设计组织时,是非常宝贵的背景信息。

And so having better maps of how work currently gets done is an incredibly valuable piece of context as you redesign your organization around AI and agents.

所以这部分我同意。

So all that part I agree with.

我担心的是一种我经常看到的流程映射的内在假设。

Where I get concerned is around an inherent assumption of process mapping that I see pretty often.

这种假设是,智能体将像人类一样做事。

The assumption is that agents are going to do things the same way that humans do.

我认为这很不太可能成真。

I think that that's very unlikely to be true.

事实上,我认为人为地限制智能体按照旧工作流程的模式做事,在很多情况下是错误的做法,而不是例如给他们目标,阐明他们能做什么和不能做什么的边界,然后让他们自己摸索。

And in fact, I think artificially constraining agents to do things along the pattern of an old workflow is in many cases the wrong approach, as opposed to, for example, giving them the goal and articulating the guardrails of what they can and can't do and letting them figure it out from there.

再说一次,这并不意味着流程映射没有价值,但我们必须理解并做好准备:未来做某事的最佳方式不一定只是过去方式的更高效版本。

Now, again, that doesn't make process mapping not valuable, but we have to understand and prepare for the reality that the best way to do something in the future will not necessarily just look like an efficient version of the way that we did it in the past.

AI原生公司的第二个特征是非常没有争议的,那就是给每个人一个“日常工具”,也就是说,组织中的每个人都可以在工作中使用像Grokbot、Claude Cowork或ChatGPT这样的日常工具。

Feature number 2 of an AI-native company is highly uncontroversial, and that is to give everyone a daily driver, i.e., everyone in the organization gets to use a daily driver harness such as Grokbot, Claude Cowork, or ChatGPT at work.

有趣的是,九到十个月前,当你看到“日常工具”这个词时,你会以为它只是指访问一个前沿模型。

Now, what's interesting here is that 9 months ago, 10 months ago, When you saw the term daily driver, you would have assumed you meant just access to a frontier model.

换句话说,人们可以使用ChatGPT或Claude。

In other words, people have access to ChatGPT or Claude.

但Alex谈论的是一个特定的工作工具,一个模型在其中运行的环境,专门为高级知识工作和编码而设计,无论是对软件工程师还是非软件工程师来说,他们现在在完成工作时也会使用代码。

But what Alex is talking about is a specific work harness, an environment in which models operate that is designed specifically for advanced knowledge work and coding, whether that's for software engineers or for non-software engineers who are now using code as part of the way that they do their job.

习惯使用一个工具意味着习惯使用上下文。

Getting comfortable with a harness means getting comfortable with context.

这意味着能够理解并组织技能,以及知道如何配置不同工具的访问权限。

It means being able to understand and organize skills, as well as understanding how to provision access to different tools.

就像我说的,这没什么争议。

Like I said, nothing controversial here.

我要指出的一点是,我认为我们会越来越多地看到人们自己构建工具,通常基于开源基础,比如刚刚推出的DeepSeek的工具,因为他们想要最大的灵活性,不想处理像投资Cursor然后被卖掉,之后由于收购再也不能通过它访问某些模型这类事情。

The one thing that I will note is that I think that we are going to increasingly see people rolling their own harnesses, often on the basis of an open source foundation, for example, DeepSeek's harness that just came out, because they're gonna want peak flexibility and they're not gonna wanna deal with things like investing in Cursor only to have it sold and no longer being able to access certain models through it because of that sale.

第三个特征,很多评论中提到,用AI现在最喜欢的词来说,是“承重墙”,支撑着其他特征。

Feature number 3 is one that many in the comments noted is, to use AI's favorite term right now, Load-bearing for the rest of the features.

核心理念是构建一个统一的智能层,将结构化和非结构化数据、文档以及业务逻辑聚合到一个可查询的单一事实来源中,并在此基础上构建智能体工作。

The idea is to build one intelligence layer, to aggregate structured and unstructured data, documents, and business logic into a single source of truth that is queryable and that agentic work can be built on top of.

毫无疑问,AI原生组织会擅长组织智能体工作所需的上下文。

It is unquestionable that AI-native organizations are going to get good at organizing the context their agents need to work.

上下文管理现在是,未来也将继续是这次组织转型时期的一个重要学科。

Context management is and will continue to be a major discipline in this new org transformation period.

如果要说我有什么异议,纯粹是为了对话有趣,我其实更喜欢用“网状结构”或“格状结构”而不是“单一层”的比喻,因为我认为特别是在大型组织中,试图为所有事情建立一个单一事实来源,而不是可以有接口相互连接、智能体可以遍历的多个事实来源,甚至可能发现并尝试在人工支持下调和事实来源之间的差异,这或许更能准确反映顶尖大型组织未来的情况。

To the extent that I have quibbles here, which is really just for the sake of interesting conversation, I kind of like the metaphor of a mesh or lattice rather than a single layer, because I think especially as you get into larger organizations, trying to have a single source of truth for everything rather than sources of truth that can interface with one another and that agents can traverse, perhaps even uncovering and trying to reconcile with human support differences in sources of truth, is perhaps a more accurate reflection of how this is going to look with the biggest organizations.

但显然实质性的观点仍然成立。

But obviously the substantive point underneath remains.

AI原生组织的第四个特征,在过去几个月里一直是本节目讨论的热门话题,那就是使用模型路由来优化整个业务中每个成功任务的成本。

Feature 4 of agent-native organizations is one that has been a big subject of conversation on this show for the last few months, which is about using model routing to optimize cost per successful task across the business.

我认为这是对的,但这不仅仅是任务路由的问题。

I think this is right, but it's not just a matter of task routing.

我认为任务路由是整个模型架构的一部分,这个架构旨在适应不同类型的任务,具有适应性和灵活性。

I think that task routing is part and parcel of an overall model architecture that is designed to be adaptable and flexible to different types of tasks.

我认为在某些情况下,它可以简化为使用一个路由器,但在很多情况下,也会涉及一个更大的架构,基于将任务难度与模型能力匹配的想法。

I think in some cases that will be reducible to using a router, But in many cases will also implicate a larger architecture based around that idea of matching task difficulty to model capability.

第五个特征,他描述的方式很有意思。

Number 5 is an interesting one in the way that he frames it.

他说AI原生组织会将上下文视为代码。

He says that AI-native organizations will treat context as code.

他们会保持架构文档和惯例的更新,同时将有计划的前期准备作为操作纪律的一部分。

They'll keep architecture documents and conventions updated while making diligent upfront planning part of the operating discipline.

换句话说,他们不会把上下文只当作必需的背景信息,而是当作智能体构建的实际基础。

In other words, they will treat context not just as the background info that is required, but as the actual foundations upon which agents are building.

这种区别很微妙但很重要,我觉得它反映了一个观点:这种AI原生性不仅体现在运营流程上,还关乎思维方式的转变。

It's a subtle but important distinction, and I think reflects the idea that a lot of this AI nativeness is not just in operational process, but also comes down to mindset shifts as well.

说到这个,第六个特征完全是关于思维方式的。

Speaking of, feature number 6 is totally about mindset.

愿意每三个月就抛弃你已构建的一切,从第一性原理重新构想工作流程。

Be willing to throw away everything you've built every 3 months and reimagine the workflows from first principles.

我觉得“抛弃一切”可能有点夸张,但确实我们需要在设计这些新系统时,假设变化而非停滞。

Now, I think throwing away everything you've built might be slightly dramatic, But it is absolutely the case that we need to design these new systems for assumptions of change rather than assumptions of stasis.

系统更新的频率——不管是三个月、六个月、九个月还是一年,或者是一种持续更新加上更大规模的周期性重构——无论实际情况怎样,变化就是常态,组织必须比现在更习惯于这种变化。

The frequency with which systems will need to be updated, whether it's 3 months or 6 months or 9 months or a year, or a sort of perpetual update marked by bigger periodic reimaginings, however it actually plays out, change is the name of the game, and it's something that organizations are gonna have to get way, way more comfortable with than they are today.

这包括不要迷恋你几个月前想出的那个绝妙方法,因为如果AI公司做得好,技术进步就意味着我们会有更简单或更强大的新方法。

And that includes not getting attached to the exciting way that you figured out to do something just a couple of months ago, Because if the AI companies do their jobs well, advances should mean that we have new ways to do things that are either easier or more powerful.

第七个特征是关于组织内的信息共享。

Feature number 7 is about info sharing across the organization.

Alex说,AI原生组织会使用技能分发系统来管理智能体行为,并通过在整个工作流程中触发一致的技能来提升token效率。

Alex says that AI-native organizations will use a skills distribution system to manage agent behavior and improve token efficiency by triggering consistent skills throughout the workflow.

换句话说,组织分发的是技能,而不仅仅是提示词。

In other words, organizations will distribute skills, not just prompts.

我认为这也代表着一个更大的转变,即智能体管理这一学科开始走上前台。

I think this as well is emblematic of a bigger shift, which is the discipline of agent management coming to the fore.

技能是智能体系统的关键,因此在整个组织内——而非局限于个人——去改进、共享、访问这些技能,将变得越来越重要。

Skills are a key aspect of agentic systems, and so having ways to improve them, share them, access them, etc., across the organization, not just within the silo of any individual, is going to be increasingly important.

第八个特征是关于技术人员与非技术人员之间关系的变化。

Feature number 8 is about the changing relationship between technical and non-technical team members.

我觉得我们已经基本过了那个阶段——人们会以为我们谈“氛围编程”或把代码用于知识工作时,就意味着市场营销和人力资源的人突然要取代现有工程师成为软件开发者。

I think we're mostly past the days where people think that when we talk about vibe coding or using code for knowledge work, we somehow mean that all of a sudden the folks in marketing and HR are going to be the software developers instead of the existing engineers.

并不是这样。

That's not what's happening.

但实际情况是,那些知识工作者首次能够自己构建东西来辅助工作。

But what is happening is both that those knowledge workers are for the first time able to build things themselves as a way to help do their job.

其次,他们比过去更有能力参与到产品和工程讨论中。

And second, they have more ability to contribute to product and engineering discussions than they might have in the past.

Alex清单中AI原生公司的第八个特征是将意图与实现分离,把技术实现与高层规格分开,这样非技术人员就能以一种智能体可转化为实施计划的格式做出贡献。

Feature 8 of AI-native companies from Alex's list is to separate intent from implementation, to keep technical implementation separate from high-level specifications so non-technical staff can contribute in a format agents can turn into implementation plans.

我认为实现这一点的方式之一,是当我们看到更多智能体从共享空间被触发时,这自然就是某些协作发生的地方。

I think one of the ways that this will play out is as we see more agents triggered from shared spaces, that's a natural place for some of this to happen.

例如,当Claude Tag发布时,其中一个引人注目的点是Anthropic技术团队的成员说,他们现在就是用这种方式启动很多构建工作的。

For example, when Claude Tag was announced, one of the more remarkable things about it was members of the Anthropic technical team saying that that was how they initiated a lot of their building now.

说到“很多”,我脑子里记得的数字大概是60%或者类似夸张的数字。

And by a lot, I mean the number that sticks out in my head is like 60% or something ridiculous like that.

如果智能体从共享空间被触发,这就为不同的人创造了更多参与那些对话的机会。

If agents are being triggered from shared spaces, that creates more of an opportunity for different people to contribute to those conversations.

但这本身也会带来不同类型的负担和新的系统需求,这就是Alex在这里说的。

But that in and of itself is going to create a different type of burden and new types of system requirements, which is what Alex is talking about here.

第九个特征是成本效率方面的:AI原生公司会把每次被接受的pull request的成本作为关键软件指标,并通过更好的token效率来降低它。

Feature number 9 is a cost efficiency feature where AI-native companies will make cost per accepted pull request a key software metric and drive it down through better token efficiency.

我认为我们才刚刚开始摸索什么是智能体交付的关键指标。

I think we are just at the beginning of the period of figuring out what the key metrics of agentic delivery are.

对我们来说,目前清楚的是,无论我们最终确定什么指标,很可能都包括某种完整性概念以及单位完整性的成本,这样才能更公平地比较模型和工具组合。

And what's clear to us at this point is that whatever the metrics we land on are, it's likely that they include some sense of completeness as well as cost per completeness in order to be able to better compare model-harness combos in a more apples-to-apples kinda way.
M1
M113:21

现阶段,问题已经不是公司是否在积极使用AI了。

At this point, it's no longer a question of whether companies are actively using AI.

而用好AI,则是另一回事。

Using it well, on the other hand, is a whole different story.

不过,Robots and Pencils这家公司,我可以指出来说它确实是为此时代而建的。

Robots and Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用型AI工程公司,直接与客户合作解决对业务至关重要的问题,而不是做一些停留在PPT上的实验。

They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每个合作项目都是从客户实际需要的成果倒推开始的。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域里区分真正的AI工程和那些噪音,这就是关键所在。

If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker.
M1
M114:32

特性10可能是很多组织进展最大的特性之一,就是构建agent原生的开发系统,让成群的编码agent进行规划、编写、测试、审查和发布代码,而人类则定义意图和验收标准。

Feature 10 is perhaps one of the features that many organizations are the farthest along with, which is to build agent-native development systems, letting fleets of coding agents plan, write, test, review, and ship code while humans define intent and acceptance criteria.

这比一年前争议少多了,但显然还有很多组织在使用传统流程,尽管未来几年很可能发生不可避免的转变。

This is a lot less controversial than it would have been a year ago, but obviously there are still many organizations that are using a traditional process despite there likely being an inevitable shift that will happen in the coming years.

特性11再次触及了token效率这个子主题,将工作业务分为规划阶段和执行阶段,使用更高成本的模型进行大量规划,然后用更便宜、更快的模型来执行。

Feature 11 once again gets at the sub-theme of token efficiency, dividing the business of work into planning phases and execution phases, using higher effort models for heavy planning and then executing with cheaper and faster models.

好处是,如果我们在特性4(关于设计高效的token架构)上做好了,这种劳动分工自然就会从这些系统中产生。

The good thing will be that if we've done our job with feature number 4, about designing efficient token architectures, this is the type of division of labor that should naturally fall out of those systems.

特性12乍一看术语挺多的。

Feature number 12 feels at first glance fairly term-heavy.

它是使用CLI工具来解析markdown文件中的元数据并遍历依赖关系,让agent能够精确控制输入的token使用量。

It's use CLI tools to parse metadata in markdown files and traverse dependency relationships, allowing agents to be precise about input token usage.

但实际上,这是一种设计更token高效系统的具体技术方式。

But really, this is a technically specific way of designing a more token-efficient system.

基本上,Alex主张的是以这样一种方式组织公司知识:agent只加载它们需要的部分,而不是整个知识库。

Basically, what Alex is arguing for is organizing a company's knowledge in a way that agents can only load the slice they need rather than the whole thing.

这感觉和渐进式披露的概念有关,这是一种信息架构模式,将复杂性逐步、按顺序展开,以免在agent工作时产生过多的上下文开销。

It feels related to the idea of progressive disclosure, which is an information architecture pattern where complexity is unpacked gradually and in sequence in order to not create too much context overhead when an agent is working.

现在,除了让agent工作得更好之外,显然还有成本方面的问题。

Now, in addition to just making agents work better, there are obviously cost dimensions of that as well.

我很高兴Alex把它包括进来,因为到目前为止我们都是在非常高的层面上讨论。

And why I'm glad Alex included it is that so far we've been operating at a really high level.

但这正是一个开始深入细节的例子。

But this is an example of where you start to get granular.

并且实际采取不同的做法,即确保agent可以参考的元数据——用来判断某个内容是否有用——正好位于它能访问的知识文件的顶部,这样它就不会在不必要的内容上浪费上下文窗口。

And actually do things differently, i.e., making sure that the metadata that an agent can reference to understand whether something is useful is right there at the top of knowledge files that it has access to so that it doesn't waste context window on things it doesn't need.

特性13开始涉及组织中的具体部门。

With feature number 13, we're starting to get into specific parts of the organization.

他建议财务将更加持续地运行,把会计和记录保持转向持续流程,以更紧密的节奏重新调整预测。

He suggests that finance will run more continuously, moving accounting and recordkeeping towards continuous processes, resetting forecasts on a much tighter cadence.

OpenAI的CFO Sarah Friar最近确实写过她如何在OpenAI内部做到这一点,以及做出这种改变需要怎样的思维转变和技术纪律。

OpenAI CFO Sarah Friar actually recently wrote about how she had done this inside OpenAI, and about what a mindset shift and a technological discipline it took to make this sort of change.

Alex还建议,在AI原生的组织中,其他部门——特别是运行整个公司的更大的agent操作系统——能够访问这些财务模型,这样在制定战略和战术时,它们可以成为逻辑的一部分。

Alex also suggests that in AI-native organizations, other parts of the organization, and specifically the larger agentic operating system through which it runs, will have access to those financial models so that that can be part of the logic as strategy and tactics are designed.

AI原生公司的特性15是公民开发者SDLC(软件开发生命周期)。

AI-native company feature 15 is the citizen developers SDLC.

这是Alex在其他地方也谈到过的方法,它让非技术员工能够使用编码工具构建一个解决方案,并从想法到生产实现,其中内置了公司的治理、访问权限、版本控制和软件约定。

It's an approach that Alex has talked about elsewhere as well, that enables non-technical employees to take a solution that they are building with coding tools from idea to production with the company's governance, access, versioning, and software conventions built into it.

这基本上是一个协调非工程师所做的事情与工程师构建方式的过程。

It's basically a process of reconciling the things that the non-engineers are making with the way that engineers build.

再次强调,这绝不是以任何形式取代软件工程,而是为了让人们为自己、自己的团队甚至某些客户(根据他们接触的客户生命周期部分)构建的新事物,能够与工程组织无缝衔接。

Again, not with the idea of replacing software engineering in any way, shape, or form, but in order to have the new things that people are building for themselves or their teams or even some segment of customers based on the part of the customer lifecycle that they touch, to have that all contiguous with the engineering organization.

这再次反映了在这个新的AI原生空间中,组织不同部分之间正在变化的关系。

It reflects again that shifting relationship between different parts of the organization in this new AI-native space.

特性16涉及了我们在本周早些时候在节目中分享的网络研讨会中讨论的循环工程。

Feature 16 gets to the sort of loop engineering that we covered in the webinar that I shared on the show earlier this week.

让非工程工作流通过外部性能指标和内部评估从先前的运行中学习,从而实现自我改进。

Make non-engineering workflows self-improving by learning from previous runs through external performance metrics and internal evaluations.

循环的思想是,我们不是给agent提示,而是给它们一个目标以及关于它们能做什么的限制,并设计一个过程,让它们可以反复循环直到达成目标。

The idea of loops is that instead of prompting agents, we give them a goal and bumpers around what they can do and design a process that they can loop through over and over again until they achieve that goal.

循环的必要条件之一是某个可验证的成功指标,它是客观的而非主观的。比如说,我需要在这次测试中达到X%的结果,这比“我们的界面需要看起来好看”要更可定义,是一个结果目标。

One of the necessary requirements of a loop is some verifiable success metric that is objective rather than subjective, i.e., I need to achieve an X percentage result on this test is a lot more definable, an outcome goal, than is our interface needs to look good.

AI原生组织将擅长创造那种清晰的成功指标,不仅针对简单的确定性任务,而且针对更广泛的知识工作任务,这些任务本身不一定具备那种成功标准。

AI-native organizations are going to be good at creating those sort of clear metrics of success, not just for the easy deterministic tasks, but for the broader array of knowledge work tasks that don't necessarily have that sort of success criteria built in natively.

特性17实际上包含两部分。

Feature 17 is really 2 parts.

第一个是用某种 AI 投资回报率框架,了解组织对其 AI 项目的期望,并能据此进行衡量。

One is to use some AI ROI framework to have an idea of what the organization is looking for out of its AI efforts and to be able to measure against that.

第二部分是 Alex 对如何建立这套体系比较主观的看法,他推荐采用实验、规模化、优化几个阶段,并在基础设施、创新和效率上分别下注。

The second part is a little bit more opinionated from Alex about the way to set that up, with his recommendation being experimental scaling and optimization phases and bets placed across infrastructure innovation and efficiency.

不管最后你用哪些阶段,以及怎样组织不同类型的工作,我觉得一个重要的建议是:要有一个复杂的 ROI 架构,能理解每个项目的目标各不相同,但组织仍然能在同一个框架下对不同项目进行评判。

Whatever the phases that you end up using and the way that you organize different types of efforts, I think the big recommendation here is to have a complex ROI architecture that can understand the goal of different efforts as being different from one another, but the organization having the ability to judge them even if they are different, all within the same framework.

特性 18 就是我之前说的「奇异博士」理论在代理工作上的实现。

Feature 18 is my Doctor Strange theory of agentic work come to life.

核心想法是,AI 原生组织会用代理集群部署大量、大量、大量的付费营销创意变体用于测试,可能是几百个,也可能是几千个。

The idea is that AI-native organizations will use agent swarms to deploy many, many, many, perhaps hundreds, perhaps thousands of paid marketing creative variations for testing.

在增加广告支出之前先做这些测试。

Before increasing spend on ads.

最初思考这个问题时,我低估了一件事(顺便说一句,我现在仍然觉得这是完全不可避免的),那就是短期内的计算限制会制约这种方法的可行性。

One of the things that I underestimated when I was first thinking about that, which by the way I still think is completely inevitable, was the way in which compute constraints in the short term would limit the viability of that sort of approach.

现在我们已经跨过了能力门槛——很多模型已经足够好,能真正完成这类创意工作,而且六个月后它们的价格会比现在便宜得多——我觉得到那时,这种实验就会开始变得比较常见了。

Now that we've crossed this capability threshold where many, many models are good enough right now to actually do this sort of creative work and will be vanishingly cheaper than they are right now 6 months from now, I think that's when you'll start to see this sort of experimentation become a little bit more normalized.

我觉得这会是特别、特别有意思的一件事。

And I think it's going to be super, super interesting to see.

有趣的是——这远远超出了我们这次讨论的范围——我几乎看到了一个营销杠铃:一端是「奇异博士」那种疯狂的代理集群,另一端是完全依赖人类品味、不看数据的品牌营销活动。

Fascinatingly, and it's way beyond the scope of this particular conversation, I almost see a marketing barbell where you are going to have just Doctor Strange crazy agentic swarms on one end of the spectrum and utter number-denying human taste for brand campaigns on the other.

基本上就是 Rick Rubin 在一端,机器在另一端,而且不知怎么搞的就是能行得通。

Basically Rick Rubin on one side and machines on the other, and somehow it'll work.

特性 19 是另一个具体的营销建议:每周审计、重写和生成针对 SEO 和 AEO 优化的文章。

Feature 19 is another specific marketing recommendation of auditing, rewriting, and generating SEO and AEO optimized articles every week.

然后衡量哪些奏效了。

Then measuring to see whether any of it worked.

我觉得更广泛的想法是,既然内容生产的成本已经降下来了,以内容为核心的策略就有很多特别有意思的实验空间。

I think the broader idea is that there is just so much interesting room for experimentation with content-based strategies now that the cost of producing content has gone down.

跟很多这些建议一样,这里的思路是:AI 原生组织不只是把事情做完,而是围绕这些事情建立学习系统,以便未来做得更好。

Like so many of these recommendations, the idea here is that the AI-native organization is not just going to do the thing, but to build learning systems around the thing to do it even better in the future.

特性 20,我们进入网络安全领域。过去几个月已经证明,这个领域极其重要。

With feature 20, we're getting into cybersecurity, something that has obviously proven itself to be extraordinarily important over the past several months.

Alex 认为,AI 原生组织会用 AI 对抗 AI,构建代理式网络安全系统来抵御 AI 驱动的威胁。

Alex suggests that AI-native organizations are going to fight AI with AI using agentic cybersecurity systems built to defend the organization against AI-powered threats.

我觉得这绝对没错,但要搞清楚组织以及合法防御方到底能拥有什么样的网络安全能力,以及这些能力如何配置,还有大量问题需要解决。

Now, I think this is absolutely true, but boy, is there a lot to figure out about exactly what type of cybersecurity capabilities organizations and legitimate defenders are going to have access to and how that's going to be provisioned.

既然我们已经跨过了一些关键的网络安全门槛,在短期内,这些将成为 AI 公司和政府面临的最重要的设计及政策问题。

These are going to be some of the most important design and policy questions for AI companies and governments in the immediate term now that we've crossed some of these critical cyber thresholds.

我们进入最后三分之一了,我接下来会稍微加快节奏。

We're in our final third, and I'll pick up the speed a little bit from here.

特性 21 是另一种 token 效率和成本管理的方式:将强化学习环境与第一方数据结合,对开源模型进行微调,用于那些需要以合理成本实现前沿性能的高量级流程。

Feature 21 is another approach to token efficiency and cost management, combining a reinforcement learning gym with first-party data to fine-tune open-source models for high-volume processes that need state-of-the-art performance at reasonable cost.

基本上,我们现在所处的世界,有大量可定制、可后训练的开源权重模型,它们非常接近前沿水平,这为 AI 原生组织带来了很多新机会。

Basically, we now live in a world where the prevalence of customizable and post-trainable open weights models that are very near the frontier opens up a lot of new opportunities for AI-native organizations.

其中很多组织会发现,这种模型管理方式对它们确实有用。

Many of which will find that this sort of model discipline is actually going to be useful for them.

不过我得说,我不认为所有组织都会突然开始自己动手训练模型。

Still, I would say that I don't believe that every organization is all of a sudden going to be rolling their own models.

我觉得各个实验室和云计算巨头会想出很多很多办法来解决成本效率问题。

I think that there are going to be lots and lots of ways that all the labs and hyperscalers try to deal with cost efficiency.

但如果你是一个确实拥有这种技术能力的组织,那这绝对是一个你现在可以去探索、借此获得优势的方向。

But to the extent that you are an organization that actually has this technical capability, it's definitely a place that you could be exploring to get an edge right now.

特性 22,我们称之为「人类三明治」——在大多数流程的首尾环节保留人类判断和参与。

Feature 22, we'll call the human sandwich, keeping human touch and judgment at the first and final mile of most processes.

如果说即使在 AI 原生组织中,工作三明治的两端也应该一直有人类存在,那我完全同意。

If the argument is that even in AI-native organizations, there should always be people on either side of the work sandwich, that I agree wholeheartedly with.

我比较不确定的是,在流程中间,我们该在哪些环节恰到好处地引入人类干预。

What I'm less sure about is where we'll find the right intervention points for humans in the middle of processes.

是的,会有很多流程几乎完全靠代理完成,但我不认为我们已经确切知道哪些模式是最合适的。

Yes, there will be many where it pretty much all happens agentically, but I don't think that we know exactly the right patterns.

所以要说这种太笼统的观点,我其实很难有太多信心。

So to speak of this one too generally feels harder for me to have a lot of confidence around.

第23条特征,我觉得争议性就小一些了。

Feature number 23, I think, is a little bit less arguable.

就是把评估当作核心基础设施来建设。

Which is the idea of making evals core infrastructure.

Alex举的具体例子是,每当新模型出来时,用一套固定的机制去测试它们的成本和性能,看是否符合公司的核心流程。

Now, the specific example Alex gives is whenever new models arrive, use a standing apparatus to test their cost and performance against the company's core processes.

但我认为,现在大家其实或多或少都进入了评估这个领域。

But I just think everyone is now in the eval business more generally.

再回到整个架构——用循环代替提示——你必须建立一套基于某种标准的评估机制,否则循环根本跑不起来。

Again, going back to the entire architecture of looping instead of prompting, you have to basically build in evaluation against some standard or else the loops don't work.

用评估的思维来思考问题,本来就是这门学科的新组成部分。

Thinking in terms of evals is just a new part of the discipline.

而且它会渗透到组织中一些你意想不到的角落。

And it's going to get to some weird parts of the organization that you wouldn't necessarily expect.

第24条特征,在我看来也同样毫无争议,那就是人人都是建设者,AI原生组织会把建设能力纳入每一个岗位,特别是高管层。

Feature 24, again, totally unarguable to me, is that everyone is a builder and that AI-native organizations will make building part of every role, including and especially perhaps those held by C-level executives.

我不确定每个知识工作者都会从自己做工作,变成管理一群替他们工作的代理。

I am not sure that every single person and every single knowledge worker go from doing their work to managing agents that do their work.

这肯定会占据很大一部分工作量,但要说全部都是,那范围也太广了。

I think that'll be a big chunk of work, but again, saying all of it is a pretty big swath to cover.

我更有信心说的是,具备建设能力——会写代码、做原型、开发产品、完成工作——现在已经成为一项关键技能了。

What I feel much more confident saying is that the capacity to build, to use code, to develop prototypes, to develop products, to do your work is now a critical capability.

当然,随着时间的推移,人和他们所用的代码之间的距离会越来越远,因为产品会逐步抽象掉技术细节,但这并不会让人们不再成为建设者,只要他们还能通过建设来解决问题、创造新机会。

Now, of course, over time, the space between people and the code that they use will get farther and farther as products obscure and abstract the technical details away, but that won't make people less of builders if they can still build things to solve their problems and create new opportunities.

这一点既是实际行动,也是思维方式的转变,而且对企业如何组织自身有相当大的影响。

This one is both practical action and mindset shift and has some pretty big implications for how enterprises even organize themselves.

第25条特征,可以说是第1条特征的双胞胎。

Feature 25 is kind of the twin of feature 1.

记录一切值得学习的东西,因为组织没有捕捉到的信息,就无法转化成AI能处理的工作。

Record everything worth learning from, because what the organization does not capture cannot be turned into AI-enabled work.

这也涉及到评估和上下文的传递。

This also gets at evals, context transmission.

我们只是需要进入一个捕获更多信息的范式,我想这会让现在每个Zoom会议里那75个会议纪要记录员很开心吧。

We just need to live in a paradigm of capturing a lot more, much to the delight, I assume, of the 75 meeting note-takers that show up in every Zoom call you have now.

第26条特征我很喜欢,因为大家讨论得还不够多。

Feature 26, I love because people don't talk about this enough.

治理常常被视为创新的阻碍,但AI原生组织会把治理当作转型的合作伙伴。

Governance is so frequently seen as a blocker of innovation, but AI-native organizations are going to treat governance as a transformation partner.

换句话说,他们会把治理看作一种释放创新的途径。

They're going to treat it, in other words, as a way to unlock innovation.

这只有在法务、HR和IT部门与AI议程的负责人紧密协作,共同制定既能解决问题又能释放新工作类型的赋能政策时,才能实现。

That's only going to work if you have legal, HR, and IT work in lockstep with the owners of the AI agenda, helping to design the enabling policies that address issues while unlocking new types of work.

我认为,这会在很大程度上成为真正优秀的AI原生组织与那些仍然把AI生硬地嫁接在旧工作方式上的组织的关键区别。

This, I think, in many ways will be a key hallmark of truly great AI-native organizations versus those that are still kind of bolting AI onto old ways of working.

第27条特征呼应了持续转型的理念。

Feature 27 hearkens back to the idea of constant transformation.

AI原生组织会倾向于先于别人颠覆自己。

AI-native organizations will maintain a bias towards disrupting the company before someone else does it for you.

这不仅仅关乎你如何完成当前的工作,更深入到你应该做什么工作。

This extends beyond just how you do your current work, but cuts all the way to what work you should be doing.

我认为AI带来的巨大变革,不仅仅是提高完成现有任务的效率,更在于发现那些你之前可能没涉足、但现在AI让你有能力去做的、既正交又一致的新机会。

I think a lot of the big transformation of AI is not just going to be in the efficiency with which you do today's tasks, but finding those sort of orthogonal and aligned opportunities that you might not have gotten into yet, but which AI now enables you to go do.

第28条特征,在技术上与“治理是赋能而非阻碍”这一点相对应,可以总结为“先建护栏,再做功能”。

Feature 28 is kind of the technical twin to governance as unlock instead of governance as blocker, and you could sum up as guardrails before features.

代理会继承提问者的权限,这些权限是在数据层强制的。

Agents inherit the permissions of whoever is asking, and those permissions are enforced in the data layer.

你一开始就把护栏建好,以后就不需要每次重新争论了。

You build the guardrails in, you don't have to relitigate it every time.

功能29本身就可以单独做成一期节目,它的核心思想是——自主权必须靠挣来。

Feature 29 could be an entire show on its own, and it's the idea that autonomy must be earned.

这些复杂而强大的智能体不会一开始就被赋予完全自主权,而是像爬梯子一样——先是观察,然后提建议,接着在批准后行动,最后自主行动——直到它们能够在既定边界内运行完整的流程。

These sophisticated, powerful agents are not given full autonomy right away, but climb a ladder— observation, suggestion, acting with approval, acting alone— until they can run whole workflows inside a defined boundary.

即使我们想识别所有东西,也不一定非要一次性全部完成。

Even if we're trying to identify everything, we don't have to identify everything all at once.

有一句老话说‘一分预防胜过十分治疗’,我觉得这里非常适用。

An old adage about an ounce of prevention being worth a pound of cure I think is very applicable here.

功能30再次强调了在这个新的智能体时代,围绕你所做的一切的信息系统——它追踪每个输出到对应的提示词、模型数据和审批人,从而让人类反馈能落实到具体的东西上。

Feature 30 once again is about the information system that surrounds everything that you do in this new agentic era, tracing every output to its prompt, model data, and approver so human feedback attaches to something specific.

而不是模模糊糊觉得哪儿不对劲。

Not to a vague sense that something is off.

我们有能力创建更全面、更复杂的系统来围绕我们的工作。

We have the ability to create much more comprehensive and complex systems that surround our work.

而且如果我们这样做,会带来各种各样的好处,从而改善后续的工作。

And if we do so, it comes with all sorts of benefits that improve the next work to be done after that.

以上就是亚历克斯的清单,里面有很多值得深思的内容。

So that is Alex's list with a lot of great food for thought in there.

至于遗漏了什么,我在评论中看到的最佳答案之一来自很多人,但可以概括为宾迪·贾米尔写的一句话:清晰的归属和问责制。

When it comes to what's missing, one of the best answers I saw in the comments came from a number of people, but was summed up by Bindi Jamil, who writes, clear ownership and accountability.

人工智能可以自动化很多事情,但依然需要有人对结果负责。

AI can automate a lot, but someone still needs to own the outcome.

每个AI工作流都应该有明确的所有者、可衡量的目标,以及当出问题时负责的人。

Every AI workflow should have a clear owner, measurable goal, and someone responsible when things go wrong.

最优秀的AI原生公司不会只问‘人工智能能做这个吗?’

The best AI-native companies won't just ask, can AI do this?

他们还会问‘谁拥有这个结果?’

They'll also ask who owns the result.

朋友们,这无异于一种新的管理学,它不仅适用于现有的管理者,也适用于所有人,因为每个人都既在成为智能体的管理者,也在成为自己工作的执行者。

This, my friends, is nothing short of a new management discipline, and it's a management discipline that applies not just to the current managers, but to everyone, because everyone is becoming a manager of agents as well as an implementer of their own work.

正因如此,很多内容需要在实践中经受压力测试,通过经验和失败来学习,然后总结成我们可以共享的集体智慧。

For that reason, a lot of this is going to have to be pressure tested in practice, learned through experience and failure, and then summed up as collective wisdom that we can all share.

希望这里有一些好的思路,能帮你让你的公司变得更具AI原生性。

Hopefully there is some good food for thought in here about how you help your company become more AI-native.

好了,今天的《AI日报》到此结束。

For now, though, that is going to do it for today's AI Daily Brief.

一如既往感谢您的收听或观看,下次见,祝好。

Appreciate you listening or watching as always, and until next time, peace.
已剔除 3 处广告(点击展开查看)
M1
M11:07广告 · 已剔除

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

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

And to learn more about sponsoring the show, send us a note at [email protected].

Finally, we're coming up on your last chance to sign up for the latest cohorts of the next executive agent training programs from Superintelligent.

So if you listen to this episode and decide that you too want to be more AI native and think those could help, you can check them out at training.besuper.ai.

M1
M111:46广告 · 已剔除

Every AI coding tool on the market does the same thing first: it starts writing code.

Blitzy does the opposite.

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

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

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building.

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

Blitzy never guesses.

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

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

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

See for yourself at blitzy.com.

That's B-L-I-T-Z-Y dot com.

Here's a harsh truth.

Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized.

Half of companies have AI tools, but only 12% use them for business value.

Most employees are still using AI to summarize meeting notes.

If you're the one responsible for AI adoption at your company, you need Section.

Section is a platform that helps you manage AI transformation across your entire organization.

It coaches employees on real use cases.

Tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value.

The result?

You go from rolling out tools to driving measurable AI value.

Your employees move from meeting summaries to solving actual business problems.

And you can prove the ROI.

Stop guessing if your AI investment is working.

Check out Section at section.ai.

That's S-E-C-T-I-O-N dot A-I.

M1
M113:47广告 · 已剔除

Head to robotsandpencils.com.

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

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

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

Marketing agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and update the CRM.

Ops agent chases the paperwork and tracks the budget.

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

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

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.