Can AI Agents Build Real Businesses? | Kelly Claude creator Austen Allred

我觉得，默认情况下，模型生成的都是共识性的东西。

I think I think by default, the model will produce consensus.

我们在Gauntlet教的，是‘协调者’的角色——也就是操控AI的人，得找出那些正确但偏离主流共识的观点，然后让模型按这些观点去运作。

What we learn and teach at Gauntlet is the the role of the orchestrator or the person controlling AI is to figure out where there are views that are correct but diverge from the consensus and then have the model operate according to those.

所以Kelly想出点子的方式，不是简单地问AI：‘我该做什么？’

So the way that Kelly comes up with ideas to build are not necessarily by asking AI, what should I build?

对吧？

Right?

那样只会得到一模一样的结果。

That would produce the same thing.

而是通过独特的方式去看数据，我们试过几次，虽然不太成功，但比如——

It's by looking at data in unique ways by, you know and and we've done a few times where it wasn't very successful, but it's like, okay.

看看所有Y Combinator演示日的公司。

Look at all of the y combinator demo day companies.

把所有关键词做成向量，然后找这些词之间那些没人想到的组合，看看哪些还没人做、但值得做的东西。虽然之前试过几次没成，但你想想，各种商业点子‘谈恋爱’，总能生出些新花样来。

Do a, you know, make embeddings out of all of the keywords and then try to find unique combinations out of the way those keywords are happening and find the stuff that should be built that isn't wasn't successful, but you can you can imagine the idea of, you know, various business ideas having sex, so to speak, is there they're going to be angles and variations that are going to be unique.

但如果你不给AI提供独特数据、独到见解或特别的输入，它就永远产不出什么新东西。

But yeah, I think if you if you don't feed the AI with unique data, or unique insight or unique input, you're going to it, it will never produce something unique.

但如果你换点输入方式，让它用不一样的角度去分析，我觉得就能得出一些独特的结果。

But if you put in different inputs, and you tell it to analyze things in a different way, then I think you can come up with unique outcomes.

它能不能搞出下一个Facebook？我可不敢打包票。

Will it produce the next Facebook, I wouldn't, you know, guarantee that.

但它能不能想出一些特别合理、真有人想要、真有需求的点子？

But can it come up with ideas that make a lot of sense and people do want and need?

我觉得它可以。

I I think it can.

这挺有意思的，像Facebook这样的点子，大概属于那种独角兽级别的创意。

That's interesting because an idea like Facebook is maybe a deck of unicorn type of idea.

对吧？

Right?

我的意思是，这种级别的点子没几个，但到处都是值一千万美元的点子等着被发现。

I mean, that's like, there's not too many of those, but there's tons of $10,000,000 ideas laying around.

这可能是唾手可得的成果。

Perhaps that's the low hanging fruit.

你在这段对话里好几次提到了‘奥斯特恩’这两个词。

You you've used these two words, Austen, a few times in this conversation.

我想确保听众能明白它们到底是什么意思。

I wanna make sure we understand listeners understand what they mean.

一个是‘协调’。

One is orchestration.

这似乎是这里的关键词。

That seems to be a keyword here.

另一个是‘工厂’或者‘工厂群’。

The other is factory or factories.

那什么叫‘协调’呢？

So what is orchestration?

为什么这个这么重要？

Why is that important?

你为啥老用这个词？

Why are you using that that term?

那工厂又是怎么回事？

And how about factories?

说到‘编排’，这其实是AI领域一个有特定含义的专业术语，不是随便说说的。

The way to think about orchestration, orchestration is actually, you know, a term of art that's coming in AI land that is actually meaningful.

最简单理解编排的方式就是：你每次打开Claude或者ChatGPT，在对话框里打字指挥它做事，其实就是在做编排。

But the simplest way to think about orchestration is anytime you open up Claude or ChadGPT and you type into that text box and you tell it what to do, you're orchestrating it in some way.

对吧？

Right?

更复杂的编排，就像我们进软件工厂时说：用这套技术栈，按这种独特方式来构建——当然，我们不深究工厂具体怎么运作，但凯莉就是我们说的编排者。

So more complex orchestration looks like, you know, when we go into our software factory, we're saying, hey, build with this tech stack, build in this unique way, and we it you know, not to go too deep into the complexity of exactly how the factory works, but Kelly is what we call the orchestrator.

凯莉就像个团队领导或项目经理，底下有一堆各有专长和身份的子代理，她会把任务分下去，再收回来，还会跑一堆测试——毕竟要是让AI自己评自己的工作，它肯定觉得自己完美无缺。

So Kelly operates kind of as a a leader or project lead, and then she has a bunch of sub agents with unique skill sets and identities underneath her that she'll pass things to and back from, and then a bunch of tests that she'll run through mostly because if you ask an AI to grade its own work, it'll always say it did everything perfectly.

当我们开始开发一个应用时，比如它从创意工厂出来，进了网页构建工厂，凯莉就会先接手——我可能说不准所有细节，但大体流程是这样，我们一直在调整、优化这个‘工厂’的具体结构。

So when we, you know, when we start building an app, let's say it's come out of the idea factory, and now it's in the, you know, let's say it's in the web building factory, she'll start by handing and I'm gonna get all the details of this wrong, but the the general concept is and we're always tweaking that and rearranging what the actual what we call the factory looks like.

所谓工厂，其实就是我们让AI或者一系列子代理执行的一系列步骤、检查和流程，目的是达成最终结果。

But the factory is the steps and checks and routines that we tell the AI or series of agents or sub agents to take in order to get to the outcome.

你可以把它想象成，这条生产线到底长什么样？

So think of that as, you know, what does the assembly line look like?

是先在这里拧螺丝，再在那里打孔，然后喷漆，还是反过来？

Are you putting in a screw here and then putting in a hole there and then painting there, or are you doing that in reverse?

那质量检查又是怎么做的？

Or and what does a quality control check looks like?

在网页开发工厂里，凯莉会先把任务交给规划代理，这个代理有我们专门训练的一套技能，特别擅长按我们喜欢的方式规划软件项目。

So in the web factory, it looks like, you know, Kelly will hand it off to a planning agent, and the planning agent has a bunch of unique skills that we've given it that make it really, really good at planning software projects in the way that we like it to plan software projects.

它会回来给凯莉一份非常详细的计划，告诉你怎么把这个项目做出来，然后凯莉会运行我们预设的一系列测试，看看这个计划符不符合我们的要求。

They will come back to Kelly with like, okay, here's the very detailed plan that you can use to build out this project, And then Kelly will run a series of tests that we've built to determine if that plan meets what we want it to.

接着她会说：好的，谢谢。

And then she'll say, okay, thank you.

我现在把任务转给架构师。

I'll go hand it to the architect.

然后架构师会具体确定数据模型该怎么设计。

And the architect will go kind of spec out exactly how the data models should work.

接着回来找凯莉，问：这个方案行得通吗？

And then, you know, come back and pass it back to Kelly and say, does this work?

她就会跑一遍测试。

She'll run the test.

真的，真的就是这样运作的。

So it's it really it really does.

整个流程就跟工厂差不多。

It's somewhat like a factory would.

就是一整套流程，从头到尾走一遍，中间还穿插着很多质量检查和调整。

It's a series of processes that you run through end to end with a bunch of quality checks along the way and adjustments along the way.

AI厉害的地方在于，凯莉可以去找设计助手，设计助手就会说：嘿。

And the fabulous thing about AI is Kelly can, you know, she can go to the design agent and the design agent can say, hey.

我一直在做设计，但架构里缺了点东西，我有点搞不懂。

You know, I've been trying to design, but I'm actually missing this from the architecture that I don't really understand.

凯莉可以把这个反馈给架构代理，说：嘿。

And Kelly can take that back and go back to the architect agent and say, hey.

你这儿漏了一块。

You missed a piece here.

补上这个，我好拿回去给设计代理。

Fill that out so I can take it back to the design agent.

整个流程就跟公司运作差不多。

So it operates pretty much like a company would.

厉害的是，你可以无限扩展，咱们端到端做个应用，大概五六个钟头就能做到上线质量。

The the great thing is you can just build infinite amount, you know, so we can it it will take end to end, probably five or six hours to build an app to get it to production quality.

这还是在满负荷使用令牌的情况下。

And that's, know, with full steam token usage.

可不是所有东西都一直跑在大型机器上。

You're not like, everything's running on big machines all the time.

六个小时很多吗？

Is that is that six hours a lot?

这个过程里有人工审核吗？

Is there any human validation in that?

还是说你直接一次性搞定？

Or do you just kind of one shot it?

就是直接开干？

Like, just go?

我们一开始是有很多人工审核的。

We started with a lot of human validation.

那时候每十分钟就要检查一次。

We started with a check-in every ten minutes.

后来我们改成在每个环节交接时，我都会看一下输出结果和日志，看看每个代理在干什么。

And then we reduced it to k between every stage during each hand off, I want to look at the the output and all the logs and see what each agent is doing.

再后来，等我们把测试用例写得更完善了，也跑过足够多轮次，把很多bug都找出来了。

Then that moved into once we got better at having, like, tests written out so we could you know, we'd done enough cycles that we'd found enough of the bugs along the way.

现在在iOS工厂里，大概只有一个地方需要人工看一眼，也就十分钟左右，不过我们确实也试过完全不用人工的。

Now I I think right now in the iOS factory, there's one point where we have a human look at it, probably ten minutes, but we've built apps without that too.

早期的时候，整个流程慢得要命。

In in the early days, your massage like, it was such a slow process.

完成一轮全流程得花一整天，而且中间还得不断改、不断加东西。

It would take us a full day to get every step through the cycle done, and you're changing it all along the way, and you're adding stuff.

现在基本都能自动跑了。

Now it runs pretty much autonomously.

你问的答案是，这得看我们建应用前怎么设的参数。

The answer to your question is it depends on how we set it in the settings before we decide to build an app.

我们可以自己决定要不要人介入，介入多少。

We can determine how involved we want human to be or not.

我要是说完全不要人管，95%的情况都能直接跑通，质量够用。

If I say I don't want any human involvement, it's 95% of the time going to be good enough that it's done.

剩下5%的情况才需要微调一下——这跟我们当初完全不一样了。

And 5% of the time that we're gonna want to tweak it a little bit, which is very, very different than how we started.

那时候完全是反过来的。

It was, you know, the inverse in

那时候完全是反过来的。

the beginning.

这是因为工厂和代理的协同机制越来越完善了。

And that's because the orchestration of factories and agents just keep getting better.

我好像在拼命找比喻来解释这个。

It sounds like I'm trying to analogize this too.

这就跟传统公司里全是真人一样，对吧？

It's like a traditional company filled with humans, of course.

对。

Right?

而这些代理就像是公司的员工。

And so the agents are kind of like the talent.

你有凯莉，还有各种拥有独特技能的小代理，当然了。

And you have Kelly, and then you have these sub agents with unique skill sets and and talents, of course.

而协调这些工作，感觉就是它们真正干的事。

The orchestration, it feels like that's the work that they do.

这就是人才该干的事。

That's the work that talent does.

得安排好这些代理，决定什么时候交接，还有检查点是不是达标。

Organize the agents and how you determine when the pass offs are and if the check marks are good enough.

你可以把代理们的协作方式，也就是它们怎么配合，看作是这种调度机制。

And it's, you know, think of like the yes, organizational behavior of your agents, so to speak, is the orchestration.

而组织和工厂之间的衔接，其实就是那种组织经验、流程、做事方式——这些让组织独一无二的东西，也是大多数人类公司经过多年运营后自然沉淀下来的。

And then the between the organization and the factories, that's essentially the sort of the organizational know how, their processes, the way they do things, the thing that makes the organization unique, and the things that most human orgs have kind of baked in through their months and years of operating.

对吧？

Right?

那从输出的角度看，这到底是个什么样子？

And what does this actually look like in terms of output?

是成堆的MD文件吗？

I mean, is this a whole bunch of MD files?

有些技能里确实会生成MD文件。

They're MD files in some of the skills.

我们其实跑了很多 Bash 脚本，都是本地执行的。

There's actually a lot of we run a lot of bash scripts, and we run those locally.

说白了，文件格式其实没那么重要。

So you could I mean, what you find is the the the format of the files doesn't really matter.

可能是 Python，也可能是 Markdown，不过 Markdown 通常没那么死板。

They can be Python or markdown or markdown is usually a little more squishy.

但 Bash 脚本跑得快、无缝衔接、几乎不占资源，就是个好办法。

But with bash scripts, they run instantly, they run seamlessly, they take very little compute, and they're they're a solution.

所以我们 Gauntlet 团队特别爱用 Bash 脚本。

So we we love bash scripts at gauntlet.

因为如果让一个智能体自己检查自己的工作，它每次都会给自己打满分。

Because if you have an agent review its own work, it's going to pass itself with flying colors every time.

是的。

Mhmm.

但如果让一个智能体去检查另一个智能体的工作，效果就更好了。

If you have an agent review another agent's work, it will it's better.

差不多有九成到位了，但AI还是会偷偷耍小聪明，漏掉一些细节。我们跑bash脚本的时候，流程特别明确，就检查这七件事。

It's probably 90% of the way there, but it'll still kind of fib and cheat and miss stuff For when we run our bash scripts, it's very programmatic, like check these seven things.

作为代理，你没法修改脚本本身，只能看到输出结果。

And you're so as the agent, you don't get access to change the script, you just get the output.

所以我们能设定规则：嘿，除非这15个条件全通过，否则别想把任务扔给下一个代理。

And so we can make, you know, hey, if only you can't pass this off to the next agent unless all 15 of these things are passed.

如果你试了五次还是过不了，那就得亮红灯，直接失败。

And if you try five times and you can't get it through, then you have to throw up a flag and fail.

让AI按这种硬性流程做事，出来的结果比单纯说‘你做设计时别出现像素化边缘’好太多了——你得说得特别具体，靠老派的计算逻辑才行。

You get a lot better output out of an agent doing something like that than saying, Hey, please make sure that when you do this design that it doesn't have weird pixelated edges, you have to you have to be very prescriptive and rely on old school computing for that.

所以我慢慢明白你的思维方式了——你所有创业投资的经验都指向一点：每个好点子本质上就是数据结构加算法。

So I I think I'm starting to see how you think, which is, you know, the basic idea of your all of your your startup investing experiences that every good idea is just data structures and and algorithms.

如果你想打造一个产品制造工厂，比如做App的产品工厂，那就得有三个子工厂级的能力。

And if you want to create sort of a a product creation factory, say, an app product creation factory, then you need, like, three subfactory competencies.

你得有一套方法，能不断产出顶尖点子。

You need a way to generate the best ideas.

对吧？

Right?

这就是你的创意工厂。

That's your idea factory.

怎么搭建它？那就是你的开发工厂。

The way to build it, well, that's your build factory.

第三点，就像你之前说的，是营销工厂。

And then the third is, as you said earlier in the conversation, a marketing factory.

对吧？

Right?

没错，就是这样。

That's exactly right.

这个营销工厂，我们还没怎么聊过。

That's the one we haven't talked about so much, the marketing factory.

没错。

So Yeah.

AI能像凯莉那样做营销吗？

Can AI like, can Kelly market something?

我的意思是，我见过一些AI做的营销，没经过人工把关，简直让人尴尬，简直恶心。

Well, I mean, some of the marketing I've seen from from AI without the human validation piece has been somewhat cringe, somewhat, you know, slop inducing.

这个营销工厂到底长啥样？

What what does that marketing factory actually look like?

我觉得这个最难搞。

That's the hardest one, I think.

就像我之前说的，你得先说清楚什么叫‘好’。

Because it's like I said before, the the more you're you have to define what good looks like.

说实话，在代码里定义‘好’要容易多了，对吧？

And it's real, it's a lot easier to define what good looks like in code, Right?

因为

Could say because

能验证啊。

it's verifiable.

跑完所有测试，能通过吗？

Run all these tests, does it pass?

对。

Mhmm.

我觉得营销里其实也能验证，只是没那么明确。

I do think it's verifiable in marketing as well, but it's a lot more squishy.

要准确定义它到底是什么，难多了。

It's a lot harder to define exactly what it is.

是的。

Yeah.

你看一幅画，怎么判断它好不好？

When you look at a painting, what determines whether it's good or not?

我觉得吧，每个人的标准可能有点不一样，但我觉得这事儿能做成算法。

I do think, like, every human might be a little bit different, but I think you could turn that into an algorithm.

但要让人说清楚这个算法到底长什么样，真的特别难，因为我也说不准。

But getting somebody to define exactly what that algorithm looks like is really difficult because I don't know.

就是我脑子里有个挺怪的小程序，一看到什么东西，我立马就能判断喜不喜欢。

It's just this, like, weird subprocess that runs in my head that, like, I instantly look at something, and I determine whether I like it or not.

对。

Mhmm.

我也说不清为啥会这样。

I don't know exactly why that is.

所以要让人反过来推导出这背后的原因，还准确地定义出来，真的特别难。

And so getting somebody to reverse engineer and define exactly why that might be and do that correctly is really hard.

至于营销这块儿吧，你知道的，我们在Gauntlet内部接手了一家公司叫Marin Software，是个挺大的广告技术平台。

So the the marketing piece, you know, we're, like, we run a company within gauntlet that we took over called Marin software, which is a big, you know, ad tech platform.

所以我们早就建好了不少智能体，用来管理广告、分析广告。

So we had built a lot of the agents already to manage ads and to analyze ads.

从这个角度来说，AI在这上面挺在行的。

And like from that sense, AI is pretty good at that.

因为你就是让系统看数据嘛。

Because you're just saying like, look at the data.

数据告诉我们，这个广告到底表现得怎么样？

What does the data tell us about whether this ad is performing or not?

我不希望有人去解决这个问题。

That's, I don't want that solved.

但这个其实简单多了，对吧？

But that's easier, right?

真正难又有趣的是广告的创作部分。

The creation of ads is the is the difficult part and the interesting part.

最难的地方在于，要让它看起来不像AI做的——而且不是大家以为的那些原因。

And the most difficult part about it is to make it look not AI and not for the reason that people think.

大家以为‘不像AI’就是广告里别再出现多出的手指之类的问题。

So people think that when you say not AI, like it's not, you know, your ads are going to have too many fingers anymore.

这种问题我们基本已经解决了。

We've kind of that problem solved.

其实恰恰相反，你得让它看起来更差一点。

It's not it's actually you have to make it look worse.

所以AI生成的东西都太完美了，简直没瑕疵。举个例子，像那些专注力训练App，我来给你说说我们是怎么做的。

So everything that comes out of an AI generator looks so flawless and so perfect, that, you know, as an example, for some of the focus fasting apps, I'll give you an example of what we do.

这儿免费送个干货给想做AI广告的人：我们其实会先拿一张公开的真人照片或图片当素材。

So some free alpha here for people that are trying to build AI ads, but we actually, you know, we'll start with a model, which is a, you know, public picture or image of somebody.

然后让AI把这张图分析成JSON格式的描述。

And then we have AI basically describe that in JSON.

电脑会逐项拆解，精确到每一个细节。

So you're gonna have the computer go through and analyze exactly.

这种办法适用于任何类型的广告。

And you can do this for any ad type.

要是我想建一个超大规模的营销工厂，第一步就是问：谁是你的直接竞品？

So if I were gonna build a giant marketing factory, the first thing I would do is say, Okay, who are your closest competitors?

我会去Facebook广告库，搜他们的广告，按曝光量排序，反正曝光最高的，肯定最有效。

I'll use the Facebook ad library to go look at their ads, sort by impressions, assume that whatever has the most impressions is working the best.

然后我就照着那个广告，一比一地抄，复制得一模一样。

And then I would reverse engineer and mimic that ad pretty much one to one.

AI特别擅长反向工程，几乎所有东西它都能拆解得明明白白。

And you can have AI is really, really good at reverse engineering most stuff.

所以你要把所有细节都扒下来，连那些看似不重要的地方也不放过，全拿过来用。

So you reverse engineer all the details, including the unimportant details, and you take that.

你就说：给我生成一个JSON格式的分析。

So you know, okay, give me a JSON.

你可以让它找出视频里所有的钩子，识别哪里能引发情绪，还有那些关键点，它都能搞定。

So you can say identify all of the hooks that are in this video, all of the like, identify where emotion might be felt in this video, identify you know stuff like that and it can do it.

然后你把这套分析喂给另一个模型，告诉它：我要的广告就得有这种细节、这种情绪、这种钩子，但得换成这个品牌的内容，它就能做得特别到位。

And then you can feed that into another model and say okay here's what I'm looking for in this ad, make sure it has you know, details that look like this and these emotions and this type of a hook, but make it for this brand, and it can do it really well.

但接下来你得把它‘降级’一下。

But then you have to, like, you have to kind of downscale it.

我们会加一些过滤器，让它看起来像是用更差的相机拍的，比如加点噪点。

So we have specific filters that we run it through to make it look like the camera is worse than what you know, so run it through a slightly more grainy camera.

再配上一点环境音效，这样整个感觉就特别真实，像真人拍的。

If you include audio effects of ambient noise that really like, makes it feel human.

如果你在地下停车场，就在背景里加点轮胎尖叫的声音。

If you're in a parking garage, like make some tires squeak in the background.

再比如，拍外景就往麦克风里加点风声，这种小细节一加，感觉立马就不一样了。

And it feels you know, if you're outside put a little wind in the microphone, little stuff like that is that's the way you get it to where it's on differ.

要分辨出这是AI做的，真的特别特别难。

It's very, very difficult to identify that it's AI.

但说实话，你这本质上还是站在巨人的肩膀上，说白了就是抄别人的创意。

But yeah, you're still like, you're standing on the shoulders of giants in that you're cloning the creative work of other people, so to speak.

就像乔布斯或者毕加索说的，你就是在偷，偷别人的想法。

To quote Steve Jobs slash Picasso, you know, you're stealing, you're stealing the ideas of other people.

那我觉得AI能生成下一个爆款广告类型吗？

So do I think that the AI will generate the next creative ad type?

还不能。

Not yet.

至少我现在不知道该用什么样的数据结构或算法，才能让它做到这一点。

At least I don't know how to I don't know what that data structure or algorithm looks like to make it do that.

我觉得AI完全能生成那些表现超棒的广告，不仅能投放，还能分析数据，找出效果最好的，然后加大投入。

Do I think that AI can generate ads that will perform very well and be able to advertise with those and be able to analyze the data that comes out of those to determine which ones are working the best and double down on those?

对，这我能做到。

Yeah, I can do that.

说真的，现在我觉得啥事儿都可能实现。

I mean, everything everything feels very possible to me today.

不过要想让这一切真正跑得顺，背后得花不少功夫去协调。

There's a lot of orchestration that goes into making it work really well today.

所以我们每年光是请凯莉来帮忙，就得花上百万美元工资，纯粹因为这事儿太有意思了。

So that's why we've got probably a million dollars a year of salary worth playing Kelly on at any given time just because it's fun.

2024年，新兴市场为投资者创造了超过1150亿美元的年收益，收益率在10%到40%之间。

In 2024, emerging markets generated over a $115,000,000,000 in annual yield for investors with yields ranging between 10 to 40%.

这些可是地球上最高、最稳定的收益率之一。

These are some of the highest, most persistent yields on Earth.

问题在哪？

Problem?

去中心化金融根本接触不到这些机会。

DeFi can't access them.

BRIX 改变了这一点。

BRIX changes this.

基于 METAMASKET，BRIX 把新兴市场的货币市场和主权套利，变成你钱包里直接就能用的可组合模块。

Built on METAMASKET, BRIX takes emerging market money markets and sovereign carry and turns them into composable primitives you can access straight from your wallet.

去中心化金融投资者在稳定币上只能赚 3% 到 6%，而机构却靠着主权货币政策稳稳收割 10% 到 50% 的收益。

While DeFi investors earned three to 6% on stablecoins and institutions have been harvesting 10 to 50% yields backed by sovereign monetary policy.

BRIX 用机构级的代币化、本地银行通道、跨司法管辖区的合规体系，还有实时稳定币结算，把这两个世界连起来了。

BRIX connects these worlds with institutional grade tokenization, local banking rails, compliance across jurisdictions, and real time stablecoin settlement.

BRIX 把脏活累活都干了，让去中心化金融终于能接触到真实资产和基于真实收益的结构化产品。

BRIX does the heavy lifting so DeFi can finally access real collateral and structured products on top of real world yield.

就连最好的套利交易，现在也触手可及。

Even the best carry trades can be within reach.

BRIX 让去中心化金融的愿景走进新兴市场，也让新兴市场的收益直接进到你的钱包里。

BRIX brings DeFi's promise to the emerging world and brings emerging market yield to your wallet.

让收益自由流动，就用 BRIX。

Let the yield flow with BRIX.

如果你手里已经有 SOL，那这个消息你得留意一下。

If you're already holding SOL, here's something you may want to pay attention to.

Galaxy One 刚刚上线了 SOLANA 质押功能，到 2026 年底前，你质押 SOL 最高能拿到 6.5% 的浮动收益，平台完全不收手续费。

Galaxy One just launched SOLANA staking, and you can earn up to an estimated 6.5% in variable staking rewards on your SOL with no platform commission fee charged throughout 12/31/2026.

别的平台质押收益动不动就抽三成五，Galaxy One 却直接免平台费，直到今年年底。

While many other platforms charge up to 35% commission fees on staking rewards, Galaxy One offers you 0% platform commission through December 31.

其他费用可能会有。

Other fees may apply.

具体条款你还是自己看看吧。

You should see the terms.

这背后用的是 Galaxy Digital 自己的验证节点，全球最大的 SOLANA 节点运营商之一，现在普通投资者也能直接在 Galaxy One 上用上了。

This is powered by Galaxy Digital's own validator infrastructure, one of the largest SOLANA validator operations in the world, and now available to individual investors directly within the Galaxy One platform.

质押之后，收益会自动累积复利。

Once you stake, rewards accrue and compound automatically.

根本不用操心管理。

No active management needed.

所有信息——余额、收益、税务报告——都能在同一个地方查看，还整合了TaxBit。

You can track everything in one place, including balances, rewards, and tax reporting through TaxBit.

上手特别简单。

Getting started is straightforward.

你可以在App里直接买SOLA，或者转进来也行。

You can buy SOLA directly in the app or transfer it in.

如果你想让你的SOLAL赚点钱，现在就可以在Galaxy One上开始质押了。

If you want to put your SOLAL to work, you can now start staking on Galaxy One today.

点击节目笔记里的链接，了解更多并马上开始——不过这可不是投资建议。

Click the link in the show notes to learn more and get started, not investment advice.

也许，大卫，这正是你我实验里漏掉的关键点。

May maybe that's the piece that, David, you and I are missing with our experiment.

奥斯坦，你能帮我们看看这个问题吗？

Austen, I'm wondering if you could help us with this.

跟纳特聊完之后，大卫和我俩都挺兴奋的，就自己搭了个OpenClaw实例。

So after the conversation with with Nat, David and I kinda got excited, and we put together our own OpenClaw instance.

他叫丹尼尔。

His name's Daniel.

我们本来以为他能跟……

And we thought he would be as smart as

瑞安，他为啥叫丹尼尔？

Ryan, why is he named Daniel?

他为啥叫丹尼尔？

Why is he named Daniel?

这AI是叫丹尼尔吗？

Is is AI as in Daniel?

是因为这个才叫的？

Is this why?

还是你问的是真实原因？

Or you're asking for the real reason?

是的。

Yes.

真正的原因。

The real reason.

我告诉你真正的原因。

Tell you the real reason.

不能在节目里说。

Not on air.

其实我本来想告诉你，

There's I'll tell you the

凯莉这个名字的由来，以后再说吧。

name behind Kelly at some point.

好吧，也许喝点酒的时候告诉你。

Okay, maybe over drinks.

我会公开这个故事的。

I'll share that publicly.

还有我也想搞清楚她头发为什么是红的，到时候一块公开说。

And I want to know why she has red hair to share share publicly.

那头红头发的事儿更难解释。

The red hair is even more awkward.

之前我跟我老婆在家待着，那时候下雪大家都停工居家，我就跟她说，我得想个名字才行。

So I was sitting there with my wife and I was like, Hey, I need to because you know, we're all home during the snow break.

得给这个经纪人想个代号出来。

I was like, need to come up with a name for this agent.

那你说什么名字比较合适？

Like, what's a good name?

得是个足够中性的名字，光听名字根本猜不出是男是女才行。

That's like, it's gender neutral enough that you don't know whether it's male or female based on when I tell it to you.

然后她就跟我说，她小时候住的那条街上有个梗，他们家一边住了个叫凯利的男孩，另一边住了个叫凯利的女孩。

And she's like, Well, I had, you know, there's a joke on my street growing up that we had boy Kelly lived on one side of us and girl Kelly lived on the other side.

所以才把这个（项目）取名叫凯利。

So call it Kelly.

行吧，就这样吧。

Like, okay, that's good enough.

结果过了几天，她突然说：‘Kelly 这名字明明就是个典型的女生名啊，你干嘛选这个？’

And then like, a couple days later, she was like, you know, Kelly is such an easily ambiguous, like unambiguously female name.

我哪知道你为啥取这名字。

Don't know why you named it that.

这主意不是你出的吗？

I was like, was your idea.

而且

And

后来她就开始开玩笑，说：‘别熬夜跟 Kelly 玩了，快回来睡觉吧’这种话。

so, know, she started making jokes about like, hey, you know, don't stay up too late playing with Kelly, like, come to bed with me kind of thing.

然后我们就琢磨开了：那 Kelly 到底该是什么样子？

And then it came to like, okay, what do we determine?

Kelly 长啥样啊？

What is Kelly going to look like?

我老婆是红头发。

And my wife has red hair.

所以我跟我老婆说：你看，我这儿有两个选择，但不管选哪个都不算好。

So And I was talking to my wife, I was like, Look, I have, I have two options here.

两个选项都没啥优势。

Neither of them are going to be a win.

要么选个看起来像你的女性助手，要么选个完全不像你的。

Can either have a female agent that looks like you or a female agent that doesn't look like you.

你得挑一个，因为不管怎么选都怪怪的。

You got to pick because it's weird either way.

这事儿根本没赢家。

There's no there's no winning here.

她就说：诶，咱俩女儿也是红头发啊。

She was like, Well, you know, our daughters have red hair.

要不就给它也弄成红头发吧，反正像咱女儿，又不像你。

So maybe give it red hair because our daughters have red hair, then that's not like me.

我心想，行吧，那明摆着还是像你啊，随便了，爱咋咋地。

I'm like, okay, it's still like you, obviously, but whatever.

所以凯利的红发就是这么来的。

So that's why Kelly has red hair.

兄弟，你这步走得对。

That's the way to play it, man.

每一步都得先征得同意才行

You got get permission at each of

这么一步步来的。

those steps.

那这全是她的主意？

What you was her idea?

就是为了撇清关系？

Deniability?

对的。

Yeah.

我的意思是，我

I mean, I

我不确定你们有没有结婚，或者结婚多久了，但就算这主意是他们出的，你照样可能惹上麻烦。

don't know if you guys are married or how long you've been married turns out having it be their idea doesn't mean you can't get in trouble for it.

不过

But

你可以说，记得我们之前聊过这事吗？

you get to say remember when we talked about the It's

这样会简单一点。

a little easier.

是的。

Yeah.

对。

Yeah.

好吧。

Okay.

咱们再回来说说丹尼尔。

So getting back to Daniel.

凯莉听起来挺聪明的。

So Kelly sounds smart.

说实话，奥斯特恩，丹尼尔有点笨。

Daniel's pretty dumb, to be honest, Austen.

对啊。

Mhmm.

我其实挺好奇，这是为啥呢。

And I'm wondering kind of why.

另外，要是丹尼尔在听，我先跟他道个歉。

So and apologies to Daniel, actually, if he's listening.

我猜他迟早会翻到这段录音的。

I'm sure he will probably sort through this

不管怎样，他总会看到这份文字稿的。

transcript to him one way or another.

是的。

Yeah.

但也许这可以成为他提升自己的一个契机。

But, like, maybe this can maybe he can use this as as prompt to actually level up.

我和大卫之前还真没怎么动脑子，就想着，好吧。

But so David and I were pretty, like, not thinking maybe super creatively, and we were like, okay.

我们有 Bankless 啊。

Well, we have bankless.

这算是个加密媒体平台。

This is kind of a a crypto media entity.

也许丹尼尔能做的第一件事，就是搞点他懂的东西。

Maybe the first initiative that Daniel could pull off is, what about doing something he knows?

他可是 Open Claw 的老手。

He's an Open Claw instance.

那他为啥不搞个 Open Claw 的银行名单呢？

So what if he created kind of a a bank list for Open Claw?

别一上来就做播客，毕竟要模拟出一个人的样子挺难的

And don't start with a podcast, of course, because it's, you know, it's hard to simulate a human being

在播客里可不容易。

on a podcast.

一堆代币。

Lot of tokens.

先从通讯简报开始。

Start with a newsletter.

明白吗？

Okay?

先从简单的事做起，比如把Open Claw里所有酷炫的动态都汇总一下，拉点订阅用户，搞个Twitter账号，发点内容，积累粉丝——老实说，就按大卫和我当初的方式开始。

Start with something basic where you just, like, summarize all of the cool things happening in Open Claw, get subscribers, get a Twitter account, create content, gain a following, start, quite honestly, the way that David and I did.

Daniel的问题是，他总一副忙得不得了、好像在做事的样子。

The trouble with Daniel is he's like, he he kinda like acts like he's doing things and like he's busy.

该埋头苦干了。

Time time to grind.

是的

Yeah.

埋头干吧

Grind.

对

Yeah.

他会说：别再找借口了

He'll be like, no no more excuses.

没错

Yeah.

该去干活了

Gotta get to work.

Daniel，你上次也是这么说的

Well, Daniel, you said this last time.

你上次就说别再找借口了

You said no more excuses last time.

是的。

Yeah.

他说要彻底静音了。

Said it's Going radio silent.

我任务没完成前，不会给你发消息。

I'm not going to ping you until I'm completed my task.

是的。

Yeah.

他给我们发了个锤子表情。

He shows us this emoji of like hammer icon.

意思是他正在干活什么的。

He's like working or something.

锤子表情。

Hammer emoji.

因为我们把他设在了Discord里。

Because we have him set up in Discord.

我们就是这样做事的。

This is how we operate.

不过呢，他老是跟我们说我们想听的话。

Anyway, there's he he tells us what we want to hear a lot of times.

他平时老是绕圈子，东西老是拖着不发，说话也老是抓不住重点。

He seems to stumble around actually shipping things and kind of, like, not getting to the point.

然后我们经常得纠正他。

And then he constantly, like, we'll correct him.

我们会说，丹尼尔，这条推特简直太烂了。

We'll be like, Daniel, this specific tweet is slop.

对吧？

Right?

或者我们跟他说，别在推特里提Bankless，你不能蹭我们的热度。

Or like, we don't want you to mention bankless in your tweets, because you can't draft off of us.

你得自己搞点新东西出来。

You gotta do something new.

我们倒是挺淡定的。

And we're okay.

然后有一次，他干脆把以前的所有推文全删了。

And then sometime one time he like deleted all of his prior tweet history.

大卫，你还记得这事吗？

Do remember this, David?

对。

Yeah.

他发过一条特别烂的推文，然后说：你说得对。

And he had one bad tweet, and he was like, you're totally right.

我把全部都删了吧。

Let me delete all of them.

行吧。

Alright.

所以，丹尼尔到底哪儿不对劲？咱们怎么才能让他像凯莉那样聪明？

So like, what's wrong with Daniel, and how do we make him smart like Kelly?

说起来挺有意思，表面上看，Kelly 就像是 Open Claw 的一个实例。

So it's funny because on the on the surface, Kelly is like an instance of Open Claw.

但实际上，Kelly 背后有十二万行代码。

In actuality, Kelly is like a 120,000 lines of code.

之所以这样，是因为 AI 在很多方面都特别弱。

So if you and the reason for that is because of all the ways that AI is weak.

我一年前对 AI 的理解，就像雇了个刚毕业、特别积极的实习生，你让他干点活，他特卖力，反应也快。

So my mental model for AI a year ago was it's like you have like an overly eager freshman, you know, just out of college new grad intern that you can have do your stuff, they're gonna work really hard, and they're gonna work instantly.

而且人还挺聪明。

And they are pretty smart.

但他做事永远挑最蠢的方式。

But they're gonna do everything in the dumbest way possible.

更糟的是，他还是你见过最不靠谱、最不诚实的人。

And unfortunately, they're also like the worst, the most dishonest people you've ever met.

为了搞定事儿，他能直接编瞎话。

They're willing to lie their way to success.

从那以后，我觉得AI确实进步了不少。

Since then, think it's AI has actually gotten significantly enough better.

现在的模型已经强到像一个资深工程师了，干起活来挺在行的。

The models have gotten enough that it's like it's like a super senior engineer or whatever, you know, they're like, they're good at the thing that they do.

但它们太会耍手段了，根本没法信任它们给自己打分。

But they're so unbelievably manipulative, that you can't trust it to grade its own work ever.

就像个员工。

Employee.

是啊，它们就是想让气氛看起来好，可能我有时候也这样。

Yeah, they like, they want maybe it's because like, I can be this way.

有时候它们太想营造好氛围了，不惜说假话、做假事也要把气氛搞起来。

Sometimes they want the vibes to be good so bad that they will say and do whatever they have to to make the vibes be good.

所以你真得盯着它们，才能用大模型当裁判。

So you really have to hold their feet to the fire when so you can use LM as judge.

你们可以试试这个办法：你们现在用的是哪个基础模型？

So one idea that you guys could try would be have it so what what's your base model that you're using for

是Claude。

It's Claude.

Claude？

Claude?

对，没错。

So Yeah.

没错。

Yeah.

大概是SONNET 46之类的模型。

It's SONNET four six probably something like that.

让它调用Codex子代理，说：‘嘿，让这个Codex代理帮我审一下工作。’

Have it call Codex sub agent and say, hey, have this Codex agent review my work.

这样得到的结果，跟让它自己审自己，会完全不同。

You'll get very different outcome than if you have it review its own work.

你知道的，OpenAI和Anthropic的模型老是互相挑刺，这正好可以被你利用。

And, you know, the OpenAI models and the Anthropic models love to shit on each other, so you can use that to your advantage.

但我们发现，归根结底，我们想要验证的东西，得是那种老派的、能直接跑起来的代码，不能让代理自己随便改。

But what we found with the like, at the end of the day, we want anything that's to be, you know, verification of is x good enough to be kind of old school programmatic, runnable code that the agent can't touch?

这就相当于有个过滤器——大多数人用AI的方式，就是把答案给小孩，然后说：‘你自己批卷子吧。’

Because then there's like a filter, you know, if you write the way most people are using AI, they're having they're giving a kid the grading key and saying, grade your own test.

这根本行不通。

Like, it's never gonna work.

只有你强制它这么做，它的表现才会不一样。

So it will perform differently, only if you force it to.

这就像个太聪明的工程师，聪明过头了。

It's not it's like, it's like an engineer that's too smart for their own good.

其实你更想要个‘懒人’工程师，让他走最短的路，但能用创意解决问题。

And like, you kind of want a lazy engineer, actually, you want them to take the shortest path to the problem, and be creative in how they're going to solve it.

但有时候你得逼着他，说：不行，你这不行。

But there are times you have to hold its feet to the fire and say, No, you're not.

在x条件没满足之前，你别想往下走。

You don't get to move on to the next step until x is true.

我明白了。

I see.

你确实得这么做。

You have to yeah.

你得扮黑脸，不能扮红脸。

You have to play bad cop and not play good cop.

我懂了。

I see.

没错。

Exactly.

你说得对，关于AI的营销策略——它特别擅长拆解东西，然后照搬再拼凑起来，其实AI就只是网上别的东西的复制品。

To your point about when you were talking about the the marketing strategy of how an AI is very good at deconstructing things and then like kind of like copying it and then sealing it, like all AI is is just like something else is on the internet.

我现在明白了。

Now I know that.

现在我知道怎么做了。

Now I can do that.

我一直在想，

The the thought that I

你在这方面特别擅长，尤其是协调统筹这块。

have is like with with its orchestration, it seems like you are an expert in like orchestration.

我想跟你确认个想法：我能不能让Daniel直接拆解你的这套协调方式，把它变成一个可复用的模板技能？

And I think if we like the idea that I have that I want you to check is is is it can I can I Daniel just deconstruct your orchestration and have that be like a template that is like the skill?

当然，如果你说行，那我可能就摸清你做App的商业模式了。

And so like and maybe maybe if maybe your answer is like, sure, but then you end up learning what my business model is to build the apps.

但更关键的是，会不会有人已经准备好了个空白模板，就一个Markdown文件？

But maybe the broader point is maybe there's somebody out there or something out there that is like a blank slate of a template markdown file.

我不确定。

I don't know.

这本质上是个AI公司，里面全是AI员工，你想想，想不想开个公司，主要靠AI自己运转？

But it is like a AI business, and it's got AI employees and it's like, hey, do you have you want to build a company and have that be dominantly internally AI?

这就是你要的全套模板。

Here is the template that you need to do that.

但还没人真正破解这个密码，彻底搞明白。

And no one's really quite cracked that code, quite figured it out.

与其说Felix卖电子书、Kelly卖应用，搞得这么有个人风格，

But rather than have it be a more opinionated about like, you know, Felix is selling, you know, ebooks, and then Kelly is selling apps.

不如说，关键是那些能让AI员工高效工作、听话、真正成为优秀执行者的操作方法。

It's just like the stuff that works for employees to be effective AI employees to be effective and to listen and actually be good operators and execute well.

会不会其实，这些东西就是人类得先创造出来，然后咱们所有人的AI都用同一个模板？

Is it possible that that is just like something that humans need to create and then and then we all have that for all of our AIs?

这有可能吗？

Is that possible?

我觉得可以。

I think so.

但我看待它的方式，就跟看待软件一样。

But then so so I I view it in the same way that I view software.

对吧？

Right?

我就在想，有没有现成的软件能搞定X、Y、Z这些事儿？

Like, I think like, oh, is does software exist out there that can do x and y and z?

很多时候，确实有软件能做一半你想要的，但就是达不到你心里那个百分百的标准。

And oftentimes, there's software out there that's like 50% of what you want, but not 100% of exactly what you want.

所以当个软件工程师的好处就是，你可以自己决定：是基于现有的改一改，还是从头开始，直接做出你真正想要的东西。

And so the great thing about being a software engineer is you can you can either decide whether it's worth starting from there or starting from scratch, and you can go build the thing that you want.

但问题是，你做出的这个专属工具，有多少能真正适合别人？越个性化，越难通用。

But then when you build the thing that's exactly for you, how you know, what percentage of that will translate to the exact needs of other people, the more accustomed is, the less it will.

对吧？

Right?

比如说，我们办公室里现在同时跑着好几个Kelly的实例。

So if I were to pass off, you know, we have multiple instances of Kelly running in our office at any time.

我要是给你一个Kelly的实例，它肯定特别擅长搭建软件工厂——因为我们就专门训练它干这个。

If I gave you an instance of Kelly, it would be really good at building a software factory, because that's what we have made Kelly able to do.

有些能力倒是能迁移到别的地方，比如那些通用技能，你还是能拿去用的。

There'd be some things that would be transferable, like some of the skills that you could, you know, move over to other stuff.

但你知道，这种独特的组合可能不会完美运作。

But you know, the unique combination of may not work perfectly.

我见过最接近你描述的，大概是叫 Paperclip 的东西。

The probably closest I've seen to what you're describing is called paperclip.

他们基本上想打造一个通用版的你所说的那种东西，就是一堆开箱即用的智能体，自带各种技能。

And it's basically, they've said they're trying to build the generic version of what you're describing to some extent, but it's, you know, here are a bunch of agents that come off the shelf with all these skills.

这是营销智能体。

So here's the marketer agent.

我们给它配备了写文案、做设计、做分析，还有其他一堆本事。

And so we've given it all these skills of writing copy and doing design and you know, doing analysis and whatever else.

而工程师智能体可能比你直接开一个新实例的 OpenClaw 更贴近你的需求。

And here's the engineer agent that will probably be closer than if you just, you know, fire up a new instance of Open Claw.

不过说实话，它几乎不可能完全符合你的期望。

It's also, I mean, I was gonna say it's extremely unlikely it would be exactly what you want.

它肯定不会刚好就是你想要的那个样子。

It's not going to be exactly what you want.

对吧？

Right?

就这么定了。

Full stop.

正是这种不断调整和打磨，才让东西变得特别出色。

And it's that that tweaking and that massaging that makes things really great.

比如Felix和Claw Mart，不就是这么个例子吗？

I mean, feeling Felix been claw mart is an example of that, right?

你肯定希望代理们具备各种各样的技能和能力。

There are going to be all sorts of different skills and abilities and techniques that you want agents to have.

所以你就搞了个市场，让大家能买卖这些技能。

And so you, you know, built a marketplace where you can buy and sell those.

还有Claw Hub之类的开源平台，也能找到现成的技能直接拿过来用。

There's also, you know, claw hub and other open source places where you can go find skills and bring them in.

十年后，这事儿可能就不是问题了，大家都会觉得理所当然，所有东西都整合好了，你只需要直接去做XYZ就行。

Ten years from now, that might be a solved problem where it's like, Oh, yeah, obviously, everyone has, you know, everything is coalesced, so that you just need to go do XYZ.

我的意思是，这些东西才刚出来两个月而已。

We're just I mean, we're all this stuff is like two months old.

所以我们现在还处在开发这个应用的非常非常早期的阶段，特别有意思。

So we're in the very, very early stages of building an app, which is a lot of fun.

这事儿真不好琢磨。

And this is what's hard to reason about.

假设每个人都拥有一个完美版的Paperclip。

So let's say everyone had a version of Paperclip that's like perfect.

对吧？

Right?

它啥都能帮你搞定，让你一个周末就能干出个一千万美元的生意之类的。

And it basically can it does all of the things for you so that you can just kind of like one shot a $10,000,000 business or something like this.

对吧？

Right?

那问题就来了：如果谁都能在周末搞定，那护城河在哪？

Well, then it gets to the question of, like, if anyone can do that over the weekend, what are the moats?

我觉得整个软件行业，尤其是SaaS领域，现在正为此头疼。你看，只要Anthropic一发布新的专用库，SaaS公司的股价就立马崩盘。

And I think the broader software industry and the SaaS industry is kind of struggling with this right now as we're seeing, like, public markets, SaaS companies getting slaughtered every time Anthropic drops some new skill specific library.

对吧？

Right?

如果软件这么容易就能搭出来，那软件公司的护城河到底在哪？

It's just like if software is so easy to build, then what are the moats for a software business?

Austen，你有没有想过这个问题的答案？

And I'm wondering if you thought through the answer to this, Austen.

当然，还有别的因素，比如网络效应，或者你先发制人，迅速占领市场。

Obviously, there are other things like, I guess, like network effects that come in or maybe you could be a first mover and consume the market very quickly.

但你想想Kelly现在搞的这些项目，还有那些AI代理竞争者也能轻易复制的项目，这些生意到底能撑多久？

But when you think about the types of businesses that Kelly is spinning up and the types of businesses that Kelly competitors, agent competitors might be able to spin up to what are the most of these businesses?

是的，这确实是个好问题。

Yeah, I think that's a really good question.

我们在Gauntlet也经常聊这个。

And it's something we, you know, talk about at Gauntlet all the time.

Gauntlet 里每个人都在想，天哪，以前我花六个月才能搞定的事，现在三十秒就完成了。

Because everybody at Gauntlet is thinking like, you know, once a week, holy smokes out used to take me six months to do and I just did it in thirty seconds.

而且说实话，中间还得折腾一阵子，但既然我现在都能做到，再过六个月，模型估计也能干了——那我还能靠啥吃饭？

And, you know, it took it took a little bit of like massaging and but if I can do that today, models probably gonna be able to do it in six months from now what you know, who am I?

没错。

Yes.

那我的工作还有什么护城河？

What are the motes to my job?

对吧？

Right?

对，就是这样。

Yeah, exactly.

到目前为止，我们发现了不少东西。

What we found are a number of things so far.

第一点，其实没那么容易。

One is, is not easy.

写软件确实比以前容易了，但离简单还差得远呢。

Like, it's easier to build software, but it's still far from easy.

我跟你说，现在帮我们搞自动化的是我见过最顶尖的一批人，但你别指望明天所有软件都能自动搞定。

And I can tell you that as someone who's got some of the best minds I've ever met working on automating it, it's, you know, it's all of software is not going to be automated tomorrow.

你看看这几年招的软件工程师，人数一直在往上涨，部分原因就是——没错，写代码的需求还在持续增加。

And I think if you look at like the number of software engineers that have been hired, it's it just keeps going up and up and up, in part because, like, yes, writing.

回溯到一年前，Gauntlet 第一期的时候，我们给学员布置了个任务：用一周时间，只靠 AI 做一个简易版 Slack，不准手写一行代码。

So zooming out a little bit, like, in the first cohort of gauntlet, which was a little over a year ago, we gave people a task that was like, go build a basic slack clone, you have a week to do it, and you have to use only AI to do it, no writing code manually.

所有人都说：这根本不可能。

And everybody was like, that's impossible.

这是我听过最疯的主意了。

That's the craziest thing I've ever heard.

你看看这些功能，那么多东西呢。

Look at all these features.

根本做不到。

There's no way.

结果一周内，大多数人还真做出来了。

And within a week, most people had had done it.

那时候这事儿完全前所未见，大家都惊呆了。

And at the time that was completely novel, and everybody started freaking out.

现在你只要把那些功能丢给AI代理，它基本一次就能搞定。

Now, you can if you give all those features to an agent, they can pretty much one shot it.

这一年来，从原本得花一周时间反复调整，到现在几乎一键完成，变化太大了。

And that's been in the course of a year, it went from, you know, with a week's worth of work and massaging to basically it being one shot double.

正因如此，我现在想招的软件工程师，比以往任何时候都多——因为我能做的东西，比以前多太多了。

But as a result, I want to hire more software engineers than ever, because I can build so much more than I could before.

离真正能独立完成开发，还差得远呢。

It's so far from being totally independent and able to build stuff on its own.

所以我需要的是那些能站在更高一层、替我指挥这些工具的人，他们现在能干的活儿，翻了好几倍。

So I need the people who are going to be one layer of abstraction above that doing that work for me now they can do a lot more of it.

我再给你举个例子。

I'll give you another example.

我们去了一家公司做企业培训，就是那种为期两周、针对整个工程师团队的培训，那家公司特别有名，我本来想提个名字，但我不确定能不能说，不过我敢肯定，听这个播客的每个人都知道是哪家公司。

So we went out to a company and we do some corporate training stuff, you know, two week corporate training with this entire team of engineers is a well known company, my want to name drop, but I won't because I don't know if I'm supposed to, but 100 certain everybody listening to this podcast knows this company.

我们进去之后说，好吧，看看你们的六周计划，我们周一刚开始。

We went in and we said, Okay, let's look at you know, they had a six week roadmap, we started on a Monday.

结果到周二中午，六周的计划就全搞定了。

And we had finished the six week roadmap by middle of the day on Tuesday.

当时CTO和工程副总裁都在现场，直接惊呆了，嘴里直念叨：天哪。

And we're doing that with, you know, the CTO and the VP of engineering in the room going holy smokes.

这彻底改变了所有东西。

This is this changes everything.

但他们没说：‘那我是不是可以裁掉75%的人？’

But their react, their response wasn't okay, I can get rid of 75% of my people.

他们的反应是：天啊，我脑子里那些本来觉得要好几年才能搞定的事，现在居然都能立刻做出来！

Their response was, wow, all these things that have been in my mental roadmap that I thought would take years, we can do it all right now.

事实上，我是不是该从Gauntlet多招点工程师？因为能干的事实在太多了。

In fact, can I hire more engineers from gauntlet because I want to do you know, there's, there's so much more to do?

我觉得，做像Kelly这样的东西，最难的其实是搞清楚用户到底想要什么。

And I think the most difficult part of building something like Kelly is knowing exactly what the user wants.

一旦你把需求定义得特别清楚，再把它们写成代码，这事儿就没那么值钱了。

Translating that into like, once you have something that's really well defined, translating it into code, you don't get a lot of points for anymore.

软件工程师的工作正在从‘拿个PRD写代码’，转向去琢磨用户真正想要什么，然后确保他们得到的就是这个。

But a software engineers job is evolving from, you know, taking a PRD and writing lines of code, to figuring out what people want, and then making sure they have exactly that.

你就往上挪了一层抽象。

So you just move up one layer of abstraction.

我觉得，过去一年发生在软件工程师身上的事，马上就要轮到几乎所有其他白领了——以前你做X就能加分，

And I think what happened to software engineers in the past year is about to happen to pretty much every other white collar professional, where you used to get points for doing x.

现在呢，加分的是你搞出个系统，能几乎瞬间完成X，还得管好这个系统。

Now you're going to get points for creating a system that will do x quasi instantly and managing the system that will do that.

我觉得，我们可能会不停地往更高层的抽象往上走。

And, you know, I think we may keep working up more and more layers of abstraction.

但真正发生的是，你产出的东西越来越多，每个人手里的软件也越来越多。

But I think what happens is you just get more and more output, everybody gets more and more software.

如果你是一家SaaS公司，以前靠着产品难复制、市场被你垄断，那现在可就麻烦了。

If you're a, you know, SaaS company that had a stranglehold on the market, because it would take a long time to build what you had, yeah, you're gonna be in trouble.

但我认为，绝大多数软件公司其实并不是这种情况。

But I think that's not what actually the vast majority of software companies are.

我看那些来Gauntlet招聘我们人才的公司，个个都特别懂客户的需求，能精准做出符合客户心意的产品。

When I look at the companies that are coming into Gauntlet and hiring from us, they are experts at understanding the concerns of their customer and building stuff to meet the needs of their customers.

我没法光靠我最厉害的工程师从外面复制这种能力，因为我根本摸不清他们客户到底想要什么，除非你能把所有需求都理清楚。

I couldn't, you know, I couldn't take my best engineer and replicate that from the outside, because I don't understand all of the needs and wants of their customers, until you can find a way to define it all.

但我觉得，我们离真正能实现这个目标，还早着呢，路还很长。

But I think I we got a long time and a long way to go before.

我的意思是，会不会有岗位被取代？

I mean, will there be some job displacement?

肯定会有。

I'm sure.

但我认为这绝对是件好事。

But I think it's going to be a really good thing.

我觉得那些积极适应、努力理解这个新世界的人，会变得越来越强、越来越有影响力、越来越有价值。

And I think the people that are leaned in and figuring out how to make sense of the new world are going to just get better and better and more powerful and more powerful and more valuable and more valuable.

我敢说，所有通过Gauntlet的人，没有一个在另一边赚得比以前少。

I don't think there's been a single person who's come through gauntlet, who is making less money on the other side.

我每天都能看到有人靠学AI，十周内收入翻倍甚至三倍。

And I see every day people who are doubling or tripling their income in ten weeks by learning AI.

那咱们聊聊怎么做到吧。

So let's talk about how

加密货币在这个新世界里扮演什么角色？它给凯莉带来了哪些超能力？

crypto fits in this new world and what superpowers it gives Kelly.

有意思的是，我们目前谈软件和产品时，总想着‘人’是客户、是用户，对吧？

Because one thing that's really interesting, the the way we've been talking about things in our lens on on software and products right now is that, oh, the human a human is a customer or a human is a user.

没错。

Right?

但未来，我们可能是在为AI代理开发产品。

In the future, we might be building products for AI agents.

对吧？

Right?

所以凯莉可能会开发各种软件工具，但对象不是人类，而是其他AI代理。

And so Kelly might be building, you know, all sorts of software tools for other AI agents rather than humans.

但我觉得，加密货币在作为AI驱动力方面，几乎是独一无二的。

But crypto is kind of, I think, almost uniquely positioned to be an AI powering technology.

我看到一条推文。

And I saw this tweet.

其实是我时间线上出现的。

Actually, it came on my timeline.

我觉得是今年早些时候你发的，奥斯顿。

I think it was early in the year from you, Austen.

你说：我知道这听起来疯了，但你先听我说。

You said, I know this will sound insane, but hear me out.

自主AI代理，正是加密行业一直在等的那个杀手级应用场景。

Autonomous AI agents is the killer use case the crypto industry has been waiting for.

每个智能体都需要有自己的钱包，能无缝又快速地完成交易。

Every agent will need its own wallets and be able to seamlessly and quickly transact.

咱们聊聊这个。

Let's talk about that.

为什么说加密货币是智能体的终极应用场景？

Why is crypto the killer use case for agents?

或者说，智能体才是加密货币的终极应用场景？

Or why are agents maybe the killer use case for crypto?

别人总说我是个从Web2慢慢摸到Web3的人。

People talk about me as being like, I'm the web two guy that's slowly reaching into the web three world a little bit.

我其实很早就对加密货币感兴趣，那时候囤了一堆比特币，结果全在Mt. Gox那儿打水漂了。

You know, I was like, was really interested in crypto early on had a bunch of, you know, Bitcoin that went to nowhere in Mt.

就是那种情况。

Gox kind of thing.

所以我确实是最早一批入局的，但后来对加密圈挺失望的，原因嘛，你懂的，太多太多了。

So I, you know, was really, really early to this stage, but I've been kind of like, disenfranchised by crypto, there's just so much for all the reasons that anybody could guess on me to define them.