#246：AI 问答——来自 MAICON 2026 的现场报道：通用人工智能、AI 成本与 GPT 时代的终结

而且就像，不用再让它生成一个HTML文件，这是我们经常做的事情，比如，嘿，把这个转成HTML文件。

And it's just like, instead of asking it to build an HTML file, which is something we commonly do is like, hey, turn this into an HTML file.

你可以把HTML文件发布出去，然后分享给别人。

You can publish the HTML file and share with people.

然后就是，哦，好的。

It's like, oh, Okay.

嗯，那挺酷的。

Well, that's pretty cool.

那可能会改变我们做事的方式。

That might change the way we do something.

所以我也说不准。

So I don't know.

对我们来说，关键就是这个动态学习的概念。

For us, it's about this dynamic learning idea.

就是说，我们会更频繁地，比如每周或每月，去了解这些模型和平台有什么新东西。

Like, we're just going to push a lot more into like regular weekly, monthly cadence of what is new in these models and platforms.

这对我们来说意味着什么？

What does it mean to us?

它如何影响我们的工作？

How does it affect our work?

然后我们会通过我们的AI学院把这些带给人们，让它更像实时学习。

And then we're going to try and bring that to people through our AI Academy and make it more like real time learning.

我在Google Gemini的研讨会上尝试做的一件很酷的事情，但你可以把它应用到任何模型上，就是有一个完整的环节，关于如何把它当作一个真正的战略思维伙伴来使用，并且通过一个有步骤的、过程驱动的方式来提问或对话，因为很多人只是写几个词，或者不来回对话，或者不带着真正聪明的问题和洞察力去进行战略性的讨论……

One thing that was really cool too that I tried to do in my workshop on Google Gemini, but you could apply this to any model, is there was a whole segment on using this and approaching this more as a true strategic thought partner and going through, like, a measured process driven query or conversation with the tool because so many people just write a few words or don't converse back and forth or don't have a strategic discussion with it with really smart questions and insight...

从中获取见解，就像他们对待顾问或顾问一样。

Get insights from it same as they would as, like, a consultant or adviser or whatever.

我的观点是，明天即使Gemini的一切都变了，也没关系。

My point here is that it doesn't matter tomorrow if everything about Gemini changes.

这个概念真的非常重要。

That concept is really, really important.

所以也要关注一些更持久的策略，比如我们的用例模型、基于问题的模型、创新模型。

So even focusing also on some of these more evergreen strategies, like our use case model, the problem based model, innovation model

这个是我大概2018年创建的。

Which I created in, like, 2018.

我们现在还在用这个来教学。

We still teach based on it.

现在依然非常适用。

Still extremely relevant.

很适用。

Relevant.

我觉得这类东西真的非常宝贵。

That kind of thing is also really invaluable, I think.

教学框架。

Teaching framework.

一个很好的例子就是……

And a really good example of this is...

所以Mike正在做这个Gemini……

So Mike's doing this Gemini...

我们怎么称呼它们？

What do we call them?

Gemini for business？

Gemini for business?

Gemini for network？

Gemini for network?

Gemini for business。

Gemini for business.

所以上周，谷歌举办了一个活动，大概一周前，甚至不到一周前……

And so last week, Google holds an event, like, a week, not even a week ago from...

他们……

They were...

而且它……

And it...

他们怎么称呼这个活动？

What do they call it?

Gemini……

Gemini...

那个新的智能体工具叫什么来着？

What's the new agent harness thing?

其实它就还叫Gemini，你要是觉得这让人困惑，那你不是一个人。

It's technically just called Gemini, which if this is confusing to you, join the club.

现在好像变成个什么工具了。

Like a harness now or something.

有点像那种智能体框架

It's like kind of an agentic harness

之类的东西。

of thing.

把Gemini重新包装成了别的东西。

Rebranded Gemini as something else.

所以那是在Mike的研讨会之前六天，或者四天前。

And so that was six days before Mike's workshop or four days before it.

所以我正在给迈克发消息，嘿。

So I'm I'm messaging Mike, hey.

你会根据这个改变工作坊的内容吗？

You're change the workshop based on this?

然后昨天在你的工作坊期间，如果我没记错的话，他们发布了氩模型。

And then during your workshop yesterday, if I'm not mistaken, they released argon model.

是的。

Yes.

Gemini四氩变成了，就像，

The Gemini four argon became, like,.

所以即使在迈克做现场工作坊的实时学习环境中，在这四天里有两件事可能影响他要在工作坊里教的内容。

So it's like even in a real time learning environment where Mike's doing a live workshop, you have in this four day span two things that could affect what he's gonna teach in the workshop.

好的。

Alright.

下一个问题很有意思，我觉得对很多听众来说都很相关。

This next question is fascinating, very relevant, I think, to several people listening.

那Claude或者ChatGPT，或者随便什么工具，它们的价格会有什么变化吗？

What, if anything, is going to happen to the cost of Claude or ChatGPT or insert whatever tool here?

尤其是那两家，一旦Anthropic和OpenAI上市，必须开始盈利了，我们会不会看到它们开始大幅涨价？

But those two, especially, once Anthropic and OpenAI go public and they have to start turning a profit, are we gonna see them start jacking up prices?

不会。

No.

那这怎么运作呢？

How's that gonna work?

它们做不到。

They can't.

这个价格……

The the the price of the...

所以现在的顶级模型，比如Opus 5.5，我觉得，是不是……

So the Frontier model today, so take, like, an Opus 5.5, I think, is that...

那可能是它们最强大的模型了。

That's probably their most power.

是Fable。

It's Fable.

我不知道Fable能不能用。

I don't know if Fable works.

在Opus 5上。

On the Opus five.

好的。

Alright.

那我们就假设Opus 5.5是最强的。

So let's say Opus 5.5 is it.

这周挺忙的。

It's been a busy week.

十二个月后，它的成本会便宜十倍。

It's gonna cost 10 x less in twelve months.

所以到某个时候，你会作为组织说，好吧。

And and so at some point, like, you're gonna say as an organization, like, okay.

这个模型已经够好了。

This model is good enough.

我在播客上公开说过这件事。

Like, I said this publicly on the podcast.

我当时觉得Sonnet 4.6就很好。

Like, I thought Sonnet 4.6.

如果你告诉我这是我能用的最后一个模型，我会高兴得不得了。

Like, if you told me this is the last model I'm ever going to have, I would have been ecstatic.

我当时就想，太好了。

And I'm like, great.

我们就用这个吧。

Like, we'll just ride with this.

我们整个公司就围绕这个模型来建。

We'll build the entire company around this model.

而且我敢肯定，那个模型的成本比刚出来时便宜了五倍，因为现在他们已经有Sonnet 5点几的版本了。

And I'm sure that model costs five x less than it did when it first came out because now they're on whatever Sonnet five point something.

所以我觉得，对于企业的大多数使用场景来说，只要愿意用上一代最先进的模型，成本就会越来越低。

So I think the the vast majority of use cases for businesses, it's just gonna get cheaper as long as you're willing to use the previous state of the art model.

所以这将会……

And so that's gonna...

这些实验室之间的竞争会迫使这种情况发生。

The competitive nature of these labs is going to force that to happen.

他们不会一直把价格抬高的。

They're not gonna just keep making it more expensive.

所以他们会做的是，把更先进的模型卖给那些真正需要超强推理能力的公司。

So what they're gonna do is they have to sell the more advanced model to the companies that need the really, really advanced reasoning.

但对他们来说，关键是他们其实并不在乎这个。

But the key for them is they actually don't care.

他们只需要把智能体推广开来，让公司里的每个人都用上这些工具。

They just need to infuse agents and get everyone in companies using these things.

所以你想，Anthropic 和 OpenAI 都要上市了，估值大概在一万亿到两万亿美元之间。

So if you think about how, I mean, Anthropic and Open AI are both gonna IPO at somewhere between 1 and $2,000,000,000,000.

然而，你去任何一家大型企业看看，几乎都没有哪家能全面采用这些工具，让每个人都每天把它融入工作流程。

And yet when you go around any major enterprise, there are almost zero of them that have full scale adoption of these tools to where everyone is infusing it into their workflow every day.

所以这些模型的市场潜力还远远没有被挖掘出来，即使每个token的成本降低十倍，你也会发现……

So the amount of untapped market potential for these models is so vast that even if you're paying 10 x less per token to do it, you're...

随着你在整个公司推广使用，对token的需求几乎是无限的。

You almost have an infinite need for tokens as you expand the use of it across the company.

我觉得他们并不在乎这个。

I don't think they care.

就是说……

Like...

而且开源模型也会变得很接近，所以他们必须接受这一点，还得考虑未来的商业模式会是什么样。

And plus open weight models are gonna be comparable, so they have to accept that, and they have to, like, think about what the business model is going to be.

Anthropic的IPO会让我们真正看清商业模式到底是什么。

Anthropics IPO is gonna be a really interesting look into what really is the business model.

比如，什么才是可持续的？

Like, what is sustainable here?

这些公司到底有什么护城河？

What are the moats that these companies have?

这也跟刚才说的有点关系。

Somewhat related too.

今年在MAICON上特别有意思。

It's been interesting, especially this year at MAICON.

大家显然都在用像Muse、Grokbod或者Instinct这样的工具。

Everyone is obviously using things like Muse or Grokbod or Instinct.

还有Dots。

Dots.

反正就这些吧。

Whatever.

我意识到我们现在还处在模型本身很重要的阶段，但有一半的对话里，有人会说，Grock bot做了这个或者那个。

And I realize we're still in the time where the model powering that matters, but half the conversation, someone's like, well, Grock bot did this or that.

我甚至都不知道那是什么模型。

I don't even know what model

Grock 已经上线了。

Grock is on.

我觉得他们现在用的是 Opus。

I think they're using Opus now.

他们是不是在考虑用图形模型？

Are they thinking using graphics models?

但重点是，我也不知道。

But the point is I don't know.

所以更像是，如果你……

And so it's more like if you...

这几乎有可能，至少在个人代理层面，会被抽象化，就像你根本不在乎用哪个模型。

It's almost possibly, at least at the personal agent level, going to kind of get abstracted away where it's like, you kind of don't care.

就像我的代理能正常工作就行。

It's just like my agent works and stuff.

我相当确信，明年这个时候的 MAICON，普通用户的体验将不再需要选择模型。

I'm fairly confident that like this time next year at MAICON, the the average user's experience, you will not be choosing your model.

我觉得他们早就知道，选模型基本上是开发者的事。

Like I think they've known for a long time that model choice is basically a developer thing.

如果你把这个给任何一家普通公司的总监或副总裁，跟他们说，嘿，这是你的助手，这里有12个模型，你写邮件、做方案或者干别的什么的时候可以选。

And that if you give this to any director, VP at any normal company and say, hey, here's your copilot, Here's your 12 models you can choose from when you're writing your emails or building your proposals or doing whatever.

他们大概只会说，行吧。

They're gonna be like, okay.

我就用默认的那个好了。

Like, I'll just use whatever it's turned on to.

我根本不知道有什么区别。

I don't even know.

有时候我打开ChatGPT，看到什么6.1，什么低中高，我都搞不清楚。

Like, sometimes I'll go into ChatGPT and it's like 6.1 sole light, medium, high.

我甚至都不知道它们之间有什么差别了。

Like, I don't even know the difference anymore.

所以我也说不准。

So I don't know.

我觉得模型选择只是四年前的一个遗留问题，不知道为什么他们就是不肯直接一刀切掉。

I think the model choice is just like an artifact of, you know, four years ago that for some reason, they haven't just ripped the Band Aid off.

我……

I...

我猜，不知道谷歌会不会先这么做。

My guess is, I don't know if Google's gonna do it first.

总会有人先这么做的，就是你根本不知道自己用的是什么，他们会根据你的提示词来路由。

Like, somebody's gonna do it first where you you just don't know what you're using, and they're gonna route it based on your prompt.

下一个问题是，我们应该怎么看待AI能做什么和AI应该做什么？

This next question is phrased, how should we be thinking about what AI can do versus what AI should do?

比如说，这对你来说很相关，因为AI能写书不代表它就应该写书。

So for instance, this is relevant to you, just because AI can write a book doesn't mean it should write a book.

你怎么决定哪些事值得自己亲自做，即使AI能帮你做？

How do you decide what's worth doing yourself even when AI could do it for you?

我不知道这个问题有没有一个统一的答案。

I don't I don't know that there's like one answer to this.

写作这个例子我总是喜欢用，因为对我来说，写作就是思考。

So the writing example is always a good one that I use because like for me, writing is thinking.

这是我处理信息的方式。

It's how I process information.

就像我准备播客的时候，我之所以每篇文章都读、每个播客都听、每个视频都看，就是因为做这些工作本身就是意义所在。

It's like when I prepare for the podcast, it's why I actually read every article, listen to every podcast, podcast, watch every video myself because doing the work itself is the whole point of it.

我能坐在这里回答你抛出的任何问题，是因为我花了十五年研究这个领域，读了所有这些资料，做了这些工作。

Like my ability to sit here and just answer whatever question you throw is because for like fifteen years, I've been studying the field and reading all these things and doing the work.

如果这十五年我让Claude（如果当时有的话）帮我总结所有东西，然后给我一些要点让我背下来跟你复述，那做这些工作的意义就没了。

If I'd spent fifteen years asking Claude, had it existed, to just summarize everything for me and then spit out some bullet points that I could say and regurgitate to you, like, you've lost the point of doing the work.

所以我觉得，部分问题就是，你最初为什么要做这些工作？

So So I think some of it is just like, well, why are you doing the work in the first place?

目的真的只是为了生成一个像信息收集那样的输出吗？

Is the purpose literally just to create an output that's like information gathering?

所以我在工作坊里用了一个例子，就是关于AI Score的，这周我们刚把它公开测试，我需要创建产品FAQ页面。

So I think I use the example in my workshop of for AI score, which we just put into public beta this week, I needed to create the product FAQ page.

其实没人在乎我是不是真的写了那个产品FAQ页面。

Nobody cares if, like, I actually wrote the product FAQ page.

从真实性的角度来看，他们不会说，我不觉得保罗真的写了这个页面。

Like, from an authenticity perspective, they're not saying they're saying like, I don't think Paul actually wrote this page.

我的……

My...

第一，页面上根本没有我的名字，所以你甚至不知道是我创建的。

One, my name's not even on it, so you don't even know that I'm the one created it.

第二，那个页面的全部目的就是给有疑问的人提供关于产品如何运作的信息。

Two, the whole purpose of that is to provide information to someone if they have a question about how the product works.

它不是用来听我的个人意见或我们对某事的看法的。

It is not to hear my personal opinion or our perspective on something.

所以在这种情况下，这显然是个合适的应用场景。

So in that instance, it's like, that's an obvious use case.

就像……

It's like...

而且它……

And it...

我就是这么做的。

It's how I did it.

我直接打开Claude，那个我做了七个月的Building AI项目，跟它说，把我们七个月一起开发这项技术的所有上下文都拿上，把我们之前一起做的面向公众的营销页面也拿上，这里有团队问我的80个关于产品如何运作的问题，以及人们可能想了解的内容。

I literally went into Claude, my seven month project I've had with the Building AI score and said, take all seven months of context of our work together, developing this technology, take the public facing marketing page that we already created together, and here's 80 questions the team asked me about how the product works and what things people might want

把它变成一个产品FAQ页面，就像你在写一个一流的SaaS产品FAQ页面那样，语气要非常平易近人。

It into a product FAQ page as though you were writing a best in class, like, SaaS product FAQ page and make it super approachable tone.

搞定。

Boom.

完美。

Nailed it.

就像……

Like...

而且它……

And it...

大概十到三十秒就写出来了，我花十个小时都写不出来。

In, like, ten to thirty seconds wrote something that I wouldn't have been able to write in ten hours.

所以我觉得关键是要学会判断什么时候用AI没问题，比如我只是想从品牌那里获取信息和事实，跟我读一本书，假设是你写的，你投入了思考，这些确实是你自己的话，那期待就是真实的。

And so I think it's just like getting good at that judgment of when is something fine to just use AI because the expectation is I'm just trying to get information and facts and something from a brand, for example, versus I'm reading a book that I'm assuming you wrote, that you put thought into this, like these are actually your words, then the expectation is that it's authentic.

所以再说一遍，没有完美的答案，但我认为组织制定这些框架并教给人们很重要，尤其是年轻员工，他们还没有那种判断力，不知道什么时候可以交上去。

And so, you know, again, there's no perfect answer, but I think it's important that organizations develop these frameworks and teach people on these frameworks because especially with younger employees, they're not gonna have that judgment yet of when is it okay to turn this in.

所以我老板让我写一份战略文件，我就想，天哪，Claude写得比我好多了。

So my boss asked me to come up with a strategic document, and I was like, man, Claude's better writing these than me.

所以你就让Claude写，然后交上去。

So you just have Claude write it and you hand it in.

然后老板说，进来吧。

And the boss says, like, come on in.

我们来谈谈你刚写的这份文件。

Let's talk about this document you just created.

比如，我不同意这三点。

Like, I don't agree with these three points.

你为什么选这个？

Why did you pick this?

然后那个人就说，我也不知道。

And the person's like, I have no idea.

是Claude写的。

Claude wrote it.

就是，我我根本不知道我们为什么选这个。

Like, I don't I don't know why we picked it.

你已经失去了重点。

You've lost the point.

重点是要完成工作。

The point was to do the work.

重点是要能为自己提交的东西负责。

The point was to, like, be able to stand behind something you've submitted.

你提到这个例子真有意思。

It's funny you mentioned that example.

我们认识一个人，我就不点名了，他笑着跟我讲了个故事，说他们有个员工就干了这么件事，而且那人连面试问题都没等到，因为这个人一看就知道，这20页的文件明显是Claude直接生成的，完全没经过人手。

Someone we know who I will leave unnamed was telling me and laughing as they told me a story about how one of their employees did exactly that thing, except they didn't even get to the interview question because this person knew immediately this 20 page document was very clearly immediately out of the mouth of Claude and, like, not touched at all by a human.

然后他就说，嘿。

And so they said, hey.

给我这20页的五个要点。

Give me five bullet points about these 20 pages.

然后那人就说，关于什么的？

And they were like, about what?

就是，随便里面任何内容。

And, like, literally anything in it.

任何一点。

Any single thing.

然后那人就说，什么意思？

And they're like, well, what do you mean?

然后他们，他们根本说不出来。

And then they they couldn't do it.

然后他们说，好吧。

And they're like, alright.

他们需要你回去重做这个。

They need you to go redo this.

然后，哪一页？

And, like, which page?

全部。

All of them.

你懂吧？

You know?

所以这个……

So this...

重点是，这种事情已经在发生了。

The point is this stuff is happening already.

我不知道那条线在哪里，但显然，需要有一条线。

I don't know where that line is, but clearly, some line needs

必须得定下一条线，而且得成为文化的一部分。

to has to be established, and it has to be part of the culture.

对。

Yes.

好。

Okay.

下一个问题。

Next one.

通用人工智能，也就是AGI，对普通人来说到底是什么样的？

What does AGI, artificial general intelligence, really look like to the regular person?

有人说我们现在已经有AGI了，但AI在很多基本的事情上还是不行。

Some people say we have AGI now, but AI still struggles with tons of basic things.

比如，AGI在普通人的生活中到底会是什么样子？

Like, what would AGI actually look like in an ordinary person's life?

我们怎么才能看出区别呢？

Like, how would we even know the difference?

回到我们开头提到的凯文·鲁斯的新书《AGI编年史》，因为我刚听完他的演讲，他确实讲了个挺有意思的事，关于“感受AGI”以及它到底是什么鬼。

So going going back to how we started this Kevin Roos' new book, AGI Chronicles, since I just sat through his talk, he actually went through this pretty entertaining thing about feel the AGI and what the hell is it actually.

所以“感受AGI”这个说法，特别是OpenAI在用。

So feel the AGI is sort of like open AI in particular uses this.

他还说，OpenAI的联合创始人伊利亚·苏茨克弗，以前会走到别人桌前，像尤达那样说：“感受AGI。”

And he was actually saying that Ilya Suttkevur, who was a cofounder of Open AI, used to go around to people's desks and just say, like, Yoda, like, feel the AGI.

就像……

Like...

就这样。

And that was it.

然后凯文采访了伊利亚，问他：“你说‘感受AGI’是什么意思？”

And so Kevin interviewed Ilya, and he's like, what do you mean when you say feel the AGI?

“AGI是什么？我们怎么知道自己感受到了AGI？”

Like, what is AGI, and how would we know if we're feeling the AGI?

然后他讲了个特别搞笑的故事，关于伊利亚，他回答问题的时候基本上像个哲学家。

And so he told this hilarious story about, like, Ilya, who's basically, like, a philosopher when he answers questions.

他说，这有点像你看到地球放大到一个人在海滩上什么的，然后突然一下子缩小到整个宇宙，背景里还放着巴赫的音乐，然后他……

He's like, oh, you know, it's kinda like when, you know, you see that you zoomed in on Earth like someone on a beach or whatever, and then all of a sudden you're like zoomed out to the universe and you hear Bach playing in the background and and then he...

这就是他的回答。

Like, that was the answer.

凯文就说，我完全不知道那是什么意思。

And Kevin's like, I don't know what the hell that means.

你在说什么呢？

Like, what are you talking about?

所以我觉得，对于AGI，我现在的理解是，当模型的能力大致相当于你团队里那些普通水平的成员时。

So I think tangibly for AGI, the way I think about it is when the model is roughly the equivalent of the b players on your team.

这是我最近想到的最简单、最直观的理解方式。

It's kinda like the simplest, most tangible way I've I've thought about it recently.

比如说，如果我们跟营销人员聊天，你看着你团队里的那些人，然后说，好。

So if we're talking to marketers and you look across a team of people on your staff and you say, okay.

这些人做这20项任务，我要拿Claude Opus 5.5或者OpenAI的6.1 Astra，让这个AI也做同样的20项任务。

Like, these people do these 20 tasks, and I'm gonna take Claude Opus 5.5 or 6.1 Astra from OpenAI, and I'm gonna give that AI those same 20 tasks.

它能不能达到你团队里普通人的水平？

Is it on par with like an average person on your team?

按照这个定义，我们一年前就已经有AGI了。

Under that definition, we had AGI like a year ago.

其实我们大多数人做的任务里，很少有它达不到普通员工水平的，可能质量上需要一些修改，或者需要第二稿。

Like, there are very few tasks that most of us do where it's not already, like, b player level in terms of quality, which means might need some edits, might need a second draft.

你可能需要调整一下提示词，但基本上它已经能产出任何部门普通员工能做出的工作成果了。

You might need to iterate on the prompts, but, like, it's generally capable of creating the work product that an average person across pretty much any department would create.

所以我不觉得……

So I don't...

凯文展示了OpenAI和Anthropic的不同定义，其中Dario那个著名的说法是，AGI就像是一个由天才组成的数据中心。

I mean, Kevin showed, like, the different definitions from OpenAI and Anthropic, which is, a data center of geniuses is, like, Dario's infamous way of explaining it.

他还展示了谷歌DeepMind的定义，而且……

He showed Google DeepMind, and it...

虽然大家意见不一，但我一直是这么想的。

Like, no one agrees, but that's how I always thought about it.

最直观的就是拿员工来比，就是说，它能不能达到你普通员工的水平？

It's the most tangible is equated to workers, and it's like, is it on par with your average worker?

然后下一级的人工超级智能，就是它能不能比你最优秀的员工还厉害？

And then the next level of artificial superintelligence is, is it better than your best worker?

就是说，它基本上是不是已经超越人类了？

Like, is it pretty much just superhuman?

然后我……

And I'm...

我今天早上在主题演讲里也说了，我基本上认为，如果你直接快进十二个月，假设这个模型……

I said on stage, I think, this morning in my keynote, I'm kind of under the assumption that if you just fast forward twelve months, just assume the model...

你给了它正确的数据，你以一种聪明的方式使用它，就是说，你知道怎么跟这些模型合作。

That you've given it the right data, that you're you're you're using the model in an intelligent way, like, you know how to work with these models.

十二个月后我们拥有的模型，会达到A级水平甚至更高。

The models we're gonna have in twelve months are are gonna be a player level and beyond it.

基本上你在公司任何部门做的任何任务，它都能胜任。

Pretty much every task you would do in any department of a company.

所以我很难接受这不算AGI，如果按它最初的定义来看的话。

And so I have a hard time accepting that that isn't AGI in terms of what it was originally meant to be.

我觉得如果你回到十年前、二十年前，把今天的模型给当时研究AGI的AI研究人员看，他们肯定会说，这已经超越了。

I think if you rewound ten, twenty years ago and you gave AI researchers who were thinking about AGI then today's models, they would be like, this is beyond it.

这明显就是超级智能。

Like, this is so obviously superintelligence.

可现在我们却在这儿争论一个已经变得没什么意义的术语。

And yet here we are just debating a term that's become, you know, basically meaning.

这个标准好像也在变，因为我自己想想，出去散步的时候用语音模式，特别是配合编码器。

It seems like that yardstick is just moving too because, like, I even think about going on a walk and using voice mode specifically with codecs.

不只是语音模式，而是语音模式加上编码器，让我家里的电脑帮我创建文件、在网上做事。

Voice mode alone, but, like, voice mode with, like, codecs where my computer back home is creating files, doing stuff for me on the Internet.

人们可能也会这么说个人代理。

People probably say this about personal agents too.

但能直接说，去帮我做这三件事，这种感觉。

But, like, being able to talk and say, go do these three things for me.

把这个改一下，那个改一下，还有这个幻灯片。

Do this to that, this to this slide deck.

去把这封邮件发了。

Go send this email.

就像说，嘿。

It's like, hey.

我正在处理呢。

I'm working on that.

太好了。

Great.

好的。

Okay.

都搞定了。

It's all done.

接下来呢？

What's next?

比如说，你要是把那段录音发给二十年前的人，他们肯定会觉得，这简直是科幻电影里的东西。

Like, if you sent that recording to somebody twenty years ago, they would be like, this is in sci fi movies this way.

感觉我们基本上已经达到那个水平了。

Like, it it feels like we're basically there.

对，关键就在于，你用它完成了一个工作任务，而且它做得很好。

Well, and the key there is, like, you did that for a work product, and it nailed it.

但没过两分钟，你突然想起来，哎呀，糟糕。

But then like two minutes later, you're like, oh, no.

我忘了报税了。

I forgot to do my taxes.

你能不能帮我把这些文件里需要的东西都整理出来，然后把PDF表格填好，等我回家的时候准备好？

Like, could you go in and like pull everything I need out of these files and and fill out the PDF for me and have it ready for me when I get home?

还有，哎呀，我忘了该给老婆送结婚纪念日礼物了。

And oh, I I forgot I was supposed to send an anniversary gift to my wife.

你能不能帮我处理一下，这怎么可能？

Could you like go, how is that not?

就像我们最初对通用人工智能的设想那样。

Like, what we initially thought AGI would be.

现在一个系统就能基本完成你要求的任何事。

It's a single system that can now do basically anything you ask it to do.

好的。

Okay.

下一个问题。

Next question.

如果人工智能能解决世界上最难的数学问题，就像最近解决千禧年大奖难题那样，那为什么它不能解决由它自己造成的失业问题呢？

If AI can solve the world's most advanced math problems like it's been recently doing with the Millennium Prize, why can't it solve also the problem of unemployment caused by AI?

所以……

So...

这是个很大的问题，但显然，我们不能按个按钮就解决失业问题。

And this is a big question, but, obviously, we cannot just click a button and solve unemployment.

但关键是，人工智能能不能帮我们解决这类问题，哪怕这些问题可能是它自己造成的？

But the point here is could AI help us solve problems like this that it might also end up creating?

我……

I...

这个……

The

实验室认为答案是肯定的。

labs think the answer is yes.

就是说，实验室相信如果我们能走到最后那一步……

Like, the labs believe that if we just get to the end game...

我不知道，我管它叫终局。

I don't I call it the end game.

也许那才是游戏的开始。

It might be the beginning of the game.

但如果我们真的达到了超级智能的水平，那所有这些问题都能解决。

But, like, if we get to the point where we truly have superintelligence, then it'll fix all of this.

所以如果我们回到五年前，所有大型AI实验室、云公司、科技公司，都会承诺实现净零排放。

So if we go back, like, five years ago, all the big AI labs, the cloud companies, the tech companies, everyone is gonna be like net zero emissions.

所以当时一切都围绕着我们要拯救环境。

And so it was all about like, we're gonna save the environment.

我们要把排放到世界上的所有碳都清除掉。

We're gonna like get rid of all the carbon that we're putting out into the world.

到了某个时候，他们突然意识到，我们实际上要建一万个数据中心，这反而会增加排放。

At some point, they're like, wait a second, we actually have to build like 10,000 data centers that are actually going to increase emissions.

但别担心，因为我们要造出超级智能，到2030年它就能想出办法解决我们制造的所有环境问题。

But don't worry about it because we're gonna build super intelligence and then by 2030, it'll actually figure out how to fix all the environmental stuff that we're gonna create.

所以他们做的很多事情，本质上就是觉得，只要我们投入足够的算力，等它足够强大了，它自然会替我们解决。

And so a lot of what they're doing is truly, it'll figure it out for us once we throw enough compute at it and once it gets good enough.

它一直在解决的数学问题，部分原因在于你可以验证答案对不对。

So the mathematic stuff that it's been solving is in part because you can prove if it's right or not.

所以对于有没有答案，是成功还是失败，是很明确的。

So there is, like, an obvious did this work or did this not in terms of having an answer.

这也是为什么很多时候可以用游戏来训练它，因为有得分。

It's why you can train it in games a lot of times because there's there's a score.

就是有办法去追踪记录。

There's like a way to keep track.

当你尝试解决一些棘手的问题，比如经济政策该怎么定，就业问题怎么处理时，

When you try and solve something messy like what do we do with economic policy and what do we do with jobs?

没有那种“我们拼命干，把十万块GPU砸上去一周就能得到正确答案”的情况。

There is no, like, we can work really hard and throw a 100,000 GPUs at it for a week and get to the right answer.

我们根本不知道正确答案是什么。

We don't know what the right answer is.

我们还得真的去测试，把它放到社会里去实践，看看行不行。

We still actually have to test it and put it out into society and see if it works.

所以，比如今天闭幕主题演讲的安德鲁·杨，我想他会谈到一些像全民基本收入或全民高收入这样的方案，作为应对人们可能失业的一种方式。

So, like, Andrew Yang today at the closing keynote, I think he's gonna talk about some of the things around, like, universal basic income or universal high income as one way that we can address the fact that people might not have jobs.

但这并不能解决我同时也失去了人生目标的问题。

That doesn't solve the fact that I also now have no purpose in my life.

对吧，好吧。

Like, okay.

谢谢每个月给我五千美元，但是我还是得面对家人，跟他们说我没有……

Thanks for my $5,000 a month, but, like, I still have to go and look at my family and be like, I don't have...

我两年都找不到工作。

I can't find work for two years.

所以这真的是一个很棘手的问题，涉及社会的很多方面，这也是为什么会有像全民基本服务这样的其他想法。

Like, so it's just a really messy thing that that covers a lot of areas of society, and that's why you have other ideas like universal basic services.

对，没错。

Like, okay.

我们也给人们提供医疗服务吧。

Let's provide health care to people too.

所以除了直接给钱让人活着之外，真正值得测试的好点子其实不多。

And so there's a lack of really good ideas that are worth testing other than let's just give people money to be alive.

所以我觉得，首先政府需要有解决问题的意愿，要承认这个问题正在出现，要承担起这个责任，并且说我们会主动寻找解决方案。

So I think that one, you need the will from the government to want to solve this problem, to accept that there is an emerging problem, to own that problem and to say we are going to proactively look for solutions.

现在的政治环境根本不愿意承认人工智能对就业有任何影响。

We are not in a current political climate that wants to admit that AI has any impact on jobs at all.

戴维·萨克斯，他最近算是政府非官方的代言人，你去看看他过去四个月的推特，看看他有多少次在抨击那些暗示AI会取代工作的人。

David Sachs, who's, you know, kind of a voice of the administration unofficially these days, just go look at his ex feed from the last four months and see how many times he's blasted people for implying that AI will take jobs.

所以这……

So it...

而这基本上就是政府的态度。

And that is the that is the position largely of the administration.

所以至少在美国，我们不愿意承认这可能会出问题，因为那会失去选票。

So at least in The United States, we are unwilling to acknowledge that it could go wrong because it would cost votes.

所以这……

So it...

就像，首先你得有决心去做，然后你才能真正把那些愿意提出难题的人聚在一起，开始尝试解决。

It's just like, first, you have to have the will to do it, then you actually have to bring the minds together who are willing to ask the hard questions to start trying to solve it.

而让我担心的是，我们就像……

And that's what worries me is like we're just...

我们甚至还没迈出第一步。

We're not even out of the starting gate doing that.

这也是我请安德鲁来的原因，因为他是少数几个真正提出‘我们可以试试这个’的人，不管你是否同意他的立场。

And it's why I wanted Andrew here because he's one of the few people, whether you agree with his position or not, who has actually taken a position of here's something we could try.

而这一点，无论是政客还是AI实验室的领导者，都缺乏。

And that's just lacking from politicians and from AI lab leaders.

他们甚至没有提出真正的想法。

They're not even offering real ideas.

只是……

It's just...

只是说‘它会变得丰富，会很棒’。

It'll just be abundant and it'll be awesome.

然后等我们到了那一步再看会发生什么。

And we'll see what happens when we get there.

好的。

Okay.

下一个话题。

Next up.

怎么……

How...

随着人们在职业生涯中不断前进，你觉得他们会如何与AI同行？

As people progress through their career, how do you see them traveling with AI?

比如，他们用AI做的事情，有多少能从一个职位带到另一个职位？

Like, how much of what they do with AI will they be able to take from position to position?

从法律上讲，一点都不能。

Legally, none.

所以，如果你想想雇佣法和我们为组织工作时签的那些合同，如果你创造了什么，那知识产权就属于你的雇主。

So, I mean, if you think about employment law and the contracts we all sign when we work for organizations, if if you create something, that intellectual property belongs to your employer.

所以，如果你是一家公司里一个前瞻性的AI总监，或者你是一个编排者，就像你在构建所有这些，你在制定所有这些策略，你构建了所有这些代理，还有所有那些工作流程等等，当你走出门时，你什么都不拥有。

So if you're an AI forward, you know, director within a a company and you build or you're an orchestrator, like you're building all this stuff, you're devising all these strategies and you build all these agents and you know, all these workflows and everything and you own none of it when you walk out the door.

所以这……

So that's...

挺棘手的。

It's tricky.

但这也涉及到你在那家公司时的个人形象权。

But that also exists within, like, your likeness when you're at that company.

所以我们可以拿你和我来举例。

So, I mean, we could do this with you and I.

就像，谁拥有我们的形象权……

It's like, who owns our likeness to...

如果他们想把它做成虚拟形象。

If they wanted to turn it into, like, avatars.

比如，如果他们决定，嘿。

Like, if they decide, hey.

哦，迈克，有你们这些人类真好，但我们要把你们做成数字版本。

Oh, Mike, it was great having you humans, but, like, we're actually gonna spin up some digital versions of you guys.

还有人的想法。

And people ideas.

我知道。

I know.

我们已经有我们的婴儿版了。

We already have the baby versions of us.

我不记得是哪个

I don't remember what

那一集了。

episode that was.

我已经……

I'm already...

我已经知道这会搞出点事情来。

I already know this is gonna result in something.

对。

Yes.

抱歉。

Sorry.

但其实我们已经在合同里处理过这个问题了。

But we've addressed this actually in our contracts.

所以再说一遍，关键是要有AI前瞻性，把这些问题想清楚，这就是为什么你需要律师。

So, you know, again, it's it's about being AI forward and thinking this stuff through, and this is why you need your attorneys.

这就是为什么你需要有AI前瞻性的领导者，他们懂得去问这些问题。

It's why you need to have AI forward leaders who understand to ask these questions.

这事儿挺乱的，但简单来说，答案是：什么都没有。

It's it's it's messy, but the short answer is nothing.

你在公司创造的任何东西，离开公司时都不归你所有。

Nothing you create for a company do you own when you leave that company.

跟这个有点相关的是，这里有个后续问题，就是你有没有见过公司考虑怎么重新培训或整合那些来自使用不同模型公司的人？

You know, somewhat related to this was a follow on question here of, like, have you seen companies consider how to retrain or integrate people who come from a company using different models?

比如，是不是有很大的适应过程？

Like, is there a big adjustment?

我以前真没怎么想过这个，但我能理解这个问题从哪来。

I've never really totally thought about this, but I could see where this is coming from.

从Copilot公司跳槽到Claude公司，或者反过来，员工是不是需要很大的适应？

Is there a big adjustment for employees coming from Copilot shop to Claude, vice versa, etcetera?

我不知道。

I don't know.

这个我没听说过。

I I haven't heard of that.

我觉得这基本上可以归入整体培训的范畴。

I mean, I think it would just fall generally under the overall training as a whole.

根据我的经验……

And my experience has been...

迈克，这个你可以补充一下。

And, Mike, you can jump in on this one.

如果你知道怎么跟AI助手对话，也知道怎么构建AI代理，那你已经掌握得差不多了，基本上就能轻松上手……

If you know how to talk to AI assistants and if you know how to build AI agents, you're that far along, you can go model the...

……轻松上手我们的AI系统。

Model our AI system AI system pretty easily.

其实提示词还是一样的。

Like, it's still prompting the same.

如果你被困在一个企业版的、被阉割的Copilot里，所有最有用的功能都被关掉了，那你无论如何都完蛋了。

Now if you're stuck in a, like, a neutered version of Copilot within an enterprise that's turned off all, like, the most valuable features, you're screwed either way.

这其实无所谓。

Like, that doesn't really matter.

这才是最大的问题。

And that's the biggest concern.

所以如果我是福特的一个AI员工，我要跳槽去别的地方，而且……

So if I was an AI Ford employee that I was making a move somewhere else and I...

假设我习惯了用Claude，但以后得用Copilot，我可能更关心的是，我能用它做什么？

Let's say I was really used to Claude and and I was gonna have to use Copilot, I would I would be more interested in like, what what am I gonna be allowed to do with it?

我的token有没有限制？

Do I have a limitation on my tokens?

因为一旦你习惯了在工作中使用这些工具，最大的问题就是，如果你去了一家限制你使用这些工具的公司，那就麻烦了。

Because once you're used to using these tools as part of your work, that's gonna be the biggest problem is if you now go to a company that limits your ability to use these tools.

我觉得更多的情况会是反过来的——那些想在公司里做创新的人，晚上和周末用自己的个人账户用Claude，他们会跳槽去那些欢迎他们、允许他们使用这些工具的公司。

I think it's gonna be a lot more of the other way where people who are trying to be the innovators within organizations who are using Claude on their personal account at nights and weekends, they're gonna jump ship and go to a company that's gonna welcome them in and allow them to use those tools.

我觉得就是这样。

I think that's exactly it.

大多数时候，关键不在于每个模型的细微差别，而在于你实际被允许做什么、不被允许做什么。

It's less about the nuances of each model most of the time and what you're actually allowed to do versus not allowed to do.

我觉得尤其是对一些顶尖人才来说，这会成为——如果不是最重要的因素的话——因为我们离那个点已经很近了……

I think for especially some top talent out there is going to become, if not one of the top thing because we're very close to...

如果你每天都在用这些工具，并且和它们深度融合，你一天都离不开它们。

If you're using these tools and embedded with them day in, day out, you can't live without them for one.

但第二点，这也是你对工作感到兴奋的原因，对。

But two, it's like the reason you're excited Yeah.

关于你的工作。

About your work.

所以放弃这些，不仅仅是薪酬考虑或职业发展的问题。

So giving that up is more than just something like a comp consideration or a career.

而且，我认为考虑到事情变化的速度，这也会限制你的职业发展。

Also, it's like a a limiting career move, I would argue, with how fast things move.

如果你不能做得足够多，你就是在自缚手脚。

You're hamstringing yourself if you can't do enough.

具体怎么做，就看你自己了。

Whatever that looks like for you.

如果我去面试，我会问：你能分享一下你们的AI政策吗？或者至少，大概的允许范围是什么？

I mean, if I was interviewing for a job, I would say, can can you share your AI policy with me or at least, like, the general frameworks of what's allowed?

你们有负责任的人工智能原则吗？

Do you have responsible AI principles?

这些原则在你们的文化里是怎么落地的？

How are those kinda tall within the culture?

我会有什么工具可以用？

What tools will I have access to?

有没有限制？

Are there limits?

你们有没有一个连接……

Do you have a connect...

我，我，我正在找市场营销方面的工作。

I'm I'm I'm looking for a job in marketing.

你们的人工智能和你们的客户关系管理系统有连接吗？

Is your AI connected to your CRM?

还是说，你们根本不做那些？

Or like, do you guys not do all that off?

这些就是我会作为一个人工智能从业者去问的问题。

Those are the things I would be asking as an AI for practitioner.

就像，我想知道，我能不能胜任你们要求我做的这份工作？

It's like, would wanna know, am I gonna be able to do the job you're asking me to do?

好的。

Alright.

我们还有时间，可能再问一个问题。

We have time for maybe one more question here.

现在我想回到之前的话题，因为我们在播客里确实聊过一点，但在MAICON上已经提到很多次了。

Now I wanna kinda come back to this because we did talk about it a little bit on the podcast, but it's come up so many times already at MAICON.

有人说，我教过别人怎么在ChatGPT里创建GPT。

Someone said, I trained people how to create GPTs in ChatGPT.

现在GPT没了或者快没了，我不知道该用什么了。

Now that GPTs are gone or going away, I don't know what to use now.

那你建议用什么替代呢？

Instead, what do you suggest as a replacement?

他们建议的是用插件来替代GPT。

So what they're suggesting are plug ins for replacement to GPTs.

根据我的经验，这行不通。

My experience is that does not work.

比如，我……

Like, I I don't...

那是一次性的体验，而且效果不好。

It was a a one off experience, and it was not good.

所以肯定没有他们预想的那么直观。

So it's certainly not as intuitive as I think they intended it to be.

我觉得这取决于你的目标是什么。

I think it depends on what your goal with it is.

所以如果你想公开分享，那你可能更倾向于用类似网站或者……

So if it is something you wanna publicly share, then you're probably moving more in the direction of something like sites or...

我也不太确定。

I don't know.

如果是内部工具，用作AI助手，那你可能会把它做成技能或项目。

If it's an internal tool that you're gonna use as an AI assistant, you might be building it as a skill or a project.

如果是重复性操作，你可能会把它变成智能体。

If it's a repetitive action, you might be turning it into an agent.

我觉得我们正在进入一个阶段，就是看实际用途，你可能需要审查你建过的所有GPT，然后问自己，它是用来干什么的？

Like, I think we're just moving into the realm where it just depends on what the actual usage is, and you're probably gonna need to audit any GPTs you've built and said, what is it used for?

谁在用？

Who uses it?

对。

Yep.

然后根据OpenAI目前提供的功能，把它转化成最合适的形式，不过这些功能三个月后可能就变了。

And then map it to what is the best thing to turn it into based on what OpenAI is currently offering that they might change in three months.

就像，我……

Like, I...

你也不知道。

You don't know.

这挺有意思的。

It's interesting.

这几乎就像你开始看到一个决策树一样，对。

It's almost like you start to see you start to see a decision tree kind of Yeah.

在这里显现出来。

Manifest here.

不幸的是。

Unfortunately.

这挺有意思的。

It's interesting.

我……

I...

我跟人聊天的时候，感觉如果你只是自己用GPT，而且用得还不深，就想要一些可重复的指令，那你就去学怎么建技能，跟AI聊聊怎么做。

When I talk to people, it's like, if you're just using a GPT for yourself and you're not that far along and you just want, like, repeatable instructions, go learn how to build skills and talk with AI about how to do it.

这以后会变得很重要。

That's gonna become important in the future.

但在这之前，直接把指令放到一个项目里，自己用就行了。

But in the meantime, just put the instructions into a project and use that yourself.

效果差不多。

It's gonna be roughly the same thing.

虽然不是完美的，但这是最接近的类比了。

It's not perfect, but it's the closest analog.

技能，我觉得，理想情况下是你开始走的那条路，而且它们不难建。

Skills is the, ideally, I think, the path you start moving down, and they're not hard to build.

难的是把它们做对，但那本来就需要很多次迭代。

They're hard to get right, but that's gonna require a lot of iteration anyway.

然后，对，如果你要进入……

And then, yeah, if you're getting into...

你需要别人公开使用这个，祝你好运，因为那是个大事，他们突然就撤了你的台。

You need people publicly to use this, good luck because that's a huge thing that they pulled the rug out from under you.

也许是个应用。

Maybe it's an app.

也许是你可以和团队分享的项目，但那样他们也能编辑指令。

Maybe it's projects you could share with your team, but then they can also edit the instructions.

所以，你不想那样。

So it's like, you don't want that.

所以谁知道会怎么发展呢？

So who knows where that head?

就像他们把一个本来很简单直观、谁都能学会的东西，我们都已经教了。

It's like they took a really simple intuitive thing that anybody could be taught, and we've taught it.

比如，我们办过工作坊，教那些刚开始接触AI的人怎么构建GPT。

Like, we ran workshops teaching people who are just starting an AI how to build GPTs.

就像，哇，这太棒了。

It's like, oh, this is brilliant.

我懂了。

Like, I get it.

我明白了。

I see it.

现在就是把它做大。

And now it's just like, just blow it up.

就是那个特别让人着迷的点。

Like, the one thing that was super into it.

好了，保罗。

Alright, Paul.

我们的时间到了，问题也问完了。

That comes to the end of our time, end of our questions.

感谢你愿意回答这些问题，我知道你在MAICON肯定有很多问题要回答，也有很多演讲要做，但到目前为止这个活动真的很棒。

Appreciate you being game to answer all these, especially I'm sure you're getting tons of questions, have tons of speaking to do at MAICON, but it's been an awesome event so far.

谢谢你来做这个。

Thanks for doing this.

感谢在场的每一位。

Thanks everyone who's been here.

我知道这里有很多播客听众。

I know we have a bunch of podcast listeners here.

我们其实……

We're actually...

等这期节目播出的时候，我们可能正在播客里提到的那个直播午餐会上。

When this is airing, we're we're probably gonna be in our live lunch that we talked about on the podcast.

所以，是的。

So Yeah.

有很多播客听众加入了我们。

You know, we got a bunch of podcast listeners that have joined us.

明天我们会和他们一起做一个专属的午餐问答环节，对你们来说就是今天。

We're gonna do a kind of exclusive lunch q and a with them tomorrow, which will be today for you all.

不过，是的，谢谢大家的收听。

But, yeah, thanks everybody for listening.

如果你在现场，也谢谢你的到来。

Thanks for being here if you're here.

如果你不在现场，我们很希望能在MAICON 2027见到你，我想我们很快会公布日期，如果还没公布的话。

And if you're not here, we'd love to see you at MAICON 2027, which I would imagine we'll be announcing the dates for if we haven't already.

我们公布日期了吗？

Do we announce the dates?

我觉得还没有。

I don't think so.

好的。

Okay.

我先不公布，因为我知道。

I won't announce them yet because I know.

我知道。

I I know.

不过我敢肯定是在明年十月。

I'm pretty sure it's next October, though.

好的。

Alright.

好的。

Alright.

谢谢大家。

Thanks, everyone.

谢谢保罗。

Thanks, Paul.

感谢收听AI问答。

Thanks for listening to AI Answers.

想继续学习的话，可以访问smarterx.ai，那里有按需课程、即将开课的班级，以及实用资源，能指导你的AI学习之旅。

To keep learning, visit smarterx.ai, where you'll find on demand courses, upcoming classes, and practical resources to guide your AI journey.

如果你有想问的问题，想用在未来的节目里，我们很乐意听到。

And And if you've got a question for a future episode, we'd love to hear it.

今天就到这里。

That's it for now.

继续探索，继续提出关于AI的好问题。

Continue exploring and keep asking great questions about AI.