OpenAI联合创始人：在你感觉准备好之前，就开始用AI进行构建 | Greg Brockman

我们会看到创业精神的大规模复兴。

We're gonna see this massive renaissance of entrepreneurship.

我觉得，这就要来了。

Like, I think it is coming.

我觉得未来一到两年内，它会真正爆发。

I think it's really gonna kick in over the next year to two years.

这位是格雷格，OpenAI的联合创始人，OpenAI就是开发ChatGPT的公司，每周有超过十亿用户在使用它。

This is Greg, cofounder of OpenAI, the company behind ChatGPT, which more than a billion users use every week.

我挺擅长写软件的。

I'm pretty good at writing software.

我现在不自己写软件了。

I don't write software anymore.

我指挥AI来写软件。

I direct an AI for it to write software.

我会提供大量的反馈和指导。

I provide lots of feedback and guidance.

当人们在开发软件时，你如何制定策略，确保所构建的东西不会被新模型淘汰？

When people are building software, how do you strategize around building something that's not gonna be made obsolete with the new models?

我正要给你看我最近做的一个东西。

I'm just about to show you something that I've built recently.

我对AI的野心够大吗？

Am I ambitious enough with AI?

嗯，我觉得对AI的野心永远都不够大，因为我们总能做得更多。

Well, I think the answer to being ambitious enough will always be no because there's always more we can do.

对于愿意提升自己智能体技能、并且只有一个小时的人来说，这个周末他们该从哪里开始？

For someone who is willing to step up their agent game and they have one hour, where do they start this weekend?

我会说首先

I'd say first of all

本视频由HubSpot赞助。

This video is sponsored by HubSpot.

我想聊聊所有实际的东西。

I want to talk all things practical.

我最近和一位诺贝尔经济学奖得主聊天，有一个话题反复出现，就是我们的AI模型已经很强大了，但缺乏实际应用。

I was recently talking to Nobel Prize laureate in economics, and one theme kept coming up that we have such capable AI models, but we lack applications.

你觉得你今年推出的产品会如何改变这个局面，帮助人们认识到AI的潜力？

What do you think that you're launching this year will help change that and help people realize what's the power of AI?

嗯，首先很有意思的是，看到有多少人在用ChatGPT。

Well, it's been very interesting to see, first of all, how many people use ChatGPT.

我们每周有12亿活跃用户。

We have 1,200,000,000 weekly active users.

每周有大约3亿人用Chat来查询健康问题，帮助他们的日常生活。

We have like 300,000,000 people who use Chat every single week for health queries, right, to help them in their personal lives.

但同时，这些使用场景中有多少只是浅尝辄止呢？

At the same time, how many of those use cases just scratch the surface?

从数据中我们可以看到，如果人们用Chat做三种不同的用途，也就是RealLife Chat能为你做的三件不同的事，你就会成为重度用户，你会留下来，继续使用Chat，并且想用得更多。

And we see this in the metrics that if people use chat for three different use cases, so three different things that RealLife Chat could do for you, you become a power user, you become, you retain, you keep using Chat and you wanna use it more.

但要达到这三种用途，确实是个很大的挑战。

But getting to those three use cases, it's such a hard challenge.

从某种程度上说，这也有道理，因为，好吧，这个AI什么都能做，但它具体能做什么呢？

And in some ways, it kinda makes sense because, okay, so here's this AI that can do anything, but what's one thing it can do, right?

就像有时候

It's like sometimes

有时候，那个神奇的东西是什么？

sometimes What's the magic thing?

对，没错。

Exactly, right?

所以，真正找到那个能打动你的神奇之处。

And so really finding that magic that speaks to you.

但同时，我觉得这也显示了巨大的机会，一个几乎空白的领域，你在和AI对话，而AI知道自己能做什么。

But also at the same time, I think it shows this massive opportunity, this massive space that's almost a gap where you're talking to an AI and the AI knows what it's capable of.

它可以向你建议，嘿，你让我做了这件事，其实我还能做更多，对吧？

And it could suggest to you, hey, you asked me to do this thing for you, actually, I could do even more, right?

这里有个主意，我可以把整个演示文稿变成电子表格。

Here's an idea, I can make all presentation a spreadsheet.

所以，这其实是我们一直在探索的事情。

So, that's actually been something we've been exploring.

但我觉得这里有个关键点，就是你对AI的期望其实是简单，对吧？

But I feel like there's something here about the fact that what you want out of AI is simplicity, right?

你不想要一个满是按钮、滑块和模型选择器的界面，那些都不需要。

You don't want an interface that has buttons and sliders and model pickers, like none of that.

那不是我们被承诺的AI。

That's not the AI we were promised.

你想要的是能对话、能解决问题、能帮你达成目标的东西，而且它们甚至能主动出击。

You want something you can talk to that can solve problems, can help you, can help solve goals, but they can actually even be proactive.

它们可以对你说，嘿，你收件箱里有封邮件，你过去几天一直没回那个人，其实我做了些研究，发现问题的正确答案是这个，我已经起草好了回复。

They can say to you, hey, you got this email that's in your inbox that you've been ignoring this person for the past couple of days and that actually I did some research and it turns out that here's the right answer to the question and I've drafted an answer.

你同意吗？

Do you approve?

顺便说一句，我刚才提到的那个场景，不是理论上的。

And by the way, the use case I just mentioned, it's not theoretical.

用Dots的时候，我确实就遇到了这个情况，就在我们开始录这个播客之前。

With dots, I actually got that exactly, this use case happening to me right before we started this podcast.

所以我觉得现在AI已经能够变得更加主动，能帮你省去很多琢磨系统能干什么的麻烦，真正帮到你，让你能实现你想做的任何事情。

And so I think we're at a point now where AI is actually able to be much more proactive and to remove so much of the burden of figuring out what is this system capable of, and to really help you and then help you be able to achieve whatever it is that you want to do.

还有没有其他一年前想都不敢想，但现在可以实现，而且大家应该去试试的用例？

Are there any other use cases that were unimaginable a year ago that are possible now and that people should be trying out?

我觉得太多了，对吧？

Well, I think so many, right?

从很多方面来说，我建议大家去探索的方式就是，你得去玩一玩这个系统，或者看看它能做什么。

There's a whole spectrum of them from, and I think the way that I always recommend going to discovery is you gotta play with the system, or you gotta see what it's capable of.

说到编程，这些模型现在写软件的水平，基本上比我认识的任何人都强，对吧？

And so when it comes to coding, like these models, they're now better at writing software than basically anyone I know, right?

我写软件也挺厉害的。

I'm pretty good at writing software.

不过我现在已经不写软件了。

Like, don't write software anymore.

我直接指挥AI来写软件。

I direct an AI for it to write software.

我会提供很多反馈和指导。

I provide lots of feedback and guidance.

所以做个小游戏总是很有趣，能快速看到效果，比如看看花了多长时间完成。

So it's always fun to like make a little game and just kind of get this quick hit on just seeing like, okay, how long did it take to deliver on that?

然后再要求更复杂的东西。

Then ask for something more complex.

我觉得现在Astra在Blender和三维建模方面非常出色，所以让它根据照片生成三维模型这类事情也挺好玩的。

I think that we're at a point now where Astra is so good at Blender, so good at three d modeling, and so it's kind of fun to ask it to produce three d models of a photo and those kinds of things.

你可以看到它用软件，比如用微软画图来勾勒照片，但你也可以要求更实际的东西。

You can see it using software, like using Microsoft Paint to sketch out a photo, but you can ask for more real things too.

比如我知道有个朋友想建客房，就用Astra生成了CAD图纸，然后拿给承包商，现在已经在建了。

Like I know, for example, a friend who was designing or wanted to have a guesthouse built and used Astra to actually produce CAD plans and then brought those to a contractor, and now it's being built.

所以这真的是一件很有赋能感的事情。

And so it's a very empowering thing.

通常你得请别人来做。

Normally, you'd have to hire someone else.

比如，我不会做CAD图纸。

Like, I don't know how to make CAD plans.

但现在我们口袋里就有这种智能，能帮我们做这些事情。

But now we have this intelligence in our pocket that can actually do these kinds of things for us.

对我来说，我特别喜欢尝试把AI应用到生活的不同方面。

To me, one of the things I really like to do is to try to push the AI in different dimensions of my life.

所以我觉得语音模式让AI变得非常容易使用，对吧？

And so I think that voice mode is something that makes AI so accessible, right?

你只要进入语音模式，开始跟它说话就行了。

You just go into voice mode and you start talking to it.

现在语音模式已经连接到了各种工具。

Voice mode now is hooked up to tools.

所以你可以让它做任何聊天能做的事情。

And so you can actually ask it to do anything that your chat can do.

如果你把AI连接到你的邮箱和日历，现在你就可以问它，我今天有什么安排？

And if you're hooked up to your email, to your calendar, now you can ask it about, what's my schedule for today?

或者，比如，我想腾出一些时间来做深度思考，我该怎么做？

Or like, what should I, you know, I wanna free up some time for some deep thought.

我该怎么安排？

How can I do that?

它实际上可以了解你日历中每个部分的情况。

And it can actually go and have context on what's going on in each part of your calendar.

我觉得，对我来说，有一种活动就是退一步思考，想想有什么比AI目前为我做的更雄心勃勃的事情可以请它去做。

And I think that in my mind that there's this activity of just stepping back and thinking about what's a slightly more ambitious thing I can ask my AI to do than it's done for me to date.

让我们看看它能不能行。

Let's see if it works.

我正要给你看一个我最近做的东西。

And I'm just about to show you something that I've built recently.

我觉得我们正处在一个每个人都能打造个人生产力应用的时代，而且不止一个。

I think we're in this era where everyone can build a personal productivity app, multiple.

我做的这个应用，其实就是让它帮我准备播客。

And this one that I've built, so I basically asked it to create an app that's gonna prepare me for the podcast.

我告诉它我要采访谁、采访多长时间，还有我的成功指标是什么，基本上就是YouTube上的观看量。

I tell it who I'm interviewing for how long and what's my success metric, and it's basically the views on YouTube.

然后它就会自动生成一个播客。

And then it goes ahead and creates a podcast.

它会做研究，建议要讨论的话题。

It does the research, suggests topic to discuss.

这很棒，但我怎么让它从我的经验中学习呢？

This is great, but how do I make it learn from my experience, right?

有没有什么地方我可以再进一步？

And is there anywhere I can push this further?

因为我要录这段对话，它会到这里来，然后它会检查，现在它正在检查我之前的对话。

Because I'm going to record this conversation, it's going to come here, and it's going to check now, it's checking my previous conversations.

但你觉得这里缺了什么，我对AI的期望够高吗？

But what do you think is missing here, and am I ambitious enough with AI?

嗯，我觉得关于是否足够有野心，答案永远是否定的，因为我们总能做得更多。

Well, think the answer to being ambitious enough will always be no, because there's always more we can do.

而这正是AI的美妙之处，它真正关乎解决问题，帮助你实现你想要达成的目标。

And that's the beauty of AI, is that it's really about solving problems, helping you achieve what you want to achieve.

你能想象的东西是没有上限的，对吧？

There's no ceiling to what you might imagine, right?

所以关键在于调整，当你能够实现一件事，也许你可以梦想得更大一些。

And so it's really just about sort of tuning that you were able to achieve a thing means maybe you can dream a little bit bigger.

所以在我看来，是的，这真是一个超酷的应用。

And so in my mind, yeah, think is a super cool app.

首先，就做背景研究而言，也许在构建这个应用的时候，AI已经去查看了我所有做过的采访。

So first of all, in terms of just doing the background research, and maybe when this was being built, the AI already went and looked at all my interviews that I've ever done.

而且……

And...

确实。

Did.

啊，好了。

Ah, there we go.

它把所有背景资料都查了，是的，它

It did all the background Yeah, it

它确实查了背景资料。

It did the background.

唯一的问题是它还是连不上社交媒体，但我觉得那是个插件的问题。

The only thing is it still can't connect to social media, but I think it's a plug in thing.

对。

Yep.

但然后，是的，它浏览了你的对话记录。

But then, yes, it went through your conversations.

好了。

There we go.

然后它

Then It

什么都有。

has everything.

所有OpenAI过去的新闻稿。

All of OpenAI's past press releases.

哦，这里有个问题标签。

Oh, there we got a questions tab.

对，我们有个问题标签。

I like Yeah, we got a questions tab.

没错。

That's right.

标题和缩略图已经在这里了，这些是我可以复制的问题。

And that's the title and thumbnail already here, and these are the questions that I can copy.

是的。

Yes.

我很好奇，在后期处理和最终视频制作完成时，它能帮上多少忙，对吧？

I'd be curious how much it helps with the post processing and the actual production of the final video once it's done, right?

因为这东西掌握了我所有的背景信息，关于OpenAI的，还有你过去节目里讲过的那些内容，那你想怎么剪辑呢？

Because this thing has all this context on me, on OpenAI, on the kinds of things that you've covered in your past episodes, and then how do you want to cut it?

我听说Astra在视频后期制作方面有很多好的应用场景，而且它能主动去做这些事情。

Like, I've heard there's so many good use cases of Astra helping with post production for videos and proactively being able to do that.

所以我觉得，如果能有一个应用，你唯一需要做的就是看看这些问题，确保你喜欢它们，那我会非常好奇。

So I think actually I'd be very curious if you could have an app where it's just like the only thing you have to do is basically look over the questions, make sure you like them.

如果你不喜欢，就提供一些反馈，它能做得更好。

And if you don't, provide some feedback, and it can do better.

你出现，真正进行采访，然后其他所有事情都处理好了。

You show up, you actually do the interview, and then everything else is taken care of.

现在从采访部分到视频真正在YouTube上发布，你需要投入多少工作？

How much work do you put in right now from the interview part to something actually being live on YouTube?

哦，工作量大得很。

Oh, a lot of work.

我有一个大团队，但我们也记录了很多，而且我做这行已经十二年了。

I have a big team, but it's also, we've documented a lot, and I've been doing this for twelve years.

所以，我们手头有所有参考过的文件，现在用AI来构建就更容易了，但即便如此，每一集至少还是要花四十个小时。

So, we have all the files that we referenced to, so it's easier to build with AI now, but it's still, every episode is at least forty hours.

哇。

Wow.

如果你能把这段时间省下来，对吧，所有的流程都自动处理好了，那你就能花更多时间真正思考什么才是合适的采访，以及所有其他相关的事情，或者如果你能扩大规模，你就能做更多的采访，是啊，那你会用这些多出来的时间做什么呢？

And if you could get back that time, right, so that kind of all of the mechanics are taken care of, so you can spend so much time really thinking about what the right interview is and all of the other things that go into this, or really thinking about if you could scale, you could do way more interviews, yeah, what would you do with that additional time?

这个问题应该问那些在日常工作中应用AI的小企业主。

That's the question to you with small businesses that are applying AI for their day to day jobs.

企业主的角色是什么？

What is the role of the business owner?

他们应该更多思考什么？

What should they be thinking more of?

现在业务的真正瓶颈在哪里？

And where's actually the bottleneck for the business now?

我觉得有一个根本性的问题。

I think that there's something fundamental.

首先，我们看到很多小企业之所以能创办起来，是因为人们觉得有了Chatuchi BT就安心了，对吧？

So first of we see so many small businesses that are being started because people feel secure with the fact that they have Chatuchi BT, right?

以前要创业，需要很多专业知识，很难起步，对吧？

That there's so many pieces of expertise that if you want to get a business off the ground, it's hard, right?

但现在你口袋里就有一个博士级别的智能，一个世界级的专家，一个商业顾问。

But you actually have a PhD level intelligence, a world expert, business consultant.

所有这些现在都免费放在你口袋里。

You have all these things in your pocket now for free.

这真的很了不起，对吧？

It's pretty remarkable, right?

如果你愿意，也有能力付费，你还能获得更深入的见解和更多的计算能力来解决任何问题。

If you're willing, if you're able to pay, you can get even so much more depth and so much more compute towards any problem you want.

所以，小企业比以往任何时候都更容易创办，我们确实看到了这一点。

So small businesses, easier to start than ever, and we're seeing it very concretely.

而且，当你和这些小企业主交谈时，你会发现他们外包给AI的任务非常多，当然有些是他们知道必须做的，但还有很多任务是以前根本没人会去做的。

And then people are able to, when you talk to many these small business owners, there's just so many tasks that they outsource to an AI that, certainly there's the ones that they knew that they had to do, but there's so many tasks that no one would have done it before.

所以他们的商业策略或者思考如何制作这份文件时，会有一些小漏洞。

And so little holes in their business strategy or thinking about how do you actually produce this document.

另外，很多小企业主甚至没有意识到自己还能做些什么。

The other thing too is small business owners who don't even realize what they could be doing.

比如我们有个员工跟我说，他在一家古董表店买表，然后开始和店主聊起AI。

Like one of our employees was telling me that he was shopping for watches in this like antique watch store, and that he started talking to the owner about AI.

店主就说，哦，AI，我什么都用不上。

And the owner was like, oh, AI, I can't use that for anything.

你知道，我开的是表店。

Like, you know, I'm a watch store.

然后他说，不不，AI对每个人都有帮助。

And he was like, no, no, AI can help everyone.

让我给你演示一下。

Let me show you.

于是他就聊了一些店主遇到的问题，店主说，嗯，他确实没有网站。

And so he talked about some of the problems he has, and he's like, Well, yeah, he doesn't really have a website.

所以就直接把它弄出来，那是最简单的部分。就像跟AI聊聊天，说‘给我建个网站，建个超棒的网站’就行了。

And so just pull That's up the easiest part Exactly, it to like And just like chatty bitty voice, being like, Build me a website, build an awesome website.

然后店主说，嘿，我每块表都有自己的故事，对吧？

And then the owner said, Hey, I also have, each watch has its own story, right?

所以这就变得很简单了，比如跟AI说每块表背后的故事，最后你就能得到一个漂亮的展示，然后你会意识到，技术就是这样变得活生生的，而平时它对很多人来说几乎是遥不可及的。

And so it's just like, then it becomes very easy, like, okay, talk to the AI about, here's the story behind each watch, and you end up with this beautiful visualization of everything, and you just realize that there's this way that technology becomes alive, that normally it's almost inaccessible to so many people.

所以我觉得，这真的在推动你思考：如果你雇了全世界最厉害的程序员，你会让他做什么？

And so I think that this really pushing on if you had like the world's best programmer on your payroll, what would you put them to work on?

而现在你就能做到这一点。

And that now you can do it.

我记得你在某个地方提到过，小企业真正需要开始培养的是野心，对吧？

I think you mentioned somewhere that it's actually ambition that small businesses have to start working on, right?

对，对。

Yes, yes.

我觉得这真的很有意思。

I think that is I think it's fascinating.

我同意。

I agree.

我觉得这真的是关于我们都能提高抱负的上限，对吧？

And I think it is really about we can all raise a ceiling of ambition, right?

我们都能完成更多事情。

We can all accomplish more.

那些真正投入其中的人，我觉得我们已经开始看到非常了不起的成果了。

People who really lean into that, I think we're already starting to see really remarkable things.

顺便说一句，我再给你讲个例子，我们有两个盲人兄弟，他们用我们的技术，既用在日常生活里，也用在工作中。

I'll tell you another example, by the way, is that we have, there are these two brothers who are blind who use our technology, and that they use it to, both in their daily life and for their work.

甚至像把线缆插到电脑正确端口这样的任务，对吧？

And even for tasks like being able to plug the cable onto the right port on a computer, right?

你可以用ChatGPT来获取这些信息，而以前他们得找人帮忙。

You can use ChatGPT to actually basically provide that information when normally they would have to ask someone for help.

所以我觉得这种赋能，能帮助人们更加独立，能按照自己的意愿生活，这一切都在发生。

And so I think this kind of empowerment, being able to help people be able to be more independent, be able to live their life the way that they want, All of that, it's happening.

现在正在大规模地发生。

It's happening right now at massive scale.

当有人想用软件做点什么的时候，OpenAI 在工作和生产力方面投入这么多，你觉得什么样的应用才值得付费？

When someone is thinking of building something in software, when OpenAI is building so much around like work and productivity, what do you think makes an app worth paying for?

以前做应用的一些环节确实正在变得大众化。

There's definitely parts of what app building used to be that's becoming commoditized.

每个人都能用上了。

It's becoming accessible to everyone.

这正在大规模地普及。

It's becoming massively democratized.

我觉得这是一件非常重要而且美好的事情，对吧？

And I think that's a really important thing and beautiful thing, right?

就是可以有更多的软件出现。

That there can just be more software.

所以，接下来，什么才是差异化，这个问题就变得更加重要了。

And so then the question of, yeah, what is differentiation becomes even more important.

我觉得，以后那些能以独特方式创造价值的软件，会有比以往更多的机会。

I think that there is going to be more room than ever for software that adds value in unique ways.

其中一部分，真的就是深入理解问题本身。

And some of this is really about deeply understanding the problem.

还有一部分，是深入了解你的用户，对吧？

Some of this is deeply understanding your users, right?

比如你想做教育相关的东西，那就有老师、家长、学生，对吧？

You think about something like if you want to build something for education, there's teachers, there's parents, there's students, right?

涉及的各方利益相关者太多了，你真的得好好考虑。

That there's so many stakeholders that you have to really think about.

还有学校管理层。

There's the administrators.

你怎么打造一个产品，既能提供准确的信息，又有合适的边界，用起来恰到好处，而且你打算卖给谁？

How do you build a product that actually gives the right information, has the right guardrails, that is useful in the right ways, and who are you selling to?

这些都是问题。

All these questions.

所以我觉得这些人为因素，比如生意的核心是什么，你发现了什么，你的洞察是什么，你看世界的独特方式是什么。

And so I think that these like human factors of what is the core of the business, what is it that you have discovered, what is it that's your insight, what is your unique way of looking at the world?

而且我觉得，你如何实现这些想法，怎么去...

And I think that the mechanics of how you instantiate that, how does...

你知道，你非得用C++或者Swift之类的语言写吗？

You know, do you have to write in C plus plus or Swift or something?

谁在乎这个呢？

Like, who cares about that?

这从来就不应该是障碍。

Like, that was never what the barrier should have been.

我看到YC的初创公司，最近的批次有60%是B2B的。

I've seen that YC startups, the recent batches are 60% B2B.

我的直觉告诉我，有了这些工具，如果我想成为企业家，我在消费领域看到了很多问题。

And my intuition is telling me with these tools available to us, if I want to become an entrepreneur, I've seen so many problems in consumer.

你觉得我们会不会进入一个越来越多的人做消费类生意的时代？

Do you think we're gonna enter this era when more and more people are building consumer businesses?

嗯，我觉得企业和消费者之间的界限会变得模糊。

Well, I think that the line between what enterprise and consumer are is going to blur.

我觉得公司的定义在某种程度上会变得不那么清晰，或者说不再像以前那样，公司就是一大群人以某种方式组织起来。

I think that what a company is is going to become just less well defined in some ways, or it's going to be less like, okay, company is like, it's like big masses of humans that are organized in certain ways.

我觉得人们会被赋予巨大的能力，对吧？

I think people are going to be so empowered, right?

我们会看到创业的大复兴。

We're going to see this massive renaissance of entrepreneurship.

我觉得这就要来了。

Like, I think it is coming.

我们已经看到了一些苗头，但真正爆发可能还要一两年。

I think we're seeing the leading edges of it, but it's really, I think, really going to kick in over the next year to two years.

至于你是卖给消费者还是企业，有些东西在区分上还是会保持不变的。

And I think that the question of are you selling to consumers or enterprises, there are some things that will remain true in terms of the split.

比如，我觉得企业销售的本质就是关系。

Like I think that when I think about what enterprise sales is, it's fundamentally about relationships.

但如果你把它看作关系，那么你之所以不能向消费者做企业销售，是因为消费者太多了，对吧？

But if you think about it as relationships, it's like, well, the reason that you can't really do enterprise sales to consumers is because there's just too many consumers, right?

规模太大，成本太高。

It's just too high of a scale, it's too expensive.

但如果AI真的能放大你建立关系的能力，建立真正有意义的关系，大规模地真正关心每个人，那么也许企业销售的运作方式甚至对更多面向消费者的业务也变得有意义。

But if you have AI that really amplifies your ability to have relationships, really meaningful ones, really take care of people at massive scale, then actually maybe the enterprise sales motion becomes something that makes sense even for more consumer businesses.

但企业的运作机制，我认为会演变。

But the mechanics of what a business is, I think, will evolve.

我认为你与用户互动的方式会演变。

I think that the how you relate to your users will evolve.

但根本的东西——你希望提供价值，与接受价值的人建立关系，并且是双向的交流——所有这些都会变得更加核心，无论哪个领域。

But the fundamentals of that you want to be delivering value and you want to build a relationship with the people who are receiving that value and have it be a bidirectional exchange, all of that is going to become much more front and center regardless of which domain.

在我们继续之前，有一件事。

One thing before we keep going.

如果你想随时了解这些创始人在此所说的内容，我们会把每次对话中的用例和提示整理到我的通讯《Future Proof》中。

If you want to stay updated on all the things these founders say here, we take the use cases and prompts from every conversation and put them into my newsletter called Future Proof.

当然，我也会把我自己的那些精彩用例放进去。

And of course, we'll also put my own beautiful use cases there.

比如说，我们会整理出反复出现的七项技能，就是未来几年大家会需要的那几种。

For example, we'll put together the seven skills that came up again and again, the ones people will need in the next few years.

这些技能我和我的团队每天都在用。

These are the skills that my team and I are also using daily.

通过描述里的链接订阅吧。

Subscribe through the link in the description.

你会收到那一期，还能看看我们之前的通讯，看看你错过了什么。

You'll get that one, and you can also read our previous newsletters and see what you missed.

当人们在开发软件时，你们是怎么制定策略，来确保做出来的东西不会被新模型淘汰呢？

When people are building software, how do you strategize around building something that's not going to be made obsolete with the new models?

我觉得有些活动会随着模型变得更聪明而扩展得很好，而有些活动则是在绕开模型的局限性。

There are some activities I think scale well as the models get smarter and some activities that are kind of working around limitations in the model.

第二种很难持久。

The second one is pretty hard to be durable.

所以我觉得有很多创业公司确实专注于这一点。

And so I think there are various points been startups that really focus on it.

我们拿模型，加一些提示词，还有一些手工定制的东西。

We take the model, we add some prompts, we have some very handcrafted whatever.

这并不是说在某个时间点这样做没有价值。

And it's not to say that there's a point in time where that is valuable.

有时候甚至跟界面有关，对吧？

And sometimes it's even about the interface, right?

在每一级模型能力下，几乎都会出现一个新的界面。

That at each level of model capability, there's almost a new interface that suggests itself.

2022年的时候，我们有一个足够个性化的AI，值得去读它的回答，就是GPT-4。

In 2022, we had an AI that was personalized enough that it was worthwhile reading the responses, the GBD4.

所以当然，你想做一个ChatGPT 3.5和GPT-4。

So it's like, of course, you want to make a chat GBD3.5 and GBD4.

你知道，我觉得到2025年底，编码代理时代才真正开始。

You know, end of twenty twenty five, I'd say, is when the coding agent era really started.

那时候的感觉就是，AI从占你软件开发工作的20%，一下子变成了占80%。

And that was like you had AI that kind of shifted from being 20% of your software production to 80% of your software production.

所以，你需要一个完全不同的产品来抓住这个趋势，这才是大家都会去用的东西。

And so, you just want a very different product that captures that, that thing, this is the thing that everyone's going to utilize.

现在我觉得我们开始进入个性化代理的时代了。

Now I think we're starting to enter this era of the personalized agent.

所以你会有一个基于云的AI，而且它正在被应用到各种不同的设备形态中。

So you're going to have a cloud based AI, and this is being ratified, number of different form factors.

OpenAI，我们刚刚发布了Dots。

OpenAI, we just released Dots.

我对它超级兴奋。

I'm super excited about it.

我觉得它有最聪明的模型来支撑这种技术，而且它真的非常有用。

I think it's got the smartest model backing up this kind of technology and it's so useful.

对我来说特别有用。

It's useful to me.

我觉得这对很多人都会很有用。

Think it's be useful for lots of people.

所以我认为核心问题是，随着模型变得更聪明，你要考虑在哪里扩展。

And so I think that the core thing is thinking about where will you scale as the models get smarter.

而且，这其中有部分是关于理解行业的。

And again, some of this is about understanding the industry.

还有一些是关于建立关系的，比如你想想医院，那里有那么多不同的利益相关者，对吧？

Some of this is about building relationships with back to, if you think about hospitals, right, there's so many different stakeholders there, right?

有病人、医生、医院管理层，还有保险公司。

That there's the patient, the doctor, the hospital admin, there's the insurance provider.

有那么多不同的参与方。

There's so many different parties.

那么，你如何在每一方之间建立信任，并提供一个对他们有用的产品呢？

And so how do you build trust across each of those and have a product that's useful to them?

然后，即使你已经在那个行业里，如果你有一个更好的模型，你已经有这些客户，你的产品实际上会因此变得更好、更有防御性。

And then actually be able to, even if you're, yeah, when you think about it, if you have a model that's just even better, if you are already in that industry, you have these customers that actually your product gets better and more defensible as a result.

你有没有特别期待看到初创公司用模型来做什么？

Is there anything you are particularly looking forward to seeing startups use models to build something?

我觉得领域专业知识也非常关键，因为有一件事很容易被忽略，那就是像OpenAI这样的公司，我们拥有的技术非常广泛，对吧？

I think the domain expertise is also really key because one thing that's easy to miss is that as something like OpenAI, we have this technology that's so broad, right?

它几乎触及经济增长和发展的方方面面，还有人们的个人生活、工作生活。

It kind of touches every single aspect of economic growth and development, people in their personal lives, work lives.

每个人都能找到适合自己的东西，真正能帮到他们的东西，帮他们实现目标的东西。

There's something for everyone, something that can really help them, something to help them achieve their goals.

但通常在任何特定的垂直领域，我们离最优还差得很远，对吧？如果你把经济看作一个分形结构，就像你放大看，你会说，好的，这里是医疗保健领域。

But often in any specific vertical, we're so far from optimal, right, that if you almost think of the economy as this fractal thing, right, where it's like you zoom in, you're like, okay, here's a sector of like healthcare.

然后你再放大，仔细看看有多少不同的公司，有多少不同的做事方式。

And then you zoom in, you like really look at the number of different companies and the number of different ways of doing things.

不管你选哪个领域，都有太多容易摘的果实了。

And no matter what sector you pick, there's just so much low hanging fruit.

这其实是我在打造OpenAI过程中遇到的最有趣、最令人惊讶的事情之一，就是意识到经济中存在着如此多的低效之处，而你可以通过更高效的流程，把时间还给人们，真正创造价值。

That's actually been one of the most interesting and surprising things for me in building OpenAI is to realize that there's just so much suboptimality in the economy in terms of how you can actually have much more efficient process that gives people their time back, that actually delivers the value.

跨领域连接并不是我们经常做的事情，对吧？

Being able to connect across different fields is not something we do very much, right?

人们都是专精于某个领域的。

It's like people specialize.

现在，把它们发展成公司是有经济意义的，因为五年前，它们可能只是公司的一个功能。

And now it makes economic sense to build them out because five years ago, they would be like a feature of a company.

现在，你可以围绕这个功能建立一个小公司。

Now you can build a small company around that feature.

没错。

Exactly.

而且我认为，即使是这些传统的行业和领域，也主要是因为我们的工具限制。

And I think that even these like classic sectors and domains are really because of the limitation of our tools.

再说，想想看，你怎么才能在任何领域成为专家呢？

Again, you think about how do you become good, like an expert in any field.

因为你去上学，专攻某个方向，而随着人文学科学习更多的东西，专精的范围就变得越来越窄。

Because you like, you go to school, you like specialize, and as humanities learn more things, what it means to specialize becomes narrower and narrower.

比如现在你是个生物学家，拿到博士学位，可能就专攻某一个具体的反应，对吧？

Like if you're a biologist now, you get a PhD, there's like one specific reaction that you got your PhD on, right?

而且这个反应跟整个系统是脱节的。

And it's like it's not connected to the whole system.

我觉得在创业的时候，能不能搞出更多这种跨界融合的东西，这是个问题。

And I think that when it comes to building a business, the question of can you have these much more hybrid things?

比如教育科技就是个例子，把懂教育的人和懂技术的人凑在一起，他们就能创新，做出新东西来。

So you think about things like ed tech as an example of like you take people who understand about education, they understand technology, they mash them up, and that they're able to innovate and do new things.

我在想，会不会还有其他我们从来没想过的领域，其实特别适合融合在一起。

I wonder if we're going to find that there's going to be other domains that we've never thought about that actually mesh together super well.

完全同意。

Totally.

这就是创业的未来，把不同领域匹配起来，找到合适的解决方案。

That's the future of entrepreneurship, matching those domains and finding the right solution.

你能给我讲讲你手头的一个比较成熟的方案，让我好有个参照来对比吗？

Can you talk to me about a setup that you have that is sophisticated enough for me to compare this?

因为有时候人们会在评论区骂我，问我像你这样的人关于晨间简报代理的问题，所以我们就不谈晨间简报代理了。

Because sometimes people bash me in the comments for asking people like you about morning briefing agents, so let's not do morning briefing agent.

我们来谈谈能帮你管理公司、提高效率的东西。

Let's do something that helps you run a company and be efficient.

这是人们看到后会认为对他们的生产力有影响的时刻。

It's something that people see, they will think it's an moment for their productivity.

我也不想贬低晨间简报代理，

I would not poo poo the morning briefing agent too

太

much

因为我觉得晨间简报代理有很多深度。

because I think that morning briefing agent has so much depth to it.

有一种情况是，好吧，是的，它给了我一些信息，但说实话，我自己刷一下通知也能看到。

There's the like, okay, yeah, it's like, it kind of gave me some info where it's like, honestly, I could have just scrolled the notification myself.

还有一种版本是，它会告诉你，这些是待处理的事项，对吧？

There's the version of it where it says, like, here are the things that are pending, right?

比如，这里有一些你还没回复的Slack消息，你真的应该按照这个优先级顺序去回复这些人。

Like, here are these Slacks that you have not answered, and you really should get back to these people with this prioritization order.

再进一步，就是我已经做了研究。

There's the next level of, I have done the research.

比如有个问题，关于我们在某个特定市场的策略应该是什么。

There's this question about what our strategy should be in some particular market.

而且我真的去看了所有的演示文稿，这里是你做决定时需要关注的东西。

And I've actually gone and I even looked at all of the presentations, and here's the things you need to be paying attention to to make this decision.

第三种情况，我觉得就是我们目前所处的阶段，对吧？

That third thing is like where I think we're at, right?

就是那种AI，能真正在多个领域做非常复杂的研究，然后给我一个非常精简的版本，告诉我应该关注什么。

It's like the AI that really can do very sophisticated research across a bunch of different things and provide a very compressed version of what I should be paying attention to.

我真正期待的下一个步骤，虽然我们还没完全做到，就是所谓的晨间简报代理，它会说，嘿，我实际上联系了这些人，比如你的同事，还有这个外部合作伙伴，我把所有事情都整理得井井有条。

The thing I'm actually very excited about as a next step that we're not quite there yet is the morning briefing agent, so to speak, that is like, hey, I actually called, like, you know, I called up these people, like I talked to, you know, these coworkers of yours, I talked to this external partner, and like, I got everything into perfect order.

比如，我只需要你批准这个采购订单，然后我们就可以开始新的项目了。

Like, I just need you to approve this purchase order, and like, we're ready to go on the new whatever.

真正主动、能替你办事的AI，这个时代就要来了。

AI that really is proactive and able to do things on your behalf, like that's coming.

而且我觉得，Dots这边我们其实已经有这个技术了。

And I think with Dots, we actually have the technology for it.

所以我觉得，我们很快会看到大家真正开始用起来。

So I think we're gonna start seeing people really start to utilize that.

我觉得我们还有改进的空间，诸如此类的。

I think we have improvements to make and things like that.

但在我看来，这些用例每一个都有很深的挖掘空间。

But I'd say, in my mind, any of these use cases, there's so much depth to them.

而对我来说，这正是AI的关键所在。

And that is, to me, the key thing about AI.

如果你只看AI写软件这个标签，好吧，确实，我们已经有能写软件的AI五年了。

If you just look at the label of an AI that writes software, okay, sure, we've had an AI that writes software for five years now.

但今天AI写的软件，跟五年前AI写的软件，完全是两码事。

But AI the right software today, totally different from AI the right software five years ago.

所以我对应用的深度非常欣赏，即使标签上写的一样，大家都听过一百遍了。

And so I have a lot of appreciation for the depth of application that's possible, even if the label on the tin is the same thing and everyone's heard it 100 times.

当你看到OpenAI的这些用例时，你们有没有内部指标来衡量模型进步如何真正转化为人们的生产力？

When you look at these use cases at OpenAI, do you have an internal metric how advances in models are actually translating into people's productivity?

我们也发布相关的博客文章。

We publish blog posts on this as well.

我们做了不少工作，研究token使用情况，以及它如何与代码行数、提交、PR等关联。

So we've done a decent amount of work looking at token usage and looking at how it correlates with like lines of code or commits and PRs and things like that.

再说一次，这些都是代理指标。

And again, all these are proxy metrics.

它们并不完美。

They're not perfect.

但我认为，另一个判断AI是否真正改变生产力的方法是，如果内部出现故障，比如员工无法访问模型，他们会多么沮丧。

But I think that another way that you can tell is AI really changing productivity and how much is if you ever have an internal outage, like if internally people can't get access to a model, how much they feel upset.

每个人都会说，我没法工作了。

Everyone's like, can't do my work.

然后你会想，好吧，六个月前你还没有这种东西，现在你说没有它两个小时就干不了活？

And you're like, okay, six months ago you had nothing like this, and now you're saying you can't do your work when you don't have it for two hours?

这怎么可能呢？

Like, how is that possible?

但我们经常看到这种情况。

But we see this all the time.

这真的太酷了。

It's so, so cool.

我自己也有这种感觉。

And I feel it for myself.

比如，我经常用的晨间简报代理，就是每当我有关于OpenAI的问题，我就直接问我的代理。

Like, the kinds of things on the morning briefing agent access that I do a lot of is I, like, anytime I have a question about OpenAI, I just ask my agent.

我不知道为什么一开始还要做别的，因为它太高效了。

Like, I don't know why I would do anything else to start because it's just so efficient.

有时候我甚至觉得自己有点懒，比如有人给我发了一条消息，里面有个文档，我就想模糊地记一下它在哪里。

It's like often, and sometimes I even feel like I'm like being a little bit lazy of like, there's like some message that someone has sent me with a doc in it, I'm just like trying to vaguely remember where was it.

所以，它跟所有东西都连着呢。

So, it's connected to everything.

什么都连得上。

Connects to everything.

连你的邮件、Slack，还有你用的其他工具。

Connects your mail, Slack, whatever you're using.

没错。

Absolutely.

而且它掌握所有信息，所以你需要什么，直接问它就行。

And it has all the information, so whenever you need something, just ask it.

对。

That's right.

而且它有

And it has the

是的，而且通过电脑操作，它真正打通了最后一步，能连接到我所有的企业信息。

Yes, and with computer use, it really closes the last mile of being able to connect to all my enterprise context.

所以这就让它变得特别有用，因为我真的有一个这样的幕僚长，而且我实际上还有一个非常出色的人类幕僚长，对吧，我能够给她更多的支持，因为我不需要再问那些基本问题了，现在我的AI就能回答。

And so then that makes it so useful because I really have this chief of staff, and I actually have a human chief of staff who's also incredible, right, in that I kind of am able to give her so much more leverage by not having to ask these basic questions that now my AI can answer.

而且其中一些所谓的基本问题，其实也挺复杂的。

And some of these, call it basic questions, are also quite sophisticated.

比如，问一些关于不同业务指标的问题，对吧？

Like, for example, asking questions about different business metrics, right?

所以它实际上需要去查询我们的数据仓库，然后得出结果，比如你在看某个市场的增长情况，或者留存曲线之类的。

So it actually has to then go and issue queries into our data warehouse and come up with like, you're looking at like the growth in some market or the retention curves or something like that.

我记得在上一份工作，在Stripe的时候，我们有一个数据团队，你得给他们发邮件，然后一周后他们才会回复查询结果，现在我就直接问AI，它马上就能搞定。

And I remember in my last job when I was at Stripe that we had a data team that you would send like an email to, and then a week later they would come back with an answer to that query, and now I just ask the AI and it just does it.

太不可思议了。

It's incredible.

你有没有见过人们没有充分利用token的情况？

Do you ever see areas where people are under utilizing tokens?

你想告诉大家，比如，多用点token在这上面吗？

Do you want to tell people like, use more tokens for this?

哦，当然。

Oh, absolutely.

而且我觉得这有两个方面。

And look, I think there's two dimensions to it.

一个是曾经有个“疯狂堆token”的时代，就是人们盲目地说多扔点token进去，这显然不是主要目标。

So one is that there was this era of token maxing, right, where it's just like people were just blind to saying throw more tokens at it, which is clearly not a primary objective.

但它确实是个不错的参考指标。

But it is a good proxy metric.

如果你不刻意去优化token数量，你可能会发现生产力的提升和token数量的增加是相关的。

If you're not trying to optimize tokens, you're probably going to then see improved productivity correlated with increased number of tokens.

我们确实看到了这一点。

And we really do see it.

我们有排行榜，会关注这个，我觉得那些使用最熟练的人确实能做出非常了不起的事情。

We have leaderboards, we kind of look at that, and I think that we see the people who tend to be the most sophisticated in their usage are just able to do really incredible things.

当然，这都取决于具体领域等等，但我很喜欢看到，比如我们的财务团队，他们真的全面转向了AI优先，我觉得有很多事情，比如以前结账要花很长时间的工作流程，现在快多了，对吧？

Again, it all depends on the domain and things like that, but I love seeing, for example, in our finance team, like that they've really gone full AI first, and I think that there's just like so many things where, like workflows that used to take a long time for closing the books that are now just so much faster, right?

而且这些都不是个例，大家都感受到了效率的提升，以及它带来的帮助有多大。

And none of those are like, everyone just feels that increased productivity and how helpful it is.

我们通讯团队里有一些早期使用者。

We have some early adopters from our comms team.

我觉得他们现在能轻松应对各种情况，比如你有一个活动，很多记者要来，你要安排座位表，还要搞清楚每个人的饮食偏好，而一个人就能搞定所有这些，因为有了 Codex 在后台帮你处理大量工作。

I think that they're just like being able to manage, like you have an event with a bunch of journalists coming and you're putting together a seating chart and trying to figure out everyone's dietary preferences, and being able to manage all of that as one person because you have codex that's actually taking care of so much of the back end for you.

这些从今年年初就开始了。

That's all been happening since early this year.

所以我们在知识工作的效率提升方面，到处都能看到这样的用例。

So we see these kinds of use cases for knowledge work productivity everywhere.

然后，当然，我觉得还有一个很容易被忽视、而且我觉得很快就会对所有人产生影响的部分，就是科学发现方面，对吧？

And then, course, the thing that I think is also very easy to miss, and I think is actually coming soon for everyone, is the scientific discovery side of things, right?

我们投入了大量算力去解决纳维-斯托克斯方程这个千禧年难题，动用了大约一万个智能体，跑了好几天。

That we put a ton of compute into solving Navier Stokes, this big millennium problem, and that was like 10,000 agents for some number of days.

这确实需要很大的算力。

Like, it's a lot of compute.