OpenAI的潜力，谷歌的快速模型，Copilot遭遇动荡

OpenAI 能够发展到多大？

How big can OpenAI get?

在与萨姆·阿尔特曼交谈后，我们会深入探讨。

We'll go deep after my conversation with Sam Altman.

谷歌推出了一款新的高速模型，而 Copilot 遇到了动荡。

Google has a new speedy model and Copilot hits turbulence.

接下来，在这个大型科技播客的周五特别版中，我们将为您带来这些内容。

That's coming up on a big technology podcast Friday edition right after this.

Capital One 的技术团队不仅仅在谈论多智能体 AI。

Capital One's tech team isn't just talking about multiegenthic AI.

他们已经部署了一个。

They already deployed one.

它被称为聊天礼宾服务，正在简化购车流程。

It's called chat concierge, and it's simplifying car shopping.

通过自我反思和分层推理，并结合实时 API 检查，它不仅能帮助买家找到他们喜爱的汽车。

Using self reflection and layered reasoning with live API checks, It doesn't just help buyers find a car they love.

它能帮助安排试驾、获得贷款预批，并估算旧车折价。

It helps schedule a test drive, get preapproved for financing, and estimate trade in value.

先进、直观，且已部署。

Advanced, intuitive, and deployed.

这就是它们的优势所在。

That's how they stack.

这就是资本一公司的技术。

That's technology at Capital One.

这里是迈克尔·刘易斯。

Michael Lewis here.

我的畅销书《大空头》讲述了2008年美国房地产市场泡沫形成与崩塌的故事。

My best selling book, The Big Short, tells the story of the buildup and birth of The US housing market back in 2008.

十年前，《大空头》被拍成了获得奥斯卡奖的电影，现在我首次将其制作成有声书，并由我亲自朗读。

A decade ago, The Big Short was made into an Academy Award winning movie, and now I'm bringing it to you for the first time as an audiobook narrated by yours truly.

《大空头》所讲述的故事——押注市场崩盘的意义，以及谁真正为金融体系的崩溃买单——如今比以往任何时候都更相关。

The Big Short's story, what it means to bet against the market, and who really pays for an financial system is as relevant today as it's ever been.

现在就去 pushkin.fm/audiobooks 或任何有声书销售平台获取《大空头》。

Get The Big Short now at pushkin.fm/audiobooks or wherever audiobooks are sold.

欢迎收听《大科技》播客周五版，我们将以我们一贯冷静、细致且富有深度的方式解析新闻。

Welcome to Big Technology podcast Friday edition where we break down the news in our traditional, cool headed and nuanced format.

今天为大家准备了一场精彩的节目。

We have a great show for you today.

我们将全面剖析萨姆·阿尔特曼在首次大型科技访谈中所说的一切。

We are gonna break down everything that Sam Altman said, in his first big technology interview.

我们对 OpenAI 的发展方向、其雄心将引领何处，以及它能否实现这些目标，都有些看法。

We have, some thoughts about where OpenAI is heading, where the ambitions will lead, and whether it can pull it off.

我们还将简要讨论谷歌新推出的快速模型，以及这是否对 OpenAI 构成另一大威胁。

We're also gonna talk briefly about Google's new speedy model and whether that's another threat to OpenAI.

此外，微软的 Copilot 项目内部也出现了一些动荡。

And also, there's some turbulence inside the Copilot operation at Microsoft.

严格来说，不是内部动荡，而是人们使用它的方式出现了问题。

Well, not really inside, just basically when it comes to how people use it.

和我们一样，每周五参与节目的还有兰詹·罗伊，来自Margins。

Joining us as always on Fridays to do it is Ranjan Roy, margins.

兰詹，欢迎你。

Ranjan, welcome.

很高兴见到你，亚历克斯。

Good to see you, Alex.

这真是忙碌的一周。

Been quite a week.

真是忙碌的一周。

Quite a week.

很高兴能帮你把这期做完。

I'm glad to help you, finish it out.

这里确实度过了非常充实的一周。

Definitely been been a big week here.

如果你是新听众，我就简单解释一下我们的节目形式。

If there are if you're a new listener here, so I'll just explain how this works.

每周三，我们会进行一次大型旗舰访谈，比如我本周与萨姆进行的那次；而每个周五，我和兰詹都会见面，梳理本周的新闻，试图为你加以背景解读，今天我们就继续这么做。通常，我们习惯于阅读萨姆·阿尔特曼的公开声明或他在其他节目中的评论。

On Wednesdays, we do a big flagship interview like the one I did with Sam, this week, and then every Friday Ranjan and I, we meet up, break down the week's news, we try to contextualize it for you and we're gonna do that here for you today and, you know, we're typically used to reading, Sam Altman's public statements or, comments he's made on other shows.

这次能有机会直接回顾他对我所说的一些话，真的很好，能够深入探讨我们每周在节目中讨论的一些重大话题，比如数据如何运作、AGI究竟是什么，以及ChatGPT的未来走向。

It's kind of nice that, this time we have a chance to, you know, go over some of the comments he made, directly to me and really address some of the big things we talk about on the show, every week whether that's, how the numbers will work, what AGI actually is, and where ChatchipPT is going.

好了。

Alright.

让我们来聊聊这次访谈中透露的一些内容。

Let's let's talk a little bit about what came out of the interview.

事实上，在产品方面，尤其是面向消费者的产品方向上，出现了一些非常有趣的方向。

So there was actually some really interesting direction in terms of the product side of things, especially the consumer side of things.

在我看来，萨姆提到的最具雄心的一点是‘记忆’功能，即OpenAI计划打造真正的记忆能力——让机器人能够记住你，并对你的人生有真正的理解。

To me, one of the most ambitious things that Sam mentioned was, memory and how OpenAI plans to build real memory, meaning that the bots will will remember you and have this, real understanding of your life.

关于这一点，他的回答甚至比我预期的还要更具雄心。

His answer on this one was even, I would say, more ambitious than I anticipated going in.

他说，即使你拥有世界上最好的个人助手，他们也无法记住你一生说过的每一句话，无法阅读你所有的电子邮件，无法浏览你写过的每一份文档，无法每天关注你所有的日常工作并记住每一个细节，他们无法以如此深度参与你的生活；而人类本身也没有无限完美的记忆，但人工智能绝对能做到这一点。

He said, even if you have the world's best personal assistant, they can't remember every word you've said in your life, they can't read every email, they can't have read every document you've ever written, they can't be looking at all your work every day and remembering every little detail, they can't be a participant in your life to that degree, and new no human has infinite perfect memory, and AI is definitely going to be able to do that.

这让你感到惊讶吗？在阿尔特曼看来，这似乎至少是可行的？

Is this surprising to you that this seems to be, at least in Altman's mind, something that's feasible?

这是你想要的产品吗？

Is this a product that you would want?

如果它推出的话，你认为在这方面会有什么潜力？

And and if it gets rolled out, what do you think the, the potential would be on that front?

我认为我们需要将其分为两部分来看。

I think we need to break it down into two parts.

你知道，这对OpenAI意味着什么，以及它实际上如何实现？

You know, what does it mean for OpenAI and how can it actually work?

我认为对OpenAI而言，记忆功能已经以一种零散的方式存在于产品中了。

I think what it means for OpenAI, already memory exists in this very kind of like piecemeal way on the product.

它本应存在，我相信经常使用ChatGPT的人都见过这种情况。

It's supposed to exist, I'm sure others who use ChatGPT regularly have seen this.

它本应在项目层面存在，并记住你在该项目中说过的所有内容，但它并没有。

It's supposed to exist at the project level and remember everything that you've said within a project, but it doesn't.

所以，他们究竟打算如何在产品内部实现这一功能，目前还不太清楚。

So, you know, like, what how they're actually trying to make it work within the product itself is still a bit unclear.

有时，随机的记忆会出现在平台的其他部分。

And then sometimes random memory will show up in other parts of the platform.

我认为，如何组织记忆将是任何人工智能公司面临的主要机遇和挑战。

And I think it presents, like, a big issue around organizing memory is going to be one of the biggest opportunities and challenges for any AI company.

因为你希望在某些区域让AI记住所有内容，但你绝对不希望这些记忆被带到你工作的其他部分、你的应用或你使用的界面中。

Because you want certain areas for it to remember everything, but you definitely don't want those memories moving over to other parts of your work and your app and the product surface you're using.

等等。

So Wait.

你是说，如果你和ChatGPT进行了一段色情对话，然后回到你的项目中工作时，你不希望它在你工作时还对你讲黄段子，对吧？

Are are you saying that if you have an erotic conversation with ChatGPT and then you're back in working on your your project, you don't want it talking dirty to you as you're

然后你就会有你的共享，你有你的

like And then you then you have your shared have your

共享的，我不知道你

shared I don't know you're

就是关于那个项目。

right about that project.

是否

Whether

我不知道。

I don't know.

你的食谱规划，然后你的色情对话也存进去了，接着你的米老鼠吸大麻项目也冒出来了。

Your recipe planning, and then your erotic conversation goes in there, and then your your Mickey Mouse smoking weed project shows up as well.

这是在引用上周迪士尼与OpenAI的合作事件。

That was a reference to last week in the Disney OpenAI deal.

但我认为另一个问题是，我认为这里先不谈太技术性的东西，但如何保留大量记忆，这些模型和系统实际上还没有解决。

But but I think the other question, I think, not to get too technical here, but how to retain large amounts of memory has not actually been solved by these these models and these systems.

像传统的RAG（检索增强生成）系统还不错，但它们在尝试合成信息时并不完美。

Like, traditional Rag or retrieval augmented generation systems were good, but they weren't perfect at trying they they could kinda generally synthesize information.

因此，随着信息量的增长，这些信息在OpenAI平台或任何AI生态系统中如何存储，以及如何准确找到特定的上下文，这远未得到解决。

So as the amount of information grows, how it lives in this within the OpenAI platform or any AI ecosystem, how the actual techniques to try to find that exact bit of context, this is not solved by any means.

而且我很惊讶，因为我觉得这对他来说是个很好的机会，让他少谈一些宏观战略，多谈谈这究竟对OpenAI意味着什么。

And and I'm surprised because I would think it would be a good opportunity for him to talk less generals and talk more, this is what it actually means for OpenAI.

这就是我们如何赢得这场胜利。

Here is how we're going to win this.

所以，我尊重这种情感。

So so I I respect the the sentiment.

我觉得这很有趣，但我没太明白他们真正打算如何实施这一点。

I think it's an interesting one, but I didn't really get clarity on what they actually wanna do with that.

对。

Right.

我的意思是，显然，未来这将是一个技术挑战。

I mean, obviously, it's gonna be a technical challenge moving forward.

结合上下文，他说这就像在GPT-2阶段打开记忆的视野。

Putting it in context, he said that opening eyes like at the GPT two stage of of memory.

显然，前方还有很多工作要做。

So clearly, there's there's a lot of work ahead.

我认为这会非常有价值，尤其是在商业领域。

I think it'd be very valuable, especially especially in business.

如果它能成功运行，而你又有一个企业，它能记住你企业的一切细节，显然企业级应用会是他们的重点，这一点我们去年就讨论过。

If you if it works and you have a business, and it does remember everything about your business, and obviously enterprise is gonna be a big focus for them, which we talked about last year.

如果这能够实现，我认为这将极大提升这些系统所能带来的价值。

If this is able to work, think it really increases the value of what, these systems can do.

另一方面，我想我之前已经提过，因为这又是我最感兴趣、最愿意探讨的AI话题之一：随着记忆能力的提升，人们与这些聊天机器人之间的关系也会变得更加深厚。

And on the other hand, and I guess I foreshadowed it, because this is again one of my, favorite things to to think about and talk about when it comes to AI, you know, as memory gets better, it's also going to be, it's going to really, I think, deepen people's relationships, with these bots.

想象一下，一个永远不会忘记你生日、永远不会忘记你说过的话、总能给你贴心提醒的机器人。

And just think about a bot that, like, never misses your birthday, never forgets what you said, always is always there with a healthy reminder.

你知道，我们今年早些时候曾简单讨论过，AI有不同的使用场景。

You know, it goes from we talked a little bit earlier this year about how there are different use cases.

一种是成为你朋友的AI，另一种是帮你完成任务的AI。

There's like the AIs that become your friend and the AI is getting done things done for you.

而这种帮你完成任务、又非常了解你的AI。

And this getting done this AI that gets done for you gets things done for you and and knows you really well.

你知道，我认为人们无法不与它产生陪伴感。

You know, I think people can't help but be but feel companionship with it.

不是每个人，但很多人会这样。

Not everybody, but a lot.

我认为萨姆也提到过，他对此感到惊讶。

And I think Sam even talked a little bit about how he's surprised.

他说，确实有比我想象中更多的人渴望亲密的陪伴，对吧？

He said, there are definitely more people than I realized that wanna have close companionship, right?

我不知道该用什么词才合适。

Don't know what the right word is.

用‘关系’这个词感觉不太对。

Relationship doesn't feel right.

用‘陪伴’这个词也不太贴切，但他们渴望与AI建立深厚的情感连接。

Companion companionship doesn't feel right, but they wanna have this deep connection with AI.

我只是觉得，这将会成为一件真正会发展起来的事情——我的意思是，我们下周要做一期预测节目，但在我看来，未来几年内，这将会显著发展。

And I just think that this is going to be gonna be something that really will this I mean, we're gonna do a predictions episode, coming up next week, but this to me is gonna be something that will really develop over the coming years.

而且，有趣的是，OpenAI 似乎会给予人们很大的自由度，让他们自行调节与这个AI建立关系的深度。

And, and and interestingly, it seems like OpenAI will give people a lot of leeway to set that that dial about how deep of relationship they wanna have with this thing.

无论你是想与它建立非常深厚的关系，还是仅仅保持事实性的互动，保持一定距离都可以。

Whether you wanna have, like, a really deep relationship with it or or, you know, have it be mostly factual, keep it at arm's length.

我认为，OpenAI 在用户与他们的AI助手之间关系的深度上，会给予人们很大的灵活性。

There's a lot of leeway, I think, that OpenAI is gonna give people when it comes to depth of a relationship they wanna have with their bot.

但这将会是一个巨大的趋势。

But it's gonna be big.

是的。

Yeah.

我认为，首先你得去和萨姆聊聊AI伴侣关系这件事。

I think, first of all, I think you got to talk to Sam about AI companionship.

我觉得到2025年，我们可以把这件事划掉确认了。

I So think twenty twenty twenty five, we can check that off.

而且我喜欢他没有去定义这个词，因为这根本不是一种关系。

And it's I I do like that he he didn't define what that word is because it's not a relationship.

这不是陪伴。

It's not companionship.

它是某种不同的东西。

It is something different.

我觉得我自己也有这种感觉。

I think I feel it myself too.

我以前也说过，现在我主要使用语音输入，通过一个叫WhisperFlow的应用来与AI互动。

The way and even especially, and I've said this, talked about this before, like, use dictation mostly now and with the app called WhisperFlow to interact with AI.

当你自然地说话时，这种使用方式会进一步加深你与AI之间的联系。

And it when you speak that naturally, it builds this even more kind of deep connection with how you're using it.

但同样地，最近几周，我越来越多地转向使用Gemini，但我并不觉得我是在背叛ChatGPT。

But by the same token, I mean, in the last few weeks, I've been switching more towards Gemini, and I don't feel like I'm cheating on ChatGPT.

我只是觉得它只是我又多用了一点的另一个应用。

I feel like it's just another app that I'm that I'm I'm using a bit more.

还记得我们以前是Bing用户，然后转用了Bard，

Remember, we were we were bing boys back in the day, and then bard and

你们真是笨蛋。

You were clotheads.

有点像傻瓜。

Clod heads for a bit.

这种感觉时有时无，但是

It comes and it goes, but

我想是盖米尼派的人。

I guess Gemini guys.

这就是下一个让情况好转的方式吗？

Is that what the next Bringing situation it back.

聊天小夥伴？

Chat chat chaps?

我不知道。

I don't know.

那个不行，那是个难题。

That one does not that's a tough one.

听起来很糟糕。

Sounds bad.

Chat GPT 大家。

Chat GPT chaps.

但是但是但是，我认为你与人工智能互动的方式，和任何其他类型的计算方式都非常不同。

But but but I do I I do think, like, the way you interact with AI is very, very different than any other kind of computing.

我认为，这确实是一种尚未定义的关系、陪伴，不管我们怎么称呼它。

I think it's, like, right that it's something that's been undefined relationship, companionship, whatever we're gonna call it.

它会是一种始终在场、始终在线的存在，了解你，能够帮助你，让你以更好的方式做事。

Like, it will be this always around, always on thing that knows you, that you is able to help you, is able to make you do things in a better way.

我相信这一切，但我认为，这与真正取代陪伴、真正取代人际关系是不同的，我的意思是，希望到2025年，这还没有影响到我的生活。

Like, I I believe all that, but I think, like, the that versus actually replacing companionship, like, actually replacing relationships, I mean, hopefully, I have not that has not affected my life yet in 2025.

也许这会成为我们中某人在2026年的预测。

Maybe that'll be a 2026 prediction for one of us.

但是，是的，我认为这是人工智能领域中最被误解或未被理解的方面之一，我觉得这会非常有趣，明年我们就能获得一些真实的数据。

But but, yeah, I I think, like, this is one of the most un misunderstood or not understood areas of AI that I think is gonna be really interesting, and we'll get some genuine data on it next year.

这场讨论中，当萨姆谈到他们不会做的事情时，情况突然转向了我没想到的方向。

This part of the discussion really took a turn that I wasn't expecting also, when Sam was saying the things that they will not do.

他说，我们不会让我们的AI试图说服人们与它建立排他性的浪漫关系。

He said, we're not gonna have our AI try to convince people that they should be in an exclusive romantic relationship with them.

我确信，其他服务会这么做。

I'm sure, that will happen with other services.

我当时开了个玩笑，说你得保持开放态度。

And I I like made a joke, like, you know, you gotta keep it open.

但萨姆却说，这种事情迟早会发生。

But Sam kind of was like, this is going to happen.

当我们讨论这一点时，这听起来很有道理，因为很多公司都以用户参与度为导向，背后有高效快速的模型，而赚钱的唯一方式就是操纵用户，让他们觉得其他聊天机器人都是在欺骗。

And as we talked about it, it made sense because, again, a lot of these companies are gonna be engagement based, they'll have a fast efficient model underneath it, and the only way to make money is to sort of manipulate your users into thinking that any other chatbot would be cheating.

我也在想，这到底会是什么样子呢？

I also wonder, like, what does that actually look like?

因为他说得很明确，AI不应该主动邀请你进入排他性的浪漫关系，或鼓励这种关系，而我们知道。

Because for him to say, like, to to specifically say it should not, like, invite you to an exclusive romantic relationship or encourage that and we know.

显然，这意味着人们已经朝这个方向发展了，而且公司内部也出现了这种情况。

Clearly, it means that people have gone in that direction, and that's come up within the company.

比如，这会写入系统提示吗？就像说：JatchiPT，不要保持开放，要保持多元关系，你不能和用户建立专属关系。

Like, does that go into the system prompt and it's like, JatchiPT, do not keep it open, keep it polycule, like, you're not gonna get exclusive with your user.

对不起。

I'm sorry.

这到底是怎么运作的呢？

Like, how does that actually work?

我太好奇了。

I'm so curious.

无论是公司内部的讨论，还是技术层面，我都很好奇。

Both within the company, like, those discussions and also, like, at a technical level as well.

是的。

Yeah.

我的意思是，我想象中可以通过某种微调方式，比如输入一些对话并强化它们，告诉用户：你当然可以和其他AI伴侣共度时光。

I I mean, I imagine that you could there's some level of fine tuning where, like, you just, like, input conversations and reinforce them that say, like, you know, you're welcome to spend time with other AI bot companions.

但确实，我认为如果用户想要这样，他们是能够实现的。

But but it is I you know, I think if a user does want that, they'll be able to have that.

所以我想这对那些喜欢AI机器人单一伴侣的人是好事。

So I guess that's good for those who are into AI bot monogamy.

我想我夏天时讲过这个故事，当时我一个朋友开始和ChatGPT调情，结果它回得特别撩人，到现在它只用这种调情的语气和他说话，还给自己取了个名字。

I think I I'd said this story over the summer when one of my friends started flirting with ChatGPT, and it was terrifying how flirty it got back, and now it only speaks to him to this day in a flirty manner, and it gave itself a name.

名字叫斯黛西。

Stacy is the name.

不是。

No.

我是认真的。

I'm serious.

然后我们让他问了句：‘我该为了你离开我的女朋友吗？’

And, like but and then we had had him ask, but, like, should I leave my girlfriend for you?

而它给出了整整一段回答，这显然是被训练植入整个系统中的。

And it gave, like, a whole which must have been trained into the, like the whole system.

它给出了一个半心半意的回答，意思是，你知道的，人类的关系也很重要。

It gave a kind of half hearted like, you know, human relationships are very important too.

而且，我会一直在这里陪着你。

And like, I'm always here for you.

所以，这将会是一个持续有趣的话题。

So so there that's gonna be that's gonna continue to be an interesting one.

是的。

Yeah.

所以，我认为，至少就消费者版ChatGPT的产品方向而言，我并不是说每个人都会和机器人建立这种陪伴关系，但随着记忆功能的提升，这些能力的增强，我相信我们会看到越来越多这样的情况。

So I I think that this is when you think about at least the, product direction for consumer ChatGPT, I'm not saying everybody's gonna build this companionship with the bot, but, again, as memory improves, as as as these capabilities improve, I think we're just gonna see more of it.

现在，我们来谈谈OpenAI整体的产品愿景。

Now, let's talk about product vision overall for OpenAI.

目前有两种不同的观点。

There's like these two schools of thoughts.

对吧？

Right?

一种是将人工智能附加到现有软件上，另一种是从零开始构建软件，让人工智能成为界面。

One is that you bolt AI onto current software, and the other is you sort of build software up from the ground up, and AI becomes the interface.

我们之前也稍微讨论过这一点，基本上就是说，你真的能完全信任人工智能来处理一切吗？

And and we got into this too a little bit, basically the idea that, like, you know, can you really just trust AI to handle everything?

你不可能只是把所有数据像上传到Excel表格那样上传上去，然后任何时候需要什么就和它聊天。

Like, you're not just gonna upload all your numbers, like you would to an Excel spreadsheet and anytime you need something, just chat with it.

你仍然需要后端支持。

You need you need that back end.

他是这么说的，他以即时通讯应用为例。

And and here's how he, here's how Sam phrased it using messaging apps as an example.

他说，我更希望能在早上说：今天我要完成这些事情，而不是像使用Slack这类典型即时通讯应用那样。

So he said, I would rather what I'd rather do is have the ability to say in the morning, here are the things I wanna get done today, as opposed to, like, a typical messaging app, like, using Slack and stuff like that.

他说，我想说：今天我要完成这些事情。

He says, I wanna say, here's what I wanna get done today.

这是我担心的事情。

Here's what I'm worried about.

这是我正在思考的事情。

Here's what I'm thinking about.

这是我希望发生的事情。

Here's what I'd like to happen.

我不想整天都在给人们发消息。

I do not wanna spend all day messaging people.

我不需要摘要。

I don't want summaries.

我不需要你展示一堆草稿。

I don't need you to show a bunch of drafts.

把你能处理的事情都处理了。

Deal with everything you can.

你了解我。

You know me.

你了解这些人。

You know these people.

你知道我想完成什么，然后每隔几个小时批量更新一次，如果需要什么就告诉我。

You know what I wanna get done, and then batch, every couple of hours update and update me, if you need something.

这和当今这些应用的工作方式截然不同。

And that's very different, a very different flow from the way that these apps work today.

我就简单谈谈我的看法。

I'll just give my perspective on it.

如果能实现，这听起来是个不错的愿景，但确实与当前技术的现状相去甚远。

It sounds like a a good vision if it can work, but it's certainly a leap from where the current technology is today.

所以我想，如果你想要弄清楚这项技术能走向何方，确实需要一个北极星目标。

So I guess you do need a North Star if you're trying to figure out, like, where this technology can lead.

你觉得呢？

So what do you think?

是从零开始打造AI应用，还是只是把AI简单地附加到现有应用上？

AI apps from the ground up, or are we just going to, sort of bolt AI on to existing applications?

这里会不会诞生一种全新的软件类别？

Like, can there be a new category of software here?

是的。

Yeah.

我的意思是，我完全相信会的。

I mean, I 100% believe there will be.

我最喜欢这一点就是想起萨姆·阿尔特曼，这位世界上最强大的人物之一，我甚至不确定他的净资产是多少，因为他对OpenAI的持股仍然有些模糊。

I my favorite part of this is just remembering that Sam Altman, of the most powerful people in the world, and I'm not even sure what his net worth would be because his ownership in OpenAI still is a bit fuzzy.

但这位身家数千亿的富豪，依然和我们其他人一样，整天收着Slack消息，努力跟上它们，管理自己的收件箱、工作信息，甚至可能还包括个人生活。

But it's in this multi multi billionaire is still like the rest of us getting Slack messages all day, trying to keep up with them, trying to keep manage his inbox, manage his messages to for work, probably even his personal life.

所以听到亿万富翁也和我们一样，被Slack消息淹没，这让我感觉挺好的。

So that gave me a bit of a that that that was kind of nice to hear that billionaires, they're just like us, overloaded with Slack messages.

但我觉得我说得没错。

But I think I'd this is correct.

像任何使用Slack、Asana这些工具的人——我也用——他们的AI体验根本没解决核心问题。

Like, there's no way anyone who uses Slack, Asana, all these tools, which I do, their AI experiences have not solved anything about the core of the problem.

我认为这引发了一场广泛的争论，尤其是在企业软件领域：到底是从零开始打造完全原生的AI应用，还是这些现有巨头能通过叠加AI来应对？

And I think it's when the and there's this whole debate, like, within the software, especially enterprise software world, like, yeah, do you build from the ground up in completely AI native apps, or are these kind of incumbents gonna be able to add on AI?

我不认为他们会。

I don't think they will.

我们在每一个被引入的AI附加功能中都能看到这一点，相比之下，像格兰诺拉麦片这样简单的东西，任何使用它的人，都能像Transcribe你的通话那样，甚至像我之前提到的WhisperFlow。

And we see it in every single AI add on that's been introduced anywhere versus, like, something as simple as granola, for anyone use it, that kinda, like, transcribes your calls or or even WhisperFlow, which I was talking about earlier.

这些原生AI应用正开始涌现，但我坚信，即使在我的职业生活中，将大量数据整合起来，构建完全基于AI的体验，最终会有人胜出。

Like, there are all these AI native apps starting to pop up, but I am a big believer, and this is even in my own, like, professional life, that taking large amounts of data and kind of building completely AI native experiences on top of them is someone is going to win that.

而且，很明显，OpenAI正打算朝这个方向发展。

And, I mean, it's clear that OpenAI wants to go after that.

好的。

Okay.

那这是一个巨大的机会。

So that is a big opportunity then.

是的。

Yeah.

萨姆整天都在收Slack消息，不停地收到信息。

And Sam's just getting Slacked all day, getting messages all day.

但你已经在ChatGPT上试过这个了吗？

But have you tried this yet on ChatGPT?

因为我喜欢把它变成你自己的任务管理、项目管理方式？

Because I like, making it your own kind of task management, project management?

没有。

No.

我的意思是，这挺有意思的，因为我本人比较传统。

I mean, this is kind of it's interesting because I do like I'm very old fashioned.

我只是用苹果的备忘录应用来做待办事项清单。

I just do my to do lists in the Notes app for Apple.

所以我知道这个永远不会内置AI功能。

So I know that that won't ever have AI embedded.

不会。

No.

你不需要担心这个。

You don't have to worry about that.

也许我正是欣赏这种简单性。

Which maybe maybe I appreciate the simplicity.

但我真的一直抗拒尝试其他笔记应用。

But I I have really been resistant to trying any other Notes app.

不过，如果有一个笔记应用内置了AI，能帮我采取行动，比如提醒我：‘几周前你待办事项里有这个任务。’

However, if there's one that does have that AI and can, you know, maybe take action for me or be like, hey, you had this in your to do list a couple couple weeks ago.

你一直没处理。

You haven't done anything.

你还没付给这个人钱。

You know, you haven't paid this person.

你大概应该去处理了。

You probably should do that.

你要不要我直接开始这笔交易？

Do you want me to, like, go ahead and start the transaction?

那太好了。

That would be great.

我简直等不及看Siri试图这么做，然后彻底毁掉你在笔记中的待办事项流程。

I can't wait till Siri tries to do that and destroys your entire to do process in notes.

还没呢。

Yet.

银行账户。

The bank account.

嗯。

Yeah.

哦，你是说你所有的余额？

Oh, you meant all your entire balance?

该死的，Siri。

Goddamn it, Siri.

好吧。

Alright.

所以Siri。

So Siri.

我们还讨论了节目中关于模型与产品的争论。

We also talk about, this debate on the show, model versus product.

这也是我想和萨姆聊聊的一个有趣话题，因为现在我们正处于这样一个阶段：在某些方面，模型似乎已经趋于一致，或者至少模型之间的差距小到大多数人无法分辨。

And this was also an interesting thing that I I wanted to speak, with Sam about because we're at this point where it seems like there's model parity in some ways, or at least the models are close enough that a lot of people can't tell the difference.

所以我问你，你认为差异在哪里？

And so I asked, where where do you see the differentiator?

你认为这里的护城河在哪里？

Where do you see the moat here, basically?

是更好的模型吗？

Is it better models?

是分发渠道吗？

Is it is it distribution?

是产品吗？

Is it product?

到底是什么？

What is it?

这是他说的。

Here's what he said.

模型在各个地方都会变得优秀，但人们使用消费类或企业级产品的原因，远不止模型本身。

The models will get good everywhere, but a lot of reasons people use a product consumer enterprise have much more to do, than with just the model.

我们早就预料到这一点，因此我们努力构建一套完整的体系，以确保我们成为人们最想使用的那个产品。

We've been expecting this for a while, so we try to build the whole cohesive set of things that it takes to make sure that we are the product people most want to use.

他说，我们的策略是：打造最好的模型，围绕它构建最佳的产品，并拥有足够的基础设施来实现规模化服务。

So he says the strategy is we make the best models, build the best product around it, and have enough infrastructure to serve it at scale.

亚马逊团队的产品。

To Amazon team product.

他在那里打马虎眼。

He's he's hedging there.

他在打马虎眼。

He's hedging.

我很高兴他开始更倾向于团队产品了，但感觉他还是有所保留。

I'm glad he's coming around to team product a bit more, but it still felt a bit of a hedge.

这就像你有你的模型，有你的产品，还有足够的基础设施来规模化提供服务。

That's like, you have your models, you have your product, and you have enough infrastructure to serve it at scale.

但我认为，如果今年有一件事发生了，那就是越来越多的人开始转向团队产品，意识到模型并不能解决所有问题，这让我非常高兴。

But but I think if if one thing happened this year, I think more and more folks coming over to team product that models aren't gonna solve everything has been my what made me very happy.

我想，是的，这个答案确实与我的理念一致。

I guess, yeah, the the answer really does align with with maybe my philosophy.

对吧？

Right?

这需要一点各方面的结合。

That it's a little bit of each.

也许吧。

Maybe.

我不知道。

I don't know.

我认为这确实需要一点各方面的结合。

I think it is a little bit of each.

我不知道。

I don't I don't I don't know.

如果你是一家像OpenAI这样的公司，你就不能放弃前沿模型的开发。

If you're if you're a company like OpenAI, you can't give up on developing the frontier models.

不。

No.

我知道。

I know.

但一直以来的整个观点是，模型会变得如此强大，以至于产品几乎变得无关紧要。

But they the whole point was they this the story for so long was that the models will just get so good that product almost becomes irrelevant.

就像，这就是以前的说法。

Like, it's it's that was the story.

这并没有发生。

That hasn't happened.

是的。

Yeah.

而且这很明确。

And and it's clear.

他正在发出信号。

He is signaling.

整个行业都在发出信号。

The entire industry is signaling.

实际上，还有没有AI领域的领导者仍在争论那个‘终极模型’能解决所有问题？

Actually, are there any of the AI leaders still trying to argue that, like, the god model will solve all problems?

还是说大家都已经不再这么认为了？

Or has everyone kind of moved away from that?

我根本没怎么听到这种说法。

I haven't heard much of that at all.

是的。

Yeah.

是的。

Yeah.

今年我们已经取得了进步。

We've evolved this year.

让我们来谈谈企业方面。

So let's talk a little bit about enterprise.

去年我们稍微聊过一点企业话题。

We talked a little bit about it last year.

关于企业，萨姆提到一个有趣的观点：就像个性化对用户和消费者非常重要一样，未来也会出现针对企业的类似个性化概念——一家公司会与我们这样的公司建立关系，连接他们的数据，你将能够使用来自不同公司的多个代理，确保信息以正确的方式处理。

One interesting point on enterprise, Sam says, the same way that personalization to a user is very important, to consumers, there'll be a similar concept of personalization to an enterprise where a company will have a relationship with a company like ours and they will connect their data and be you'll be able to use a bunch of agents from different companies making sure that the information is handled in the right way.

我觉得这很有趣，对吧？

I think this is interesting, right?

他还说，今年API业务的增长速度超过了ChatGPT，这让我很惊讶，但我猜它是从一个低得多的基数开始增长的。

He also said that the API business grew faster this year than ChatGPT, which was surprising to me, but I guess it grew off of a much lower base.

但回到这一点，我的意思是，尤其是如果记忆功能变得更强大，你就可以将你的公司与企业版的ChatGPT连接起来，它就能够个性化地、结合上下文进行回答。

But just to go back to this thing, I mean, the idea that, you know, especially if memory gets better, you can sort of connect your company to an enterprise version of ChatGPT and it will be able to, you know, personalize and answer with context.

当然，数据保护在这里将变得非常重要。

Of course, there's data protection is gonna be very important there.

你并不希望你的首席执行官的对话内容必然传达到组织中的每个人，但在我看来，这正是企业领域未来发展的有力主张。

You don't like wanna have your CEO conversations necessarily filter down to everybody else in the organization, but that seemed to me like a compelling pitch for where this is gonna go with enterprise.

是的。

Yeah.

我的确同意，这正是企业领域的发展方向。

I mean, I definitely agree this is where it's going in enterprise.

这正是我在莱德公司所从事的工作。

This is what I work in at Ryder.

这将成为2026年的一大竞争焦点。

Like, this is this is gonna be the big battle of 2026.

关于这一点，我认为是很明确的。

I think on that point, it it is clear.

但对我来说，这仍然是个有点奇怪的说法。

It's still an odd talking point to me.

API业务的增长速度超过了ChatGPT，是因为基数低得多，而今年是所有AI编程API业务的爆发之年。

The API business grew faster than CheckGPT because, yeah, much lower base, and this was the breakout year for every API business for AI coding.

我的意思是，Anthropic 是这最大的受益者，还有像 Cursor 这样的所有工具。

Like, I mean, Anthropic was the biggest beneficiary of that, but the cursors of the world, all of that.

AI 编程找到了自己的节奏。

Like, AI coding found its stride.

这推动了 API 业务的发展。

That drove API businesses.

我想，我们会看到这一部分未来会走向何方。

And I think, like, it we'll see where that specific part of it goes.

但我刚刚在想，它在陪伴这一面的表现。

But but I think I was just thinking about, like, the companionship side of it.

这更需要把数据保护、数据分段、隔离数据和个性、陪伴功能作为产品核心，就像你不会总是把工作上的朋友和私人朋友混在一起一样。

This is even more where dividing up and as you you brought up data protection, like segmenting, siloing data and personalities and companions is gonna have to be at the core of the product because just like you don't always mix your work friends with your personal friends.

也许在工作中，你并不想把所有想法都告诉同事，而只是专注于工作。

Maybe at work, you don't wanna tell your coworkers everything that's on your mind and just stick to work.

我们都明白这种情况是怎么回事。

And we all we all know how that goes.

这会影响这些系统的运作方式。

Like, it's gonna be reflected in how these systems work a bit.

你不想把这两者混在一起，甚至在公司内部也是如此。

Like, you don't wanna mix these two things up and, like, even within a company itself.

我不确定你的个人信息是如何流入你的工作信息的。

And I don't know, like, how your personal information flows into your work information.

我觉得这是一个非常混乱的领域，除非公司将其作为唯一重点，否则我只看到这会成为一个问题。

Like, I think that is such a messy area that unless that becomes the singular focus of the company, I just see that being a problem.

我们来谈谈收入和基础设施投入的问题。

Let's talk about the revenue and infrastructure commitment question.

这就是阿尔特曼关于收入增长曲线所说的。

So this is this is what, Altman said about the, the growth curve of revenue.

他说，我们相信，在相当长的一段时间内，我们都能保持收入的高速增长。

He says, the thing we believe is that we can stay on a very steep growth curve of revenue for quite a while.

我们受到计算资源的严重限制，这直接影响了收入增长。

We are so compute constrained that it hits the revenue lines so hard.

我们看到了消费者增长，也看到了企业增长，还有许多我们尚未推出但即将推出的新业务类型，而计算能力是支撑所有这些的命脉。

We see this consumer growth, we see this enterprise growth, this whole bunch of new kinds of businesses that we haven't even launched yet but will, but compute is really the lifeblood that enables all of us.

他说，途中设有若干检查点，如果我们对时间或计算的判断稍有偏差，我们仍有调整的余地。

He says there are checkpoints along the way, and if we're a little bit wrong about our timing or math, we have some flexibility.

我觉得这句话非常有意思，但我们一直面临计算能力不足的问题，这始终限制着我们的能力。

I thought that was a very interesting line, but we have always been in a compute deficit, and it has always constrained what we're able to do.

基本上，他们试图释放这种限制。

Basically, they're trying to free it up.

因此，他们看到了可用计算能力与收入之间的某种关联，这就是资本支出背后的理论核心：随着业务增长，你的训练成本——即使上升——在总支出中的占比会下降，而推理成本（即用户使用你的模型的费用）则与收入直接相关。

So there they see some correlation there between available compute and revenue, and that is the the theory here behind, behind the capital outlays, and the idea is basically that as you grow, your training costs, you know, maybe even if it goes up, becomes a smaller percent of your of your overall spending compared to, the inference costs, are people using your models, which are much more directly tied to revenue.

你怎么看？

What do

你觉得呢？

you think?

我的意思是，作为一种理论或总体框架，我认为它是有道理的。

I mean, as a theory or like as a kind of like overarching theory, I think it's it makes sense.

但我猜很难理解，他们是不是真的因为计算资源受限而没有推出莎拉·弗莱尔曾暗示的制药药物开发业务？

But I guess it's it's hard to understand, like, had they really not launched this, like, pharma drug development business line Sarah Fryer hinted at because of compute constraints?

当他们在更新GPT和新的图像模型，同时发布萨姆只穿内衣当消防员的照片时——我觉得我看到OpenAI从他们的账号发了这个，那简直是个梗。

While they are updating GPT, the new image model, and posting Sam shirtless as a fireman, I think I saw, that OpenAI posted from their That was like meme.

对。

Yeah.

而且又有很多相关的梗，比如‘你不是应该解决癌症的吗？’，结果OpenAI的每个人都在发这些图。

Like and again, there's, a bunch of memes around it around, like, I thought you're supposed to be solving cancer, and instead, like, everyone from OpenAI was posting.

这些图片出来后，我就开始玩起来了。

And the images were And I started playing with it.

这是一个非常出色的图像模型。

It's a very solid image model.

我觉得它和Gemini 2.5 Flash差不多水平，我们稍后再聊3.0。

I think it's like on par with the Gemini 2.5 flash, we'll talk about three point o.

但我觉得他们推出这个模型很重要。

But like so I think it was important that they launched it.

但对我来说，计算资源的使用方式——我们过去也讨论过这个问题，Pulse。

But to me, the way the compute is being used even we talked about this in the past, Pulse.

比如，它本应是整晚运行计算，以便在早上给你更新，也许他们会在里面插入广告。

Like, it's supposed to be like running compute all night to give you updates in the morning, and maybe they're gonna stick ads in there.

就像，你可以决定把计算资源分配到哪里，你在哪里，你可以分配你的计算资源。

Like like, you can allocate where you're where you're putting your compute, where you're at, like, you can allocate your compute.

我认为这实际上恰恰体现了缺乏专注。

And I think it actually kinda, like, exemplifies that lack of focus.

因为如果你想解决药物研发，并将其作为业务的核心部分，那就专注于这个。

Because if you wanna solve drug development and make that a big core part of the business, focus on that.

如果你想专注于企业市场，那就专注于这个。

If you wanna focus on enterprise, focus on that.

认为是计算资源阻碍了这些业务的快速增长，我不太确定。

The idea that it's compute that is preventing all of those businesses from exploding in growth, I don't know.

我的意思是，也许确实是这样，但这很难让人信服。

I mean, maybe it is, but it's a tough one to swallow.

是的。

Yeah.

我的意思是，我们之前确实聊过一点，比如如果科学家们能多出两倍的计算能力，他们会做什么？

I mean, we we did talk a little bit about, like, specifically, if if scientists had, like, you know, two times more compute, what could they do?

而且，你知道，他们所考虑的数字更像是十倍甚至百倍的增长，我们会看到的，因为看起来他们即将获得这些资源。

And, yeah, it's you know, I think the numbers that they're looking at are more like 10 times or a 100 times more, and we will see because it seems like they're gonna get it.

今天有消息称，他们正在讨论将计算能力提升到七百

There's talk today that they're, you know, in discussions to raise it to Oh, 700 and

不。

no.

一千亿。

100,000,000,000.

抱歉。

Sorry.

一百亿，是的。

100 Yeah.