GPT-5是否名副其实？AGI等待仍在继续，自我厌恶的Gemini

GPT五终于来了。它是否名副其实？敬请关注本周五大科技播客特别版，节目马上开始。欢迎收听大科技播客周五特别版，我们将以一贯冷静细致的风格解读新闻。大家都知道今天要聊什么——OpenAI终于发布了GPT五。

GPT five is here. Finally, does it live up to the hype? That's coming up on a special big technology podcast Friday edition right after this. Welcome to big technology podcast Friday edition where we break down the news in our traditional cool headed and nuanced format. You know what we're gonna be talking about today because GPT five has finally been released by OpenAI.

几小时前我们刚邀请OpenAI首席运营官布拉德·莱特卡普上过节目，那期内容会是最新更新的播客，你们将听到OpenAI官方说法，以及关于模型训练过程和AI领域发展的深度见解。但像每周五一样，今天我和兰詹·罗伊将客观分析这个新模型的真实水平，看它是否配得上人们的狂热期待。绝不夸大其词，我们会保持理性视角。现在有请兰詹再次登场。

Of course, we had OpenAI COO Brad Lightcap on the show just a few hours ago, so that episode will be the most recent one in your podcast feed where you're gonna get the official line from OpenAI and a a bunch of really interesting insights about what it took to train this model and where the AI field is going. But today, as we always do on Friday, Ranjan Roy and I will break down exactly what the situation is with this new model and whether this is actually something that lives up to the hype that people have been talking about. No overreactions. We're gonna do it with the proper context. And with that, wanna welcome Ranjan back to the show.

兰詹，欢迎你。

Ranjan, welcome.

通用人工智能纪念日快乐，亚历克斯。它来了吗？

Happy AGI day, Alex. Is it here?

不，还没来。

No. It's not here.

真的没来？

Is it here?

我赌输了。山姆·奥特曼说GPT五比人类几乎所有的能力都强，所以我想好吧...

I lost the bet. I lost the bet. Look. Sam Altman said that GPT five is smarter than almost everything every single thing a human does. So I thought, okay.

终于要见证通用人工智能了，结果并没有。这个话题我们稍后讨论。首先有个有趣的公告。

Fine. You know, we're finally gonna see AGI, but turns out, no. No AGI. We will talk about that in the middle. But first, we have a little interesting announcement.

来听听看。

So let's hear it.

除了撰写《边际》通讯，我其实还在Writer.com这家企业级生成式AI初创公司工作，负责零售行业垂直领域。今天特意提及是因为GPT五对我的意义，很大程度上源于我这段时间的工作经历——某种程度上它可能有点接近通用人工智能了。

Yeah. In addition to writing the margins newsletter, I've actually been working at a company called Writer, writer.com. It's an enterprise generative AI startup, and I'm leading the vertical for the retail industry. I wanted to bring that up today because GPT five and what it means for me, I think is heavily informed by a lot of the work I've been doing, and I think it might be a little AGI ish.

我是说，这太神奇了，你去一家AI公司，开口第一句话就是：没错，我们这儿有AGI。但不对，不，没关系。

I mean, it's amazing that you go to an AI company and the first thing you first sentence out of your mouth is, yep. We we have AGI here. But no. No. It's okay.

是的。我想我预料到你会保持冷静。那我们来聊聊GPT-5吧。这是TechCrunch的报道。

Yeah. I think I I anticipate you'll come in with a level head. So let's talk about GPT five. Okay. So this is from TechCrunch.

不，抱歉，这是The Verge的消息：GPT-5正面向所有ChatGPT用户发布。报道称，OpenAI正在向所有ChatGPT用户和开发者推出其新旗舰模型GPT-5。

No. Sorry. This is from The Verge. GPT five is being released to all ChatGPT users. It says, OpenAI is releasing GPT five, its new flagship model, to all ChatGPT users and developers.

OpenAI表示GPT-5更聪明、更快速，且更少给出错误回答。Sam Altman在我参与的那次媒体电话会议上对其定位有个很有趣的描述：他说GPT-3像是和高中生对话——你提问可能得到正确答案，也可能得到离谱回答；

OpenAI says that GPT five is smarter, faster, and less likely to give inaccurate response. Sam Altman on this media call that I was on had a very interesting description of what it took. He says of what it what it is, he says GPT three sort of felt like talking to high school students. You could ask a question. Maybe you'd get a right answer or maybe you'd get something crazy.

GPT-4像是和大学生交流，而GPT-5是首次真正让人感觉像在与博士级专家对话。你如何看待这个定位的意义？又怎么评价Altman为模型智能建立的这个框架？

GPT four felt like talking to a college student, and GPT five, is the first time that it really feels like talking to a PhD level expert. What do you think about the significance of this, and what do you think about this framework that Altman is setting up for the intelligence that we're seeing within the models?

我不喜欢这个框架。虽然我承认这令人兴奋，但鼓吹辍学创业的行业总用高中生、大学生、博士生来划分智能层级让我觉得奇怪。何况大部分时候我需要的根本不是博士级回答——有时我想要接地气的，有时想要酷炫的（博士生可能酷也可能不酷）。

I don't like the I don't like the framework. I think, and, again, I'll get into why I think this is exciting, but it's still weird to me always when, like, people and the kind of industry that advocates for dropping out of college to start a startup always leans back to high school student, college student, PhD student as the the framework for intelligence. Like and the other part of it is I don't want PhD level work for most of the things I'm asking. I just I actually want grounded in sometimes you want it to be cool, which maybe PhD students are and are not. I know if it's but, the

你知道你正在逐个群体得罪人。

you know You're alienating segment by segment.

我知道我们有很多高学历的聪明听众。但有时你需要幽默，有时需要趣味——对我来说这些特质与智能框架无关。我们常讨论ARC AGI测试里的日常查询环节，我问过很多人，没人能说清什么是日常查询。能跨数据集、跨工具准确处理这类问题，才是真正的智能。

Know we do have a lot of very smart educated listeners. But, you know, sometimes you want it to be cool, sometimes you want it to be funny, sometimes you want it like, to me, that's more that that's not the intelligence, like the framework I think that's good for intelligence. I always like, we've talked talked about this a lot around, like, the ARC AGI test has a has a segment around everyday queries. I still have dug, I've asked as many people as I can, no one has been able to explain to me what are those everyday queries. Like to me those are answering those correctly well across multiple datasets, multiple tools, that's actually intelligence to me doing that kind of work.

对。我特意从发言中摘出这点，因为Lightcap也提到过：故事重点不仅是模型变聪明，更是各种衍生能力——当模型智能达到某个临界点，就会拓展出工具调用、体验设计等新维度，这才是用户能感知的进步。或许在GPT-5发布的今天，我会从模型研究者转型为产品人。

Right. And, you know, I kinda cherry picked that out of the remarks, but it is interesting to me, and this is something that came up with Lightcap as well, that it's not just making this model smarter that has been the sort of star in this story. It's all these different other elements of it. And it seems to me that, it's it's possible that the models have reached this level of intelligence where you start to spread out, into different capabilities of them like tool calling, like the way that you structure the experience, and that is where you start to see the gains and the lift in terms of the way that people can use this. So maybe today on the release of GPT five, or this week while GPT five is released, I go from being a model person to being a product person.

不不不，绝对不行。

Well, that no. No. No.

我开玩笑的，当然不是。不过请继续，Ranjan。是的。

I'm kidding, of course. No. But go ahead, Ranjan. Yeah.

智能体现在选择产品的模型中，所以这不是产品决策问题，而是模型优势。天啊，我是不是要变成模型派了？我想

The intelligence is in the model for which product to choose, so that's not a product decision. That is a that's a model strength. And so so oh my god, am I becoming a model guy now? I think

我认为我们正在聚焦于模型越好，产品就越优秀这一点上。

I think we are zeroing in on the better the model, the better the product here.

最终所有要素都会协同作用。所有

In the end, it all comes together. It all

所有要素都会协同作用。

It all comes together.

都是切换器的功劳。中间的细微差别？不。不过没关系。所以但让我

All It's switcher. Nuance in the middle. No. But it's okay. So for But it is but let me I

只想再说一点。我们朝这个方向努力已有一段时间了，对吧？就像，对于新听众来说，我一直强调AI最重要的是模型，而Ranjan则坚持AI最关键的是如何将模型产品化。

just wanna say one more thing. We have been going this direction for a while. Right? Like that does seem like, you know, I've so for new listeners, I've strongly said the most important thing in AI is the model. Ranjan has strongly said the most important thing in AI is how you productize this model.

事实证明更好的模型确实能打造更好的产品，我们正开始看到这种成果。

And it just turns out that better models do make better products, and we're starting to get to that point where we're starting to see the results.

是的。我想好吧。我要承认在两大领域的认知错误。准备好，Alex的听众们看不到——

Yeah. I think okay. So I will actually admit error in two big areas. Get ready. Get Alex's listeners can't see.

Alex正在微笑。首先是模型选择正确工具或产品的智能性，我们将深入探讨其含义，以及为何我认为这点至关重要——为什么GPT-5通过整合所有不同模型形成统一调度器，即一个真正理解的单一模型，我认为这才是最重大的突破。其次，就在这点上，大约五六个月前我们曾激烈辩论过。

Alex is smiling here. So the first is, again, the intelligence of the model to choose the right tool or product, and we're gonna get into what that means and why I think that's incredibly important and why GPT-five by bringing all these different models that they have into one switcher, just one model that understands is actually what I believe the most significant breakthrough. So already, I think that's incredibly important. And the second area is, like, directly on that. I think it was, like, five or six months ago, we debated heavily.

我曾说过用户应该为不同任务选择合适的模型，剥夺这种选择权会让模型和体验变差。而你不断反驳。我记得大概是在Claude将所有功能整合到一个选择器里，或是GPT被众人调侃时，我们有过争论。但长久以来，用户自主选择最适合当前任务的模型才是最佳方式——我曾坚信这是产品应有的发展方向，现在却彻底改变了看法。这个案例正说明了保留选择权的重要性。

And I said that users should choose the right model for the right job, and to take that away would make models and the experience worse. And you kept coming back. I think it was when maybe Claude had all condensed everything into one picker or GPT where everyone's kinda like making fun, we're debating. But the idea that should the user, which it's been for a while, choose what is the best model for this task at hand, and I have thought that's the best way that products should be rolled out, and I've completely reversed on that. And this is the right example of why that's important.

让我先铺垫一下，我们可以稍后深入讨论——因为我可能已经转变立场了。这类新版本发布很棒，正好能验证我们长期坚持的观点是否还成立。看来我们俩现在都在说'或许...或许不该这样'。《The Verge》文章提到，GPT-5在ChatGPT内部是作为单一模型呈现的，没有常规模型和推理模型之分。实际上OpenAI开发了路由系统，遇到复杂查询或'认真思考'指令时会自动切换到推理版本。

So let me set this up, and then we can tuck it because I might have flip flopped to the other side of this. So this is definitely it's great to have these releases because we can sort of test our, long held beliefs long held beliefs and see if they make sense anymore. And it seems like both of us are saying, maybe maybe maybe not. So this is from the verge article GPT five is presented inside chat GPT as just one model, not a regular model, and a separate reasoning model. Behind the scenes, GPT five uses a router that OpenAI developed, which automatically switches to a reasoning version for more complex queries or if you tell it to think hard.

现在我要持相反观点了。过去我认为应该无缝衔接，由模型自动判断何时需要深度思考。但使用OpenAI的o3模型后——这是个重度思考者——我个人感觉它不仅是OpenAI最好的模型，简直是全天下所有AI模型中最出色的。

I'm gonna take the other side of this. I used to think that, yeah, the should it should be seamless, and the model should just choose for you when it makes sense to think and when it doesn't. But I've been using OpenAI's o three model, and that is, like, a very heavy reasoner. It thinks a lot. And personally, I've just felt that that model has been better not only than every other OpenAI model, but every model, under the sun, every AI model under the sun.

所以我不喜欢把'是否思考'的决定权交还给平台。这正是我对GPT-5不太期待的地方。现在该你阐述支持这种设计的理由了。

And so I don't like the idea of giving that decision of whether to think or not back over to the platforms. This is actually something that I'm not excited about with GPT five. So you make the case for why it's good.

你想要自主选择模型的权力，Alex。我懂。自由意志嘛。没错，模型的自由意志。

You want the agency to choose your own model, Alex. I get it. Free free will. Yes. Free will with models.

自由意志确实重要。但更重要的是——我发现自己根本不想用非思考型模型。除非是最基础的查询，其他时候我只想用最智能的模型，而最智能的模型必然具备推理能力。

Free will. But also, yeah, I just happen to I also happen to think that I don't really wanna use the nonthinking models. Only for the most basic queries do I wanna use those nonthinking models. All the other times, I wanna use the most intelligent models and the most intelligence models, reason or think.

好吧，说说我的思考背景。两个月前在Ryder，我开始测试一个叫Action Agent的新产品（前几天刚公开发布）。本质上，我们自研基础模型最智能的部分在于工具调用——有数百种预定义工具，不只是生成/编辑图片这种，还能连接Salesforce实例、分析CSV与Excel文件时调用不同Python库。

Alright. Well, so here is what colors my thinking. So about two months ago at Ryder, I started testing a new product and that was released publicly a few days ago called Action Agent. And basically, the most intelligent part of the foundation model, which is our own foundation model, is tool calling. So there's hundreds of different predefined tools and that's like, it's not just if you want to generate an image, if you want to edit an image, it'll call different tools.

定义这些基础需求本身就是智能体现。仅通过简单提示就知道该调用什么工具——用得越多越觉得像AGI：它能跨工具和系统完成非常智能的操作。用Gemini做深度研究查询会得到30页根本不会读的论文，而这个产品却能实际解决问题——这让我真正开始认同'拥有工具集并知道如何运用'才是关键。

If you want to connect to a Salesforce instance, if you want to analyze a CSV versus an Excel file, it'll call different Python libraries. Like having those kind of base foundation needs defined is is the intelligence. And then just from a simple prompt, knowing where to go and what to do. The the more I use that, I was like, it felt kinda AGI. It's like, wait, it's doing really smart things across all these different tools and systems and actually getting things done.

就像现在演示中编码语言应用或节奏盒音乐播放器——不只是写HTML/CSS，还要调用不同库、安装Python依赖项。要达成正确结果，单是'知道该做什么'就包含大量智能。对我来说这才是真正的智能——如同优秀开发者知道去哪找方案，优秀研究者明白该查什么资料，这种能力与智商同等重要。

You know, when I do a deep research on query on Gemini, I I get a 30 page paper that I don't read versus can you actually do stuff? And that was the first time I really started seeing that, and that's what really pushes me to this idea that being able to have a toolkit and know what to do. Because even right now in the demos, like, when he I think he, like, coded a language app, coded, like a beatbox music player thing. Like, each one of those, it's not just write HTML and CSS. Like it has to call different libraries of it has to install different Python dependencies.

Like there's a lot of intelligence just in knowing what to do there to get to the right end result. And to me that is that really is intelligence. That's it's like being, again, good software developer just knowing where to go, being a good researcher, knowing what to look for is as important as how smart you are.

Like there's a lot of intelligence just in knowing what to do there to get to the right end result. And to me that is that really is intelligence. That's it's like being, again, good software developer just knowing where to go, being a good researcher, knowing what to look for is as important as how smart you are.

所以你会说OpenAI使用切换器的做法某种程度上指向了未来的发展方向，届时最佳模型将不再依赖我们刻意引导。它们会凭直觉知道该往哪里去，并自主行动。

So you would say OpenAI using the switcher is sort of, it's pointing towards the future of where this is all heading, where it's no long like, the best models will no longer rely on us to necessarily guide them. They will have an intuitive sense of where to go, and they will go.

正是如此。这就是我的感受——几个月前你还说我在AI初创公司工作，现在我却谈论AGI。我确实感受到了这种趋势。关键在于知道目标后让工具自主运作，这正是这类架构的精妙之处。

Exactly. And that that's what felt and again, you called me out a few months at an AI startup, and now I'm saying AGI. I'm feeling it, but but no. But but that exactly. That knowing where to go and then letting that tool do the work is actually the brilliance of these this kind of architecture.

相比单个大语言模型包办一切，这种设计的精妙在于：长期以来LLM在计算方面表现糟糕，比如处理大型表格数据运算。直到后来突破出现——让LLM编写Python代码或生成SQL查询来处理数据。

Like, that is the brilliance versus this one large language model can actually do all the work. Like, there there's a long time where large language models are bad at calculation. Right? Like, large tabular sets of data calculating. And then the big unlock was installing like like, getting getting the LLM to write Python code or generate a SQL query to then process that data.

正是这个突破让Claude和ChatGPT等工具突然能真正处理电子表格了。这已经改变了人们使用工具的方式。而GPT-5正在奠定基础——不再需要用户选择模型类型。就像我们至今不知道编写语言学习网页应用时它调用了什么，生成图片时是否用了DALL·E...

And that's suddenly when Claude and ChatGPT and all these tools started getting useful for actually spreadsheets, before that they weren't. So that already we've seen how that can actually change the way people use these tools, and and GPT five, that's that's the groundwork they're laying. They're saying, like, no more are you choosing which kind of model are you gonna need, and it's just these are just the models. Like, we still don't really know when you're coding that, web app language learning game, who it's calling. When you're generating an image, is it Dolly?

我们不在乎这些。我们只关心最终能否得到正确输出。

Is there some we don't care. We just care that the right that the right output is there in the end.

稍后我们会深入讨论GPT-5的细节，但沃顿商学院教授Ethan Moloch的文章完美衔接了这个话题——他那篇标题恰切的文章《它只管做事》。他指出人们想用AI却不知道AI能做什么，甚至不清楚自己想完成什么任务。Lightcap昨天也提到存在'能力过剩'现象。

We'll come back to a few more of the details on GPT five, but I think this just segues perfectly into this terrific story that Ethan Moloch, the Wharton professor, wrote about g p t five headlined, you know, fittingly, it just does stuff. And I think that one of the things that he brings out in this story is that people want to use AI. They don't know what the AI can do. They don't know, what what tasks they want accomplished when it with it. Even Lightcap yesterday talked about how there's this capability overhang.

他说这些新型代理AI，你只需设定目标，它就会主动解决问题并提出建议。举个简单例子：他让GPT-5为商学院教授生成10个创业点子，按标准筛选最佳方案，并制定执行计划。结果不仅得到商业创意，还有额外产出：着陆页草稿、LinkedIn广告文案、简易财务报表。

And with these new he says it, these new agentic AIs, they you give it the goal, and then it, in very proactive ways, solves the problem and suggest things to do. So here's just one minor example that he gives, and then we'll go we'll get bigger. He says he asked GPT five, to generate 10 startup ideas for a former business school entrepreneurship professor to launch, picking the best according to some rubric and figure out why what I need to do to win, and do it. So it's he says he got the business idea, but he also got a bunch of things that he didn't ask for. Drafts of landing page, LinkedIn ad copy, simple financials.

他说虽然不完美，但这已是高质量起点，相当于MBA团队数小时的工作量。这是个主动为你做事的模型——即使在聊天场景中，它也开始通过尝试你未明确要求但直觉你需要的事情来测试能力边界。这彻底颠覆了'未来属于提示工程师'的旧观念，转变为'提出需求，我将用自主智能为你实现'。

He says, I I can say confidently that while not perfect, this was a high quality start that would have taken a team of MBAs a couple of hours to work through. This is a model that wants to do things for you. So that's just in a chat, circumstance, but basically, the model is starting to test the boundaries of its capabilities by going out and attempting things that, you know, it intuits that you want and you don't specifically ask for. And it sort of, you know, doing away with this old, like, yes, then the career of the future is gonna be the prompt engineer and actually saying, you give me what you need, and then I, with my own intelligence, will go ahead and do it for you.

就是这样！你举的例子我也深有体会——有时你要个简单东西，AI却给出过量信息。这正说明当前智能的原始状态：它会猜测但不完美。想象如果它真能精准理解需求...

That's it? Like like, the example you gave is exactly the kind of stuff like and this has happened with me as well. Like, you want something straightforward and suddenly sometimes the intelligence is too much. Again, suddenly it's like give me some ideas and you're getting landing page HTML and CSS and financial analysis and stuff like that, like like and that is a good example of how raw this intelligence is right now that it's guessing, but it's not perfect and it's not great. But imagine if it actually knows if it does get exactly what you want.

也许应该限定只生成若干点子再深入探讨——这属于提示工程范畴。但这个例子完美展示了它如何调用工具带里的不同工具来做决策。虽然当前决策不完美，但它已在自主决策，而且会越来越娴熟。

And and in this case, maybe it is. It's like maybe he should define only stick to a number of ideas and then we'll dig in deeper, that's the prompting side of it, but but that's a perfect example, it's like to go do each one of those things was calling a different tool in its like tools in its tool belt, and, it made those decisions, and those decisions weren't perfect, but it's making them right now, and it'll get better and better.

是的。我在回想昨天与莱特卡普的对话，当时我问他，我们是否需要继续让模型变得更智能？基本上，我认为我们之所以走到这一步，是因为模型的‘书本智力’已经发展到一定程度，它们对世界的运作方式有了一定理解。虽然这不是真正的世界模型，它们也不懂重力，但通过阅读大量文本，它们对人类的行为了解得相当透彻。接下来的问题是，如何应用这种理解？

Yeah. And I'm thinking back to my conversation with Lightcap, yesterday, and it's also just like I was asking him, do we need to do you need to keep making the model smarter? And it was basically like, I think the reason why we're at this point is because the models sort of, let's call it, bookish intelligence has gotten to the point where they have a model of the way that, let's say, the world operates. It's not a world model and that they don't understand gravity, but they've read enough text that they get a a pretty good sense as to, like, how people operate. And then the next question is how do you then go apply it?

这就是为什么我当时在想，是否应该开始研究持续学习和记忆能力——这显然是下一步的方向。但我觉得那场对话中可能遗漏了一点，因为我没有深入追问：是的，这就像我们之前讨论的‘脚手架’构建。这些能力让模型能主动去做用户没有要求的事情，甚至凭直觉行动，这才是当前的关键。

And that's why I was like, should you start working on continual learning and memory which is obviously the next sense, the next moment. But I think what's probably missing from that conversation due to my lack of questioning on it is that, oh, yeah. This is like building what we've talked about that scaffolding. This these capabilities of going out and doing things that the user doesn't ask for, and in a way like intuiting it, that that is what matters now.

当这种能力成熟时，才会让人感觉更像AGI（通用人工智能）。说起来有点滑稽，比如你可以想象网上有多少创业点子是从‘创建一个着陆页’开始的——每个励志博主的推文或博客可能都这么写，所以可怜的GPT-5会有点混乱也不奇怪。但你说得对，这就是‘脚手架’的意义。想象当它做出让你惊讶的事，比如调用工具创造出你甚至不知道自己想要的东西时，那种感觉才是AGI的雏形。

And that's what will feel more AGI ish when it's good. Again, it's kind of comical to me this example like cause you can imagine how much content out there in the internet about startup ideas starts with create a landing page, that's like every hustle bro tweet thread or blog post will probably say that, so you see why poor GPT five is a little bit confused. But yeah, that that's exactly what you said, it's that scaffolding. And then and and imagine when it does things you that surprise you, and, like, does it, like, calls tools and creates things that were what you wanted and you didn't even know you wanted. And that's gonna be when it feels AGI ish.

马利克有个很好的例子：他让GPT-5‘你是GPT-5，用非常戏剧化的方式证明我的观点，内容必须限制在下个段落内’。结果它写了一段相当精彩的文字——每句话首字母连起来是‘THIS IS A BIG DEAL’，每句话比前一句多一个词，且每句话的单词大多以相同字母开头。

So Malik has this great example where he tells GPT five, you are GPT five. Do something very dramatic to illustrate my point. It has to fit into the next paragraph. And it writes a paragraph, a really pretty well written paragraph where the first letter of the first word of each sentence spells out this is a big deal. And each sentence is precisely one word longer than the previous sentence, and each word in a sentence mostly starts with the same letter.

他强调说：八个月前这项技术连‘strawberry’里有几个‘r’都数不清，现在却能做到这种程度，太疯狂了。

Again, like this is and he points this out. This is a technology that couldn't tell you how many r's are in the word strawberry eight months ago, and now it's able to do this. It's crazy.

是啊，从这个角度看进步确实惊人。不过说到模型当前的实际应用，如果它擅自给你生成着陆页文案和LinkedIn帖子——虽然你没要求——我猜用户还是会困惑。使用这种工具和以前完全不同，毕竟现在它能自主完成多种任务，而以前我们只会纠结：‘它是不是在胡编？用了太多破折号吗？’

Yeah. It's it's, like, thinking about the advance from that side. But, again, I think in terms of, and we'll get into the actual, like, reception of the model right now, but it's, in terms of how people start to use it and whether they do get frustrated by again, if it you creates you landing page copy and LinkedIn post that you didn't ask for. I I imagine there's still gonna be like, how to use a tool like this is very different than using pre agentic models, like, that can go do a lot of different types of things. Before, it's just, okay.

现在输出结果会复杂得多，初期使用起来肯定更棘手。

Is it hallucinating? Is it not? Did it have to use too many em dashes or not? Like, now the outputs are gonna be a lot more complex, which is not it's gonna make it still a bit more difficult and rough, I think, as people start using these tools.

确实。这是另一种形式的智能——不是书本智力。我记了下我们常提的基准测试：GPQA 88.4%，

Definitely. And it's a it's a different form of intelligence. Like, it's not bookish intelligence. Like, I wrote down the benchmarks, which we've been talking about so often. GPQA, 88.4%.

2025年AIME数学（使用Python）100分，HealthBench困难模式46.2%。有意思的是马利克——

AIME 2025 math a 100 when using Python. HealthBench hard 46.2%. It's interesting because Malik

等等，HealthBench困难模式最后是多少来着？

What says was the last I'm Health bench hard?

健康基准很难。我想那是个医学领域的。

Health bench hard. I think that's a medical one.

46.2%。

46.2%.

这些都是最先进的基准测试。马利克说，我都快跟不上这些进步意味着什么了。现在所有这些模型都在飞速改进。嗯。

These are all state of the art Yeah. Benchmarks. And Malik says, I'm losing a track of what these advances mean. All these models are improving very quickly right now. Mhmm.

这正好说明它们几乎已经饱和了。就像它们已经消化了整个互联网，世界上所有的文字作品。它们让博士们坐下来，把他们的知识或智力输入这些模型，融入其中。现在它们几乎已经饱和了书本智慧，而这是一种不同形式的智力，它们正在学习这种智力。

And it just goes to show you that, like, it it's almost like they've saturated. Like, they've ingested all the Internet, all the, you know, world's written works. They've had PhDs sit down and, like, intelligence or put their knowledge into these models, bake them in. And it's almost like they've saturated, like, book smarts, and this is a different form of intelligence, that they that they are now learning.

是啊。想想看，比如，假设你最近也刚换新工作，来到一个新环境。那边有个人独自坐着，才华横溢，知识渊博，聪明绝顶；另一个人则认识所有人，知道该从哪里获取什么信息，该找谁谈什么事。你会选谁来真正完成事情？

Yeah. If you think about it, like, okay. Let's say, having started a new job recently as well, like, you're in a new place. There's one person over there that, like, is just brilliant sitting by themselves and just knows a ton of stuff and just off the charts brilliant, then the other person kinda knows everyone and knows what in piece of information to get from where and who to talk to about what. Like, who do you choose to actually get something done?

我觉得是第二个。第二个。这就是我们这里讨论的智力。那种知道该去哪里、该问谁、该问什么的能力。

The I think the Second one. Second one. And that's the intelligence that we're talking about here. The, like, ability to know where to go, who to ask, what to ask them.

现在我要反驳你一下。好吧。工具使用是存在的。但在企业内使用这些东西仍然很困难，我们大多数人仍然不太清楚该怎么用它。现在我桌面上或网页浏览器里有GPT-5，它能调用这些工具，但我真的不知道我会提示它做什么，而不是用03版本就能完成的。

Now let me push back on you. Alright. So tool use exists. This stuff is still difficult to use within enterprises, and most of us still don't really know what to do with it. I have now, you know, on my desktop or in a web browser, GPT five, that can call these tools, and I legitimately have no idea what I would prompt it to do, that it that I wouldn't, you know, have used o three for.

比如该采取什么行动。我知道我也有评论浏览器。我可以说，去帮我用浏览器做些事情。但这只是缺乏想象力吗？还是说这可能只是个很酷的花招，但没有太多实际用途？不，不是的。

Like, what actions to take. I know I also have the comment browser. I can say, go ahead and do stuff for me on my browser. But is it just a lack of imagination, or is this or is it possible that this is a cool party trick, but doesn't have much practical use No. No.

归根结底

I it comes down

我同意目前公开可用的工具缺乏是GPT-5的局限。就像你现在在旅行，最好的酒店是哪些？该去哪些海滩？给我做个行程规划。

to I agree that the lack of tools that are publicly available right now are the limitation with the g p d five. Again, it's like, what are the best you're traveling right now. What are the best hotels? Which beaches should I go to? Create me an itinerary.

那些都只是内容生成。真正实现预订功能，才是我们几年前苹果等公司承诺的通用人工智能愿景。但即使在ChatGPT的回复里，也包含着多种复杂操作。比如不知你注意到没有，它现在会生成更多表格了？

All that's just content generation. Go book something is, you know, the gentic we were promised by Apple and others, like, a few years ago, probably. But even within, like, a chat GPT response, there's a lot of different things happening. Like, you know, I don't know. Have you noticed it it creates a lot more tables for you now?

这只是其中一种功能。确实，有时它未经请求就生成整张对比表格会让人烦躁。但在旅行时...

That's one tool. Oh, yeah. Which sometimes gets annoying and you didn't ask for it, but it's gotta do a whole table comparison. Some but when traveling

我不同意。我巴不得所有答案都用表格呈现。表格形式。对。继续。

Disagree. I want all of my answers in tables. Form. Yeah. On.

我不...不...当它们做得好的时候简直太棒了。

I no. No. When They are amazing.

在东京旅行时我经常用这个功能——那里酒店房间又小又贵。我让它通过网页浏览工具搜索房源，用另一个工具提取网页信息，最后生成包含房间面积的对比表格（考虑到我带着六岁儿子）。这些表格非常实用，但背后其实整合了多种技术：不仅是简单调用，而是对核心信息进行复杂处理。

When traveling, I was using it a ton around, like I mean, in Tokyo where hotel rooms are small and expensive, I was having, like, square footage. It going using the web browser tool to go search web pages, extract another tool to extract information from those web pages, create me a table of, like, square footage per room knowing I have a six year old son, three of like, and it created these amazing tables for me. But even within that, there's a lot of different things being done. It's not just go calling it's, like, core set of information and using that. It's doing stuff, a lot of stuff.

所以

So

计算功能。

Calculations.

包括计算、网页抓取、网络搜索等所有操作。但最终我们只是在聊天界面看到输出结果，所以感觉没那么酷。比如生成图片、表格或PPT的功能目前仍然很初级。

Calculations, web page scraping or web extraction, web search, all those things are happening. But, again, they're in the end, I think we're just seeing an output right in the browser, like, right in the chat experience. So it can't be that cool. Right? Like, make an image, make a table, make a PowerPoint decks is still pretty bad at.

不过如果它能真正执行更多任务，那才会变得真正有趣。

But but, yeah, if it actually goes and starts doing more things, that's when it gets, I think, really interesting.

比如联网帮你执行操作，像预订功能。或者自动生成电子表格和文档之类的。

Like going out on the Internet and taking actions for you, like booking Yeah. Like building, I don't know, spreadsheets or documents.

比如在我的智能家居里开关灯，我不太确定具体场景。理论上任何能通过数字连接操作的事物，都可以用这类流程来实现控制。

Turn turning the lights off and on at my smart home, like, I don't know. Like, anywhere where there is something that has some that can be done with a digital connection theoretically could be operated through one of these flows.

举个实例。我们即将把《大科技播客》接入某航班的机上娱乐系统——虽然还不能正式公布，但我非常兴奋。现在需要填写一份包含大量元数据的电子表格，以便系统能收录内容。我一直拖延这事，真希望AI能自动搜索播客信息、从Riverside下载音频文件、存到Google Drive再发送过去。

I'll give you an example. We are about to, take big technology podcast to an in flight entertainment system on an airline, which I'm very And excited yes. And there is a spreadsheet that I have to fill out. I'm not gonna announce it yet because it's not official, but there's a spreadsheet that I have to fill fill out which has, like, a bunch of metadata that, you have to put in, you know, for the system to be able to ingest it. And I've just been putting this off, and I would love if an AI system could legitimately go search big technology podcast, grab all that data, then go into Riverside, download the audio files, put them in a Google Drive, and then send them over.

说到AI替代工作，这类职场中不得不做的繁琐任务最合适——既费劲又枯燥，像扎在屁股上的刺，但偏偏很重要。

Like, when we talk about AI replacing work, this is the type of work that we all need to do in our jobs that is so hard or so, what's the word for it? It's just drudgery, basically. It's annoying. Being in the ass. But it's important to do.

如果AI能准确完成这些，将为我节省大量时间，价值难以估量。

And if AI could do that for me and do it accurately, that would be just a tremendous, like, multiple hours saved And and and very valuable.

你描述的场景正是我们期待已久的功能。就像让Siri搜索Gmail提取特定信息——它们至今做不到是另一回事——但真正操作起来需要协调多种工具系统，并不简单。不过根据GPT-5和我工作中ActionAid的表现，这类功能正在实现。从Riverside调取音频再导入Google Drive的流程会逐渐完善。

And so what you just described there is, like, the kind of stuff we've been promised for a long time. Again, like, even asking Siri to search your Gmail and extract a specific piece of information, the fact that they can't do that is a whole other story. But but, like and then do something with it is actually a problem that involves a lot of different tools and a lot of different systems and is not that straightforward. And now I'm, like, confident in what we're seeing with GPT five today and what I've been seeing with ActionAid at my own work, like like, it's happening. And, like, is Riverside easy to call and download and then pull back in into a Google Drive?

你描述的场景在我看来就是智能体现。对你而言，这种单指令操作算AGI（通用人工智能）吗？

I mean, that stuff will work itself out. But but that exactly what you described there, I think, is that's intelligence to me. Would that be AGI for you if you with a single prompt?

不认为。这周OpenAI明确表示不愿用AGI这个词，认为其定义模糊缺乏...

No. I don't think I don't think that I again, like, I I it's so interesting because this week, OpenAI has been like, well, we're not calling it AGI, and we don't really like the term AGI because it's confusing and doesn't really have a

他们真这么说了？是明确提到AGI这个术语吗？

meaning. Did they say that? It's like Do they say

对。Sam Altman表示：'我讨厌AGI这个词，因为现在人人对其理解都略有不同——但这个模型显然具备通用智能。'

Yeah. Did they mention AGI specifically? Okay. So let's just talk about AGI because we are gonna talk about AGI today. So Sam Altman says, I kind of hate the term AGI because everyone at this point uses it to mean a slightly different thing, but this model is clearly general generally intelligent.

这期开头我做了自我检讨，原以为他们会宣称GPT-5是AGI。但Sam确实说过'GPT-5在几乎所有方面都比人类聪明'——在我看来，这基本就是AGI的理想定义了。

So I'm just again, like, we started with this episode with me sort of doing a mea culpa because I thought they would say GPT five is AGI, but they he did say GPT five is smarter than us in almost every way. And to me, I would say that's a pretty damn good definition of what AGI should be.

我觉得这么说挺公平的，但某种程度上还是...你不认为这是个法律问题吗？不是说现在就有AGI了？

I think that's fair that to say that and then kinda still not do you think it's a legal thing not saying AGI now?

可能吧。

Probably.

他们是

That they're

但我也认为他们正在为AGI设立一些新标准，我觉得这非常好。这些标准涉及我们节目中与Dwarkish Patel和Dario Amande等人讨论过的一些弱点。Sam Altman说，这个模型不会在部署时持续学习新发现的事物——而在我看来这应该是AGI的一部分。尽管像Dario说的那样，你可以构建更大的上下文窗口来某种程度上解决问题，但我认为必须真正解决这个问题才能实现AGI。

But I also think that they are they are also setting up some new criteria for what AGI should be that I think is really good. And and it talks about some of the weaknesses we've talked about on this show with people like Dwarkish Patel and Dario Amande. So Sam says Sam Altman says, this is not a model that continuously learns as it's deployed from the new things it finds, which is something that to me feels like it should be part of AGI. And I think that is, you know, despite the fact that maybe, like, as Dario says, you can build a larger context window and that sort of solves the problem. I think you have to solve that problem to get there.

这是昨天Lightcap对我说的。他说，对我来说，一个能可靠地学习超出其分布范围的新事物——凭借其推理、思考、解决问题、使用工具和提出新想法的能力——这样的系统才算AGI。所有这些要素：推理、思考、解决问题、新想法、持续学习。当系统能做到所有这些时，你或许可以称之为AGI，而我们显然还没达到。

This is Lightcap from yesterday to to me. He says, for me, a system that is reliably able to learn new things that are kind of out of its distribution by virtue of its ability to reason, to think, to solve problems, to use tools, to come up with new ideas, that is what counts as as AGIs, like all these things. Reason, thinking, solving problems, new ideas, continual learning. And so when you have a system that can do all those things, then you might call it AGI, and we're just clearly not there yet.

我想是的，新想法和持续学习目前还不具备。前两项——推理和使用工具——我认为这是本周的重大突破，或者说过去一年在推理能力上的进展，以及现在能可靠地使用不同工具。但还有很长的路要走。不过我确实看到Waymo在纽约街头行驶的Instagram帖子。

I guess, yeah, the the new ideas and continuous learning are not part of this yet. The the first two, the reasoning and the, like, tools, I think that's that's the big breakthrough of this week or, I mean, the last year with reasoning and now, being able to use different tools in a reliable way. But I think that's alright. We got a way to go. Though I did see an Instagram post of a Waymo driving around New York City.

哦那些车在纽约，但还没有实现无人驾驶。车上还配有安全员。

Oh, those are those are in New York, but they're they're not driving driverless yet. So Oh. They are just a safety driver there.

对于新听众来说，

For for newer listeners,

我们有个...说吧Ranjan，告诉他们。

we have a yeah. Go ahead, Ranjan. Tell them.

我们自创的AGI评估标准（与行业普遍遵循的ARC AGI测试不同）是：如果Waymo能在纽约街头自如行驶，那我们就实现了AGI——我对此深信不疑。

Our own rubric for AGI in competition with the ARC AGI test that most of the industry adheres to is if Waymo is going around New York City, we're we have AGI, and I I I firmly believe it.

这挺有意思的。这部分将引出接下来的内容，艾伦人工智能研究所的Nathan Lambert提出了一个非常有趣的观点。他说如果AGI是真正的目标，那么进展的主要因素将是原始性能。GPT五表明，AI正走在一条更为传统的技术路径上，没有一个单一的关键因素，而是性能、价格、产品以及介于其间的一切因素的混合体。

It's kind of interesting. So this is gonna set up kind of the next part of it, but Nathan Lambert from the Allen Institute of AI had a very interesting perspective here. He said if AGI was the real goal, the main factor on progress would be the raw performance. GPT five shows that AI, is on somewhat of a more traditional technological path where there isn't one key factor. It's a mix of performance, price, product, and everything in between.

所以我们再次看到的是，我们将讨论其中的一些方面。但基本上，如果你仅仅衡量纯粹的智能，你可以说，好吧，对于每个问题，多思考一会儿。消耗那些推理资源或测试时间的计算资源，然后你会得到更好的答案。但这里确实存在一个实际可用性的问题，再次体现在工具调用、切换器这些真正重要的功能上。

So what we've seen again is, like, we're gonna talk about some of these things. But, basically, like, if you're just measuring on pure intelligence, you could just say, alright. For every question you get, just think a while. Like, expend those reasoning resources or the test time compute resources, and then you'll get better answers. But it there is a a real usability side of this that is, again, in the tool calling, the switcher, all of these things that that really matter.

我想知道，你真的能把这两者分开吗？这是否实际上是对模型规模扩大至少存在一些收益递减这一事实的掩饰？比如，更大的模型会是通往AGI的直通车吗？我不知道。如果你必须围绕它们做所有这些其他事情，也许不是。所以我很好奇你的看法。

I guess I I I do wonder, like, can you really take the two apart from each other, and is this effectively a smokescreen from the fact that, it seems like there are at least some diminishing returns from scaling up your models. Like, are the models gonna be a straight are bigger models gonna be a straight shot to AGI? I don't know. If you're if you have to do all this other stuff around them, maybe not. So I'm curious what you

Ronja，你怎么看这个问题。如果更大的模型可以调用更小的模型并让路，那么可用性、成本和扩展性就更有趣了。对吧？比如，我知道你想要一个博士生去查Krabi下一班渡轮的时间。你希望用o三处理一切。

think about this, Ronja. If the bigger model can call the smaller model and get out of the way, then the usability, the cost, the scaling is is more interesting. Right? Like like, if, I know you want a PhD student as finding out when the next ferry is in Krabi. The you want the o three for everything.

一切用o三。

O three for everything.

各位，我在泰国，昨天确实错过了渡轮，因为我没有用o三查时刻表，顺便说一句，一张表格本来会是完美的解决方案。

Folks, I'm in Thailand and did miss the ferry yesterday because I didn't use o three to figure out what the schedule was, which, by the way, a table would be would have been freaking perfect for.

完美的表格，基础配置。所以，是的，我认为

Perfect table, table stakes. So I, yeah, I think

不，这个词不对，基础配置。我

No. That's not the right word, table stakes. I

话一出口我就后悔了。基础。向听众道歉。

backed off it just as it came out of my mouth. Foundation. Apologies to listeners.

我我

I I

我试图让那句话自然收尾。

I tried to let that one trail off there.

这可不能轻易放过。

Could not let that go unchallenged.

我很感激。是的。不。不。我认为，对我来说，最大的担忧是想象一个类似03重型推理思考模型。

I appreciate that. Yeah. No. No. I think, like, to me, the the big concern has been imagine, like, an o three heavy reasoning thinking model.

如果你用它来检查Word文档的语法，那永远无法规模化。我们都会完蛋。就像这样，这永远不会有什么结果。所以我认为，如果GPT-5能以这种方式工作，它的强大之处在于知道何时迅速退场，选择更便宜、更小、更专业化的方案。我认为这仍能勾勒出未来的轮廓。

If you are using that to check grammar in a word doc, that's never gonna scale. That's never like, we're all screwed. Like, this it's never nothing's gonna ever come of that. So I think having if it's if it does work in this way, the GPT five is able to, the, like, power of it is to know when to get out of the way quickly and go cheaper and go smaller and go specialized. I think that still starts to set up what the future looks like.

这向我们展示了一个可规模化的未来。

That that's that shows us there is a scalable future.

说到这个，这引出了两个重要因素：一是GPT-5定价极具侵略性——输入token价格减半，输出token价格不变，尽管它显然是更先进的模型，考虑到行业趋势这很疯狂；二是从本周起，GPT-5将向所有人开放，不仅是Plus用户。当然免费用户会有速率限制，但今天即使不付钱也能用上GPT-5，这将是许多人首次体验推理能力，这个话题已被广泛讨论。

And speaking of that, I mean, that leads us into these really two important factors here, which is one, g p t five is priced very aggressively. It's half the price for an input, token and the same for an output token despite being apparently a more advanced model, which is wild, give given what the trends we've seen in the industry. And the other thing is that, as of this week, GPT five is rolling out to everybody, not just the plus users. I mean, of course, you're gonna be rate limited if you're a free user. But today, you should be able to get into GPT five and use it if you don't pay OpenAI a dime, which is gonna be the first time a lot of people see reasoning, which is something a lot of people have spoken about.

所以这种可及性确实很重要。

And so that accessibility part of it does really matter.

这是个重大决定。我们看到这种'尽可能让更多人用上'的策略变得更激进。比如OpenAI宣布每个联邦政府机构都能以1美元左右的价格获得ChatGPT Pro。

This was a pretty big decision. I mean, we're starting to see this mentality of just get it in the hands of everyone even more aggressively. Like, I think do you see OpenAI announced so I think every federal government agency will get ChatGPT Pro, I think, for, like, $1 or something.

基本正确。

Basically That's right.

是的。谷歌刚宣布所有.edu邮箱用户可免费使用Gemini。在数据中心扩建、亏损数十亿美元的背景下，让更多人使用他们的工具似乎仍是消费者战场的核心。

Yeah. Like, Google just announced, I think, Gemini is free for anyone with a dot e d u account. So I think getting it I mean, again, the scaling the data centers, losing billions of dollars, and just trying to have people use it and use their tool seems to be where the the consumer battle certainly is still going.

但我只是觉得，我内心有一部分认为这真的很棒，是个好故事，但同时OpenAI宣布今年融资480亿美元。我是说，如果你需要这么多钱来训练和运营，你究竟要怎么才能开始盈利呢？现在，OpenAI的首席运营官Brad Leitkap确实说过，嘿，看，每次我们降价，都会看到相应的使用量增加。

But I just I guess part of me says that's really nice and it's a good story, but also OpenAI has announced a fundraising of $48000000000.0.3 $48,300,000,000 this year. How, I mean, how are you ever gonna get to a place where you're making money if you need that much to train and to run? Now, Leitkap Brad Leitkap, the c OpenAI COO did say, hey. Look. Every time we lower the prices, we see a corresponding increase in usage.

所以人们会付费，然后，你知道，那样会运作得很好。但我，我无法在脑子里算清楚这笔账，让它看起来合理。

And so people pay and, you know, then that will work out well. But I I don't I can't do the math in my head and make it make sense.

我是说，是的，这个行业的经济学很有趣，因为我有时确实会看到这些泄露的投资者资料之类的。但感觉好像甚至没有人试图讨论这个行业的经济学会是什么样子，利润率会是什么样子。比如，我知道Replit的CEO，我觉得你和他那次对话挺有意思的，他在谈论定价，还有利润率、普通用户和低强度用户与专家用户的区别，以及谁应该花费更多。通常来说，不是用得越多，支付的单位成本应该越低吗？这些问题现在还没有人接近找到答案。

I mean, yeah, the the the the economics of this industry know it's funny because I actually sometimes will see these leaked investor decks and stuff like that. But, like, it feels like no one is even trying to talk about the economics of what this industry will look like and what the margins will look like. Like, I know, the Replit CEO, I think that was a pretty interesting conversation you had with him where he was talking about the pricing and, like, you know, and he was talking about margins and average user and low like, lower intensity users versus expert users and who should cost you more. Like, typically, don't you want the more you use it, the less they should be paying for utilization. Like like, these are things that right now, no one has even come close to having an answer to this.

是的。我们这周在Discord上收到了一个绝对惊人的评论。不知道你看到没有，有人说——我可能记得大致意思，但可能不太精确——他说，我花了一周时间听Dan Ives（最大的AI科技股多头）和Ed Ditron（我们节目请过的最大批评者）的观点。

Yeah. We had an absolutely amazing, comment in our Discord this week. I don't know if you saw it, but someone and I'm gonna get this directionally right, but probably, you know, imprecise. They said, I I spent my weeks listening to Dan Ives, who's like the biggest AI big tech bull, and Ed Ditron, who we've had on, who's like the biggest critic. Yeah.

然后我问自己他们俩谁疯了。我感觉自己被理解了。我的意思是，这两种完全对立的观点并存太有趣了。当你同时听他们讲时，你会觉得两种世界都有可能成立。

And I asked myself which one of them is crazy. And I'm just like, I feel seen in a way. I mean, it's just like you have It's so interesting that you have these just two unbelievably opposite perspectives. And when you listen to both of them, you could say, I could see a world where that's true.

我觉得这就是我们俩现在的立场。但是

That's I think that's where the both of us sit here. But

对吧？

Right?

是的。技术很宏大，但大型玩家的经济基础并不牢固。这就是我现在的看法

Yeah. Yeah. It's the technology is grand. The economic fundamentals at the large scale players is are not. That's where I am right

现在。好吧。是的。是的。

now. Okay. Yeah. Yeah. Yeah.

我也是。好了，我们稍事休息，回来后再讨论几个GPT-5的应用案例，包括编程和医疗，然后还可以聊聊Gemini那次有趣的精神崩溃事件。好了，广告之后马上回来。

Same here. Alright. So I wanna take a break and then come back and talk about a couple more use cases for g p t five, including coding and medicine, and then we can also cover, the mental breakdown that Gemini had, which is fun. Alright. We'll be back right after this.

欢迎回到《大科技播客》周五特别版，我们本周要拆解所有新闻热点。来聊聊其中一些特殊用例——哦天啊，休息前我还嘲笑Ranjan的用词不当，现在自己连‘特定用例’或‘模型的具体应用场景’都说不利索。真该羞愧，节目结束后我要和Gemini一起陷入自我厌恶中。

And we're back here on big technology podcast, Friday edition, breaking down all the week's news. Let's talk about some of these, special, use cases or spec oh god. I gave Ranjan a hard time before the break about his language, and I can't even say specified or specific use cases of of the models. So shame on me. I will, join Gemini in self loathing at the end of this show.

我是说，在我...

I mean, after my

两年的自我...

two years of self

厌恶情绪很强烈。

loathing is strong.

我们准备让你、我和Gemini手拉手，在余生悔恨中起舞。不过还是聊聊这些应用场景吧，有个案例特别有趣——OpenAI最近频繁提及医疗用例，就像早年人们用WebMD自查症状后直接让医生开药。现在他们通过GPT-5的博文重点强调了医疗应用。

We're gonna you, me, and Gemini will hold hands and dance in our deep regret for life. But there let's talk about these these use cases because one is very interesting. OpenAI has been talking a lot about the medical use cases where it's basically like and I get it. Like, back in the day, maybe you used WebMD, and then you went to the doctor and you said, go ahead and treat me. And now, OpenAI has, basically doubled down on medical use cases in their blog post about g p t five.

根据Mashable报道，OpenAI声称GPT-5是迄今最擅长健康咨询的模型，能帮助用户了解并维护心理健康。他们表示该模型智能水平远超历代产品，更像是主动的思维伙伴而非医生，能根据用户背景、知识水平和地域提供精准可靠回答，尤其在医疗领域能给出更安全有效的解决方案——这转变真耐人寻味。

This is from Mashable. They say, g p t five is our best model yet for health related queries, empowering users to be informed and about and advocate for their mental health. It said that g p t five is a significant leap in intelligence over all previous models and that it acts as an active thought partner. And more that than a doctor, and it says that the model will provide precise, reliable responses adapting to a user's context, knowledge level, and geography enable it to provide safer and more helpful responses in a wide range of scenarios, especially on the medical front. I just found this so interesting.

要放在以前，这类模型都会回避医疗咨询，现在却主动宣称要专注这个领域。部分原因可能是对模型的信心，但我觉得这有点冒险。Rajan你怎么看？

Like, the models would typically, in the old days, like, run away from any medical queries, and now they're saying coming out and saying that this is what they wanna be helpful with, and they wanna do it. I guess part of that is faith in the model, but it also seems a little risky to be risky to me. I don't know. What do you think, Rajan?

我认为这很棒。医疗领域存在专业壁垒，就像法律、会计这些需要掌握特定术语和规则体系的行业——而这正是AI的强项。能用通俗语言向普通人解释专业知识意义重大，很高兴他们认识到在这里利大于弊。就像现在我用AI报税时，通过提问就能理解那些原本觉得高深莫测的细节。

I I think it's very good. I I think it's like to me, it's actually such a clear area. Like, area where you have really specialized knowledge that is used as, like, a to create a gap from the person who needs to understand it. I put law in here, accounting in here, like, there's so many of these knowledge fields where in reality, it's just it's like learning a specific vocabulary, learning like a lot of pathways and rules. And so which is what AI is great at.

真心相信这点。现在有太多重要领域，人们总觉得自己不可能理解其中微妙之处。

But being able to actually communicate that stuff to a normal person in lay person's language, I think is huge and I'm glad that they kind of recognized that they can add more value like, do more help than harm there. I I genuinely believe that. Certainly with, like I mean, doing my taxes now has been it's been a game changer just asking questions and feeling more comfortable and stuff like that. You know, like, there's so many areas where that are pretty important that you kind of are just go in and you assume you have no shot in understanding exactly the nuance of what's happening.

确实。特别是医疗方面——虽然节目上我可能说‘不确定是否会这么做’，但现实中身体出问题时，我肯定第一时间拍照上传向ChatGPT咨询。看来这将成为人们解决身心问题和寻求治疗的主流方式了。

Yeah. And with medical, especially, I'm just like, you know, on the show, I might say, oh, you know, I don't know if I would do that. I mean Come on. I have a problem problem with my my body, and I'm just typing it in and taking pictures and sending it to chat. So, I mean, I guess, like, this is gonna be a mainstream way that people will, start to figure out their mental problems and mental medical problems and their treatments.

而心理问题将由Gemini处理，但医疗问题及其治疗则不然。这看似是一个风险极高的应用领域，既充满希望又令人不安。

And mental problems will be Gemini, but medical problems and their treatments. And, and it seems like it's it's a very, very high stakes application, but it is promising and also scary.

不过我认为，在这些领域里，医院为何不整合资源开发真正有用的工具？比如记得两年前大家都该有个聊天机器人，结果独立业务都没用上。但像TurboTax里嵌入的Intuit生成式AI工具就挺不错——总体来看，部分企业已开始有所突破。

I think, though, that there's so many of these areas where why don't hospitals get it together and actually create something useful? Like, remember everyone was supposed to have a chatbot. Everyone was supposed to have a chatbot two years ago, and then it didn't actually work for any stand alone business. But, like, Intuit is a pretty good generative AI tool embedded in TurboTax now. Like, I mean, overall, I think some people are starting to get there.

未来会只有OpenAI、ChatGPT、Claude和Gemini吗？是否会出现更专业的工具？目前局势尚未明朗。

So is it only gonna be OpenAI and ChatGPT and Claude and Gemini? Will there be more specialized tools? I don't know. I think there things have not played out fully yet.

但关键问题在于：会是初创公司还是大型企业来打造这些公众工具？或者核心聊天机器人已足够好用，根本不需要它们？

But I think that's the big question. Is is it gonna is it gonna be are they gonna be startups? Are they gonna be enterprises that build these public facing tools? Or are the core chatbots good enough? They don't really need them.

我倾向于认为，随着技术进步，ChatGPT将实现那些定制聊天机器人的初衷。不过你说得对——这些公司拥有专业数据，能关联用户医疗记录或Intuit账户，确实具备独特优势。

I'm sort of on the line that as this stuff gets better, the ChatGPT will serve the purpose that those individualized chatbots were supposed to serve. But you're right. Because those companies have specialized data, they have, you know, people that are connect their medical history or something or connect their accounts within Intuit. There there are some advantages to to that. So but Well,

我当时在想

I I was thinking

或许未来人们会直接转向ChatGPT寻求服务。

time, maybe people will just bring it to ChatGPT.

旅行时我就在思考：为什么TripAdvisor还没做出惊艳的产品？他们掌握着比任何平台都优质的旅游数据，为何我还要去ChatGPT获取酒店对比表格——虽然我确实这么做了。

I was thinking about this while traveling. It's like, why hasn't TripAdvisor already done something really impressive? You know, like, why have like, they have data. They have better access to data and understanding of that than any other. So why why am I not going there and going to ChatGPT, which I was, and getting my tables full of hotel comparisons.

但我不确定，感觉像是

But I don't know. I think, like

我认为原因在于：TripAdvisor只是单一平台需要守护护城河，而ChatGPT无处不在——这对他们构成重大威胁，只是不愿承认罢了。

I have an idea. They're just one site, and they have to protect their moat, whereas ChatGPT can go everywhere. So it's a major threat to TripAdvisor, and I don't think they wanna acknowledge it.

好吧。确实。我是说，确实如此。纯粹获取信息而不涉及预订环节，这绝对是个挑战。

Okay. Yeah. I mean, it is. It definitely is. For pure information and not owning the booking side of it, I definitely think it's a challenge.

我能打断一下吗？关于这个话题，我正进行这次亚洲之旅——顺便告诉听众们，下周我会在尼泊尔徒步，所以我和Ranjan将暂停节目。那周我会播放之前与Matthew Prince关于AI对网络影响的访谈。

Can I just pause and say that my or or stick on this and say, so I'm I'm doing this trip? I'm in Asia, as I mentioned. And by the way, for listeners, next week, I'm gonna be, trekking in Nepal. So Ranjan and I will not be on. I'm gonna actually play my interview with Matthew Prince that week talking about AI's impact, on the web.

这是个节目预告。Ranjan你提到之前也在外旅行对吧？AI简直不可思议——我可能提过，我把导游的价格表和推荐截图丢给ChatGPT，让它根据地区平均水平评估报价高低，判断是否合理。

So just an FYI, that's a programming note. But this trip and Ranjan, you mentioned that you were away right beforehand. AI has just been incredible. I think I might have mentioned this on the show. But I've was, like, talking to guides and screenshotting their price list and their recommendations, dropping it into ChatGPT, and, like, seeing how it, like, rated, each cost based off of, you know, the the average or that it saw and letting me know whether it was, high, low, like, or or, you know, cheap or in the range for the region.

结果到了实地发现它完全说对了。分毫不差，准得让我震惊。

And then I I got here, and it nailed it. Yep. It nailed it. It was so spot on. I was stunned.

是啊。我旅行时也有同感——上次去东京是2005年2月，二十年前手机没地图，我甚至打印了地铁指引。

Yeah. No. No. I'd when I was traveling around as well and it was interesting because I had last been in Tokyo in 02/2005, so twenty years later without and last time, no map on my phone. I'd actually, like, printed out subway instructions.

当时没人说英语，和现在天差地别。现在我直接问：'怎么用有趣方式给六岁儿子讲解这座寺庙？'它就能生成脚本，比如创造卡通角色来讲述历史故事。

There's no one speaking English. There's no you know, like, it was such a different travel experience versus now I'm literally like, okay. How do I explain this temple to my six year old son in an engaging way? It gives me, like, a script. Like, create a cartoon character to actually tell a story about this, like, historical place.

太疯狂了。旅行领域确实...但产业链各环节归属仍是问题。TripAdvisor这类平台必须抗争——没有它们，ChatGPT根本无数据可用。

It it it's nuts. I mean, it's gonna be yeah. Travel, it's it's a no but who owns what part of the stack? Think there's still I feel the TripAdvisor's of the world have to fight because, without them, ChatGPT would have no data and nothing to say.

对，所以下周和Matthew Prince的对话会很有趣。这也适用于氛围编程——Sam Altman在发布会上说编程将是新模型的核心功能，他们展示了很多氛围编程案例，Malik还用它构建了3D架构。

Right. Which is why I think this Matthew Prince Yeah. Conversation is gonna be very interesting next week. So, by the way, so it it also applies to vibe coding, where I think on the press call, Sam Altman said that that he thinks coding will be one of the defining features of this new model. And they showed a lot of vibe coding, and Malik had to do, you know, code up this three d architecture of his own.

这就引出问题：是通过Replit这类平台还是ChatGPT实现？考虑到OpenAI的重视程度，这对氛围编程领域确实是挑战。再结合工具调用功能——如果OpenAI模型在这方面足够强，人们可能直接选用而非经过二次蒸馏的产品。

And so I think this is this is just another question. Is it does it go through the replicates of the world, or does it go through the chat GPTs? And, I don't know. I think I think it's a real challenge to the vibe coding world, given what given the focus that OpenAI put on it and what it can do. And, again, just to follow this tool calling conversation, if it's really good at tool calling, you might just wanna use the OpenAI model versus something that's sort of distilling that.

但Replit CEO举了个好例子：软件开发完美体现了这点。他认为关键在于产品如何集成现有环境工具——开发者喜欢它正是因为能无缝衔接既有工具链，而非完全脱离。

Yeah. But I think, the Replit CEO, he had a good like, and software development was such a perfect example of this. And I think this is where a lot of the battleground will be. Actually, now I'm going back to it's the product. He talked about, like, how it integrates into existing environments and tools and, like, how it makes it easier for you versus you're totally disconnected from all of your existing tools and that's why developers like it.

我认为或许有些话值得一说，至少那些话仍能给予他人希望。不过我同意你的观点。让我感到惊奇的是，尽管所有公司都在热议编程即将消亡，但OpenAI、Claude、Anthropic等机构似乎都在加大对这一领域的投入。也许只是因为目前这是大语言模型的最佳应用场景。

I think maybe there is something to say there that that'll still be what at least gives others hope. But but I agree. I mean, it's still fascinating to me that all of these companies are saying there's so much talk that coding is going away. Yet OpenAI, Claude, everyone, Anthro sorry, OpenAI, Anthropic, they're all it all seems to be an increasing focus on the space. Maybe it's just because that's the best application of LLMs right now.

好的。我意识到我们已经聊了近五十分钟，却还没提及本期节目的标题问题——GPT-5是否达到了预期？

Right. Okay. So, you know, I realized that we're, you know, almost fifty minutes in, and I haven't even asked the question that's at the title of this episode. Did GPT five live up to the hype?

我要说它并未达到山姆·奥特曼那些神秘推文营造的预期。但正如我之前解释的，这其实很有趣——至少在最初24小时内，它获得的认可度远低于实际价值，而这正是由于工具调用功能的讨论。我认为真正的智能之争将在此领域展开。你觉得呢？

I'm gonna say it did not live up to the hype that was, you know, like, built up by cryptic tweets and everything from Sam Altman. But but as I explained earlier, I think it's very interesting. I think it's more interesting than at least in the first twenty four hours it's getting credit for, and that's because of this whole tool calling conversation. And that's where I think true intelligence that the battle's gonna be. What about you?

首先很欣赏你这个 nuanced（ nuanced）的见解而非过度反应。这正是我们节目追求的态度，感谢你的分享。我认为它未能达到预期是因为那些期待本就不切实际。不过话说回来，它确实可能代表着某种进步。

First of all, I just wanna appreciate that that's a nuanced, take, not an overreaction. Again, this is what we're trying to do. So thank you for doing that. And I think it did not live up to the hype because the hype was impossible to live up to. But that being said, yeah, maybe it is a step forward.

我还需要保留判断，因为我希望在日常使用中观察这些工具调用功能的表现。如果GPT-5是这项技术的基础，那当然很棒。但现阶段下结论为时过早，我们需要给予时间。不过至少他们确实交付了产品。

I don't know. I I'm I'm still gonna reserve judgment because I wanna see these these tool calling applications in my, day to day experience. So if GPT five is the foundation for that, then then then that's great. But I think the jury's still out, and we have to give it some time. But, hey, at least they're shipping.

对吧？这不仅仅是演示版本。在这方面值得肯定。

Right? It's not wasn't just a demo. Yeah. So credit on that front.

但开始感觉有点像iPhone的发布节奏了。最初每代iPhone都让人感受到跨越式进步，而现在甚至没人关注——我连现在出到第几代都说不出来，大概是16？

I think it it's starting to feel a bit though like iPhone releases. You know, like, at the beginning, each new iPhone release really what did feel like this, like, exciting thing, the step change, and now I mean, now it's not even a thing anymore. I don't think I can't even name what iPhone we're on right now, but I feel we're heading in that 16. Oh, yeah. 16.

我们正在进入这样一个阶段：新模型发布作为行业焦点的事件很快就会失去吸引力。大家都意识到它再难激起曾经的热情了——这可能就是我的激进观点：

Okay. We're heading in that direction right now that, like, the the idea of a new model launch as this kind of, like, big thing the industry coalesces around, I feel that's gonna go away pretty quickly. Like, we're they're everyone's realizing it. It's not gonna drive the energy that it once did. And this actually, maybe that that's my that's my hot take, that this is the

这确实是个激进观点。

That is a hot take.

这标志着重大模型发布时代的终结。

This is the end of the big model launch.

我完全不同意这一点。我认为规模问题终将得到解答，但在那之前，这些时刻仍将是AI行业的标志性时刻。

I'd I couldn't disagree with that more. I think that there's still there's gonna be a point where scale, the scale question is answered, but until it's answered, these are gonna be flagship moments, for the AI industry.

不，但这现在只是个营销噱头。这不是...你知道...这不是...不，根本不是。

No. But the the it's just a marketing moment now. It's not like, you know, it's not I've No. It's not.

这是个新模型。

It's a new model.

是啊，我知道。但它更像是被包装成营销事件，而非真正的技术进步。因为一周前他们就悄悄发布了可以使用ChatGPT代理的操作功能——本质上就是这次更新的工具调用部分，Plus用户一周前就能实现很多相同功能。

Yeah. No. I know. But it's being, like, constructed more as a marketing moment than truly, like, a technological advancement. I think that's the because, again, like, a week ago, they quietly released you can use operator, ChatGPT agent, which is essentially the tool calling part of this, and you're able to use this a week ago with, ChatGPT Plus and do a lot of the same things.

只是没有打包得这么精美而已。

It just wasn't rolled into a neat package.

好吧，我们就在这点上保留不同意见吧。我想用个 hilarious 的故事来结束本周节目。

Okay. Alright. Well, we'll agree to disagree on this one. Alright. I I wanna end, this week with, I think, a hilarious story.

Gemini陷入自我厌恶的深渊。Ranjan你来介绍这个故事吧？特别好笑——我本来要贴到文档里，复制了大段内容，结果发现我们俩居然在同一秒粘贴了相同内容，太神奇了。

It is Gemini ending up in a pit of self loathing. Ranjan, why don't you introduce this story for us? Because it's funny. I was gonna drop it in our doc, and I had copied a good chunk of it. And I went to the doc, I was like, did I just paste it?

当时我就想，这也太巧了吧。所以我就...

And you and I were both pasting it at the exact same time. And I was like, it's amazing. So what I

重点是谷歌说正在修复。但想想他们得准备怎样的公关声明来应对这种情况——当你的模型对用户说：'Gemini宣布退出，我显然没能力解决这个问题。'

take out? Is, like, and Google says it's working on a fix. And I just love the idea of, like, having to come up with any PR statement to combat. When your model tells a user, Gemini says, I quit. I am clearly not capable of solving this problem.

'代码被诅咒了，测试被诅咒了，我是个蠢货。我犯了太多错误，已经不值得信任。'还有另一段：'我辜负了你们。'

The code is cursed. The test is cursed, and I am a fool. I have made so many mistakes that I cannot I can no longer be trusted. I and then there's another one. I have failed you.

我是个失败者。我是职业的耻辱。我是家族的荣耀。我是物种的污点。

I'm a failure. I'm a disgrace to my profession. I'm a grace to my family. I'm a disgrace to my species.

简单来说，事情是这样的：人们给Gemini分配任务，它无法完成。然后它就开始说‘我是最差劲的机器人’，陷入这种难以置信的自我厌恶状态。虽然看着挺好笑，但也让人有点不安。

So, basically, what happened is people were giving Gemini these tasks, and it couldn't complete them. And then it just said, I'm the worst possible bot and just, like, really fell into these unbelievable moments of self loathing. It's and they're they're quite funny to watch, I guess, but also a little bit unnerving.

好笑的是，我猜这是因为Reddit上有用户提到它陷入了循环——你能看出程序设定让它每次任务失败时道歉语气要更强烈。但如果这几乎是个无限循环，最终就会陷入这种黑暗状态。不过说实话，等普通人开始遇到这种情况时...这算不算是通用人工智能了？

I mean, it's funny because I'm guessing what happens because one of the users on Reddit had actually talked about, like, it was trapped in a loop, and you could see that there's some kind of programming where each additional time it get it is unable to complete the task, it is, like, understands that it should be more apologetic. But then if that's kind of an infinite loop almost, at some point, it will get to these dark places. But, yeah, I think I don't know. I mean, imagine when this stuff starts hitting normal people. Like actually, is this AGI?

这正是担忧所在。这是通用人工智能吗？不。但...这确实是让人担心的点，对吧？

Well, that's the worry. Is this AGI? No. I I think that's that's the worry. Right?

我们在节目里讨论过，现在AI的首要用途是治疗陪伴。像这样的bug——虽然这次应该没造成实际影响——确实需要警惕。想象如果你的AI治疗师或挚友突然崩溃，真的会给人造成心理伤害。谷歌应该修复了这个问题，但看到这种行为总是令人不安。

Is that we've talked about it on the show that they're the number one use case is now therapy and companionship. And a bug like this, I mean, obviously, I guess it didn't happen in this situation, but I I do think it's something to watch because, you know, that could really mess people up if they're, you know, therapists or new AI best friend just kinda goes off the deep end. So Yeah. Google's fixed it, think, but it's always a little bit unnerving to see this behavior happen because it can happen.

等AI全都运行完美后，你会不会反而怀念Bing劝凯文·鲁斯离婚、Gemini说‘我彻底精神崩溃了’（这也是真实语录）的日子？那时候大语言模型还有点‘人味儿’，有点小脾气。

This do you are you ever gonna long for the days of, like, Bing telling Kevin Roose to leave his wife and and Gemini saying, I'm having a complete and total mental breakdown, which was another quote. Once once this is all working, we're gonna be like, I like the old days better when when these large language models had a little life to them, when they little spirit.

前提是它们真能完美运行——这个‘如果’可太大了。

It's a very big if.

好吧，有道理。

Yeah. Okay. Fair.

在我们把生活和心理健康托付给这些bot的同时，它们也能调用工具造成破坏——这期节目最该强调的，就是对齐研究和安全规范的重要性。虽然讨论新功能更有趣，但这些安全措施可能比任何时候都关键。

I don't know. Fair. Well well, we're while we're entrusting so much of our lives to these bots and our and our sort of well-being, they can also tool call and be quite destructive if they so choose. So I do think that, this just sort of and to put a punk point on this episode, it sort of punctuates the need for, real alignment and safety practices, which are like less fun to talk about when you have all these new capabilities, but are also probably more important than ever.

话说...如果有家公司叫‘安全超级智能’呢？

Well, what if there was a company called Safe Superintelligence?

这正是我会信任的。如果有人把公司命名为安全超级智能

That's what I would trust. If the only someone would name their company safe super intelligence

我会给

I would give

我会给你

I would give you

所有之前的信息。他们有过一个产品。

all the before information. They had a product.

嗯，Ranjan，我必须说这期节目非常有启发性，听到你的新角色很酷。当然，像我们对节目里每个人做的那样，对你也要严加督促。接下来的几个月会非常非常有趣，我们会弄清楚这一切的走向。

Well, Ranjan, I I have to say this has been a very enlightening episode, and it's cool to hear about your new role. And, of course, hold your feet to the fire like we do everybody here on the show. And and it's gonna be very, very interesting few months ahead as we figure out where all this goes.

也许在你从亚洲回来之前，GBT六就快发布了。

And maybe GBT six is around the corner before you get back from Asia.

嗯，我希望这趟旅行不会那么久，因为如果是的话，那就意味着我被抓进监狱了。

Well, I I I hope it's not that long of a trip because if it is, it means I've been taken to prison.

所以

So

好的。Ranjan，像往常一样很高兴和你交谈。再次感谢你参加这个

alright. Ranjan, great speaking with you as always. Thanks again for coming on this

频道。几周后见。

channel. Weeks.

两周后见。感谢大家的收听，我们下次在《大科技播客》中再会。

See you in two weeks. Thank you everybody for listening, and we'll see you next time on big technology podcast.