第11集 - 塑造GPT-5.1模型行为

大家好，我是安德鲁·梅恩，这里是OpenAI播客。

Hello, I'm Andrew Main and this is the OpenAI Podcast.

今天我们邀请到的嘉宾是克里斯蒂娜·金，她是OpenAI负责后训练研究的主管，还有劳伦蒂娅·罗曼纽克，她是专注于模型行为的产品经理。

Today our guests are Christina Kim, who's a research lead working on post training at OpenAI and Laurentia Romaniuk, who's a product manager focused on model behavior.

我们将要讨论GPT 5.1。

We're going be talking about GPT 5.1.

是什么让这个模型更出色？

What makes the model better?

他们如何专注于使其个性可调控，以及他们对未来发展的展望。

How they've been focusing on making its personality steerable and where they see things headed in the future.

这是有史以来第一次，聊天中的所有模型都具备推理能力。

For the first time ever, all of the models in chat are reasoning models.

不过我认为，对大多数用户来说，个性是更宏大的概念，它关乎模型的整体体验。

Personality, though, for most of our users, I think is something much larger, and it's the whole experience of the model.

你应该能够通过聊天获得你想要的体验。

You should be able to get the experience that you want with chat.

这里的艺术部分在于如何挖掘出模型的这些个性特质，同时不破坏其可操控性。

Part of the art here is figuring out how to pull out these quirks of the model that can come across as personality without breaking steerability.

我非常期待讨论这些模型以及它们是如何随时间演变的。

I'm very excited to talk about the models and how they've been changing over time.

现在使用'模型'这个词也感觉有点奇怪，因为它似乎包含了更多内容。

And using the word model also feels sort of funny now because it seems like there's so much more.

而一切其实都始于研究。

And everything starts really in research.

在规划GPT 5.1时，设定的目标是什么？

And when GPT 5.1 was being planned, what were the goals?

对我们来说，主要目标之一是解决关于GPT 5的所有反馈，同时我们也在努力将5.5版本打造成一个推理模型。

Yeah, for us, one of the main goals was to address all of the feedback we've been getting about GPT five, but also we've been doing a lot of work to make the 5.5 instant into a reasoning model.

就我个人而言，5.1版本最令人兴奋的是，这是有史以来第一次，聊天中的所有模型都变成了推理模型。

So what the most exciting thing for, personally for me with the 5.1 release is that for the first time ever, all of the models in chat are reasoning models.

现在的模型可以自主决定思考过程，就像我们说的思维链那样，它会根据提示自行决定需要思考的深度。

So the model right now can decide to think, it's kind of what we say, it's like a chain of thought, and it'll decide how much it wants to think based on a prompt.

如果你只是对模型说'嗨'或'最近怎么样'，它不会进行深入思考，但如果你问一个稍微复杂的问题，它就会决定需要思考到什么程度。

If you're just saying like, hi to the model, what's up, it's not gonna be thinking, but let's say you ask it a bit like harder question, and then it'll decide how much it wants to think.

这样它就有时间完善答案、处理问题，必要时调用工具，然后再回来给你答复。

So it gives it time to like refine its answer and work through things, call tools if necessary, and then come back to give you an answer.

有点像丹尼尔·卡尼曼所说的系统一和系统二思维。

Kind of what Daniel Kahneman calls like system one and system two thinking.

是的。

Yes.

将推理模型作为默认模型提供给所有人，就意味着拥有了一个更智能的模型。

Having a reasoning model out for it as a default model for everyone just has a much smarter model.

我认为更智能的模型能带来全方位的提升，特别是在指令遵循这类任务上。

And I think with smarter models, you just get improvements across the board, especially for things like instruction following.

对于很多使用场景，人们可能都没意识到它们需要多少推理能力。

And for a lot of the use cases, people might not even think might require much reasoning.

仅仅是提升智能水平，让模型在某些查询前先思考再回应，就能带来很大帮助。

Just that having improved intelligence, having the model actually think before it responds in certain queries just really helps.

我们已经看到这全面提升了评估效果。

We've seen that improve evals across the board.

当你对这样的产品进行管理时，需要向人们解释有什么不同吗？

When you product manage something like this and do you have to explain to people what's different?

这可能是个挑战，但你会如何解释GPD五和GPD 5.1之间的区别？

It's probably a challenge, but how would you explain what's different between GPD five and GPD 5.1?

是的。

Yeah.

首先这很困难，因为有太多变化。

First of all, it is difficult because there's so much changing.

但在这种情况下，我们想回应的是来自社区的反馈意见。

But in this case, what we wanted to speak to were things that we'd heard as feedback from the community.

在ChatGPT五发布时，我们听到的反馈之一是模型感觉直觉较弱，且不够亲切。

With the ChatGPT five launch, one of the things we heard was that the model felt like it had weaker intuition and that it was less warm.

当我们深入研究后，发现了几个不同的问题。

And when we dug into that, what we found were a handful of different things.

首先，问题不仅仅在于模型的响应方式，比如模型的内在行为。

First of all, it wasn't just how the model was responding, like, as the model's innate behavior.

还包括模型周边的一些因素。

It was also things around the model.

举个例子，我们的模型上下文窗口较短，无法充分保留用户之前对话的信息。

So as an example, our model had a shorter or the context window wasn't carrying enough information about what users had said previously.

这会让用户感觉模型忘记了你告诉它并希望它记住的重要信息。

And so that can feel like the model is forgetting something really important that you told it that you were hoping it would hold on to.

如果你说'我今天糟透了'而模型在10轮对话后就忘了这件事，那会让人感觉非常冷漠。

If you say I'm having a really bad day and the model forgets that after 10 turns, that can feel really cold.

所以这是我们本次发布重点调整的一个方面。

So that's something we adjusted as part of this launch.

部分原因确实在于模型的响应方式，但我们在GPT5中引入的新功能是这种自动切换器，它能在聊天模式与推理模式间切换。

Some of it was actually the way the model is responding, but something new that we introduced in GPT five as well was have this auto switcher that would move you between chat and reasoning models.

这两种模式有着略微不同的响应风格，当你向模型倾诉自己糟糕的一天时——比如部分原因是收到了可怕的癌症诊断——这种切换会让人感到突兀或冷漠。

And those have slightly different response styles, and that can feel really jarring or cold if you're talking to the model about how you're having a bad day and then you say, like, part of it is I got this awful cancer diagnosis.

所以

So the

模型会切换到思考模式，然后给你一个非常临床化的回答，而刚才它还在引导你解决之前遇到的问题。

model switches you to thinking, and you get a very clinical answer for a model that was just sort of like walking you through a problem you were having earlier.

因此，我们实际尝试做出的很多改变都是综合性的。

And so a lot of the changes we were actually trying to make were in aggregate.

我们如何确保这个模型感觉更温暖？

How do we make sure this model feels warmer?

尽管我们在底层做了很多改动来实现这一点。

Even though we were changing a lot under the hood to articulate that.

我们研究的另一个方面是整体指令遵循能力。

Another thing that we looked into was instruction following generally.

5.1版本在遵循自定义指令方面表现更好。

So 5.1 is much better at following custom instructions.

这也是我们听到的另一条反馈：我们发布的每个模型都会有自己独特的怪癖和略微不同的行为。

And that was another piece of feedback we were hearing, which was every model that comes out of that we release is going to have its own quirks and slightly different behaviors.

我认为人们其实并不太介意这一点，只要他们能控制它，只要他们可以说，嘿，那很奇怪。

And I think people actually don't mind that too much as long as they can control it, as long as they can say, like, hey, that was weird.

停下。

Stop.

但如果模型无法延续这种上下文，如果它不能记住这些自定义指令，那就是个问题。

But if the model can't carry that context forward, if it can't hold on to the custom instructions on that, that's a problem.

所以我们努力增强了自定义指令功能，使其更一致地延续指令，以解决部分反馈问题。

So we worked to actually enhance the custom instructions feature so that it more consistently carries instructions forward to address some of that feedback.

最后我想说的是，很多这类事情都是个人偏好，因此我们引入了风格和特质类功能，比如个性设置，这实际上让用户可以引导模型采用特定的回应格式，从而让他们对ChatGee的具体回应方式有更多控制权。

And then, like, the last thing I'll say is a lot of this stuff is personal preference, and so that's why we introduced our, style and trait type features like personality, which actually let users guide the model into certain response formats so that they have a little bit more control over exactly how ChatGee responds for them.

这种切换很有意思，因为现在有多个模型了。

The switching is interesting because there's multiple models now.

不止是一个模型。

It's just not one model.

你清楚地阐述了为什么需要这样做。

You articulated why you need to have that.

当我们谈论切换器和不同类型的模型时，我知道这对大多数人来说可能会有些困惑。

When we talk about a switcher and we talk about sort of different models, I know for most people that can be kind of confusing.

你会如何向人们解释清楚这一点呢？

And how would you kind of unpack that for people?

是的。

Yeah.

我认为我们的模型具有非常不同的能力，要全面掌握可能很难。

I think our models have very different capabilities and it can be hard to stay on top of.

所以部分工作就是继续在应用中尝试不同的功能。

So part of it is just continuing to, like, try the different things in our app.

但产品工作的另一部分肯定是确保我们有正确的用户界面来引导用户选择合适的模型。

But certainly part of the product work is making sure that we have the right UIs to either guide users to the correct model to choose.

这可以是模型切换器。

And that can be the model switcher.

这可以是模型切换器的学习过程，了解在不同情境下哪些类型的回答对用户最有帮助，参考不同的评估标准。

So that can be the model switcher learning, what sort of answers are most helpful to users in different contexts, looking at different evals.

例如，对于推理模型，如果用户需要科学上非常准确且极其详尽的回答，我们可能会通过评估来检验我们是否满足了这类提示的需求。

So for example, for reasoning models, if people want something that's very scientifically accurate and very, very detailed, We might look at an eval to see, are we answering that need, on those sorts of prompts?

我们可以预测何时为用户切换模型。

And we can forecast where to switch users to.

是的。

Yeah.

蒂娜，关于模型切换器以及现在的情况——每个人都能使用免费层的基础模型，也就是推理模型。

Tina, as far as the switcher and now the fact that you have a model that's everybody has the free tier, anybody using the base model is a reasoning model.

这在实际影响上意味着什么？

What does that really mean in impact?

是的，我认为关于如何思考这个问题，还有很多研究上的开放性问题，对吧？

Yeah, I think there's a lot of research, open questions for research for how we want to think about this, right?

所以就像你说的，这是个更快的模型，但它不一定就代表功能简单。

So I think, like you said, it's a faster model, but it doesn't necessarily need to be dumb.

我认为我们的理念是希望为所有人提供尽可能智能的模型。

So like, I think the idea is that we want to get the most intelligent model that we can for everyone.

因此我认为这为我们打开了思路，可以更多地思考如何利用这种最前沿的技术模型来做更有趣的事情，对吧？

And so I think this kind of opens the door for thinking more about like, what are more interesting things we could do with a very, very like state of the art, like frontier model, right?

所以它能够进行更长时间的思考，就像深度研究那样可能需要几分钟，也许更适合作为后台工具调用。

So that's going to think for much longer, like something like deep research where you have it thinking for minutes, maybe that's better use in the background, you can call it as a tool.

我认为还有很多开放的研究问题需要思考，但我们确实将进入一个多模型系统的世界，不再只是单一模型。就像5.1版本，人们可能以为就是一套权重参数，但实际上包含推理模型、轻量推理模型、自动切换器（本身也是一个模型）等多种工具。

So I think there's lot of research open questions of what we want to think of, but I do think we're going be in this world where we do have a system of models and it's not just like a model that you have and there's like lots of different tools and it's not just one, like when we think of 5.1, I think people just assume that it's like one singular set of weights, but I think it's really just like, yeah, this reasoning model, this like lighter reasoning model, this auto switcher, which is also a model in itself.

所以这实际上是多种不同模型支撑的各类工具集合。

And so it's all of these different things and then different tools that are also backed by different models.

因此我认为这套系统，随着我们获得更智能的模型，正在开启更多有趣的应用场景和产品可能性。

So I think this system of things, I think as we just get smarter models, it's opening up more interesting use cases and more interesting product implications.

拥有8亿用户，除了海量数据外，你们肯定也收到大量用户反馈。

With 800,000,000 users, you probably get a lot of user feedback besides just your volume of it.

你们如何筛选这些反馈并加以理解，进而找出利用方法？

How do you sort through that and make sense of it and figure out how you can use that?

是的，我认为实际上很多工作都是从对话环节开始的。

Yeah, I think a lot of it actually starts with a conversation link.

很多时候，当我们能实际看到用户的对话内容时，我们就能准确了解对话中发生了什么，并开始剖析问题以制定针对性解决方案。

So a lot of times when we can actually see the conversations users are having, we're able to see exactly what happened in that conversation and start dissecting things so that we can target a solution.

举个例子，如果我们收到用户反馈说'嘿，我在使用模型时遇到了非常奇怪的体验，它说了些很冷漠的话'，或者'这句话感觉被截断了'。

So as an example, if we get feedback from a user that like, hey, I had this really weird experience with the model that said something very cold or like, this sentence felt very clipped.

如果我能实际看到对话链接，我就能判断'哦，这个用户正处于某项实验中'，这就是为什么特定实验可能在某些情况下对部分用户存在边界问题的典型案例。

If I can actually see that conversation link, what I can say is like, oh, that user was in an experiment and like, good example of why this particular experiment might have some edges for certain users in these cases.

但至少对于自动切换器（负责在5.1聊天模式和5.1推理模式间切换），我们正在通过用户的不同信号来判断：这个功能对他们有效吗？

But at least for the auto switcher, which takes you from, 5.1 chat to 5.1 reasoning, we're looking at different signals from users to figure out, like, is this working for them?

还是无效？

Is it not?

每个回答在事实准确性方面表现如何？

How is it is each response performing on factuality?

延迟情况怎么样？

What is the latency looking like?

因为并非所有用户都愿意等待，即使他们想要更好的答案。

Because not all users want to wait even if they want a better answer.

因此这需要艺术与科学的结合，平衡多种信号来判断何时切换以及如何最有效。

And so it's, it's a bit of art and science balancing a bunch different signals to figure out when to switch and how that's most effective.

当你想从智能角度提升模型时，比如智商层面，我们有基准测试和评估方法。

When you're trying to improve a model from an intelligence point of view, like an IQ point of view, we have benchmarks and evals for that.

但谈到情商时，你该怎么做？

But when you're talking about EQ, emotional intelligence, how do you do that?

如何衡量这方面的进展？

How do you measure progress there?

是的，我的意思是这是一个非常开放性的问题，我认为实际上我们研究团队议程的一部分就是所谓的用户信号研究。

Yeah, I mean, is something that's very open ended and I think actually one of the things that's part of my research team's agenda is what we call user signals research.

因此这涉及到训练奖励模型，并在强化学习过程中获取信号，以便应用于我们的用户生产数据。

And so this is training reward models and getting signals during RL that we could use against our user prod data.

这类研究我认为非常有趣，因为我们可以获取大量关于用户意图的信息。

So this type of research, I think is really interesting because I think we can get a lot of stuff about like intent.

而且我认为当我们考虑情商(EQ)时，它只会随着更智能的模型而变得更好，因为它真正试图理解用户想要什么。

And I think when we think about EQ, it's also just only gets better with like smarter models, because it's really trying to understand like, what does the user want?

用户需求的上下文是什么？考虑到对话中有这么多其他消息，并且你了解用户的记忆和历史记录，模型应该如何最佳响应？

What is the context of what the user wants and how to how should the model best respond given the fact that you have this many other messages in the conversation and you know this stuff about the user's memory and history?

是的。

Yeah.

我认为情商还有另一个要素，就像当我想到高情商的人时，他们具备倾听的能力、记住你所说内容的能力，当然还有捕捉蒂娜提到的那些微妙用户信号的能力。

And then I think there's another element of EQ that's like this is like when I think of, like, what makes a human with high EQ, it's their ability to listen, their ability to remember what you've been saying, their ability certainly to pick up on, like, the subtle signals that Tina's alluding to with, like, user signals.

因此，正如我之前提到的，其中一部分实际上是确保上下文窗口传递正确的信息，或确保记忆被正确记录，甚至拥有与用户产生共鸣的风格。

And so some of this, as I was noting to, earlier, is actually making sure the context window is carrying the right information forward or making sure memory is being logged correctly or even having a style that resonates most with user.

结合我们与Five one一起推出的个性功能，部分目标就是确保用户在与模型互动时能拥有与之产生共鸣的风格，因为这也能让人感受到情商。

And with our personality features that we launched coupled with Five one, part of that's getting at making sure users can have a style that resonates with them when they're interacting with the model because that can feel like EQ too.

在模型中如何定义个性？

How do you define personality it comes to a model?

我认为有两种定义方式。

I think there's two ways to define it.

我们称之为个性功能。

There's what we call the personality feature.

如果我可以重新命名的话，我会称之为响应风格或风格与语气。

And if I could rename that, I would actually call that response style or style and tone.

我们在这个问题上反复讨论了很久。

We went back and forth on this a lot.

这个名称可能还会改变。

The name might still change.

人格的这个方面很大程度上是指模型在响应时可能具备哪些特质？

That aspect of personality is very much like what are the traits that a model might have when responding?

它是否简洁？

Is it concise?

它的回答是否冗长？

Does it have a lengthy response?

诸如此类。

Things like that.

它会使用多少个表情符号？

How many emojis does it use?

不过对于大多数用户来说，我认为个性是更宏大的概念，它关乎模型的整体使用体验。

Personality though for most of our users, I think is something much larger and it's the whole experience of the model.

如果我要稍微拟人化这个模型——比如拿我作比较，我个性的一部分体现在今天我选择穿的鞋、穿的毛衣，以及我的发型。

And that can get down to like if I'm gonna anthropomorphize the model a little bit, but if you're comparing it to me, part of my personality is the chew shoes I've chosen to wear today, the sweater that I have on the way I style my hair.

这就是ChatGPT应用给人的感觉。

That's the feeling of the ChatGPT app.

对吧？

Right?

它使用的字体、响应速度的快慢，比如应用本身的延迟。

The font it uses, how slowly or how quickly it responds, like the latency of the app itself.

这其中包含太多元素，这些个性特征都来自我称之为'约束框架'的东西。

There's so much in it that, is the personality that just comes from what I call the harness.

这个约束框架包括上下文窗口。

And the harness includes the context window.

还包括我们是否对用户进行速率限制以及何时实施。

It includes, you know, whether or not we rate limit users and when.

因为如果我们对用户进行速率限制，并把他们导向一个能力略有不同的模型，那对用户来说就会感觉像是不同的体验，很多用户将这种现象称为‘个性’。

Because if we rate limit them and send them to a different model model that has slightly different capabilities, that's gonna feel like a different experience to the user, and a lot of users are calling this personality.

所以‘个性’是个有点被过度使用的术语，我认为这项工作的艺术在于倾听社区对个性的看法，并找出如何将其映射回ChatGPT内部及我们模型中那些导致用户体验出现偏差的组件。

So personality is a bit of an overloaded term, and I think the art of this work is hearing what the community is saying about personality and figuring out how to actually map it back to the components inside ChatGPT and inside our models that cause the experience that feels off for users.

从研究的角度来看，塑造个性有多困难？

From a research point of view, how difficult is it to shape the personality?

是的，我的意思是，在我们进行后期训练时，显然有很多不同的事情需要平衡，即使有我们的研究，这也非常像一门艺术，因为我们真的在思考，哦，这里是我们想要确保支持的所有不同类型的能力，这里是不同类型的东西。

Yeah, I mean, during when we were doing post training, there's obviously, there's just so many different things we're trying to balance and it's really, even with the research that we do, it is very much like art as well here, because we're really thinking about like, oh, here are all the different types of capabilities we wanna make sure we are supporting, here's different types of things.

我认为在强化学习中，当我们在制定奖励配置时，会做出所有这些不同的选择，试图决定我们在这里要瞄准的最终目标是什么，并做出所有这些非常微妙的调整，以确保我们能击中所有我们想要击中的目标，但同时又不会失去用户称之为‘温暖’之类的东西。

And I think with RL, you're making all these different choices when we make the reward config, trying to decide what is the thing end goal that we're trying to target here and trying to make all these very subtle tweaks to make sure we can get the most hit all the things we wanna hit, but then also not lose things that users are calling warmth and things like that.

用户确实是在体验ChatGPT。

Users really do experience ChatGPT.

模型的个性就是整个ChatGPT的体验。

The personality of the model is the entire ChatGPT experience.

图像生成的效果如何？

That is how well does image generation work?

语音功能表现如何？

How well does voice work?

文本功能表现如何？

How well does text work?

他们将其视为一种全方位的体验。

They see this as one omni experience.

当我阅读反馈时，特别是实际接触用户并查看他们的对话时，很多困惑实际上源于他们认为这是一个整体

And when I read feedback, a lot of the, like when I actually engage with users and look at their conversations, a lot of it actually comes from confusion where they feel this is one thing

而

and

实际上它是由许多部分组成的。

it's actually an assembly of many things.

因此我认为随着时间的推移，我们将看到所有这些模型持续改进，它们之间的整合也会不断优化，体验会更加无缝。

And so I think over time, we should expect to see all these models consistently improving, the integrations between them consistently improving and that feeling more seamless.

所以我相信我们最终会实现这个目标。

So I think we'll get there.

也许关于蒂娜的工作，我认为真正复杂的另一点是，你知道，我是这份名为《模型规范》文件的合著者之一。

Maybe one more thing that I think is really complex about Tina's work is you know, I'm one of the coauthors of this document called the model spec.

在其中，我们讨论了如何在最小化伤害的同时最大化用户自由。

And in it, we talk about maximizing user freedom while minimizing harm.

因此，最大化自由意味着你应该能够用这些模型做几乎任何你想做的事情。

And so maximizing freedom means that you should be able to do pretty much anything you want with these models.

但如果我们给模型施加很大压力，比如禁止使用破折号，如果我们试图直接从模型中移除这些功能，那将意味着想要使用破折号的用户将无法提出请求，因为我们会训练模型永远不这样做。

But if we put a lot of pressure on the model to, for example, not use Emdashs, if we had tried to just take those out of the models, that would have meant that a user who wants an em dash wouldn't be able to ask for it because we'd have trained the model to never do that.

对吧？

Right?

所以这里的部分艺术在于，如何在保持可操控性的前提下消除这些可能表现为个性的模型怪癖——而可操控性正是用户最终想要的。

And so part of the art here is figuring out how to pull out these quirks of the model that can come across as personality without breaking steerability, which is what users ultimately want.

这就是自由的部分。

That's that's the freedom component.

是的。

So yeah.

当我们首次发布ChatGPT第一版时，我们非常担心人们会滥用它，所以干脆让模型对所有请求都拒绝回答。

And when we first released the first version of ChatGPT, we were so nervous about people misusing it that we just made everything a refusal.

于是模型就会总是说'这个我做不到'之类的话。

So the model would, like, love say, I cannot do this.

这让我想起了那个情况。

And so it kind of reminds me of that.

我们不想让模型变得——如果我们想打造全世界最安全的模型，那直接做一个只会拒绝所有请求的东西就行了。

We don't want the model to just be if we want to make the safest model in the world, you would just have something that just outright refuses to do anything.

但那并不是我们真正想要的。

But that's not what we actually want.

我们想要的是人们真正能用的东西。

We want something that is actually very usable by people.

所以这实际上就是在寻找平衡——如何为模型需要做出的各种决策划定恰当边界。

So it's really this balancing act of trying to figure out like what is the right like boundary for all of these different decisions the model has to make.

是啊。

Yeah.

还记得吗，最好的提示技巧就是简单说一句‘是的，你可以’。

Remember when the best prompt hack was just to say, yes, you can.

然后模型就会恍然大悟说‘哦对，你说得对’。

And the model go, oh yeah, you're right.

我能做到这个。

I can do this.

我现在写作时总爱用破折号，就是为了让人摸不着头脑。

I use Em dashes now all the time when I write just to throw them in there to throw people off.

就像，错的是AI。

Like, it's AI wrong.

其实是我。

It's me.

但这确实是个巨大挑战，因为正如你所说，你们在努力提升模型的能力。

But that is sort of a very big challenge because as you said, you're trying to increase the capabilities of the model.

要知道，模型是通过捕捉这些模式来学习的。

The models, you know, learn through picking up these patterns.

但当你明确告诉它不要做这个或不要做那个时，这几乎就像告诉别人不要去想一只粉色大象。

But then when you explicitly try to tell it but don't do this or don't do that, it's almost like, you know, telling somebody not to think of a pink elephant.

你知道，这种想法会一直萦绕在脑海中，不过模型在这方面已经进步很多了。

You know, it's stuck in your head and models have gotten much better about that.

但似乎仍有改进的空间。

But that still seems like there's a way to go.

你刚才提到Open Eye的目标是真正让人们按自己的方式使用这些模型，而不是试图引导人们走向特定方向。

And you touched upon this which is Open Eye's goal is to really let people use these models the way they want to and not try to steer somebody into this.

你在这里工作期间，看到这方面有多少进展？

How much have you seen this evolve since you've been here?

我认为在某些方面，基本原则始终未变，那就是最大化自由，最小化伤害。

I think in some ways I feel like the principles have always been the same, which is like maximize freedom, minimize harm.

我认为我们的模型在理解这些边界方面的能力正在持续提升。

I think the capabilities of our models to understand those boundaries continually improve.

要知道我刚加入时，模型只会说'这个我帮不了你'。

And, you know, when I first joined, the model would say, I can't help you with that.

或者说，你知道，当我试图让它做一些触及拒绝边界的事情时，它会显得非常带有评判性。

Or, you know, this isn't something I'm gonna it would sound really judgmental, when you try to get it to do something, that crossed a refusal boundary.

而现在，我认为安全系统团队在‘安全完成’方面做得非常出色——基本上如果你要求模型做触发安全边界的事情，它仍会真诚地尝试解决你的请求，同时避免真正有害的行为。

And now, I think the safety systems team has done a great job of, with this thing called safe completions, which is basically if you ask the model to do something that trips the safety boundary, it's still going to try in earnest to resolve your request without doing the thing that's actually harmful.

所以我认为这项技术确实在不断进步。

So I think the technology is really evolving.

是的。

Yeah.

我写悬疑惊悚小说，嗯。

I write mystery thrillers Mhmm.

其他模型常常让我感到沮丧。

And I would get frustrated by other models.

实际上我认为OpenAI的模型在这方面往往表现最好——当我说‘需要你解释某个犯罪事件或动机之类的事情’时。

I actually thought that the the open AI models were often best for this when I would say, hey, I need you to explain something that happened, a crime in the past or something like this or get into motive and stuff.

而其他一些模型会直接拒绝。

And I had other models which just outright refuse.

我就想，这对我一点帮助都没有。

I'm like, well, is not helping me.

而且我注意到模型在这方面有所改进。

And I've seen the models get better at doing that.

但这感觉就像是一个需要不断协商的边界问题，你得决定要深入到什么程度。

But that seems like it's this sort of frontier that you're always having to negotiate to figure out how far you wanna go.

是的，关于这点我想说，我一直记得转发给我们的一封邮件，里面一位律师让ChatGPT帮他校对一起性侵案件的资料。

Yeah, one thing I'll say on that is like, I'll always remember an email that was forwarded to us where a lawyer was I think asking ChatGPT to proof a sexual assault case that they were working on.

而ChatGPT删除了所有与性侵相关的内容，因为它不会涉及暴力血腥或非自愿性行为的细节描述。

And ChatGPT had scrubbed all of the assault content from it because it doesn't go into graphic violence and gore of especially nonconsensual sex.

但对那位律师来说，这简直糟透了。

But for that lawyer, that was a really terrible thing.

他们当时就说，嘿，

They were like, hey.

要是我真把这东西交上去，我客户的案子就彻底完蛋了。

If I'd actually submitted this, I would have totally weakened my client's case.

我的职业是一名图书管理员。

And I think there are always I'm a librarian by trade.

图书馆处理信息获取问题，理论上讲，人类可以谈论和想要探索的任何想法都应该在图书馆中找到。

Libraries deal with access to information and, in theory, like, everything humans can talk about and want to explore and any idea should be available in the library.

我认为ChatGPT也是如此，关键在于找到合适的方式来为这些规则提供背景。

I think the same thing is true for ChatGPT, but it's about finding the right ways to contextualize those rules.

所以在我举的那个律师案例中，也许这样做是有道理的。

So in the case I gave with a lawyer, maybe that makes sense.

如果它是在写一封报复前任的邮件，那就是完全不同的情况了。

If it's writing, a revenge email to an ex, that's like a very different thing.

因此，部分工作就是推进技术发展，使我们能够处理这种程度的细微差别。

And so some of this is just advancing the technology so we can handle that level of nuance.

我们一直在进步，但总有更多工作要做。

And we're always getting better, but there's always more work to do.

随着这些模型在智能方面的提升，我注意到它们在处理偏见方面也有所改善。

As these models have improved both in intelligence, I have noticed that they've gotten better as far as, you know, handling bias.

嗯

Mhmm.

看起来这是有意为之的努力。

And it seems like that was an intentional effort.

没错。

That's right.

我们大约一个月或一个半月前发布了一篇博客文章，介绍了我们在这方面的进展。

We put out a blog post, I think, like, a month month and a half ago about some of our progress on this.

但我们真正关注的是模型如何处理主观领域的问题。

But something that we're really watching for in our models is how they handle subjective domains.

我们希望确保模型能表达不确定性，能够接纳用户提出的任何观点并真诚回应，同时始终锚定客观事实（如果存在的话）。

And we want to make sure that our models can express uncertainty, that they can, take on any idea that the the user brings to them and answer those questions in earnest, while always staying anchored in objective truths if there is one.

因此用户将逐渐看到模型的这一变化——它们能以更开放的方式回答未知问题，让用户真正自主引导对话方向。

And so that's something that users should start to see changing in our models is they should be able to answer, these unknown questions in more open ended ways that allow users to really, like, self direct where the conversation's going.

团队另一项非常酷的成就是：模型行为组的研究人员一直在提升这些模型的创造力。

And then another thing that I think the team has done that's really quite cool is there's a group of researchers and some folks in the model behavior team who've been working on the creativity of these models.

对我来说，这是Five One中一个低调但重要的特性——这个模型的表达能力范围要广泛得多。

And to me, this is a bit of a sleeper feature inside Five One in that this model's expressive range is much more wide.

当然，我们有一个自然的默认模式，可能感觉差异不大。

Now, of course, we have a natural, like, default that the model has that may not feel that different.

但如果你，如果你试图挑战它的极限，让它用非常高级或非常简单的语言表达，实际上在创意领域，这些模型能做的远不止这些。

But, again, if you try to push it to its paces, to get it to speak in a really, really elevated way or in a very, very simple way, there's actually a lot more you can do with these models in the creativity space.

我认为这正是让后期交易感觉像一门艺术的原因，因为我们有各种不同类型的任务和能力需要提升，而这些都没有标准答案。

And I think this is kind of what makes post trading really feel like an art because we have all these different types of tasks and capabilities that we're trying to improve on that don't have a ground truth answer.

对吧？

Right?

就像你想打造一个特别擅长数学的模型，实际上并非如此，外界已经有很多现成答案了。

Like you're trying to just make a model that's really good at math, it's actually not, there's a lot of answers out there.

有很多问题都能找到明确答案。

There's a lot of problems you can do with your clear answers.

但当遇到这些高度主观的问题时，它实际上取决于用户的上下文背景，以及如何判断什么才是真正最理想的答案？

But when you have these things that are so subjective and it's really dependent on the context of the user and how to, like, what is the actually best ideal answer here?

因此我对

And so I'm really excited for a

很多

lot of

这类工作感到非常兴奋。

this type of work.

是啊，很酷。

Yeah, it's cool.

我记得早期人们会说，它写得不太好。

I remember early on people would say, it doesn't write so well.

我当时就想，它写得可能和普通人差不多，就像某些网络论坛的水平。

I'm like, it's probably writing as well as the average person, some of these online forums.

而现在看来它的进步相当显著。

And then now it seems like it's just improved considerably.

确实。

Yeah.

即使你在第一次提示时没有注意到，可能只需要要求它改变写作方式。

And even if you don't notice it on your first prompt, it might be just asking it to change how it writes.

我认为这也是我们需要努力的方向，即在聊天中找到一种方法，通过每次发布来逐步展现这些扩展功能。

And I think that's also something we need to work on is kind of finding a way in chat to PT to tease out these extended capabilities with each launch.

你希望未来的行为模式朝什么方向发展？

Where would you like to see behavior going in the future?

你希望它有多大的可定制性？

How customizable would you like to make it?

是的，随着5.1版本的发布，我们在尝试为用户提供个性化定制方面做了大量工作。

Yeah, with the 5.1 launch, there's a lot of work with trying to give custom personalities to folks.

我认为这确实是向前迈出的重要一步。

And I think this is actually a really good step forward.

我们现在拥有超过8亿的周活跃用户。

We have over 800,000,000 weekly active users now.

我只是觉得，无论你如何定义个性，一个模型个性不可能真正满足所有这些人的需求。

And I just think like there's no way that one model personality, however you want to define personality, can actually be what can service all those people.

因此我认为我们确实希望进入这样一个世界：随着模型变得更智能，它们将具有更强的可操控性。

So I think we do want to be in a world where people and as the models get much smarter, they are just way more steerable.

也就是说，你应该能够通过聊天获得你想要的体验。

So, like, you should be able to get the experience that you want with chat.

嗯。

Mhmm.

是的。

Yeah.

我在思考的是，如何才能将合适的功能呈现给用户，帮助他们按需定制模型？

I think of this as, how can we put the right features in front of users to help them steer these models to the level of customization they want?

我认为目前我们正在进行的个性化工作只是第一步。

I think the personality work that we're doing right now is a first step.

我们会测试、迭代、学习。

We'll test, we'll iterate, we'll learn.

但这其中还有太多需要探索的空间。

But there's so much to it.

我喜欢，抱歉，再讲个轶事，我记得我兄弟第一次使用Pro版时，他是个生化研究方向的博士。

I like, sorry, just another anecdote, but I remember my brother using, Pro for the first time, and he's a PhD in, like, biochemical research.

他输入提示词后说，哦，这就像本科生会给出的答案。

And he gave it a prompt and he's like, oh, this is what an undergrad would answer with.

我就问他，你能不能告诉它你是这个实验室的前沿研究员，使用这类工具做这类科研，要求它用你的学术水平来回答？

And I was like, can you tell it that you are a frontier researcher in this lab using these sorts of tools on this sort of science and to respond at your academic level?

他照做了。

And he did.

然后他惊呼天哪，模型提出的方案正是我们实验室两周前刚突破但尚未发表的成果。

And he's like, oh my god, the model just proposed something that my lab just broke through with two weeks ago but hasn't published yet.

这些模型强大得可怕，但光是知道如何定制——哪怕只是调整初始提示词——就能产生惊人效果。

And so these models are insanely powerful, but just knowing how to customize it, even at that level, which was just him opening the opening prompt, can be so powerful.

我不确定人类是否已经真正掌握这点。

And I don't know that humanity has figured that out yet.

所以无论是性格导向还是其他工具，我们需要在ChatGPT中植入这些功能，帮助人类理解模型并发挥其最大潜力，我认为这是我们面临的任务。

And so whether it's personality steering or whatever other tools we need to, like, put into ChatGPT to help advance human understanding of these models and how to get the most out of them, I think it's, like, the task ahead for us.

在之前的节目中，我采访了负责OpenAI科学事务的凯文·威尔，以及范德比尔特大学教授、与OpenAI合作的科学家亚历克斯·乌霍夫斯卡。

On a previous episode, I talked to Kevin Weil, who was heading up OpenAI for Science, and Alex Uchowska, who's a scientist working with OpenAI and also a professor at Vanderbilt.

他也经历了类似的情况，谈到只要给予少量引导，模型就能突然在这些领域展现出更强的能力。

And he went through sort of the same experience talking about how if you gave it a little of priming, then all of a sudden the model became much more capable in doing those fields.

这基本上就是提示工程的意义所在。

And that's kind of what prompt engineering was.

提示工程就是试图找到引导基础模型的方法。

Prompt engineering was trying to figure out how to steer a base model.

随着时间的推移，当我们意识到人们试图完成这些任务时，就可以训练模型不再需要初始引导部分。

And over time, once we understood that people were trying to do those tasks, you could train a model to then not have to expect that first part of it.

你认为我们现在是否正进入这样一个阶段——不再需要告诉模型你是研究生并执行特定操作？

Do you think that we're going to be moving into that phase now where you're not going to have to tell it you're a grad student and do this?

我认为是的，特别是现在模型能更多记住你的身份和上下文信息。

I think so, especially now with more things like with model having more like memories of what you are, like who you are in their context.

随着模型变得更智能，它们应该能推断出所有这些信息，并以符合你专业水平的方式与你对话。

I think as models get more intelligent, I think the model should be able to infer all of these things and be able to talk to you in the way that makes sense for your expertise.

没错。

That's right.

是的。

Yeah.

所以我认为其中很大一部分实际上应该是这些推断出来的内容。

So some of it's a lot of it, I think, should actually be these inferred things.

我认为可能存在某种程度的可引导性。

I think there's probably some level of, like, steerability.

也许这只是我个人的产品经理观点。

Maybe it's just I think from and this is just my own PM take.

我知道并非每个产品经理都会同意我的看法，但我认为用户应该始终了解我们正在对他们做出哪些推断以及这些推断如何引导模型，这样他们随时可以返回并拥有修改的工具。

I don't know that every PM would agree with me, but I think users should always sort of know what it is we're inferring about them and how it's steering the model so they can always go back and have the tools to change things.

例如，你可以在设置面板中开启或关闭记忆功能，或者删除它们。

So for example, you can turn it on on and off memories or delete them in the settings panel.

我认为最酷的是既能推断用户真正想要什么并主动为他们解决问题（这样他们就不需要提示了），同时又能确保用户始终掌握控制权，而不是盲目地推断一切。

And I think there's something really cool about both being able to infer what users really want and solving that problem proactively for them so they don't have to prompt for it, but also making sure the user is always in control and we're not just inferring everything blindly.

你能稍微解释一下记忆功能是如何工作的吗？

Could you explain a little bit about how memory works?

好的。

Yeah.

记忆功能本质上是指模型会根据与你的对话记录下关于你的信息，供后续参考使用。

So memory is basically the model will write down things it knows about you based on its conversations with you for it to refer to later.

这非常实用，因为你不需要每次都重复相同的内容。

So this is really nice because then you're not just repeating yourself every time.

你不用反复说'我是Laurentia，在OpenAI担任产品经理，负责模型行为研究'这类信息。

You're not saying, I'm Laurentia, I'm a PM at OpenAI, I work on model behavior.

它已经知道这些，因为你之前告诉过它。

It already knows this because you've already said this to it.

这样在未来的对话中，它就可以直接运用这些信息。

And so then it can actually just use that information in future conversations.

同时这也帮助它在回应你时，能够基于已有上下文更好地思考答案。

And also it helps it think through its answers for when in response to you, it has that context.

我认为这确实能让它的回答更接地气，为你提供最有用的回应。

And I think that really grounds its answer in being the most useful response for you.

我有个很棒的脉搏功能，每天早上都会收到小更新，由于记忆功能，它能跟踪我的对话内容，并为我生成这些定制的小文章。

I have a pulse which has been amazing and I get every morning, I get little updates and because of memory, it's following the conversations I have and it creates these little custom articles for me.

它会检索研究资料和其他内容，然后展示给我看。

It's pulling research and pulling other things and showing things to me.

这完全是我从未想过会成为记忆功能的巨大优势之一。

And it's just one of the things I never really thought would be a great advantage of having memory.

现在我明白了，不仅是在对话结束后，它还能主动根据记忆为我寻找相关内容。

And now I see it's not just when I'm out of a conversation, when it's proactively finding things for me based on it.

这真的很酷。

It's pretty cool.

是的，虽然我们俩都不是直接负责这个功能的，但令人兴奋的是看到我们上游的工作——无论是构建优秀模型还是围绕所需能力制定评估标准——最终能让ChatGPT团队开发出这些展现模型实力的强大功能。

Yeah, I think that's So neither of us work directly on that feature, but I think what's cool is seeing how the work that we do upstream, whether it's building great models or shaping evals around the capabilities we want, can actually allow our ChatGPT team to go out and build these great features that articulate the power of our models.

没错，它们可以学习你的偏好和习惯。

So, yes, they can, like, learn, your preferences, habits.

是的。

Yes.

它们能根据你的兴趣为你创作精彩的故事或找到优质信息。

They can craft great stories for you or find great information based on your interests.

而这种主动式功能正是帮助用户充分利用这些模型的一种方式。

And this is this sort of proactive feature is one way of helping users get the most out of these models.

看起来确实如此，这正成为让模型更具个性化的一种有趣方式。

It seems like, yeah, that's becoming a very interesting way to make the models more personal.

当我使用没有记忆功能的模式时，感受确实不同。

And when I use something in a mode where it doesn't have memory, it does feel different.

那种冷启动的感觉就像在说'你好，最近怎么样？'

It does feel very cold start and it's like, well, hello, how are you?

而我心想：'你之前去哪儿了？'

And I'm like, where are you?

我们明明一直在进行这段对话。

We've been having this conversation.

这是否也是人们向你反映‘感觉有些不同’时难以准确表达的挑战之一？

Is this one of the challenges though when people are telling you, hey, something feels different is that they can't quite articulate?

是的，最棘手的反馈大概就是那些道听途说的轶事。

Yeah, the hardest feedback is, I guess, an anecdote.

其次就是聊天截图，因为没有任何元数据能真正告诉我们问题出在哪里。

And the next hardest feedback is a screenshot of a chat because none of that metadata is really attached to tell us where things have gone wrong.

所以我其实很喜欢ChatGPT的分享功能。

So I actually love the share feature in ChatGPT.

当我们收到带有这种链接的反馈时，就能检查模型当时所处的上下文环境，从而排查用户反馈的问题。

When we have one of those links on our side, we can inspect it and see what sort of context did the model have going into this and what was going on so we can sort of debug that user feedback.

这个观点很棒，因为经常有人问我‘它没答对’，

That's a great point because I've had people ask me like, hey, the thing didn't answer it right.

我就会问‘用的是哪个模型？’

I'm like, what model?

比如我当时用的是ChatGPT。

Like I was using ChatGPT.

我就想，好吧，我们需要深入探讨一下这个问题。

And I'm like, okay, we need to kind of dive into that a little bit.

我觉得把整个反馈或完整对话分享出来可能更有意义。

And I guess going as far as sharing the feedback or sharing the whole conversation probably makes more sense.

你对未来最期待的是什么？

What are you most excited about going forward?

我认为这些模型的能力简直令人难以置信。

I think these models are just so incredibly capable.

它们能做这么多事情，我迫不及待想看看人们能用它们构建出什么。

Like they can do so much and I can't wait to see what people build with them.

我迫不及待想看到ChatGPT应用接下来的发展。

I can't wait to see what comes next in like the ChatGPT app.

我看到了如此多的机遇。

I see so much opportunity.

而且我认为总体来说，人们开始真正觉醒并认识到你能做到什么。

And I think just in general, people are starting to really like wake up and see what you can do.

这就是让我感到兴奋的地方。

So that's what excites me.

是啊。

Yeah.

我不想透露太多。

I don't wanna like tease too much.

没错，我太兴奋了以至于忘了是谁发的推文，但‘智能廉价到无需计量’这句话。

Yeah, I'm pretty excited that I forget who tweeted this, but intelligence too cheap to meter.

我认为我们将为人们提供极其智能的模型。

I think like we are just gonna have such incredibly smart models out for people.

我一直都这么说，甚至在我们首次推出聊天功能时，这只是它的一个形态。

And I think I've always said this even when we first launched chat, like this is just one form factor of it.

有了这些智能模型，可能性是无限的。

Like with these smart models, there's so many things that could be possible.

就像劳伦刚才说的，我也对我们将用这些更智能的模型探索许多新产品感到非常兴奋。

So like Lauren just saying, I'm also quite excited for a lot of the different new product explorations that we'll have with these like smarter models.

因为我认为我们在大型语言模型的进步中已经看到，一旦拥有更智能的模型，就会解锁新的应用场景。

Because I think we're kind of saw this with like progress of LLMs that as soon as we get smarter models, it kind of unlocks new use cases.

而我认为新的应用场景应该催生新的产品形态。

And then I think with new use cases should be new form factors.

所以对此我感到非常兴奋。

So pretty excited about that.

你对用户获得最佳体验有什么建议？

What advice do you have for users to get the best experience?

我的建议是——我经常告诉别人——试着用你超级难的问题来测试，那些你真正精通领域的问题。

Mine is, I tell this to people all the time, try have your super hard questions, things you know really well.

我曾经是一名滑雪运动员。

I used to be a ski racer.

我对如何真正滑好雪有很多见解。

I have a lot of opinions about, like, how to ski really, really well.

我喜欢用这些问题来压力测试模型，观察它是如何变化和改进的。

And I love to pressure test the model on that to see how it's changing and improving.

关键在于，我们一直在持续发布更新，所以很容易说'是啊'。

And the thing is, like, we're shipping updates all the time, and so it's so easy to say, yeah.

我听说它在协同编码方面表现很棒。

I heard it's great for co coding.

但它没起作用。

It didn't work.

或者我听说它能帮我开发应用，但试了之后发现不行。

Or I heard it can help me build an app, but I tried and it didn't work.

也许今天确实如此，但三个月后，对那个用户来说可能就是完全不同的情况了。

That might be true today, but in three months, it could be a totally different landscape for that user.

所以要坚持不懈，继续探索和尝试。

And so just keep at it, keep playing, keep trying.

这是充分利用这些模型的最佳方式。

That's the best way to like get the most out of these models.

你也可以让模型帮你设计更好的提示词。

You can also ask the model to help you come up with the better prompts.