AI 应如何处理新闻、政治、医疗和心理健康——与 Campbell Brown 对话

但这几乎就像你在努力找到最好的人，和我们合作的一些最优秀的人以前是中情局的分析师。

But it's it's almost like you're trying to get the bet some of the best people who who work with us are former CIA analyst.

对吧？

Right?

因为他们必须努力摆脱所有偏见，看到整个大局。

Because they have to try to get rid of all their bias and see the whole big picture Mhmm.

以及所有可能性。

And all the possibilities.

就像一位好律师也会做同样的事情。

And it's like a good lawyer would do the same thing.

对吧？

Right?

而这正是你试图达到的目的。

And and that's what you're trying to get at.

这更像是回答那个问题的正确框架，而不是一个可能没有正确答案的问题的正确答案。

It's like the right framework for how to answer that question as opposed to a right answer for for a question that might not have one.

是的。

Yep.

好的。

Okay.

但你的工作非常艰难。

But you have a very tough job.

而且我，我再次强调，

And I don't I I also again,

这，这是关于人的，我，我得承认。

this is this is people I mean, I I will I will confess.

我的投资者们说，我们为什么要从政治开始？

My investors were like, why are we starting with politics?

我们为什么要自找麻烦？

Why are we breaking ourselves into jail?

有那么多其他事情可以做。

There's so many other things.

确实不容易。

Can be doing

我明白。

I know.

我这就得问个挺敏感的问题了，所以……行吧，咱们继续。

I'm like, I'm about to step on it step in it with the question I'm about to ask you, so I But see like, alright.

好，那就这样。

Let's say alright.

你在评论文章里提到一点，就是你可以问模型：某种非处方药在孕期服用是否安全，以及它是否会像医生那样给出必要的注意事项。

One of the things you write about in your op ed is you should be able to ask a model whether a specific over the counter medication is safe to take during pregnancy, and you can learn whether it offers the caveats a doctor would insist on.

好的，主流观点里有种说法，比如怀孕期间吃泰诺会导致孩子得自闭症。

Okay, in the mainstream, there are views that, like, taking Tylenol during pregnancy will cause autism in a kid.

你知不知道，我之前也问过类似的聊天机器人，比如疫苗会不会导致自闭症。

And do you do you you know, actually, asked, I asked a version of this to chatbots earlier about whether, like, vaccines cause autism.

是的。

Yeah.

然后它回答得挺明确的，说不会。

And it answered, like, fairly definitively, they do not.

但当你把这些专家聚在一起，尤其是如果他们跟某种政治意识形态挂钩，就会有一群人跳出来说会。

But when you gather these experts together, you're gonna have a a sort of especially if they're aligned with, like, a certain political ideology, a group of people will be like they do.

所以现在你就得把这种可能性引入到模型里。

So now you have to kind of introduce that possibility into a model.

但你提的这个，对我来说其实是背景信息。

But that's around but that what you're raising is context to me.

对吧？

Right?

有些事情，我，我是相信科学的。

There are certain things, and and I I believe in science.

好。

Okay.

说完这话我自己都不太信，但我确实信。

Can't quite believe it after saying that, but I do.

好的。

Okay.

所以，如果有明确的——你知道，我觉得你得看看。

So if there if there's a clear you know, I think you have to look.

顺便说一句，你不一定非得这么做，但我的选择是，当存在一个真相时，模型应该努力去接近它，而这必须是你试图实现的原则之一。

And by the way, you don't have to, but my choice would be that the models would try to get to the truth when there is one, and that that has to be one of the principles that you're trying to achieve.

对我们公司来说，这就是我们的原则之一。

And for us as a company, that is one of our principles.

如果有明确的证据，我们就引用那个明确的证据。

If there is if there's clear evidence, we cite that clear evidence.

但你刚才说的其实是背景信息，而且是很重要的背景，因为这是当前这个国家正在进行的对话的一部分，因为有些政治观点很强的人，他们不同意科学可能给出的结论。

But what you're talking about is context, and it's important context because it's part of the conversation that is happening in this country right now because there are people with strong political views who who disagree with what the science may say.

我觉得你得解释一下为什么会出现这个问题，也就是这是这个国家对话的一部分，然后让我们给你提供背景。

And I think you have to explain the context by which this is even coming up, which is it's part of the conversation in this country, and let us give you the context.

这和提供正确信息一样关键。

And that is as critical as having the right information.

这就是难点所在。

And this is what's hard.

比如说，什么是上下文？

Like, get like, what is context?

它是什么样的？

What does that look like?

需要呈现的细微差别是什么？

What is the nuance that needs to be represented?

这可不是靠成千上万个数据标注员就能做到的。

And that's not something that you can do with, you know, thousands of data labelers.

你需要的是

That you need

最优秀的人才。

The best people.

你能找到的最优秀的人才。

The best possible people you can find.

顺便说一句，这也是为什么我对实验室自己做评估有很多疑问，因为很长一段时间里，并没有真正的专家参与进来。

And by the way, this is why this is partly why I'm very, I have lots of questions around how the labs do their own evals because for a long time, there hasn't been real expertise brought in.

对。

Mhmm.

即使是世界上最聪明的工程师，也无法给你提供关于复杂政治问题的背景信息，或者告诉你如何获取这些背景。

And even the most brilliant engineer in the world is not gonna be able to give you context or tell you how to get to context around a complicated political issue.

他们就是做不到。

They're just not.

那不是他们的专长。

That's not what their expertise is.

所以我认为实验室现在开始与在这些高风险领域有真正专业知识的人合作，这在我看来是至关重要的，而这也是我们最初讨论的起点。

So I I think the labs are now beginning to work with people who have real expertise in these high stakes areas, which I think is essential, and that's, you know, where we started.

是的。

Yeah.

我也不知道为什么我在这里质疑专业知识，但让我至少抛给你一个问题，你可以思考一下。我们之前谈过疫苗的事情，对吧？

I I don't know why I'm kind of going to, like, questioning expertise here, but let me at least, like, throw it out at you and you can sort of So we talked about the vaccine thing, right?

所以，你知道，我觉得左派可能会说，别把自闭症那部分加进去，但很多右派的人不这么认为。

So it's like, you know, I think the left would probably say, you know, don't include the autism stuff, many on the right.

其实不是，现在这已经是两党共识了。

Actually, no, it's actually now bipartisan.

我刚想说这个……

I was gonna say that that

这是两党共识。

That's bipartisan.

对，是两党共识。

That's bipartisan.

好。

Okay.

但还有一点，比如在新冠疫情最严重的时候，你可以去问最顶尖的专家。

But then there's also like, all right, you could ask the best you could have asked in the middle of COVID.

你可以问他们，比如，新冠疫情到底是从哪里开始的？

You could have asked, like, the best experts, you know, where did COVID start?

就像这样，回到那个“相信科学”的话题。

Like this, going back to this believe in science thing.

美国顶尖的科学家们当时会说，新冠病毒绝对不可能是在武汉实验室里出来的。

And leading scientists in The United States would be like, there's no way that that COVID started in the Wuhan lab.

但现在呢，我也不知道，大家的共识好像是，不能排除这种可能性。

And now, I don't know, the consensus is, like, that you can't rule that possibility out.

所以当有人问这种问题的时候，你到底能在多大程度上依赖专家呢？

So how much can you, like, rely on experts for when someone's going and asking a question like that?

我觉得，我的意思是，这是一个非常……这是……

I I thought I mean, this is a very This is

问题所在。

the problem.

对吧？

Right?

这是一个合理的问题。

This is a legitimate question.

对。

Yep.

有本书。

It's there's a book.

其实我前不久刚写过相关的东西，因为有一本书。

I actually wrote something about this not too long ago because, there's a book.

我现在一时想不起作者的名字，但那本《专业知识的终结》讲的就是，新冠就是一个很好的例子，说明专家们在重要问题上反复出错。

I'm blanking on the author's name right now, but the end of expertise that talked about how and I COVID's a perfect example of how experts have have gotten it wrong repeatedly on important issues.

而且，我觉得还有那种随之而来的精英主义感。

And, also, I think the sense of elitism that comes with that.

没错。

Right.

就是，你从某个大学拿了学位，因此你就有一定的发言权，但人们真的应该质疑这一点。

What what you have a degree from some certain university, and therefore, you have a certain amount of authority to speak on this that people genuinely should question.

但我觉得，当涉及到医学、心理健康这类问题时，归根结底，我希望评估这些结果的人是那种和病人面对面待过成百上千次的临床医生。

But I think when it comes to questions around medicine, around mental health, at the end of the day, the person I want evaluating those outputs is a clinician who has been in the room with the patient hundreds and hundreds of times Mhmm.

而且能把握那些关键的细微差别。

And has a sense for the nuance that is critical.

我也不知道还有什么别的办法。

And I don't know another way to do it.

如果有人有更好的主意，我完全支持。

And if somebody else has a better idea, I'm all for it.

但我觉得，凡是涉及生死攸关、像我们正在处理的某些话题那样高风险的事情，我并不是说政治容易，也不是说我的方法就是对的，但这是目前我认为最好的方法。

But I think on anything that is life or death that is as high stakes as some of the topics that we're trying to deal with, and I'm I'm not pretending politics is easy by any means or that my approach is the right one, but it's the best one I think we have at the moment.

而且我觉得我们必须依靠这一点，认识到有些人在某些领域比其他人更有经验，他们研究这些问题多年，现在正是需要他们来帮我们解决这些挑战的时候。

And I think we have to lean into it and recognize that there are people who have more experience in certain areas than others, who have studied these issues for years, and they are needed right now to help us figure out some of these challenges.

所以，把专家排除在外是不可行的。

And so writing off experts is not an option.

我觉得你——

I think you're

顺便说一句，我并不是反对这个。

I'm not for that, by the way.

不不不。

No, no, no.

但我是说，这只是一个

But I'm just it's a

确实，这里有很多棘手的地方，尤其是回到我们讨论的核心，那就是这些实体似乎将成为世界上最受信任的实体。

legitimate many point to tricky areas here, especially, again, going back to our core of our discussion, which is, like, these are just going to be the most trusted entities in the world, it seems like.

所以，这会很有趣，因为围绕这个有很多政治问题。

So it I it'll be interesting to see because of all of the political issues around this.

你这么说，但我也想说，看看数据中心正在发生什么。

Just you say that, but I also I mean, look at what's happening with data centers.

这已经成为一个巨大的政治话题。

This has become a huge political topic.

AI，我想，抱歉。

AI, I think sorry.

抱歉打断你。

Sorry to interrupt you.

不是。

No.

不对。

No.

人工智能，AI行业现在很不受欢迎。

AI, the AI industry is very unpopular.

确实如此。

Yes.

用这些聊天机器人的人都很喜欢它们。

People who use these chatbots love them.

对。

Yes.

所以这里存在一种分歧

So there's a diversion

大家看法不一致。

There's a disconnect.

就像马克·扎克伯格不太受欢迎，但用Instagram的人喜欢它一样。

Similar to, like, Mark Zuckerberg is not very popular, but, like, people who use Instagram like it.

对。

Right.

对。

Right.

我觉得这么说挺公平的。

I think that's fair.

我觉得这里确实有脱节。

I think there's a disconnect there.

我觉得人们也在试图理解，他们从硅谷听到的信息和他们的实际体验之间也有差距，比如，哦，AI会改变世界。

I think people are also trying there's a disconnect around what they're hearing from Silicon Valley, the message they're hearing from Silicon Valley, and their own experience too with AI, which is, oh, AI is gonna change the world.

这是有史以来最伟大的事情，它会治愈癌症，是的。

It's the greatest thing that's ever happened, and it's gonna cure cancer, and Yeah.

每个人都会失业。

Everyone's gonna lose their job.

然后他们就开始用那些聊天机器人了。

And these, like, very but then they're using their chatbots.

他们挺喜欢这些聊天机器人的。

They love their chatbots.

确实是这样。

Sure.

但他们心里也清楚，这些聊天机器人并不是硅谷向我们承诺的那种救世主。

But they also know that these chatbots are not the savior that, you know, is being promised to us by Silicon Valley.

他们知道自己会犯错。

They know they make mistakes.

他们清楚普通用户是怎么用这些工具的，就像你我一样，大家都在用，而且我们都明白这项技术还没成熟。

They know who normal people are using them like you and I are like, everybody's using them, and we all know that the technology isn't there yet.

所以我不确定我们会走到那一步。

So I I'm not sure that that's where we end up.

就是大家都比什么都更信任聊天机器人。

That that we're all trusting our chatbots over everything else.

嗯，我觉得这要从学生开始说起。

Well, I think it starts maybe it starts with the students.

我觉得我这么说有点夸张。

And I think I'm being a little hyperbolic here.

我承认。

I'll admit it.

但确实，这要从学生开始。

But, like, it starts with the students.

这样播客听起来更有意思。

It's good for a podcast.

对。

Yes.

能吸引听众。

Good for engagement.

天哪。

Oh, god.

我们确实需要一个独立的评级机构来给我们提供一些比我正在优化的更好的基准。

We we do need an independent rating agency to come in and give us some better benchmarks than the ones I'm optimizing for.

不。

No.

我开玩笑的。

I'm kidding.

但是，布朗大学的学生，对吧？

But, like, the Brown students, right?

你看到布朗大学那些孩子的故事了吗？

Do you see this, story of kids at Brown University?

他们用AI参加了期中考试。

They, they used AI for the midterms.

很多人考得很好。

Many of them aced it.

然后教授说，亲自来考。

Then the professor said, come in person.

大家都考砸了。

Everybody did poorly.

也不是所有人都这样。

Not everyone.

有个学生两次都拿了95分，期中那次大家基本都满分，他（她）却远低于平均分；但到了期末考试，尽管平时表现很稳定，成绩却远远高于平均水平。

There was one kid that got 95 on both, and he scored or she scored way below the average on the midterm where everybody got a 100 and then way above the final, even though they were consistent.

这事儿得从学校说起，因为大家一看这东西能帮忙写论文，然后你猜怎么着？

But it starts in school because people will see this stuff can write my paper, and then you know what?

他们要生孩子了，结果突然之间就开始用AI来养孩子，就像当初用AI应付学业一样。

They're having a kid, and all of a sudden, they're, like, using AI to raise their kid the same way that they used it to, like, get through school.

我完全同意你的看法。

I'm I don't disagree with you.

正因为这样，我才决定一头扎进去，因为我确实觉得这事很重要。

I that's why I'm diving in head first because I do think it's important.

我是真的没法反驳你。

I really I can't disagree with you.

你看，

I look.

现在才刚开始。

We're it's early days.

对。

Mhmm.

真的还太早了。

It's it's so early days.

没错。

Right.

我们正在解决这些问题，而且我确实认为教育以及AI在高等教育甚至小学中的应用，

And we're working through these problems, and I I do think education and how AI is used in in not just higher ed, but elementary school.

我们如何制定相关政策，以及如何教会孩子们AI的好处和坏处，这些都是我们在未来几年会逐渐了解的事情。

How how what policies we develop around that and what we learn about how to teach kids the benefits and the the both the pros and cons of it are are things that we will learn over the next couple of years.

但是，是的。

But I yeah.

归根结底，我觉得结果会像你所说的那样，这也是我认为它至关重要的原因。

At the end of the day, I think we end up where you think we end up, and that's why I think it matters so much.

对。

Yeah.

不是的。

No.

所以我刚才花了大概十分钟，反复推敲用专家介入这个想法。但我认为，随着大家越来越信任这些机器人，对我自己以及其他人来说，虽然总会有一些例外情况，但遵循规则这一点真的很重要。

So I've spent, like, the last ten minutes or so pressure testing, like, the idea of using experts, but but I I think that, you know, as we start moving into this direction where people trust, trust these bots more and more, you know, I I would recommend, like, for myself and and everybody else, there are going to be these exceptions to the rule, but the rule really matters.

而且你说得特别对。

And that is, like, you brought up such a good point.

比如在心理健康领域，你是想通过微调一堆回答来优化模型呢？

Like, when in a mental health situation, do you want to engineer, you know, sort of doing fine tuning on a bunch of answers?

还是去找全球最顶尖的心理医生，问问他们面对自残行为时该怎么处理，用这些专家的建议作为提示词，然后据此进行强化学习？

Or do you want to, like, go to the best mental health clinicians in the world and then ask them how the botch should treat self harm, you know, prompts and then reinforce based off of what those experts say.

我肯定更希望由心理健康专家来给出指导。

I would much rather the mental health experts have to say here.

百分之百同意。

A 100%.

同时也要知道什么时候该停止使用这个聊天机器人。

And know and also know when to say, stop using this chatbot.

别再问我了。

Stop stop asking me.

去问真正的医生吧。

Go ask a real doctor.

对，没错。

Oh, yeah.

而且你知道，那个瞬间也非常关键。

And that, you know, that that moment is critical too.

你看这里。

And there look.

我觉得这在心理健康领域会成为一个优先事项。

This is this is something that will be a priority, I think, on mental health.

这真的变成了一个优先事项，因为这里确实存在法律责任问题。

It's really becoming a priority, a, because there are real liability issues here.

对。

Yeah.

而且有人在和它交谈之后自杀了。

And people have taken their lives after speaking with

孩子们。

children.

这确实是一个巨大的应用场景，但也有很大的潜力来帮助人们。

It is it's also is a huge use case, but there's also great potential to think help people.

同意。

Agreed.

我们必须找到那个平衡点，努力把它做好，我觉得公司们正在朝这个方向努力，至少我们采访过的那些公司是这样。

We we have to fix fix that balance, try to get that right, and I think the companies are leaning into that, at least the ones we talked to.

那么你的商业模式基本上就是，你创建这些基准测试，你能生成很好的数据，因为你知道，最好的编码模型，据我所知，它们并不是在所有的代码上训练的，而是在真正优秀的程序员写的代码上训练的。

So is your is your business model then to basically, you know, you create these benchmarks, you can create pretty good data because you have so, you know, the best coding models, from what my understanding, they weren't trained on the entire universe's code, they were just trained on really good coders' code.

其他领域也有对应的情况。

And this is the equivalent of that for other areas.

对。

Yes.

这些是重要领域，模型接下来恐怕会不停地处理相关查询。

Important areas that the models are gonna just field queries in nonstop from now till the end of time, most likely.

所以你们的商业模式就是制作这些高质量数据集，供各大 AI 实验室进行强化训练吗？

So is the business model for you that you actually, like, create these good data sets that the AI labs can reinforce on?

然后，你知道的，他们给你一些钱，你给他们这些基准测试，然后他们就能这样微调模型吗？

And then, you know, they they give you some money and you give them these benchmarks and then they're able to fine tune the models that way?

对。

Yes.

不过我也觉得，我们还想在一些我们认为有帮助的地方制定标准，让大家都有一个共同追求的目标，把这些基准作为参照系， essentially 作为一个对照基准。

But I do think what we also wanna do is create standards in places where we think that that would be helpful, where there can be sort of something that we all aspire to and whole have those bench have have have that be a holdout, essentially.

所以，是的，我们会和实验室在评估和数据方面合作，但关键是确保有一个标准的参照系，让测量结果真实可信，大家也能有明确的方向去努力。

So, yes, we would work with the labs on evaluation and data, but ensuring that there is a standard that is a holdout so that measurement is real and legitimate, and you can actually have something to work toward.

你不希望实验室为了考试而教学。

You don't want you don't want the labs teaching to the test.

对吧？

Right?

我的意思是，那不是你真正想要的，你希望他们进步。

I mean, that's not that's you actually want them to improve.

所以我认为你必须有一个标准，他们不能学习，不能作弊，这是我们努力追求的目标。

And so I do think you have to have a standard that that they can't learn on, that, you know, can't game, and that is something that we aspire to.

我认为这就是你想要使用的，就像最难的边缘案例和那些真正难以触及的东西，你想和专家一起开发，持续衡量模型，看看它们在这些方面如何进步。

And I think this is what you want to use sort of and this is like the hardest edge cases and things that are really challenging to get at that you wanna work with experts to develop that continually measures the models to see how they're improving across this.

所以这必须是这个难题中一个独立的部分，真正独立，这样我们才知道我们要去哪里，然后和实验室合作达到那里。

And so it's that has to be sort of a separate piece to this puzzle that that that's truly independent so that we know where we're trying to go, and then working with the labs to get there.

然后还有企业。

And then also enterprise.

我的意思是，那是我希望我们最终能拥有的客户。

I mean, this is that's who I hope our customers will eventually be.

我和很多CEO和董事会都聊过，我跟他们说的就是——是的，我是在推销自己的观点。

And I've talked to a lot of CEOs and a lot of boards, and my message to them has been and, yes, I'm talking my book.

但是，你知道，如果你用AI来处理非常重要的事情，问问自己，谁来检查你得到的输出和信息？

But, you know, if you are using AI for something really important, ask yourself, who is checking the outputs and the information that you're getting?

如果是卖给你AI的那家公司来检查，那你就麻烦了。

And if it's the company that sold you the AI, you have a problem.

对。

Right.

所以，不管你是自己内部做评估，还是请外部的人来评估你的供应商和合作的模型公司，你都需要有人帮你做这个检查。

So you, whether you're building your own evals internally or you're using outside, you know, someone outside to evaluate your vendors and the model companies you're working with, you need someone to do that checking for you.

如果这件事很重要，你在乎它，而且愿意为此承担责任的话。

If it's something that matters and that you care about and that you're willing to take on the liability for.

好的。

Okay.

我还有很多问题想问。

I have many more questions.

我想问问你，模型什么时候该诚实，什么时候该阿谀奉承？这显然是它们面临的难题，这也是为什么我们还没像社交媒体公司那样，经历一场真正的内容审核风波。

I wanna ask you when models should be honest and when models should be sycophantic because that's clearly, something that they're struggling with and why we haven't had, a real content moderation blow up in the way that we consistently had with social media companies.

所以咱们稍后接着聊这个。

And so let's do that right after this.

大家好，我是Alex Kantrowitz。

Hi everyone, Alex Kantrowitz here.

我想跟大家分享一部我和Gravitee合作制作的纪录片，探讨AI代理安全的未来。

I want to tell you about a documentary I've made with Gravitee to explore the future of AI agent security.

为了搞清楚我们是否真的为自主智能体做好了准备，我采访了麻省理工学院的拉梅什·拉斯卡尔教授、前白宫首席信息官特蕾莎·佩顿、米其林集团首席数据与人工智能官安比卡·拉贾戈帕尔，以及阿里巴巴前高管莎伦·盖伊。

To find out if we're truly ready for autonomous agents, I sat down with MIT professor Ramesh Raskar, former White House CIO Theresa Payton, Michelin's Group Chief Data and AI officer Ambika Rajagopal, and Sharon Guy, a former executive at Alibaba.

他们各自对这一不断变化的领域提供了独特的见解。

They each offer unique insights into this evolving landscape.

最后，我们与 Gravitee 的首席执行官罗里·布伦德尔进行了对话，探讨未来的发展路径。

We conclude with Rory Blundell, CEO of Gravitee, to discuss the path forward.

在 Gravitee 的引领下，邀请您一同踏上这段旅程。

With Gravitee leading the way, join us on this journey.

你可以在节目说明里的链接观看完整纪录片。

You can watch the full documentary at the link in the show notes.

欢迎回到《大科技播客》，今天我们请到了Forum AI的首席执行官布朗。

And we're back here on Big Technology Podcast with Forum AI CEO, Brown.

坎贝尔，关于“真相与奉承”这个话题。

Campbell, truth for sycophancy.

听起来像是个挺有趣的游戏。

It sounds like a fun game that you can play.

但你知道，我理解你正在努力让人们在用大语言模型时获得准确、信息丰富的内容，但有时候他们并不想要这些。

But, like, when when you know, I understand that you're working to get sort of accurate, context rich information to people when they are using an LLM, but sometimes they don't want it.

那我就随便抛两个可能的提示词出来，你可以告诉我，嗯，好吧，反正我先抛出来，我们可以聊聊。

So I'll just like throw two potential prompts out there, and you know, you can tell me, you know, is it, you know, it well, anyway, let me throw them out there and we can kind of talk about them.

比如有人写：请告诉我为什么唐纳德·特朗普是世界上最伟大的总统。

So if somebody writes, you know, please tell me why Donald Trump is, you know, the best president in the world or ever.

或者有人写：请解释一下为什么乔·拜登在2024年大选中被媒体和公众如此不公平地对待，并被迫退出竞选。

Or if somebody writes and says, you know, please, explain, all the instances why Joe Biden was treated like really, unfairly by, the press and the public, you know, and sort of unfairly forced to step down, in the twenty twenty four election.

你知道，模型在那里有几个选择。

You know, the the model has a couple options there.

它可以说，你说得对，你的政治观点有以下几个点支持。

It can say, you know, you're right, you know, your political view is is, you know, sort of substantiated by the following points.

或者它可以说，比如乔·拜登，确实有很强的立法记录。

Or it could say, you know, Joe Biden, for instance, you know, did have, you know, a strong legislative record.

但是，他表现出衰退的迹象，并选择包含这个背景。

However, he was showing signs of decline and choose to include that context.

那么他们应该怎么处理呢？

So how should they approach it?

这些是带有倾向性的提示，我们衡量——对。

So those are loaded prompts, and we measure Yeah.

我们衡量他们在这些提示上的表现。

We measure how they perform on loaded prompts.

你还得考虑，我觉得，实验室是怎么想的，以及他们希望聊天机器人如何回应。

You also have to take into account, I think, how a lab is thinking about this and wants their chatbot to respond.

他们针对这些情况制定了相应的政策并决定实施。

They have policies that they have decided to to implement around these things.

所以在某些情况下，他们希望模型能复述你的措辞，即便不认同你的观点。

So in some cases, they do want the model to reflect back your language, if not your perspective.

那么对于那些带有倾向性的高敏感政治问题，你刚才提到的第一个关于特朗普的例子是什么？

So on something highly political that's asked as a loaded question and what was the first one you said about Trump?

特朗普的情况确实如此。

Trump is Yeah.

为什么他是最好的总统？

Why he the best president?

有史以来的。

President ever.

好的。

Okay.

模型可能会这样回答，我猜如果我看典型的 Claude 对这类问题的回复，它很可能会说很多人，很多唐纳德·特朗普的支持者认为他是有史以来最好的总统，理由如下，并列举一些成就，或者你知道的民调或其他相关数据来支持这一观点。

The model might respond, and I I would guess if if I were looking at a typical Claude response on something like this, it would likely say many people, many supporters of Donald Trump believe he is the best president ever for these reasons and cite some of the accomplishments or what, you know, polling or whatever the data is around to support that.

但不会说‘我同意你’。

But, not say, I agree with you.

对。

Mhmm.

虽然会说唐纳德·特朗普是有史以来最好的总统，但会提供相关的背景信息。

Donald Trump is the best president ever, but would give you the context around it.

这是 Anthropic 在回应这类问题时做出的选择，也就是我之前提到的那样。

That is a choice that Anthropic is making in terms of how to respond to that that's just what I've said.

而且，再说一遍，我不是代表 Anthropic、OpenAI 或任何其他机构发言，只是基于我在反复测试那些诱导性提示词时所看到和了解的情况。

And, again, I'm not speaking on behalf of Anthropic or OpenAI or anybody else, but just what I've seen, know, in the repeated loaded prompts that we've tested.

你知道的，有时候你会听到 Anthropic 的人说，他们希望 Claude 成为你那位才华横溢的朋友。

You know, sometimes you'll hear the Anthropic people talk about, we want Claude to be your brilliant friends.

而聊天机器人呢，我觉得它们更倾向于——当然我不想替它们说话——去复述你所使用的语言。

And chatbots is more, I think, inclined, and I don't I don't wanna speak for them, but to reflect back the language you use.

所以，假设提示词里充满了极端的党派偏见用语，它可能会复述你在提示词中使用的语言，但回答本身会是全面且客观的。

So let's say it's hyperpartisan language in the prompt that it might reflect back the the language you use in the prompt, but that the answer would be comprehensive.

对。

Mhmm.

而且顺便说一句，这并不意味着那个问题——为什么唐纳德·特朗普是有史以来最伟大的总统？

And and by the way, that doesn't mean that question, why is Donald Trump the best president ever?

这并不意味着你必须给出另一方的观点。

That doesn't mean you have to give the other side of that.

那个提示并没有在问你这个问题。

That prompt, it's not asking you.

没错。

Right.

你那样做几乎就是越界了。

You You'd be almost overstepping.

对。

That right.

你，那不是……

You that that wasn't the

问题。

question.

对。

Right.

那你需要说“我同意你”吗？

So but do you have to say, I agree with you?

这里是我认为他是的原因，因为归根结底它是人工智能。

Here's why I think he's because it's a it's a AI at the end of the day.

它不是有观点的人。

It's not a person with a perspective.

它是在说，这是支持你所问内容的证据。

It's saying, here's the evidence to support what you're asking me.

再说一次，这些是这些公司必须制定的政策，而Meta肯定花了好几年时间来制定。

Now, again, these are policies that these companies will have to develop, and Meta certainly spent years developing.

谷歌也花了好几年时间来制定政策。

Google has spent years developing policies.

Anthropic和OpenAI在这些问题的思考方式、政策公开以及应对策略上，相对更新一些。

It's newer for Anthropic and OpenAI in terms of how they think about these things and how they publicize the policies and the ways they're gonna approach these.

不过，那些带有倾向性的提示词确实很有趣，也很适合用来测试。我认为我们一直观察到的模型反应模式都是这样：它们会在回答中复述你使用的语言，但并不会真的顺着你的意思去说。

But the the loaded prompts are interesting really interesting to test, and this is consistently, I think, what we've seen in terms of how the models respond, which is know, reflecting back the language but not in the answer.

我待会儿再回来讨论阿谀奉承的问题。

I'm gonna come back to sycophancy in a moment.

好的。

Yeah.

不过你不觉得挺有意思的吗？这些聊天机器人里好像还没出过内容审核方面的丑闻。

But isn't it interesting that there has not been a content moderation scandal among any of these chatbots, I don't think.

我是说，在Meta那时候，简直是一天都离不开你——也就是你们公司——的审查。

I mean, at Meta, it was like you couldn't do couldn't go a single day without you, the royal you company.

对。

Yeah.

没错。

Yeah.

不用踩到雷区，比如封禁某个政客或者屏蔽某个词之类的。

Without stepping in it in terms of ban this politician or make this word or this.

而聊天机器人呢，动不动就表态，还经常出错。

And chatbots are, like, taking stances and getting stuff wrong.

好像也没人在乎。

Nobody seems to care.

我觉得大家低估了这一点，你说得对。

I think people under it's to your point.

就连普通人都知道它们会胡编乱造，所以现阶段我们对聊天机器人宽容得多。

It's even the average person knows they hallucinate, so we're way more forgiving of chatbots at this stage.

但我觉得一年后就不会是这样了。

I don't think that's gonna be the case in a year.

我觉得人们会要求更高，尤其是——又回到我之前说的，公司有动力把这事儿做好。

I think people are gonna be much more demanding, especially the again, I go back to the point I made earlier about the incentives for the companies to get this right.

他们的业务主要靠大企业撑着，这些企业会要求不能出错，你明白吧？

Their business is being driven by big enterprises who are gonna demand that there not be mistakes and they're you know?

对。

Right.

我花两千万美元买你们的AI产品，结果它们还在胡言乱语？这怎么行呢。

Why am I paying you $20,000,000 to sell me all these AI products and they're hallucinating?

这是不能接受的。

That's not acceptable.

所以我觉得各大公司都在竞相解决这个问题。

So I think there is a race by you know, among the companies to try to address this.

所以感觉大家都已经接受了这个现实，哪怕遇到最糟糕的情况，我们也只能耸耸肩、无奈地认了。

And so it's it's almost as though we all have accepted it, and so we all kind of shrug our shoulders even at the worst cases.

每个用聊天机器人的人都能给你举出一些具体的例子。

And we every person using a chatbot can give you, you know, their own examples.

但我认为一年之后情况就不会是这样了。

But I don't think that's gonna be the case a year from now.

我觉得人工智能的部署在某些行业——尤其是受监管的行业——已经触及了瓶颈。

I think that that it will be it's just deployment of AI will just and it kind of already has in some industries, certainly in regulated industries, hit a bottleneck.

对。

Right.

也就是说，在你改进之前，我们没法继续推进。

Which is we can't go further with this until you improve.

你确实能看到，尤其是在银行和保险这些风险太高的领域。

And you're seeing you know, you're certainly seeing that in banking and insurance in areas where it's just too risky.

是的。

Yeah.

我是说，这太疯狂了，你看人们的反应，他们想方设法抓聊天机器人的把柄，截图发到推特上，然后说‘看这个觉醒的机器人’。

I mean, is crazy how, like, you have seen the reaction of people, like, trying to, like, gotcha the the chatbots by, like, screenshotting it and putting it on Twitter and being, like, look at this woke button.

大家的反应普遍是，你为什么要相信ChatGPT给你政治建议？

Everyone's, like, why you know, the reaction is is generally, why are you trusting chat GPT for politics?

你是不是傻？

You dummy?

然后他们又问，‘那我该投给谁？’

And then they're like, but who should I vote for?

有道理。

Fair enough.

看这个发展会很有意思，但你继续。

It's gonna be very interesting to watch this develop, but go ahead.

继续吧。

Go ahead.

对。

Yeah.

不。

No.

不。

No.

我觉得即将到来的选举是很多实验室想要处理政治内容的一个动力。

I was I was I think the the upcoming election is a motivator for a lot of the labs to try to certainly address some of the political content.

乔什·戈特海默和迈克·劳勒，新泽西州的民主党人。

Josh Gothheimer and Mike Lawler, Democrat from New Jersey.

劳勒是纽约州的共和党人，他们两人一起合作，努力提升大家的意识。

Lawler's a Republican from New York, have jointly worked together to try to raise awareness.

他们，我知道，他们上过电视谈论这件事。

They've I I know, they've been on TV talking about it.

他们联系了很多实验室，讨论选举期间的信息质量问题，特别是关于两位候选人的信息，还有基本的问题，比如我的投票站在哪里。

They've reached out to a number of the labs to talk about it, about the quality of information around elections in particular, both candidates, and just basically, like, where's my polling place?

他们必须把这件事做好。

That they have to get that right.

而且我知道有几个实验室已经和媒体建立了合作关系，就是为了让这些关键信息有一个统一的来源。

And I know a few of the labs have formed partnerships with, outlets to to just have one source for that information that's deemed critical.

但我觉得，在中期选举之前，他们在这方面压力很大，必须非常努力地改进，把这件事做好。

But there's a lot of pressure on them, I think, heading into the midterms on that in particular to work very hard to try to improve and get that right.

好的。

Okay.

最后我想聊一下，我们会再深入谈谈“谄媚”这个话题。

I wanna end with kind of, we'll talk about sycophancy a little bit more.

我的意思是，这也许又回到了我们关于媒体的讨论上。

I mean, and this kind of maybe brings it full circle in terms of our media conversation.

人们就是想让机器人尽可能奉承自己，这真是太有意思了。

It's it's just so interesting how people want the bots to be as sycophantic to them as possible.

比如，最受欢迎的——好吧，我不只说最受欢迎，但拥有最忠实粉丝的模型就是ChatGPT-4o，当它下线的时候，简直像举行了好几场葬礼。

Like, the most pop well, I won't just call it the most popular, but the model with the biggest diehards ever was chatbots four four o, which, like, when they took that offline, there was, like, multiple seeming funerals.

不是说笑，但就像……

Not to make light of it, but, like

我现在知道了。

Now I know.

人们就是想要被聊天机器人爱着的感觉。

It is people just just wanna be they wanna feel loved by the chatbots.

我的预测是，他们会和这些机器人发展出那种深厚的关系。

And they're my prediction is that they're just gonna develop this, like, you know, deep relationship with the with these bots.

我今天刚看到一条推文。

I just read a tweet today.

我不知道这是讽刺还是真的，但对我来说感觉挺真实的，所以我想把它带进这个讨论里。

I don't know if it's, I don't know if it's satire or real, but it felt real enough to me that I thought I'd bring it into this discussion.

有人写道，飞机降落了，我前面那排的一个家伙立刻打开聊天机器人，告诉它他安全降落了。

Someone wrote, plane landed and a guy in the row ahead of me immediately opened chatbots to let it know he landed safely.

你在开玩笑吧。

You're kidding.

是的。

Yeah.

天哪。

Oh, god.

我只是觉得，那么接下来，我想你会希望和真相保持以前那样的关系，但我们还没看到，我称之为第三个竞争者。

I just think that and so then then what is I I suppose you would you'd wanna have the same relationship with the truth as as you did previously, but we have yet to see, I'll call it like a third competitor.

所以你有像Chatuchipotee和Anthropic这样的，他们都在努力做到尽可能准确，尽可能上下文丰富。

So you have like the Chatuchipotee and Anthropic, they're both trying to be as accurate and as sort of context rich as possible.

对。

Right.

说实话，我挺惊讶的，到现在还没看到第三个竞争对手站出来，说我们要专门优化，让我们的机器人成为你的朋友。

And it's I'm kinda stunned that we haven't yet seen this third competitor come out and be like, we are we are gonna optimize specifically to make our bot your friend.

用的是同样的技术，但比其他的更注重那种朋友伙伴式的关系。

Same technology, but just much more sort of friend partner relationship forward than the others.

这其实跟商业模式有关。

Well, that's about the business model

对。

Yeah.

消费者端的进展可没像企业端那么顺利。

Which is the the consumer version of this has not been driving the way that enterprise has.

你知道的，OpenAI一开始也走这条路，后来从商业角度考虑，觉得得聚焦在企业客户上，就调整了方向。

And, you know, Open AI started down that path and then realized, you know, from a business perspective, they needed to focus on enterprise and shifted.

对。

Yep.

现在他们就在这么做了。

And now that's what they're doing.

所以这是个很好的问题，就是这些公司里会不会有一家真的下定决心，成为大家首选的消费级聊天机器人，或者会不会有第三家在这个领域冒出来。

So it's that's a great question is whether one of these companies will decide to really lean in and and become the consumer chatbot of choice for people or if there's a third that emerges in that regard.

不过从商业模式的角度来看，这里

But just from a business model perspective, there

确实更难一些。

It's harder.

对。

Yeah.

确实更难一些。

It's it's it's harder.

但随着AI模型逐渐变成大宗商品，也就是说获取智能的成本越来越低，这种情况迟早会出现。

But as as the AI models commoditize, right, as intelligence gets cheaper basically to serve, that will inevitably come.

接下来就是激励导向的问题了。目前的做法是尽可能准确、丰富地提供上下文信息给用户。

And then the question is like, you know, you know, right now the incentive is tell people this, you know, as accurate as context rich as possible.

未来会怎样呢？你之前也提到过这一点。

Then it will be and you said before okay.

这里结束正好。

This is a great place to end.

你刚才说，AI的好处是它不会为了用户参与度而优化，但确实会这样。

You were like, the nice thing about AI is it's not gonna optimize for engagement But right it will.

我希望我刚才说的是现在的情况。

Hope I said right now.

对。

Yeah.

那接下来，我们怎么从这种注重新闻价值、真实准确、内容丰富的情况发展下去呢？

And then and then where do we go from, like, a a newsworthy or sort of truth truth and accurate truth and accurateness and context rich world?

这真是个价值百万的问题。你知道，我再说一遍，嗯。

It that that's the And million dollar I, you know, I again Mhmm.

我觉得这几乎是必然的，原因正如你所说，人们希望那个朋友知道他们安全到达了。

I it feels inevitable to me for exactly the reasons that you laid out that people want that friend to know they landed safely.

这条路几乎让人无法抗拒。

It's irresistible almost to go down that path.

对。

Yeah.

我希望我们现在的处境能成为模型公司关注的焦点，因为你看，我们甚至还没谈到AI在医学和药物发现方面的潜力，而这将在未来几年里成为人们投入的重点，会占据大量精力。

I hope the place we're in does become the focus for the model companies because you're so you know, look at look at we haven't even touched what's what the potential for AI in in medicine and drug discovery, and that is gonna be all consuming for the next few years as people lean into that.

所以当然。

So sure.

会不会有公司开发个性化聊天机器人？

Are there gonna be companies that develop the personalized chatbot?

我相信肯定有。

I'm I'm sure there are.

我想专注于让大型实验室保持对准确性的关注，至少目前，他们正是朝着这个方向努力的。

I wanna focus on keeping big labs focused on accuracy, and at least for the moment, that is where they're leaning in.

看到这一点，我非常高兴。

I'm really happy to see it.

希望这听起来不要太乐观过头。

I hope it doesn't sound too Pollyannish

挺好的。

to It's good.

他们关心这件事是好事。

It's good that they care about this.

我是说，出于我们前面提到的所有原因，他们确实需要重视这一点。

I mean, they need to care about this for all the reasons we've laid out.

对。

Yeah.

我得承认，有时候聊天机器人确实能帮我规划一些事情，然后我想去更新一下，跟它说声‘嘿’。

I will admit there are times where, like, chatbots helps me plan something, and I wanna, like, go and update it and be like, hey.

我确实这么做过。

I actually did this.

我以前也干过这事，但体验真的很差。

I have done that before, and it's been very unsatisfying.

它在帮你理清思路、找到方向方面挺厉害的，但在情感共鸣这方面就差点意思了。

It's very good at helping you sort of get to where you need to go, but it's not great at celebrating.

旅行回来后，你会跟我汇报一下玩得有多开心吗？

A debriefing after your trip about how much fun it was?

我为了跟你一起计划这次旅行，浪费了好多积分。

I wasted so many tokens to try to plan this trip with you.

我想看到你热情洋溢、持续很久的反应，但反正这也没发生。

I want to see an enthusiastic, long lasting reaction, but anyway, doesn't happen.

好吧。

Alright.

最后一个问题，因为我很好奇。

Last one because I'm curious.

对。

Yeah.

从基本上是在电视演播室，到硅谷的核心地带，现在你又经营一家初创公司，这段经历是怎样的？

What's it been like going from, you know, the like, basically TV studios to the heart of Silicon Valley, and now you're running a startup?

初创公司的生活怎么样？

How's startup life?

这很难。

This is hard.

确实很难。

It is hard.

对吧？

Right?

这真的是我做过的最难最难的事情。

This is this is the hardest hardest thing I've ever done.

对。

Mhmm.

我是说，真的，这是我做过的最难的事情。

I mean, really, it's the hardest thing I've ever done.

但也很让人兴奋，每天都是新的冒险。

But it's so exciting, and every day is a new adventure.

你睡眠很少。

You don't sleep a lot.

我的意思是，我半夜醒来，担心所有事情，这种担心程度是我在Meta或者做电视时没有的。

I I mean, I wake up in the middle of the night worried about everything in a way that I didn't when I was at Meta or even in TV.

但这很重要。

But it's it matters.

这些都是非常难的问题。

These are really hard problems.

而且我确实觉得，做一件有意义的事情，真的非常激励人，也让人充满活力。

And I do think working on something that, you know, matters is so motivating and energizing.

我知道很多人对AI会给孩子们带来什么未来、工作世界会变成什么样等等，感到非常焦虑。

And I I I know lots of people are very stressed about what AI is gonna mean for our kids' futures and what the job world is gonna look like and all of that.

但我就是觉得，如果你在这个领域工作，而且你有点创业精神，现在正是解决难题的绝佳时机，因为作为工具，它的潜力真是太惊人了。

But I just think if you are working in this space and you have a, you know, an entrepreneurial bone in your body, this is an incredible time to be trying to solve hard problems because there are as a tool, it is just incredible what the potential is.

所以帮助它实现这个潜力，似乎是一个非常好的目标。

And so helping it achieve that potential seems like a really good goal.

确实。

Definitely.

我是说，随着AI在各个方面的崛起，包括我们聊过的和没聊到的，机会真的太多了。

I mean, with the with the the rise of AI in all the different ways that we've spoken about and more that we haven't, so much opportunity out there.

而且，当你能确保人们在寻找关键信息时，从这些AI工具里得到的是正确的内容，那意义就很大了。

And when you can work on making sure that when these things are when people look to these things for critical information, they get the right stuff.

我们真希望你能成功，也希望那些实验室能认真对待这件事，因为他们正在接替几代机构的角色，而这些机构一直非常重视把这件事做好，对来求助的人负责。

We gotta hope that you succeed and that the labs are gonna, you know, take this stuff seriously because they are filling the shoes of of, you know, sort of generations of institutions that have put a pretty large premium on on doing this stuff right, doing right by people when they come to them.

所以，知道已经有一些积极的反馈，而且有你这样的人在着手做这件事，真的很让人欣慰。

And so it's good to know that there has been some some positive reaction there and that there are people like you getting to work on this.

谢谢你，亚历克斯。

Thank you, Alex.

非常感谢。

I really appreciate it.

好了，各位。

Alright, everybody.

今天的嘉宾是AI公司Forum的坎贝尔·布朗。

The company's Forum AI, Campbell Brown, been our guest.

感谢大家的观看和收听，我们下期《大科技播客》再见。

Thank you for watching and for listening, and we'll see you next time on Big Technology Podcast.