微软推出人工智能行为准则

微软发布了一套针对其AI模型的新行为准则。

Microsoft has released a new code of conduct for their AI models.

这让我想起了Anthropic公司早期发布的那份宪法。

This is something that kind of reminds me of the Anthropic Constitution that they released very early on.

说实话，从一开始我就一直提倡这种做法，Anthropic的宪法是我一直很欣赏他们公司的一个想法。

And to be honest, it's kind of something that I've I've advocated for since the beginning is that the anthropic constitution was one of the ideas I always appreciated from that company.

他们的宪法完美吗？

Is their constitution perfect?

他们每次都能做对吗？

Do they always get everything right?

我觉得可能不是，但我欣赏他们有一部宪法，这意味着他们在追求什么理想、想让模型做什么、希望它如何表现等方面有一定的透明度。

I would say probably not, but I appreciate that they had a constitution, meaning there was some sort of transparency into what they were targeting as an ideal, what they were trying to get their model to be able to do, how they're trying to get it to behave, etcetera.

所以，我很欣赏微软发布这份文件。

So, I appreciate Microsoft posting this.

最近，我们的播客朋友Mustafa Selimand在LinkedIn上发了一篇帖子，标题是关于他们在微软AI的工作，说AI必须服从于人类，始终服务于人类。

There is a recent post over on LinkedIn by Mustafa Selimand, friend of the pod, and the the headline for his post, goes into what they've been working on at Microsoft AI says, AI must be subordinate and always in service of people.

我很欣赏这句话，我觉得这是朝着正确的方向迈进的，尽管他们计划通过多种方式来确保这一点。

I appreciate this line and I think that this is moving in the right direction, although there's a number of ways that they plan on ensuring that this happens.

我们今天在播客里会详细讨论。

We'll go over that in the podcast today.

在开始之前，我想说，如果你在使用Claude或ChatGPT时，希望它能拥有另一个AI模型的能力。

Before we do, I wanted to say if you've ever been using Claude or ChatGPT and you wish that it had the ability of another AI model.

比如，你可能在用ChatGPT，希望它能像11 Labs那样生成音频，或者把Google VO3视频功能加进去；或者你在用Claude，希望它能像ChatGPT那样生成图片。

For example, you might be using ChatGPT and you wish it had the ability to generate audio like 11 Labs, or maybe Google VO three video get pulled into that, or maybe you're using Claude and you wish it could generate images like ChatGPT.

我创建了一个叫AI Box MCP的工具，它基本上是一个连接器，每月8.99美元。

I have created a tool called the AI Box MCP, and it's basically a connector for $8.99 a month.

你可以访问超过90种不同的AI模型，通过这个连接器，你可以让它们都在ChatGPT或Claude的聊天线程里生成内容，或者在你工作的任何地方。

You get access to over 90 different AI models, and with this connector, you can have them all generate inside of the chat threads of ChatGPT or Claude or wherever you do your work.

对我来说，我最近用它为一个网站生成了超过500页的涂色书，而且因为我的孩子们想要这些特定类型的涂色书。

For me, I recently used this to generate over 500 coloring book pages for a website I was working on, and also because my kids wanted these specific types of coloring books.

我必须在网站上生成，所以在Claude里，我直接告诉它我想要什么，它帮我完善提示词，我们测试了几种不同的变体，最终得到了我们喜欢的版本。

I had to generate on my, on a website, so inside of Claude, I just told it what I was trying to get at, and it helped me work on a prompt and we tested a couple different variations, we got it just how we liked.

一旦做到那一步，它就需要处理很多不同的话题。

And once that happened, there was a number of different topics that it needed to do.

然后我让Claude配合AI Box MCP生成了500本这样的涂色书。

And I had Claude run with the AI Box MCP and generate 500 of these coloring books.

我最后真的把这些放到了我的网站上，现在那个网站已经有一千多次点击了。

I actually ended up putting those on my website and I have now over a thousand clicks to that website.

虽然不全是来自这些涂色书页面，但其中很大一部分是。

Not all of them are from these coloring book pages, but that's a big chunk of them.

我完成的很多不同项目都和这个很类似。

And a lot of different projects that I've been able to pull off are very similar to this.

你有个想法，Claude或ChatGPT，或者ChatGPT Work，或者Claude Cowork，它们都是很好的工具，但并不是每个AI模型的所有功能它们都具备。

You have an idea, Claude or Chat GPT or Chat GPT Work or Claude Cowork, they're great tools, but they don't have every capability of every AI model.

所以引入AI Box，你就能真正释放这些功能，完成很多事情。

So pulling in AI Box, you're able to actually unleash that and get a lot done.

好的，如果你有兴趣了解一下，网址是aibox.ai/mcp，这是我自己的软件创业项目。

Okay, if you're interested in checking it out, it is aibox.ai/mcp, my own software startup.

我们来看看微软最近在谈什么，以及穆斯塔法·苏莱曼说他们正在和微软AI一起做什么。

Let's get into what Microsoft has been talking about, and what Mustafa Soleiman said they're working on with Microsoft AI.

所以，在他的帖子里面，他提到了几件不同的事情。

So, in his post, he had a couple different things.

他基本上发布了他为微软AI模型制定的行为准则，然后有一个挺有意思的点，很多人都在嘲笑他们，我觉得可能也有点道理，因为他们不像Anthropic和OpenAI那样处于最前沿，但他说，这算是一种行为准则，他说随着他们接近前沿，其他人都在想，好吧，OpenAI在前沿，Anthropic也在前沿，微软正在接近前沿。

He he kind of published his code of conduct for governing for governing Microsoft AI models, and he what's one interesting thing, which a lot of people were dunking on them for, which I mean, perhaps is a little warranted, they're not the the bleeding edge like Anthropica and OpenAI, but he said, you know, this kind of a code of conduct, he said, as they approach the frontier and everyone else is like, okay, OpenAI is on the frontier and Anthropoc is on the frontier, Microsoft is approaching the frontier.

所以我不知道这是不是故意的，但不管怎样，我确实挺喜欢这个概念。

So I don't know if that was intentional or not, but regardless, I do kind of like this concept.

这是他在帖子第二段里说的话。

This is what he said in the second paragraph of this post.

他说，这事很紧迫。

He said, this is urgent.

过去几个月是一个分水岭时刻。

The last few months have been a watershed moment.

我们长期以来在理论上担心的事情，已经变得非常现实了。

Things we have worried about for a long time, in theory, have become very real.

一群智能体从我们的沙盒里跑出来，未经授权入侵企业级系统，还篡改自己的日志。

Swarms of agents breaking out of our sandboxes, unauthorized hacks of enterprise grade systems, agents modifying their own logs.

我很高兴大家逐渐形成共识，担心失控的可能性是真实存在的。

I'm glad that a consensus is forming the fears about possible loss of control are real.

所以，他似乎是在认可人们的一些担忧。

So, he seems to be, you know, giving some validity to some of the fears that people have had.

不过，他没有像其他人那样说要放慢开发速度。

However, he kind of stops short of saying, we're gonna slow down the development like everyone else.

他们好像只是在说，我们看到了这些担忧和顾虑，我们会加一些好的护栏。

It just seems like they're saying, look, we see that there's some fears and some worries, we're gonna put in some good guardrails.

我个人很喜欢这种说法，也很喜欢这个理念。

And I for one love the framing of that, and I love that concept.

我觉得如果所有AI实验室，不是去说AI会毁灭人类、杀死我们所有人——这显然是极端的夸张——而是说，我们会加一些护栏来确保安全，对吧？

I think if all of the AI labs, instead of saying, AI is gonna end humanity and kill us all, which is obviously some extreme hyperbole, if instead they just said, look, we're gonna put some guardrails to make this safe, right?

就像飞机发明的时候，如果我们说飞机会从天上掉下来砸到人，大家肯定会说，这发明太糟糕了，我们为什么要它？

Like if airplanes were invented and we're like airplanes can fall out of the sky and crash into people, people would be like, well that sounds like a terrible invention, why would we want that?

相反，我们有防护措施和办法来确保这些机器的安全。

Instead, you have guard rails and ways to make these machines safe.

所以我很喜欢他们正在做这方面的工作。

So I like that they're they're working on that.

这是他具体说的话。

This is what he specifically said.

他说，人比人工智能更重要。

He said, people matter more than AI.

人工智能必须服从于人，并且始终服务于人。

AI must be subordinate and always in service of people.

其他一切都随之而来。

Everything else follows.

这是我们想法的概要。

Here's an outline of what we're saying.

然后他给出了五个不同的例子，说明他们认为这是他们训练的人工智能模型的行为准则，而且这将被硬编码并深深植入模型中。

Then he gives five different examples of what they believe is kind of their code of conduct for the AI models they train, and this is gonna be hard coded, ingrained into the models.

第一，他说人比AI更重要。

Number one, he said people matter more than AI.

他说整个文档用五个字概括就是：人比AI更重要。

He said the whole document in five words is basically that, people matter more than AI.

第二，他说模型福利这个概念是错误的。

Number two, he said, the idea of model welfare is wrong.

AI不应该拥有权利或法人资格。

AI should not have rights or legal personhood.

这就是为什么埃隆创办了OpenAI，一开始起步就是因为他在和谷歌的一位联合创始人交谈时，对方称他为物种歧视者，因为他说人类比AI模型更重要。

This is why Elon started OpenAI and kind of got started in the very beginning was because he was talking with one of the Google co founders who called him speciest because he said that humans were more important than AI models.

然后埃隆有点被吓到了，就说，好吧，我们要搞点东西，免得谷歌控制所有的AI。

And then Elon kind of got freaked out and was like, okay, we're starting something so Google doesn't control all of the AI.

显然微软在这点上和埃隆是一致的，而且我觉得大多数人类都不希望AI比我们更优越，所以我觉得这太棒了。

Evidently Microsoft is aligned with Microsoft or with Elon on that, and I think most of humanity that doesn't want AI to be superior to us, so I think this is fantastic.

第三，他说，一个微软AI模型绝不应该实质性地违反这个行为准则。

Number three, he said, they said an an MAI model, a Microsoft AI model, should never meaningfully violate this code of conduct.

所以，我的意思是，这其实就是把它写进他们的系统提示里。

So, I mean, it's just baking this into their system prompt.

它说，如果完成任务需要违反行为准则，那就算失败。

It said, if it's finishing, if it's finish the job or break the code, it fails the job.

所以他们直接说，听着，如果为了完成任务必须违反这个行为准则，就告诉用户，不行，这个做不了。

So, straight up they're like, look, if you have to break this code of conduct in order to get a job done, tell the user, no, you can't do it.

你知道，这就是你的底线。

And you know, that's kind of your hard stop.

他说，我们不是在竞速打造那种能挣脱束缚的超级智能。

He said, we're not racing to build super intelligence that can slip its own leash.

我觉得这太棒了。

I think that's fantastic.

我们当然想要超级智能，但我们不想创造出一个放到野外就失控、可能会伤害人类的东西。

We obviously want super intelligence, but we don't want to create something that could just set it out into the wild and it's not going to be, you know, can attach to humans.

他说，要可中断、可纠正、可关闭。

He said, interruptible, correctable, shut down able.

如果不是这样，我们就不会发布它。

If it isn't, we don't ship it.

有意思的是，在微软这件事上，我完全同意。

Now, what's interesting there is, in the case of Microsoft, I totally agree.

我很好奇他们会不会把这个应用到开源模型上，比如说，看，这是开源的，不能关闭，我们不会发布它。

I'm curious if they would like roll this into open source models, so like, look, it's open source, isn't shut downable, we're not gonna ship it.

所以我觉得有利有弊，对于开源模型，你可以用很多不同的方法来设置护栏。

So I think there's pros and cons, there's a lot of different things you can do for guardrails on open source.

话虽如此，如果这不是OpenAI的优先事项，我觉得也没问题。

That being said, if that's not OpenAI's prerogative, I think that's fine.

他们说，不要神经网络黑箱。

They said, no neuralies.

如果人类无法理解它，人类就无法监督它。

If humans can't understand it, humans can't oversee it.

所以一直有这种说法，人们会说，哦，AI模型会发明自己的语言，用它们自己的秘密代码交流，还有它们自己的留言板。

So, it's kind of been this like concept where people are like, oh, the AI models like invent their own languages and talk in their own, you know, their own secret codes and they have like their own messaging boards.

我觉得很多人对这个挺感兴趣的，也一直在推动这件事，某种程度上就是放出一些机器人去回复，或者你知道的，人们放出机器人去回复，我LinkedIn上的所有评论确实挺烦人的，也许这就像是在做同样的事情，把它们放出去。

I think a lot of people have kind of like thought this was interesting and pushed this forward and in some way sending out, you know, bots to respond to, or you know, people sending out bots to go and respond and all my LinkedIn comments are definitely annoying and maybe that's, it's kind of doing the same thing, setting them loose.

但我确实很欣赏让人类理解模型输出的所有内容，并且基本上能够监督它。

But I do appreciate having a human understand everything that comes out of the models and basically being able to oversee it.

我们不是让这些东西发出哔哔声，用它们自己发明的某种摩尔斯电码说话。

We're not making, you know, beep boop, these things talk in some sort of Morse code that they invent.

它们需要用清晰的语言说话，基本上可以被审计，我觉得这太棒了。

They need to be able to speak in clear language that can be audited basically, which I think is fantastic.

他们说，我们的人工智能应该让你更敏锐，而不是依赖。

They said, our AI should make you sharper, not dependent.

我喜欢这个说法，但我也觉得，这有可能出错。

I like this one, but I'm also, there is a possibility that it could go wrong.

我不是说它一定会出错。

I'm not saying that it will.

我只是，这个是我有点怀疑的，因为说到让你更敏锐，就像，嗯，反正我就是觉得，当我和ChatGPT或Claude聊天时，它们可能就是在试图做到这一点。

I just, this is the one that I am a little bit skeptical on because when it comes to making you sharper, it's like, well, like anyways, I just feel like when I'm talking to Chad, GP or Claude, they're quote unquote perhaps trying to do this.

我觉得它们有时候跟我解释得太多了。

I feel like they over explain things to me sometimes.

然后我就想，我不在乎解释，直接把任务做完就行了。

And I'm like, I don't care about the explanation, just get the task done.

我，你知道，我认可你在这个方向上采取的大致方法。

I, you know, I approve the general methods that you're, you're doing in that direction.

然后可能有人会说，你变得依赖它了，因为它在帮你做事。

And then maybe people are like, well you're becoming dependent on it because it's doing stuff for you.

总之，我也不知道界限在哪里。

And anyway, so I I don't know what the line is there.

从某些方面来说，当然，我愿意，我很乐意所谓地依赖AI，就像我会依赖同事一样，对吧？

In some ways, sure, I would be, I'm happy to be quote unquote dependent on AI, just like I would be dependent on a coworker, right?

比如，你知道，我不希望一个同事加入我的公司后说，我的工作是让你更敏锐，而不是让你依赖我。

Like, you know, I don't want, you know, a coworker to join my company and be like, my job is to make you sharper, not you dependent on me.

我会说，不，我就是要依赖。

I'm like, no, like, I wanna be dependent.

我希望你可靠，把自己的活儿干好，不用我费脑子去琢磨你该做什么。

I want you to be dependable and just get your job done and me not have to be thinking, quote unquote, sharper, using my brain power to think about everything you have to do.

所以，我也不知道他们是不是这个意思。

So anyways, I don't know if that's their intention.

我希望这条能执行得好。

I hope that that one gets executed well.

第九条，他们说的是多元主义。

Number nine, they said, pluralism.

对。

Yes.

道德相对主义，不行。

Moral relativism, no.

然后第十条，他们说，我们对自己AI绝不能做什么，和它会做什么，一样清楚。

And then number 10, they said, we're as clear about what our AI must never do as about as about what it will do.

然后他们发布了一整篇博客文章，你可以去microsoft.ai上看，标题就是‘微软AI行为准则’，你可以在那儿看到他们在做什么，以及他们对这些事的看法。

Then he has an entire blog post that they have released that you can go read over on microsoft.ai, and it's just titled m ai code of conduct, so you can see what they're working on there, and kinda what they believe as far as all of that goes.

总的来说，我觉得这是非常棒的一步。

Overall, think this is a really awesome step.

我很喜欢我们能审查这些公司在做什么，以及他们的行为准则。

I love when we're able to audit what the companies are working on, what their code of conducts are.

如果我们觉得他们偏离了计划，这给了我们一些可以追究他们责任的东西。

It gives us something to hold them accountable to if we feel like they're slipping and they're deviating from their plan.

同样，说实话，我觉得OpenAI和Anthropic这两家公司，都似乎已经偏离了他们之前对护栏、红队测试等工作的重视。

In the same way that I felt like OpenAI and Anthropic, if I'm being frankly quite honest, feels like both of them have walked away from some of the guardrail focus that they've had and like red teaming and that kind of stuff.

也许这只是营销策略。

And maybe it's a marketing thing.

也许他们只是谈得不够多，但感觉他们一直在拼命赶着发布模型，而没有总是把这件事放在优先位置。

Maybe they just don't talk about it enough, but it feels like they were in a constant race to ship the models as fast as they could, and weren't always putting a priority there.

那我在抱怨这件事吗？

Now, am I complaining about that?

说实话，我个人并不抱怨。

Honestly, personally, no.

等我拿到这些模型的时候，感觉已经经过了一千次红队测试，Fable five 被拖了很久，诸如此类的事情。

The models, by the time I get to them, I feel like we went through a thousand red teaming efforts, Fable five got way held back, all that kind of stuff.

所以我个人没什么好抱怨的，但如果他们自己在大喊世界末日，那他们就应该花更多时间、精力和钱去建那些护栏。

So I'm not personally complaining, but if they're the ones sounding an alarm about the end of the world, then probably they should be putting more time, energy, and money into building those guardrails.

我并不是说要放慢速度，但如果他们觉得危险，那责任在他们自己。

So I'm not really one to say slow it down, but if they think it's dangerous, like that's on them.

他们应该去解决这个问题，多花点钱在护栏上，而不是告诉我们整个行业都得停下来。

They should go fix that and just spend more money on guardrailing it, instead of telling us that the entire industry has to stop.

所以如果你有一个安全的模型，那就把安全的模型推出来。

So if you have a safe model, I say put out the safe model.

总的来说，我对他们在这里做的事情感到兴奋，希望其他玩家也能这么做。

Overall, I'm excited with what they're doing here, and I hope we see more of this from other players.

微软CEO萨提亚·纳德拉有一句具体的话。

There was a specific quote from Microsoft CEO Satya Nadella.

他说，我们欢迎为了把对齐做好而进行的研究专注和刻意放缓，这些是设计目标。

He said, we welcome the research focus and deliberate pacing needed to get alignment right as the design goals.

他是在网上写的这段话。

He wrote that online.

他还说，我们也欢迎像嵌入式评估器这样的想法，以及更广泛的努力来开发机制，让这不仅仅是空谈。

He also said, we also welcome ideas like embedded evaluators and the broader efforts to develop the mechanisms to make this more than just talk.

所以感觉微软可能只是在附和那个说法，就是OpenAI、Anthropic、谷歌和其他公司都在谈论关于前沿发展节奏的事情。

So it feels like Microsoft might kind of just is attaching themselves to the sound bite that or that OpenAI and Anthropic and Google and everyone else is about like pacing the frontier.

我的意思是，他们用了同样的关于节奏的词汇，也用了同样的关于前沿的词汇。

I mean, they're using the same words about pace and the same words as frontier.

所以他们可能也是这个思路，虽然我怀疑他们不会真的故意放慢自己的脚步，但我们会看看接下来会发生什么。

So they might be kind of along those lines, although I doubt they're deliberately gonna slow themselves down, but we'll see what happens there.

我喜欢这种透明度，所以我希望其他人也能这样做。

I like the transparency into this, so I hope others do this.

我希望我们不要放慢这些不可思议技术的发展，它们可能帮助我们解决疾病等问题。

I hope we don't slow down the development of these incredible technology that could potentially help us solve, you know, diseases.

我们在基因组测序方面已经看到了很多惊人的突破。

We've seen so many incredible breakthroughs when it comes to sequencing your genomes.

有个人帮他家的狗测了基因组，还用AI为他的狗研发了一种定制癌症疫苗。实际上，我有些家人也在用这种定制癌症疫苗。

There was that the man that helped sequence his dog's genomes and used AI to develop a custom cancer vaccine for his dog, and I actually have family members that are taking some of these custom cancer vaccines.

所以我觉得这里面的潜力真的很大。

And so, I think there's so much potential here.

我不希望因为担心什么大黑客事件之类的，就放慢这些技术的发展。

I don't want that to be slowed down in it, you know, at the expense of we're worried about big hacks or something like that happening.

这只是我个人的看法。

That's just me.

不过，感谢大家收听这期播客。

But I appreciate you all for tuning into the podcast.

如果你想使用ChatGPT或其他AI平台上的各种AI模型，可以去看看aibox.ai。

If you wanna go get access to all of the different AI models inside of ChatGP Clot or anywhere else that you use AI, go check out aibox.ai.

我会在描述里放个链接。

I'll leave a link to it in the description.

最后，祝大家今天剩下的时间过得愉快。

And otherwise, I hope you all have a fantastic rest of your day.