Constitutional AI 2026: Training Claude to Self-Critique

00:00

All right, let's get into it.

好了，我们开始吧。

00:02

Today, we are taking a really close look

今天，我们要深入剖析

00:04

at one of the biggest names in the AI game right now.

当前人工智能领域最炙手可热的巨头之一——

00:07

Claude, the AI assistant from Anthropic.

来自Anthropic公司的AI助手Claude。

00:10

You know, with hundreds of AI tools out there,

要知道，市面上已有数百种AI工具，

00:12

it feels like a new one pops up every day.

几乎每天都有新面孔出现。

00:14

So the big question is, what makes Claude so special?

那么关键问题来了：Claude究竟有何独特之处？

00:17

What's its secret sauce?

它的制胜秘诀又是什么？

00:19

So at its heart, Claude is an AI assistant

其核心而言，克劳德是一款由名为Anthropic的研究公司打造的人工智能助手。

00:22

built by a research company called Anthropic.

该公司在推出它时，怀揣着一个极其清晰的使命：

00:25

And when they launched it, they had this super clear mission.

让它有用，让它诚实——

00:28

Make it helpful, make it honest,

而最关键的一点是，让它无害。

00:30

and this is the big one, make it harmless.

这种对安全性的全面考量，没错，

00:33

And that whole idea of safety, yeah,

正是我们今天要反复探讨的主题。

00:35

that's a theme we're gonna be hitting on a lot today.

好了，先说重点。

00:38

Okay, first things first.

好，先说第一点。

00:39

If you've never used Claude before, you're gonna love this.

如果你之前没用过Claude，你一定会爱上它的。

00:42

Getting started is actually super, super simple.

上手其实超级、超级简单。

00:45

Let's just walk through the steps right now.

现在咱们就来一步步操作。

00:47

Seriously, it couldn't be easier.

说真的，没有比这更容易的了。

00:49

You just pop over to Claude.ai, sign up using your email

你只需打开Claude.ai，用邮箱注册，

00:52

and click through a quick verification.

再快速完成验证。

00:54

And bam, you're in.

然后“啪”一下，就进去了。

00:56

The chat window is right there, ready for you

聊天窗口就在那儿，随时等你开聊。

00:58

to type in your very first prompt.

以下是中文翻译：

01:00

Now, here's a quick tip.

这里有个小技巧。

01:01

And honestly, this goes for pretty much any AI you use,

**输入你的第一个提示词。**

01:04

but it's especially true for Claude.

现在，这里有个小技巧。

01:06

Be specific.

说实话，这适用于你使用的几乎所有AI，

01:08

The clearer and more detailed your prompts are,

但对Claude来说尤其如此：

01:10

the better your results are gonna be.

**要具体。**

01:12

Don't think of it like typing a few words

你的提示词越清晰、越详细，

01:14

into a search engine.

以下是中文翻译：

01:15

Think of it more like you're giving detailed instructions

把它想象成你在给专业人士、学生、研究人员下达详细指令。

01:17

to a really smart assistant who's ready to help.

**翻译成搜索引擎的查询语句。**

01:21

Okay, now we're getting to one of Claude's real superpowers.

把它想象成你在给一个非常聪明的助手下达详细指令，

01:24

This is the feature that makes it an absolute beast

这个助手随时准备提供帮助。

01:26

for professionals, students, researchers,

你可以直接扔一份又密又无聊的报告给它。

01:29

really anyone who deals with a lot of information.

好了，现在我们开始触及克劳德真正的超能力之一了。

01:32

We're talking about its ability

正是这个功能，让它成为专业人士、学生、研究人员——

01:33

to handle massive amounts of text.

处理海量文本的关键，

01:36

So, this all boils down to something called

归根结底在于一个概念——

01:38

the context window.

上下文窗口。

01:40

The easiest way to think about it

最直观的理解方式，

01:41

is like the AI's short-term memory.

就是把它看作AI的短期记忆。

01:44

And Claude's is huge.

而克劳德的记忆容量惊人——

01:46

We're talking 200,000 tokens,

足足20万个词元，

01:48

which are just tiny pieces of words.

也就是构成词语的微小单元。

01:50

But what does that actually mean?

但这到底意味着什么呢？

01:52

Well, to put it in perspective,

这么说吧，打个比方——

01:54

it's like you could hand Claude a 500-page book

就像你可以递给克劳德一本500页的书，

01:56

and it would remember the whole thing.

它能把整本书一字不差地记住，贯穿你整个对话。

01:58

Word for word for your entire chat.

那它在现实世界有什么用呢？

02:01

And the real world uses for this?

哦，用处可大了。

02:02

Oh man, they're huge.

你可以把一份枯燥冗长的报告扔给它处理。

02:04

You can throw a dense boring report at it

而且你知道最棒的是什么吗？

02:06

and get a perfect summary in seconds.

并在一秒内获得完美摘要。

02:08

Or find that one needle and a haystack piece

或从海量PDF中精准定位那根“针”——

02:11

of information inside a giant PDF.

哪怕它藏在浩如烟海的文档里。

02:13

You can even analyze hundreds of pages of code

你甚至能分析数百页代码，

02:15

or pull out the most critical points

或从冗长复杂的合同中提炼关键要点。

02:17

from a long complicated contract.

这完全是颠覆性的变革。

02:19

It's a total game changer.

而你知道最棒的是什么吗？

02:20

And you know what the best part is?

立刻就能针对那份文档提问。

02:22

Using this feature is so easy.

使用这个功能非常简单。

02:24

You just look for that little paperclip icon

你只需在聊天界面中找到那个小小的回形针图标，

02:26

right there in the chat.

点击它，上传你手头的任何文件——无论是PDF、

02:27

Click it, upload whatever you've got, a PDF,

文本文件还是Word文档，就搞定了。

02:30

a text file, a Word doc, and that's it.

然后你就可以直接开始针对这份文档提问。

02:32

You can start asking questions

真的就这么简单。

02:34

about that document right away.

就这么简单。

02:35

It really is that simple.

真的就这么简单。

02:37

Okay, before we get into the really fascinating stuff,

好的，在我们深入探讨那些真正引人入胜的内容之前——

02:40

the philosophy behind how Claude keeps itself safe,

也就是克劳德如何确保自身安全背后的哲学原理——

02:42

just a quick pause.

先稍作停顿。

02:43

If you're getting something useful out of this,

如果你觉得这些内容对你有所帮助，

02:45

it would be awesome if you could take a second

那将非常棒，如果你能花一秒钟

02:46

to hit that like button, maybe share it

点个赞，或许再分享给

02:48

with someone who'd find it interesting,

觉得有趣的人，

02:50

and of course, follow us so you don't miss out

当然，也请关注我们，这样你就不会错过后续内容。

02:51

on what's next.

接下来要聊什么？

02:53

All right, now we're going to talk about

好了，现在我们要谈谈

02:54

what really, truly sets Claude apart from the pack.

究竟是什么让Claude真正从同类中脱颖而出。

02:58

It's all about its ethical framework.

关键在于它的伦理框架。

03:00

And this isn't some feature they tacked on at the end.

这可不是什么事后添加的功能。

03:02

No, this is baked into its DNA

不，这是刻在它基因里的东西，

03:04

and it's all thanks to this incredible concept

而这一切都归功于一个了不起的概念——

03:06

they call constitutional AI.

他们称之为“宪法式人工智能”。

03:09

So what in the world is constitutional AI?

那么，宪法人工智能到底是什么呢？

03:13

Well, its anthropics totally unique way

这是Anthropic公司独树一帜的方法，

03:15

of building a safe AI from the very beginning.

旨在从源头构建安全的人工智能。

03:19

See, instead of having humans constantly tell the AI

不同于让人类不断告诉AI

03:22

that was a good answer or that was a bad answer,

“这个回答好”或“那个回答差”，

03:25

they give the AI a rulebook, a set of core principles,

他们给AI一本规则手册，一套核心原则，

03:28

literally a constitution that it has to follow.

确切地说，是一部它必须遵守的“宪法”。

03:31

And this is where it gets really different

而真正的独特之处，正在于此。

03:33

from how most other models are trained.

这与大多数其他模型的训练方式不同。

03:35

You see, most models use this thing called RLHF,

你看，大多数模型采用一种名为RLHF的技术，

03:38

reinforcement learning from human feedback

即基于人类反馈的强化学习——

03:40

where people are constantly grading the AI's responses.

人们会不断对AI的回应进行评分。

03:43

But Claude, it learns to supervise itself.

但Claude不同，它学会了自我监督。

03:46

It literally uses its own constitution

它实际上会依据自身的“宪法”

03:48

to check its work, critique its answers, and make corrections.

来检查自己的输出、评判自己的答案，并做出修正。

03:51

It's basically its own ethics coach.

它本质上就是自己的伦理导师。

03:53

So how do they do this?

那么，它们是如何做到的呢？

03:54

It's basically a two-step process.

这基本上是一个两步过程。

03:57

First, in what's called the supervised phase,

首先，在所谓的监督阶段，

03:59

the model just practices.

模型只是进行练习。

04:00

It generates answers and then immediately checks them

它生成答案，然后立即对照其“宪法”进行核对，

04:03

against its constitution, correcting itself as it goes.

并在此过程中自我修正。

04:06

Then, in the second phase, the reinforcement stage,

接着，在第二阶段，即强化阶段，

04:08

it uses all that self-critique to get better and better,

它利用所有这些自我批评来不断改进，变得越来越好。

04:11

constantly reinforcing the good constitution-friendly habits.

不断强化有益于宪法的良好习惯。

04:15

And we're not talking about a few simple do's and don'ts here.

这里我们讨论的并非几条简单的注意事项。

04:18

This constitution is a serious document.

这部宪法是一份严肃的文件。

04:20

We're talking about 75 different articles,

它包含75条不同的条款，

04:23

pulling from some really heavy hitting sources,

借鉴了极具分量的来源，

04:25

like the UN's Universal Declaration of Human Rights.

例如联合国的《世界人权宣言》。

04:28

So this gives it a really solid, globally aware, ethical foundation.

因此，这赋予了它一个非常坚实、具有全球视野的道德基础。

04:32

Okay, so we've talked about how to get started

好了，我们已经讨论了如何开始。

04:34

and we've talked about the deep ethical principles it runs on.

我们探讨过它所遵循的深层伦理原则。

04:37

But what does this actually mean for you today

但这对今天的你——

04:39

trying to figure out which Claude model to use?

正纠结该用哪个Claude模型——究竟意味着什么？

04:41

Let's break down the options.

我们来逐一分析选项。

04:43

The Claude 3 family basically has a tool for every job.

Claude 3系列基本能应对各种任务。

04:46

You've got Haiku, think of it as the sprinter.

比如Haiku，你可以把它想象成短跑选手：

04:48

It's super fast, really cost-effective,

速度极快，性价比超高，

04:51

and perfect for quick, simple tasks.

非常适合处理快速、简单的任务。

04:53

Then there's Sonnet.

然后是「十四行诗」。

04:54

This is your all-rounder hitting that sweet spot

这是一款全能型选手，在速度与性能之间找到了完美的平衡点。

04:56

with a perfect balance of speed and performance.

接着是「杰作」。

04:59

And then, there's Opus.

「杰作」是重量级冠军。

05:01

Opus is the heavyweight champion.

它是你为最复杂、最烧脑的分析和研究而准备的利器。

05:02

It's the one you bring out for the most complex,

所以，当你退后一步，纵观全局时——

05:04

brain-melting analysis and research.

脑洞大开的分析与研究。

05:07

So when you step back and look at the whole picture,

所以当你退一步，纵观全局，

05:09

Claude really isn't just another chatbot.

克劳德绝非又一款普通的聊天机器人。

05:11

It's designed to be more like a partner.

它被设计得更像一位伙伴——

05:13

One that's incredibly powerful, totally focused on making

强大无比，全心致力于提升你的效率，

05:16

you more productive, and most importantly,

而最重要的是，它建立在极其坚实的伦理基础之上。

05:18

built on this really strong ethical foundation.

这便引出了核心要点：

05:21

And that really brings us to the big takeaway here.

Anthropic 从根本上试图

05:24

Anthropic is fundamentally trying

将道德指南针直接嵌入人工智能本身。

05:25

to build a moral compass right into the AI itself.

把道德指南针直接嵌入AI本身。

05:28

So I want to leave you with a question to think about.

因此，我想留给大家一个问题去思考：

05:30

What does it mean for the future?

这对未来意味着什么？

05:32

For how we work and live with these machines,

对我们如何与这些机器共事、共存意味着什么——

05:34

if we can give them an ethical constitution from the start?

如果我们从一开始就能赋予它们一套道德准则？

05:37

It's definitely something to chew on.

这绝对值得深思。

05:39

Thanks for tuning in.

感谢收看。

Play Queue