简明解释大语言模型：英语精读讲义

Imagine you happen across a short movie script that describes a scene between a person and their AI assistant. 短語動詞

想象一下，你偶然发现了一个简短的电影剧本，里面描述了一个人和他们的AI助手之间的场景。

The script has what the person asks the AI,

剧本里有这个人问AI的问题，

but the AI's response has been torn off. 短語動詞

但AI的回答部分被撕掉了。

Suppose you also have this powerful magical machine

假设你还有一台强大的神奇机器，

that can take any text and provide a sensible prediction

它可以接收任何文本并提供合理的预测，

of what word comes next.

猜出下一个词是什么。

You could then finish the script by feeding in what you have to the machine, 短語動詞

你可以把你手头上的内容输入到这台机器中来完成剧本，

seeing what it would predict to start the AI's answer,

看看它会预测出什么词作为AI回答的开头，

and then repeating this over and over 短語搭配

然后一遍又一遍地重复这个过程，

with a growing script completing the dialogue.

随着剧本不断变长，最终完成整个对话。

When you interact with a chatbot,

当你与聊天机器人互动时，

this is exactly what's happening.

发生的情况正是如此。

A large language model is a sophisticated mathematical function

大语言模型是一个复杂的数学函数，

that predicts what word comes next for any piece of text.

用于为任何文本片段预测下一个出现的词。

Instead of predicting one word with certainty, though,

然而，它并不是绝对确定地预测某一个词，

what it does is assign a probability 語法

它所做的是分配一个概率

to all possible next words.

给所有可能的下一个词。

To build a chatbot,

要构建一个聊天机器人，

what you do is lay out some text 短語動詞

你需要做的是列出一些文本，

that describes an interaction between a user and a hypothetical AI assistant.

描述用户和假设的AI助手之间的互动。

You add on whatever the user types in 短語動詞

无论用户输入什么，你都把它加进去

as the first part of that interaction,

作为这种互动的第一部分，

and then you have the model repeatedly predict the next word

然后让模型反复预测下一个词，

that such a hypothetical AI assistant would say in response,

也就是这个假设的AI助手会用来作为回答的词，

and that's what's presented to the user.

这就是展示给用户的内容。

In doing this,

在执行这个过程中，

the output tends to look a lot more natural 語法

生成的输出往往看起来自然得多，

if you allow it to select less likely words

如果你允许它选择那些概率较低的词，

along the way at random. 短語搭配

在这个过程中随机选择。

So what this means is even though the model itself is deterministic, 語法

那么这意味着，尽管模型本身是确定性的，

a given prompt typically gives a different answer each time it's run. 用詞

同一个提示在每次运行时，通常都会给出不同的答案。

Models learn how to make these predictions

模型学会如何做出这些预测

by processing an enormous amount of text,

是通过处理海量的文本，

typically pulled from the internet.

这些文本通常来自互联网。

For a standard human to read the amount of text that was used to train GPT-3,

对于一个普通人来说，要读完用于训练GPT-3的文本量，

for example,

例如，

if they read non-stop 24/7,

如果他们24小时不停地阅读，

it would take over 2600 years.

这需要超过2600年。

Larger models since then train on much, much more.

从那以后，更大的模型在比这多得多的数据上进行训练。

You can think of training a little bit like tuning the dials on a big machine. 語法

你可以把训练看作有点像在一台大机器上调表盘。

The way that a language model behaves

语言模型的表现方式

is entirely determined by these many different continuous values,

完全由这许多不同的连续数值决定，

usually called parameters or weights.

这些数值通常被称为参数或权重。

Changing those parameters will change the probabilities

改变这些参数将改变概率，

that the model gives for the next word on a given input.

即模型在给定输入下为下一个词提供的概率。

What puts the large in large language model 習語

大语言模型之所以能被称为“大”，

is how they can have hundreds of billions of these parameters.

是因为它们可以拥有数千亿个这样的参数。

No human ever deliberately sets those parameters.

从来没有人类刻意去设定这些参数。

Instead, they begin at random,

相反，它们从随机值开始，

meaning the model just outputs gibberish,

这意味着模型一开始只会输出胡言乱语，

but they're repeatedly refined based on many example pieces of text.

但它们会根据许多文本文本示例被反复微调。

One of these training examples could be just a handful of words, 短語搭配

这些训练示例之一可能只有几个词，

or it could be thousands,

或者也可能有成千上万个词，

but in either case,

但无论是哪种情况，

the way this works is to pass in all but the last word from that example 語法

它的工作原理都是将除了该示例最后一个词之外的所有内容输入

into the model and compare the prediction

到模型中，并比较模型的预测结果

that it makes with the true last word from the example.

与示例中真实的最后一个词。

An algorithm called backpropagation is used to tweak all of the parameters

一种被称为反向传播的算法被用来调整所有的参数，

in such a way that it makes the model 語法

其调整的方式是使模型

a little more likely to choose the true last word

稍微更有可能选择真实的最后一个词，

and a little less likely to choose all the others.

而稍微减少选择所有其他词的可能性。

When you do this for many, many trillions of examples,

当你对成千上万亿个示例执行此操作时，

not only does the model start to give more accurate predictions on the training data, 語法

模型不仅开始在训练数据上给出更准确的预测，

but it also starts to make more reasonable predictions

而且它也开始做出更合理的预测，

on text that it's never seen before.

针对它以前从未见过的文本。

Given the huge number of parameters

考虑到庞大的参数数量

and the enormous amount of training data,

和海量的训练数据，

the scale of computation involved in training a large language model

训练一个大语言模型所涉及的计算规模

is mind-boggling. 用詞

是令人难以置信的。

To illustrate, 表達脈絡

为了说明这一点，

imagine that you could perform 1 billion additions and multiplications

想象一下，如果你能每秒执行10亿次加法和乘法

every single second.

运算。

How long do you think that it would take

你认为需要多长时间，

for you to do all of the operations

才能完成所有的运算，

involved in training the largest language models?

也就是训练最大语言模型涉及的运算？

Do you think it would take a year?

你觉得需要一年吗？

Maybe something like 10,000 years?

也许是1万年左右？

The answer is actually much more than that.

答案其实远不止于此。

It's well over 100 million years. 短語搭配

这需要远超过1亿年的时间。

This is only part of the story, though. 習語

然而，这只是故事的一部分。

This whole process is called pretraining.

这整个过程被称为预训练。

The goal of autocompleting a random passage of text from the internet

从互联网上自动补全一段随机文本的目标，

is very different from the goal of being a good AI assistant.

与成为一个优秀的AI助手的目标截然不同。

To address this, 表達脈絡

为了解决这个问题，

chatbots undergo another type of training, 短語搭配

聊天机器人经历了另一种类型的训练，

just as important, called reinforcement learning with human feedback.

同样重要，这被称为基于人类反馈的强化学习。

Workers flag unhelpful or problematic predictions, 用詞

工作人员标记出无益或有问题的预测，

and their corrections further change the model's parameters,

他们的修正会进一步改变模型的参数，

making them more likely to give predictions

使它们更有可能给出预测，

that users prefer.

也就是用户更偏好的预测。

Looking back at the pretraining, though, 表達脈絡

不过，回顾预训练，

this staggering amount of computation

这种惊人的计算量

is only made possible by using special computer chips 語法

只有通过使用特殊的计算机芯片才能实现，

that are optimized for running many, many operations in parallel, 短語搭配

这些芯片针对并行运行许多操作进行了优化，

known as GPUs.

它们被称为GPU。

However,

然而，

not all language models can be easily parallelized.

并不是所有的语言模型都能轻易地并行化。

Prior to 2017, 用詞

在2017年之前，

most language models would process text one word at a time. 短語搭配

大多数语言模型每次只能处理文本中的一个词。

But then,

但后来，

a team of researchers at Google

谷歌的一组研究人员

introduced a new model known as the Transformer.

推出了一种被称为Transformer的新模型。

Transformers don't read text from the start to the finish.

Transformer不会从头到尾阅读文本。

They soak it all in at once, 短語動詞

它们一次性吸收全部文本，

in parallel.

以并行的方式。

The very first step inside a Transformer,

在Transformer内部的第一步，

and most other language models for that matter, 表達脈絡

其实对于大多数其他语言模型也是如此，

is to associate each word with a long list of numbers.

是将每个词与一长串数字关联起来。

The reason for this is that the training process only works with continuous values, 語法

原因是训练过程只处理连续数值，

so you have to somehow encode language using numbers,

所以你必须想办法用数字来编码语言，

and each of these lists of numbers

而这每一个数字列表

may somehow encode the meaning of the corresponding word.

可能以某种方式编码了对应词的含义。

What makes Transformers unique 語法

让Transformer独一无二的是，

is their reliance on a special operation known as attention. 短語搭配

它们依赖于一种称为注意力机制的特殊操作。

This operation gives all of these lists of numbers

这个操作给所有的这些数字列表

a chance to talk to one another 習語

一个相互交流的机会，

and refine the meanings that they encode

并根据周围的上下文

based on the context around,

完善它们所编码的含义，

all done in parallel.

一切都在并行进行。

For example,

例如，

the numbers encoding the word bank

编码单词“bank”的数字，

might be changed based on the context surrounding it

可能会根据它周围的上下文被修改，

to somehow encode the more specific notion of a river bank.

从而以某种方式编码了“河岸”这一更具体的概念。

Transformers typically also include a second type of operation

Transformer通常还包括第二种操作，

known as a feed-forward neural network,

被称为前馈神经网络，

and this gives the model extra capacity

这给了模型额外的容量

to store more patterns about language learned during training.

来存储在训练期间学到的关于语言的更多模式。

All of this data repeatedly flows through many different iterations

所有这些数据反复流经这两种基本操作的

of these two fundamental operations,

许多不同迭代中，

and as it does so, 表達脈絡

在这个过程中，

the hope is that each list of numbers

人们期望每个数字列表

is enriched to encode whatever information might be needed 語法

都能得到丰富，从而编码任何可能需要的信息，

to make an accurate prediction

以做出准确的预测，

of what word follows in the passage.

预测段落中接下来的词是什么。

At the end,

在最后，

one final function is performed on the last vector in this sequence,

序列中最后一个向量会执行最后一个函数，

which now has had a chance to be influenced by all the other context 語法

这个向量现在已经有机会受到来自输入文本的

from the input text,

所有其他上下文的影响，

as well as everything the model learned during training,

以及模型在训练期间学到的所有内容的影响，

to produce a prediction of the next word.

从而产生对下一个词的预测。

Again,

同样，

the model's prediction looks like a probability for every possible next word.

模型的预测看起来像是每个可能的下一个词的概率。

Although researchers design the framework for how each of these steps work, 表達脈絡

尽管研究人员设计了每个步骤如何运行的框架，

it's important to understand that the specific behavior is an emergent phenomenon 用詞

但重要的是要理解，具体的行为是一种涌现现象，

based on how those hundreds of billions of parameters

基于那数千亿个参数

are tuned during training.

在训练期间是如何被调整的。

This makes it incredibly challenging to determine 語法

这使得确定模型为什么会做出它具体的预测

why the model makes the exact predictions that it does.

变得极其具有挑战性。

What you can see is that when you use large language model predictions to autocomplete a prompt, 表達脈絡

你能看到的是，当你使用大语言模型的预测来自动补全提示时，

the words that it generates

它生成的词

are uncannily fluent, 用詞

异常流畅，

fascinating,

令人着迷，

and even useful.

甚至非常有用。

If you're a new viewer

如果你是一位新观众，

and you're curious about more details on how Transformers and attention work,

并且对Transformer和注意力机制的具体工作原理感到好奇，

boy do I have some material for you. 語法

哇，我正好为你准备了一些资料。

One option is to jump into a series I made about deep learning 短語動詞

一个选择是去看我制作的关于深度学习的系列视频，

where we visualize and motivate the details of attention 用詞

在那个系列中，我们将注意力机制的细节可视化，并解释了它的动机，

and all the other steps in a Transformer.

以及Transformer中的所有其他步骤。

But also, on my second channel,

另外，在我的第二个频道上，

I just posted a talk that I gave a couple months ago 用詞

我刚发布了几个月前发表的一场演讲，

about this topic for the company TNG in Munich.

是关于这个主题的，当时在慕尼黑的TNG公司讲的。

Sometimes I actually prefer the content that I make as a casual talk

有时候，我其实更喜欢我作为非正式演讲所制作的内容，

rather than a produced video, 語法

而不是制作精良的视频，

but I leave it up to you which one of these feels like the better follow-on. 習語

但我让你来决定这两者中哪个更适合作为后续内容。