簡明解釋大型語言模型：英語精讀講義

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

想像一下，你偶然發現了一個簡短的電影劇本，裡面描述了一個人和他們的AI助理之間的場景。

The script has what the person asks the AI,

劇本裡有這個人問AI的問題，

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

但AI的回答部分被撕掉了。

Suppose you also have this powerful magical machine

假設你還有一台強大的神奇機器，

that can take any text and provide a sensible prediction

它可以接收任何文本並提供合理的預測，

of what word comes next.

猜出下一個詞是什麼。

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

你可以把你手頭上的內容輸入到這台機器中來完成劇本，

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

看看它會預測出什麼詞作為AI回答的開頭，

and then repeating this over and over 片語搭配

然後一遍又一遍地重複這個過程，

with a growing script completing the dialogue.

隨著劇本不斷變長，最終完成整個對話。

When you interact with a chatbot,

當你與聊天機器人互動時，

this is exactly what's happening.

發生的情況正是如此。

A large language model is a sophisticated mathematical function

大型語言模型是一個複雜的數學函數，

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

用於為任何文本片段預測下一個出現的詞。

Instead of predicting one word with certainty, though,

然而，它並不是絕對確定地預測某一個詞，

what it does is assign a probability 文法

它所做的是分配一個機率

to all possible next words.

給所有可能的下一個詞。

To build a chatbot,

要建立一個聊天機器人，

what you do is lay out some text 片語動詞

你需要做的是列出一些文本，

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

描述使用者和假設的AI助理之間的互動。

You add on whatever the user types in 片語動詞

無論使用者輸入什麼，你都把它加進去

as the first part of that interaction,

作為這種互動的第一部分，

and then you have the model repeatedly predict the next word

然後讓模型反覆預測下一個詞，

that such a hypothetical AI assistant would say in response,

也就是這個假設的AI助理會用來作為回答的詞，

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

這就是展示給使用者的內容。

In doing this,

在執行這個過程中，

the output tends to look a lot more natural 文法

產生的輸出往往看起來自然得多，

if you allow it to select less likely words

如果你允許它選擇那些機率較低的詞，

along the way at random. 片語搭配

在這個過程中隨機選擇。

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

那麼這意味著，儘管模型本身是確定性的，

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

同一個提示在每次運行時，通常都會給出不同的答案。

Models learn how to make these predictions

模型學會如何做出這些預測

by processing an enormous amount of text,

是透過處理海量的文本，

typically pulled from the internet.

這些文本通常來自網際網路。

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

對於一個普通人來說，要讀完用於訓練GPT-3的文本量，

for example,

例如，

if they read non-stop 24/7,

如果他們24小時不停地閱讀，

it would take over 2600 years.

這需要超過2600年。

Larger models since then train on much, much more.

從那以後，更大的模型在比這多得多的資料上進行訓練。

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

你可以把訓練看作有點像在一台大機器上調轉盤。

The way that a language model behaves

語言模型的表現方式

is entirely determined by these many different continuous values,

完全由這許多不同的連續數值決定，

usually called parameters or weights.

這些數值通常被稱為參數或權重。

Changing those parameters will change the probabilities

改變這些參數將改變機率，

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

即模型在給定輸入下為下一個詞提供的機率。

What puts the large in large language model 慣用語

大型語言模型之所以能被稱為「大」，

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

是因為它們可以擁有數千億個這樣的參數。

No human ever deliberately sets those parameters.

從來沒有人類刻意去設定這些參數。

Instead, they begin at random,

相反地，它們從隨機值開始，

meaning the model just outputs gibberish,

這意味著模型一開始只會輸出胡言亂語，

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

但它們會根據許多文本範例被反覆微調。

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

這些訓練範例之一可能只有幾個詞，

or it could be thousands,

或者也可能有成千上萬個詞，

but in either case,

但無論是哪種情況，

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

它的工作原理都是將除了該範例最後一個詞之外的所有內容輸入

into the model and compare the prediction

到模型中，並比較模型的預測結果

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

與範例中真實的最後一個詞。

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

一種被稱為反向傳播的演算法被用來調整所有的參數，

in such a way that it makes the model 文法

其調整的方式是使模型

a little more likely to choose the true last word

稍微更有可能選擇真實的最後一個詞，

and a little less likely to choose all the others.

而稍微減少選擇所有其他詞的可能性。

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

當你對成千上萬億個範例執行此操作時，

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

模型不僅開始在訓練資料上給出更準確的預測，

but it also starts to make more reasonable predictions

而且它也開始做出更合理的預測，

on text that it's never seen before.

針對它以前從未見過的文本。

Given the huge number of parameters

考慮到龐大的參數數量

and the enormous amount of training data,

和海量的訓練資料，

the scale of computation involved in training a large language model

訓練一個大型語言模型所涉及的運算規模

is mind-boggling. 用詞

是令人難以置信的。

To illustrate, 表達脈絡

為了說明這一點，

imagine that you could perform 1 billion additions and multiplications

想像一下，如果你能每秒執行10億次加法和乘法

every single second.

運算。

How long do you think that it would take

你認為需要多長時間，

for you to do all of the operations

才能完成所有的運算，

involved in training the largest language models?

也就是訓練最大語言模型涉及的運算？

Do you think it would take a year?

你覺得需要一年嗎？

Maybe something like 10,000 years?

也許是1萬年左右？

The answer is actually much more than that.

答案其實遠不止於此。

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

這需要遠超過1億年的時間。

This is only part of the story, though. 慣用語

然而，這只是故事的一部分。

This whole process is called pretraining.

這整個過程被稱為預訓練。

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

從網際網路上自動補齊一段隨機文本的目標，

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

與成為一個優秀的AI助理的目標截然不同。

To address this, 表達脈絡

為了解決這個問題，

chatbots undergo another type of training, 片語搭配

聊天機器人經歷了另一種類型的訓練，

just as important, called reinforcement learning with human feedback.

同樣重要，這被稱為基於人類回饋的增強學習。

Workers flag unhelpful or problematic predictions, 用詞

工作人員標記出無益或有問題的預測，

and their corrections further change the model's parameters,

他們的修正會進一步改變模型的參數，

making them more likely to give predictions

使它們更有可能給出預測，

that users prefer.

也就是使用者更偏好的預測。

Looking back at the pretraining, though, 表達脈絡

不過，回顧預訓練，

this staggering amount of computation

這種驚人的運算量

is only made possible by using special computer chips 文法

只有透過使用特殊的電腦晶片才能實現，

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

這些晶片針對平行執行許多操作進行了最佳化，

known as GPUs.

它們被稱為GPU。

However,

然而，

not all language models can be easily parallelized.

並不是所有的語言模型都能輕易地平行化。

Prior to 2017, 用詞

在2017年之前，

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

大多數語言模型每次只能處理文本中的一個詞。

But then,

但後來，

a team of researchers at Google

Google的一組研究人員

introduced a new model known as the Transformer.

推出了一種被稱為Transformer的新模型。

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

Transformer不會從頭到尾閱讀文本。

They soak it all in at once, 片語動詞

它們一次性吸收全部文本，

in parallel.

以平行的方式。

The very first step inside a Transformer,

在Transformer內部的第一步，

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

其實對於大多數其他語言模型也是如此，

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

是將每個詞與一長串數字關聯起來。

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

原因是訓練過程只處理連續數值，

so you have to somehow encode language using numbers,

所以你必須想辦法用數字來編碼語言，

and each of these lists of numbers

而這每一個數字列表

may somehow encode the meaning of the corresponding word.

可能以某種方式編碼了對應詞的含義。

What makes Transformers unique 文法

讓Transformer獨一無二的是，

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

它們依賴於一種稱為注意力機制的特殊操作。

This operation gives all of these lists of numbers

這個操作給所有的這些數字列表

a chance to talk to one another 慣用語

一個相互交流的機會，

and refine the meanings that they encode

並根據周圍的上下文

based on the context around,

完善它們所編碼的含義，

all done in parallel.

一切都在平行進行。

For example,

例如，

the numbers encoding the word bank

編碼單詞「bank」的數字，

might be changed based on the context surrounding it

可能會根據它周圍的上下文被修改，

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

從而以某種方式編碼了「河岸」這一更具體的概念。

Transformers typically also include a second type of operation

Transformer通常還包括第二種操作，

known as a feed-forward neural network,

被稱為前饋神經網路，

and this gives the model extra capacity

這給了模型額外的容量

to store more patterns about language learned during training.

來儲存在訓練期間學到的關於語言的更多模式。

All of this data repeatedly flows through many different iterations

所有這些資料反覆流經這兩種基本操作的

of these two fundamental operations,

許多不同迭代中，

and as it does so, 表達脈絡

在這個過程中，

the hope is that each list of numbers

人們期望每個數字列表

is enriched to encode whatever information might be needed 文法

都能得到豐富，從而編碼任何可能需要的資訊，

to make an accurate prediction

以做出準確的預測，

of what word follows in the passage.

預測段落中接下來的詞是什麼。

At the end,

在最後，

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

序列中最後一個向量會執行最後一個函數，

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

這個向量現在已經有機會受到來自輸入文本的

from the input text,

所有其他上下文的影響，

as well as everything the model learned during training,

以及模型在訓練期間學到的所有內容的影響，

to produce a prediction of the next word.

從而產生對下一個詞的預測。

Again,

同樣，

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

模型的預測看起來像是每個可能的下一個詞的機率。

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

儘管研究人員設計了每個步驟如何運行的框架，

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

但重要的是要理解，具體的行為是一種湧現現象，

based on how those hundreds of billions of parameters

基於那數千億個參數

are tuned during training.

在訓練期間是如何被調整的。

This makes it incredibly challenging to determine 文法

這使得確定模型為什麼會做出它具體的預測

why the model makes the exact predictions that it does.

變得極其具有挑戰性。

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，當你使用大型語言模型的預測來自動補齊提示時，

the words that it generates

它生成的詞

are uncannily fluent, 用詞

異常流暢，

fascinating,

令人著迷，

and even useful.

甚至非常有用。

If you're a new viewer

如果你是一位新觀眾，

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

並且對Transformer和注意力機制的具體工作原理感到好奇，

boy do I have some material for you. 文法

哇，我正好為你準備了一些資料。

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

一個選擇是去看我製作的關於深度學習的系列影片，

where we visualize and motivate the details of attention 用詞

在那個系列中，我們將注意力機制的細節視覺化，並解釋了它的動機，

and all the other steps in a Transformer.

以及Transformer中的所有其他步驟。

But also, on my second channel,

另外，在我的第二個頻道上，

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

我剛發布了幾個月前發表的一場演講，

about this topic for the company TNG in Munich.

是關於這個主題的，當時在慕尼黑的TNG公司講的。

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

有時候，我其實更喜歡我作為非正式演講所製作的內容，

rather than a produced video, 文法

而不是製作精良的影片，

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

但我讓你來決定這兩者中哪個更適合作為後續內容。