LARGE LANGUAGE MODELS (LLMs)

Type in the code given

HOW LARGE
LANGUAGE
MODEL (LLM)

WORKS

V06
HOW XIAN NENG

JANANEE A/P SUBRAMANIAM

SITI NABILAH BINTI MUHAMAD ZIN

ICE-BREAKING

SESSION

“GUESS THE AI”

Word Cloud

If you want to ask AI a question, what would it be?

INTRODUCTION

Large Language Models (LLMs)

type of artificial intelligence

significantly transformed the landscape of natural
language processing (NLP)

These models are trained on vast amounts of text
data

allowing them to understand, generate, and
manipulate human language with remarkable
proficiency

NLP

Text generation, translation, and

sentiment analysis.

MAR

CONVERSATIONAL AI
Powering chatbots and virtual

assistants for natural

interactions.

TRANSFER LEARNING

Fine-tuning for specific tasks,

saving time and resources.

MULTI MODAL
APPLICATIONS

Processing text, images, and

audio for tasks like image

captioning.

DESCRIBE THE USE OF FOUNDATION MODEL OF

LLM

CONTENT CREATION

Assisting in writing, generating

articles, and automating

reports

MAR

PERSONALIZATION
Tailoring recommendations
based on user preferences

ACCESIBILITY

Enhancing tools for individuals

with disabilities (e.g., speech-to-

text).

RESEARCH AND
DEVELOPMENT

Supporting innovation in AI

through versatile applications

DESCRIBE THE USE OF FOUNDATION MODEL OF

LLM

KEY COMPONENTS OF LARGE LANGUAGE MODELS

(LLMS)

Lorna Alvarado

the process of breaking down text into
smaller units called tokens, which can
be words, subwords, or characters.

the process of converting text tokens
(words, subwords, or characters) into dense
numerical vectors (fixed-length arrays of
numbers).

deep learning model architecture designed
to handle sequential data, particularly well-
suited for natural language processing tasks

TOKENIZATION

EMBEDDING

TRADITIONAL MODEL

ATTENTION
MECHANISM

TRANSFORMER

TRADITIONAL MODEL

Attention Mechanism

Tokenization

Word Embedding

TRANSFORMER

WHAT

IS?

PURPOSE?

TOKENIZATION?

Tokenization is the process of dividing

text into smaller units called tokens,
which can be words, sub-words, or

characters

Helps machines understand and

process natural language by converting

unstructured text into structured,

analyzable units.

TOKENIZATION

How many meaning units are there in the sentence?

Let's learn about tokenisation and how LLM works.

Let's learnabout tokenisationand how LLMworks.

Word-Level Tokenisation

SubWord Level Tokenisation

Let'slearnabout tokenisationand how LLMworks.

TRY TO TOKENISE THE FOLLOWING

EXERCISE:

Becky enjoyed teaching three-year-olds.

word: “Becky” “enjoyed” “teaching” “three-

year-olds”

Multiple Choice

"Becky enjoyed teaching three-year-olds"

How many sub-words are there in the sentence?

• 11

• 12

TRY TO TOKENISE THE FOLLOWING

EXERCISE:

Becky enjoyed teaching three-year-olds.

word: “Becky” “enjoyed” “teaching” “three-

year-olds”

sub-word:

"Be", "##cky", "enjoy", "##ed", "teach",

"##ing", "three", "-", "year", "-", "olds"

TRY TO TOKENISE THE FOLLOWING

EXERCISE:

Becky enjoyed teaching three-year-olds.

word: “Becky” “enjoyed” “teaching” “three-

year-olds”

sub-word:

"Be", "##cky", "enjoy", "##ed", "teach",
"##ing", "three", "-", "year", "-", "old" “s"

Embedding - Converting meaning to numbers

How do you make computer understand meaning?

i) Break down words into “meaning units”
ii) Assign “meaning units” to a number
iii) Put all these numbers into a vector

King: Human, male
Queen: Human, Female
Ox: Animal, male
Cow: Animal Female

King: [0.1, 0.0]
Queen: [0.1, 1.0]
Ox: [0.2, 0.0]
Cow: [0.2, 1.0]

This will enable the machines to group words based on their meaning:
All words that has the meaning of female, just look at the second vector value.

Multiple Choice

Which of the following might represent “prince”?

• [0.34, 0.12]

• [0.25, 0.65]

• [0.56, 0.70]

• [0.78, 0.21]

Is there a problem with this algorithm?

The Problem

i) What if the same word has completely different meaning?

“bank”, “bank”

2) What if they are close enough but different?

“kid”, “kid”, “kid”?

Solution: Contextualisation

WHY CONTEXT

MATTERS:

Words can mean different things in different sentences. For

example, the word “bank” could mean:

A financial institution: “I need to go to the bank.”

The side of a river: “We sat by the river bank.”

money

bank

river bank

shore

financial

side of river

Word map

embedding space

embedding space

vectors

vectors

vectors

vectors

vectors

water

money

bank

financial

Word map

embedding space

vectors

vectors

money: [0.45, -0.87, 1.02, -1.23, 0.76]

bank: [0.46, -0.89, 1.03, -1.20, 0.74]

Since these two vectors are very close
(the numbers are similar), the AI knows
that "money" and "bank" are related.

You can think of it like a set of
coordinates that help the AI

understand relationships between

words.

vectors

How does LLM know what to focus on?

The cat lay on the carpet because it was tired

• Are all the words equally important?

• Attention Mechanism

  • Attention Score

Reorder

Order the following words according to Attention Score from High to Low for the sentence:

The cat lay on the carpet because it is tired.

When the LLM is trying to figure out the meaning of "it"

• Cat

• Lay

• tired

• carpet

• because

1

2

3

4

5

• 1
• 2
• 3
• 4
• 5

How does Attention Mechanism work?

​Think of two sentences that use the same word in different meaning.

How does Attention Mechanism work?

​Think of two sentences that use the same word in different meaning.

​I go to the bank to get some money
I go to the bank to fish

How does Attention Mechanism work?

Use AI to learn about AI:

​I go to the bank to get some money
I go to the bank to fish

​ChatGPT Prompt:
Explain how you differentiate these two sentences using Attention Mechanism?

How do LLM predict

Activity:
Think of a word to complete the following sentence that doesn't make sense:
The cat lay on the carpet because it is ________.

Ask the AI to explain itself

​Activity:

Prompt 1:
Complete the following sentence with a word that doesn't make sense.


Prompt 2:

Open Ended

Share your prompt!

Ask the AI to explain itself (Use Chatgpt)

​Activity:

Prompt 1:
Complete the following sentence with a word that doesn't make sense.


Prompt 2:
How do you use Attention Mechanism to determine that this word doesn't make sense?

The LLM doesn’t just guess randomly—it uses everything it knows

to pick the best word based on what’s already been said.

LLMS (LARGE LANGUAGE MODELS)

Let’s say you start a sentence: “Once upon a…”
Now, the LLM needs to predict what comes next. To do that, it plays a
word-guessing game!

Look at the context: The LLM sees "Once upon a…" and knows this
usually starts a fairy tale.

1.

Guess the next word: It thinks, “What word would make sense here?”
Words like "time," "princess," or "story" might all be possibilities.

2.

STEP-BY-STEP: HOW THE LLM
PREDICTS THE NEXT WORD

4. Assign probabilities: What word to choose next?
For example:

“time” might have a high probability (because “Once upon a time” is a
common phrase).
“princess” might have a lower probability (it could make sense but isn't
as common as "time").
“banana” would have a very low probability (it doesn’t really fit a fairy
tale).

STEP-BY-STEP: HOW THE LLM
PREDICTS THE NEXT WORD

5. Pick the word with the highest probability: The LLM chooses
the word that has the highest score—in this case, it will likely choose
“time.”

So, the sentence becomes: “Once upon a time…”

This process continues, word by word, until the sentence is
complete!

THE PROCESS CONTINUES:

Imagine LLMs (Large Language Models) read 100 sentences on:

The sky is ____.

90 sentences continues with “blue”, another 10 sentences continue
with “grey”.

HOW LLMS CALCULATES

THE PROBABILITY

(how likely something is to happen.)

The cat which was very fluffy jumped over the fence.

The cat jumped over the fence.

How to use LLMs to produce a picture:

How to use LLMs to produce a text:

Multiple Choice

Imagine you’re writing a magical story with an LLM. You start with:
“The wizard waved his wand and turned the…”
Which of these words do you think the LLM is most likely to choose next?

• Castle

• Car

Multiple Choice

Pretend the LLM is a word detective trying to complete the sentence:
“The dog ran quickly to the…”
The LLM looks at the sentence and guesses the next word. Which of these words would probably get the highest probability?

• Park

• Kitchen

SUMMARY

KEY COMPONENTS OF LARGE LANGUAGE MODELS

(LLMS)

Lorna Alvarado

the process of breaking down text into
smaller units called tokens, which can
be words, subwords, or characters.

the process of converting text tokens
(words, subwords, or characters) into dense
numerical vectors (fixed-length arrays of
numbers).

deep learning model architecture designed
to handle sequential data, particularly well-
suited for natural language processing tasks

TOKENIZATION

EMBEDDING

TRADITIONAL MODEL

ATTENTION
MECHANISM

TRANSFORMER

To understand human

language

LLMS (LARGE LANGUAGE MODELS)

To produce materials in
natural human language

LLMS (LARGE LANGUAGE MODELS)

Poll

Do you understand now what is Large Language Models (LLMs)?

• Yes

• No