1. Maths ❌

2. Stats ❌

3. Linear Equation ❌

4. Probability ❌

5. Calculus ❌

GENERATIVE-AI => Generating something

What is GPT?

GENERATED PRE-TRAINED TRANSFORM

The name of Chat GPT is what it does; it does not differ, just in short form

Generative: means it generates, for example, it generates, translation

1. Hello => Namaste

2. Now => Abhi

Pretrained: means it’s already trained for that

Transformer: Just like a translate app

Sequence is very important because it changes everything:

1. Development is more interesting than Coding( many agree on this)

   Coding is more interesting than Development (Reversed version, few will agree on this)

Working of an AI (taking inputs)??

Step -1 (Tokenisation)

Each letter is treated as an individual token because the system does not have inherent knowledge of English or numbers.

website: Tiktokenizer

Now in this, each single letter is taken differently, but in this situation, BEST (one token) DEVELOPER (other token) is also called vocab

Step 2 (Vector Embedding)

Website: Embedding projector - visualization of high-dimensional data

HOW VECTOR EMBEDDING ACTUALLY LOOKS

⬆️ Overview of the website

LET’S TAKE PROGRAMMERS

THIS WHEN WE CLICK THE PROGRAMMERS

WE ACTAULLY GET THIS OUT OF IT

WHAT IT ACTUALLY HAS ;

POSITIONAL ENCODING

Very important, as the position is important to choose for every word

Not every word does not mean what it means in English

1. I have an apple (it’s the actual apple), I have the apple (this is an iPhone)

Google-based -research paper: 1706.03762

1. Inputs + Positional Encoding

• Input Embedding: Converts each input token (like a word or subword) into a vector of fixed dimension.

• Positional Encoding: Since Transformers don’t have recurrence or convolution, positional encodings are added to help the model understand the order of tokens in a sequence.

Final result:
Input Embedding + Positional Encoding = Input to the Encoder Blocks

2. Encoder Block (repeated N times)

The encoder consists of N identical layers (often 6 in the original Transformer paper). Each layer has two main components:

A. Multi-Head Self-Attention (Yellow block)

• The model compares each word to every other word in the input to capture contextual relationships.

• "Multi-head" means this is done multiple times in parallel with different parameter sets, allowing the model to learn different types of relationships.

Phase 2

INTERFENCING

Example: I am a developer

This line has ended; this should be known

Training

training the model until it learns the correct thing, as it also makes mistakes sometimes

For example, I ask AI what is the best language to learn DSA for jobs?

answer: Python

(Now I would disagree with it until the AI gives me the correct answer, this is how it will know what it should give when someone asks it about the same thing)