Generative Ai creates brand-new content like text, images, music or code, instead of just analyzing existing data. this creative ability makes it revolutionary for businesses and is changing industries rapidly .

AI (Artificial Intelligence)

AI is a broad field of computer science focused on building systems that can perform tasks that normally require human intelligence. AI focuses on understanding, analyzing, or deciding based on data—not creating new content.

What AI can do:

• Recognize patterns
• Predict outcomes
• Classify images or text
• Recommend products
• Detect fraud
• Play games like chess or Go

Examples of traditional AI:

• Spam filters
• Google Maps route optimization
• Face recognition
• Recommendation systems (Netflix, Amazon)

GenAI (Generative AI)

GenAI focuses on generation — producing new content using patterns learned from data. Generative AI is a subset of AI that can create new content such as:

• Text
• Images
• Code
• Audio
• Video
• Designs

It uses models like GPT, DALL·E, Claude, Gemini, Stable Diffusion, etc.

What GenAI can do:

• Write emails, blogs, or code
• Create images and designs
• Generate human‑like conversations
• Produce music or videos
• Create synthetic data
• Build chatbots and copilots

What is Large Language Model(LLM)

Massive AI models trained on enormous volumes of text data have become exceptionally powerful at generating human-like language. These advanced systems learn patterns, context, and relationships between words, allowing them to produce natural and meaningful responses.

Before the introduction of Transformer architecture, artificial intelligence struggled to understand long context and maintain continuity in conversations. Earlier AI models often lost track of information across longer paragraphs, which limited their real-world usability.

Today, modern Large Language Models (LLMs) can maintain conversations, understand context, and remember information from earlier parts of a discussion. This ability makes them highly valuable for business applications such as customer support, content creation, automation, and intelligent decision-making.

Role of Generative AI in Business

1. Automate content creation
2. Personalize Customer experience
3. Speed up software development
4. Complex Data Analysis.

What is Token

• LLM process text as tokens, not complete words
• A token =word fragment, whole word, or punctuation mark .
• Example Generative = 1token, running = run + Ing
• Cost and speed measured by token count 
• Critical for managing GenAi application performance

What is Embedding

• Embeddings=long list of numbers representing words
• Captures meaning and context numerically 
• Similar meanings =numerically close embeddings

An embedding is a way to convert text (or images, audio, etc.) into numbers so that a computer can understand meaning, similarity, and relationships.

Meaning converted into math (These numbers are usually stored as a vector)

Key Features of Embeddings

1. Meaning-based representation

Embeddings capture semantic meaning, not just keywords.

2. High‑dimensional vectors

Typically hundreds or thousands of numbers that represent concepts.

3. Similar inputs → similar vectors

Useful for:

• Search
• Recommendations
• Clustering
• Deduplication

4. Language agnostic

Embeddings work across different languages and can group similar meanings even in different languages.

5. Efficient for large-scale search

Ideal for semantic search, RAG (Retrieval-Augmented Generation), and knowledge retrieval.

6. Used heavily in GenAI

Behind tasks like:

• Chatbot memory
• Document search
• Code search
• Personalization
• Context retrieval for LLMs

What is Attention ?.

Attention is a mechanism that helps AI decide which words are most important when understanding or generating text. When a model reads a sentence, not all words matter equally.

Example:
In the sentence: “The dog that chased the cat was hungry.”

To understand “was hungry,” the model must pay attention to “dog”, not “cat.” Attention helps the model figure out these relationships automatically.

Key Features of Attention

1. Focus on relevant information

The model gives higher weight to important words or tokens.

2. Captures relationships

Helps understand:

• context
• dependencies
• meaning across long sentences

3. Works even with long text

That’s why AI models can handle big paragraphs better than older models.

4. Basis of Transformers

Transformers (like GPT, BERT) rely heavily on Self-Attention.

5. Makes models more accurate It improves translation, summarization, Q&A, and text generation.

Foundation Models and Model families

Foundation Model: A very Large Ai model trained on massive broad datasets to perform a wide range of general tasks, like a highly educated generalist That can be easily adapted to many specific tasks without training from scratch .

Choose the right tool for the job based on cost and performance. This flexibility allows you to optimize costs while getting high quality output for various applications.

1. What is a Proprietary Model?

A proprietary model is:

• Created and owned by a company
• Closed-source (the inner working, weights, and training data are not publicly available)
• Used under license, often paid
• Fully controlled by the company

Examples

• OpenAI GPT‑4 / GPT‑5
• Anthropic Claude
• Google Gemini
• Microsoft Copilot foundation models

 Key Features

• Usually state‑of‑the‑art performance
• High accuracy, safety, reliability
• Professional support
• No access to model weights
• Limited customization
• Pay‑per‑use (API pricing)

 2. What is an Open Model?

An open model is:

• Released publicly
• Often open-source or partially open-source
• Model weights are available
• You can run it, customize it, or fine‑tune it locally or on cloud

Examples

• LLaMA / Llama 3
• Mistral / Mixtral
• Falcon
• Gemma
• Phi‑3 (open model family)

Key Features

• Transparent (weights available)
• Can be self‑hosted (your own servers)
• Allows full customization and fine‑tuning
• Lower cost in long-term
• Community-driven improvements
• May have weaker safety or guardrails depending on version

Concept of Model

Foundation models are generalists that can be specialized for specific tasks . this specialization fon turns a powerful general tool into  a highly effective business solution.

• Gemini Nano
• Gemini Pro
• Gemini Ultra

Role of the Model Garden

Model Garden on Vertex Ai is a centralized hub where you can discover test and deploy a wide variety of models. it simplifies model selection and deployment, giving you a single to manage all AI models your organization uses. A key tool for any Gen AI leader

Difference between Parameters vs Hyperparameters

1. Parameters (Model Parameters)

These are learned automatically by the model during training.

Examples

• Weights and biases in neural networks
• Coefficients in linear regression
• Centroid positions in K‑Means
• Decision boundaries learned by an SVM

Key Points

• The model adjusts these values to minimize loss.
• They directly determine the model’s predictions.
• They change every epoch/iteration during training.

2. Hyper parameters

These are set by you before training begins.

Examples

• Learning rate
• Number of layers / number of neurons
• Batch size
• Number of trees in Random Forest
• k in k‑NN
• Regularization strength (λ)

Key Points

• They are not learned by the model.
• They control how the model learns.
• Tuned using methods like Grid Search, Random Search .

Hyper parameter: Learning Rate

• Controls how much model adjusts parameters
• Too high: model overshoots and never settles
• Too low: training tasks forever
• Finding perfect learning rate is critical for fine tuning

Batch Size

When training a machine learning model, it learns by looking at examples. But it doesn’t look at all examples at once and it doesn’t look at them one by one .

What is Batch Size?

Batch Size = how many training samples the model sees before it updates its internal parameters.

• You can read one page and take notes (tiny batch)
• or read a whole chapter and then take notes (big batch)
• or read the whole book and then take notes (very big batch)

A batch is just “how much stuff you process before making an update.”

 Small Batch Size (16, 32)

Pros

• Uses less memory (works on smaller GPUs)
• More frequent updates → model adapts faster
• Often generalizes better (a little randomness helps learning)

Cons

• Training is noisier (updates bounce around)
• Might take longer to converge

Analogy: Taking notes after every few pages → fast feedback but sometimes messy.

Large Batch Size (256, 512, 1024)

Pros

• Training is more stable and smooth
• Faster on big hardware (GPUs/TPUs love large batches)
• Good for very large datasets

Cons

• Needs more memory
• Can overfit or get stuck in worse solutions
• Sometimes ends up with poorer generalization

Analogy: Reading a whole chapter before taking notes → smooth but sometimes harder to learn effectively.

Role Of the Epoch :

• One epoch =one complete pass-through entire training dataset
• Fine tuning runs for specific number of epochs
• Too few epochs: insufficient learning
• Too many epochs : risk of over fitting to train data 

Overfitting :

• Training: High performance
• Real world :poor performance 

Underfitting  :

• Training : poor performance
• Real world  :poor performance

Hyper parameters Tunning on Vertex Ai

• Vertext Ai vizier automates hyperparameter search
• Smart algorithms find optimal settings
• Eliminnate manual trial and error testing 
• Save significant time and computing costs  .

Importance of Hyperparameters 

• Hyperparameters directelly impact cost and quality 
• Optimization reduces fine tanning time and expense 
• Enables higher performing specialized model
• Critical for informed resource allocation decision
• Ensures efficient and effective Gen Ai initiatives.