Topics
21 topics, explained simply — the concept first, then the mechanics, the maths, and a worked example.
Introduction to Generative AI – Use Cases and Types
What generative AI actually does, the main model families, and where it's used in the real world
Generative vs Discriminative Models
Learning the joint distribution P(x,y) versus learning the decision boundary P(y|x) directly
Variational Autoencoders (VAE) – Concept, Architecture, and Loss
A probabilistic autoencoder that learns a smooth latent space you can actually sample from
Generative Adversarial Networks (GAN) – Concept and Applications
Two networks locked in a game — one forging fakes, one hunting them down — until the fakes win
GAN Architecture
Inside the Generator and Discriminator networks, layer by layer, for one training iteration
GAN Training Challenges and Loss Functions
Why GANs are hard to train, and the practical tricks that keep them from falling apart
Conditional GAN and Applications
Feeding a GAN a label so you can control exactly what it generates, instead of leaving it to chance
StyleGAN & Modern GAN Variants
Separating 'what to generate' from 'how to control it' via a mapping network and per-layer style injection
Diffusion Models (DDPM, Stable Diffusion) – Overview and Applications
Learn to destroy an image with noise, then learn to undo it one step at a time
Practical Hands-on: GAN or VAE Implementation
What the training loop actually looks like, and the three bugs that bite everyone once
Introduction to Transformers – Motivation for Generative AI
Why generative AI gave up on recurrence and bet everything on attention
Encoder Architecture
Embed, attend, feed-forward, normalize — repeat N times
Prompt Engineering – Zero-shot, One-shot, Few-shot & Chain-of-Thought Prompting
Same task, four prompts — how much example and reasoning scaffolding you give changes everything
Embeddings and Vector Representations for LLMs
Turning words into points in space so 'similar meaning' becomes 'nearby vectors'
Retrieval-Augmented Generation (RAG) – Architecture and Applications
Grounding generation in retrieved documents instead of relying only on frozen model weights
Self-Attention – Concept, Step-by-Step Example, and Hands-on Visualization
Every token looks at every other token and mixes their values by relevance
Multi-Head Attention
Several smaller attentions run in parallel so the model can learn different kinds of relationships at once
Positional Encoding
Injecting order back into a mechanism that only sees content, not sequence position
Decoder Architecture
Masked self-attention, cross-attention, and feed-forward, stacked and repeated one token at a time
Cross-attention
Query from the decoder, key and value from the encoder — how the decoder reads the full input
Transformer Training and Scaling Challenges
Quadratic attention cost, huge compute needs, and fragile optimization at scale — and how to manage them