2. The Russian Doll Analogy
Think of Russian Matryoshka dolls — those charming wooden dolls that nest inside one another, each smaller than the last.
- Outermost Doll: Artificial Intelligence (AI)
- Next Layer: Machine Learning (ML)
- Deeper Still: Deep Learning (DL)
- At the Core: Generative AI
They're not competing technologies — they're layers of intelligence, each building upon the last.
3. The Outermost Layer: Artificial Intelligence
Artificial Intelligence (AI): This isn't a single technology — it's an umbrella term covering any technique that enables machines to mimic human intelligence.
The Ultimate Goal
Build machines that can think, reason, understand language, and solve problems — sometimes even better than humans can.
6. AI in the Art World: A Debate
Théâtre d'Opéra Spatial
In 2022, the painting that won first prize at the Colorado art exhibition wasn't created by a human, but by an AI called Midjourney.
Questions it raised:
- Can we call an AI-created work 'art'?
- Who is the true artist?
- Is creativity exclusive to humans?
8. Real World Example: Spam Detection
The Traditional Approach Fails: Create a rule like "block emails mentioning 'prize'" and scammers simply write "pr!ze" or "p-r-i-z-e" to bypass it.
The Machine Learning Advantage:
- "This person has never sent you an email before."
- "This email came from Nigeria at 2 AM."
- "There are lots of dollar signs ($$$) in the subject."
It synthesizes these patterns into a decision: "This looks like spam!"
11. The Hidden Labor Behind AI
Here's the dirty secret: The hardest part of supervised learning isn't fancy algorithms — it's the tedious human work of data labeling.
Example: Autonomous Vehicles
Teams of people manually draw bounding boxes around pedestrians, cars, and traffic signs in millions of video frames. It's painstaking work.
Behind every "smart" AI is an army of human labelers, often working through platforms like Amazon Mechanical Turk.
14. Reinforcement Learning: Trial, Error, Reward
The AI learns through experience — like a child discovering that touching a hot stove hurts. Every action has consequences.
Example: Autonomous Taxi
- Drop off passenger safely → +10 points
- Run a red light → -100 points
- Through millions of simulations, it learns which behaviors maximize rewards.
19. Vanishing Gradient Problem
The "Telephone Game" Problem
During training, the learning signal (gradient) must travel backward through all layers. But with each layer, it gets weaker and weaker until it vanishes into nothing.
Remember the "telephone game" where a message gets garbled as it passes through many people? Same problem here!
This roadblock stalled deep learning for decades. Modern fixes like ReLU activations and LSTM cells finally cracked it.
20. GANs: A Creative Duel
Generative Adversarial Networks
Ian Goodfellow's 2014 breakthrough: teaching AI to create by making two neural networks compete against each other!
- Generator (The Forger): Creates synthetic data trying to pass as real.
- Discriminator (The Detective): Tries to spot the fakes from the real deal.
21. The Training Arms Race
How They Improve Together
Round 1: Generator creates a crude blob. Discriminator easily spots it as fake.
After thousands of rounds: Generator gets better at faking. Discriminator sharpens its detection. Eventually, the fakes become indistinguishable from reality!
The Dark Side
- Deepfakes: Frighteningly realistic fake videos that can manipulate public opinion.
- Mode Collapse: Sometimes the generator gets stuck creating variations of the same thing over and over.
22. Transformers & Large Language Models
The architecture powering ChatGPT, Gemini, and the AI writing revolution.
- Next-Token Prediction: At its core, it's just predicting the next word based on what came before. Simple idea, profound results.
- Attention Mechanism: The breakthrough that lets AI understand context. It learns which words to "pay attention to".
Example: In "The animal didn't cross the street because it was too tired", attention helps determine "it" refers to the animal, not the street.