GENAI ✦ NOTES
0 / 0
GENERATIVE AI ✦ FUNDAMENTALS

Generative
& AI

How machines learned to create. Tap anything with a   to reveal its explanation, one idea at a time.

01 — FUNDAMENTALS

The basics

Generative AI is a class of models that learn patterns from data and generate new samples that follow those patterns. Tap each card to uncover one core idea.

02 — WHERE IT FITS

The layers

Each field sits inside the one before it. Open them from the outside in, like unpacking a set of nesting dolls.

Artificial Intelligence

TAP
Machines doing tasks that normally need human intelligence. That includes hand-written rules, not only learning.

Machine Learning

TAP
Systems that learn patterns from data instead of being explicitly programmed. Generative models are part of machine learning too.

Deep Learning

TAP
Machine learning built on neural networks with many layers.

Generative AI

TAP
Models that generate new content such as text, images, audio, video and code. Modern generative AI mostly uses deep learning (transformers, diffusion models, GANs, VAEs), although simpler generative models existed long before deep learning.
03 — CONTRAST

Two mindsets

Start with the picture: play with it to see each approach in action (the data is simulated for teaching). Then guess how each side would answer the questions below and tap to check.

04 — ARCHITECTURES

Model zoo

Five families of generative models, which work in quite different ways. They also overlap: LLMs are a kind of autoregressive model, and some systems combine families. Pick one to see how it works, what it makes, and examples.

↑ Pick a model to open its card
05 — MECHANICS

Under the hood

A model's life has two phases. Switch between them, then open each question below.

SIMULATED DEMO · NOT A REAL MODEL

Be the language model

A simplified, simulated next-token exercise. The probabilities below are predefined, illustrative values written for teaching. They do not come from a real language model or from any training data. A real LLM calculates a probability for every token in its vocabulary, using its trained parameters and the full preceding context. Tap a word, or let it sample.

Why multiply the probabilities? TAP
P(x) = ∏t=1…n P(xt | x1, …, xt−1)

The probability of a whole sequence of tokens is the product of each token's probability given the tokens before it. That's the chain rule of probability, and it holds exactly. Autoregressive LLMs are trained to estimate each of these conditional probabilities. Pretraining typically means doing this over very large amounts of text: predict the next token, compare with the actual one, adjust the parameters. The text supplies its own targets, which is why this is called self-supervised learning. In the demo above, each step's number comes from a hand-written table, so the product is illustrative only. (Encoder models like BERT pretrain differently, by filling in masked words. They are built for understanding tasks such as search and classification rather than for generating text.)

06 — END TO END

The pipeline

From data to the apps you use: a simplified, typical sequence. The exact pipeline varies between models, and stages can repeat or overlap. Swipe and tap each stage.

07 — RECALL

Test yourself

Nine questions covering everything above. You get an explanation after each answer.