The Ilmverse Dialogue ← Resources

How does Generative AI work?

The machine
that learned
to write.

scroll
The Ilmverse Dialogue Resource 01

01 / 12  ·  The prompt

You type. It answers.

Everything between those two moments is the story.

>

02 / 12  ·  Tokens

First, it's shattered.

The model reads tokens, not words. Fragments, each with a numeric ID.

03 / 12  ·  Embeddings

A word's meaning comes from its neighbours.

The model has seen every token in millions of sentences. From the words it tends to sit beside, it builds a vector, a list of numbers that encode what this word means.

poem
about

A list of numbers. Each one a learned shade of what this word means.

04 / 12  ·  Similarity

Similar ideas, similar fingerprints.

Words we use interchangeably end up with nearly identical vectors.

05 / 12  ·  Space

Meaning becomes geometry.

Stack those vectors together and patterns emerge. Weather words cluster, food words cluster, emotions cluster.

06 / 12  ·  Attention

Every word watches every other.

A transformer reads the whole sentence at once. Each token decides which others matter, all in parallel.

07 / 12  ·  Context

Same word. Different meaning.

Attention is how the model tells the difference.

08 / 12  ·  Depth

One pass isn't enough.

The same sentence flows through the attention machinery dozens of times. Each pass refines the picture. Grammar, tone, reference, agreement, all emerging as the stack climbs.

richer representation
plain embeddings

09 / 12  ·  Prediction

All of it, to guess one word.

A probability over every token in the vocabulary.

The cat sat on the

10 / 12  ·  Temperature

A dial between safe and strange.

That probability distribution has a knob. Turn it down and the model picks the most likely word every time. Turn it up and it starts gambling.

balanced 0.70
0.0   deterministic 1.8   chaotic
The cat sat on

Same model. Same prompt. Same probabilities. Just a different appetite for risk.

11 / 12  ·  Generation

Predict. Append. Repeat.

No plan. No outline. Just the next word, over and over.

12 / 12  ·  Limits

Which is why it lies so well.

The model knows what text usually follows what text, not what's true.

plausible & true
"The Eiffel Tower was completed in 1889."
plausible & false
"It was dismantled during WWII and rebuilt in 1946."

A map of language.

Not a mind. Not the world. Just the patterns, and like any map, most useful when you remember it isn't the territory.

Take the AI Readiness Quiz

■   end   ■