Chapter 2 · How LLMs actually work

It's autocomplete.
Just… enormous.

Next-token prediction, temperature, and the context window — the humble trick every agent is built on.

The one-sentence version

An LLM is autocomplete, scaled up absurdly.

It doesn't "know" facts or "understand" your question the way you do. Underneath, it does one thing, over and over:

predict the next token.

Everything an agent can do — tools, memory, reasoning — is built on top of this single move.

Step zero

Models don't read letters. They read tokens.

  • Your text is chopped into tokens — often whole words, sometimes word-pieces.
  • Each token maps to an integer ID, because a neural net only does math on numbers.
  • A space usually rides along with the word after it.

"Attention is all you need!" → 7 tokens → 15139 · 2939 · 318 · 477 · 345 · 761 · 0

The shape of the machine

Tokens in → one token out.

Goes in

The whole sequence so far, as integer IDs — your prompt, plus everything generated up to now.

Comes out

A single guess: a probability score for every possible next token — often 100,000+ of them.

One pass of the model = one token's worth of prediction.

Inside a single guess

A probability over every token. Pick one.

"The cat sat on the ___" — the model's top candidates for the next token:

mat 41% floor 15% sofa 11% rug 9% table 8% …plus ~100,000 more tokens sharing the last 16%

The whole engine

Predict. Append. Repeat.

  • Pick a token from the distribution and stick it on the end.
  • Feed the whole longer sequence back in and predict the next one.
  • This loop is called autoregression — and it's all a language model does.

A paragraph is just this step, run a few hundred times.

The catch

It predicts likely — not true.

Because it only ever chases likely-sounding text, an LLM can be completely confident and completely wrong. That's a hallucination.

Tools

Let it check reality instead of guessing.

Memory

Give it facts it can look up, not invent.

Guardrails

Catch the confident-but-wrong answers.

This one fact is why the rest of this course exists.

Dial #1

Temperature: the creativity knob.

The model hands you a distribution — but how boldly should it pick? Temperature reshapes those probabilities before sampling.

  • Turn it down → the top choice dominates.
  • Turn it up → the odds flatten out.

Same model, same prompt — only the boldness changes.

Two extremes

Low temp vs. high temp.

low · 0.1

Focused

Always grabs the front-runner. Safe, repetitive, near-deterministic. Ask twice, get the same answer.

high · 1.5

Creative

Long-shots get a real chance. Surprising, varied, riskier. Ask twice, get two different answers.

Low temp ≈ a calculator. High temp ≈ a brainstormer.

Dial #2

The context window: its short-term memory.

A model can only "see" a fixed number of tokens at once. Add more, and the oldest fall off the edge:

dropped plan context window a trip to Japan in spring now ← older newer →

Why it forgets you

The model has no memory of its own.

  • It isn't remembering your earlier messages — it's re-reading whatever still fits in the window.
  • Slide past the window and that context is gone for good.

The fix: give the model an external memory and a search tool.

retrieval-augmented generation (RAG) · vector databases · the agent loop

Chapter 2 in one breath

Tokens in, one token out — on a loop.

An LLM tokenizes your text, predicts the next token from a probability distribution, and repeats. Temperature decides how boldly it picks; the context window decides how much it can see.

It has no memory of its own — and closing that gap is what turns a language model into an agent.

Next up

Chapter 3 · Chatbots vs Agents.

We've got a next-token predictor with two dials and no memory. Now we give it goals, tools, and a loop — and watch a chatbot become an agent.

Press S for speaker notes · F for fullscreen · to advance.