verbatome
Verbatome / Chapter 12IV · From model to chatbot12 min read · ~30 min hands-on

In this lesson

Trace a sentence yourself

Take a brand-new phrase through one coherent toy model, recover the whole course from memory, and separate the model from the chatbot around it.

By the end, you can

  • Trace a brand-new phrase from text through tokens, vectors, attention, block output, logits, and decoding.
  • Distinguish trained weights, runtime context, one model call, and external orchestration in a chatbot.
  • Explain why shared representations can transfer evidence beyond exact-context counts.

Before you start: Confident, wrong, and augmented

For eleven chapters, one sentence held your hand: eat an orange carried you from counting to attention to training. You do not need it anymore - and this chapter proves it. The tracer holds three phrases the course has never touched, plus a composer for building your own context. Every attention weight, logit, and generated token comes from one coherent computation, and your job is to narrate all of it.

Observe the whole challenge

Your job has two lanes. First, explain the inner model from text to next token. Second, explain the outer chatbot that formats messages, retrieves evidence, calls the model repeatedly, and decides when to stop. Never use “the AI” as one foggy box when you can name the responsible layer.

Predict the boundary after logits

Your bet

The Transformer has produced logits for one position. Which job still happens outside the model itself?

Nobody grades this. Commit, and let the instrument settle it.
Pipeline tracer / brand-new transfer task

Choose one of three genuinely new phrases, or compose your own context. Every number comes from one trained vocabulary, one frozen block, and one decoder.

Context source

1. Text

taste an

A human string enters the wrapper. The model still cannot read characters directly.

Coherent toy computation. The block is frozen; all vocabulary rows were trained by gradient descent on chapter 8’s same 22 visible examples. New phrase vectors remain handcrafted, so this demonstrates transfer without claiming real-language quality.

Manipulate a brand-new trace

Choose a phrase. Before opening each stage, say what representation should appear next. Then reveal it and connect the visible numbers:

  1. Text is split into supported token IDs.
  2. Lookup rows and position signals enter the residual stream.
  3. Causal attention mixes only available positions.
  4. The final position leaves the block as one context vector.
  5. Dot products with vocabulary rows become logits and softmax shares.
  6. A decoding policy selects one token, appends it, and repeats.

Run a second phrase. The values change, but the operations do not. That invariance is the transferable mental model this course has been building.

Explain both loops from memory

Final challenge Close the tracer. In two sentences, explain to a friend how an LLM turns text into its next token and how that model becomes the chatbot they use. Produce your explanation before you see the model answer; matching its wording does not matter, but every responsible stage does.

Explain it without the course

Two sentences. One inner model, one outer chatbot.

Close the tracer. Explain how text becomes the next token, then how repeated model calls become the chatbot you use. Write from memory before comparing.

0 characters committed. Write at least 80 characters so both loops have room to appear.

Now test transfer rather than merely recognizing a paragraph. Rebuild the six-stage mechanism and sort every system responsibility into its proper layer.

Final performance deck · task 1 / 6

Four transfer decisions, one pipeline reconstruction, and one system-boundary sort. Every task must be solved.

A colleague says the assistant “looked up Lyon in its weights” after you pasted a museum registry. What actually changed?

Reflect with one synthesis

Checkpoint

Which description correctly separates a chatbot's three main levers?

Which description correctly separates a chatbot's three main levers?

Connect the course to your next conversation

When you open a real assistant, watch the layers. Your messages and chat history fill a finite context window. A template marks their roles. The model emits one conditional distribution at a time. A decoder chooses tokens. The application may search, run code, call tools, check policy, or ask the model again. Under all that product behavior remains the game from lesson 1: given the visible past, what token comes next?

The cat from chapter 2 can finally retire. You no longer need a fixture sentence, because you can now trace any sentence - including this one.

Sources and scope
  • No new mechanism is introduced here; this chapter synthesizes chapters 1-11 and their cited sources.
  • The finale trains all eight vocabulary rows by real gradient descent on chapter 8’s same 22 visible examples, then freezes them. The three new phrase vectors remain handcrafted teaching inputs.
  • Passing the in-product tasks records demonstrated progress; the moderated protocol in docs/learner-sessions.md remains the learning-efficacy gate.
  • Content and claims reviewed on July 20, 2026.