In this lesson
Trace a sentence yourself
Take a new phrase through one coherent toy, show how trained rows transfer beyond exact counts, and separate a Transformer step from the system around it.
By the end, you can
- Trace a new phrase through tokens, vectors, attention, block output, logits, and decoding.
- Distinguish trained weights, runtime context, one Transformer forward pass, and the serving or application layer around it.
- Explain how shared trained rows can transfer evidence beyond exact-context counts in this toy.
Before you start: Confident, wrong, and augmented
For thirteen chapters, eat an orange held your hand. The tracer now lets go. Three unseen phrases and a composer wait behind one final door; every displayed number belongs to one coherent toy computation.
Your job: narrate the causal chain, bet by bet, without borrowing the answer from the old sentence.
Observe the whole challenge
The final door opens onto two lanes. The inner one is a single Transformer forward pass, tokens to logits. The outer one is serving and application code: it assembles context, decodes, runs tools, repeats steps, and stops.
“The AI did it” no longer earns partial credit. Name the responsible layer.
Predict the boundary after logits
Your bet
A Transformer forward pass has produced logits for one position. Which job still happens outside that pass?
Commit before manipulating the instrument. This choice is final.
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.
1. Text
A human string enters the wrapper. The model still cannot read characters directly.
Coherent toy: the block is frozen and all vocabulary rows were trained on chapter 8’s 22 visible examples. New phrase vectors remain handcrafted. This shows transfer inside the fixture, not real-language quality.
Manipulate a brand-new trace
Choose one phrase, or compose your own context from the token buttons. Before revealing a stage, say which representation should appear next. Then open it and connect the visible numbers.
- A human string enters the wrapper as text.
- The tokenizer emits supported IDs and lookup returns one vector per ID.
- Causal attention mixes information from available positions.
- The final position leaves the block as one context vector.
- Vocabulary dot products become logits and softmax shares.
- A decoding policy chooses a token; serving or application code appends it and requests another pass.
Stages 5 and 6 stay locked until you correctly name the winning token from the block output. Commit before the logits appear.
Run a second phrase. No exact count row exists for either string, yet the same trained vocabulary rows score both final vectors. The values change; the operations and weights hold.
This is transfer inside the toy. Its eight vocabulary rows were trained on 22 visible examples, then frozen. The new phrase vectors remain handcrafted, and no answer-specific rule was added for the finale.
Shared rows can act similarly on nearby representations - evidence about this fixture, not about real-language competence - so an exact string the toy never saw can still attract a useful bet.
Explain both loops from memory
Close the tracer. In two sentences, one per lane, explain how one forward pass turns context into logits and how serving plus application code turn repeated steps into a chatbot.
Write before opening the rubric. Matching its wording does not matter; assigning every stage to the right owner does.
Explain it without the course
Two sentences. One inner model, one outer chatbot.
Close the tracer. Explain how one forward pass produces logits, then how serving and application code turn repeated steps into the chatbot you use. Write before comparing.
Now test transfer. Diagnose four new situations, 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?
Pick an answer first.
Connect the course to your next conversation
When you open a real assistant, watch for the two lanes. The outer one packs messages, roles, history, evidence, and intermediate state into runtime context; the inner one turns that visible past into logits for one position.
Serving code decodes those logits and may repeat the pass; an API can bundle many such steps into one request. Validating evidence, searching, running code, calling tools, enforcing policy - that work commonly sits in the application.
Under the whole product remains lesson 1’s wager: given the available past, which token comes next?
The cat from chapter 1 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 production mechanism is introduced here; the chapter synthesizes lessons 1-13 and their cited sources.
- The finale trains all eight vocabulary rows by gradient descent on chapter 8’s 22 visible examples, then freezes them. The three new phrase vectors remain handcrafted inputs.
- The unseen phrases test whether shared trained rows act on new representations inside this toy. They do not establish generalization quality for a production language model.
- Passing the in-product tasks records demonstrated progress; the moderated protocol in
docs/learner-sessions.mdremains the learning-efficacy gate. - Content and claims reviewed on July 28, 2026.