Ten Thousand Nudges: How Training Works
By the end you'll be able to
- Describe training as: guess, compare, nudge, repeat, at vast scale.
- Explain that nobody writes the rules: they condense out of practice, then freeze when training stops.
- Distinguish “the model was trained on X” from “the model learns from me”.
- Predict
- Train
- Decide
- Field note
⚡ 30-second warm-up: from Lesson 3, why did Kerry's 80-page guideline hit a limit, and what was the risk?
Everything is metered in tokens. Past the budget, the tail can be silently dropped. A summary of half a document then looks identical to a summary of all of it.
Watch a machine get better at prediction without anyone programming the answers in.
Start with an impossibility
Nobody taught the model grammar. Nobody could: there are too many rules, and most of them we don't even know ourselves. Yet there it is, writing flawless sentences.
Commit to a prediction
How did the rules get in there?
This lesson changes one thing: you'll stop thinking "someone put the answers in". Instead, picture billions of tiny corrections during training, condensing into rules. Once training stopped, the dials froze.
Discover the mechanism
Take Lesson 1's machine, but set all its internal settings to random. Show it a fragment of real text with the ending hidden. It guesses the next token, badly. Compare that guess with the real continuation. Then nudge every internal setting a tiny amount, in the direction that would have made the true token slightly more likely.
One nudge achieves almost nothing on its own. But repeat it across trillions of fragments, on warehouse-scale computers, over weeks. The nudges compound into everything you saw in Act 1: grammar, style, facts-as-patterns, code. This is training.
Two consequences matter at work. First, once training stops, the settings freeze, so answering your questions later doesn't change them. Practice is over: the machine is only performing. Second, the examples themselves aren't stored inside. What remains is the skill they built, in the way your old maths homework is gone, but the arithmetic stays with you.
Real training pipelines run in stages. Broad practice on everything comes first, then specialist practice, then coaching from human feedback about which answers people prefer (more in this lesson's Go deeper). "Reasoning" models get extra practice at step-by-step working. The practise-nudge-freeze core is what matters most here; Lesson 6 opens the machine to show what the "settings" actually are.No one can tell you the physics of a good delivery. You bowl. You see where the ball lands. You adjust your grip, just slightly, and do it again, ten thousand times. Skill builds up in your body, not as a set of memorised instructions.
- Each delivery = one prediction attempt.
- Where the ball landed vs the target = comparing the guess with the real next token.
- The tiny grip adjustment = updating the internal settings.
- Muscle memory = the trained model. The coach choosing the drills = the builders choosing the training data, a human decision with human accountability.
Train a machine yourself
This is a real, tiny learner, not a canned animation. It runs in your browser, starts out knowing nothing, and learns letter patterns from a small text as you feed it. It works on the same principle as a frontier model, just roughly a hundred million times smaller. Nothing leaves this page.
Its writing:
Where this bites at work
Yesterday Kerry corrected her assistant's error about a supplement interaction. Today, in a fresh chat, it repeats the same mistake. It isn't being stubborn. Her correction never touched the frozen dials; it lived only in that conversation's context, and that context is gone now.
Using a model is not training it. Whether a provider uses conversations to train future models is a policy setting worth checking. For Kerry there's a second layer too: client health details count as sensitive information under Australian privacy law. So what she types is a disclosure decision regardless of any training policy (Lesson 19 builds the full toolkit).
A bowler wants to improve and knows what the target is. The model has no goals: only a mathematical procedure that reduces error. And "muscle memory" is a body metaphor. Nothing in the machine feels practised, or remembers practising. Keep the mechanism, drop the mind.
Now make the call
Kerry's assistant repeated yesterday's corrected error in today's fresh chat. She has ten minutes before her next client. What does she do?
Try 'correct it again' too. Understanding why it fails makes the freeze visible.
Apply it to your work
Find the instruction you keep re-typing every session, such as your house style, your role, or your programme's context. Put it into your tool's saved-instructions feature instead. That's the correct fix for a machine that never learns from you.
Prove it to yourself
Two quick questions, untimed and retryable. They count toward ★ Mastered (completed, decision, and self-check all correct). This is what stops the feeling of "that made sense" evaporating by next week.
Explain it in your own words
Write two or three sentences for a colleague. Then compare against the rubric below; it checks ideas, not wording. Only you see your note.
✓ Lesson 5 stamped: Emergence Witness
You trained a machine with your own hands and watched skill condense out of practice. You also know the dials froze before you ever typed a word. Lesson 6 opens the box to show what those dials actually are.
Evidence & review — how we know what this lesson claims
| Claim | Type | Basis | Review risk |
|---|---|---|---|
| Training iteratively reduces prediction error over example data (gradient-based learning) | How it works | Primary research literature | Low |
| Deployed model weights are frozen; conversations do not update them in-session | How it works | Architecture documentation | Low |
| Whether providers train future models on user conversations is a per-product policy setting | Product behaviour | Provider policies | High — reviewed quarterly |
| The in-page learner | Teaching device | Real n-gram learner, labelled as vastly smaller than a neural model | Medium |
| The bowling-by-feel analogy | Useful mental model | Limits stated in the lesson | Low |
Concept review due: January 2027. Product-behaviour claims: quarterly.
Where you are in the machine
You're inside the engine, watching how its dials got set. Select any layer to see its role, its lessons, and your live progress.
The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).
Next: what exactly are the “dials”?
Practice nudged billions of “internal settings”. But settings of what? Lesson 6 opens the machine and finds not a brain, but stacked layers of astonishingly simple arithmetic.