Layers of Sieves: Neural Networks
By the end you'll be able to
- Describe a neural network as layers of simple units, each transforming its input and passing it on.
- Explain that weights are the adjustable settings training nudged — billions of them, none meaningful alone.
- Explain why a specific wrong “fact” can't simply be found and deleted inside the model.
- Predict
- Probe
- Decide
- Field note
⚡ 30-second warm-up — From Lesson 5: Kerry corrected the chatbot yesterday. Why is the same error back today?
The dials are frozen — her correction lived only in that conversation's context, which is gone. Using a model is not training it.
Open the machine and find not a brain, but stacked layers of astonishingly simple arithmetic.
Start with a misleading word
"Neural network" makes you picture a brain. What's actually inside is closer to an enormous spreadsheet: multiply, add, pass it on. So here's the puzzle — how could stacked arithmetic tell sincerity from sarcasm?
Commit to a prediction
Where does "sarcasm detection" live inside the machine?
This lesson changes one thing: you'll stop picturing a digital brain with rules inside and start picturing layers of simple arithmetic with billions of learned dials — where no single dial holds any rule at all.
Discover the mechanism
Inside the machine, your tokens become long lists of numbers (much more on this next lesson). Those numbers flow through dozens of layers. Each layer does humble arithmetic: multiply by stored settings, add, keep what's positive, pass it on.
The stored settings are the weights — the dials Lesson 5's practice nudged. Billions of them. No dial means anything on its own, and no engineer set any dial by hand: each holds whatever value made prediction slightly less wrong across trillions of examples.
What stacking buys is depth of abstraction: early layers respond to surface features of text; deeper layers to things like tone, topic and structure. "Neural" honours a loose biological inspiration from the 1940s — it is not a claim that a brain lives in the box.
"each layer detects more abstract features" is a deliberately tidy story — real layers mix roles, and reading what a given layer truly does is an active research field, not a solved one (this lesson's Go deeper). The missing ingredient — how words share information across a sentence — is Lesson 8's famous mechanism.Progressively finer sieves: gravel out first, then grit, then flecks — gold at the end. No single sieve finds gold; the sequence does.
- Each sieve = a layer. Its mesh size = weights, set by practice (Lesson 5).
- Material flowing through = the evolving representation of your text.
- Gold = the final candidate scores (Lesson 4's raffle tickets).
- Where it breaks: sieves only discard; layers can construct new features from combinations. And one sieve line serves every topic at once — no cooking-sieve vs legal-sieve.
Probe the layers yourself
Conceptual illustration — a simplified demonstration, not readings from a real model. Two nearly identical sentences enter the stack. Drag the depth probe and watch when — and where — the machine's internal picture of them separates.
A: “That's great — a payrise!” B: “That's just great — another Monday.”
Layer 1: nearly identical — both are cheerful-looking words about ordinary things.
Where this bites at work
Maya's compliance team asks: "Show us the exact rule the assistant used to word that advice summary." She can now explain why that request, taken literally, may have no answer: the behaviour lives spread across billions of interacting dials, not in a rule that can be printed.
What a professional offers instead: evidence about behaviour — systematic testing, guardrails, documented limits and human review of consequential outputs. Process assurance where rule-extraction isn't available (Lesson 21 completes this).
What remains uncertain Don't over-learn "nobody can read the dials": interpretability research is making real progress, and targeted edits exist in the lab — they're just unreliable at frontier scale today. The durable claim is narrower: per-answer rule extraction is not generally available, so consequential use needs process safeguards, not promised explanations.
Now make the call
Maya's compliance lead insists: “Before launch, give us the rule the model applies when it words fee disclosures.” The launch review is Friday. What does Maya do?
Try the 'ask the model to explain itself' option — the introspection trap gets a corporate wardrobe in this lesson.
Apply it to your work
Draft the two-line answer you'd give a board member who asks “why did the AI word it that way?”: honest about the limit (no per-answer rule readout), concrete about the assurance (systematic testing plus named human review). Keep it in your checklist starter.
Prove it to yourself
Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct). This is what defeats the feeling of "that made sense" evaporating by next week.
Explain it in your own words
Two or three sentences for a colleague. Then compare against the rubric — it checks ideas, not wording, and only you see your note.
✓ Lesson 6 stamped: Layer Prober
You watched two near-identical sentences part company layer by layer — and you can now explain to anyone why there's no rule book inside the box. Lesson 7 shows where meaning lives instead.
Evidence & review — how we know what this lesson claims
| Claim | Type | Basis | Review risk |
|---|---|---|---|
| Neural networks compute via layered weighted transformations | How it works | Primary research literature | Low |
| Deeper layers respond to progressively more abstract features | Useful mental model | Simplified from interpretability research; stated as tidy story | Medium |
| Per-answer rule extraction is not generally available at frontier scale | How it works (current state) | Interpretability research status | High — field moving fast |
| The depth-probe demonstration | Teaching device | Simulation, labelled; real probing is open research | Medium |
| The sieve-line 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
Still inside the engine — now you know its substance: layers and dials. Select any layer for details and 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: the most beautiful idea in the box
Your tokens become “long lists of numbers”, we said. Those lists are one of the loveliest ideas in AI — numbers that behave like meaning. Lesson 7 hands you the map.