Contents

Eight sections. Thirty-two notes.

Written in public, a section at a time. Notes in black are written; grey ones are next. Sections one to three are the foundation; after that, follow your curiosity.

do a practical to back the theory.

Opening
1. 00Why this notebook exists
Section 1 What is actually happening when you type a question
Section 2 Why it can be confidently wrong
1. 05Hallucination is the same mechanism that makes the model useful Practical 05
2. 06Why a system that writes poetry struggles to multiply Practical 06
3. 07The confidence is in the prose, not in the knowledge
4. 08Sycophancy: models tend to agree with you
5. B2Build 02 · The ten-minute review of an AI-written draft
Section 3 Context: the most important word nobody defines
1. 09A context window is a whiteboard, not a memory
2. 10Lost in the middle: where in a long document the model reads worst
3. 11The system prompt: the instructions you never see
4. 12The model is not the product
5. B3Build 03 · Ask a long document. Then check the answer. written early
Section 4 Giving the model knowledge
1. 13Retrieval: a closed-book exam versus an open-book one
2. 14Embeddings: how a machine decides two sentences mean the same thing
3. 15Fine-tuning changes how a model behaves far more than what it knows
4. 16Retrieval, fine-tuning, or a longer prompt: what each is good at
5. B4Build 04 · The same helper, for a document too long to paste
Section 5 Agents: what changes when the model can act
1. 17An agent is a loop: think, act, observe, repeat
2. 18Why errors compound: 95% per step across ten steps is a coin flip
3. 19Prompt injection: the security problem everyone outside security is ignoring
4. 20Human-in-the-loop is a design decision, not a slogan
5. B5Build 05 · An agent that drafts but never sends
Section 6 How do you know it's working?
1. 21"It looks good" is not evidence: what an eval actually is
2. 22Building a twenty-question eval set
3. 23Benchmarks: what they measure, why leaderboards leak
4. 24The model under your product changed and nobody told you
5. B6Build 06 · A twenty-question eval for one task you actually do
Section 7 The economics of a token
1. 25What inference actually costs, and why an answer is priced differently from a question
2. 26Why reasoning models are slow
3. 27Small models versus large: why bigger isn't always the answer
4. 28What "open" actually means: weights, data, source
5. B7Build 07 · A cost sheet for one real workflow
Section 8 How the model got its manners
1. 29Pretraining versus post-training: where knowledge comes from, where behaviour comes from
2. 30RLHF in plain language: why models sound the way they do
3. 31Reasoning models: what happens in the "thinking", and what doesn't
4. 32Why the model is better in English, and what that means for Indian languages
5. B8Build 08 · A ten-question check in your own language