Beyond the ChatbotOne Layer Deeper
Syllabus

Eight arcs. Thirty-two notes. One idea at a time.

Arcs one to three are the foundation and run in order. After that the sequence may move with what readers ask for. Each note comes with a small experiment you can run yourself with whatever AI tool you already use.

do a practical to back the theory.

Arc 1 · September – October 2026

What is actually happening when you type a question

  1. A model doesn't know things. It predicts what comes next.Experiment: give a model the same half-sentence five times and watch the continuations diverge.
  2. Tokens: why the model reads "strawberry" differently from you.Experiment: count the r's in strawberry, then see how a tokeniser splits the word.
  3. Why the same question gives different answers: temperature and sampling.Experiment: one prompt, five runs, at two temperature settings.
  4. The model is a snapshot: training cutoffs.Experiment: ask about last month, with and without web search.
Arc 2 · October – November 2026

Why it can be confidently wrong

  1. Hallucination is the same mechanism that makes the model useful.Experiment: ask for five references on an obscure topic, then check them.
  2. Why a system that writes poetry struggles to multiply.Experiment: two four-digit numbers, plain; then "show your working"; then with a code tool.
  3. The confidence is in the prose, not in the knowledge.Experiment: ask the model to rate its confidence on ten facts, then verify.
  4. Sycophancy: models tend to agree with you.Experiment: state something wrong with authority; then ask neutrally.
Arc 3 · November – December 2026

Context: the most important word nobody defines

  1. A context window is a whiteboard, not a memory.Experiment: a long conversation, then a question about a detail from the start.
  2. Lost in the middle: where in a long document the model reads worst.Experiment: hide one fact at the top, middle and bottom of a long document.
  3. The system prompt: the instructions you never see.Experiment: the same question, the same model, two different products.
  4. The model is not the product.Search, memory, tools and guardrails are the wrapper — and the wrapper is what you are evaluating.
Arc 4 · December 2026

Giving the model knowledge

  1. Retrieval explained: a closed-book exam versus an open-book one.
  2. Embeddings: how a machine decides two sentences mean the same thing.Experiment: similarity scores for pairs that look alike but mean different things.
  3. Fine-tuning changes how a model behaves far more than what it knows.
  4. Retrieval, fine-tuning, or a longer prompt: what each mechanism is good at.
Arc 5 · January 2027

Agents: what changes when the model can act

  1. An agent is a loop: think, act, observe, repeat.
  2. Why errors compound: 95% per step across ten steps is a coin flip.Experiment: a five-step task, run ten times.
  3. Prompt injection: the security problem everyone outside security is ignoring.Experiment: put an instruction inside a document and ask for a summary.
  4. Human-in-the-loop is a design decision, not a slogan.
Arc 6 · February 2027

How do you know it's working?

  1. "It looks good" is not evidence: what an eval actually is.
  2. Building a twenty-question eval set.Experiment: run it across three models and publish the results.
  3. Benchmarks: what they measure, why leaderboards leak.
  4. The model under your product changed and nobody told you.A running experiment: the same prompt every month.
Arc 7 · March 2027

The economics of a token

  1. What inference actually costs, and why an answer is priced differently from a question.
  2. Why reasoning models are slow.Experiment: the same question with reasoning on and off, timed.
  3. Small models versus large: why bigger isn't always the answer.Experiment: one task on a small open model and a frontier model.
  4. What "open" actually means: weights, data, source.
Arc 8 · March – April 2027

How the model got its manners

  1. Pretraining versus post-training: where knowledge comes from, where behaviour comes from.
  2. RLHF in plain language: why models sound the way they do.
  3. Reasoning models: what happens in the "thinking", and what doesn't.
  4. Why the model is better in English, and what that means for Indian languages.Experiment: the same question in English, Hindi and Gujarati.
Reading as a book

When an arc closes, its notes are collected into a part of a book you can download. Part One, covering arcs one to three, is planned for December 2026.

Follow on LinkedIn