Engineering

AI Foundations: Start Here

Four short series explaining how these systems actually work — what the model is doing, what your coding tool is doing around it, how to work with it day to day, and how to ship it in your own product. Which one to read depends on what surprised you.


This page exists because two posts went out that named a whole layer nobody had written down, and the people who said they could not follow them were right.

Start from what surprised you

Pick the line that sounds like your week. Each goes straight to the post that answers it. Each of those four also opens by naming what it assumes — usually one earlier post — so you will know within a paragraph whether to read that one first. Start where the question is; back up only when the post tells you to.

If none of those is it, the four tracks below are sorted by who they are for, not by difficulty. Read the one that matches you and stop.

Track 1 — what the model is doing

For any engineer who uses AI and has never looked inside. Assumes you can code and have used a chat assistant; nothing else. It is the broadest entry point, and the other three tracks lean on it.

  1. What a Model Actually Is — tokens, one guess at a time, and why it never looks anything up.
  2. Why It Sounds Certain When It’s Wrong — hallucination as the mechanism working as designed, plus what temperature does.
  3. The Context Window Is the Entire Memory — stateless model, whole conversation resent every turn, and what “it forgot” means.
  4. How It Learned to Follow Instructions — pretraining versus post-training, and why a system prompt works at all.
  5. Embeddings, and Why Retrieval Beats Memorising — vectors by hand, retrieval in thirty lines, then broken on purpose.
  6. Reading a Model Card Without Guessing — parameters, context length, price, speed, and benchmarks that mislead.

Track 2 — what your coding tool is doing

For developers using an AI coding tool daily who hit behaviour they cannot explain — the readers who got lost. Assumes track 1, or a willingness to follow its links. Each post explains the mechanism in general terms first, then shows the same thing concretely in Claude Code.

  1. What Actually Happens When You Press Enter — the loop end to end, in about forty lines you can run.
  2. A Tool Call Is a Request, Not an Action — the model never runs anything; your harness decides whether to honour it.
  3. MCP, Explained Without the Spec — what a server offers, which side your editor is on, what crosses the wire.
  4. Build a Small MCP Server and Watch the Wire — a hundred lines, two read-only tools, every message logged.
  5. What Can Go Wrong With an MCP Server — the attack classes the specification itself names, and what to check before installing one.
  6. Where the Context Comes From — your message is the smallest part of the request; compaction throws some of the rest away.
  7. Who Is Allowed to Do What — asking an agent not to, versus actually stopping it.
  8. How a Skill Gets Found and Loaded — the name and the description together are the trigger, and two similar ones fight.
  9. Subagents: Why Isolation Is the Point — a fresh context with a narrow brief; parallelism is the side effect.
  10. The Whole Thing, Annotated — one real session, first prompt to landed commits, every mechanism labelled where it fires.

Track 3 — using it well, day to day

For working developers, and the tech leads deciding whether this is worth it. Assumes nothing, and a non-engineer can follow it. Two posts:

  1. Is It Actually Making You Faster? — what to measure, what the numbers hide, and how to run the comparison honestly for two weeks.
  2. Rolling It Out to a Team — which configuration belongs to the repository rather than the person, who owns the shared skills, and the adoption failures that turn out to be social.

The track is short because most of its ground was already published, and those two link out to it rather than restate it. Already written:

Track 4 — putting AI in your own product

For engineers shipping a feature with a model behind it. A different reader from the other three: not somebody using an AI tool on their work, but somebody putting a model inside something other people will use. Assumes tracks 1 and 2. It is the only track whose code you keep: an agent loop and an eval harness, both written by hand, both short enough to read in one sitting.

Four posts:

  1. Writing the Agent Loop Yourself — a step budget, an approval gate, reasoning printed each turn, and the four ways it actually went wrong.
  2. Knowing Whether It Works — twenty cases where the assertion is “good enough” rather than equality, and a real regression the suite catches.
  3. Making the Model Yours — and When Not To — a real fine-tune, before and after, and what happened when I tried to teach it a fact instead of a manner.
  4. Running It in Production — streaming, the two cost levers, a ceiling that trips, injection against a service, and what you may legally send.

The three guides — AI architecture, training, fine-tuning and evaluation and MLOps and production operations — are the breadth-first catalogues this track points into. They assume you already know the shape of the problem; these four posts are where you get it.

The reference pages

Three pages here are not rungs on any ladder. They are meant to sit open in a second tab while you read the rest.

What this series is not

It explains mechanisms and where they break. It is not a security assessment, not a compliance framework, and not legal advice. If you are making a deployment decision — whether a model may touch customer data, what has to be logged, who signs it off — start instead with NIST AI 600-1, the Generative AI Profile of the AI Risk Management Framework (July 2024), and with a CISO’s guide to agentic AI (Anthropic, July 2026) for the organisational version of the same question. Understanding the machine is a prerequisite for those conversations, not a substitute for them.

References and further reading

Riddam Jain

Staff Engineer · Amsterdam

I write about cloud and application architecture, AI, and leading engineers. If this was useful, let's connect.