Learn AI, for people who already ship software
Read it once, in order,
and go one layer deeper without losing your place.
Every technical term in the prose is a link that opens a short explainer beside the text instead of navigating away. 2 tracks, 23 topics, 99 terms, about 2 hours. It assumes you know distributed systems, caching and API design, and assumes nothing about machine learning.
Start with AI fundamentals. Once its first phase is done, Classifying with small models opens and runs alongside the rest.
Track 1 · 4 phases · 0/22 done
AI fundamentals
From tokens to production, in dependency order. Four phases, each of which makes the next one make sense.
How does text become something a model can compute on?
Before any model does anything interesting, your strings have to become numbers. This phase is about that conversion and everything it quietly decides for you.
Phase 2 · The model
🔒 LockedWhat is actually happening inside, and what does that cost?
A transformer is a fixed amount of arithmetic per token. Once you can count that arithmetic, latency, memory and price stop being mysterious.
Unlocks at 4 of 4 topics in Phase 1 — 0 done. Unlock anyway.
Phase 3 · Building with it
🔒 LockedHow do I compose a system around a non-deterministic component?
The model is one call in a larger system. This phase is the engineering around it: what you put in the context, what you get back, and how you know it works.
Unlocks at 5 of 6 topics in Phase 2 — 0 done. Unlock anyway.
Phase 4 · Systems
🔒 LockedHow do I run this in production?
Loops, tools, queues, budgets and the failure modes that only show up under load. Ordinary distributed systems work, with one unusual component in the middle.
Unlocks at 5 of 6 topics in Phase 3 — 0 done. Unlock anyway.
Track 2 · 6 modules · 0/1 done
Classifying with small models
Small encoders, prototypes, calibrated confidence and honest metrics. It builds on Phase 1 (embeddings, vector space, encoders) and runs alongside the rest of the fundamentals.
🔒 Unlocks at 4 of 4 topics in Phase 1 — 0 done.
- Module A · FramingShould this be a model at all, and when should it stay quiet?0 of 2 written
- Module B · Labels and representationWhat does the model see, and what does it promise to return?0 of 4 written
- Module C · Making a decisionHow do scores become a confident answer, or an honest "not sure"?0/1 done · 1 of 6 written
- Module D · Knowing it worksHow do I know it is right, and right for whom?0 of 6 written
- Module E · Exploring data by handWhat is actually in the data before I label it?0 of 6 written
- Module F · Beyond bi-encoders, and shippingWhat do I reach for when similarity is not enough, and how do I ship it?0 of 3 written
Examples · 30 problem statements
All examples →Situations you could actually be handed, worked for how to think rather than how to solve. These 6 need no prerequisites — start with them if you want to know what the material is for before reading it.
Is this an AI problem at all?
Three questions to ask before naming a model, because most requests that arrive as "add AI" resolve on the first one.
A spam filter and a chatbot are not the same kind of thing
The word "AI" covers two products with opposite shapes — one has a right answer and one has better and worse ones — and the shape decides everything downstream.
The brilliant intern with no memory
The single most useful analogy for working with a model — someone who has read everything, remembers nothing between conversations, and sounds equally sure either way.
Where would the answer come from?
Every answer a model gives has one of three sources — its training, your prompt, or a tool — and asking which one, before building, decides most of the architecture.
The cost of being wrong decides the design
The same model sits behind an autocomplete and a legal notice; what changes is what a wrong answer costs, and that — not the difficulty — sets how much checking goes around it.
The flaky test that isn't flaky
A test that passes four times in five is not broken — it was written for a function and the thing under test is a distribution, and every habit you have about testing needs one amendment.
Projects · one full build
All projects →The practical end of a track: every step from an empty folder to something that runs, including the parts that go wrong. Not gated.
Local-first
Progress, highlights and notes live in this browser. No account, no server, no telemetry. Export them to Markdown or JSON whenever you like.
Level-gated, not locked
A phase or module opens at 80% of the one before it. If you already know the material, there is an unlock switch in settings — the progress data stays meaningful either way.
A diagram per concept
Every topic carries one hand-drawn picture of the mental model, placed before the detail rather than after it, because a diagram is scaffolding and not decoration.