I want to learn how AI works, and I'm starting from nothing. Where do I start?
Here. You're in the right place, and it costs nothing. If you want to learn how AI systems work and you are starting from nothing — no maths, no coding, no background — this page is the honest answer: the order to learn things in, what you actually need first, and what you can safely ignore for now.
Free · no account · no tracking · last updated 30 July 2026
The short answer
Start with the AI basics — how a language model reads text at all — and go in order. The course that does this is called Field Foundations; the name just means the foundations of the AI field. It is free, it assumes nothing, and Module 1 is below.
Start with how a language model reads text at all. Not with opinions about AI, not with the news, not with a framework. The machine first. Almost everything confusing about AI systems becomes ordinary once you know what is actually happening inside them, and the first step of that is smaller than you think.
Start with Module 1 — Tokenization →
What you need before you begin
No maths. Not none forever — but none to start, and none for a long while. Where a piece of maths genuinely earns its place, it gets explained where it is used rather than assumed.
No coding. Same answer. You can understand how these systems work, and read the research, without writing a line of code. Coding becomes necessary if you want to do empirical research rather than follow it — and that is a later problem.
No degree, no application, no cohort. There is nothing to be accepted into.
The thing that actually stops most people isn't difficulty — it's not knowing the order.
The field looks like a wall of terms: transformers, attention, RLHF, alignment, interpretability, mesa-optimization. It isn't a wall. It's a sequence, and each piece is reasonable once the one before it is in place. Most people bounce because they start in the middle, where nothing can make sense yet.
The order
How text becomes something a model can use
A model has never seen a letter. Text gets chopped into chunks and turned into numbers first, and that single fact explains a surprising amount of odd model behaviour — including why they miscount letters in a word. Module 1 — Tokenization.
How a number becomes a meaning
Those numbers are just addresses. Embeddings turn each one into a position in a space of meaning, so that related things sit near each other. This is where meaning first enters the machine. Module 2 — Embeddings.
How words decide which other words matter
Attention — the mechanism the whole modern era is built on, and the one most people have heard of without knowing what it does.
The repeating block, and choosing the next word
Stack the pieces, run them many times, and turn the result into an actual word. That completes one pass through the machine — and at that point you can read most technical writing about how models work.
Then: how it learns, how it is shaped, and where alignment comes in
Training, fine-tuning, human feedback, and the failure modes — reward hacking, and the difference between asking for the wrong thing and getting a model that only appears to want the right one. That is the part the safety field argues about, and it needs everything above it to make sense.
How to actually learn it, rather than just read it
This matters more than which materials you pick, and almost nobody teaching themselves does it.
Read one section, then stop and write what it said in your own words. Not copied sentences — yours. If you can only copy it, you don't have it yet.
Then close everything and say it out loud from memory. This is the part that feels bad and works. Re-reading produces a comfortable feeling of familiarity that is not the same as knowing, and the gap between the two is where most self-teaching quietly fails.
Then come back to it days later, cold. The gap is the point, not a flaw in your memory.
Every module here is built for that loop. Each one ends with a button that copies the whole module into a prompt for any AI assistant — it will discuss it with you and answer your questions — and a second button that hands over the module's own test questions and asks the assistant to grade you honestly, closed book. You don't need a subscription; free tiers are fine.
What makes this different from the other options
There are good programmes out there and you should know what they are. BlueDot Impact runs cohort-based courses — genuinely good, and application-gated with fixed start dates. ARENA is excellent and assumes you can already write machine-learning code. AISafety.com maintains a large directory of links.
What none of them is, is a taught path that is open right now, assumes nothing, and never closes. That is the gap this fills. If one of those fits you better, take it — the point is that you start somewhere, not that you start here.
Who wrote this. I'm an independent alignment and safety researcher running a twelve-month behavioural study of an AI agent. These are the materials I needed when I started and couldn't find — built properly, and given away.
If something here is wrong or unclear, write to me. You're talking to me directly, not to a team.