Three jobs sit behind the phrase “working in AI”, and they need different things. This is the whole syllabus for each of them, written out: what to learn, in what order, roughly how long it takes, and — the part nobody tells you — what you can safely skip.
Everything here is written, not collected. 87 topics across 279 sections, each with an explanation, a syllabus you can follow, practice that ends in working code, and the questions an interview will ask about it. Free, no account, nothing to sign up for.
The most expensive mistake a newcomer makes is studying the wrong syllabus entirely, because the field’s job titles do not map onto its topics. Each route below is written out in full, in order.
You build products on top of models somebody else trained: agents, retrieval systems, assistants, anything where a language model is a component rather than the project.
Foundations · Working with models · Agents · Retrieval · Internals · Knowing it works · Shipping it · The interview
Read the routeYou train, evaluate and ship models: the person who owns a model from the data that feeds it to the service that serves it, and who is called when its numbers move.
Foundations · The maths · Data · Classical ML · Deep learning · Scale & evals · Production · The interview · Specialisms
Read the routeYou work on what models can do rather than what can be built with them: pretraining, alignment, interpretability, safety, and the questions nobody has settled yet.
Foundations · The maths you need · Learning from data · Deep learning · Language models · Research practice · The interview · The frontier
Read the routeRather than describe it: this is the opening of Attention, conceptually, one of 4 sections in Transformers & attention, unedited.
Every current language model is a transformer, and the part of a transformer that does the work is attention. This section explains it in plain language with no equations, because being able to say what attention does in five minutes is a genuinely separate skill from being able to implement it — the first is asked in interviews constantly and the second is asked rarely. The implementation, with the shapes and the arithmetic, is the from-scratch section later in this topic.
To understand a word you often need another word far away. In the sentence about the trophy not fitting in the suitcase because it was too big, working out what "it" refers to means looking back at two candidate nouns and deciding which one the word "big" makes sensible. Change "big" to "small" and the answer flips to the other noun. Nothing local to the word "it" contains that information.
Older sequence models processed one step at a time and carried a summary forward, so distant information degraded. They also could not be parallelised across the sequence, because step ten could not begin until step nine finished. The sequence-models topic measures exactly how badly the first problem bites: in a plain recurrent network the gradient reaching fifty steps back is around 5e-19, which is not small but absent.
Attention lets each position look directly at all the others in the same sequence — self-attention — and decide how much each one matters for it. Distance costs nothing: the fiftieth word and the first are one step apart, exactly as the fiftieth and the forty-ninth are. And because every position can do this at the same time, the whole sequence is processed in parallel, which is what made training at scale possible.
Several topics carry an explicit warning that they are commonly over-studied for a given role, with the reason. A roadmap that only adds things cannot help you finish.
A topic’s hour figure has to equal the sum of its sections, or the site refuses to build. So do the dependency links, the route ordering, and every cross-reference on the site.
No link dumps. Every explanation, code sample and interview answer was written for its place in the syllabus. External resources appear where they genuinely beat what could be written here, with a line saying why that one.
Where the field disagrees with itself, this says so rather than picking a side. Where the writing is thin relative to the hours, it says that too.
Every question is attached to the section that teaches its answer, so a question you cannot answer has somewhere obvious to go. Follow-up chains included, because single questions are easy to memorise and the second and third are where understanding shows.
Open the question bankNot everything is a syllabus. These are the pages you come back to rather than read once.
7 guided builds — a GPT from scratch, RAG with citations, an eval harness — where every step links to the section that teaches it.
The round structures at frontier labs, big tech, scale-ups and startups: what each tests, what it does not, and the non-coding rounds nobody prepares for.
Every term this site uses, defined in a sentence or two, each pointing at the material that teaches it properly.
RAG or fine-tuning, an agent or a pipeline, open weights or an API — weighed with the costs of both sides stated.
Transformer memory arithmetic, hyperparameters, latency budgets, GPU specifications, and which metric to use when.
All 87 of them with summaries, depths and the roles that need them — for when the map is the wrong tool.
This map will not get you a job, and reading it will not either. It tells you what to learn, in what order, and what to skip. That is a real problem worth solving and it is not the same as being ready.
What actually gets you hired
Every topic has a Now actually do it section, and there are 180 of them. That is where the hours go. The reading is the map; the practice is the journey.
Depth is uneven, deliberately
The AI Engineer route is written in the most depth. The ML Engineer and Research routes are thinner in places and should be read as orientation plus a reading list rather than as complete teaching. Where a topic is shallow relative to its hours, the external resources are doing the work — and 19 topics on the AI Engineer route currently have none, so on those you are on your own.
It is also not only depth. Whole areas are missing from routes that arguably want them: the Research route carries no evaluation topic and no inference-economics topic, and nothing anywhere mentions JAX. Take a route as a strong spine, not a complete syllabus, and read around it.
Treat confident claims here, including these, as things to verify rather than accept.
For completeness, since this page is about what the site does and does not do: page views are counted through Vercel Web Analytics. It sets no cookies and there is nothing to sign in to, but it is not nothing — a pseudonymous visitor hash is derived from request data against a salt that rotates daily. Nothing you mark as done is part of that; it never leaves your browser.
All 87 of them, grouped by subject rather than by route, because most topics sit on more than one.