A roadmap for learning AI

The problem with learning AI in the current era is not scarcity of material — it is the opposite. There are thousands of courses, videos, and threads, and no obvious order to take them in, so it is easy to spend months busy without getting anywhere. What actually helps is a coherent sequence: a small number of foundations that everything else rests on, learned in an order where each stage makes the next one comprehensible. This is that roadmap. It is opinionated about order and about method — I lean hard toward building projects over collecting tutorials — but the structure reflects how the field genuinely stacks up rather than any personal shortcut.

Learn in the right order

AI education has a genuine dependency graph: some things are prerequisites for others, and skipping ahead produces the illusion of progress without the substance. The reliable sequence:

1. 1

   Mathematical foundations

   Linear algebra (vectors, matrices — the language of data and models), calculus (derivatives and gradients — how models learn), and probability and statistics (how we reason under uncertainty and evaluate results). You do not need all of it before starting — enough to begin, then deepen as needed.

2. 2

   Programming

   Python is the lingua franca of AI. Get comfortable with the language and the core numerical and data-handling libraries. Fluency here removes friction from everything downstream.

3. 3

   Classical machine learning

   Before neural networks, understand the fundamentals: what training, features, generalisation, overfitting, and evaluation actually mean. These concepts underpin everything, and they are far clearer in simpler models than in deep ones.

4. 4

   Deep learning

   Once the fundamentals are solid, move to neural networks — how they are built, trained, and applied to images, language, and beyond. This is where the modern frontier lives, but it rests entirely on the layers beneath it.

5. 5

   Specialisation

   AI is too large to master whole. Once you have the foundations, go deep in one area — computer vision, natural language processing, or another — rather than staying shallow across all of them.

Projects over tutorials

Here is the single highest-leverage habit, and the one I would defend most strongly: build things. Watching a course gives you the comfortable feeling of understanding while leaving you unable to do very much unaided. Building a project — even a small, imperfect one, end to end — forces you to confront the parts that tutorials smooth over: messy data, choices with no obvious right answer, the debugging that is where real learning happens.

The pattern that works is to alternate: learn a concept, then immediately apply it to something concrete. Reproduce a simple result, then modify it. Take a dataset you actually care about and try to do something with it. A portfolio of projects you built and can explain is worth more than a stack of course certificates — both for what it teaches you and for how it demonstrates real ability to others.

Read papers, and read them early

You do not need to wait until you are an expert to start reading research — you need to start badly and improve. Even a rough first pass at a paper builds vocabulary and a sense of how the field argues with itself. Pair this roadmap with a deliberate approach to reading the literature so the primary sources become an asset rather than an obstacle; that skill compounds with everything else you learn.

Consistency beats intensity

Learning AI is a long game, and the people who succeed are rarely the ones who studied hardest for a month — they are the ones who kept going for a year. A steady, sustainable rhythm beats heroic bursts that end in burnout, because the material genuinely takes time to internalise and compounds only if you stay in contact with it. Treat it as a marathon: pick a pace you can hold, protect the habit, and let the accumulation do the work. The frontier will keep moving, so the real skill you are building is not any single technique but the durable ability to keep learning the next one.

Practical takeaway

Follow the dependency order — foundations, programming, classical ML, deep learning, then a specialisation — without skipping the boring early layers, since they are what make the exciting later ones make sense. Refuse to get trapped mastering math in the abstract; learn enough to move, and deepen on demand. Above all, build: alternate every concept with a project, however small, because doing is where understanding actually forms. Start reading papers before you feel ready, and optimise for a consistent pace you can sustain for years. There is no shortcut around the foundations — but there is a clear path through them, and following it in order is what turns the overwhelming amount of available material from a source of paralysis into a plan.

Editorial note — A conceptual learning roadmap. First person reflects genuine learning preference (projects over tutorials, consistency over intensity); no fabricated credentials, timelines, salary figures, or personal outcomes are claimed. Curriculum order is a widely held convention, not a measured result.