Learning AI in 2026: what you actually need before you start

Two things have happened at once. AI became genuinely useful, and it became genuinely easy to make something that looks like AI work without understanding any of it.

Both are true, and the gap between them is where most beginners get stuck. So before you start this track or any other, here is what the work actually requires.

The maths question

Everyone asks it, so: you need less than a maths degree and more than nothing.

Specifically you need to be comfortable with:

• Linear algebra — vectors, matrices, and what multiplying them does. Not proofs. You need to understand what shape your data is and why the model complained about it.
• Probability and statistics — distributions, conditional probability, and above all what an average hides. Most bad models are bad because someone trusted an average.
• Calculus — enough to know what a gradient is and why the training loop follows it downhill. You will not compute derivatives by hand after week three.

That is genuinely it. If you can follow a formula and are willing to be confused for a few days, you have enough. If you failed maths at school and never went back, this track will be harder for you than the marketing on any institute's website suggests, including ours — but it is doable, and the fix is practice rather than talent.

Using an API is not the same as understanding models

You can build something impressive this weekend by calling a language model API. That is a real skill and worth having. It is also not what companies mean when they hire an AI engineer.

The difference shows up the moment something goes wrong. Your chatbot gives a confidently wrong answer. Now what? If your entire mental model is "send text, receive text", you have no next move. If you understand retrieval, evaluation and what the model is actually doing, you can diagnose it: the retrieval pulled the wrong document, or the prompt allowed an unsupported claim, or your evaluation never tested this case.

That diagnostic ability is the job. It is why our curriculum spends time on classical machine learning before touching large language models — not out of academic loyalty, but because the habits you build there (measure honestly, split your data properly, distrust a good score) are the habits that stop you shipping something broken later.

What the first month honestly feels like

Week one is fine. Python, notebooks, loading a dataset. Satisfying.

Week two you meet real data. It has missing values, inconsistent formats, and a column that means two different things depending on the row. You will spend more time cleaning than modelling, and you will wonder when the AI part starts.

Week three you build your first model and it works. Then you realise it works suspiciously well, and eventually discover you accidentally trained it on the answer. This is called data leakage. Everyone does it once. Finding it yourself is a genuinely important moment.

Week four something clicks and the pace picks up.

We say this out loud because students who expect week two to be exciting decide they are not smart enough. They are; the second week is just boring. The people who quit are almost never the ones who could not do it.

The four projects that matter more than the certificate

By the end you should be able to point at:

1. A prediction model with an honest write-up — including what did not work and how you know the score is real.
2. Something with your own data. Anyone can run a tutorial on a famous dataset. Collecting and cleaning your own is what shows capability.
3. A retrieval system over real documents — the most commercially useful thing a junior AI engineer can build right now.
4. One deployed thing. A model in a notebook is homework. A model behind an interface someone else can use is work.

That is the portfolio. It matters more than any certificate, ours included.

Where this leads

Machine learning engineer, data scientist, AI engineer, NLP engineer — and a very common path where an existing analyst or developer moves sideways into AI work at the company they already work for. That last route is underrated and often the fastest.

What we will not tell you is that a six-month track makes you a senior AI engineer. It makes you someone who can genuinely build, evaluate and deploy models, and who understands enough to keep learning without a teacher. That is the realistic outcome, and it is worth a lot.