Most people who decide to learn AI from scratch begin in the same place: a course catalogue. They scroll through forty options, read three syllabi, download nothing, and close the tab. Two weeks later they do it again.
That loop is the single most common reason beginners stall out for a year and still can’t explain what a loss function does. The fix isn’t a better course. It’s a sequence — a rough order of operations that tells you what to touch in week one, week six, and month five.
First, Split “AI” Into Two Different Jobs
The phrase covers everything from using ChatGPT to writing a transformer from scratch, and those are not the same skill. Decide which one you’re actually after, because the first 20 hours look completely different.
- AI as a tool. Prompting, automation, building workflows with the OpenAI or Anthropic APIs, connecting models to spreadsheets and documents. You can be useful within a month. No calculus required.
- AI as a discipline. Training models, tuning them, evaluating them, reading papers. This is the six-to-twelve-month track, and it does involve maths and a fair amount of debugging.
Plenty of people want the second and only need the first. Starting with the tool side is not a cop-out — it gives you working intuition about what models are bad at, which makes the theory land much harder later.
A Timeline That Reflects How Learning Actually Goes
Weeks 1–3: Vocabulary and one small notebook
You need about fifteen terms before anything else makes sense: training data, features, labels, overfitting, inference, parameters, embeddings, tokens, fine-tuning, validation, loss, gradient, epoch, prompt, context window. Write your own one-sentence definitions. If you can explain overfitting to a friend using a memorisation analogy, you’re ready to move on.
Weeks 4–8: Python, pandas, and your first model
Not “learn Python” in the abstract — learn the specific 15% of Python that data work uses. Lists, dictionaries, functions, loops, and importing libraries. Then pandas for loading a CSV, scikit-learn for fitting a model on it, and matplotlib so you can see what happened. A decision tree classifier on the Titanic dataset is a cliché because it works: 891 rows, a handful of columns, and a score you can improve by hand.
Weeks 9–16: Neural networks and one real project
PyTorch or Keras, pick one and stay. Train a small image classifier on a Kaggle dataset. Expect to be confused about tensors and shapes for about ten days. Then expect it to click suddenly and permanently.
Months 4–6: Language models and shipping something
Hugging Face’s course is free and genuinely good here. Build a retrieval-augmented chatbot over a folder of your own PDFs — fifty lines of code, and you’ll learn embeddings, chunking, and why the model confidently invents things. That last lesson is worth more than three certificates.
The Maths Question, Answered Honestly
You do not need a maths degree to start, and anyone who tells you to spend six months on linear algebra before touching a dataset is handing you a reason to quit rather than a plan.
What you actually need early on: averages, standard deviation, the idea of a vector, dot products, and a rough sense of what a derivative measures. That’s maybe eight hours of Khan Academy. Add matrix multiplication when you hit neural network layers, and probability distributions when you start caring about sampling temperature. Learn each topic at the moment it blocks you, not two years before you need it.
The Traps That Eat Six Months
Tutorial hell. If you’ve watched nine courses and built zero things, the problem isn’t knowledge, it’s output. Every tutorial should end with you modifying it until it breaks.
Tool-hopping. Beginners switch from TensorFlow to PyTorch to some new framework every fortnight because a YouTube thumbnail told them to. Pick one stack in January and don’t look at benchmarks again until March.
Reading the news instead of the docs. Following every model release feels productive and teaches almost nothing. An hour in official documentation beats a week of announcement threads.
GPU envy. You do not need hardware. Google Colab gives you a free GPU for a few hours a day, and every project in your first six months fits comfortably inside that.
Where to Find Instruction Worth Your Time
Structure helps, but only when it’s the right structure. If you’re unsure what belongs in your first month, this breakdown of an AI for beginners course and what to learn first is a good way to avoid spending week one on the history of the perceptron.
Don’t want to wait for a signup flow before you write your first line of code? There are solid free AI courses with no signup that get you into a browser-based notebook inside five minutes, which matters more than it sounds — friction kills momentum.
Once you’re past the basics and want to build rather than read, look for free AI coding courses that end with a working project instead of a certificate. A course that makes you deploy something imperfect beats a polished lecture series every time.
And if you’re choosing between paid options later on, this honest ranking of the top AI courses and the ones to skip will save you a few hundred pounds on material you can get free elsewhere.
Build Things That Are Too Small to Be Impressive
This is the part people skip, and it’s the part that works. Your first five projects should each take an afternoon and be slightly embarrassing to show anyone:
- A script that sorts your downloads folder into categories using a text classifier.
- A model that predicts whether a used car listing is overpriced, trained on scraped data.
- An image classifier that tells your houseplants apart — 40 photos, six classes, 89% accuracy is a fine result.
- A summariser that turns a 4,000-word article into five bullets through an API call.
- A chatbot that answers questions about your own notes, with citations back to the source file.
Finishing five small things teaches you more than one ambitious half-built project, because each one forces you to handle real data, real errors, and the gap between a tutorial and a messy CSV with missing values.
What to Do in the Next Seven Days
Open a free Colab notebook tonight and load a CSV you actually care about — your Spotify history, your bank export, a public dataset on air quality near your postcode. Print the first five rows and the column types. That’s it. That’s day one, and it’s more than most people do in a month.
On day three, fit a linear regression on two of those columns and look at the error. On day five, break it deliberately: feed it a column of text and read the error message properly instead of googling the fix. On day seven, write three sentences about what surprised you.
Repeat that rhythm for twelve weeks and you will have a working vocabulary, a Python habit, five finished notebooks, and a clear sense of which direction to push next — language models, computer vision, or the tooling side. That’s what learning AI from scratch actually looks like in practice: not a curriculum you finish, but a loop of small, concrete, mildly frustrating experiments you keep running.