How to Learn AI and Machine Learning for Free: A Realistic Roadmap
A realistic, free path into AI and machine learning: what to learn first, the best free courses and videos, the tools you need, and projects that stand out.
You can learn AI and machine learning properly without paying for a single course. The best material in the world, from Stanford, MIT, Harvard, 3Blue1Brown and Andrej Karpathy, is free online. The hard part is not finding resources but knowing what to learn, in what order, and what to skip.
That is what the Hamropedia AI & ML Roadmap is for: 8 phases and 98 topics, in order, with hand-picked free resources for each. This article explains how to use it.
The big picture: 8 phases
| Phase | What you learn | Why it matters |
|---|---|---|
| 0. Your toolkit | Python, NumPy, Git, the command line | Every later phase uses them daily |
| 1. Math foundations | Linear algebra, calculus, probability | The language every model is written in |
| 2. Classical ML | Regression, trees, evaluation, built in NumPy | Understanding why models work and fail |
| 3. Deep learning | Neural networks, backprop, PyTorch | The engine behind modern AI |
| 4. Transformers & LLMs | Attention, tokenisers, building a small GPT | How ChatGPT-style models work inside |
| 5. AI engineering | RAG, evaluation, agents, deployment | The core of most AI engineer jobs |
| 6. Systems & scale | GPUs, memory, speed, multi-GPU training | Making models fast and cheap |
| 7. Generative models & RL | VAEs, diffusion, reinforcement learning | Image generators and RLHF |
Five rules that matter more than any course
- Go in order. Skipping maths to "get to the fun part" is the most common reason people stall at deep learning.
- One resource per topic. Collecting courses feels like progress, but it isn't. On the roadmap, one resource per topic is marked Worth it. Start there.
- Build, then rebuild. Code along with a video, then close it and write the code again from a blank file. The second time is where you learn.
- Explain it out loud. If you can't explain gradient descent to a friend in two sentences, you don't know it yet.
- Revisit old topics. Every couple of weeks, test yourself on earlier topics and untick anything you've forgotten. The roadmap's tick boxes are saved in your browser for exactly this.
The best free resources (a starter set)
- Maths, visually: 3Blue1Brown: Essence of Linear Algebra and Essence of Calculus.
- Maths, written: Mathematics for Machine Learning (free book).
- Classical ML: StatQuest videos and An Introduction to Statistical Learning (free book, Python edition).
- Deep learning from scratch: Andrej Karpathy: Neural Networks: Zero to Hero.
- Deep learning book: Understanding Deep Learning by Simon Prince (free PDF).
- Transformers: The Illustrated Transformer and Karpathy's "Let's build GPT".
- AI engineering: Your AI Product Needs Evals by Hamel Husain.
The full list, with 370+ resources sorted by topic and type (video, interactive, reading, code, paper), is on the roadmap.
What you need
- A normal laptop. Phases 0 to 3 run fine on any recent computer.
- Python (free) and a code editor such as VS Code (free).
- Free GPU time when you need it: Google Colab and Kaggle notebooks both offer free GPU hours, which is enough for this roadmap's projects.
- Time: consistency beats intensity. One focused hour a day, every day, goes further than a ten-hour weekend once a month.
Projects that make you stand out
Employers and universities remember projects that solve a real problem, especially with local data. Ideas:
- Devanagari handwriting recogniser using the public Devanagari Handwritten Character Dataset (Phase 3).
- Nepali news topic classifier: scrape headlines, label a few hundred, train and evaluate (Phases 2–3).
- Romanized-to-Nepali typing model that learns from examples instead of rules. Compare it with our rule-based typing tool (Phase 4).
- A RAG assistant that answers questions about NOC or WES rules from official documents, with an evaluation set you wrote yourself (Phase 5).
Put each one on GitHub with a short README: the problem, the data, how you measured success, and what didn't work.
Common questions
Do I need a computer science degree? No. You need patience with maths and lots of practice. Many working ML engineers are self-taught or came from other fields.
Python or something else? Python. Almost every AI library and course uses it.
Should I start with ChatGPT-style apps (Phase 5) because that's where the jobs are? You can build simple apps early, and that's motivating. But to debug them, evaluate them and stand out in interviews, you need Phases 1–4. Do both: build small apps while working through the foundations.
Is AI going to replace programmers? AI is changing what programmers spend their time on. People who understand how models work, and can judge when they are wrong, are in a strong position. This roadmap is designed to make you one of them.
Start now: Phase 0: Your toolkit → · Also read: Nepal's National AI Policy, explained