How to Learn AI from Scratch: A Free Step-by-Step Beginner’s Guide

Who This Guide Is For

This guide is for people who have heard the phrase “learn AI” a hundred times and still aren’t sure where to start. You don’t need a mathematics PhD. You do need patience, a working computer, and about an hour a day. The path described here is deliberately modular — you can stop after Stage 3 and already have practical skills that employers want.

What “Learning AI” Actually Means

Artificial intelligence is an umbrella term. In practice, the skill tree splits into three branches:

• Machine learning (ML): Training statistical models on data to make predictions.
• Deep learning (DL): A subset of ML using neural networks with many layers.
• AI engineering: Building applications on top of pre-trained models via APIs (like GPT-4 or Claude).

Most job postings labeled “AI” in 2025 are really asking for ML or AI engineering skills, not novel deep learning research. Knowing this prevents the common mistake of jumping straight to PyTorch tutorials before you can write a Python function.

Stage 1: Python Fluency (4-6 weeks)

Step 1: Choose one beginner Python course and finish it

The best free options in 2025 are Harvard CS50P (rigorous, with auto-graded problem sets) and the official Python tutorial at python.org. Avoid tutorial hell — pick one, do every exercise, move on. You need to be comfortable with lists, dictionaries, functions, classes, and file I/O before touching any ML library.

Step 2: Learn NumPy and Pandas

Nearly all ML data work flows through these two libraries. Kaggle’s free Pandas micro-course takes about 4 hours and gives you immediate hands-on practice with real datasets. Do the exercises — reading documentation is not the same as writing code.

Common mistakes at this stage

• Jumping to neural networks before understanding Python data types.
• Watching tutorial videos without coding along.
• Treating “I ran the code” as equivalent to “I understand the code.”

Stage 2: Math Foundations (3-5 weeks, parallel with Python)

Step 3: Linear algebra and probability

You need two things: an intuition for matrix multiplication (how data transforms), and basic probability (how models express uncertainty). 3Blue1Brown’s “Essence of Linear Algebra” series on YouTube (free) builds geometric intuition in roughly 3 hours. For probability, the first four chapters of “Probability and Statistics for Data Science” by Blitzstein and Hwang are freely available as PDFs from Harvard.

You do not need calculus mastery before starting ML — most practitioners use automatic differentiation frameworks that compute gradients for you. A conceptual understanding of derivatives suffices for Stage 3.

Stage 3: Classical Machine Learning (8-12 weeks)

Step 4: Andrew Ng’s Machine Learning Specialization

This three-course series on Coursera (free to audit without a certificate) is the single most recommended on-ramp in the field. Ng is a co-founder of Google Brain and former chief scientist at Baidu. The 2022 updated version uses Python and scikit-learn rather than the old MATLAB version. Expect 8-10 hours per week. By the end, you will have implemented linear regression, logistic regression, decision trees, and neural networks from near-scratch.

Step 5: Build three projects with real data

Kaggle’s competition archive contains hundreds of beginner-friendly datasets. The Titanic survival prediction, house price prediction, and digit recognizer competitions are rites of passage. The goal at this stage isn’t winning — it’s developing the muscle memory of loading data, handling missing values, training a model, and evaluating it honestly. Document your work in a public GitHub repository or a Kaggle notebook. This becomes your portfolio.

Stage 4: Deep Learning (optional, 12+ weeks)

Step 6: fast.ai’s Practical Deep Learning for Coders

If your goal requires neural networks — image classification, NLP, generative models — fast.ai’s free course is uniquely well-designed. It teaches top-down: you train a working image classifier on Lesson 1, then learn the underlying math progressively. This top-down philosophy is the opposite of most academic curricula and dramatically reduces the time to first working result.

For more mathematical depth, DeepLearning.AI’s five-course Deep Learning Specialization (also free to audit on Coursera) fills in the theory that fast.ai skips.

Stage 5: AI Engineering (the fastest path to employment)

Step 7: Build with APIs before building from scratch

In 2025, a significant fraction of AI work involves integrating large language models, vision models, or speech models via API rather than training new models. OpenAI’s, Anthropic’s, and Google’s developer documentation are free. Building a working application — a summarization tool, a document Q&A system, a content classifier — using an existing API demonstrates practical AI skill to employers without requiring GPU hardware or months of training runs. See the AI learning category for platform reviews and project ideas.

Honest Time Estimates

At 1 hour per day: Stages 1-3 take roughly 6-9 months. Stage 4 adds another 4-6 months. Stage 5 can be started in parallel with Stage 3 and takes 2-3 months. The total path to a competitive junior ML or AI engineer portfolio is 9-18 months for most self-taught learners — faster if you have prior programming experience, slower if you are learning Python and math simultaneously. Anyone claiming you can “learn AI in 30 days” is selling a course, not teaching a skill.

Free Resource Summary

• Python: Harvard CS50P (cs50.harvard.edu/python)
• Data manipulation: Kaggle Learn — Pandas, NumPy
• Math: 3Blue1Brown linear algebra (YouTube), Blitzstein probability (Harvard)
• Classical ML: Andrew Ng Machine Learning Specialization (Coursera audit)
• Deep learning: fast.ai Practical Deep Learning for Coders
• Projects and datasets: Kaggle, UCI ML Repository, Hugging Face Datasets

Also explore our comparison of the best AI learning platforms to find the right structured course for your specific goals and budget.