You are starting from zero and want to know what to learn, in what order, and when to move on.
Reviewed
Roughly four to six months at five to seven hours a week; less if you already code
You will use AI tools well, write basic Python, explain how a model learns from data, and have shipped one small AI app you can show other people.
Most AI roadmaps open with a semester of linear algebra, and most people who follow them stop in the first month. This one runs in the order that keeps you going: use the tools first, then learn just enough code and maths to understand them, then train models, then build with them.
You do not need every stage. If you will never write code, stages 1 and 7 are the whole path for you. If you already program, start at stage 3 and move quickly. Each stage ends in something you can check, so you know when to move on instead of guessing.
Start by using a chat assistant every day on tasks you already know how to judge: drafting messages, summarising notes, cleaning up a spreadsheet, explaining a concept you are studying. Volume on familiar work teaches you faster than any course, because you can see exactly where the output goes wrong.
Keep a short log of every bad answer and why it was bad. Within a few weeks the pattern is clear: the model guesses when you leave things unspecified, and it states wrong facts with full confidence. Knowing that is what separates people who use AI well from people who give up on it.
You are ready to move on when you have handed one recurring task to AI for good and can explain, in a sentence, why a specific answer failed.
Python is the language almost all AI work is done in. You need the basics, not mastery: variables, lists and dictionaries, loops, functions, reading files, installing a package and running a notebook.
Learn it by automating something you did by hand in stage 1. A script that cleans a file you get every week is worth more than any number of exercises, because you will keep using it.
You are ready to move on when you can write a script that reads a CSV file, fixes its messy columns and prints a summary, without copying the whole thing from an assistant.
You need three ideas, at the level of intuition rather than proofs: vectors and matrices, because data and model weights are stored that way; derivatives and gradients, because that is how a model learns; and basic probability and statistics, because model outputs are probabilities and results need measuring.
Do not wait to finish the maths before moving on. Learn each idea when the next stage needs it; it sticks far better with a problem attached.
You are ready to move on when you can explain in your own words what a gradient tells a model to do, and why a score measured on the training data proves very little.
Machine learning is how a model learns patterns from examples instead of following rules someone wrote. The core ideas are few: training and test data, features and labels, overfitting, and choosing a metric that matches what you care about.
Work on real datasets from the start. Kaggle has thousands of public ones, and scikit-learn lets you train a first model in a dozen lines of Python.
You are ready to move on when you have trained a model on a public dataset and beaten a simple baseline on data the model never saw during training.
Neural networks are behind image recognition, speech and the large language models you used in stage 1. Learn what a training loop does, how a network is built from layers, and how fine-tuning adapts a model someone else trained to your own data.
Pick one framework and stay with it. PyTorch is the most common choice today, and most open models ship with PyTorch code.
You are ready to move on when you have fine-tuned a pretrained image or text model on a small dataset of your own and can say how accurate it is and where it fails.
Most AI products today are built on top of a large language model rather than trained from scratch. The skills here are practical: calling a model through an API, getting structured output you can rely on, retrieval-augmented generation (RAG) so the model answers from your documents, evals that tell you whether a change made things better, and agents that use tools.
Build one small thing end to end rather than many demos. A question-answering app over a set of documents you know well is a good first project, because you can tell when its answers are wrong.
You are ready to move on when that app is live, other people can use it, and it passes a test set of twenty questions you wrote before you built it.
Skills open doors only when someone can see them. Put each project on GitHub with a README that says what it does, how well it works and what you would do next, and write a short post about what you learned.
Then test yourself alongside other people: hackathons, open-source issues, internships and community projects. A hackathon weekend teaches more about shipping under pressure than a month of solo study, and it is a common way to meet future teammates.
This stage has no finish line. A useful checkpoint is two public projects with clear READMEs and one hackathon entry, finished or not.