A friend of mine learned enough machine learning to get hired as an engineer on a budget of about $40. No bootcamp, no master’s, no employer paying for anything. He watched Stanford lectures on YouTube during his commute, ran his first models on a free Kaggle notebook, and applied for jobs eighteen months later.
The material that makes a story like that possible is genuinely free and genuinely good. Stanford posts its lectures. Google publishes its own internal training. The researchers who wrote the transformer paper also wrote an interactive textbook you can read in a browser tab. Almost none of it is hidden.
The problem is navigation. Search for free AI learning resources and you get pages of listicles recycling the same six links, half of which are paywalled by lesson three or reference an API that was deprecated in 2023. So treat this as a working map instead, roughly in the order I’d work through it.
Start with the people who build the models
Grant Sanderson’s four-part neural network series on 3Blue1Brown runs about an hour in total. It gives you the geometric intuition for what a weight and a gradient actually are, which is the thing most courses skip before dumping notation on you.
From there, Andrej Karpathy’s Neural Networks: Zero to Hero series on YouTube is the single highest-value free resource in the field. He builds micrograd, then makemore, then a working GPT, typing every line on screen. You watch him hit a bug, stare at it, and fix it. That’s closer to real engineering than any polished lecture.
fast.ai’s Practical Deep Learning for Coders takes the opposite approach: you get a working image classifier in lesson one and learn why it works afterwards. Videos, notebooks and the book are all free on the site.
Google’s Machine Learning Crash Course is around 15 hours and worth doing for the vocabulary alone, especially now that it includes modules on large language models and prompting. Hugging Face runs its own free courses too, and their NLP course is what most working engineers used to get comfortable with transformers before fine-tuning anything.
The specialist academies that barely get mentioned
Once you’re past the fundamentals, the free material gets oddly specific, and that’s where it becomes useful.
LangChain Academy
Free, and focused on building rather than theory. If you want a structured path into LangChain Academy’s agent and LangGraph material, it’s a solid few weekends of work and doesn’t cost anything beyond your patience with documentation.
Weights & Biases Academy
Experiment tracking, hyperparameter sweeps, evaluating LLM outputs. The kind of unglamorous MLOps skill that separates people who finish projects from people who have notebooks full of dead ends. W&B Academy’s free training is short, practical, and built around the tool you’d use at work anyway.
Elements of AI
A University of Helsinki course with over a million enrolments. No coding required, roughly six weeks at a few hours a week. It’s the one I recommend to managers and journalists who need to understand what these systems can and can’t do before they make decisions about them.
Kaggle and Colab: the free compute nobody budgets for
You can read about gradient descent for months without learning much. At some point you need a GPU, and the free tier is better than most people assume. Google Colab hands out a T4 with session limits, which is enough to fine-tune a small model with LoRA. Kaggle notebooks give you roughly 30 GPU hours a week.
Kaggle Learn is underrated too. Intro to Machine Learning takes about three hours. Intro to Deep Learning is similar. Then enter the Titanic competition, then House Prices. A public leaderboard that tells you your model is worse than a random forest teaches faster than any certificate.
If you teach, someone has already built the thing you need
Teachers get a rough deal from generic AI courses, which tend to assume you want to become an engineer. The better option is material written for the classroom: lesson planning, assessment design, and honest guidance on why AI detectors don’t work. This roundup of free AI courses built for teachers covers the ones that don’t waste a staff meeting.
Books, papers and newsletters worth the bandwidth
- Dive into Deep Learning (d2l.ai) — free, interactive, with runnable code in every chapter. Used as a course text at hundreds of universities.
- Speech and Language Processing by Jurafsky and Martin — the standard NLP reference, with draft chapters posted free as they’re revised.
- arXiv — the actual papers. Start with Attention Is All You Need, then the InstructGPT paper on RLHF. Read the abstract and figures first; the maths can wait.
- The Batch from DeepLearning.AI, Ahead of AI by Sebastian Raschka, and Jack Clark’s Import AI — three newsletters that will keep you current without the hype cycle.
How to spot a free resource that will cost you a month
Free isn’t the same as cheap. A bad course costs you the weeks you’d have spent on a good one, so run a few checks before you commit.
- It mentions a certificate before it mentions a project you’ll build.
- The last update predates the current model generation. Tooling moves fast; stale screenshots waste real time.
- Every lesson ends in an email capture before the content arrives.
- Nobody is named as the instructor, or the instructor has no work you can find.
- It’s listicle-shaped: 47 tools, two sentences each, no depth on any of them.
A simpler test: does it make you write code or evaluate a model within the first hour? If you’re still reading definitions in week two, close the tab.
A rough eight-week route through the free stuff
Weeks one and two: Google’s crash course plus the 3Blue1Brown series. Weeks three and four: fast.ai lessons one to three, then the Titanic competition on Kaggle. Weeks five and six: Karpathy’s series up to the makemore videos, then the Hugging Face NLP course. Weeks seven and eight: pick one specialist track, build something small end to end, and write down what broke. That last part matters more than the syllabus.
When free stops being enough
Free has real gaps. There’s nobody to unblock you at 11pm, no deadline, and no cohort pushing you forward. If you’ve done eight honest weeks and you’re still going, it’s reasonable to spend something. The middle ground is a curated paid course in the $20 to $50 range, and there are more of those worth buying than you’d expect — this breakdown of cheap AI courses that earn their price covers the ones people actually finish.
If you need structure and accountability, a bootcamp can be worth it, though the $15,000 sticker price usually isn’t. There are shorter, cheaper programmes now, and affordable AI bootcamps are worth comparing against simply joining a study group and following the plan above with other people.
Start free regardless. Give it six weeks of real effort before you pay anyone. Most people who buy a course quit in week two, and the free version of the same syllabus would have told them that for nothing.