How to Learn AI from Scratch: Free Resources Guide

Learning AI from scratch used to mean paying for a bootcamp or a graduate program. It doesn't anymore. Model labs, universities and cloud providers publish free teaching material, so access isn't the bottleneck. Order is. Pick the wrong sequence and you end up watching tutorials instead of building things that work.

This guide is a 90-day order that works, using free resources, and it ends with an application you built and published yourself. Every resource below was checked on September 27, 2026 and was free to use that day. Where a course sells an optional extra, such as a certificate, the table says so.

Pick your goal first

"Learn AI" covers three different goals, and they shouldn't all start with the same course list.

• Use AI in the job you have. For marketing, sales, operations or another non-engineering role, you don't need the full path. Take Elements of AI, one or two DeepLearning.AI short courses on prompting and retrieval, and then build something in a tool you'd use at work: an n8n or Zapier workflow, a Custom GPT, or a Claude Project. Measure it against the task it replaces. For sales roles, breaking into AI sales maps the entry points.
• Build applications on top of models. Follow the 90-day path below. It assumes no programming background and teaches Python in month one.
• Study how models work. The path below is a start, but research and machine learning engineering need more math and theory. The courses for that are at the end of this guide.

Why you can do this without paying

Free material now covers every stage. The University of Helsinki's Elements of AI teaches concepts without code. Harvard's CS50P teaches Python. Google's Machine Learning Crash Course covers core machine learning and includes a module on large language models. Hugging Face, DeepLearning.AI and Anthropic publish free courses on building with models, and MIT OpenCourseWare and Stanford's course sites fill in theory.

The catch is order. Jump straight into building an AI agent without basic Python, or without knowing what tokens (the chunks of text a model reads) and embeddings (lists of numbers that capture meaning) are, and you hit a wall in week one. The path below avoids that.

The 90-day path

Month one covers foundations: what AI is, basic Python and core machine learning concepts. Month two covers how today's language models work, how to call one from code, and how to check what comes back. Month three is shipping: build, publish and document one real project.

The weeks are a planning guide, not proof of progress. Each month ends with something you should be able to do. Move on when you can do it, not when the calendar says so.

Month 1: Foundations

Weeks 1 to 2: Take Elements of AI, created by the University of Helsinki and MinnaLearn. It's free and needs "no complicated math or programming," and it gives you the vocabulary every later course assumes: supervised and unsupervised learning, neural networks and the societal questions around them. Do the Introduction to AI part now. The Building AI part expects basic Python, so save it.

Weeks 3 to 4: Learn enough Python to be useful. Harvard's CS50P is free through OpenCourseWare and runs from functions and loops to reading and writing files, libraries and tests. Codecademy's Learn Python 3 is a shorter option: the lessons are free, while practice projects and the certificate need a paid plan. Focus on lists, dictionaries, functions, reading and writing files, and basic API calls. You can skip advanced features like decorators and async for now.

By the end of month 1, you should be able to write a Python script that calls an API, reads the JSON it sends back and prints a result.

Month 2: Calling models from code

Weeks 5 to 6: Google's Machine Learning Crash Course. It's free, and its programming exercises run in Colab in the browser, which needs a Google account. Do every exercise, not just the videos. The exercises are the point.

Weeks 7 to 8: Start calling models from code. The site's agent-building course is 17 free lessons in four stages. Its first stage, "Define the boundaries," takes you from sending an API request and handling its failures to prompt contracts with fixtures, tool calling and MCP, a standard way to plug a model into outside tools. The labs run locally on synthetic data, so you don't need a paid API key to reproduce their results.

Alongside that, pick one of these to see what's inside the model:

• The Hugging Face LLM course, which is "completely free and without ads." It teaches you to load a model from the Hub, fine-tune it and share the result. It expects good Python and some deep learning, so it fits better after the crash course than before it.
• DeepLearning.AI's short courses on prompting, retrieval, agents and evaluation. Many run one to two hours. Course pages say access is "free for a limited time during the DeepLearning.AI learning platform beta," and a paid Pro plan adds certificates.

By the end of month 2, you should be able to send a request to a model, check whether the answer matches the format you asked for, and explain why a bad answer was bad.

Month 3: Ship something real

This is the month most beginners skip. Courses feel like progress. Building feels like failure.

But the only way to learn AI is to build something, get stuck, fix it and ship it.

Use the agent course's second and third stages as your guide: retrieval with access filters, scoped memory, human approval and honest evaluation. Then pick one project:

1. A document Q&A system over your own PDFs that cites the passage it used and says when it can't answer.
2. A Telegram or Discord bot that summarizes articles you forward it, with a check that rejects summaries missing the source link.
3. A fine-tuned Hugging Face model that classifies emails into your own categories, with a held-back test set.
4. A small app that takes a URL, fetches the page and returns a structured summary that validates against a schema.

If you'd rather follow a spec, the course's ticket workflow capstone connects delivery, retrieval and approval in one build.

Set a hard deadline: published to a public repository by day 90. Explain what you built in a README. The README is what a hiring manager or client reads. The code is proof you wrote it.

Every free resource in one table

Tools you'll use, all with free plans

Sign up for these and get comfortable with them. Each has a free option that covers learning, with the limits noted.

GitHub for code hosting and version control. Learn git init, git add, git commit and git push, and you can handle most of what you'll need.

Google Colab for notebooks in the browser. Its FAQ says Colab is free to use, but GPU access on the free version is "heavily restricted" and not guaranteed. That's enough for course exercises, not for training large models.

Hugging Face for models, datasets and demo hosting. Per its Spaces documentation, static Spaces are free, and a free personal account can host up to two Gradio Spaces. Other compute-backed Spaces need a paid plan.

A model API for your own projects. Google AI Studio is free to use and the Gemini API has a free tier. Anthropic's pricing page says new API users get a small amount of free credit for testing. Pick one provider and focus.

An editor. VS Code is free. Cursor has a free Hobby plan with limited agent requests, and Windsurf, now renamed Devin Desktop, lists a free plan. Both help once you can read the code they write.

Three beginner mistakes

Three failure patterns show up again and again.

The first is getting stuck in tutorials. You finish course after course, never build anything of your own, and six months later you can't answer "what did you actually make?" The fix is the month-3 project rule. Stop watching once you can build.

The second is tool obsession. You spend three weeks deciding between PyTorch and TensorFlow, Claude and GPT, LangChain and direct SDK calls. None of that matters in month 1. Use whatever the course you're taking uses, and switch later if you have a real reason to.

The third is skipping the math. You don't need a math degree, but you do need basic comfort with vectors, matrices, derivatives and probability. If you've never seen these, 3Blue1Brown's free linear algebra and calculus lessons are the gentlest start.

Build a project someone can check

If your goal is a job or freelance income, the courses are necessary but not enough. What gets attention is a public project with a clear write-up and proof that it works.

Each project should answer four questions: what problem does this solve, how does it work, how do you know it works, and what would you do differently. For AI projects, the third question means an evaluation set someone else can rerun. The AI engineer path lists what that evidence looks like.

Ninety days and one project gets you foundations, not a job offer. Postings set experience bars. Baseten's AI Engineer posting, for example, asks for five or more years building and shipping software. The career transition guide maps what you already do to a role and shows employer-posted pay.

What to do after day 90

Pick a specialization based on what you enjoyed:

• LLM application development: finish the agent course, then go deeper on retrieval, evaluation and one agent framework.
• Machine learning engineering: take the full fast.ai course, then learn MLOps fundamentals such as experiment tracking and model serving.
• Research: Stanford's CS224N (NLP) and CS231n (vision) publish their course materials.
• AI for business: go deeper on workflow automation tools such as n8n, Zapier or Make, and build agents that solve a specific business problem.

Related Guides

• How to Become an AI Engineer
• How to Transition Into an AI Career
• The Best AI Newsletters to Subscribe To