How to Learn AI from Scratch: Free Resources Guide
ZarifThree years ago, learning AI from scratch meant a $10,000 bootcamp or a graduate program. In 2026, it does not. Every model lab, university, and major cloud provider has open-sourced their teaching material. The bottleneck is no longer access. It is curation. There are now more free AI courses than any human can finish, and the wrong sequence is what keeps most beginners stuck on tutorials forever instead of building things that work.
This guide is the sequence. Ninety days, zero dollars, and at the end you will have built and deployed a working AI application. Every resource named below is free as of May 2026 and verified to be current. No paid upsells masked as free trials.
Learning AI from scratch in 2026 means building the skills to use, customize, and deploy machine learning and generative AI systems, starting from no programming background and progressing through hands-on projects rather than passive video consumption.
TL;DR
- A practical zero-to-deployed path takes about 90 days at 8 to 10 hours per week using only free resources
- Start with Elements of AI (no code required) for concepts, then Python via the Codecademy free track, then Google's Machine Learning Crash Course
- Hugging Face hosts thousands of free pretrained models and a free course; this is where you actually start building
- The two free courses that matter most for generative AI are DeepLearning.AI's short courses and Anthropic's free Claude Skills training
- Spend at least 50 percent of your time building, not watching videos; one shipped project beats 40 hours of passive learning
Why You Can Actually Do This Without Paying
The economics of AI education flipped in 2024 when the major labs realized that more developers building on their APIs meant more revenue, and the cheapest customer acquisition channel was teaching the next generation for free. As of 2026, OpenAI, Anthropic, Google, Microsoft, and Hugging Face all maintain free curriculum that is better than what most paid bootcamps offered three years ago. Add in MIT OpenCourseWare, Stanford CS courses on YouTube, and the University of Helsinki's Elements of AI, and you have a complete syllabus.
The catch is sequencing. A beginner who jumps straight into building an LLM agent without understanding tokens, embeddings, or basic Python will hit a wall in week one and quit. The path below avoids that.
The 90-Day Path
The plan splits into three months. Month one is foundations: what AI is, basic Python, and core machine learning concepts. Month two is the modern AI stack: transformers, LLMs, prompting, and embeddings. Month three is shipping: build, deploy, and document one real project.
Month 1: Foundations
Week 1 to 2: Take Elements of AI from the University of Helsinki. It is free, requires no programming, and gives you the conceptual scaffolding (supervised vs unsupervised learning, neural networks, ethics) that every later course assumes you already have. Budget 8 to 12 hours total across two weeks. Skip nothing.
Week 3 to 4: Learn enough Python to be useful. The free Codecademy Python 3 track or the free CS50P from Harvard via edX both work. Focus on lists, dictionaries, functions, file I/O, and basic API calls. You do not need decorators or async yet.
By the end of month 1 you should be able to write a Python script that hits an API, parses the JSON response, and prints a result.
Month 2: The Modern AI Stack
Week 5 to 6: Google's Machine Learning Crash Course. It is free, well-paced, and built around TensorFlow with practical exercises. Do every exercise, not just the watching. This is where most learners cheat themselves; the exercises are the entire point.
Week 7 to 8: Hugging Face's free course on transformers and large language models. By the end you should be able to load a pretrained model, run inference, and fine-tune a small model on a custom dataset. Hugging Face is also where you will spend most of your future career, so getting comfortable with the platform is itself the lesson.
Concurrently, work through DeepLearning.AI's short courses. Most are 60 to 90 minutes, free, and cover specific practical topics: prompt engineering, RAG, agent design, evaluation. Pick the four most relevant to what you want to build and finish them.
Month 3: Ship Something Real
This is the month most beginners skip. They keep taking courses because courses feel like progress and building feels like failure. The only way to actually learn AI is to build a thing, get stuck, debug, and ship.
Pick one project from this list:
- A document Q and A system over your own PDFs using a free embedding model and Claude or GPT for generation.
- A Telegram or Discord bot that summarizes news articles you forward it.
- A fine-tuned Hugging Face model that classifies emails into custom categories.
- A Streamlit app that takes a URL, scrapes the page, and returns a structured summary.
Set a hard deadline: project shipped to a public URL or repo by day 90. Document what you built in a public README. The README is the artifact you show to a hiring manager or client; the code is just the proof you can write it.
The Free Resource Stack
| Resource | What You Learn | Time | Best For |
|---|---|---|---|
| Elements of AI (University of Helsinki) | AI concepts, no code | 10 to 15 hours | Absolute beginners |
| Codecademy Python 3 (free track) | Working Python skills | 20 to 25 hours | Non-programmers |
| CS50P (Harvard via edX) | Python fundamentals plus problem solving | 40 to 60 hours | Learners who want depth |
| Google Machine Learning Crash Course | Core ML, TensorFlow basics | 15 hours | Bridging concepts to code |
| Hugging Face Course | Transformers, LLMs, fine-tuning | 20 to 30 hours | Modern AI practitioners |
| DeepLearning.AI Short Courses | Targeted practical skills | 1 to 2 hours each | Just-in-time learning |
| Anthropic Claude Skills (free training) | Building with Claude | 5 to 10 hours | LLM application developers |
| fast.ai Practical Deep Learning | Deep learning, top-down approach | 40 to 60 hours | Career-track learners |
Tools You Will Use (All Free Tier)
Get accounts and basic familiarity with these. All have a free tier sufficient for learning.
GitHub for code hosting and version control. Learn git init, git add, git commit, and git push and you are 80 percent of the way.
Google Colab for free GPU compute. You do not need to buy a GPU to learn deep learning; Colab gives you one in the browser.
Hugging Face for models, datasets, and Spaces (free hosting for demo apps).
OpenAI, Anthropic, or Google AI Studio for API access. OpenAI gives free credits for new accounts; Anthropic and Google offer free tiers on Claude and Gemini that are generous enough to learn on. Pick one provider and focus.
VS Code for editing. The Cursor and Windsurf editors are also free for individuals and ship with built-in AI assistance that doubles your learning speed once you know basic Python.
Use the AI you are learning to learn the AI. Pick one model (Claude or ChatGPT) and treat it as your tutor. When a course confuses you, paste the confusing paragraph in and ask "explain this to me as if I have only finished week 3 of a Python course." When your code breaks, paste the error in and ask for the most likely cause. This single habit cuts the typical 90-day path to roughly 60 days for most learners.
Common Beginner Mistakes
Three failure patterns dominate.
The first is tutorial purgatory. You finish course after course, never build anything original, and after six months you cannot answer "what did you actually make?" The cure 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, between Claude and GPT, between LangChain and direct SDK calls. None of those choices matter at month 1. Pick whatever the course you are taking uses, and switch later if you have a real reason.
The third is skipping math. You do not need a math degree, but you do need basic comfort with vectors, matrices, derivatives, and probability. If you have never seen these, work through 3Blue1Brown's free YouTube series on linear algebra and calculus. Two hours of his videos is worth a semester of textbook.
Building a Portfolio That Gets You Hired or Funded
If your goal is a job or a freelance income, the courses are necessary but not sufficient. What gets attention is a public portfolio of three to five small, working projects with clear writeups.
Each project should answer: what problem does this solve, how does it work technically, what did I learn, and what would I do differently. Hiring managers spend 30 seconds on each repo. Make those 30 seconds count.
Publish on GitHub with a clean README and a deployed demo (Hugging Face Spaces or Vercel are free and easy). Cross-post the writeup on Medium, LinkedIn, or your own site for SEO. The combination of repo plus deployed demo plus writeup is what converts applicants to interviews in 2026.
What to Do After Day 90
If you finished the path and shipped a project, you are now ahead of most "AI engineers" who got the title in 2024 and have not built anything since. Choose a specialization based on what you enjoyed:
- LLM application development: go deeper on RAG, agent frameworks (LangGraph, CrewAI), and evaluations.
- Machine learning engineering: go deeper on the fast.ai full course, then learn MLOps fundamentals (MLflow, model serving).
- Research: go deeper on Stanford CS224N (NLP) and CS231N (vision) on YouTube; both are free.
- AI for business: go deeper on workflow automation tools (Make.com, n8n, Zapier) and building agents that solve specific business problems.
The first 90 days is the hardest because every concept is new. Months 4 through 12 compound; you will pick up new techniques in days that used to take weeks.
FAQs
Do I need a math or computer science background to learn AI?
No. Elements of AI requires no math or programming, and the path above is built for absolute beginners. You will need to pick up basic Python and some comfort with vectors and probability along the way, but both are addressable inside the 90-day window. The people who fail at learning AI usually fail because they never build, not because they lacked the math.
How much money do I really need to spend?
Zero, if you stick to the free resources in this guide. The only optional spend is API credits beyond the free tiers (typically $5 to $20 a month if you build something interesting) and a domain name if you want to deploy a project under a custom URL.
Should I learn machine learning or generative AI first?
In 2026, generative AI is where most of the new jobs and the highest leverage live, but you cannot fully understand it without a base in machine learning. The path above gives you the ML foundation in month 1 and 2, then pivots to LLMs and generative AI in month 2 and 3. Skipping the ML base means you can use APIs but cannot debug them when they fail.
Can I get an AI job after 90 days of self-study?
For most people, no. Ninety days plus one shipped project gets you to junior-friendly territory but not to a senior or specialized role. To convert to a paid role, plan on 6 to 12 months total: the 90-day path plus 3 to 6 months of building progressively harder projects, contributing to open source, and writing publicly about what you learn. The full year still costs zero dollars.
Which programming language should I learn?
Python. Python is the dominant language for AI and machine learning, every major library and framework is Python-first, and the Hugging Face ecosystem assumes it. JavaScript is useful later if you build front-end interfaces, but Python is non-negotiable for the work itself.
What about AI for non-technical careers?
If your goal is to use AI in marketing, sales, operations, or another non-engineering role, you can compress the path. Skip months 1 and 2 of programming-heavy material and instead focus on Elements of AI, the DeepLearning.AI short courses on prompt engineering and RAG, and hands-on projects in tools like Make.com, n8n, ChatGPT Custom GPTs, or Claude Projects. Six weeks instead of 12, with measurable workplace impact at the end.
