The best free AI resource is the one that matches your current prerequisites and ends in work you can test. For machine-learning foundations, begin with Google’s Machine Learning Crash Course. For code-first deep learning, evaluate fast.ai. For academic theory, use MIT OpenCourseWare. For transformer and LLM practice, use Hugging Face. For short notebook exercises, use Kaggle Learn.
This is not a ranking. The five official destinations returned HTTP 200 on July 17, 2026, but “free” can describe learning material without guaranteeing a certificate, unlimited compute, permanent access or support.
Compare the Five Verified Resources
| Resource | Best fit | Check before starting | Evidence to produce |
|---|
| Google ML Crash Course | Core ML concepts and exercises | Programming, algebra and current module prerequisites | Completed exercises plus one baseline model |
| fast.ai Practical Deep Learning | Code-first deep-learning projects | Python experience, setup and compute requirements | Notebook, evaluation and error examples |
| MIT OpenCourseWare 6.034 | Classical AI concepts and academic depth | Older course date, mathematical level and time commitment | Solved problems and concept notes |
| Hugging Face LLM Course | Transformers, datasets and LLM tooling | Python and deep-learning prerequisites by chapter | Small model or dataset workflow with model card |
| Kaggle Learn | Short notebook-based practice | Account, notebook environment and course availability | Completed notebook adapted to a new dataset |
The resource is input. Your tested artifact is the learning evidence.
Path A: You Do Not Code Yet
Do not begin by copying a large model tutorial you cannot explain. Start with a no-code evaluation project:
- Pick one narrow task, such as extracting a deadline from an email.
- Write 20 fictional examples, including missing and ambiguous deadlines.
- Define the expected answer before using an AI assistant.
- Record output, pass/fail and failure type in a spreadsheet.
- Change one instruction and rerun the same examples.
You will learn inputs, expected behavior, evaluation and error analysis—the same habits needed later in code. Then learn enough Python to load a CSV, transform values and calculate a metric.
Use the beginner project briefs to keep the first deliverable bounded.
Path B: You Know Basic Python
Start with Google Machine Learning Crash Course or a short Kaggle Learn sequence. Your first coded project should contain:
- a documented dataset source;
- a train/test or time-based split where appropriate;
- a simple baseline;
- one metric chosen before tuning;
- several failed examples;
- an environment or dependency file;
- a README explaining how to reproduce the result.
Avoid jumping between five courses. Finish one small classification, regression or evaluation task first.
Path C: You Want Practical Deep Learning
fast.ai’s course is designed around building. Read the current official prerequisites rather than relying on an old blog summary. For every lesson you use:
- Run the provided example.
- Replace the dataset or task with a permitted alternative.
- State what changed.
- Compare the result with a naive baseline.
- Save failure examples and limitations.
A notebook that only reproduces the instructor’s output is practice, not independent portfolio proof.
Path D: You Want LLM and Transformer Skills
The Hugging Face LLM Course provides a structured entry into transformer tooling. Use its current chapter order and prerequisites. A credible learning artifact might include:
- dataset provenance and license;
- preprocessing steps;
- selected model and task;
- held-out evaluation;
- inference examples;
- error categories;
- a model or project card documenting intended and excluded uses.
Do not upload private, licensed or personal data merely because a notebook accepts file uploads.
Path E: You Want Deeper AI Theory
MIT OpenCourseWare’s 6.034 materials offer classical artificial-intelligence topics. The course date matters: use it for durable concepts and problem solving, not as a source of current product or framework instructions.
For each topic, write three things:
- the problem the method solves;
- a small worked example;
- a situation where the method is the wrong choice.
Theory becomes useful when you can connect it to a decision in a real system.
A Six-Week Self-Taught AI Roadmap
Week 1: prerequisites and baseline
Choose one path, test the required environment and write a small baseline artifact. Record what you already know and what blocks you.
Week 2: core lessons
Complete only the lessons needed for your first deliverable. Keep a glossary of concepts you can explain with an example.
Week 3: minimum project
Build the smallest end-to-end version: one input, one transformation or model, one output and one metric.
Week 4: evaluation
Create held-out or difficult cases. Save failures. Compare against a simple baseline and resist changing several variables at once.
Week 5: reproduction
Re-run the project from a clean environment. Add dependencies, setup steps, a fixed seed where relevant and sample data others may legally use.
Week 6: portfolio evidence
Publish only what you may share. Include the problem, data, method, metric, failures, limitations and next experiment. A short reproducible project is stronger evidence than a collection of completion screenshots.
How to Evaluate a Free Resource
Score each candidate before investing weeks:
| Criterion | Question |
|---|
| Provenance | Is this the current official page? |
| Prerequisites | Can I meet the stated coding, math and hardware requirements? |
| Deliverable | Will I build or solve something inspectable? |
| Evaluation | Does the path teach testing and failure analysis? |
| Access | Which material, compute, support and credentials are actually free? |
| Freshness | Are framework instructions still current? |
| Safety | Can I practice without exposing private data or credentials? |
Reject a resource when its main promise is a salary, guaranteed job, universal timeline or unexplained certificate value.
What “Free” Does Not Prove
Free learning access does not automatically include:
- a verified certificate;
- hosted GPU time without limits;
- mentor support;
- permanent access;
- permission to reuse every dataset;
- suitability for your region or accessibility needs;
- readiness for a specific job.
Confirm current terms on the official destination. This guide intentionally does not publish hard-coded quotas or certificate prices.
Portfolio Checklist
- Narrow problem and intended user
- Data source, permission and exclusions
- Baseline and chosen method
- Test split or evaluation set
- Metric and error examples
- Reproduction commands
- Limitations and unsafe uses
- No secrets or private data
- One screenshot or demo tied to the measured result
If coding-agent practice fits your path, use the OpenAI Codex tutorial to apply the same evidence standard to repository changes.
Choose one resource, write the six-week deliverable before beginning, and evaluate progress by what you can reproduce—not by how many videos you watched.
