AI For Beginners

What it is

AI For Beginners is a free curriculum from Microsoft, kept as a GitHub repository under the MIT license. Its README calls it a 12-week, 24-lesson course with practical lessons, quizzes and labs. The lessons start with the history of AI and symbolic AI, then cover neural networks, computer vision and language models. They end with genetic algorithms, reinforcement learning, multi-agent systems and AI ethics, plus an extra lesson on multi-modal networks, CLIP and VQGAN. The code lives in Jupyter notebooks, and the README says the course covers tools like TensorFlow and PyTorch. Where a lesson offers both, the setup page says you can follow whichever framework you like.

1. Fork the repository, then clone it
2. Build the ai4beg environment with conda
3. Take a lesson’s pre-lecture quiz and read its text
4. Run its notebook and try the challenges
5. Take the post-lecture quiz, then do the lab

Who it suits

Learners who want a structured path through AI and can read a little Python. The README calls the curriculum beginner-friendly, and the first of its beginner examples lists Python basics as the prerequisite. The README also says its coverage of recent models may be a bit lacking in the state of the art.

Skip it if you want the newest techniques or a business view of AI. The README leaves out business cases, classic machine learning, cloud ML services, chat bots and the deep mathematics, and names other places to learn them.

Where to use it

The curriculum is Markdown lessons, Jupyter notebooks and labs. The How to Run page says you need to run Python code in its Jupyter notebooks, and gives local, container and cloud routes.

• Your own computer: Miniconda with the ai4beg environment, then VS Code or Jupyter in your browser
• A dev container: the repository’s .devcontainer folder, which VS Code offers to reopen the code in
• GitHub Codespaces: a browser-based VS Code started from the Code button
• Binder: free cloud Jupyter, with some web access blocked and basic compute
• A cloud GPU: a Data Science Virtual Machine, an Azure Machine Learning workspace or Google Colab
• GitHub itself: read the text and look through the code without running anything

The page says model training can be painfully slow without a GPU, so the later lessons are where a cloud GPU helps most.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with the steps from Microsoft’s setup pages. The main route is a clone and a conda environment. The README tells you to fork the repository with the Fork button at the top right, and then clone it.

1. Install Miniconda.
2. Clone the repository and create the environment, as below.
3. Open the folder in VS Code and accept the Python extension it suggests, or start Jupyter in the folder.
4. Open a lesson’s README, then its notebook.

git clone http://github.com/microsoft/ai-for-beginners
cd ai-for-beginners
conda env create --name ai4beg --file .devcontainer/environment.yml
conda activate ai4beg

The repository holds translations in 50+ languages, which makes a clone large. The README gives a sparse checkout that leaves them out.

git clone --filter=blob:none --sparse https://github.com/microsoft/AI-For-Beginners.git
cd AI-For-Beginners
git sparse-checkout set --no-cone '/*' '!translations' '!translated_images'

How to use it

Start at Lesson 01 or anywhere you like. Take a lesson’s pre-lecture quiz, read its text and open its notebooks. In VS Code you open the .ipynb file. For Jupyter in the browser, start it in the course folder:

jupyter notebook

Then open any .ipynb file and work in it. The README’s table lists no notebook or lab for Lesson 01. The lesson itself has a pre-lecture and a post-lecture quiz, a challenge and an assignment, Game Jam. Labs, where a lesson has one, sit in that lesson’s lab folder.

Example

You have an evening and a laptop with Python. You clone the repository and run the first beginner example from the examples folder:

cd examples
python 01-hello-ai-world.py

The script teaches a tiny model the rule y = 2x from five examples. It prints its progress every 20 rounds, then tests the model on four new numbers. The starting weight is random, so only the initial guess line changes from run to run. By round 20 the weight prints as 2.00 and all four tests match the true answer. Change learning_rate (line 30 of the file) and run it again to see what training does.

For newcomers and experts

If you are new. A Jupyter notebook is a page that mixes text with code you run one cell at a time. conda keeps each project’s Python packages in their own environment, so this course’s packages do not clash with your other work. A neural network, the subject of Lessons 03 to 05, is a model that learns its settings from examples. Lesson 01 has a section on weak AI versus strong AI. If setup feels heavy, start with the four beginner examples, or read the lessons on GitHub with nothing installed.

For experts. The environment file pins matplotlib 3.9, numpy 1.26, requests 2.32 and scipy 1.13. It takes PyTorch with torchtext, torchvision and torchdata from the pytorch channel and OpenCV from conda-forge. The root environment.yml reads the same as .devcontainer/environment.yml, and its pip step installs the requirements.txt beside it, which pins tensorflow 2.17.0, keras 3.15.0, gym 0.26.2 and tensorboard 2.17.1.

Not every lesson has both frameworks. The README’s table lists TensorFlow only for Lessons 11 and 19, and PyTorch only for Lesson 20. Lesson 08’s TensorFlow link opens Lesson 05’s IntroKerasTF.ipynb. The dev container uses Microsoft’s Miniconda image and builds the same environment from environment.yml after it is created. The quiz app in etc/quiz-app is a separate npm project: npm install, then npm run serve.

Good to know

The notebooks run on your machine or on the cloud service you pick. The lessons’ quiz links open the online quiz app at ff-quizzes.netlify.app, unless you run the quiz app yourself. Some notebooks fetch models or datasets from the public internet when they run, and the How to Run page says Binder blocks some of that.

• Cost: the How to Run page calls Binder free and says Google Colab comes with some free GPU support. It gives no prices for Codespaces, Azure VMs or Azure Machine Learning workspaces
• Windows: if conda fails on gdk-pixbuf with a gbk codec error, set PYTHONUTF8 to 1 and try again, as the How to Run page says
• Pages disagree: the Getting Started page says its quiz app holds 50 quizzes, and the quiz app README says 40. The README says 24 lessons, while its table also lists a 25th in Extras. Its online quiz link and the Getting Started page’s link point to different hosts
• Age: large language models get one lesson in the table, Lesson 20, and the README says the course may lag the state of the art
• Tracking links: the README’s Microsoft Learn links carry tracking IDs, so a click carries them
• Removal: the pages give no uninstall steps. Deleting the folder is Mindket’s suggestion. The How to Run page removes the environment with conda deactivate, then conda env remove --name ai4beg
• License: MIT. The LICENSE file reads “Copyright (c) Microsoft Corporation.”