Choose your AI course: five university resources and where to start

Use this guide to choose a starting point, open the right free materials and complete one small learning session. You'll also get a prompt that turns your background into a specific study route using the courses in the carousel.

Choose the question you want to answer

If Python functions, loops and reading an error message are still unfamiliar, start with the relevant prerequisite. Choosing a foundation that fits your level will make the university material much easier to use.

The public resources below are free to study. That doesn't make you an enrolled university student or include every certificate, assessment service, cloud computer or current classroom recording.

1. Harvard CS50 AI: build while learning

Open CS50's Introduction to Artificial Intelligence with Python. Harvard provides seven weeks of OpenCourseWare and recommends CS50x or at least a year of Python experience. The subjects include search, uncertainty, optimisation, learning, neural networks and language.

Start with Search. Watch a section, then explain the idea before tackling its associated project. Follow the course's own setup and academic-honesty requirements if you submit work for feedback. Use AI assistance only in ways the course permits.

Try this first: draw a small network of six places and paths between them. Pick a starting point and destination. Trace the order in which you would explore the locations without revisiting the same place indefinitely. This is an original warm-up, not a supplied assignment answer.

You understand the session when: you can explain what information the search procedure keeps and why the order of exploration matters.

Use the OpenCourseWare route for public material. The site separately explains certificates, edX submission access and credit options.

2. Stanford CS231n: understand image models

The CS231n course page covers deep learning for computer vision. Its public course notes give you a practical place to begin with image classification and neural networks. Python, NumPy, calculus, linear algebra and basic probability are expected.

Open the image-classification notes before the advanced architecture sections. Follow how a labelled collection of images becomes a classification problem. Pay attention to the separation between training, validation and test data.

Try this first: choose a small imaginary application, such as distinguishing product photos from screenshots. Describe what counts as each label and list three ambiguous examples. Then explain how images from the same photoshoot could accidentally make a test set too similar to the training set.

You understand the session when: you can explain why you need examples the model hasn't trained on, even if its training result looks excellent.

The course page links older public lecture recordings. Current Canvas recordings are restricted to enrolled students, so use the public links rather than expecting a university login.

3. MIT 6.S191: connect lectures to code labs

MIT's Introduction to Deep Learning publishes lectures, slides and code labs. The 2026 material includes music generation and language-model fine-tuning. The stated maths prerequisites are derivatives, matrix multiplication and the chain rule; Python familiarity is helpful.

Start with the introductory lecture, then open Software Lab 1 from the same course page. Read the notebook's setup cells and instructions before running the whole thing. Work through one section and identify its input, trainable settings and output.

Try this first: after a small example runs, change one setting and describe what you expect it to affect. Keep the original output so you can compare them. If you're unsure what a setting means, ask for an explanation before changing several at once.

You understand the session when: you can point to the data, the model and the loss in the example, and explain their different jobs.

The materials are public. A hosted notebook's available computing resources and usage limits are a separate matter. Start with the provided small examples before spending money on hardware.

4. Stanford CS229: work through the reasoning

Use the public Autumn 2018 CS229 syllabus and notes. This archive includes linear regression, classification, model selection, learning theory and review material. The archive year matters: you're studying enduring foundations, not joining the current classroom.

Begin with the linear-algebra and probability review links if the notation feels unfamiliar. Then open the supervised-learning notes and work through a tiny regression example. Later, the bias/variance and advice-on-applying-ML materials help you think about why a model fails.

Try this first: use the calculation below to connect a prediction rule to its error. Once it makes sense, compare the notation in the course notes with your own example.

You understand the session when: you can explain the difference between a parameter the model learns and a setting you choose for training.

Some old assignment or video links may require university services. The direct public notes are the intended entry point here.

5. Stanford CS224N: follow the language-model foundations

CS224N focuses on natural language processing with deep learning. The official page links a free complete 2024 lecture playlist. It expects Python, linear algebra, calculus and probability, with machine learning knowledge making the material easier to follow.

Start with word representations before jumping to the most advanced language-model lectures. Track how a text input becomes a numerical representation and how the training task defines what the model is being asked to predict.

Try this first: take the sentences “I deposited money at the bank” and “We sat beside the river bank”. Explain why the same written word needs context. Then connect that observation to the representations discussed in the lecture.

You understand the session when: you can describe the model's input, prediction target and training signal without using “it just understands” as the explanation.

Use the public playlist and materials. Paid professional programmes and current enrolled-student recordings are separate offerings.

A worked example: connect a prediction to its error

Suppose your practice data is (x, y) = (1, 2), (2, 4), (3, 6). These are made-up learning examples.

Try the rule prediction = 1.5 × x. It predicts 1.5, 3, 4.5. Subtract the target from each prediction to get -0.5, -1, -1.5. Square those errors to get 0.25, 1, 2.25. Their average is 3.5 / 3, about 1.167.

Now try prediction = 2 × x. The average squared error is zero on these three examples. You've just compared two parameter choices using a loss function. You haven't shown that the second rule works on real-world data. That would require a meaningful dataset and appropriate unseen examples.

This small calculation makes the later lecture notation much less mysterious. The letters are describing actions you can already follow.

Get a course route matched to your starting point

Paste this into Claude and replace the three background lines. If Claude can't open a page, paste its relevant course description instead.

Help me choose one AI course and produce my first three study sessions.

My Python experience: [what I can currently build]
My maths comfort: [algebra / vectors / derivatives / probability]
What I want to understand or build: [specific outcome]
Time per session: [minutes]

Choose from Harvard CS50 AI (https://cs50.harvard.edu/ai/), Stanford CS231n (https://cs231n.stanford.edu/ and https://cs231n.github.io/), MIT 6.S191 (https://introtodeeplearning.com/), the public 2018 CS229 notes (https://cs229.stanford.edu/syllabus-autumn2018.html), and CS224N's public materials (https://web.stanford.edu/class/cs224n/).

Read accessible official descriptions. Recommend one primary course and explain the fit. Identify any prerequisite gap before recommending a lecture. For each session, give the exact linked starting material, one concept, an original worked example and one question that checks understanding. Give the worked solution after the question so I can try it first.

Keep free materials distinct from paid certificates, cloud compute and enrolled-student access. Don't invent lesson titles, completion promises or links. Respect the course's academic-honesty policy and don't solve assessed assignments for me.

Keep the learning manageable

If you spend a whole session fighting an installation, switch to the lecture and a paper example, then return to setup with the exact error message. If the equations lose you, identify the first unfamiliar operation and practise that one. Your next step is to open one course and complete one example you can explain back, then build from there.