A weekly roadmap + prerequisites + practice plan
If you use ChatGPT, Claude, or Gemini every day and you want to understand what is actually happening inside LLMs, this Stanford series is one of the cleanest paths I have found.
This README gives you:
- All 9 lecture links (working)
- A simple 9-week plan (1 lecture per week)
- Prerequisites you should know before you start
- A practice workflow so you learn by building, not just watching
Rule: don’t binge. One lecture per week + hands-on notes + one small build.
https://www.youtube.com/watch?v=Ub3GoFaUcds
https://www.youtube.com/watch?v=yT84Y5zCnaA
https://www.youtube.com/watch?v=Q5baLehv5So
For each lecture, do this:
- Watch the lecture (no multitasking)
- Write 7 to 10 bullet notes in your own words
- Ask an LLM to quiz you (and explain what you got wrong)
- Build one tiny thing (30 to 90 minutes)
- Publish something small (optional but powerful)
- a short LinkedIn post
- a short note
- a tiny Kaggle notebook
- a mini GitHub repo
This is how you go from “I watched it” to “I can apply it”.
You do not need to be perfect. You need basic comfort.
You should be comfortable with:
- functions, loops, list and dicts
- reading files, working in notebooks
- basic debugging (print, exceptions)
Good free resources:
- Kaggle Learn: Python
https://www.kaggle.com/learn/python - Official Python tutorial
https://docs.python.org/3/tutorial/
Minimum practice (do this once before Week 1):
- write a small script that loads text, cleans it, counts words, prints top 20
You do not need heavy math. You need the practical pieces.
Linear algebra:
- vectors and matrices
- dot product
- cosine similarity (used in embeddings)
- matrix multiplication shape logic
Probability:
- probability basics
- expectation and variance intuition
- cross entropy as “how wrong your predicted probabilities are”
Good free resources:
- 3Blue1Brown Linear Algebra (visual)
https://www.3blue1brown.com/topics/linear-algebra - StatQuest (great for intuition)
https://www.youtube.com/@statquest
Minimum practice:
- implement cosine similarity in Python
- compare 3 vectors and find nearest neighbor
You should understand:
- train / validation / test
- overfitting vs underfitting
- loss, optimization (basic concept)
- metrics
Good free resources:
- Google Machine Learning Crash Course
https://developers.google.com/machine-learning/crash-course
Minimum practice:
- train a simple model once (even logistic regression) and understand what changed
You should know:
- what a layer is
- what backprop means (conceptually)
- embeddings as learnable vectors
- attention at a high level (the series will teach it deeper)
Good free resources:
- fast.ai (practical)
https://course.fast.ai/ - CS231n notes (strong fundamentals)
https://cs231n.github.io/
Minimum practice:
- use PyTorch to fit a tiny model on dummy data (even random points)
Week 1 to 3:
- build a tiny tokenizer and count token frequency
- implement attention on a toy example and print attention weights
- implement cosine similarity and retrieval for embeddings
Week 4 to 7:
- compare full fine-tuning vs LoRA (read, summarize, count parameters)
- write a small report: SFT vs RLHF vs DPO in your own words
- replicate a tiny alignment toy example (conceptual is fine)
Week 8:
- build a toy RAG:
- store 5 docs
- retrieve top 2
- answer based on retrieved text
Week 9:
- evaluate outputs:
- create 20 test prompts
- score with simple rubric
- compare two models
Lecture #: Title
Date:
- 1 lecture per week
- 2 short practice sessions in the week (45 minutes each)
- 1 note writeup session (30 minutes)
You will finish in 9 weeks with real understanding.
- Stanford Online / Stanford University for making the content public
- Afshine Amidi + Shervine Amidi for the curriculum and teaching
- Credit for the original share that triggered this README goes to Raunak Kumar’s LinkedIn post