Articles / llmvideosfree-resources
Learn how LLMs work from free videos: a weekend plan
Seven free videos that explain how LLMs work, split into a Saturday and a Sunday, plus a nine-video track for people who want to build a GPT in code.
You can learn how a large language model works in a weekend without writing any code. Seven videos on YouTube cover it well, and they come from two channels: 3Blue1Brown, which explains the maths with animation, and Andrej Karpathy, who helped build these systems and explains how they are trained.
This plan follows the How LLMs work, in a weekend path on GyaanStack Videos. Total watching time is about eight hours. At the end there's a second track, Build a GPT yourself, for when you want to write the code.
Saturday: what's inside the model#
About two and a half hours. Watch in this order, because each video uses words the previous one introduced.
1. Large Language Models explained brieflyThe big picture in seven minutes. If you watch only one video on this list, watch this one.
2. [1hr Talk] Intro to Large Language ModelsOne hour that ties pretraining, fine-tuning and security together. It's a talk, not a lecture, so it moves quickly and doesn't need maths.
3. Transformers, the tech behind LLMsWhat a transformer does with your words, from tokens to embeddings to the prediction of the next token.
4. Attention in transformers, step-by-stepHow attention lets words take meaning from each other. This is the core idea in every modern LLM, and the animation makes it much easier to follow than the paper.
5. How might LLMs store factsWhere facts live inside the model. After video 4 you know how words share information; this one covers the layers that hold what the model knows.
Sunday: how it's trained and how to use it#
About five and a half hours, in two long videos. Split each one across the day.
6. Deep Dive into LLMs like ChatGPTThe full pipeline from internet text to a chat assistant: collecting data, tokenizing, pretraining, then the post-training that turns a text predictor into an assistant. It goes deeper than video 2 on every step.
7. How I use LLMsHow an expert uses LLMs day to day. After six videos about how the model works, this one is about getting good work out of it.
A short check before you move on#
If you can answer these in your own words, the weekend worked:
- What is a token, and why does the model see your text as tokens and not letters?
- What does attention do, in one sentence?
- What's the difference between a pretrained model and the assistant you chat with?
- Why can a model state something false with full confidence?
Write the answers down. Explaining it on paper shows the gaps faster than rewatching.
If you only have two hours#
Watch video 1 (7 min), video 2 (59 min) and video 4 (26 min). Together they give you the big picture, the training story and the one mechanism that makes transformers work. You can do the long Sunday videos later, one chapter at a time.
Who this is for#
The weekend plan suits anyone who uses LLMs at work and wants to know why they behave the way they do: product managers, analysts, and engineers who haven't worked on models before. None of the seven videos needs code. If you already know what attention is, skip to Sunday.
Track two: build a GPT yourself#
This is the Build a GPT yourself path on GyaanStack Videos. It's for people who know some Python and want to write every piece. It's nine videos and about sixteen hours, so plan for a few weeks of evenings, not a weekend. Videos 3 and 4 from Saturday are steps 5 and 6 here, so you can skip them.
Pictures first
- But what is a neural network? (3Blue1Brown, 18 min). See what a neural network is before touching code.
- Gradient descent, how neural networks learn (3Blue1Brown, 20 min). The idea behind every training loop.
Then code
- The spelled-out intro to neural networks and backpropagation: building micrograd (Andrej Karpathy, 2 h 25 min). Write backpropagation by hand. Everything later rests on it.
- The spelled-out intro to language modeling: building makemore (Andrej Karpathy, 1 h 57 min). Your first language model, one character at a time.
- Transformers, the tech behind LLMs (3Blue1Brown, 27 min). Skip if you watched it on Saturday.
- Attention in transformers, step-by-step (3Blue1Brown, 26 min). Skip if you watched it on Saturday.
- Let's build GPT: from scratch, in code, spelled out. (Andrej Karpathy, 1 h 56 min). The GPT in code, line by line.
- Let's build the GPT Tokenizer (Andrej Karpathy, 2 h 13 min). Tokenizers are the hidden cause of many LLM quirks.
- Let's reproduce GPT-2 (124M) (Andrej Karpathy, 4 h 1 min). Train a real GPT-2 the way labs do. Expect to need a GPU for the full training run.
Type the code yourself while you watch instead of copying the finished repo, and run it after each section.
Where to go next#
If you want to apply this instead of going deeper into the model, the LLM engineering roadmap picks up from here with prompting, RAG and agents. The written resources for these topics are in the GyaanStack LLM foundations and LLM subjects.