Start with a short visual model of LLMs, then go deeper into training, attention, systems, and hands-on practice.
Start here. This short 3Blue1Brown explainer gives the cleanest first mental model: a chatbot completes a conversation by repeatedly predicting likely next tokens from the current context.
7:58 · 3Blue1Brown · Short visual foundation
A chatbot answer can be framed as completing a dialogue, one token at a time.
The model scores many possible next tokens, then the system chooses from those scores.
Allowing less likely tokens can make answers feel natural, but also less repeatable.
Pretraining learns broad patterns; later feedback helps shape assistant behavior.
Attention lets each token draw on the other tokens in its context, weighted by relevance; in chat models, a prediction can never use tokens that come after it.
Now that you know LLMs predict the next word, where do they learn those patterns? This video shows the pretraining process: how an enormous amount of text gets curated, cleaned, split into tokens, and fed to GPUs for months.
15:53 · Attention Visualized · 3.8K views
A model that learns by next-token prediction: it covers up the next token (a word or word piece) and guesses it, across its whole tokenized training text.
Most raw web text is filtered out as duplicate, spam, or low quality; only a fraction becomes training text. The exact amounts differ by model.
Thousands of GPUs running for months. No teacher — just text.
The deepest dive. Self-attention is the core mechanism that lets each token (a word or word piece) draw on the other tokens in its context, weighted by relevance; in chat models, a prediction can never use tokens that come after it. This is what makes modern AI different from everything that came before.
13:01 · Visual AI · 19K+ views
Each token queries the other tokens in its context (in chat models, never later ones) to understand meaning.
Query, Key, Value — three projections that compute relevance between words.
Self-attention solved the long-range dependency problem that held AI back for years.
Now that you understand how LLMs work, see how engineers build real products on top of them. This Stanford course summary covers the full technology stack from base models to multi-agent systems.
27:24 · Gary Chen · 58K+ views · Based on Stanford's "Beyond LLM" course
No domain knowledge, outdated info, hard to control, struggles with long context.
A core skill for every AI user. Includes two usage styles: delegating whole tasks vs. collaborating step by step.
Retrieval-Augmented Generation — give the model access to your own documents.
Train the model on your data to specialize it for specific tasks.
LLMs that can plan, use tools, and take actions — not just answer questions.
Multiple AI agents collaborating on complex tasks, each with a specialized role.
Now that you understand the theory, try training a real AI model yourself. Google's Teachable Machine lets you build an image, sound, or pose classifier in minutes — no coding required. This short video shows how it works in a classroom setting.
5:50 · STEM Learning By Doing · Advanced / Optional
Collect examples, label them, and watch the model learn the patterns.
See where the model gets confused — that's how you build intuition for AI limits.
Go to teachablemachine.withgoogle.com and train your own classifier.