How Does an LLM Work?
Understanding the Complex: What actually happens when you give a machine a question and it answers? 11 lessons, ~150 minutes, free.
How Does an LLM Work?
All 11 lessons in order.
What's This Actually About?
A scene from the night ChatGPT launched — and the central question this course answers: what actually happens when you give a machine a question and it answers?
Why Should I Care?
Three concrete reasons this topic is worth 150 minutes: LLMs are already embedded in your life, key decisions are being made right now, and the public debate is badly calibrated.
The Basics You Need
Four concepts you need before the mechanism lessons: tokens, neural networks, training, probability distributions — and the anchor example that runs through Lessons 4–6.
How Does It Work? (Part 1): From Text to Numbers
How a sentence becomes tokens, tokens become vectors, and vectors travel through 96 transformer layers. Part 1 of the mechanistic explanation.
How Does It Work? (Part 2): Attention
Why the model knows *France* matters more than *the*. The attention mechanism explained: queries, keys, values, multi-head attention.
How Does It Work? (Part 3): Knowledge from Statistics
How the model 'knows' that Paris is the capital of France — when nobody told it directly. Training, RLHF, and emergence. Full resolution of the anchor example.
Who's Doing This? Why? Who's Paying?
The people: Altman, Hassabis, Amodei, Hinton, LeCun, Gebru, DeepSeek. The money. Three competing development models: US VC, Chinese state, European regulatory.
What's Contested? What Don't We Know?
Three live controversies: do LLMs understand? Is emergence real? Does scaling lead to AGI? What's overhyped, underhyped, and what nobody actually knows.
What's Next?
Three timelines to never confuse. Five concrete milestones to watch. And what would genuinely surprise me — tools for tracking the field.
What If...?
Calibrated speculation in both directions: three positive, three negative scenarios — each with a condition analysis. What would need to be true for each to happen?
What Do You Take Away?
Six core ideas you should be able to explain. Honest limits. Where to read next. And why you shouldn't trust us — or anyone — blindly.