🎯 Core Goals
- Introduce the LLM as an input/output black box.
- Build curiosity about what happens inside with a predefined guess quiz.
- Bridge to the sandwich model in 3.2.
Text goes in. Text comes out. That’s the whole interface — but what’s actually happening inside that box? Let’s find out.
👁️ Visuals & Interactives
💬
Your Message
Text In
→
📦
LLM
??? What's inside?
→
🤖
Reply
Text Out
What's actually happening inside that box?
👇 Next up: If it just predicts words... how does it "remember" your whole conversation? That's what we explore next.
📝 Key Concepts
- The Black Box: From the outside, an LLM is simple — text in, text out. The chat interface hides all the machinery.
- Next-Token Prediction: The LLM’s core job is to predict the most statistically likely next word/token, given everything before it. It does this repeatedly until the reply is complete.
- Not a Search Engine: LLMs don’t look things up in real-time. They generate text from learned patterns — powerful, but also why they can confidently say things that are wrong.
- The Interface Illusion: The chat app makes it look like a flowing conversation. What actually happens each turn? That’s what we explore next.
If it just predicts words… how does it “remember” your whole conversation? That’s the trick — every time you hit Send, something surprising happens under the hood. Let’s look at it next.