How LLMs Work¶
An LLM does one thing: predict the next token. Everything you experience — answers, code, apparent reasoning — is that loop run repeatedly. Understanding the loop explains the strengths, the costs, and the failure modes.
flowchart LR
J["Junior: the loop and parameters"] --> M["Middle: distillation and quantization"]
M --> S["Senior: versions and drift"]
S --> P["Professional: model policy"]
Levels¶
| Level | Guide | You are done when |
|---|---|---|
| Junior | The loop and parameters | You can explain next-token prediction, what "7B" measures, and why the model doesn't learn from your conversation. |
| Middle | Distillation and quantization | You can explain why a distilled small model beats a generic one and the memory/quality trade of quantization. |
| Senior | Versions and drift | You can protect an app from silent model upgrades and design for cheap model swaps. |
| Professional | Model policy | You can write an org-wide policy for approved models, pinning, and upgrade evaluation. |
Practice rule¶
When output surprises you, don't ask "why is it being difficult" — ask "what tokens did it see, and what is the most probable continuation of those tokens." The loop explains almost every weird answer.
Related¶
- Tokens and Context — the unit the loop operates on, and what each one costs.
- Temperature and Sampling — how the "pick the next token" step gets its randomness.