How it works

What’s inside, and what it costs. Four slides.

What an LLM is

TRAINING: HOW THE MODEL IS MADE 1 Pre-training Reads the internet Predicts the next word 2FeedbackLearns from examplesPeople rank answers 3RLReinforcement learning:tries, gets checked,rewarded when right WHEN YOU USE IT4ThinkingWorks it out firstLonger: better,slower, costs more

Training makes the model. Thinking happens on your question: the longer it thinks, the better the answer, and the more of your plan it uses.

Prompt engineering (RIP)

Kojima et al. (2022), Large Language Models are Zero-Shot Reasoners
  • 2022 trick: add “let’s think step by step”, and maths goes from 18% to 79%.
  • Now it’s built in: that’s the thinking stage.
  • Magic words and templates? Mostly noise now.

What an agent is

  • An LLM with tools: it reads and writes files, runs programs, searches the web.
  • On your computer, in the folder you start it in.
  • It asks before it does anything big. You say go.
  • A chat window only sees what you paste into it. The agent works on the files themselves.

Claude Code and Codex are agents.

Usage limits

Every answer costs computing power. The strong models and long thinking cost the most.

  • Claude Pro runs out sooner within a five-hour window. ChatGPT Plus runs out sooner over a week. Higher plans give many times more.
  • What eats it: long sessions, big files, the strongest model, long thinking.
  • See where you stand: /usage in Claude Code, /status in Codex.
  • Hit the limit? Wait for the reset, switch to a Fast model with /model, or pair up. Hitting it every day? A higher plan is the fix.