Skip to content

Chapter 00 · 2 min read · Updated

Start Here: Planning and Reasoning, Explained Simply

Five papers trace the climb from a model that answers instantly to one that reasons step by step, acts with tools, checks its work, and learns by practice.

Prefer one idea per screen? Read it as 5 slides →

In folder 01 you learned how a language model is built, scaled, aligned, and measured. The result was a model that is fluent and knowledgeable. But fluent is not the same as smart. A model can write a beautiful sentence and still get a two-step math problem wrong, because answering instantly, in one breath, is not how hard problems get solved.

This folder is about the leap from a model that answers to a model that thinks. The five papers here trace how researchers taught models to reason step by step, to act in the world by using tools, to check their own work, to learn reasoning through trial and error, and finally to handle problems far larger than their memory.

00.1

How to use these explanations

Read the chapters in order; each one builds on the last. Bold linked terms like chain of thought jump to the glossary, where every key idea is explained from scratch. If a term feels like it belongs to folder 01 (such as how a Transformer works), you will find it in that folder's glossary; this folder's glossary focuses on reasoning.

You do not need to have memorized folder 01, but it helps to know one thing from it: a model is fundamentally a next-word predictor. Everything in this folder is a clever way of getting more careful thinking out of that simple engine.

00.2

The big picture

The story of this folder is a climb up five steps, each one giving the model a new thinking power.

flowchart TD
    A[1. Think out loud<br/>Chain of Thought] --> B[2. Think and act<br/>ReAct, tool use]
    B --> C[3. Check the work<br/>Verify step by step]
    C --> D[4. Learn to reason by practice<br/>DeepSeek-R1]
    D --> E[5. Think beyond memory limits<br/>Recursive Language Models]

Here is how the five papers map onto that climb.

StepNew powerPaper
Think out loudShow the steps, get better answersChain-of-Thought Prompting
Think and actMix reasoning with using toolsReAct
Check the workReward each step, not just the final answerLet's Verify Step by Step
Learn by practiceUse reinforcement learning to grow reasoningDeepSeek-R1
Beat memory limitsProcess inputs far bigger than the context windowRecursive Language Models

00.4

One idea to hold in your head

A model that answers instantly is using a fixed, small amount of effort per question, no matter how hard the question is. The single thread running through this entire folder is the idea of spending more effort on harder problems: more steps, more tool calls, more checking, more practice, more recursion. Once you see that pattern, every paper here becomes a different answer to the same question, "how do we let a model think harder when it needs to?"

Ready? Start with Chapter 1: Chain of Thought.