Your 30-day RL learning roadmap
How AI learns to reason, use tools, and complete real tasks. One practical lesson a day, in six stages.
Start with one idea
During reinforcement learning, a model makes attempts, receives reward-derived feedback, and is updated through a training algorithm. Simply retrying a task does not necessarily update the model.
The six stages
- 1Understand learningDays 1–6RL, pretraining, demonstrations, updates, agents and context
- 2Understand feedbackDays 7–12Rewards, preferences, checkable outcomes, PPO, GRPO and DPO
- 3Understand practiceDays 13–18Environments, rollouts, resets, verifiers, reward hacking and difficulty
- 4Understand capabilitiesDays 19–24Coding, tools, screens, conversations, long tasks and reasoning
- 5Judge the evidenceDays 25–28Generalization, setup failures, reliability and computational budgets
- 6Put it togetherDays 29–30An inspectable environment and the whole training pipeline
Your daily checklist
Each lesson is a short video. Work down the list as they come out.
Stage 1 · Days 1–6
Understand learning
- 01How AI learns without seeing the solution
- 02Three ways an AI model learns
- 03Show the answer or score the attempt?
- 04What actually changes when AI learns?
- 05How AI chooses its next action
- 06Did it learn, or read your correction?
Stage 2 · Days 7–12
Understand feedback
- 07What should we reward an AI for?
- 08Who decides which answer is better?
- 09When a test can grade the AI
- 10Why training updates need care
- 11Why AI tries the same question repeatedly
- 12Learning directly from preferences
Stage 3 · Days 13–18
Understand practice
- 13Where AI practises real tasks
- 14One task, different attempts
- 15Why every attempt needs a clean start
- 16The AI says done. How do we check?
- 17Full marks, wrong result
- 18Choosing tasks that teach something
Stage 4 · Days 19–24
Understand capabilities
- 19How coding agents practise fixing bugs
- 20Access to tools is only the beginning
- 21How AI finds the right button
- 22When the user changes their mind
- 23Which step caused the failure?
- 24Thinking longer versus training differently
Stage 5 · Days 25–28
Judge the evidence
- 25Skill or a leaked exam?
- 26Did the model fail, or did the setup break?
- 27One success versus dependable success
- 28What did that successful attempt cost?
Stage 6 · Days 29–30
Put it together
- 29Build a practice task you can inspect
- 30How an AI agent gets trained, end to end
After each episode, ask
- 1What is the task?
- 2What does the system observe?
- 3What can it do?
- 4How is success checked?
- 5Is this training or evaluation?
By the end, you should be able to explain the pipeline and inspect a small task setup. This introduction does not replace implementation practice or a full RL course.