InstructGPT was the breakthrough that showed RLHF could align language models to follow human instructions safely. Background: GPT-3 was powerful but often unhelpful, verbose, or produced harmful content. Didnt follow instructions well. Approach: Fine-tune GPT-3 using RLHF (Reinforcement Learning from Human Feedback). Three-step process. Step 1 - SFT: Supervised fine-tuning on human-written demonstrations of helpful responses. Step 2 - RM: Train reward model on human comparisons of model outputs (which response is better). Step 3 - PPO: Use reward model to provide feedback signal for reinforcement learning (Proximal Policy Optimization). Results: 1.3B InstructGPT preferred over 175B GPT-3 despite 100x fewer parameters. More helpful, less harmful. Key insights: Human feedback more valuable than scale alone. Smaller aligned models beat larger unaligned ones. Impact: Foundation for ChatGPT (InstructGPT + dialogue), established RLHF as standard for LLM alignment. Legacy: Every major LLM now uses instruction tuning and human feedback. Transformed how LLMs are deployed.