AI Safety & Alignment

A grounded introduction to **AI safety and alignment**: what it means to get AI systems to reliably do what we intend, and why that turns out to be hard. We cover how today's language models are actually aligned (SFT, RLHF, constitutional methods, red-teaming, and evals), where those methods fall short, and how to reason about risks without hype. By the end you'll be able to spot reward hacking, tell near-term harms from speculative ones, and explain why oversight and evaluation remain open problems.